Modelling seasonal habitat suitability and connectivity ...

Modelling seasonal habitat suitability and connectivity

for feral pigs in northern Australia:

towards risk-based management of infectious animal diseases

with wildlife hosts

Jens G. Froese

M.A., Albert-Ludwigs-Universität Freiburg, Germany

A thesis submitted for the degree of Doctor of Philosophy at

The University of Queensland in 2017

School of Agriculture and Food Sciences

i

Abstract

Infectious animal diseases are a major biosecurity threat in an increasingly connected world.

Wildlife hosts are a well-recognized risk factor for disease introduction, establishment and spread.

Northern Australia is vulnerable to disease incursions from neighbouring countries, and widespread

invasive feral pigs (Sus scrofa) can seriously complicate post-border disease management. The aim

of this thesis was to generate new regional-scale spatial knowledge of feral pig populations in

northern Australia to inform risk-based management of directly transmitted infectious animal

diseases for which feral pigs are a host.

Due to environmental variability and empirical knowledge gaps across this vast region, I

adopted a resource-based modelling approach, based on expert knowledge but rooted in landscape

ecological theory, to answer three research questions at multiple levels of biological organisation.

Specifically, I conceptualized feral pigs in northern Australia as a metapopulation and the landscape

as displaying a patch-corridor-matrix structure. At the level of individual feral pig breeding herds, I

explored the selection of supplementary and complementary resources within home ranges. At the

level of local subpopulations, consisting of several herds with adjacent or overlapping home ranges,

I used a habitat suitability modelling approach to investigate the distribution of potential patches of

breeding habitat emerging from the interactions between resources and home range movements. At

the metapopulation level, I examined potential dispersal pathways between many such patches

using a habitat connectivity modelling approach. As feral pig movements and distributional patterns

vary with conditions, I applied models to two seasonal scenarios (wet and dry) corresponding to

northern Australia’s annual rainfall cycle.

This thesis contributed methodological advances and new ecological insights. I developed a

novel combined methodology, spatial pattern suitability analysis, for capturing feral pigs’ resource-

seeking home range movements based on expert-elicited response-to-pattern curves and spatial

moving window analysis. Based on landscape ecological principles, this methodology improves the

application of resource-based Bayesian networks models to mobile animals. I found that habitat

suitability for persistent feral pig breeding in northern Australia is dependent on spatial interactions

between four key habitat requirements: water and food resources as well as protection from heat

and from disturbance. Through scenario analysis and empirical validation I showed that habitat

suitability at the regional scale is most reliably modelled as a function of distance to supplementary

and complementary resource patches. When applied to a wet season and a dry season scenario,

mapped model results indicated that the spatial distribution of feral pig habitat patches varied

markedly. Importantly, empirically validated findings suggest that dry season conditions restrict

overall habitat suitability for feral pig breeding in northern Australia more than previously thought.

ii

This thesis also provides the first attempt to describe seasonal habitat connectivity in the entire

northern Australian feral pig population. By linking model assumptions to gender-specific

differences in dispersal ability, I showed that dry season connectivity between habitat patches was

limited for breeding herds, but less constrained for solitary males. Due to greater resource

abundance, wet season connectivity was greater irrespective of dispersal ability. Three broad types

of habitat patches were identified: some were always isolated; some were highly connected to form

large habitat components; and some were mostly isolated but became connected to large

components during the wet season or for wide-ranging males.

By linking the landscape ecological research perspective on feral pigs to a landscape

epidemiological perspective on directly transmitted infectious diseases, risk areas for the

establishment, spread and persistence of disease in wildlife hosts could be identified. I used the

example of classical swine fever to illustrate these links. Following introduction, successful

establishment of such a disease is contingent on locally dense host populations in patches of

breeding habitat; subsequent disease spread requires host dispersal between infected and susceptible

subpopulations; and long-term disease persistence depends on a lasting supply of susceptible

individuals within a regionally connected host metapopulation. Effective post-border disease

management should capitalize on these links. For example, early detection surveillance activities

could be targeted in habitat patches and designed so that each connected habitat component is

sampled. In the event of an incursion, patch connectivity may be used to establish containment

zones, focus population control and support delineation of disease-free compartments. Moreover,

host risk could be combined with disease-specific introduction pathways, transmission rates and

other parameters to generate deeper, dynamic insights into disease-host interactions for better

incursion preparedness.

In conclusion, the research contained in this thesis provides, for the first time, a complete and

coherent, spatially-explicit, seasonally-specific and regional-scale picture of areas most at risk of

disease establishment (via host habitat suitability) and spread (via host habitat connectivity) in feral

pigs in northern Australia. The resource-based modelling approach is transparent and flexible, and

could be applied to other invasive species and wildlife hosts of infectious animal diseases,

especially in data-constrained situations and for wide-ranging species.

iii

Declaration by author

This thesis is composed of my original work, and contains no material previously published or

written by another person except where due reference has been made in the text. I have clearly

stated the contribution by others to jointly-authored works that I have included in my thesis.

I have clearly stated the contribution of others to my thesis as a whole, including statistical

assistance, survey design, data analysis, significant technical procedures, professional editorial

advice, and any other original research work used or reported in my thesis. The content of my thesis

is the result of work I have carried out since the commencement of my research higher degree

candidature and does not include a substantial part of work that has been submitted to qualify for

the award of any other degree or diploma in any university or other tertiary institution. I have

clearly stated which parts of my thesis, if any, have been submitted to qualify for another award.

I acknowledge that an electronic copy of my thesis must be lodged with the University Library and,

subject to the policy and procedures of The University of Queensland, the thesis be made available

for research and study in accordance with the Copyright Act 1968 unless a period of embargo has

been approved by the Dean of the Graduate School.

I acknowledge that copyright of all material contained in my thesis resides with the copyright

holder(s) of that material. Where appropriate I have obtained copyright permission from the

copyright holder to reproduce material in this thesis.

iv

Publications during candidature

Peer-reviewed publications

Froese, JG, Smith, CS, Durr, PA, McAlpine, CA & van Klinken, RD (2017). "Modelling seasonal

habitat suitability for wide-ranging species: invasive wild pigs in northern Australia." PLoS

ONE: e0177018. doi: 10.1371/journal.pone.0177018

Froese, JG, Murray, JV, Perry, JJ & van Klinken, RD (2015). "Spatial modelling to estimate the risk

of feral pigs to pig farm biosecurity in south-eastern Australia." Animal Production Science

55(12): 1456.

Conference proceedings

Froese, JG, Smith, CS, Durr, PA & van Klinken, RD (2016). "Integrating landscape structure into

participatory Bayesian network models of habitat suitability." in S Sauvage, JM Sánchez-Pérez

& AE Rizzoli (eds). Proceedings of the 8th International Congress on Environmental

Modelling and Software. Toulouse: p. 838.

Froese, JG, Smith, CS, McAlpine, CA, Durr, PA & van Klinken, RD (2015). "Modelling habitat

suitability and connectivity of feral pigs for exotic disease surveillance in northern Australia."

in T Weber, MJ McPhee & RS Anderssen (eds). MODSIM2015, 21st International Congress

on Modelling and Simulation. Modelling and Simulation Society of Australia and New

Zealand, Broadbeach: p. 279.

Froese, JG, Smith, CS, McAlpine, CA, Durr, PA & van Klinken, RD (2015). "Moving window

analysis links landscape-scale resource utilization to habitat suitability models of feral pigs in

northern Australia." in T Weber, MJ McPhee & RS Anderssen (eds). MODSIM2015, 21st

International Congress on Modelling and Simulation. Modelling and Simulation Society of

Australia and New Zealand, Broadbeach: pp. 1352-1358.

Froese, J, Smith, C, Durr, PA & van Klinken, RD (2014). "Where can all the pigs be found?

Harnessing expert knowledge for the spatial modelling of feral pig distribution and abundance

in northern Australia." in MN Gentle (ed). Program and abstracts 16th Australasian

Vertebrate Pest Conference. Brisbane: p. 70.

v

Scientific reports

Murray, J, Froese, J, Perry, J, Navarro Garcia, J & van Klinken, R (2015). Impact modelling for

rabbits and feral pigs in the QMDB. CSIRO Biosecurity Flagship, Brisbane.

Murray, JV, Froese, J, Perry, J & van Klinken, R (2015). Assessing the risk of feral pigs interacting

with domestic pig herds in southeastern Australia. Final report to Australian Pork Limited

(APL), Project 2114/486. CSIRO Biosecurity Flagship, Brisbane.

vi

Publications included in this thesis

Froese, JG, Smith, CS, Durr, PA, McAlpine, CA & van Klinken, RD (2017). "Modelling seasonal

habitat suitability for wide-ranging species: invasive wild pigs in northern Australia." PLoS

ONE: e0177018. doi: 10.1371/journal.pone.0177018 – incorporated as Chapter 4

Contributor Statement of contribution (PPL 4.20.04 Authorship)

Jens Froese (Candidate) Conceived idea and designed methodology (40%)

Collected data (90%)

Analysed data (100%)

Wrote the paper (80%)

Carl Smith Conceived idea and designed methodology (20%)


Clive McAlpine Conceived idea and designed methodology (10%)


Peter Durr Conceived idea and designed methodology (10%)

Collected data (10%)


Rieks van Klinken Conceived idea and designed methodology (20%)


vii

Contributions by others to the thesis

A panel of feral pig experts not only helped parameterize the models contained in this thesis, they

also fundamentally shaped the research design and methodology by highlighting the need to take a

landscape ecological view on feral pigs at the outset of the project. Individual experts are

acknowledged by name in the following section. Human ethics approval to conduct expert

elicitation was obtained from the CSIRO Human Ethics Committee and the UQ School of

Agriculture and Food Sciences Ethics Committee (Appendix 1). Justin Perry (CSIRO) contributed R

scripts and helped with access to CSIRO high performance computing facilities needed for some of

the spatial analyses. Justine Murray (CSIRO) contributed expert elicitation methodologies on which

some of the data collection was based and co-facilitated an expert workshop.

Statement of parts of the thesis submitted to qualify for the award of another degree

None

viii

Acknowledgements

Embarking on a PhD is a time of great enthusiasm and anticipation of what lies ahead. Many

experienced advisors – university representatives, supervisors, colleagues and fellow PhD students

– offer much needed guidance about the inevitable challenges in such a long-term project and the

bumpy path that lies ahead in the journey towards completion. I have now been on this path for the

best part of four years. At times, it has been cruisy. At times, it has been exhilarating. At other

times, it has been boring. A few times it has been so dreadfully bumpy that I wanted to turn back.

On the journey I have learned a lot about myself both as an aspiring scientist and as a person. Most

of all, I have learned that there are a lot of people travelling along with you. You may be driving the

car, but others are reading the map, paying for the fuel and providing you with company. I have

enjoyed the support of many wonderful people, while being conscious that I have given back little. I

can only offer my heartfelt gratitude and hope to repay the debts in the personal and professional

life that lies ahead.

Ahead of all, I would like to deeply thank my family: my wonderful wife Kieran for allowing

me to focus on my work when it was most needed while keeping our children alive, advancing her

own career as the best midwife in Brisbane and looking fine at the same time; and my beautiful

daughters Flora, Roxy and Juno for making parenting look easy. I love you all and forever will.

Completion would not have been possible without my outstanding and generous advisory team.

Carl Smith, my principal advisor, has a door that is literally always open! Best of all, whatever the

issue – a technical glitch, methodological question or existential crisis – he always seemed to have

an answer or idea on how to tackle it. Rieks van Klinken was the ‘conceptual engine’ behind my

research, always reminding me of the broader picture when I got bogged down yet again in some

minor detail. He also provided me with invaluable opportunities outside the thesis in his research

team at CSIRO. Peter Durr and Clive McAlpine had a somewhat smaller role in this team, but I was

continuously surprised how much effort both were willing to invest in providing advice, revising

manuscripts, and even attending expert workshops, even though we saw each other so little. I

sincerely thank you all, Carl, Rieks, Peter and Clive, for your support and guidance.

Next, I must offer my deepest gratitude to a panel of feral pig experts, who kindly donated a

significant amount of time to share their knowledge and without whom this research would not have

been possible. Thank you Jim Mitchell (FeralFix), Justin Perry (CSIRO), Col Dollery (Queensland

Parks and Wildlife Service), Andrew Hartwig (Cape York Landcare), Scott Middleton (Gulf

Catchments Biosecurity & Agribusiness Innovation), Tim Kerlin (Northern Australia Quarantine

Strategy), Peter Caley and Cameron Fletcher (both CSIRO), Cassandra Wittwer and Joe Schmidt

(both Northern Australia Quarantine Strategy), Les Harrigan (Rinyirru Land Trust), Glen Sheppard

ix

(Cook Shire Council), Bart Dryden and Jamie Livingston (Terrain Natural Resource Management),

Travis Sydes (Far North Queensland Regional Organisation of Councils), Dave Berman, Darren

Marshall and Vanessa MacDonald (Queensland Murray Darling Committee). I also thank the

Balkanu Cape York Development Corporation, Queensland Parks and Wildlife Service, Northern

Australia Quarantine Strategy and Atlas of Living Australia who shared data used for model

validation, and the Queensland Murray Darling Committee for employing my services on a related

project. It has been great fun and wonderfully rewarding working with you guys.

Further thanks go to my colleagues and fellow researchers: Justin Perry, Russell Richards and

Miguel Angel Villamil Castro helped with R analyses. Justine Murray helped with expert

elicitation. Justin Perry, Justine Murray, Javi Navarro, Sam Nicol and Hawthorne Beyer kindly

reviewed my manuscripts. Tony Pople agreed to chair all my milestone review meetings and Bill

Ellis and Shu Fukai offered their feedback on one milestone occasion each. Last but not least, Grant

Hamilton from the Quantitative Applied Spatial Ecology Group at the Queensland University of

Technology, thank you for trusting in my ability to embark on a full-time position as Research

Associate while working towards completion of my PhD thesis, and for giving me the space that I

needed in the final weeks.

Finally, I would like to thank all funding bodies who supported my research, foremost The

University of Queensland and CSIRO, who topped up an Australian Postgraduate Award with a

generously resourced Integrated Natural Resource Management scholarship. Additional support in

attending two amazing Australian Pig Science Association Conferences in Melbourne was provided

by Australian Pork Limited – thank you for showing such interest in my work.

P.S. And thank Sus scrofa for being such an awesome, interesting study species! I know you

shouldn’t be here in the Australian environment. I also hate to say that my research may contribute,

in a small way, to the effective killing of many of you. Although I haven’t handled you, and only

seen a few, after years of computer modelling I feel that I know you intimately! That is largely

thanks to the many knowledgeable scientists who have studied you before me and the experienced

practitioners who continue to handle you in the wild. For these insights I am deeply grateful.

x

Keywords

Sus scrofa, landscape ecology, infectious animal disease, wildlife host, habitat suitability, habitat

connectivity, Bayesian network, Circuitscape, expert elicitation, spatial analysis, northern Australia

Australian and New Zealand Standard Research Classifications (ANZSRC)

ANZSRC code: 050103, Invasive Species Ecology, 25%

ANZSRC code: 050104, Landscape Ecology, 50%

ANZSRC code: 070205, Animal Protection (Pests and Pathogens), 25%

Fields of Research (FoR) Classification

FoR code: 0501, Ecological Applications, 75%

FoR code: 0702, Animal Production, 25%

xi

Table of Contents

Abstract ................................................................................................................................................ i

Table of Contents .............................................................................................................................. xi

List of Figures .................................................................................................................................. xvi

List of Tables .................................................................................................................................. xvii

List of Abbreviations .................................................................................................................... xviii

Chapter 1 Introduction ................................................................................................................... 19

1.1 Research problem ................................................................................................................... 19

1.1.1 Biosecurity threats from infectious animal diseases with wildlife hosts ..................... 19

1.1.2 Infectious animal diseases in feral pigs ....................................................................... 20

1.1.3 Risk-based disease management in northern Australia ............................................... 21

1.1.4 A landscape ecological and epidemiological research perspective ............................. 22

1.2 Research aim ........................................................................................................................... 23

1.3 Research questions .................................................................................................................. 23

1.4 Study region ............................................................................................................................ 23

1.5 Research approach .................................................................................................................. 25

1.6 Thesis structure ....................................................................................................................... 27

1.7 Research ethics approval ........................................................................................................ 29

Chapter 2 Literature review .......................................................................................................... 30

Summary ............................................................................................................................................ 30

2.1 Theoretical concepts in landscape ecology ............................................................................. 30

2.1.1 Habitat suitability ........................................................................................................ 30

2.1.2 Patch-corridor-matrix model and metapopulation ecology ......................................... 32

2.1.3 Scaling in space and time ............................................................................................ 33

2.1.4 Landscape heterogeneity and home ranges ................................................................. 35

2.1.5 Habitat connectivity .................................................................................................... 36

2.2 Landscape ecology of feral pigs ............................................................................................. 37

2.2.1 Distribution and abundance ......................................................................................... 37

2.2.2 Social organisation and reproduction .......................................................................... 38

2.2.3 Resource selection and home range movements ......................................................... 39

xii

2.2.4 Metapopulation structure and dispersal ....................................................................... 40

2.3 Landscape epidemiology of infectious diseases in feral pigs ................................................. 41

2.4 Habitat suitability models ....................................................................................................... 43

2.4.1 Modelling frameworks ................................................................................................ 43

2.4.2 Bayesian networks ....................................................................................................... 46

2.4.3 Landscape variables .................................................................................................... 47

2.5 Habitat connectivity models ................................................................................................... 49

2.5.1 Modelling frameworks ................................................................................................ 49

2.5.2 Habitat patches ............................................................................................................ 50

2.5.3 Matrix resistance ......................................................................................................... 51

2.6 Expert elicitation ..................................................................................................................... 52

Chapter 3 Integrating landscape structure improves habitat models of mobile animals: feral

pigs in Australia ............................................................................................................................... 54

Summary ............................................................................................................................................ 54

3.1 Abstract ................................................................................................................................... 54

3.2 Introduction ............................................................................................................................. 55

3.3 Methods .................................................................................................................................. 58

3.3.1 Study region ................................................................................................................ 58

3.3.2 Modelling approach ..................................................................................................... 59

3.3.3 Resource quality models ............................................................................................. 61

3.3.4 Spatial pattern suitability analysis ............................................................................... 61

3.3.5 Habitat suitability model ............................................................................................. 63

3.3.6 Model evaluation and validation ................................................................................. 64

3.4 Results..................................................................................................................................... 65

3.4.1 Spatial pattern suitability analysis ............................................................................... 65


3.5 Discussion ............................................................................................................................... 70

3.5.1 Ecological significance................................................................................................ 70

3.5.2 Benefits of the modelling approach ............................................................................. 72

3.5.3 Limitations................................................................................................................... 73

3.5.4 Conclusion ................................................................................................................... 74

3.6 Appendices ............................................................................................................................. 74

xiii

Chapter 4 Modelling seasonal habitat suitability for wide-ranging species: invasive feral pigs

in northern Australia ....................................................................................................................... 75

Summary ............................................................................................................................................ 75

4.1 Abstract ................................................................................................................................... 75

4.2 Introduction ............................................................................................................................. 76

4.3 Materials and methods ............................................................................................................ 77

4.3.1 Study region ................................................................................................................ 77

4.3.2 Habitat suitability model ............................................................................................. 78


4.4 Results..................................................................................................................................... 91

4.4.1 Sensitivity analysis ...................................................................................................... 91

4.4.2 Predictive performance................................................................................................ 91

4.4.3 Seasonal habitat suitability .......................................................................................... 94

4.5 Discussion ............................................................................................................................... 96

4.5.1 Seasonal habitat suitability .......................................................................................... 97


4.5.3 Management implications ......................................................................................... 100

4.6 Appendices ........................................................................................................................... 101

Chapter 5 Modelling habitat connectivity for biosecurity: the risk of infectious disease spread

in feral pigs in northern Australia ................................................................................................ 102

Summary .......................................................................................................................................... 102

5.1 Abstract ................................................................................................................................. 102

5.2 Introduction ........................................................................................................................... 103

5.3 Materials and methods .......................................................................................................... 105

5.3.1 Study region and species ........................................................................................... 105

5.3.2 Habitat connectivity model ....................................................................................... 105

5.3.3 Model evaluation ....................................................................................................... 110

5.4 Results................................................................................................................................... 111

5.4.1 Matrix connectivity ................................................................................................... 111

5.4.2 Patch connectivity ..................................................................................................... 113

5.5 Discussion ............................................................................................................................. 118

xiv

5.5.1 Matrix connectivity ................................................................................................... 118

5.5.2 Patch connectivity ..................................................................................................... 119

5.5.3 Implications for disease management ....................................................................... 121

5.6 Appendices ........................................................................................................................... 122

Chapter 6 General discussion and conclusions .......................................................................... 123

Summary .......................................................................................................................................... 123

6.1 Research question 1 – resource selection by individual breeding herds .............................. 123

6.1.1 Main findings ............................................................................................................ 123

6.1.2 Significance and advances......................................................................................... 124

6.2 Research question 2 – seasonal habitat patches for subpopulations ..................................... 125

6.2.1 Main findings ............................................................................................................ 126


6.3 Research question 3 – seasonal patch connectivity for metapopulations ............................. 128

6.3.1 Main findings ............................................................................................................ 128


6.4 Synthesis and applications .................................................................................................... 130

6.4.1 Contributions to the thesis aim .................................................................................. 130

6.4.2 Applications to risk-based disease management ....................................................... 133

6.5 Limitations and future research ............................................................................................ 136

6.5.1 Empirical research ..................................................................................................... 136

6.5.2 Improved habitat suitability and connectivity models .............................................. 138

6.5.3 Spatiotemporal dynamics .......................................................................................... 140

6.5.4 Optimization models ................................................................................................. 141

6.5.5 Other study systems and applications ....................................................................... 142

References ....................................................................................................................................... 143

xv

Appendices ...................................................................................................................................... 165

Appendix 1 Human ethics approval ................................................................................................ 165

Appendix 3.1 Model variables – parameters .................................................................................. 167

Appendix 3.2 Methodology – spatial pattern suitability analysis ................................................... 170

Appendix 3.3 Methodology – validation ........................................................................................ 180

Appendix 3.4 Validation plots – model scenarios .......................................................................... 190

Appendix 4.1 Model variables – parameters (supplements Fig 4.2) .............................................. 191

Appendix 4.2 Model variables – spatial data (supplements Table 4.2) .......................................... 209

Appendix 4.3 Methodology – spatial pattern suitability analysis ................................................... 214

Appendix 4.4 Methodology – validation ........................................................................................ 220

Appendix 4.5 Validation maps – seasonal habitat suitability ......................................................... 229

Appendix 4.6 Additional analyses – seasonal habitat suitability .................................................... 231

Appendix 5.1 Methodology – omnidirectional current density ...................................................... 233

Appendix 5.2 Methodology – patch connectivity ........................................................................... 239

Appendix 5.3 Analysis – tests of significance ................................................................................ 241

xvi

List of Figures

Fig 1.1 Study region in northern Australia.................................................................................... 24

Fig 1.2 Schematic illustration of the multi-level research approach. .......................................... 26

Fig 1.3 Thesis structure.. ................................................................................................................. 29

Fig 2.1 The link between a mobile species’ behavioural levels and corresponding spatial and

temporal scales of analysis. ............................................................................................................. 34

Fig 2.2 Approaches to modelling species distributions, or ‘habitat’, in geographical and

environmental space. ....................................................................................................................... 43

Fig 3.1 Resource-seeking home range movements of mobile animals. ........................................ 56

Fig 3.2 Study region in north-eastern Australia............................................................................ 58

Fig 3.3 Methodology for modelling habitat suitability for feral pig breeding. ........................... 59

Fig 3.4 Computing resource suitability indices using spatial pattern suitability analysis. ....... 63

Fig 3.5 Averaged expert-elicited response-to-pattern curves relating structural patterns of

resource quality to functional suitability indices. ......................................................................... 66

Fig 3.6 Predicted-to-expected (P/E) ratio and corresponding habitat suitability maps for three

alternative models in the Lakefield area. ....................................................................................... 69

Fig 4.1 Study region in northern Australia.................................................................................... 78

Fig 4.2 Feral pig habitat suitability model. .................................................................................... 79

Fig 4.3 Expert-elicited resource suitability in response to distance. ........................................... 88

Fig 4.4 Sensitivity of habitat suitability to four habitat variables and expert opinion. ............. 91

Fig 4.5 Validation plots for individual expert and averaged seasonal habitat suitability

models. ............................................................................................................................................... 93

Fig 4.6 Seasonal habitat suitability for feral pig breeding in northern Australia. ..................... 94

Fig 4.7 Share of modelled suitable habitat found in different vegetation types and land use

classes for each seasonal scenario. .................................................................................................. 96

Fig 5.1 Study region in northern Australia.................................................................................. 105

Fig 5.2 Methodology for modelling habitat connectivity for feral pigs. .................................... 107

Fig 5.3 Validation plots for seasonal habitat suitability models. ............................................... 109

xvii

Fig 5.4 Seasonal matrix connectivity in northern Australia (for moderate ‘matrix resistance

(ResDisp)’). ...................................................................................................................................... 112

Fig 5.5 Seasonal patch connectivity for feral pigs in northern Australia. ................................ 115

Fig 5.6 Interactions between ‘corridor delineation threshold (CDThresh)’ and ‘inter-patch

dispersal distance (DispDist)’ at each level of ‘matrix resistance (ResDisp)’ for the dry (a-c)

and wet (d-f) season........................................................................................................................ 115

Fig 5.7 Seasonal patch connectivity for two selected habitat patches of feral pigs in northern

Australia. ......................................................................................................................................... 118

List of Tables

Table 2.1 Sources of bias in expert elicitation and strategies or techniques for minimizing

them. .................................................................................................................................................. 53

Table 3.1 Validation data sets with ancillary information. .......................................................... 64

Table 3.2 Performance of the eleven habitat suitability models against three validation data

sets. .................................................................................................................................................... 67

Table 4.1 Bayesian network model variables and their states, with definitions......................... 84

Table 4.2 Spatial data proxies linked to model explanatory variables and methods for

reclassifying data attributes into state-specific categories. .......................................................... 86

Table 4.3 Validation data sets with ancillary information. .......................................................... 90

Table 4.4 Validation metrics for individual expert and averaged seasonal habitat suitability

models. ............................................................................................................................................... 92

Table 4.5 Amount of feral pig habitat in each habitat suitability index class per seasonal

scenario. ............................................................................................................................................ 95

Table 5.1 Amount of dispersal habitat in the study region. ....................................................... 113

xviii

List of Abbreviations

Abbreviation Full term

ALA Atlas of Living Australia

ALUM Australian land use and management classification

AUSVETPLAN Australian Veterinary Emergency Plan

BN Bayesian network

CBI Continuous Boyce Index

CPT Conditional probability table

CSF Classical Swine Fever

CSIRO Commonwealth Scientific and Industrial Research Organisation

DOI Digital object identifier

ESRI Environmental Systems Research Institute

FGDB File geodatabase

GIS Geographic information system

HSI Habitat suitability index

MVG / MVS Present major vegetation groups / subgroups

NAQS Northern Australia Quarantine Strategy

OCD Omnidirectional current density

P/E ratio Predicted-to-expected ratio

SDM Species distribution model

UQ The University of Queensland

19

Chapter 1 Introduction

1.1 Research problem

1.1.1 Biosecurity threats from infectious animal diseases with wildlife hosts

Infectious animal diseases are a major biosecurity1 threat in an increasingly connected world.

Disease epidemics may seriously impact human, livestock, and wildlife health and incur large costs

on industry and government. Many infectious diseases exist in a host-pathogen continuum

involving wildlife, domestic animals, and humans (Daszak et al. 2000). Over 60% of newly

reported infectious diseases in humans between 1940 and 2004 were zoonotic, and most of these

originated in wildlife (Jones et al. 2008). For example, bats are important reservoir hosts of several

dangerous emerging pathogens such as the rabies, Nipah or Hendra viruses (Calisher et al. 2006).

Migratory birds may contribute to global spread of highly pathogenic avian influenza H5N1, with

devastating consequences to human health and the poultry industry (Olsen et al. 2006; Altizer et al.

2011). Where humans are not implicated in the host-pathogen continuum, emerging animal diseases

may have significant economic consequences, both directly on production and indirectly on market

access and trade. For example, 6.5 million animals were slaughtered in the United Kingdom in

response to an outbreak of foot-and-mouth disease in 2001 and combined costs to industry and

government were estimated at ₤8 billion (National Audit Office 2002).

Australia is currently free from many of these ‘emergency animal diseases’ (Emergency

Animal Disease Response Agreement 2012). Disease-free status has been attributed to both

geographical isolation and a well-functioning, multi-layered biosecurity system (Beale et al. 2008).

This is highly beneficial to Australia’s agricultural industries in terms of livestock welfare,

production advantages and access to competitive domestic and international markets (East et al.

2013; Brookes et al. 2014). However, freedom from disease cannot be guaranteed. Increasing

transboundary movements of humans, animals and animal products carry unavoidable risks of

disease introduction and necessitate effective management of incursions (Beale et al. 2008).

Wildlife hosts may contribute to disease emergence, establishment, spread and persistence and are

recognized as an important focus of risk-based disease management (Daszak et al. 2000; Reisen

2010). In Australia, invasive wildlife poses a particular threat, introducing disease to susceptible

domestic herds of the same species or complicating incursion management in widespread,

uncontrolled and expanding wild populations (Animal Health Australia 2011).

1 Usage of term ‘biosecurity’ varies considerably between disciplines and regions. Here, I adopt an Australian definition of biosecurity as “the protection of the economy, environment and human health from the negative impacts associated with entry, establishment or spread of exotic pests (including weeds) and diseases” (Beale et al. 2008, p. 1).

20

1.1.2 Infectious animal diseases in feral pigs

Feral pigs (Sus scrofa) are one of Australia’s worst invaders and among the continent’s most

abundant and widespread terrestrial mammals (West 2008). The species is a susceptible host to

numerous endemic and exotic animal diseases (Animal Health Australia 2011). I use one of these,

classical swine fever (CSF), to illustrate the role of feral pigs as a risk factor and the impacts of an

incursion. CSF is a highly contagious directly transmitted disease, whose potential introduction is of

major concern to the Australian pork industry (Brookes et al. 2014). CSF is classified as an

emergency animal disease in Australia (Emergency Animal Disease Response Agreement 2012),

and as a notifiable disease internationally (World Organisation for Animal Health 2016). While

Australia is currently free from CSF, the disease is occurring throughout the world, including

neighbouring countries in South East Asia (Animal Health Australia 2012; Leslie et al. 2014).

CSF is caused by a virus of the family Flaviviridae and affects only S. scrofa. The disease is

highly contagious in both domestic and feral pigs and can spread by direct contact between live

animals as well as ingestion of contaminated animal products or transfer of contaminated equipment

(Animal Health Australia 2012; World Organisation for Animal Health 2016). Mortality rates and

clinical signs are highly variable: acute forms have an infective period of two weeks and cause high

mortality and morbidity; low virulent strains, however, may cause chronic disease and prolonged

infectiousness with few clinical signs, enabling lasting disease persistence within a susceptible host

population (Kramer-Schadt et al. 2007; Animal Health Australia 2012).

Large outbreaks of CSF have occurred in Europe in the 1990s, resulting, for example, in the

culling of 11 million pigs and economic losses of over US$2 billion in the Netherlands (Artois et al.

2002). An epidemiological analysis of 327 CSF outbreaks in Germany during the 1990s has

identified infected wild boar as the source of 60% of all primary outbreaks (n = 93) (Fritzemeier et

al. 2000). In Australia, the risks associated with feral pigs during a CSF incursion could pertain to

both disease transmission into domestic herds or establishment within wild populations (Animal

Health Australia 2012). Pearson (2012) found that the risk of exposure by domestic herds to directly

transmitted diseases carried by surrounding feral herds is low but not negligible. However, such a

disease may readily establish and spread within widespread feral pig populations themselves

(Animal Health Australia 2012; Cowled et al. 2012). Then, international obligations would require

comprehensive surveillance of feral pigs, additional measures for separating feral from domestic

herds and active control of the disease in feral pigs through containment, vaccination and culling

(Animal Health Australia 2012; World Organisation for Animal Health 2016). Ultimately, these

combined measures are aimed at demonstrating freedom from disease – across the country, or

within epidemiological ‘zones’ or ‘compartments’ (referring to animal populations that are

21

functionally separated by geographical or management boundaries and can be attributed with a

distinct health status) – so that the safety of trade can be ensured (Scott et al. 2006).

1.1.3 Risk-based disease management in northern Australia

Tropical northern Australia is particularly vulnerable to disease introduction from neighbouring

countries due to its vast coastline, remoteness and traditional transboundary movements (Australian

National Audit Office 2012). For CSF, the most likely pathways of introduction to the North are

illegal importation of infected live animals or animal products from commercial shipping, foreign

fishing or yachting vessels. For other infectious animal diseases, important entry risk factors are

wind-borne insects, migratory birds or contaminated fomites (Cookson et al. 2012; East et al.

2013). Northern Australia also contains large and widespread feral pig populations. Its wetland and

floodplain habitats hold among the highest known local feral pig densities in Australia (Choquenot

et al. 1996). There is a particular need for effective post-border disease management (i.e. actions

aimed at actually or potentially introduced pathogens) in feral pigs at all stages of invasion,

including early detection surveillance, preparedness and incursion response (Beale et al. 2008).

Early detection surveillance is currently managed by the Northern Australia Quarantine

Strategy (NAQS) program (Australian National Audit Office 2012). Surveillance activities are

prioritised according to a risk-based framework that considers the likelihood of disease occurrence

across broad risk areas by combining the risks of introduction, establishment and spread (Cookson

et al. 2012; East et al. 2013). For CSF, surveillance predominantly involves aerial surveys and

representative sampling of feral pigs in high risk areas (Cookson pers. comm.). Incursion

preparedness in Australia has been boosted through the Wildlife Exotic Disease Preparedness

Program, which funded research on disease modelling and diagnostic tools, host ecology, disease

epidemiology and novel techniques for population control between 1999 and 2007 (Henderson

2008). Numerous simulation models have investigated the dynamics of potential disease epidemics,

including CSF, specifically in feral pigs and given important insights on their management (Pech &

Hone 1988; Hone et al. 1992; Doran & Laffan 2005; Milne et al. 2008; Cowled et al. 2012; Leslie

et al. 2014). Response to incursions of emergency animal diseases in wildlife hosts is guided by the

Emergency Animal Disease Response Agreement (2012) and the Australian Veterinary Emergency

Plan (AUSVETPLAN) disease strategies (e.g. for CSF; Animal Health Australia 2012) as well as the

Wild Animal Response Strategy (Animal Health Australia 2011). Response involves a range of

measures including: collecting information about susceptible wild animal populations, carrying out

disease surveillance, preventing spread and containing the disease, controlling susceptible

populations to enable disease fadeout via culling or vaccination, and demonstrating freedom from

disease (Animal Health Australia 2011 & 2012).

22

Across all elements of this multi-layered post-border disease management system, the

importance of feral pig distribution, abundance, dispersal as well as seasonal and behavioural

factors for the establishment and spread of infectious diseases has often been highlighted (Cowled

& Garner 2008; Animal Health Australia 2011). Yet, spatially-explicit, regional-scale2 knowledge

of feral pigs in northern Australia, which duly accounts for seasonal and ecological effects, remains

limited. In this thesis I aimed to fill this knowledge gap.

1.1.4 A landscape ecological and epidemiological research perspective

I adopted a landscape epidemiological perspective on infectious diseases and a landscape

ecological perspective on wildlife hosts. Such an ‘integrated landscape approach’ has been widely

applied in conservation biology but is underutilized in biosecurity research (Glen et al. 2013). In

this framework, I viewed feral pigs in northern Australia as a metapopulation, with local

subpopulations in patches of suitable breeding habitat that are regionally connected by dispersal

corridors3 (Forman 1995; Hanski 1998; Hastings 2014). A metapopulation perspective on feral pigs

in Australia is supported by landscape-genetic evidence of population structuring, especially in

heterogeneous landscapes (Hampton et al. 2004; Cowled et al. 2008; Lopez et al. 2014).

This metapopulation structure also affects the risk of invasion and persistence by directly

transmitted diseases such as CSF (Anderson et al. 1986; Riley 2007). Following introduction,

successful disease establishment is contingent on locally dense breeding populations with high

contact rates that let an infectious pathogen’s basic reproductive rate R0 exceed 1. Subsequent

disease spread requires feral pig dispersal between infected and susceptible subpopulations. Disease

persistence depends on a lasting supply of susceptible individuals within a regionally connected

feral pig metapopulation (Anderson et al. 1986; Kramer-Schadt et al. 2007). The distinction

between rapid local establishment and slower regional disease spread (Cowled & Garner 2008) is

supported by Zanardi et al.‘s (2003) description of two distinct CSF outbreaks in adjacent wild boar

subpopulations separated by unfavourable habitat in Italy, or Hampton et al.’s (2006) analysis of

infectious waterborne pathogens excreted by feral pigs in Western Australia. Simulation models of

disease dynamics in feral pigs have further suggested that seasonal and gender-specific differences

2 Here, ‘regional-scale’ refers to a geographical area that is consistent with the total extent of feral pigs’, or another large mammal species’ natural range (i.e. thousands to millions of km2). Consistent with the usage in the landscape ecological literature (e.g. Mayor et al. 2009; Yackulic et al. 2016, see also Fig 2.1), this is larger than the landscape scale (i.e. tens to hundreds of km2, consistent with the extent of a local population’s range), but smaller than the global scale (i.e. millions to hundreds of millions of km2, exceeding the extent of most species’ ranges). 3 Here, ‘patches’ or ‘habitat patches’ refer to spatially contiguous habitat areas of sufficient quality and size to allow for feral pig breeding. A habitat patch is inhabited by a subpopulation that persists locally independent of other subpopulations within the broader metapopulation. For further explanation of these concepts see Sections 2.1 and 2.2.

23

in feral pig habitat use, home range movements and dispersal ability may affect disease

establishment and spread (Doran & Laffan 2005; Milne et al. 2008).

1.2 Research aim

The aim of this thesis was to generate spatially-explicit, seasonally-specific and regional-scale

knowledge of feral pig populations in northern Australia to inform risk-based management of

directly transmitted infectious animal diseases with feral pigs as a wildlife host.

1.3 Research questions

To achieve this aim, I address three specific research questions through spatially-explicit,

resource-based models at multiple biological levels.

Research question 1

How can habitat suitability for mobile species such as feral pigs be reliably modelled at the

regional scale, given uncertainty about the processes of habitat selection, in particular the resource-

seeking home range movements by individual breeding herds? Answering this question will provide

a suitable methodology for investigating research question 2.

Research question 2

How are patches of suitable feral pig breeding habitat that potentially support locally dense

populations distributed in northern Australia, and is their spatial distribution influenced by seasonal

effects? Answering this question will provide an indicative measure of the seasonal establishment

risk of directly transmitted infectious diseases within feral pigs.

Research question 3

How are patches of suitable feral pig breeding habitat connected by potential dispersal

pathways to form a regionally connected metapopulation, and is their connectivity influenced by

seasonal or gender-specific effects? Answering this question will provide an indicative measure of

the seasonal spread risk of directly transmitted infectious diseases within feral pigs.

1.4 Study region

Our study region extended across tropical Australia north of the Tropic of Capricorn (Fig 1.1).

Similar to a previous study by Cowled et al. (2009), I included all Australian bioregions and

subregions (Department of the Environment 2012) that are fully or partly within agro-climatic

zones I (hot, seasonally wet / dry), J (hot, wet) or H (semi-arid transition to desert climates)

(Hutchinson et al. 2005). The study region encompassed 1.76 million km2 across three jurisdictions:

Queensland, the Northern Territory and Western Australia (Fig 1.1). Northern Australia has a

24

tropical climate, alternating between a wet season (November / December to April / May) and a dry

season (May / June to October / November). Rainfall and primary productivity broadly decline

along a north-south, and to a lesser extent an east-west, gradient (Guerschman et al. 2009; Jones et

al. 2009). Monsoonal savanna woodlands and semi-arid grasslands that are extensively utilized for

cattle grazing are interspersed with riverine channels, coastal wetlands and rainforest fragments

(Fox et al. 2001). Low-lying grasslands are seasonally inundated, transforming into vast floodplains

during the wet season. Intensive uses and population centres are concentrated in fertile coastal

lowlands. Drier inland areas are among the most sparsely populated on earth (Cowled et al. 2009).

Fig 1.1 Study region in northern Australia.

All of the study region appears climatically suitable for feral pigs and has mostly been invaded.

Arid desert regions with insufficient rainfall were not included in this study. Feral pigs are reported

to be widespread in the east and localised in the north and west, where they are expected to expand

their range (Cowled et al. 2009; West 2008). A range of management activities are conducted

throughout northern Australia to mitigate feral pig impacts, including lethal (aerial shooting, poison

baiting, trapping) and non-lethal (exclusion fencing) methods. Yet, efforts are hampered by the

region’s remoteness and inaccessibility. While continuous management has effectively reduced

local densities, there is little evidence of sustained population reduction (Bengsen et al. 2014).

25

1.5 Research approach

The applied aim of this thesis was to enable targeted, risk-based management of infectious

animal diseases with feral pigs as a wildlife host by generating new spatially-explicit knowledge of

the host population. As feral pigs are widespread throughout much of the study region in northern

Australia (West 2008), and an incursion of a disease such as CSF may occur at many different

locations, especially along the coastline (Cookson et al. 2012), effective post-border disease

management requires information at the regional scale. Further, feral pig populations strongly

respond to variability in environmental conditions (Choquenot et al. 1996), which in turn will affect

disease establishment and spread (Milne et al. 2008). Given northern Australia’s annual rainfall

cycle, host information should therefore be seasonally-specific. Unfortunately, continuous empirical

data is rarely available over large areas and for multiple temporal scenarios (Stephens et al. 2015).

This gap in knowledge is also true for feral pigs in northern Australia (Cowled & Garner 2008).

Collecting detailed information through field studies was out of the scope of this thesis.

Instead, I adopted a modelling approach, rooted in landscape ecological theory, to answer three

research questions at multiple biological levels in space and time, whereby ‘level’ refers to the

hierarchically structured organisation or observed response of a species (McGarigal et al. 2016, for

details see Section 2.1 and Section 2.2). Here I focused on three organisational levels (Fig 1.2): At

the level of individual feral pig breeding herds, I explored the selection of resources within home

ranges. At the level of local subpopulations (i.e. many herds), I used a habitat suitability modelling

approach to investigate the distribution of potential patches of breeding habitat emerging from the

interactions between resources and home range movements. At the metapopulation level, I

examined potential dispersal pathways between such patches using a habitat connectivity modelling

approach. Similar combinations of habitat suitability and connectivity models have been used to

study mobile species’ use of entire landscapes (Cianfrani et al. 2013; Dickson et al. 2013; Stewart-

Koster et al. 2015).

26

Fig 1.2 Schematic illustration of the multi-level research approach. At the local subpopulation level, the

distribution of suitable habitat patches (panels B, red polygons) depends on the selection of resources (shown

as coloured patches) by individual breeding herds in home ranges (panels A, red circles). At the regional

metapopulation level, habitat patches (panels C, connector lines) are connected by dispersal. At all levels,

two seasonal scenarios (wet vs. dry) accounted for temporal variability in environmental conditions.

In this thesis, I took a resource-based modelling approach (Hartemink et al. 2015) – building

on the ecological or behavioural factors influencing the spatial population patterns of feral pigs –

and applied models to two seasonal scenarios (wet and dry). Due to empirical uncertainty about the

processes of habitat selection and dispersal at the regional extent, models relied heavily on expert

knowledge and, where possible, subsequent evaluation of assumptions against empirical data.

Importantly, models did not simulate actual population dynamics, spread or abundance in a

spatiotemporal continuum. This level of detail was rejected partly because feral pigs in northern

Australia are subject to frequent and unpredictable control activities that strongly affect actual

population dynamics and distributional patterns. The adopted resource- and scenario-based research

approach was ultimately deemed most appropriate, and cost-effective, for generating regional-scale

spatial knowledge on feral pigs that may usefully inform strategic disease management decisions.

27

1.6 Thesis structure

This thesis is partly comprised of publications in peer-reviewed scientific journals (Chapter 4 is

published and Chapters 3 and 5 are being prepared for submission). These were reproduced here in

their published or prepared manuscript form, with only minor modifications4. First, in publications,

the study species Sus scrofa was referred to by the internationally recognized term ‘wild pig’; here,

this was substituted by ‘feral pig’, which is more familiar to Australian readers. Second, journal

requirements on formatting and referencing were streamlined to conform to thesis submission

guidelines. Third, all references and appendices were collated at the end of the thesis5.

In Chapter 1 I introduced the biosecurity threat of infectious animal diseases with wildlife

hosts. I highlighted that limited spatially-explicit knowledge of wildlife hosts at the regional scale is

a major hindrance to effective post-border disease management at all stages of invasion. The aim of

this thesis is to fill this knowledge gap for feral pigs in northern Australia. I identified three research

questions that need to be answered to achieve this aim and provided a brief overview of the study

region. As empirical studies are prohibitive at the regional scale, I address these questions through

landscape ecological models of habitat suitability and connectivity at multiple biological levels.

In Chapter 2 I review the conceptual and methodological underpinnings of this thesis. First, I

review five major theoretical concepts in landscape ecology. Next, I review current knowledge on

the landscape ecology of feral pigs in northern Australia, highlighting that there are still significant

gaps in our understanding across spatial and temporal scales. Then I link this landscape ecological

framework to a landscape epidemiological perspective on infectious diseases in feral pigs. Finally, I

review the literature relating to the adopted research methodologies. This includes spatially-explicit

approaches to modelling habitat suitability and connectivity as well as methods for eliciting robust

knowledge from experts for model development and parameterization.

In Chapter 3 I develop a novel combined methodology, spatial pattern suitability analysis, for

integrating resource-seeking home range movements into habitat models of mobile animals. This

involves measuring structural patterns of resource quality at the home range scale and then relating

these measures to functional values from expert-elicited response-to-pattern curves. I use scenario

analysis and empirical validation in a subsection of the study region in northern Queensland to

evaluate whether this methodology improves model performance and how structural patterns should

4 Due to the stand-alone nature of journal papers, some repetition in introductory remarks, methodological descriptions and associated figures, and discussion of findings does occur throughout this thesis. This is especially the case for Chapters 3 and 4, which used a similar methodology for an exploratory (Chapter 3) and an applied (Chapter 4) analysis. 5 Appendices were numbered according to chapters and order of first mention, i.e. Appendix 1 supplements Chapter 1, Appendices 3.1 – 3.4 supplement Chapter 3, and Chapter 2 does not have an appendix.

28

be measured for feral pig breeding herds. This chapter addresses research question 1, providing a

methodology for reliably modelling habitat suitability for feral pigs at the regional scale under

uncertainty about the processes of habitat selection by individual breeding herds. It is being

prepared for submission to Ecological Modelling.

In Chapter 4 I apply the methodology from Chapter 3 to model and map seasonal habitat

suitability for feral pig breeding and persistence in northern Australia. I provide a detailed

description of the resource-based, spatially-explicit modelling approach using expert-elicited

Bayesian networks and spatial pattern suitability analysis to account for resource-seeking home

range movements. I compare modelled habitat suitability for a wet season and a dry season scenario

and empirically validate model accuracy against four independent distributional data sets per

scenario. This chapter addresses research question 2, providing an indicative measure of the

seasonal establishment risk of directly transmitted infectious diseases within feral pigs in northern

Australia. It has been published in PLoS ONE.

In Chapter 5 I model and map seasonal habitat connectivity for feral pigs in northern Australia.

I apply the results from Chapter 4 to delineate contiguous patches of feral pig breeding habitat,

parameterize matrix resistance to inter-patch dispersal, and model seasonal connectivity in the

landscape matrix using a circuit-theoretic approach. Then, I estimate connectivity between patches

and evaluate its sensitivity to a range of model assumptions that are linked to gender-specific

differences in feral pig dispersal ability. This chapter addresses research question 3, providing an

indicative measure of the seasonal spread risk of directly transmitted infectious diseases within feral

pigs in northern Australia. It is being prepared for submission to Journal of Applied Ecology.

In Chapter 6 I relate the results from Chapters 3 to 5 to the aims of this thesis. For each of the

three research questions, I discuss the main findings of my research as well as their methodological,

ecological and applied significance. Next, I synthesize how the adopted multi-level modelling

approach and its outputs can help to inform risk-based management of infectious animal diseases

for which feral pigs are a wildlife host. I conclude by summarizing limitations and giving

recommendations for future research that may further improve on this thesis’ findings.

29

Fig 1.3 Thesis structure. Broad descriptions of chapter content, types of results and associated publications

(published or being prepared for submission) are also given.

1.7 Research ethics approval

The research published in this thesis was conducted with human ethics approval from the CSIRO

Human Ethics Committee (Approval 075/13, 27 Aug 2013) and the UQ School of Agriculture and

Food Sciences Ethics Committee (9 Oct 2013).

Chapter 3 Develop and validate methodology for integrating home range movements

into feral pig habitat models

Publications Thesis chapter (description) Results

Ecol Modelling (in prep.)

PLoS ONE (2017)

Chapter 4 Model and validate seasonal habitat suitability for feral pigs

in northern Australia

Chapter 6 Discuss findings and applications to risk-based management

of infectious animal diseases in feral pigs and recommend future research

Models & maps

Metho-dology

Chapter 2 Review study system and theoretical framework (wildlife hosts and infectious diseases) and research methods

(habitat and connectivity models, elicitation)

Chapter 1 Introduce research problem –infectious animal diseases in feral pigs

Chapter 5 Model and evaluate seasonal habitat connectivity for feral pigs

in northern Australia

J Applied Ecol

(in prep.) Models & maps

30

Chapter 2 Literature review

Summary

In this chapter I review the conceptual and methodological underpinnings of this thesis. First, I

review five major theoretical concepts in landscape ecology. Next, I review current knowledge on

the landscape ecology of feral pigs in northern Australia, highlighting that there are still significant

gaps in our understanding across spatial and temporal scales. Then I link this landscape ecological

framework to a landscape epidemiological perspective on infectious diseases in feral pigs. Finally, I

review the literature relating to the adopted research methodologies. This includes spatially-explicit

approaches to modelling habitat suitability and connectivity as well as methods for eliciting robust

knowledge from experts for model development and parameterization.

2.1 Theoretical concepts in landscape ecology

Turner & Gardner (2015) broadly state that the discipline of “landscape ecology emphasizes

the interaction between spatial pattern and ecological process – that is, the causes and consequences

of spatial heterogeneity across a range of scales” (p. 2). With regard to an individual organism, this

means that reproduction, dispersal or other behavioural responses are influenced by the way abiotic

and biotic factors are distributed in the landscape and the organism’s responses in turn affect its

own distributional patterns and abundance (Levin 1992). In essence, landscape ecology combines

the spatial focus on observed patterns of biogeographers with the functional focus on underlying

processes of ecologists (Turner & Gardner 2015). The integration of these perspectives – pattern

and process, geography and ecology, structure and function – was also a primary motivation behind

this thesis. Excellent syntheses of the varied theoretical concepts and research approaches in

landscape ecology have been provided by Turner & Gardner (2015), Farina (2006), Forman (1995)

and Wu (2013). It is out of the scope of this thesis to repeat these discussions here. Instead, I will

focus on five broad theoretical concepts, which underpin the adopted research approach.

2.1.1 Habitat suitability

The first concept is the ‘habitat’. To define its use in this thesis, I must review diverse and

often ambiguous research perspectives, which fundamentally also reflect the dichotomy between

pattern and process emphasized above. Understanding the geographic distribution of species has

historically been of interest to ecologists and biogeographers alike. Earliest investigations have

focused on explaining observed distributional patterns from permissive environmental and climatic

conditions (von Humboldt & Bonpland 1807; Grinnell 1917). Hutchinson (1957) formalized

Grinnell’s insights as an “n-dimensional hypervolume, every point in which corresponds to a state

of the environment which would permit [… a species] to exist indefinitely.” (p. 416). He coined this

31

multi-dimensional volume of environmental requirements the ‘fundamental niche’ (since also

known as the ‘Grinnellian niche’; Leibold 1995; Pulliam 2000) and contrasted it with the ‘realized

niche’ (since also known as the ‘Hutchinsonian niche’; Leibold 1995; Pulliam 2000), which was

supposed to be narrower due to the effects of competition. Importantly, Hutchinson’s (1957)

“Concluding remarks” turned the focus from the description of observed distributional patterns to

the underlying, species-specific processes or mechanisms from which these patterns emerge.

Exactly how ecological processes in multi-dimensional ‘environmental (niche) space’ translate into

distributional patterns in two-dimensional ‘geographical space’ remains a complex question, and

depends on both the species and the environment (Hutchinson 1957; Pulliam 2000; Guisan &

Thuiller 2005; Hirzel & Le Lay 2008). Broadly, however, a species fundamental niche can be

linked to its potential distribution (where it may persist in a permissive environment), while the

realized niche gives rise to its actual distribution (where it currently occurs) (Pearson 2007).

The related ‘habitat’ concept has been used predominantly, and often ambiguously, in animal

studies. Firstly, Hall et al. (1997) or Morrison et al. (2006) stress that habitat should be an

organism-centric concept and not be confused with ‘habitat type’ (structurally similar landscape

elements, e.g. a type of vegetation). Secondly, some authors have defined habitat as the physical

area an animal potentially or actually occupies in geographical space (the descriptive ‘distribution’

perspective; Hall et al. 1997; Kearney 2006; Morrison et al. 2006), while others have referred to the

environmental requirements needed for persistence in environmental space (the mechanistic ‘niche’

perspective; Pulliam 2000; Hirzel & Le Lay 2008). Confusion arises when considering that: (a)

some suitable ‘niche habitat’ may actually not be occupied (e.g. due to historical or biogeographical

range limits; short-term extinctions due to stochastic events or, for invasive species, population

control activities; limitations on dispersal and re-colonization; or biotic interactions with other

species); (b) some unsuitable ‘niche habitat’ may consistently be occupied (e.g. due to continuous

immigration from suitable ‘source’ habitat into unsuitable ‘sink’ habitat; or due to regular

movements across unsuitable habitat); (c) permissive environmental conditions as well as actual

occupancy may change over time (e.g. due to climatic conditions and resource availability); and (d)

spatial heterogeneity and behaviour may influence habitat use (e.g. the amount of habitat may be

too small or fragmented to allow persistence or even-short term occupancy; or mobile species may

satisfy different requirements in separate locations, i.e. have a spatially disjunct ‘niche habitat’)

(Pulliam 2000; Guisan & Thuiller 2005; Hirzel & Le Lay 2008; Yackulic & Ginsberg 2016).

Here, I attempted to avoid such confusion by adopting a descriptive definition of habitat as a

geographical place, but limiting its application to feral pigs’ potential distribution and persistence.

In line with the resource-based approach (Hartemink et al. 2015), I defined ‘habitat’ as the

32

geographical area in which a species can potentially persist without immigration based on the

availability of all resources and conditions required for survival and reproduction.

Habitat thus defined can vary in ‘suitability’, which I used here equivalent to the term ‘habitat

quality’ that is preferred by some authors (Hall et al. 1997; Beutel & Beeton 1999; Stephens et al.

2015). I adopted the definition by Hall et al. (1997) who state that habitat quality (here ‘suitability’)

“should be considered a continuous variable, ranging from low to medium to high, based on

[the quality of] resources available [and conditions encountered] for survival, reproduction,

and persistence” (p. 178).

According to these definitions, my investigations of habitat suitability in feral pigs included only

factors that may impact on potential occupancy: (a) abiotic habitat factors such as consumable

resources and climatic conditions; (b) limiting factors such as the presence of competitors, predators

or anthropogenic disturbances; (c) spatiotemporal variability in the availability and quality of these

habitat factors; and (d) behavioural factors, in particular feral pigs’ use of home ranges to satisfy

heterogeneously distributed habitat requirements and avoid disturbances (Section 2.2.3). I excluded

from the analyses additional habitat factors that may only affect actual occupancy or persistence: (e)

dynamic biotic interactions within feral pig populations (e.g. reproductive, metapopulation, source-

sink or density-dependent dynamics; or demographic stochasticity) or with other species (e.g.

trophic, predator-prey or community interactions); and (f) historical anthropogenic factors such as

species introduction, translocation, colonization or population control.

2.1.2 Patch-corridor-matrix model and metapopulation ecology

The second concept is the ‘patch-corridor-matrix’ model (Forman 1995). In this simple spatial

model, the landscape is viewed as a mosaic, with every point being part of either a ‘patch’, a

‘corridor’, or the background ‘matrix’ (Forman 1995, McGarigal et al. 2012). While conceptually

simple, operationalising patches and corridors is by no means trivial (e.g. Girvetz & Greco 2007,

Beier et al. 2008, Galpern et al. 2011). One approach is to regard patches and corridors as

structurally distinct landscape features of a contiguous shape, whereby corridors may be

distinguished from patch polygons by their linear form (e.g. two forest stands connected by a

vegetated stream). Applications in conservation biology and reserve design have often focused on

connecting such ‘natural landscape blocks’ (= contiguous patches, Beier et al. 2011) of broad

preservation value by means of wildlife ‘linkages’ (= linear corridors, Beier et al. 2008).

However, in many landscapes these distinctions may not be obvious. Even where structurally

separate features may be discerned by a human observer, it is widely acknowledged that there is no

single set of patches in a landscape. Rather, “the contrast between a patch[, or a corridor,] and the

33

surrounding matrix is dependent on the perceptual abilities and behavioral responses of a focus

organism” (Girvetz & Greco 2007, p. 1132). From this functional, organism-centric perspective, the

patch-matrix dichotomy is best described in terms of discontinuities in habitat quality or patterns of

use (McGarigal et al. 2012). For mobile animals such as feral pigs, which use home ranges to

access resources at different locations (Börger et al. 2008; Section 2.1.3), meaningful habitat

patches must be delineated functionally. While there are many approaches for doing this (Girvetz &

Greco 2007, Galpern et al. 2011), all aim at identifying patches that are “good enough, big enough,

and close enough together” to function as habitat (Beier et al. 2008, p. 844; Section 2.5.2).

The patch-corridor-matrix model, when functionally defined, can also be linked to the theory

of ‘metapopulation dynamics’ (Hanski 1998). Then, unstable local populations inhabit discrete

patches of suitable breeding habitat that are embedded in a matrix of unsuitable habitat. Certain

areas in the matrix function as dispersal corridors between patches. Metapopulation persistence

depends on local population dynamics and movements in the patch network (Hanski 1998; Hastings

2014). Glen et al. (2013) argue that biosecurity applications and invasive species management

could benefit from using the metapopulation paradigm in reverse, focusing on eradication of local

populations and disruption of inter-patch dispersal. In this thesis I echo their call.

2.1.3 Scaling in space and time

The link to metapopulation ecology points to the importance of scale as a third fundamental

concept to be discussed here (Levin 1992). From a functional perspective, habitat patch networks in

a given landscape vary not only by focal species, but also display a hierarchically nested structure

depending on the biological level and time horizon investigated (Girvetz & Greco 2007). According

to McGarigal et al. (2016), “‘level’ refers to a constructed organizational hierarchy […in] (a) the

environment […] or (b) the focal organism’s behaviour” (p. 1164). For example, a species’

behavioural hierarchy is often described in terms of Johnson’s (1980) four ‘orders’ of habitat

selection (1st order = population range, 2nd order = home range, 3rd order = resource patches such as

feeding sites, 4th order = individual resources such as food items). Each of these levels must be

investigated at different spatial and temporal scales (Mayor et al. 2009; McGarigal et al. 2016).

However, what constitutes a suitable spatiotemporal scale of analysis, including ‘grain’ (=

resolution, smallest unit of observation) and ‘extent’ (= largest entity of analysis), is by no means

clear. Here, Addicott et al. (1987) introduced the general concept of the ‘ecological neighbourhood’

as “the region within which […an] organism is active or has some influence during […an]

appropriate period of time” (p. 341). Addicott et al. (1987) argued that each ecological process has

its own unique ecological neighbourhood and the choice of process will determine the appropriate

timeframe and scale of measurement. Yackulic & Ginsberg (2016) provided an excellent synthesis

34

of how a species’ biological and behavioural levels can be linked with ecological processes and

suitable grains and extents of analysis to explain distributional patterns consistent with the

metapopulation perspective on mobile species adopted here (Fig 2.1): individuals or social groups

(level) select and move between (process) resource patches (grain) within daily home ranges

(extent); local subpopulations (level) form where several home ranges or territories (grain) intersect

to form contiguous habitat patches (extent) that allow for in situ persistence (process);

metapopulations (level) form where several local populations (grain) are connected via dispersal

across geographical ranges (extent) to allow for species persistence (process); ecological processes

and distributional patterns at each level may fluctuate periodically (temporal grain) and stabilize, or

shift, over longer time horizons (extent). A large number of ‘focal patch’ or ‘multi-scale’ studies

have been devoted to finding the “right” ecological neighbourhood at each of these biological levels

(Section 2.4.3; Holland et al. 2004; Jackson & Fahrig 2015; McGarigal et al. 2016).

Fig 2.1 The link between a mobile species’ behavioural levels and corresponding spatial and temporal

scales of analysis. (Source: Mayor et al. 2009, Fig 1, modified after Yackulic & Ginsberg 2016)

When studying habitat suitability for local feral pig subpopulations (research question 2), I

assumed that the most relevant ecological neighbourhood is the home range of individual breeding

herds, i.e. suitability for breeding and persistence is regulated by the same scale-dependent

variations in resource availability that influence habitat selection by individual breeding herds

(Yackulic & Ginsberg 2016). This is because the dominant process at the level of subpopulations is

social aggregation, where individual breeding herds interact with adjacent herds in overlapping

home ranges to form contiguous subpopulations within a habitat patch (Section 2.2.2; Choquenot et

35

al. 1996; Hone 2012; Yackulic & Ginsberg 2016). In territorial animal populations characterized by

repulsion, the ecological neighbourhoods of individuals and local populations may be quite

different (Yackulic & Ginsberg 2016).

2.1.4 Landscape heterogeneity and home ranges

Addicott et al. (1987) also refer to the importance of ‘environmental patterning’ in ecological

neighbourhoods. This can be illustrated with mobile animals’ home range behaviour. Home ranges

can be broadly defined as the “area traversed by an individual in its normal activities of food

gathering, mating, and caring for young” (Burt 1943, p. 352). While many theoretical and

operational questions about home range behaviour remain, the conceptual basis of animals’ space

use is widely agreed upon: mobile animals use home ranges to satisfy a number of heterogeneously

distributed habitat requirements (Börger et al. 2008; Powell & Mitchell 2012). Spatial patterns of,

and relationships between, required resources and other habitat factors – for example the location of

predators in relation to hiding places or the availability of sufficient feeding as well as breeding

sites – will affect space use, and habitat suitability.

The spatial heterogeneity arising from such patterns and relationships has also been referred to

as ‘landscape structure’ (Kupfer 2012; Turner & Gardner 2015). A range of ‘landscape metrics’ for

describing aspects of landscape structure have been developed over the past two decades

(McGarigal & Marks 1995; Uuemaa et al. 2009; McGarigal et al.2012). It is out of the scope of this

thesis to comprehensively review these, yet metrics can be classed into two general categories.

Metrics of landscape ‘composition’ refer to the variety and abundance of elements in a landscape

(e.g. resource or habitat patches), and include measures of amount, density, proportional cover or

diversity. Metrics of landscape ‘configuration’ refer to the spatial position and arrangement of

landscape elements. They can be used to measure both the patterning of individual resources and

the spatial relationships between different types of resources or habitat patches. Typically,

configuration metrics measure properties such as distance / proximity, patch geometry, edges

between patches, fragmentation, aggregation / clumping or isolation (McGarigal & Marks 1995).

Importantly, all metrics of landscape structure, per se, are independent of ecological function, that

is, they carry no implicit meaning to the way a species uses and perceives its environment (Kupfer

2012; Turner & Gardner 2015).

Here, Dunning et al. (1992) described three fundamental ecological processes that link

structural heterogeneity within ecological neighbourhoods (Addicott et al. 1987), or more

specifically mobile animals’ home ranges, to habitat quality (here ‘suitability’). Resource

‘complementation’ occurs when an individual selects resources required for different purposes (e.g.

forage and breeding sites) at different locations within its home range. Therefore, habitat suitability

36

of a focal site depends on the availability of all required resources within the accessible home range

neighbourhood. The individual may also be able to select different types of ‘supplementary’

resources that can be substituted for the same purpose (e.g. different food items). Here, habitat

suitability depends on the quality of the best substitutable resources that are available. Lastly,

habitat suitability is influenced by ‘neighbourhood effects’, that is, the distance of complementary

and supplementary resources from the focal site will affect their value to a mobile animal.

2.1.5 Habitat connectivity

While ‘patches’ have frequently been investigated at all biological levels (e.g. resource, home

range or habitat patch), ‘corridors’ have typically been studied at the metapopulation level in terms

of dispersal within a patch network6. In the expanding field of connectivity research, which broadly

aims at studying how patches are linked to one another, two broad perspective have emerged

(Tischendorf & Fahrig 2000; Moilanen & Hanski 2001; Kool et al. 2013; Fletcher et al. 2016a):

Metapopulation studies have focused on the population dynamics within and between patches that

are ‘isolated-by-distance’ and investigated how linked patches facilitate metapopulation persistence

(Pulliam & Danielson 1991; Hanski 1998; Hastings 2014). The ‘landscape connectivity’ (here

referred to as ‘habitat connectivity’) perspective has focused on landscape heterogeneity, both

within patches and in the surrounding matrix, and studied how this heterogeneity influences

connectivity between habitat patches that are ‘isolated-by-resistance’ (McRae 2006; Turner &

Gardner 2015). Thus, ‘connectivity’ is conceptualised as an attribute of the landscape, rather than

populations, and has been defined as “the degree to which a landscape facilitates or impedes

movement of organisms among […habitat] patches” (Tischendorf & Fahrig 2000, p. 7).

Nevertheless, habitat connectivity is not independent of the organism, nor the biological level,

investigated (Tischendorf & Fahrig 2000). Kool et al. (2013) distinguished between three analytical

perspectives on connectivity: ‘Structural connectivity’ refers to the physical attributes and

heterogeneity in a landscape. It can be directly measured using landscape metrics – ranging from

simple Euclidean distance to complex measures of landscape composition and configuration

(Kindlmann & Burel 2008) – and carries no implicit ecological meaning. ‘Functional connectivity’,

on the contrary, considers structural heterogeneity as well as species-specific behavioural responses

(Tischendorf & Fahrig 2000; Kindlmann & Burel 2008). A critical concept in studies of functional

connectivity is ‘matrix resistance’. It refers to the permeability of landscape elements outside of

habitat patches to movement or the costs (energy expenditure, mortality risk) associated with

6 Yet, movements between resource patches within a home range, or movements between home range patches within a local population, could also be conceptualized as ‘corridors’ (e.g. Mitchell & Powell 2012).

37

traversing them. Matrix resistance, and functional connectivity, of a given landscape will vary

according to the requirements, abilities and behavioural responses of an organism (McRae 2006;

Kindlmann & Burel 2008; Zeller et al. 2012). Depending on the biological level investigated, Kool

et al. (2013) and Fletcher et al. (2016a) further distinguish between ‘functional connectivity’ as

defined by Tischendorf & Fahrig (2000), which studies dispersal movements between local

populations over moderate time frames (months to years, Fig 2.1), and ‘genetic connectivity’, which

studies gene flow within or between metapopulations (decades to centuries, Fig 2.1). The ecological

neighbourhood of dispersal or gene flow processes, and corresponding spatiotemporal scales of

analysis, may be quite different (Fletcher et al. 2016a). Finally, Fletcher et al. (2016a) contrast

functional ‘potential connectivity’ (where matrix permeability allows species movement) and

‘realized / actual connectivity’ (where movement currently occurs) (Calabrese & Fagan 2004).

When modelling habitat connectivity for regional feral pig metapopulations (research question

3), I focused on potential functional connectivity. I did not attempt to analyse actual dispersal

movements at a particular point in time (Fletcher et al. 2016a), gene flow over long time horizons

(Fletcher et al. 2016a), or dynamic population-level responses such as source-sink relationships

between patches (Pulliam & Danielson 1991; Dunning et al. 1992) or metapopulation persistence

(Hanski 1998; Hastings 2014). The ecological neighbourhood of connectivity analyses in this thesis

was determined by the ability of individuals to disperse between local subpopulations.

2.2 Landscape ecology of feral pigs

Feral pigs are descendants of domesticated Eurasian wild boar (Sus scrofa). Both wild and

domesticated forms of S. scrofa have been introduced by early settlers to all continents and many

oceanic islands. The species is widely considered as one of the worst invasive species throughout its

introduced range (Barrios-Garcia & Ballari 2012). In Australia, after two centuries of recurrent

introductions, translocations and natural dispersal, feral pigs are now among the most abundant and

widespread terrestrial mammals (West 2008). Consequently, the species’ ecology, impacts and

management options have been widely studied (Choquenot et al. 1996; Hone 2012; Bengsen et al.

2014). Below I review the published knowledge on key ecological processes and spatial patterns

from the individual to the metapopulation level, highlighting that there are still significant gaps in

our understanding of feral pig populations in Australia across spatial and temporal scales.

2.2.1 Distribution and abundance

The distribution and abundance of feral pigs in Australia has been investigated along three

broad lines of enquiry. First, a large number of field studies have empirically investigated local

habitat use and population densities in response to a range of environmental and biotic factors, often

38

dynamically across seasons or years (Hone 1990a; Caley 1993 & 1997; Corbett 1995; Choquenot

1998; Mitchell 2002; Choquenot & Ruscoe 2003; Mitchell et al. 2009). For example, Caley (1993)

and Hone (1990a) showed that feral pig distribution and abundance varies considerably between

seasons and habitat types of varying productivity. Yet, the detailed empirical observations from

such studies remained confined to a few study sites and were rarely integrated with spatial data sets.

Hone (1990b) consolidated five such published field studies to estimate the total size of the feral pig

population in Australia. Yet, with a limited sample size and no spatial integration, this estimate

remained enormously uncertain (95% confidence interval: 3.5 to 23.5 million pigs).

In response to a lack of spatially-explicit information, the second line of enquiry sought to map

the broad distribution and abundance of feral pigs at the regional or continental extent (Mitchell et

al. 1982; Wilson et al. 1992; West 2008; Department of Agriculture and Fisheries 2015). Due to

methodological and resourcing constraints, these efforts typically relied on expert-derived relative

estimates per coarse land parcel and were poorly validated, which limited their usefulness for on-

ground applications (Cowled & Garner 2008). Other ongoing mapping exercises such as the Atlas

of Living Australia (2016) or FeralPigScan (2016) collate and map occurrence records across

Australia from a range of sources, including historical surveys or citizen science. All studies in this

group share two further shortcomings: they were unable to fill data or knowledge gaps and temporal

variability was not considered.

A third type of study – habitat models, which infer species distributions from environmental

predictor variables (Section 2.4.1) – can overcome the limitations of simple mapping exercises. For

feral pigs in northern Australia, one such model was developed by Cowled and colleagues using a

statistical (Generalized Additive Modelling) approach (Cowled & Giannini 2007; Cowled et al.

2009). However, the authors suggested that the model markedly underestimated feral pig

distributions in areas for which no presence / absence data were available. Such extrapolation errors

are a common limitation of statistical models, which rely purely on data with limited consideration

of ecological and behavioural processes (Elith et al. 2010). Further, the model was calibrated from

aggregate occurrence records and did not account for temporal variability.

2.2.2 Social organisation and reproduction

While broad distributional patterns have been mapped and modelled over regional extents,

Cowled & Garner (2008) suggest that, at a finer grain, feral pigs are “rarely distributed

homogenously […], since they have strong social tendencies” (p. 202). There are two distinct social

entities in any feral pig population: related adult sows, sub-adults and juveniles typically live in

social herds called sounders; mature boars usually disperse from their natal area and remain solitary

(Choquenot et al. 1996; Gabor et al. 1999; Mitchell 2008). Female herd sizes can vary considerably

39

between different environments and seasons, and are not stable over time. Choquenot et al. (1996)

reviewed published herd sizes, concluding that they typically contain fewer than 10, but under

exceptional circumstances up to 100 animals. These breeding herds display a highly fluid ‘fission-

fusion’ dynamic, with sub-groups splitting off, merging and exchanging individuals across space

and time (Gabor et al. 1999; Spencer et al. 2005). While Gabor et al. (1999) found that sounders

living in adjacent habitat patches did rarely interact, there is little further evidence for territorial

behaviour in either females or males (Choquenot et al. 1996; Hone 2012). In fact, genetic analyses

have shown that individual boars frequently move between sounders for mating and sows accept

multiple mating partners (Gabor et al. 1999; Spencer et al. 2005).

Feral pigs have an extraordinary reproductive capacity, comparable to rabbits rather than other

large mammals in Australia (Choquenot et al. 1996; Mitchell 2008). Under favourable conditions,

breeding may occur throughout the year and sows may produce two weaned litters per year. Each

litter contains on average 5 to 6 piglets and requires 2 to 3 months lactation to weaning (Choquenot

et al. 1996). Results from a number of studies investigating population dynamics of feral pigs in

various environments indicated that populations can increase up to twofold (instantaneous rate of

increase ~0.7) within one year (Choquenot et al. 1996; Hone 1990a; Caley 1993).

2.2.3 Resource selection and home range movements

Site-scale resource selection and habitat use of feral pigs has also been studied widely in the

Australian context. A species’ habitat (resource) requirements are determined by its physiological

attributes. Due to their compact shape and few sweat glands, feral pigs overheat easily (Mitchell

2008). This can have a negative effect on survival (Choquenot & Ruscoe 2003) and fertility (Greer

1983) in both sows and boars. The species’ diet is omnivorous. Yet, pregnant or lactating sows in

particular require a relatively high intake in digestible energy and crude protein (Choquenot et al.

1996). Preferred food items include fresh green vegetation and a range of other protein-rich plant

(fruit, seeds, grains, roots and bulbs) and animal material (eggs, invertebrates) to satisfy nutritional

requirements (Choquenot et al. 1996; Ross 2009). Despite being highly adaptable to a range of local

habitat conditions, resource selection analyses have, unsurprisingly, established that feral pigs

prefer habitats which satisfy their physiological needs for thermoregulation (Dexter 1998), frequent

hydration (Caley 1997) and a protein-rich diet (Caley 1997; Dexter 1998; Wurster et al. 2012).

After Dunning (1992), these resource requirements are both complementary (i.e. are all needed for

different purposes as described above) and supplementary (i.e. can be opportunistically substituted

with resources serving the same purpose, e.g. shady tree vs. muddy wallow for thermoregulation).

Importantly, feral pigs’ habitat use and distributional patterns depend both on the availability

and patterning of resources as well as behavioural factors. Like all large and mobile animals, feral

40

pigs are able to access heterogeneously distributed resources at different locations and times within

the boundaries of their home range (Powell & Mitchell 2012). A number of home range analyses

have found that feral pigs are relatively sedentary for their size and occupy well-defined home

ranges. There were consistent gender-specific and seasonal differences in home range movements

(Caley 1997; Dexter 1999; Mitchell et al. 2009). Dexter (1999) found that female home ranges

decreased significantly under drought conditions. This was interpreted as a “compromise between

the need to forage widely for food and the need to stay close to reliable cover and water” (Dexter

1999, p. 759). Male home ranges were significantly larger and, interestingly, did not vary with

conditions. Such findings suggest that boars are less movement-constrained by daily requirements

for water and protein-rich food (Choquenot et al. 1996). Overall, male and female home range sizes

ranged widely between 1-40 km2 and 1-20 km2 respectively (Choquenot et al. 1996; Hone 2012).

Despite these detailed findings on resource selection and home range movements, few studies

have addressed the role of landscape structure and scaling, i.e. the patterns and spatial relationships

between resources within home range neighbourhoods. Here, a rare insight stems from Choquenot

& Ruscoe (2003), who explicitly investigated resource complementation at different scales and

found that food resources and thermal refuge must co-occur within 5 km to facilitate feral pig

persistence. Caley (1993) investigated food supplementation with cereal crops and found that feral

pig density in riverine woodlands increased almost four-fold when crops were available within a 5

km radius. These site-scale findings are consistent with the observed 20 km2 home range limit for

breeding herds (Hone 2012). Yet, the influence of resource composition and configuration in the

home range on broad distributional patterns of feral pigs has not been studied so far.

2.2.4 Metapopulation structure and dispersal

Several studies have analysed the population structure of feral pigs using landscape genetic

evidence. In a first study covering 4,000 km2 in south-western Queensland, Cowled et al. (2006)

found a genetically contiguous population. When repeating the study over a much larger (500,000

km2) area within the same region, Cowled et al. (2008) did find genetic differentiation into five

subpopulations. A simple Euclidean distance measure explained 50% of the variability in

differentiation. Interestingly, evidence for mixing was most prominent in drier landscapes. This

suggested that feral pigs may utilize rivers and floodplains to disperse widely within unfavourable

landscapes while they are less inclined to disperse where resources are available locally (Cowled et

al. 2008). In a fragmented landscape in south-western Western Australia, Hampton et al. (2004)

revealed genetic fine-grained differentiation at the catchment level. There was high gene flow along

water courses but little population mixing between neighbouring catchments as little as 25 km apart

(Hampton et al. 2004). Likewise, Lopez et al. (2014) identified strong population-structuring at a

41

scale of 25 to 35 km in the coastal lowlands of wet-tropical northern Queensland. Unfortunately,

while population structuring has been extensively documented, dispersal processes responsible for

these emergent patterns often remained unclear. In particular, the influence of landscape

heterogeneity in the matrix on inter-patch connectivity has not been investigated.

In summary, empirical studies indicate that the patch-corridor-matrix and metapopulation

frameworks can adequately describe populations over regional extents. Distributional patterns are

spatially clustered because feral pigs congregate in social breeding herds (also called ‘sounders’ or

‘mobs’). Herds display a dynamic group structure and interact frequently with adjacent herds to

form contiguous local populations within a habitat patch. Contact between subpopulations in

separate habitat patches is generally rare, as the intervening matrix habitat may present a major

obstacle to resource-constrained females. Yet, besides the energetic and survival costs associated

with moving through poor quality habitat, no absolute barriers to feral pig dispersal have been

identified in Australia (Choquenot et al. 1996; Cowled & Garner 2008; Cowled et al. 2008; Hone

2012; Lopez et al. 2014). Hence, subpopulations may be frequently linked via male dispersal

movements, as these individuals are less constrained by daily resource requirements (Choquenot et

al. 1996; Gabor et al. 1999). Importantly, research gaps exist with regard to inter-patch dispersal,

metapopulation dynamics, and connectivity within the habitat patch network.

2.3 Landscape epidemiology of infectious diseases in feral pigs

The field of ‘landscape epidemiology’ (also termed ‘spatial epidemiology’) has recently re-

emerged from earlier work and received much attention since (Ostfeld et al. 2005). Essentially, it

stipulates the concept of the ‘nidus of pathogen transmission’, where “vector [if applicable], host,

and pathogen populations intersect within a permissive environment” (Reisen 2010, p. 463). That is,

spatial patterns of disease distribution and spread will depend on the spatial patterns of host (and

vector) distribution and the spatial relationships between hosts from the individual to the population

level (Ostfeld et al. 2005; Hartemink et al. 2015). As these patterns and relationships are the

primary focus of landscape ecology (Turner & Gardner 2015; Section 2.1), the disciplines are

closely linked (Reisen 2010). Anderson et al. (1986) distinguished three phases of invasion by

infectious diseases: establishment, persistence and spread. For directly transmitted (as opposed to

indirectly transmitted – e.g. vector-, air-, or soil-borne) diseases with a single host, these phases can

be linked to the general processes in host metapopulations described by Hastings (2014): local

population growth and persistence (Section 2.1.1), connectivity and dispersal (Section 2.1.5), and

regional metapopulation persistence. Following introduction, initial disease establishment is

contingent on locally dense host populations with high contact rates that let an infectious pathogen’s

basic reproductive rate R0 exceed 1. Subsequent disease spread requires host dispersal between

42

infected and susceptible populations. Disease persistence depends on a lasting supply of susceptible

individuals. This could pertain to short-term persistence in a closed subpopulation or long-term

persistence in a regionally connected metapopulation (Anderson et al. 1986; Kramer-Schadt et al.

2007; Riley 2007). In the following I illustrate these links using the example of the directly

transmitted disease classical swine fever (CSF) and its sole host species Sus scrofa.

Feral pigs congregate in social breeding herds that interact frequently with herds in adjacent or

overlapping home ranges to form contiguous, persistent local populations within patches of suitable

breeding habitat (Section 2.2.2). CSF is highly contagious and spreads mainly by direct contact

between live individuals, especially between piglets or from mothers to their young (Kramer-Schadt

et al. 2007; Animal Health Australia 2012). Kramer-Schadt et al. (2007) observed that within-herd

transmission of CSF depends on a persistent supply of young susceptible animals. Therefore,

Cowled & Garner (2008) suggested that local feral pig populations could be regarded as single

epidemiological units, where host density is sufficiently high, and interaction sufficiently frequent,

to let CSF’s basic reproductive rate R0 exceed 1 (Anderson et al. 1986).

Contact between separated feral pig populations is generally rare, as the intervening matrix

habitat may present a major obstacle to dispersal. Yet, contact rates may vary by gender and season,

with resource-constrained female breeders having a lower dispersal ability than often solitary males,

especially under adverse conditions (Section 2.2.4). Kramer-Schadt et al. (2007) observed that CSF

spread between infected and susceptible local populations may be enhanced by increased “contact

during the rutting season, male dispersers and establishment of new social groups” (p. 4). The view

that disease spread in feral pigs may be limited by spatial clustering is also supported by Zanardi et

al.’s (2003) study in Italy. They found that two adjacent populations of wild boar that were

separated by unfavourable habitat experienced two temporally distinct CSF epidemics.

Finally, while contact may be rare, especially during adverse conditions, local feral pig

populations are by no means isolated. Populations may merge, or dispersal movements increase,

during favourable conditions (Section 2.2.3). Distant populations may also be linked by linear

landscape features (Section 2.2.4). In some landscapes, feral pigs may even be distributed

homogeneously (Hone 2012). CSF, on the other hand, occurs in strains of highly variable

infectiousness and impact on the host organism (Section 1.1.2). Low virulent strains may cause

chronic disease and prolonged infectiousness with few clinical signs, enabling long-term disease

persistence within a feral pig metapopulation that seasonally connects local infected and susceptible

subpopulations (Anderson et al. 1986; Kramer-Schadt et al. 2007; Animal Health Australia 2012).

In summary, classical swine fever in feral pigs appears an ideal host-pathogen system to follow

Ostfeld et al.’s (2005) call for “[more studies that consider] the types, sizes and positions of

43

landscape elements (e.g. habitat patches, physical or biotic gradients, and type of matrix

surrounding patches) and their connectivity […as] potentially important drivers of [disease] risk or

incidence […, and that explore] the importance of landscape composition (number and types of

patches) and configuration (spatial relationships among patches) to disease dynamics” (p. 334).

2.4 Habitat suitability models

2.4.1 Modelling frameworks

With increased availability of spatial environmental data, and continued development of new

analytical methods, an enormous body of literature has emerged on the modelling of species

distributions, or ‘habitats’. In Fig 2.2 I attempt to summarize the major approaches based on their

treatment of explanatory variables, formulation of relationships, and modelled responses in either

geographic or environmental space (Section 2.1.1). I emphasize that these categories are arbitrary,

as many hybrid approaches continuously emerge in all lines of research (Franklin 2010; Gallien et

al. 2010), and broadly outline approaches and applications below.

Fig 2.2 Approaches to modelling species distributions, or ‘habitat’, in geographical and environmental

space. The diagram illustrates schematically (by connected points) how different components of the

modelling workflow are treated in either or both geographical and environmental space. Input variables

44

(covariates) can relate to abiotic and biotic spatial entities (e.g. resource patches), or to abiotic and biotic

functional requirements for an organism. The modelling process or algorithm may focus on spatial patterns

of and relationships between, or on the ecological function of and interactions between, abiotic and biotic

variables. The model output (response variable) may describe a species’ functional niche, or its spatial

distributional patterns. The location of each connected point on the abiotic-biotic variable / pattern / process,

fundamental-realized niche, and potential-actual distribution axes is approximate. Placement in the middle of

each axis means that the approach can accommodate either or both extremes.

Statistical habitat models are also known as ‘species distribution models’ (Elith & Leathwick

2009). They correlate a response variable (typically presence, presence / absence or abundance) and

a number of explanatory variables (covariates), both observed in geographical space, via assumed

but non-specified functional links in environmental space to infer the response at non-sampled

locations in geographical space. Neither do they model relationships between variables in

geographical space, nor do they make explicit reference to the functional niche requirements of an

organism. Approaches are methodologically and terminologically diverse (Dormann et al. 2012;

Gallien et al. 2010; Guisan & Thuiller 2005; Venette et al. 2010). Relatively simple ‘bioclimatic

envelope’ models have combined species presence records with abiotic (climatic) variables to

approximate a species potential distribution, usually over broad areas (Guisan & Thuiller 2005;

Venette et al. 2010). More complex methods have integrated species presence or presence / absence

records and abiotic (e.g. vegetation, land use, physical landscape characteristics) or biotic (e.g.

predators, areas with known competitive advantage) covariates, modelled relationships using

machine-learning or regression-based techniques, and been used to estimate potential or actual

distributions over broad or narrow study areas (Guisan & Thuiller 2005; Phillips et al. 2006; Elith

& Leathwick 2009; Jiménez-Valverde et al. 2011). Resource selection functions have quantified

animals’ preferential selection of abiotic resources or habitats by comparing observed patterns of

use with availability and used these insights to estimate actual habitat use, occupancy, or abundance

(Johnson 1980; Manly et al. 2002; Boyce et al. 2016). Although statistical habitat models have also

been referred to as ‘niche models’ (Elith & Leathwick 2009), I concur with Kearney (2006), who

reserves the term ‘niche’ for models that explicitly focus on functional processes.

Some spatially-explicit approaches model only spatial relationships between response and

explanatory variables, with no reference to underlying ecological processes. These models have

been applied to estimate a species’ actual distribution from incomplete presence data, or a mobile

animal’s home range ‘utilization distribution’ from incomplete telemetry data (Powell & Mitchell

2012). Techniques include range maps (Elith & Leathwick 2009), spatial interpolation (Bahn & Gill

2007), and statistical estimators such as convex hulls or α-hulls (Phillips et al. 2008), or kernel

45

smoothing / density estimators (Gormley et al. 2011; Kie et al. 2011). Recent geostatistical methods

such as regression-kriging or geographically weighted regression have emerged to bridge the gap

between modelling spatial or functional relationships (Austin 2007; Hengl et al. 2009).

Statistical techniques can only make robust inferences within the environmental or spatial

gradients contained in the data. Extrapolation to new geographical regions or novel conditions is

fraught with danger (Elith & Leathwick 2009; Paton & Matthiopoulos 2016). To predict species’

potential distribution across regions that have not been fully sampled, a range of approaches that

infer patterns from underlying processes, rather than processes from observed patterns, have been

developed. These have been termed ‘process-based’, ‘resource-based’, ‘deductive’ or ‘mechanistic’

models (Kearney & Porter 2009; Gallien et al. 2010; Venette et al. 2010; Hartemink et al. 2015).

‘Mechanistic niche models’ have focused on the biophysical niche requirements of organisms

and on that basis modelled how abiotic conditions interact to affect fitness (Phillips et al. 2008;

Kearney & Porter 2009; Kearney et al. 2009). As the entire modelling process is based in

environmental space, it has been claimed that such models can truly capture a species’ fundamental

niche (Kearney 2006). Biophysical requirements can subsequently be linked to spatial data proxies

and the modelled niche projected into space as a description of the organism’s potential distribution.

‘Resource-based habitat models’ have also focused on the processes in environmental space

from which habitat patterns emerge. However, different to niche models, resource-based models

have taken a broader view on an organism’s habitat requirements and modelled interactions

between abiotic, behavioural and temporal factors (Hartemink et al. 2015). Resource-based habitat

models have typically not accounted for intra- or inter-specific interactions (Dijak & Rittenhouse

2009). Notable techniques in this framework include habitat suitability index (HSI) models (Brooks

1997; Dijak & Rittenhouse 2009) or Bayesian network models (Section 2.4.2). By linking habitat

requirements to spatial data proxies, models can directly estimate habitat quality or suitability,

which in turn can be related to potential distribution in geographical space. Resource-based models

may not predict species’ actual distributions well (Dijak & Rittenhouse 2009).

Finally, all of the models discussed above essentially provide static snapshots of an organism’s

fundamental / realized niche or potential / actual distribution. Yet, each of these responses, and in

particular the realized niche / actual distribution end of the spectrum, may vary across time. Further,

they may all be influenced by biotic interactions within populations (e.g. density-dependent

population dynamics or demographic stochasticity) or with other species (e.g. trophic, predator-prey

or community interactions). A variety of ‘dynamic simulation models’ have been developed to

account for biotic interactions, often without reference to geographical space (i.e. spatially-implicit

models). Notable examples include ‘population viability analysis’, ‘metapopulation models’,

46

‘matrix models’, or ‘individual-based models’ (Hanski 1998; Larson et al. 2009; Franklin 2010;

Gallien et al. 2010). Spatially-explicit forms of these dynamic models often integrate abiotic and

structural variables (e.g. patch size, or outputs from other habitat models) and may therefore be used

to infer aspects of either the fundamental or realized niche. ‘Mechanistic home range analyses’ are a

widely used spatially-explicit application of the individual-based modelling framework. By

modelling the process of individual movement in environmental space (as a function of behavioural

responses to abiotic and biotic and habitat variables that can be linked to geographical space), such

models can make inferences about home range behaviour (Mitchell & Powell 2004; Börger et al.

2008; Mitchell & Powell 2012; Moorcroft 2012). When projected into geographical space,

simulation models are most useful for inferring dynamic response variables such as actual space

use, abundance, persistence, or population viability.

2.4.2 Bayesian networks

‘Bayesian networks’ are a general, flexible approach and have been applied in diverse contexts

(McCann et al. 2006; Wilhere 2012; Fenton & Neil 2013; Landuyt et al. 2013). Bayesian network

models are graphical influence diagrams, in which explanatory variables (‘parent nodes’) are linked

by causal relationships to response variables (‘child nodes’). Each model variable has at least two

mutually exclusive states, which exhaustively capture its total range of values or conditions. Causal

relationships and interactions between several parent nodes are quantified in ‘conditional

probability tables (CPT)’ behind each child node in the network (Cain 2001; Chen & Pollino 2012;

Fenton & Neil 2013). Because CPTs calculate responses as probability distributions based on

Bayes’ Theorem, they explicitly capture uncertainties about these relationships (Chen & Pollino

2012; Wilhere 2012). CPTs can be parameterized from data (empirical or model-derived) or expert

knowledge (published or elicited).

When applied to habitat modelling, Bayesian networks can be used in either a statistical data-

driven, or process-based knowledge-supported framework (Boets et al. 2015). ‘Naïve Bayesian

classifiers’ learn the joint probability of a response variable from its relationships to a range of

explanatory variables in a training data set, similarly to logistic regression or other statistical

machine learning techniques (Lorena et al. 2011; Boets et al. 2015). Bayesian networks have far

more frequently been applied as resource- or process-based habitat models, where they offer unique

strengths (Marcot 2006; McCann et al. 2006; McNay et al. 2006; Douglas & Newton 2014;

Tantipisanuh et al. 2014; Hamilton et al. 2015). Just like most other habitat models (Section 2.4.1),

Bayesian networks model relationships in environmental space. Marcot et al. (2001) has proposed a

spatially-explicit framework, which was subsequently refined by Smith et al. (2007) and explained

in detail by van Klinken et al. (2015). Here, habitat suitability was conditional on a set of habitat

47

variables representing resource requirements (for animals) or invasion processes (for invasive

plants). Each habitat variable was itself influenced by several measurable explanatory variables and

each explanatory variable was linked to one or more remotely sensed or mapped spatial data

proxies.

The advantages of the resource-based, spatially-explicit Bayesian network approach include: a

robust statistical framework for modelling interactions between habitat variables based on species

ecology; flexible data needs, including the ability to integrate unpublished expert knowledge; and

explicit representation and propagation of uncertainty throughout the model, which allows

computation of habitat suitability as a probabilistic index (Uusitalo 2007; Wilhere 2012; Landuyt et

al. 2013). There are also several important limitations. First, model parameterization can be

challenging when the model includes many interactions between parent nodes because a conditional

probability distribution must be specified for each combination of states in all parents linked to a

child node (Cain 2001; Chen & Pollino 2012). If using expert knowledge, simplifying elicitation

techniques must be carefully considered (Cain 2001; Fenton et al. 2007; Section 2.6). Second,

Bayesian networks represent temporal dynamics poorly (McCann et al. 2006). Yet, they lend

themselves to a scenario approach, which has previously been applied to future climates (Murray et

al. 2012), land management (Smith et al. 2012) or population control (Murray et al. 2014), and

could equally be used to represent a series of time horizons (such as the seasonal scenarios used in

this thesis; Chapter 4). Third, while integration with spatial data layers has been frequent, it has

been applied in a non-standardized and often complex form (Johnson et al. 2012a). Yet, this

problem may be alleviated in future by recent technical advances (Landuyt et al. 2015).

2.4.3 Landscape variables

A further important limitation of Bayesian networks, which they share with most other habitat

modelling approaches outlined in Section 2.4.1, is that they are inherently spatially ‘neutral’ or

‘implicit’ (Johnson et al. 2012a). While both explanatory and response variables are routinely

projected into geographical space, relationships are modelled purely in environmental space and are

therefore “blind” to landscape heterogeneity and spatial interactions between variables (Elith &

Leathwick 2009; Yackulic & Ginsberg 2016). For mobile species such as feral pigs, which use

complementary and supplementary resources at different locations within heterogeneous home

ranges, integrating spatial relationships between habitat variables into the modelling process is

critical (Guisan & Thuiller 2005; McGarigal et al. 2016). While several methodologies for doing

this have been developed, these are not routinely integrated into either habitat modelling approach

(Beck & Suring 2009; Moorcroft 2012; McGarigal et al. 2016; Yackulic & Ginsberg 2016). To my

knowledge, resource-based Bayesian network habitat models have rarely addressed this issue

48

(Landuyt et al. 2015). Broadly, three methodologies for integrating ‘landscape variables’ that

measure spatial heterogeneity using various metrics of landscape composition or configuration in

home ranges, or more generally ecological neighbourhoods (Addicott et al. 1987; Holland & Yang

2016), have emerged:

Statistical resource selection functions and empirical habitat studies have adopted a ‘multi-

scale’ approach, whereby landscape variables were computed in circular areas of different radii

around a sampling site. ‘Scale-optimized’ variables that best explain a given response were then

selected and relationships between them modelled in environmental space (Holland et al. 2004;

Jackson & Fahrig 2015; McGarigal et al. 2016; Miguet et al. 2016). In resource selection functions,

for example, multiple scales effectively constrained the area in which resource availability was

measured and subsequently contrasted with actual use (DeCesare et al. 2012; Martin & Fahrig

2012; Laforge et al. 2016; Paton & Matthiopoulos 2016). Step selection functions have further

allowed this area of availability to shift following an animals’ movements or weighted availability

in the circle according to behavioural criteria, but have usually not used multiple scales of

measurement (Arthur et al. 1996; Rhodes et al. 2005; Moorcroft 2012).

Some recent statistical species distribution models have also adopted the multi-scale approach.

These computed landscape variables in moving windows that are incrementally centred on each

‘focal pixel’ within a study area, thereby essentially converting spatially independent variables into

landscape-scale variables that summarize information contained in each pixel’s neighbourhood

(Guisan & Thuiller 2005; Yackulic & Ginsberg 2016). Generally, scale-optimized landscape

variables were then selected in a univariate model and the relationships between all variables

investigated in a multivariate model (Bellamy et al. 2015; Ducci et al. 2015). One drawback of all

statistical approaches is that the functional relevance of included landscape variables has not always

been explicit and meaning is derived solely from correlations. This may be problematic because

scale-dependent insights may not be transferrable across larger study areas (Paton & Matthiopoulos

2016

Therefore, resource-based ‘landscape HSI models’ have been applied to large landscapes.

These studies used published or expert a priori knowledge to select landscape variables measured at

a specified scale, e.g. based on an animal’s known home range size or habitat use (Mitchell et al.

2002; Dijak et al. 2007; Dijak & Rittenhouse 2009). Similar to the statistical approaches described

above, neighbourhood summary variables were then computed in moving windows for each focal

pixel. Importantly, these structural measures were related to functional suitability indices,

describing the value of a measured spatial pattern to a given species, using expert knowledge or

response functions that were assumed to apply across the entire study area (Dijak & Rittenhouse

49

2009). The main disadvantage of this method is that landscape variables are rarely empirically

scale-optimized and their relationship to habitat suitability or species occurrence rarely validated

(Beck & Suhring 2009; McGarigal et al. 2016).

2.5 Habitat connectivity models

2.5.1 Modelling frameworks

In accordance with the breadth of conceptual perspectives on landscape connectivity (here

‘habitat connectivity’; Section 2.1.5), a variety of modelling frameworks have been developed.

Kool et al. (2013) and Fletcher et al. (2016a) provided excellent syntheses of the most commonly

used approaches. As opposed to the habitat models discussed in Section 2.4.1, most frameworks

explicitly incorporate spatial relationships between habitat patches in geographical space. They

differ, however, in their treatment of structural patterns, and functional dispersal processes, in the

intervening matrix. Approaches also differ in focus, describing connectivity as either an attribute of

the patch or the matrix (Moilanen & Hanski 2001; Calabrese & Fagan 2004; Fletcher et al. 2016a).

Here, I refer to measures of ‘matrix connectivity’ or ‘patch connectivity’ to emphasize this

distinction (Calabrese & Fagan 2004; Kindlmann & Burel 2008; Rayfield et al. 2011).

‘Spatial metapopulation models’ have focused on population dynamics within habitat patches,

depending on their location in relation to other patches (Hanski 1998; Fletcher et al. 2016a).

However, patch connectivity has typically been measured in terms of simple ‘isolation-by-distance’

(e.g. using dispersal kernels and measuring nearest neighbour distance) or patch aggregation (e.g.

using buffer measures) (Moilanen & Nieminen 2002; Calabrese & Fagan 2004; Kindlmann & Burel

2009). Landscape genetic studies have similarly compared genetic distances with Euclidean

distances between discrete subpopulations (McRae & Beier 2007). In either approach, structural

heterogeneity in the matrix and its influence on dispersal was rarely integrated (Kool et al. 2013).

On the other end of the spectrum, ‘spatial pattern indices’ have focused solely on quantifying

aspects of landscape structure in both patches and the matrix. They provided measures of structural

matrix connectivity without explicit reference to an organism or dispersal process (Calabrese &

Fagan 2004; Kindlmann & Burel 2009).

Recently, graph-theoretic approaches have become the dominant paradigm in connectivity

research (Urban et al. 2009; Fletcher et al. 2016a). The basic concept is well-defined by Rayfield et

al. (2011): “In its most basic form, a graph is a set of nodes, some pairs of which are joined by

links. [When applied to habitat patch networks,] a “habitat graph” [is] a collection of nodes (habitat

patches) and links that connect pairs of nodes (representing the potential or frequency of movement

between habitat patches)” (pp. 847 & 849). The framework is extremely flexible and has been used

50

to measure structural or functional connectivity as an attribute of individual patches (‘nodes’),

dispersal corridors (‘links’) or the entire patch network, depending on how nodes and links are

represented (Galpern et al. 2011; Rayfield et al. 2011). For example, links can be directional or

non-directional and informed by Euclidean distance or matrix resistance (Section 2.5.3). Matrix

resistance to dispersal has typically been integrated into patch network graphs via ‘least cost path’

algorithms (Galpern et al. 2011; Etherington & Holland 2013). These algorithms determined the

optimal, most efficient route of dispersal between two nodes (patches) “as a function of the distance

travelled and the costs traversed” (Etherington & Holland 2013, p. 1223).

However, real organisms may rarely use the least cost path or “optimal” route of dispersal.

Further, Moilanen (2011) highlighted that patch network graphs are limited by arbitrary patch

delineation and dispersal thresholds. ‘Circuit-theoretic models’, which are also based on graph and

network theory and have mostly been implemented in the Circuitscape modelling environment

(McRae et al. 2008; McRae et al. 2013), have received increasing attention. Circuitscape treats the

landscape analogous to an electrical circuit, passing current between pairs of nodes (patches)

through an intervening network of resistors (matrix pixels). Circuitscape has been mostly used to

compute the resistance distance of a patch as an improved measure of genetic connectivity (McRae

& Beier 2007). However, the algorithm also computes the current density in each matrix pixel,

which can be interpreted as the probability of a species moving through that pixel via random walk

theory (McRae et al. 2008). As relative flow rates are measured, Circuitscape can identify multiple

alternative connections between nodes rather than a single least cost path. Further, recent technical

developments have allowed computation of movement probabilities (i.e.’ matrix connectivity’)

across large landscapes independent of habitat patches (Koen et al. 2014; Pelletier et al. 2014).

Finally, a disadvantage of all graph-theoretic models is that they are static representations of

reality and typically do not integrate population dynamics in nodes or spatiotemporal variability in

links. More complex ‘dynamic network models’ (Ferrari et al. 2014), ‘spatial metapopulation

models’ (Hanski 1998; Holland et al. 2007; Lurgi et al. 2016), ‘individual-based models’

(Lookingbill et al. 2010; Kanagaraj et al. 2013) or ‘spatially structured diffusion models’ may be

best suited when the goal is to measure actual connectivity and its effects on metapopulation

persistence (Moilanen 2011; Kool et al. 2013; Hastings 2014; Fletcher et al. 2016a).

2.5.2 Habitat patches

Most connectivity modelling frameworks and underlying ecological theories require that

habitat patches are spatially defined as a basic unit of analysis (Girvetz & Greco 2007). Early

approaches have regarded patches as structurally distinct landscape features and attempted to

delineate them based on rules of spatial contiguity. However, structural patches have little basis in

51

ecological theory (Section 2.1.2). Contiguity rules are also sensitive to the spatial grain of analysis,

disregard patch configuration (e.g. shape and edge effects) and composition (e.g. size), and may

overemphasize the disruptive force of narrow intervening areas of poor quality habitat (Beier et al.

2007; Girvetz & Greco 2007). Yet, until Girvetz & Greco’s (2007) important paper, relatively little

progress had been made towards functional patch delineation consistent with ecological theory.

Functional approaches to patch delineation can be applied to multiple biological levels and

time horizons (e.g. resource patches used by an individual for minutes to hours vs. habitat patches

used by local populations for months to years). Patches may also be referenced to specific time

periods to allow for dynamic shifts through time. There is a range of methods available, but all

require the definition of a three quantitative parameters (Girvetz & Greco 2007; Galpern et al. 2011,

Shirk & McRae 2013): (1) a quality threshold above which a landscape element (e.g. a spatial pixel

on a map) is considered as suitable habitat for the organism; (2) a size threshold, which is deemed

necessary for a breeding pair or population to persist without interaction with other patches; and (3)

one or several functions describing the effect of low quality matrix habitat adjacent to or within a

patch. The latter can allow for patch expansion across narrow intra-patch gaps that do not affect the

focal species, or patch removal, for example when edge effects impact negatively on the suitability

of narrow habitat “spurs” (Girvetz & Greco 2007). Parameterization usually relies on prior

empirical knowledge. For example, results from habitat suitability models may be used to define

quality thresholds, while size thresholds are commonly derived from home range estimates or

assumed breeding requirements. Intra-patch gap crossing ability or edge effects may be derived

from empirical knowledge on species mobility and habitat use (Beier et al. 2007; Shirk & McRae

2013). However, Moilanen (2011) warns that multiple thresholding may result in considerable loss

of information.

2.5.3 Matrix resistance

Separating the landscape into suitable patches and an unsuitable matrix can be problematic, for

example when generalist species utilize continuously distributed resources (Hamilton et al. 2006).

Moreover, the shift towards functional perspectives in landscape ecology has led to the recognition

that animal dispersal between patches is influenced by the variable conditions encountered in the

matrix. The purpose of ‘resistance surfaces’ is to quantify structural heterogeneity in the landscape

between patches, or indeed the entire landscape, from the perspective of a specific organism moving

through that landscape (Zeller et al. 2012). Resistance surfaces assign species-specific cost values

to each landscape element (e.g. pixel) and have also been referred to by “combinations of: cost,

friction, permeability, or resistance, and; layer, grid, map, raster, or surface” (Etherington et al.

52

2014, p. 1). They are a crucial, and sometimes the only, input for functional connectivity models

(Beier et al. 2007; Koen et al. 2014).

Resistance surfaces can be parameterized in many ways. Zeller et al. (2012) provide an

excellent review of the different approaches available. The most common types of data used for

model calibration were expert knowledge, followed by information from landscape genetic studies

and single-point observations (Zeller et al. 2012). Direct empirical measures of animals’ long range

(inter-patch) dispersal movements, while conceptually best suited to infer matrix resistance, have

been used infrequently (Zeller et al. 2012; Kool et al. 2013; Fletcher et al. 2016a). The most

common analytical approach was described by Zeller et al. (2012) as the ‘one-stage expert

approach’. Here, biologically informed expert opinion was used to assign resistance values to

landscape elements without further verification against empirical data. When used within broader

resource-based habitat modelling frameworks (Hartemink et al. 2015), landscape resistance has

often been calculated by reversing habitat suitability (Beier et al. 2007; Murray et al. 2014).

However, this relies on the assumption that habitat selection within patches and dispersal between

patches are influenced by the same environmental variables (Beier et al. 2007). For feral pigs, this

assumption may be justified, as dispersal appears to be limited solely by the (gender-specific)

energetic and survival costs associated with moving through poor quality habitat (Section 2.2.4).

2.6 Expert elicitation

Faced with uncertainty, complexity and lack of empirical data, decision-makers have always

relied on the knowledge of experts. The value of expert knowledge to inform landscape ecological

research and modelling has also increasingly been recognised (Perera et al. 2012). The process of

collecting, synthesising and analysing expert knowledge is commonly referred to as ‘expert

elicitation’. Yet, expert knowledge as well as the process of eliciting it are prone to error and bias.

Consequently, a number of recent publications have argued that expert knowledge must not be

perceived as a “cheap, readily available source of knowledge” (McBride 2013, p. 156) and

advocated a structured approach that treats expert elicitation as a form of scientific data collection

(Kuhnert et al. 2010; Johnson et al. 2012b; Martin et al. 2012; McBride et al. 2012a). Like all data,

expert knowledge has limits: it is subjective, incomplete and often unstructured. Elicitation cannot

remove uncertainty or knowledge gaps. Rather, the goal is to minimize cognitive and motivational

biases and provide “an accurate representation of an expert’s true beliefs” (McBride 2013, p. 7). A

structured elicitation approach should follow a three-step protocol: (a) problem definition,

development of an elicitation methodology and expert selection; (b) elicitation pre-training and

actual elicitation; and (c) verification, error analysis, aggregation and validation (McBride 2013).

For expert-based Bayesian networks, Marcot (2006) recommended integrating these steps into an

53

iterative elicitation process that allows for repeated evaluation of model behaviour, validation of

model accuracy and opportunities for expert revision. Within this framework, a range of strategies

and techniques can be used to avoid or minimize bias (Table 2.1).

Table 2.1 Sources of bias in expert elicitation and strategies or techniques for minimizing them as used

in this thesis. (Sources: Kuhnert et al. 2010; McBride et al. 2012a; Fenton & Neil 2013; McBride 2013)

Source of bias Elicitation strategies and techniques employed in this thesis

Motivation / elicitation fatigue: Experts

have other commitments and need to

justify time investment to themselves

and others.

⋅ Develop rapport and shared understanding / support for elicitation goals

⋅ Elicit in workshop setting to enhance professional experience

⋅ Limit time investment and number of answers required

⋅ Provide rapid feedback and follow-up to maintain interest

Anchoring bias: Expert beliefs are

influenced by a preconceived estimate

⋅ Develop a joint system understanding and conceptual models through

group consensus in a participatory setting

Uncertainty:

⋅ Experts are uncertain of their, or have

no, knowledge on elements of the

system

⋅ Experts know that there is uncertainty

inherent in elements of the system

⋅ Select experts from diverse backgrounds, preferably well-calibrated ‘expert

practitioners’ that interact with the model on a regular basis

⋅ Stimulate memories and understanding through group interaction

⋅ Maintain simplest possible conceptual model structure

⋅ Use a probabilistic framework that allows for uncertainty

⋅ Validate elicitation results against independent data

Overconfidence: Experts overestimate

the accuracy of their beliefs

⋅ Select ‘expert practitioners’

⋅ Use a probabilistic framework

Variability: Experts have different

system understanding and uncertainties

⋅ Elicit quantitative estimates from individuals and average results

⋅ Evaluate model sensitivity to different expert estimates

Dominance: Experts conform to

seniority or dominance

Groupthink: Groups want to achieve

consensus and dismiss deviations

⋅ Elicit system understanding and quantitative estimates in small groups or

individually followed by open review and discussion

⋅ Encourage group participation and consider alternative model scenarios

Linguistic: Experts interpret statements

and questions differently

⋅ Develop clear definitions and shared understanding for all system elements

⋅ Ask simple, unambiguous questions

⋅ Frame probabilities as interval judgements and interactions as weights

⋅ Document elicitation process and results to allow retrospective evaluation

Ambiguity: Experts avoid options where

they have no knowledge or uncertainties

⋅ Review model structure and behaviour in an iterative process

⋅ Delphi approach: elicit individual knowledge, share among the group and

give opportunity for individual revision

54

Chapter 3 Integrating landscape structure improves habitat models of mobile

animals: feral pigs in Australia

Summary

In this chapter I develop a novel combined methodology, spatial pattern suitability analysis, for

integrating resource-seeking home range movements into habitat models of mobile animals7. This

involves measuring structural patterns of resource quality at the home range scale and then relating

these measures to functional values from expert-elicited response-to-pattern curves. I use scenario

analysis and empirical validation in a subsection of the study region in northern Queensland to

evaluate whether this methodology improves model performance and how structural patterns should

be measured for feral pig breeding herds. This chapter addresses research question 1, providing a

methodology for reliably modelling habitat suitability for feral pigs at the regional scale under

uncertainty about the processes of habitat selection by individual breeding herds. It is being

prepared for submission to Ecological Modelling.

3.1 Abstract

Ecological theory suggests that habitat suitability for mobile animals is influenced by

landscape structure: the spatial patterns of, and relationships between, resource or habitat patches

within heterogeneous home ranges. Yet, many habitat and distribution models of such species either

neglect landscape structure or do not sufficiently evaluate alternative ways of measuring it. We

modelled habitat suitability for feral pigs in northern Australia using a resource-based approach.

Here, we aimed to (a) integrate measures of landscape structure at the home range scale into this

framework, and (b) evaluate which landscape metrics and scales of measurement yield improved

models. We developed spatial pattern suitability analysis, which measured patterns of resource

quality in moving windows and related these structural metrics to functional resource suitability

indices from expert-elicited response-to-pattern curves. We developed eleven alternative habitat

suitability models with and without integrating landscape-scale variables using one of three metrics

(distance-dependent, composition-dependent and combined distance / composition-dependent) and

analysis scales (1, 2 and 3 km radius). All models were empirically validated against independent

distributional data. Habitat suitability models which integrated landscape structure outperformed the

model which did not in 90% of evaluated scenarios. Models that measured resource suitability as a

7 A detailed description of the methodology for modelling habitat suitability using Bayesian networks, including expert elicitation procedure, model variables, and spatial data proxies, is given in Chapter 4. This is due to the timing of journal submissions. The final submission of Chapter 3 to Ecological Modelling will reflect this and refer to Chapter 4 (published in PLoS ONE) for details. Here I focus on the spatial pattern suitability analysis methodology and its evaluation.

55

function of distance performed consistently best against all validation data. Spatial pattern

suitability analysis usefully captured expert knowledge on home range movements. Through model

validation we could determine which landscape variables most improved predictions of habitat

suitability. Our approach is applicable to other mobile species, modelling frameworks and

landscape metrics.

3.2 Introduction

Spatial models of species’ habitat and distributions are now widely applied to inform

management decisions (Millspaugh & Thompson 2009; Venette et al. 2010; Guisan et al. 2013)

from local to global extents (Fletcher et al. 2016b). While methodologically diverse (Guisan &

Thuiller 2005; Beck & Suring 2009; Dormann et al. 2012), all rely on the link between an

organism’s distribution and a combination of habitat variables or environmental requirements

(Grinnell 1917; Hutchinson 1957; Leibold 1995; Pulliam 2000). At the same time, decades of

biogeographical and landscape ecological research have established that species distributions and

habitat use are influenced by landscape structure – the spatial patterns of, and relationships between,

required resources or habitat patches within heterogeneous landscapes (MacArthur & Wilson 1967;

Dunning et al. 1992; Levin 1992; Turner & Gardner 2015) – and that this influence is sensitive to

both the scale and metric used for measuring these patterns (Addicott et al. 1987; Wiens 1989;

McGarigal & Marks 1995; Holland et al. 2004). Despite the theoretical importance of landscape

structure and scale, several recent reviews have highlighted that these issues are not routinely

integrated into contemporary habitat and species distribution models (Beck & Suring 2009;

Moorcroft 2012; McGarigal et al. 2016; Yackulic & Ginsberg 2016). Here, we investigated the

integration of landscape structure into a regional-scale habitat models of mobile animals.

Mobile animals have home ranges, broadly defined as the “area traversed by an individual in

its normal activities of food gathering, mating, and caring for young” (Burt 1943, p. 352). While

many theoretical and operational questions about home range behaviour remain, the conceptual

basis of animals’ space use is widely agreed upon: mobile animals use home ranges to satisfy a

number of heterogeneously distributed habitat requirements (Börger et al. 2008; Powell & Mitchell

2012). Here, we view home ranges as a type of ‘ecological neighbourhood’ (Addicott et al. 1987) in

which resource selection by mobile animals occurs (Fig 3.1). Heterogeneity in this neighbourhood

can be measured using a range of ‘landscape metrics’, which describe aspects of landscape

‘composition’ (the variety and abundance of resources) or ‘configuration (the relative position of

resources) (McGarigal & Marks 1995; Uuemaa et al. 2009; McGarigal et al.2012). Yet, all metrics

of landscape structure, per se, are independent of ecological function, that is, they carry no implicit

meaning to the way a mobile species uses its home range (Kupfer 2012; Turner & Gardner 2015).

56

Here, Dunning et al. (1992) suggested that habitat selection in home ranges is governed by resource

‘complementation’ (several resources are selected for different purposes), ‘supplementation’

(different resources are selected for the same purpose), and ‘neighbourhood effects’ (proximal

resources are selected preferentially to distant resources). The suitability of a site to function as

habitat thus depends on the availability and patterning of complementary and supplementary

resources with the home range, and the species’ behavioural response to these patterns (Fig 3.1).

Fig 3.1 Resource-seeking home range movements of mobile animals. Consider two mobile species (red

and black squiggles) with three resource requirements, all heterogeneously distributed across a landscape

(dark red, light red and grey patches). Landscape portion (A) contains few scattered and portion (B)

abundant contiguous resource patches. Each species uses space according to its mobility and corresponding

home range size (black dashed circles). Suitability of two focal habitat patches (black squares) is influenced

by: the presence of resources in the landscape (greater in portion B than A); the scale of each species’ home

range in which resources are available (red species can access all three resources, black species only two, but

note the difference if home range sizes were reversed (grey dotted circles)); and the behavioural response of

each species to landscape structure (How much of each resource is required? Are distant resources used less

than proximate resources? Are scattered resources used less than contiguous resources?).

Broadly, three methodologies for integrating ‘landscape variables’ that measure spatial

patterning in explanatory variables into habitat models have emerged: Statistical models using

resource selection functions have adopted a ‘multi-scale’ approach, whereby landscape variables

were computed in circular areas of different radii around a sampling site, and ‘scale-optimized’

variables that best explain a response variable were selected (Holland et al. 2004; Jackson & Fahrig

2015; McGarigal et al. 2016; Miguet et al. 2016). Some recent statistical species distribution

models have also adopted the multi-scale approach, but computed landscape variables in moving

windows that are incrementally centred on each ‘focal pixel’ within a study area (Bellamy et al.

2015; Ducci et al. 2015). As scale-dependent insights from empirical multi-scale studies may not be

57

transferrable across large study areas (Paton & Matthiopoulos 2016), resource-based ‘landscape

HSI models’ have used published or expert a priori knowledge to derive suitable landscape metrics

and measurement scales (Dijak et al. 2007; Dijak & Rittenhouse 2009). Landscape variables were

then equally computed in moving windows and, importantly, related to functional suitability indices

that are assumed to hold for entire large landscapes (Dijak et al. 2007; Dijak & Rittenhouse 2009).

Here, we aimed to account for home range movements when modelling habitat suitability for

feral pig breeding and persistence in north-eastern Australia. Feral pigs are descendants of

domesticated Eurasian wild boar. The species has proven a successful invader wherever introduced,

impacting negatively on agriculture, the environment and human and animal health (Barrios-Garcia

& Ballari 2012; Bengsen et al. 2014). Regional-scale information about their distribution and

habitat suitability for persistence is needed to effectively manage these impacts. Feral pigs’ habitat

(resource) selection, space use and distributional patterns have been studied widely. However, the

influence of landscape structure and scaling has rarely been analysed. For example, resource

selection analyses have established that feral pigs prefer habitats which satisfy their physiological

needs for frequent hydration (Caley 1997), thermoregulation (Dexter 1998) and a protein-rich diet

(Giles 1980; Caley 1997; Dexter 1998; Ross 2009; Wurster et al. 2012). Yet, only Choquenot &

Ruscoe (2003) explicitly investigated resource complementation at different scales and found that

food resources and thermal refuge must co-occur within 5 km to facilitate feral pig persistence.

Home range analyses have found movements to be limited by the seasonal availability of these

resource requirements, especially in breeding herds of related sows and their young (Caley 1997;

Dexter 1999; Mitchell et al. 2009). Female home range sizes ranged widely (1-20 km2 – Hone

2012; Choquenot et al. 1996), but possible effects of landscape structure remain unclear.

We adopted a resource-based modelling framework using Bayesian networks. When used to

model habitat (Marcot et al. 2001; Smith et al. 2007; Tantipisanuh et al. 2014; van Klinken et al.

2015), this approach offers flexibility to use expert-elicited knowledge where empirical data

availability is limited, a robust statistical framework for modelling interactions between

complementary resources, and an explicit treatment of uncertainty (Uusitalo 2007; Wilhere 2012).

Our objectives were: (A) To develop a methodology for integrating landscape structure at the home

range scale into this resource-based modelling framework. We limited our analysis to three simple,

yet ecologically meaningful, landscape metrics: distance-weighted resource quality (a measure of

landscape configuration), average resource quality (a measure of landscape composition) and a

combined measure, distance-weighted average resource quality. (B) To test whether such

integration actually improves our model and, if so, which metrics and scales of measurement best

explain observed feral pig presences across the study region. Validation is critical especially for

58

expert-elicited models (van Klinken et al. 2015). We evaluated ten alternative ‘landscape models’

(which used landscape-scale habitat variables using either of the three metrics at either of three

expert-defined measurement scales) and one ‘control model’ (which used site-scale habitat

variables) against independent distributional data, expecting landscape models to perform better.

3.3 Methods

3.3.1 Study region

Our study region in north-eastern Australia (Fig 3.2) is characterised by a hot tropical climate

with a highly seasonal rainfall regime. Ecosystems are dominated by monsoonal savanna

woodlands, extensively utilized for cattle grazing and interspersed with fragments of wet tropical

rainforest, coastal wetlands and open semi-arid grasslands (Fox et al. 2001). Intensive uses are

limited to the fertile coastal lowlands and plateaus of the eastern ranges. According to previous

estimates, feral pigs are common to widespread across the region (West 2008), yet distribution and

abundance varies considerably with climatic and seasonal conditions (Choquenot et al. 1996). We

limited our habitat model to the late dry season (October to November), when resources are most

constrained and scattered across the landscape. Increased heterogeneity under such conditions may

pronounce the influence of landscape structure on habitat suitability.

Fig 3.2 Study region in north-eastern Australia. The study region is shown in grey. Locations of the

independent distributional data sets used for model validation are shown in colour (details in Table 3.1).

59

3.3.2 Modelling approach

In order to evaluate which landscape metric and scale of measurement best explains feral pig

presence in north-eastern Australia, we developed eleven alternative habitat suitability models. Our

modelling approach (Fig 3.3) adapted the conceptual framework described in Marcot et al. (2001),

Smith et al. (2007) and van Klinken et al. (2015): habitat suitability was defined in terms of feral

pigs’ ability to breed and persist over time (Morrison et al. 2006); this depended on a set of habitat

variables representing resource requirements; each habitat variable was itself influenced by several

explanatory variables, which were linked to one or several spatial data proxies.

First, we modelled the quality of each habitat variable (henceforth termed ‘resource quality’),

conditional on its explanatory variables, in separate expert-elicited Bayesian networks. Then, we

used ‘spatial pattern suitability analysis’ to capture feral pigs’ selection of supplementary resources

within home ranges. This involved measuring spatial patterns of resource quality within moving

windows and relating structural metrics to functional ‘resource suitability indices’ (SIr) from expert-

elicited response curves. Here, we computed distance- (SIDr), composition- (SICr) and combined

distance / composition-dependent (SIDCr) indices at three scales of measurement each (1, 2 and 3 km

moving window radii). Finally, we modelled habitat suitability, conditional on all habitat variables,

in another expert-elicited Bayesian network. Ten ‘landscape models’ integrated landscape-scale

variables (SIr), while one ‘control model’ used site-scale variables (xr) (Fig 3.3).

Fig 3.3 (next page) Methodology for modelling habitat suitability for feral pig breeding. Major steps

with associated methods and implementation tools. Steps 1 and 2 are illustrated using the example of food

resources. Analogous procedures were applied to the other three identified habitat variables (water,

protection from heat stress and protection from disturbance). Bayesian network results are shown as expert-

elicited probability distributions (bar graphs) and modelled index values ± standard deviation (below bar

graphs), conditional on uniformly distributed prior probabilities. Spatial pattern suitability analysis (here

shortened to PATTSI) was used to compute three alternative landscape-scale resource suitability indices

(SIr): distance- (SIDr), composition- (SIr) or combined distance / composition-dependent (SIDCr). Step 3 is

illustrated using SIDr. Table 3.2 details landscape variables used in habitat suitability models 1-10.

60

Food quantityHighModerateLow

33.333.333.3

Food qualityHighModerateLow

33.333.333.3

Food quality indexVery goodGoodModeratePoorVery poor

19.523.222.513.921.0

51.2 ± 29

Habitat suitability indexVery highHighModerateLowVery low

2.0112.026.121.338.6

33.5 ± 23

Heat protection suitability indexVery goodGoodModeratePoorVery poor

20.020.020.020.020.0

Food suitability indexVery goodGoodModeratePoorVery poor

20.020.020.020.020.0

Water suitability indexVery goodGoodModeratePoorVery poor

20.020.020.020.020.0

Disturbance protection suitability indexVery goodGoodModeratePoorVery poor

20.020.020.020.020.0

r {raster} / ArcGIS – spatial data

pre-processing

Expert-elicited Bayesian network

model

ArcGIS – output mapping

Expert-elicited response curves

r {raster} – focal (max) /

ArcGIS Spatial Analyst –

focal (mean)

Expert-elicited Bayesian network

model

1. Model resource quality indices

2. Spatial pattern suitability analysis

3. Model habitat suitability index

Step Methods Tools

Focal (max)

Focal (mean)

Focal (mean)

Scen

ario

X (w

ithou

t PAT

TSI) Sc

enar

ios

1 –

10

(with

PAT

TSI)

Scen

ario

s 1

– 10

(w

ith P

ATTS

I)

SICr SIDCr SIDr

61

3.3.3 Resource quality models

Models were calibrated by iteratively eliciting knowledge from a panel of experts with

scientific or field knowledge of feral pigs in north-eastern Australia (Marcot 2006). Elicitation

followed a structured process and used a range of strategies and techniques for minimizing sources

of bias (Table 2.1). Experts identified four resource requirements for feral pig breeding and

persistence: water, food, protection from heat stress and protection from disturbance. We modelled

resource quality as a function of several explanatory variables in probabilistic Bayesian networks

using the Norsys Netica 5.12 and AgenaRisk 6.1 software. Model explanatory variables were linked

to remotely sensed or mapped GIS data layers, which were rasterized and resampled to a common

extent and resolution (100m), then reclassified using the R ‘raster’ package (R Core Team 2015;

Hijmans 2015) and ESRI ArcGIS 10.2 software. Time-specific spatial data were averaged over two

months (October / November) and five years (2010-2014) prior to reclassification to

representatively capture late dry season conditions. We discretised resource quality into five equal

classes and assigned each with a numerical range (0-20 for the poorest class, …, 80-100 for the best

class). Spatially-explicit ‘resource quality indices’ (xr) were computed as expected values by

summing the mid-point value of each class weighted by its probability of occurrence. Accordingly,

xr could range between 10 (mid-point of the poorest class) and 90 (midpoint of the best class). An

example model is shown in Fig 3.3. Tables defining all variables, conditional probabilities and

spatial data proxies are detailed in Appendix 3.18.

3.3.4 Spatial pattern suitability analysis

Spatial pattern suitability analysis combined to methodological steps: (a) spatial moving

window analysis to measure spatial patterns of resource quality (b) expert elicitation to relate

structural metrics to functional resource suitability indices (Appendix 3.2).

3.3.4.1 Expert elicitation

We conducted semi-structured interviews with selected individuals from our panel of feral pig

experts (n = 6). First, we asked each expert to define the home range ‘ecological neighbourhood’

(Addicott et al. 1987) of feral pig breeding by specifying a ‘mobility threshold’ beyond which

resources are inaccessible. We assumed that the same threshold applies to all habitat variables.

Experts estimated mobility thresholds at either 1 km (n = 1), 2 km (n = 2) or 3 km (n = 3) which

8 Appendix tables 3.1.1 – 3.1.3 in the present version of this manuscript (prepared for submission to Ecological modelling) have since been included in the main body and appendix of a manuscript published in PLoS ONE (Chapter 4). They are included as Table 4.1, Table 4.2 and Appendix Table 4.1 in this thesis. The final submission of Chapter 3 will reflect this and refer to Chapter 4 (published in PLoS ONE) for details.

62

corresponded to assumed home ranges of approximately 3, 12 and 28 km2. These behavioural

thresholds determined the three scales of measurement evaluated in this study.

Second, we asked experts to relate structural metrics of resource patterns to their functional

value for feral pig breeding. Similar to other studies (Larson et al. 2009; Dijak & Rittenhouse

2009), we expressed these values as ‘resource suitability indices’ (SIr). Like resource quality,

suitability was discretised into five equal classes (0-20 for the poorest class, … , 80-100 for the best

class). We focused elicitation on two metrics: distance and average quality.

Distance was defined relative to the mobility threshold, and discretised into five equal distance

bands (‘very close’, ‘close’, ‘medium’, ‘far’ and ‘very far’). For each habitat variable, we asked

experts to relate each distance band to a corresponding suitability class under the assumption that

other factors (e.g. amount) do not constrain suitability. Average quality depended on both the

amount and quality of accessible resource patches. Hence, different patterns (e.g. many low quality

resources or few high quality resources) could yield similar average values. Again, we defined five

equal average quality classes (‘very high’, ‘high’, ‘moderate’, ‘low’ and ‘very low’), relative to the

maximum value (100% cover with resources of highest quality xr max). Experts related each to a

corresponding suitability class, assuming no other constraints. Thus, each expert defined two step-

wise response-to-pattern curves fDr (distance-dependent measurement) and fCr (composition-

dependent measurement) per habitat variable. Finally, we averaged response curves for each habitat

variable across experts for input into moving window analysis (Fig 3.3).

3.3.4.2 Moving window analysis

Moving window analysis has been extensively used to investigate landscape structure (Dijak et

al. 2007; Ducci et al. 2015): it computes the value of a focal pixel by summarizing the values of all

neighbouring pixels contained within a spatial analysis window. This window is incrementally

shifted and centred on each pixel within a study region, enabling analysis of fine-grained patterns

within overlapping neighbourhoods (Larson et al. 2009). The moving window size and shape must

be specified and many landscape metrics are available to compute focal pixel values (McGarigal et

al. 2012). We used circular moving windows with radii corresponding to each of the three elicited

scales of measurement (1, 2 and 3 km). Three SIr were computed in each window (Fig 3.3),

resulting in nine alternative indices overall. Distance-dependent resource suitability (SIDr) was the

highest distance-weighted resource quality index (xr) within the moving window, where weights

were derived from response curves fDr. Composition-dependent suitability (SICr) was computed by

reclassifying average resource quality according to response curves fCr. To compute combined

distance / composition-dependent suitability (SIDCr), we also used fCr to reclassify average resource

quality. However, each individual resource quality index was also weighted according to response

63

curves fDr. Analyses were implemented in R 3.2.2 (R Core Team 2015) and ESRI ArcGIS 10.2

software (Appendix 3.2).

Fig 3.4 schematically illustrates that computed SIr at a focal pixel could vary considerably

depending on how spatial patterns are measured (Froese et al. 2015). Consider spatial resource

patterns A and B, which return the same SIDr (top left-hand plot), but very different SICr and SIDCr

(middle and bottom left-hand plot). Consider pattern C, which returns moderate SIr when measured

in a large window (black boxes in left-hand plots), but SIr = 0 when measured in a smaller window

(red boxes in left-hand plots). Non-linear response curves may alter these effects somewhat.

Fig 3.4 Computing resource suitability indices using spatial pattern suitability analysis. Consider a

resource of highest quality xr max (grey pixels) and various spatial patterns around a focal pixel (black pixel):

A = close and abundant; B = close and scarce; C = far and abundant; D = far and scarce. Three alternative

resource suitability indices (SIr) are computed within large (black dashed line) or small (red dashed line)

moving windows as a function of distance (SIDr, top left-hand plot), average quality (SICr, middle plot) or

distance-weighted average quality (SIDCr, bottom plot). We plot approximate SIr of the focal pixel

corresponding to each pattern A-D and moving window size (red boxes = small, black boxes = large)

assuming linear distance- and composition-dependent responses.

3.3.5 Habitat suitability model

We modelled habitat suitability for feral pig breeding and persistence as a function of four

complementary habitat requirements. Analogous to resource quality, models were implemented as

64

expert-elicited Bayesian networks and spatially-explicit habitat suitability indices (HSI) computed

by summing the mid-point values of five discrete classes weighted by their probabilities of

occurrence (Fig 3.3; Appendix Table 4.1E). We developed eleven alternative habitat suitability

models. For nine ‘landscape models’ we computed all four SIr using the same landscape metric and

scale of measurement. As feral pigs may respond differently to each habitat variable, we added a

‘best knowledge’ landscape model (derived from a combination of expert opinion and literature

review) with resource-specific SIr (distance-dependent SID water and SID heat at 3 km radius, combined

distance / composition-dependent SIDC food and SIDC disturbance at 2 km and 1 km radius respectively).

Lastly, a ‘control’ model used site-scale resource quality indices (xr) without spatial analysis.

3.3.6 Model evaluation and validation

3.3.6.1 Validation data

We validated predicted HSI from all eleven habitat suitability models against three independent

data sets of feral pig presence sourced from collaborators (Table 3.1). We used only presence

records corresponding to breeding herds (identified as female or with a group count greater than

two) and collected during the late dry season. For each data set, we defined a validation background

(shaded in red, blue and black in Fig 3.2) representative of surveyed habitat types from existing

management units (National Park boundaries for Lakefield) or by applying a 15km buffer to

presence records (for Balkanu and NAQS).

Table 3.1 Validation data sets with ancillary information.

Name Source No. of

records

Date of

collection

Method and purpose

of collection

Background

size (km2)

Typical habitat types

Balkanu Balkanu Cape

York Development

Corporation

181 Sep-Nov

2013-14

Systematic aerial survey

and management

(shooting)

3,954 Eucalyptus woodlands

& coastal wetlands

Lakefield Queensland

Parks & Wildlife

Service

350 Oct-Dec

2009-13

Systematic aerial

management (shooting)

5,788 Eucalyptus / Melaleuca

woodlands, coastal

wetlands & grasslands

NAQS Northern Australia

Quarantine Strategy

103 Sep-Nov

2007-10

Opportunistic aerial

survey and disease

sampling (shooting)

11,630 Eucalyptus woodlands,

coastal grasslands &

chenopod scrublands

Presence records for mobile species may contain spatial error for many reasons (Hunsacker et

al. 2001; Boyce et al. 2002). As our data sets had been collected by third parties, we faced

65

considerable uncertainty about such error. Similarly, remotely sensed or mapped data layers may

contain positional inaccuracies or misrepresent landscape features due to their mapping resolution

(Hunsacker et al. 2001). Together, such spatial error may yield noisy results that obscure main data

trends. To reduce this noise, we upscaled both predicted HSI and presence records to a 1km

resolution. We subsequently thinned data points to ensure independence, allowing only one

presence record collected on the same day within a given 1 km pixel.

3.3.6.2 Validation metrics

We used the Continuous Boyce Index (CBI) method described by Hirzel et al. (2006) after

Boyce et al. (2002) to evaluate model performance. The CBI was developed specifically for

situations where a numerical prediction such as our HSI is validated against presence-only

observations (Boyce et al. 2002). HSI was partitioned into n overlapping bins b of width w, which

were incrementally shifted upwards along the total range of HSI (HSImax – HSImin) by resolution

factor r (Hirzel et al. 2006). In our analysis, we set w to 10 and r to the default 1/100th of HSImax –

HSImin. The predicted-to-expected (P/E) ratio was then computed as the (predicted) proportion of

presence records in each bin b divided by the (expected) proportion of the validation background

covered by that b and plotted against the average HSI of b. For a good model, this results in a

monotonically increasing curve (Hirzel et al. 2006). The CBI measures the Spearman rank

correlation coefficient of P/E against HSI and varies from 1 (correct model, P/E steadily increases

as HSI increases) to –1 (false model, P/E steadily decreases with increasing HSI), with values close

to zero indicating a random prediction (Hirzel et al. 2006). We computed the P/E ratio and CBI for

each of the three validation data sets independently. We also computed the total proportion of each

validation background expected to be highly or very highly suitable habitat (HSI ≥ 60). Analyses

were implemented in R 3.2.2 (R Core Team 2015; Appendix 3.3).

3.4 Results

3.4.1 Spatial pattern suitability analysis

Expert-elicited response-to-pattern curves fr approximated either of three general shapes (see

Froese et al. 2015): linear decay (SIr declines steadily with increasing distance to, or decreasing

average quality of, a resource), exponential decay (SIr declines at a faster rate) and inverted

exponential decay (SIr declines at a slower rate). Individual experts’ fr are provided in Appendix

Table 3.1.4. After averaging across experts, different responses emerged for the four habitat

variables (Fig 3.5). When describing resource suitability for feral pig breeding as a function of

distance (fDr), these differences were substantial (Fig 3.5A). For instance, at ‘medium’ distance,

water resources were still considered highly suitable (SID water = 80) while disturbance refuges

66

already had poor functional value (SID disturbance = 22). The distance-dependent suitability of food

resources and heat refuges approximated linear decay (both SID food / heat = 58 at ‘medium’ distance).

The relationship between suitability and average quality (fCr) showed similar, but less pronounced,

differences between habitat variables (Fig 3.5B). For example, while 50% average resource quality

always corresponded to high SICr (SIC water = 90, SIC heat = 86, SIC food = 77 and SIC disturbance = 70), the

functional value at 25% and 15% average quality remained considerably higher for water (SIC water =

83 and 50) and heat (SIC heat = 78 and 50) than for food (SIC food = 53 and 30) and disturbance (SIC

disturbance = 50 and 26).

Fig 3.5 Averaged expert-elicited response-to-pattern curves relating structural patterns of resource

quality to functional suitability indices. Panels A and B compare resource suitability indices for water,

food, protection from heat and protection from disturbance in response to distance (fDr) and average quality

(fCr) respectively. Distance-dependent response-to-pattern curves cross the x axis at a ‘mobility threshold’

beyond which resources are inaccessible.


Validation using the Continuous Boyce Index (CBI) method showed that integrating landscape

structure improves model performance. Overall, we evaluated 30 landscape model (1-10) /

validation data combinations using various landscape-scale habitat variables and three control

model (X) / data combinations using site-scale habitat variables. Validation metrics are shown in

Table 3.2 and corresponding P/E ratio curves in Appendix Fig 3.4. Results differed between

validation data sets. Yet, when comparing models within each data set, integrating landscape

structure improved CBI values in 27 out of 30 scenarios (Table 3.2): When validated against the

67

Balkanu data, all landscape models 1-10 performed very well (CBI 0.72 to 0.97) while the control

model performed poorly (CBI 0.12). For the NAQS data, landscape models also had higher CBIs

than the control model (CBI -0.88), but performance remained mostly poor (CBI -0.25 to 0.72). For

the Lakefield data, certain landscape variables improved and others depressed model performance

compared to the control model. Yet, the latter’s relatively high CBI (0.66) concealed erratic peaks in

the P/E ratio curve rather than reflecting a steady positive relationship between P/E and HSI (Fig

3.6C).

Table 3.2 Performance of the eleven habitat suitability models against three validation data sets.

Habitat suitability models with

landscape-scale variables

Validation metrics per validation data set

Balkanu (n =239) Lakefield (n =371) NAQS (n =103)

ID Measurement Scale HSI ≥ 60 CBI HSI ≥ 60 CBI HSI ≥ 60 CBI

1 Distance (SIDr) 1 km 34% 0.86 18% 0.85 16% 0.32

2 Distance (SIDr) 2 km 47% 0.96 27% 0.96 27% 0.72

3 Distance (SIDr) 3 km 57% 0.97 35% 0.97 38% 0.30

4 Composition (SICr) 1 km 22% 0.74 8% 0.71 5% -0.01

5 Composition (SICr) 2 km 21% 0.91 5% 0.71 5% 0.16

6 Composition (SICr) 3 km 20% 0.87 4% -0.14 4% 0.17

7 Distance / Composition (SIDCr) 1 km 21% 0.82 8% 0.80 5% -0.25

8 Distance / Composition (SIDCr) 2 km 21% 0.95 6% 0.50 5% 0.05

9 Distance / Composition (SIDCr) 3 km 21% 0.89 5% 0.47 5% 0.31

10 a Best Knowledge (various SIr) various 55% 0.88 12% 0.96 16% -0.20

X Site-scale variables (xr) N/A 6% 0.12 6% 0.66 2% -0.88

We show the Continuous Boyce Index (CBI) and proportion of validation background expected to be highly or very

highly suitable habitat (HSI ≥ 60). A CBI = 1 would indicate a perfectly accurate, a CBI ~ 0 a random, and a CBI < 0 a

false model.

a The following resource-specific landscape variables were used for landscape model 10: SID water and SID heat at 3 km,

SIDC food at 2 km and SIDC disturbance at 1 km.

Results also varied between the ten evaluated landscape models. Only landscape models

computed from distance-dependent SIDr (1-3) consistently performed well (Table 3.2 bold values)

across all scales of measurement (1, 2 and 3 km). Model 2 (using SIDr computed at 2 km moving

window radius) was the only model to yield very good results against all validation data (CBI 0.72

to 0.96). In contrast, landscape models computed from composition-dependent SICr (4-6), combined

68

distance / composition-dependent SIDCr (7-9) or resource-specific SIr (10) had mixed results (Table

3.2). While all these models performed very well against the Balkanu data (CBI 0.74 to 0.95) and

no better than a random model against the NAQS data (CBI -0.25 to 0.31), validation against the

Lakefield data highlighted differences (CBI -0.14 to 0.96). Specifically, performance was generally

below models 1-3 (except model 10) and CBI values decreased at larger scales of measurement.

Even where CBIs were high, P/E ratio curves did not display the monotonic increase characteristic

of good models (Hirzel et al. 2006, e.g. Fig 3.6B: model 5 vs. Fig 3.6A: model 2). Moreover, when

examining each scale of measurement individually (1, 2 or 3 km), landscape models computed from

SIDr always outperformed models computed from SICr or SIDCr. Despite these inconsistent overall

results for models 4-10, some individual model / data combinations were among the best of all

those evaluated (e.g. model 8 / Balkanu data: CBI 0.95 or model 10 / Lakefield data: CBI 0.96, both

with steadily increasing P/E ratio curves, see Appendix Fig 3.4).

Finally, we also analysed the proportion of validation background expected to be highly or

very highly suitable habitat (HSI ≥ 60, Table 3.2). Similar to CBI values, integrating landscape

structure increased the proportion of HSI ≥ 60 in 26 out of 30 cases and results varied between

landscape models and validation data (Table 3.2). All control model (X) / data combinations

expected very low proportions of suitable habitat (HSI ≥ 60 2% to 6%, e.g. Fig 3.6C). Models using

distance-dependent SIDr (1-3) progressively increased this value with rising scales of measurement

(up to 57% for model 3 / Balkanu background, e.g. Fig 3.6A). Models 4-9 (using SICr or SIDCr) also

mostly increased the expected proportion of HSI ≥ 60, but to a lesser degree and somewhat

differently for each validation data set (e.g. Fig 3.6B). Model 10 resulted in moderate to large

increases. Interestingly, three of the landscape models which did not expect a greater proportion of

suitable habitat than the corresponding control model (i.e. models 6, 8 and 9 for the Lakefield

background), also had reduced CBI values.

69

Fig 3.6 Predicted-to-expected (P/E) ratio and corresponding habitat suitability maps for three

alternative models in the Lakefield area. Presence records from Lakefield (see Table 3.1 for details and

Fig 3.2 for location) are also shown. Landscape variables used for modelling are shown in Table 3.2.

70

3.5 Discussion

The results of this study demonstrate that integrating landscape structure improves the

predictive performance of a process-based habitat model of feral pigs in north-eastern Australia. We

are confident in our results, as the predictions from our expert-elicited models were empirically

validated against three truly independent distributional data sets. Improved performance was

observed in 90% of the 30 evaluated landscape model / validation data combinations. However,

results varied between the ten alternative landscape models and three validation data sets. Models

that used distance-dependent resource suitability indices performed consistently best against all

data. In turn, we discuss (a) the ecological significance of observed results, (b) the benefits of our

approach using expert-elicited Bayesian networks and response-to-pattern curves, and (c)

limitations of models, spatial analyses and evaluation methods with suggestions for future research.

3.5.1 Ecological significance

Our study was driven by landscape ecological principles. Because mobile animals use home

ranges to satisfy their habitat requirements (Burt 1943; Powell & Mitchell 2012), and these

requirements are heterogeneously distributed across landscapes (Turner & Gardner 2015), spatially-

explicit habitat models must integrate landscape structure at the home range scale. We aimed to

integrate these principles into a resource-based modelling framework using expert-elicited Bayesian

networks that is suitable for regional-scale applications, in our case feral pigs in northern Australia.

Feral pigs have distinct habitat requirements affecting their ability to breed and persist – water,

food, heat refuge and disturbance cover (Choquenot et al. 1996) – which were also identified by our

panel of experts. Moreover, they are both mobile, i.e. can access required resources at different

locations in the landscape, and sedentary, i.e. restrict their resource-seeking movements to defined

home ranges (Caley 1997; Dexter 1999; Mitchell et al. 2009). The results from this study indicate

that spatial pattern suitability analysis can usefully capture feral pigs’ resource selection within

home ranges and allow for improved predictions of habitat suitability. We suggest that our control

model underestimated habitat suitability by failing to account for the use of supplementary and

complementary resources (Dunning et al. 1992), thereby limiting high HSI to locations that

provided all habitat requirements. We expect similar findings for other mobile animals.

We also sought to explore which landscape metrics and scales of measurement are best suited

to reliably modelling habitat suitability for feral pig breeding over a large study region. Here, we

evaluated ten alternative landscape models integrating different metrics and scales. Validation

results showed that not all models were equally able to explain observed feral pig presences across

the study region.

71

Firstly, results varied considerably among evaluated landscape models. Models computed from

distance-dependent SIDr accurately described observed feral pig presences at all measurement

scales. Yet, only one model (intermediate scale, r = 2 km) performed very well against all data. In

contrast, models integrating composition-dependent SICr or combined distance / composition-

dependent SIDCr displayed mixed, scale-dependent results. Both of the latter metrics emphasized the

value of large contiguous resource patches and discounted that of small scattered patches (Froese et

al. 2015). Predicted suitable habitat was therefore confined to resource-rich core areas and may

have failed to capture two types of actual habitat, and feral pig presences, in the validation data: (a)

the wider neighbourhood of core areas was not modelled as suitable especially at a narrow scale of

measurement; and (b) small resource patches used by feral pigs to move between core areas were

“averaged out” and modelled as unsuitable especially at a broad measurement scale. Individual

well-performing model / data combinations may have reflected a balance between broadening core

areas and capturing marginal resource patches under certain local conditions. In contrast, distance-

dependent measures simply extended the “accessibility” of resources to larger neighbourhoods,

increasingly so at larger measurement scales (Froese et al. 2015), thereby predicting more suitable

habitat and explaining additional feral pig presences. We suggest that this simple landscape metric

may be less prone to model error from expert uncertainty about resource utilization and movement

capacity. It could be suited to situations where empirical data is scarce or where some

‘overprediction’ of suitable habitat is preferred to the false omission of actual habitat (e.g. invasive

species management; Jiménez-Valverde et al. 2011).

Secondly, there were considerable differences between validation data sets. Most notably,

model performance was generally high against Balkanu or Lakefield and poor against NAQS data.

This may have been due to (a) error in the NAQS data, (b) error in the spatial data linked to model

explanatory variables in this area, or (c) failure of Bayesian networks models to adequately describe

local feral pig habitat (van Klinken et al. 2015). Interestingly, NAQS-validated landscape models

still improved considerably on the control model, with model 2 achieving a very high CBI. This

suggests that poor performance is unlikely due to a false model, but problems in predicting feral pig

presences from the location of resource patches alone. The NAQS area is characterised by drier

grassland and woodland habitats than the woodland and wetland-dominated Lakefield and Balkanu

areas. Under such conditions, feral pigs may more often traverse, and be detected in, marginal

habitat between scattered resource patches. Time-averaged spatial data layers may also have failed

to capture resources which were actually available at the time of data collection. The NAQS data

with its smaller sample size (n = 103) and opportunistic approach to data collection may have been

more susceptible to such bias than the comprehensive Lakefield and Balkanu data (n = 350 and

181). Landscape models 1-3, which used distance-dependent landscape variables, may have been

72

better able to reflect inter-patch movements and weaken bias. Nevertheless, other sources of error

such as local differences in habitat requirements or behaviour cannot be dismissed. Van Klinken et

al. (2015) describe a range of methods to systematically test for error and uncertainty in spatially-

explicit Bayesian network models. We suggest that empirical studies of feral pigs’ movement

ecology and habitat utilization in resource-poor environments as well as remote or field validation

of spatial data layers may contribute most to improving our models.

3.5.2 Benefits of the modelling approach

Methodologically, the application of resource-based habitat models using Bayesian network or

other frameworks to mobile animals was improved by accounting for resource-seeking home range

movements. Spatial pattern suitability analysis essentially converted site-scale into landscape-scale

variables similar to previous statistical or expert-based approaches (Dijak et al. 2007; Martin &

Fahrig 2012; Ducci et al. 2015; McGarigal et al. 2016). However, our methodology offered four

main advantages: First, resource suitability indices were computed from response-to-pattern curves.

These were explicitly referenced to feral pig behaviour (i.e. the ability of breeding herds to access

and survive on available resources), which enhances transferability across large study regions

(Larson et al. 2009). Second, eliciting this information from experts was cost-effective and provided

transparent hypotheses (Appendix Table 3.1.4) that can be easily updated if new knowledge

becomes available. The structured elicitation method for relating measured spatial patterns to

corresponding functional response (suitability) values was easily implemented and understood by

expert practitioners. It may prove a useful addition to the landscape ecology toolbox, especially in

situations where multi-scale empirical studies are not feasible due to time, resourcing or data

constraints. Scales of measurement were also elicited a priori from experts based on their field

knowledge. While validation did not indicate selection of an ‘optimal’ scale for any of the

landscape variables, expert assumptions were broadly consistent with the range of empirical home

range estimates from Australia (3-28 km2 in this study compared to 1-20 km2 in Choquenot et al.

1996). Third, landscape variables were derived from numerical site-scale indices rather than

categorical spatial proxies (e.g. vegetation type, land use). These indices directly represented the

quality of resources as perceived by feral pigs (Moorcroft 2012). They were also modelled in

Bayesian networks from a suite of spatially-explicit explanatory variables. Hence, different site-

scale conditions could yield similar quality values (e.g. food with high biomass and energy content

vs. high quality and protein content), which usefully reflected Dunning et al.’s (1992) concept of

resource supplementation. Finally, we have systematically tested how models based on different

landscape metrics and scales of measurement affect model performance. While this is common in

statistical multi-scale studies, resource-based habitat models that may be preferred for large study

73

regions or when data is limited, have rarely rigorously evaluated assumptions behind explanatory

variables (Beck & Suring 2009).

3.5.3 Limitations

3.5.3.1 Modelling approach

Firstly, we focused on three simple landscape metrics that were ecologically meaningful and

easy to relate to experts’ field knowledge. Yet, resource suitability for feral pig breeding may

depend on other metrics such as fragmentation (Jackson & Fahrig 2016), edge proximity (Dijak &

Rittenhouse 2009) or heterogeneity (McClure et al. 2015). Secondly, we asked each expert to

specify one scale of measurement applicable to all habitat variables and landscape metrics. Yet,

theory (Addicott et al. 1987; Wiens 1989) and evidence (Jackson & Fahrig 2014; Miguet et al.

2016) suggest that species may in fact respond to different habitat variables and aspects of

landscape structure at different scales. Both could be explored through further expert elicitation or

multi-scale empirical studies (Jackson & Fahrig 2015), or by evaluating other combinations of

resource-specific SIr than in our ‘best knowledge’ model. Thirdly, although widely used, circular

moving windows may not reflect the shape of feral pigs’ actual home ranges (Dexter 1999).

Knowledge from movement ecological studies (Morelle et al. 2014) in combination with technical

advances that allow for dynamic moving windows (Berry 2013) may further improve results.

Finally, we generalized local knowledge to the regional extent by averaging. While it has been

suggested that average expert estimates are more accurate than either individual or consensus

estimates (McBride et al. 2012b), model uncertainty from divergent knowledge could be further

analysed (Johnson et al. 2012b; Murray et al. 2012). In particular, future research could aim to

adapt spatial pattern suitability analysis for probabilistic habitat variables rather than our

deterministic resource quality indices xr.

3.5.3.2 Model evaluation and validation

We used the Continuous Boyce Index to evaluate model performance (Boyce et al. 2002;

Hirzel et al. 2006). While this method is well suited to presence-only validation data, it provides no

information on model specificity, i.e. its ability to correctly predict absences and minimize false

positives (Hirzel et al. 2006; Jiménez-Valverde et al. 2011). Secondly, CBI values alone may be

misleading in some instances. For example, P/E ratio curves that are steadily increasing along a flat

slope may result in a high CBI but not necessarily reflect a good model. Systematically collected

presence / absence data, especially from resource-poor environments, could facilitate confusion

matrix-based validation of sensitivity vs. specificity and is suggested as a future research priority.

74

3.5.4 Conclusion

Our results highlight that habitat models of mobile animals must integrate landscape structure.

We developed a novel combined methodology, spatial pattern suitability analysis, for capturing

feral pigs’ resource-seeking home range movements based on expert-elicited response-to-pattern

curves and spatial moving window analysis. Based on landscape ecological principles, this

methodology improved the application of resource-based Bayesian networks models to mobile

animals. Through comprehensive scenario analysis and validation, we were able to determine which

of various landscape-scale variables allowed for improved predictions of habitat suitability. Our

approach could be applied to other mobile species, modelling frameworks and landscape metrics,

for example to compute ecologically meaningful landscape predictors for statistical analyses or to

integrate expert knowledge on species mobility and space use into resource-based habitat models.

Built on a priori knowledge, our methodology may be particularly useful to support regional-scale

management decisions where empirical data is limited.

3.6 Appendices

Appendix 3.1 Expert-elicited response-to-pattern curves.

Appendix 3.2 Methodology for spatial pattern suitability analysis (annotated R code).

Appendix 3.3 Methodology for model validation (annotated R code).

Appendix 3.4 Validation plots for eleven habitat suitability models.

75

Chapter 4 Modelling seasonal habitat suitability for wide-ranging species:

invasive feral pigs in northern Australia

Summary

In this chapter I apply the methodology from Chapter 3 to model and map seasonal habitat

suitability for feral pig breeding and persistence in northern Australia. I provide a detailed

description of the resource-based, spatially-explicit modelling approach using expert-elicited

Bayesian networks and spatial pattern suitability analysis to account for resource-seeking home

range movements. I compare modelled habitat suitability for a wet season and a dry season scenario

and validate model accuracy against four independent distributional data sets per scenario. This

chapter addresses research question 2, providing an indicative measure of the seasonal

establishment risk of directly transmitted infectious diseases within feral pigs in northern Australia.

It has been published in PLoS ONE.

4.1 Abstract

Invasive wildlife often causes serious damage to the economy and agriculture as well as

environmental, human and animal health. Habitat models can fill knowledge gaps about species

distributions and assist planning to mitigate impacts. Yet, model accuracy and utility may be

compromised by small study areas and limited integration of species ecology or temporal

variability. Here we modelled seasonal habitat suitability for feral pigs, a widespread and harmful

invader, in northern Australia. We developed a resource-based, spatially-explicit and regional-scale

approach using Bayesian networks and spatial pattern suitability analysis. We integrated important

ecological factors such as variability in environmental conditions, breeding requirements and home

range movements. The habitat model was parameterized during a structured, iterative expert

elicitation process and applied to a wet season and a dry season scenario. Model performance and

uncertainty were evaluated against independent distributional data sets. Validation results showed

that an expert-averaged model accurately predicted empirical feral pig presences in northern

Australia for both seasonal scenarios. Model uncertainty was largely associated with different

expert assumptions about feral pigs’ resource-seeking home range movements. Habitat suitability

varied considerably between seasons, retracting to resource-abundant rainforest, wetland and

agricultural refuge areas during the dry season and expanding widely into surrounding grassland

floodplains, savanna woodlands and coastal shrubs during the wet season. Overall, our model

suggested that suitable feral pig habitat is less widely available in northern Australia than previously

thought. Mapped results may be used to quantify impacts, assess risks, justify management

76

investments and target control activities. Our methods are applicable to other wide-ranging species,

especially in data-poor situations.

4.2 Introduction

Where mammalian wildlife becomes invasive, it is often detrimental to the economy and

agriculture as well as environmental, human and animal health (Beale et al. 2008, Hone 2007). To

effectively mitigate impacts, spatially-explicit knowledge on invaders’ distribution and habitat use

is needed (Ostfeld et al. 2005; Venette et al. 2010). This can be particularly challenging for wide-

ranging species, as continuous empirical information is rarely available over broad geographic

regions (Stephens et al. 2015). With rapid developments in spatial environmental data availability

and new analytical methods, habitat models that infer species distributions from environmental

predictor variables have proliferated to fill knowledge gaps (Guisan & Thuiller 2005; Elith &

Leathwick 2009). Research is methodologically and terminologically diverse – depending on the

research perspective, “habitat models” are also known as “species distribution models”, “ecological

niche models”, “habitat suitability models”, “resource selection functions” and variations thereof

(Guisan & Thuiller 2005; Morrison et al. 2006; Beck & Suring 2009; Elith & Leathwick 2009;

Dormann et al. 2012; Venette et al. 2010). However, important ecological considerations such as

temporal variability or behavioural factors are often missed, especially in statistical, correlative

models. This can affect model accuracy and utility for decision-making (Franklin 2010). Here, we

developed a resource-based, spatially-explicit approach to modelling seasonal habitat suitability for

a widespread and harmful mobile invader, the feral pig (Sus scrofa), in northern Australia.

Feral pigs, originally native to Eurasia, are one of the most widespread terrestrial mammals

(Barrios-Garcia & Ballari 2012). Both wild and domesticated forms were introduced by early

settlers to all continents and many oceanic islands (Barrios-Garcia & Ballari 2012). In its introduced

range, S. scrofa is also known as feral pig, feral swine, wild hog or razorback and has often been

associated with severe negative impacts (Choquenot et al. 1996; Barrios-Garcia & Ballari 2012;

Bengsen et al. 2014). In Australia, invasive feral pigs are a major threat to unique ecosystems and

agricultural industries (Bengsen et al. 2014; Department of the Environment 2015). They are most

widespread in the tropical north, yet spatial knowledge is either empirical, detailed, and local scale

(Hone 1990a; Caley 1993 and 1997; Mitchell 2002; Mitchell et al. 2009; Wurster et al. 2012;

Elledge et al. 2013), or expert-based, coarse, broad scale, and poorly validated (Wilson et al. 1992;

West 2008; Department of Agriculture and Fisheries 2015). Improved regional-scale knowledge of

feral pig distribution could be used to delineate management units and limit re-invasion of

conservation sites following local eradication (Bengsen et al. 2014; Dexter & McLeod 2015). It

could also help assess the magnitude of environmental and economic impacts or the risk of

77

establishment of infectious animal diseases, especially when abundance estimates are derived

(Kramer-Schadt et al. 2007; Cowled & Garner 2008; Krull et al. 2016; Weber et al. in press).

Statistical habitat models for feral pigs have been developed for northern Australia (Cowled &

Giannini 2007; Cowled et al. 2009) as well as parts of Europe and North America with similar

knowledge gaps (Bosch et al. 2014; McClure et al. 2015; Morelle & Lejeune 2015; Alexander et al.

2016). However, some general limitations of statistical models are also apparent in these studies.

First, correlative models calibrated from species presence or presence / absence records can only

reliably predict species distribution within, and not outside, the environmental gradients used for

model calibration (Elith et al. 2010). Second, except for Morelle & Lejeune’s (2015) study in

Belgium, all models were calibrated from aggregate species records and did not consider temporal

variability or ecological factors. In northern Australia such models may yield misleading results

when reflecting previous research: Caley (1993) and Hone (1990a) showed that feral pig

distribution and abundance varies considerably between the wet and dry season; Caley (1997) and

Mitchell et al. (2009) found that habitat use and home range movements differ distinctly between

breeding herds (consisting of related sows and their young) and solitary boars; and Choquenot &

Ruscoe’s (2003) work suggested that feral pig persistence depends on complementary access to key

resources within the boundaries of such home ranges.

Here we adapted a resource-based modelling framework using Bayesian networks that allowed

us to address these issues. This general approach has previously been applied to habitat models

(Marcot et al. 2001; Smith et al. 2007; Tantipisanuh et al. 2014; van Klinken et al. 2015) and offers

several advantages over correlative methods: a robust statistical framework for modelling

interactions between habitat variables based on species ecology rather than distributional data;

flexible data requirements with the ability to integrate unpublished expert knowledge; and explicit

treatment of the uncertainty in parameter estimates (Uusitalo 2007; Wilhere 2012; Landuyt et al.

2013). Our objectives were: (1) to model seasonal habitat suitability for feral pig breeding and

persistence in northern Australia at the regional scale whilst integrating behavioural factors as well

as temporal variability; and (2) to rigorously evaluate accuracy and uncertainty in our expert-

elicited models by validating spatial predictions of habitat suitability against truly independent

distributional data sets.

4.3 Materials and methods

4.3.1 Study region

Our study region covered 1.76 million km2 north of the Tropic of Capricorn spanning three

Australian states (Fig 4.1). The climate is tropical with seasonal rainfall, alternating between a wet

78

and a dry season. Rainfall and primary productivity broadly decline on a north-south, and to a lesser

extent, on an east-west, gradient (Guerschman et al. 2009; Jones et al. 2009). Monsoonal savanna

woodlands and semi-arid grasslands are interspersed with riverine channels, seasonally inundated

floodplains, coastal wetlands and rainforest fragments (Fox et al. 2001). Intensive human uses are

concentrated in fertile coastal lowlands. The semi-arid inland is sparsely populated.

At a coarse scale, all of the study region appears climatically suitable for feral pigs and has

mostly been invaded (West 2008). Arid regions with insufficient rainfall were not included in our

study. Feral pigs are reported to be widespread in the east and localised in the north and west (West

2008). Highest local densities have been recorded in resource-abundant wetlands and floodplains,

yet these populations fluctuate considerably with climatic conditions (Hone 1990a; Caley 1993;

Choquenot et al. 1996). A wide range of management activities are conducted throughout northern

Australia to mitigate feral pig impacts, including lethal and non-lethal methods. Yet, effective

management is hampered by the region’s remoteness and there is little evidence of sustained

population reduction (Bengsen et al. 2014).

Fig 4.1 Study region in northern Australia. The study region is shown in grey. Locations of the

independent distributional data sets used for model validation are shown in colour. Refer to Table 4.3 for

details about these data sets.

4.3.2 Habitat suitability model

Our modelling approach consisted of three main steps (Fig 4.2). First, we modelled ‘resource

quality indices’ for a suite of habitat variables, referenced specifically to the resource requirements

79

of feral pig breeding herds, in separate Bayesian networks. Second, we used ‘spatial pattern

suitability analysis’ to capture feral pigs’ ability to access complementary resources at different

locations within their home range. Finally, we modelled a ‘habitat suitability index’ by combining

all ‘resource suitability indices’ in another Bayesian network. The model was calibrated using a

structured, iterative elicitation process (Marcot 2006) with a panel of experts. Experts were

practitioners with field knowledge of feral pigs from various localities and professional

backgrounds (Perera et al. 2012). We combined techniques for eliciting system understanding

through group consensus and for eliciting quantitative estimates from individuals with opportunities

for Delphi-style revision (Kuhnert et al. 2010; Martin et al. 2012; Murray et al. 2012; van Klinken

et al. 2015). Expert elicitation was approved by the CSIRO Human Ethics Committee (Project

075/13) and written consent obtained from all participants.

Fig 4.2 (next page) Feral pig habitat suitability model. Resource quality indices for each habitat variable

were modelled in Bayesian networks. Spatial pattern suitability analysis was used to compute resource

suitability indices as a weighted function of distance to resource patches (fDr). Habitat suitability was

modelled in another Bayesian network. An average habitat suitability index was computed and mapped (Fig

4.6) from six individual expert models. Bar graphs show expert-elicited conditional probabilities and values

below graphs show modelled index values ± standard deviation. Probabilities and indices change once

evidence about the states of each explanatory variable at a given study area pixel is inserted (i.e. prior

probabilities are no longer uniformly distributed).

80

Heat stressLowModerateHigh

33.333.333.3

50 ± 27

Heat protection quality indexVery goodGoodModeratePoorVery poor

55.814.414.55.739.49

70.3 ± 27

ExpertExpert1Expert2Expert3Expert4Expert5

20.020.020.020.020.0

Shady vegetation coverGoodModeratePoor

33.333.333.3

50 ± 27

Water quality indexVery goodGoodModeratePoorVery poor

33.48.216.294.9647.2

45.1 ± 37

Freshwater presenceYesNo

50.050.0

50 ± 25

Terrain ruggednessLowModerateHigh

33.333.333.3

50 ± 27

ExpertExpert1Expert2Expert3Expert4

25.025.025.025.0

Food qualityHighModerateLow

33.333.333.3

50 ± 27


20.020.020.020.020.0

Food quality indexVery goodGoodModeratePoorVery poor

19.523.222.513.921.0

51.2 ± 29

Food quantityHighModerateLow

33.333.333.3

50 ± 27

Water suitability indexVery goodGoodModeratePoorVery poor

20.020.020.020.020.0

50 ± 29

Food suitability indexVery goodGoodModeratePoorVery poor

20.020.020.020.020.0

50 ± 29

Habitat suitability indexVery highHighModerateLowVery low

2.0112.026.121.338.6

33.5 ± 23

ExpertExpert1Expert2Expert3Expert4Expert5Expert6

16.716.716.716.716.716.7

Disturbance protection suitability indexVery goodGoodModeratePoorVery poor

20.020.020.020.020.0

50 ± 29

Heat protection suitability indexVery goodGoodModeratePoorVery poor

20.020.020.020.020.0

50 ± 29

Disturbance protection quality indexVery goodGoodModeratePoorVery poor

18.420.020.716.224.7

48.2 ± 29

Dense vegetation coverGoodModeratePoor

33.333.333.3

50 ± 27


20.020.020.020.020.0

Disturbance stressLowModerateHigh

06.6093.4

18.7 ± 13


20.020.020.020.020.0

Frequency of controlLowHigh

0 100

25

Intensity of controlLowHigh

0 100

25

Predator presenceLowHigh

50.050.050 ± 25

Spatial pattern suitability analysis (using each expert’s response function fDr)

Spatial pattern suitability analysis (using each expert’s response function fDr)

81

4.3.2.1 Bayesian network models

We adapted the Bayesian network modelling framework proposed by Marcot et al. (2001),

refined by Smith et al. (2007) and explained in detail by van Klinken et al. (2015). Here, habitat

suitability was conditional on a set of habitat variables representing resource requirements. Each

habitat variable was itself influenced by several measurable key explanatory variables, and each

explanatory variable was linked to one or more remotely sensed or mapped spatial data proxies. Our

model was implemented in the Norsys Netica v.5.12 software.

Expert elicitation

During an initial expert workshop, a preliminary model was developed (Marcot 2006). A panel

of experts (n = 18) constructed a conceptual model, defined each model variable (habitat suitability,

habitat variables and explanatory variables including spatial data proxies) and assigned it mutually

exclusive states. We quantified causal relationships in the network by eliciting conditional

probability tables (CPTs) behind each response variable (child node). We used the CPT calculator

software (Cain 2001), which reduces the number of elicited response probabilities to key scenarios,

i.e. combination of states in explanatory variables (parent nodes), and interpolates all other

combinations. Each step was performed in break-out groups or individually, followed by panel

discussion and consensus formation (except for the CPT calculator) (Murray et al. 2012; van

Klinken et al. 2015).

Following preliminary application, sensitivity analysis and validation of the preliminary model,

we conducted semi-structured interviews with a self-selected subsample from our panel of experts

(n = 6). Model structure, spatial data proxies and evaluation results were reviewed against each

expert’s knowledge and simplified CPTs were parameterized. We asked experts to revise prior

CPTs from the preliminary model rather than parameterizing new ones (Kuhnert et al. 2010; Martin

et al. 2012). As interviews were less time-constrained than the workshop, experts could utilize

either or both of two elicitation methods that were more flexible and robust to error than the CPT

calculator (Cain 2001). Method A was implemented in the AgenaRisk v.6.1 software and made the

simplifying assumption that any response follows a truncated normal distribution (TNormal)

centred on the weighted mean of its explanatory variables (Fenton et al. 2007). In order to use this

method, we converted all model variables into “ranked nodes”, whose states were assigned with

equal intervals on a numerical scale from 0 to 100 (Fenton et al. 2007). Experts only defined: (a) the

weight of each explanatory variable, (b) overall uncertainty in making this judgement (determining

the variance of TNormal), and (c) whether the weighted mean function should be replaced by either

a weighted minimum (to describe limiting factor relationships), or maximum (to describe

substitution relationships) function (Fenton et al. 2007). Method B restricted elicitation to key

82

scenarios as in Cain (2001). However, instead of directly assigning probabilities to each state of the

response variable, we used interval judgements (Speirs-Bridge et al. 2010; McBride et al. 2012a),

asking experts for their best estimate and the outer bounds of a 95% confidence interval. To

maintain consistency with method A, we allowed only TNormal distributions centred on the best

estimate. Post-elicitation, we discretised interval judgements into probabilities for each response

state using a binning algorithm.

Habitat suitability and resource quality indices

The final Bayesian network model (Marcot 2006) is shown in Fig 4.2. Expert-elicited CPTs are

provided in Appendix Table 4.1. Definitions for all model variables and their states are given in

Table 4.1. Spatial data proxies that determined the state of each explanatory variable at each pixel in

the study region are described in Table 4.2. Experts identified four key resource requirements for

sustained feral pig breeding: food, water, protection from heat and protection from disturbance (Fig

4.2). Each of these habitat variables had five states with corresponding equal numerical intervals (0-

20 for the poorest state, …, 80-100 for the best state). For each habitat variable, we computed

expert-averaged ‘resource quality indices’ (xr) as model expected values from an equal-weighted

average CPT (Martin et al. 2012; McBride et al. 2012b; Fig 4.2) by summing the mid-point value

of each state interval weighted by its probability. Accordingly, xr could range between 10 (mid-

point of the poorest state) and 90 (mid-point of the best state) and varied spatially according to the

combination of states of explanatory variables at a given pixel. Following spatial pattern suitability

analysis of pixel-scale xWater, xFood, xHeat and xDisturbance (see below), we computed a ‘habitat

suitability index’ (HSI) from the derived landscape-scale habitat variables (Martin & Fahrig 2012;

Jackson & Fahrig 2015). The method was analogous to the resource quality indices. However, we

used each individual expert’s CPT to compute model expected HSI. This allowed us to evaluate

average results as well as model uncertainty from diverging expert knowledge.

4.3.2.2 Spatial pattern suitability analysis

In order to capture feral pigs’ ability to access their four resource requirements at different

locations within heterogeneous home ranges, we converted pixel-scale resource quality indices (xr)

into landscape-scale ‘resource suitability indices’ (SIr). Apart from various empirical estimates of

female home range sizes (1-20 km2 in Choquenot et al. 1996; Hone 2012) and a previous finding

that pasture and riverine woodlands must co-occur within 5 km (Choquenot & Ruscoe 2003), we

had limited a priori knowledge about feral pigs’ resource-seeking home range movements in

northern Australia.

83

We therefore elicited distance-dependent response-to-pattern curves (fDr) for each habitat

variable from individual experts (n = 6) during interviews (Fig 4.3 and Appendix 4.3). This

involved first specifying a feral pig ‘mobility threshold’ (or home range boundary) beyond which

resources are inaccessible to breeding herds. Experts defined these at 1 km (n = 1), 2 km (n = 2) or

3 km (n = 3), corresponding to circular home ranges of approximately 3, 12 and 28 km2. Second,

experts described the functional value of a given resource for feral pig breeding in response to

distance (Fig 4.3). Third, we applied spatial moving window analysis to compute SIr at a focal pixel

as the highest distance-weighted xr of all resources contained within an analysis window. This

window was shifted and centred on each pixel within the study region. We used circular moving

windows (Jackson & Fahrig 2015) with radii defined by elicited mobility thresholds and distance

weights defined by response curves fDr. We termed this combined methodology ‘spatial pattern

suitability analysis’ (Appendix 4.3).9

4.3.2.3 Seasonal scenarios

We applied each individual expert’s final model (Marcot 2006) to two seasonal scenarios by

linking model explanatory variables to seasonally-specific spatial data proxies (Table 4.2). We were

most interested in the late wet season (March to April), when resources required by feral pigs are

abundant and widely distributed, and the late dry season (October to November), when resources

are scarce and scattered across the region. Suitable remotely sensed or mapped spatial proxies were:

(a) discussed with experts and sourced from various agencies (Table 4.2); (b) rasterized and

resampled to a common extent and a fine resolution (1 ha) for capturing spatial heterogeneity

relevant to feral pigs (van Klinken et al. 2015); and (c) averaged for the two months corresponding

to each seasonal scenario over five years (2010-2014) to reflect average conditions. Finally, spatial

data attributes were reclassified to match the states of explanatory variables (Table 4.2 and

Appendix 4.2). Some model variables were linked to static spatial proxies without seasonal

variability (e.g. terrain ruggedness). For three variables determining Disturbance stress (Fig 4.2 and

Table 4.1), no spatial proxies were available for the study region. We applied a global (spatially

uniform) scenario [42], with “high” Intensity of control and Frequency of control and no selected

state for Predator presence (Fig 4.2). While this assumption likely overestimated disturbance in our

models, it approximated conditions under which most of the local validation data were collected.

9 A detailed description of the methodology for spatial pattern suitability analysis including theoretical framework, expert elicitation and moving window analysis is given in Chapter 3.

84

Table 4.1 Bayesian network model variables and their states, with definitions.

Node & definition State (interval) State definition

⋅ Habitat suitability index (HSI) Whether feral pig mobs are able to breed and persist

Very high (80-100) Feral pig mobs always able to breed, strong population growth

High (60-80) ... usually able to breed, population growth

Moderate (40-60) ... occasionally able to breed, population maintenance

Low (20-40) ... usually unable to breed, population decline

Very low (0-20) … always unable to breed, strong population decline

⋅ Water suitability index (SIWater) Whether there is sufficient potable water available to meet drinking requirements of feral pig mobs

Very good (80-100) Drinking water fully sufficient to meet requirements

Good (60-80) ... usually sufficient to meet requirements

Moderate (40-60) ... occasionally sufficient to meet requirements

Poor (20-40) ... usually limiting breeding and persistence

Very poor (0-20) … strongly limiting breeding and persistence

⋅ Water quality index (xWater) Presence of accessible sources of potable water

Very good (80-100) Potable water present and fully accessible

Good (60-80) ... present and usually accessible

Moderate (40-60) ... present and accessible with some difficulty

Poor (20-40) ... present and usually inaccessible

Very poor (0-20) … absent or fully inaccessible

⋅ Freshwater presence Presence of potable water

Yes (50-100) Potable water present

No (0-50) Water absent or non-potable (salty or brackish)

⋅ Terrain ruggedness Accessibility of water source due to terrain ruggedness

Low (67-100) Level and / or low-lying terrain

Moderate (33-67) Moderately rugged terrain

High (0-33) Highly rugged terrain

⋅ Food suitability index (SIFood) Whether there are sufficient food resources available to meet nutritional and energy requirements of breeding sows

Very good (80-100) Food resources fully sufficient to meet requirements





⋅ Food quality index (xFood) Presence and nutritional value of food resources

Very good (80-100) Food resources present and of very high nutritional value

Good (60-80) ... present and of high nutritional value

Moderate (40-60) ... present and of some nutritional value

Poor (20-40) ... present and of low nutritional value

Very poor (0-20) … absent or of very low nutritional value

⋅ Food quality Quality (protein content) and accessibility of food resources

High (67-100) High quality food resources and readily accessible

Moderate (33-67) Moderate quality food resources and / or restricted accessibility

Low (0-33) Low quality food resources or inaccessible

⋅ Food quantity Quantity (energy content) of food resources

High (67-100) High quantity (energy-rich) of food resources

Moderate (33-67) Moderate quantity (some energy) of food resources

Low (0-33) Low quantity (little or no energy) of food resources

85

⋅ Heat protection suitability index (SIHeat) Whether there is sufficient heat refuge available to meet protection requirements of feral pig mobs

Very good (80-100) Heat refuge fully sufficient to meet requirements





⋅ Heat protection quality index (xHeat) Presence and quality of refuge from heat stress conditions

Very good (80-100) Insignificant heat stress or refuge offers full protection

Good (60-80) Refuge offers good protection from heat stress conditions

Moderate (40-60) Refuge offers some protection ...

Poor (20-40) Refuge offers little protection ...

Very poor (0-20) Refuge offers no or very little protection ...

⋅ Heat stress Heat stress conditions from daytime temperatures

Low (67-100) Insignificant heat stress

Moderate (33-67) Some heat stress from prolonged moderate daytime temperatures

High (0-33) Significant heat stress from prolonged high daytime temperatures

⋅ Shady vegetation cover Cool microclimate provided by shady vegetation canopy

Good (67-100) Deep shading provided by a dense vegetation canopy

Moderate (33-67) Dappled shading provided by an open vegetation canopy

Poor (0-33) No or little shading provided by a very sparse vegetation canopy

⋅ Disturbance protection suitability index (SIDisturbance) Whether there is sufficient disturbance refuge available to meet protection requirements of feral pig mobs

Very good (80-100) Disturbance refuge fully sufficient to meet requirements





⋅ Disturbance protection quality index (xDisturbance) Presence and quality of refuge from disturbance stress

Very good (80-100) Insignificant disturbance stress or refuge offers full protection

Good (60-80) Refuge offers good protection from disturbance stress

Moderate (40-60) Refuge offers some protection ...

Poor (20-40) Refuge offers little protection ...

Very poor (0-20) Refuge offers no or very little protection ...

⋅ Dense vegetation cover Cover provided by understory vegetation

Good (67-100) Good cover provided by dense understory vegetation

Moderate (33-67) Moderate cover provided by open or medium-height vegetation

Poor (0-33) No or little cover provided by sparse or tall vegetation

⋅ Disturbance stress Disturbance stress from human or predator interference

Low (67-100) Insignificant disturbance stress

Moderate (33-67) Some disturbance stress

High (0-33) Significant disturbance stress

⋅ Intensity of control Low (50-100) No or low to moderate impact control activities

High (0-50) High impact control activities

⋅ Frequency of control Low (50-100) No or infrequent to occasional control activities

High (0-50) Sustained control effort (throughout and across years)

⋅ Predator presence Low (50-100) No or insignificant numbers of predators are present

High (0-50) Significant numbers of predators are present The first state listed is considered most favourable to feral pig breeding. Numerical intervals corresponding to each state

are also listed in brackets.

86 N

ode

[sea

sona

l?]

Spat

ial l

ayer

Rec

lass

ifica

tion

into

stat

e-sp

ecifi

c ca

tego

ries [

nam

e of

stat

e]

Nam

e D

ate,

file

type

So

urce

, ref

eren

ce

Rat

iona

le, m

etho

d

Bes

t M

iddl

e W

orst

Fres

hwat

er

pres

ence

[s

easo

nal]

Que

ensl

and

wet

land

dat

a V

.3 -

wet

land

are

as /

poin

ts (s

prin

gs)

2012

, ESR

I sh

apef

ile

(1:1

00,0

00)

Dep

t. of

Env

ironm

ent

& H

erita

ge P

rote

ctio

n Q

LD, E

PA 2

005

Cla

ssifi

ed a

reas

as ‘

pere

nnia

l/eph

emer

al/n

one’

from

W

etcl

ass &

WtrR

egim

e at

tribu

tes,

excl

uded

non

-pot

able

w

ater

from

Sal

inm

od a

ttrib

ute,

cla

ssifi

ed p

oint

s as

‘per

enni

al’ e

xcep

t if d

orm

ant

⋅ dry

: per

enni

al

⋅ wet

: per

enni

al

+ ep

hem

eral

[Y

es]

n/a

⋅ dry

: non

e +

ephe

mer

al

⋅ wet

: non

e [N

o]

Geo

fabr

ic S

urfa

ce

Car

togr

aphy

V.2

.1 –

H

ydro

Poin

t / D

am

2012

, FG

DB

fe

atur

e cl

ass

(1:2

50,0

00)

Bur

eau

of

Met

eoro

logy

, BO

M

2012

Cla

ssifi

ed a

s ‘pe

renn

ial’

exce

pt a

band

oned

/dry

or s

alin

e

⋅ dry

/wet

: pe

renn

ial

[Yes

] n/

a ⋅ d

ry/w

et:

none

[N

o]

Geo

fabr

ic S

urfa

ce

Car

togr

aphy

V.2

.1 –

H

ydro

Are

a / W

ater

Bod

y

2012

, FG

DB

fe

atur

e cl

ass

(1:2

50,0

00)

Bur

eau

of

Met

eoro

logy

, BO

M

2012

Cla

ssifi

ed H

ydro

Are

as a

s ‘ep

hem

eral

/non

e’ a

nd

Wat

erB

ody

as ‘p

eren

nial

/eph

emer

al/n

one’

from

A

HG

FTyp

e, P

eren

nial

& S

rcFT

ype

attri

bute

s

⋅ dry

: per

enni

al

⋅ wet

: per

enni

al

+ ep

hem

eral

[Y

es]

n/a

⋅ dry

: non

e +

ephe

mer

al

⋅ wet

: non

e [N

o]

Pres

ent M

ajor

Veg

etat

ion

Subg

roup

s (M

VS)

, NV

IS

V.4

.1

2012

, FG

DB

gr

id (1

00m

)

Fede

ral D

ept.

of th

e En

viro

nmen

t, D

EWR

20

07

W

ater

bod

ies i

nade

quat

ely

map

ped

in tr

opic

al, s

ub-

tropi

cal o

r war

m te

mpe

rate

rain

fore

st, c

lass

ified

as

wat

er p

rese

nt (n

on-li

miti

ng)

M

VS

2 &

6

[Yes

] n/

a M

VS

othe

r [N

o]

Terr

ain

rugg

edne

ss

3 se

c SR

TM d

eriv

ed

Dig

ital E

leva

tion

Mod

el,

V.1

.0

2010

, FG

DB

gr

id (3

arc

sec)

Geo

scie

nce

Aus

tralia

, G

alla

nt e

t al.

2011

Cal

cula

ted

terr

ain

rugg

edne

ss in

dex

(TR

I) u

sing

gd

alde

m a

lgor

ithm

(bas

ed o

n W

ilson

et a

l. 20

07),

clas

sifie

d vi

sual

ly a

s thr

esho

lds s

cale

dep

ende

nt

TR

I < 1

2 [L

ow]

TRI 1

2-24

[M

oder

ate]

TR

I >24

[H

igh]

Food

qu

ality

[s

easo

nal]

Pres

ent M

ajor

Veg

etat

ion

Subg

roup

s (M

VS)

, NV

IS

V.4

.1

2012

, FG

DB

gr

id (1

00m

)

Fede

ral D

ept.

of th

e En

viro

nmen

t, D

EWR

20

07

Es

timat

ed fo

od q

ualit

y in

nat

ural

syst

ems u

nder

goo

d /

poor

acc

essi

bilit

y of

bel

ow-g

roun

d pr

otei

n, e

licite

d fr

om e

xper

ts &

Mur

ray

et a

l. 20

15

C

lass

ifica

tion

see

A

ppen

dix

Tabl

e 4.

2.1

Cat

chm

ent s

cale

Lan

d U

se

of A

ustra

lia (C

LUM

) 20

15, F

GD

B

grid

(50m

)

Fede

ral D

ept.

of

Agr

icul

ture

, AB

AR

ES

2011

Estim

ated

food

qua

lity

in m

odifi

ed sy

stem

s (la

nd u

se

clas

s) u

nder

goo

d / p

oor a

cces

sibi

lity

of b

elow

-gro

und

prot

ein,

elic

ited

from

exp

erts

& M

urra

y et

al.

2015

Cla

ssifi

catio

n se

e

App

endi

x Ta

ble

4.2.

2

Mon

thly

rela

tive

soil

moi

stur

e up

per l

ayer

(W

Rel

1), h

isto

rical

run

26j /

ope

ratio

nal v

26

2015

, ESR

I flo

at (0

.05°

) C

SIR

O, R

aupa

ch e

t al.

2009

Cal

cula

ted

aver

age

WR

el1

for e

ach

scen

ario

, cla

ssifi

ed

visu

ally

into

goo

d / p

oor a

cces

sibi

lity

of b

elow

-gro

und

prot

ein

sour

ces

W

Rel

1 >

20

[Goo

d]

n/a

WR

el1

≤ 20

[P

oor]

Tab

le 4

.2 S

patia

l dat

a pr

oxie

s lin

ked

to m

odel

exp

lana

tory

var

iabl

es a

nd m

etho

ds fo

r re

clas

sify

ing

data

att

ribu

tes i

nto

stat

e-sp

ecifi

c ca

tego

ries

.

88

Fig 4.3 Expert-elicited resource suitability in response to distance. Resource suitability indices (SIr) were

computed from elicited response-to-pattern curves (fDr) for each of the four habitat variables in the model:

water (A), food (B), heat protection (C) and disturbance protection (D). Curves cross the x-axis at different

points because experts defined different home range boundaries.


4.3.3.1 Sensitivity analysis

Behaviour of each individual expert model as well as an expert-averaged consensus model was

evaluated using the “Sensitivity to findings” algorithm in the Norsys Netica 5.12 software. We

focused on the variance reduction metric recommended for numerical variables (Marcot 2012). It

assessed the relative influence of our four habitat variables on predicted habitat suitability by

89

calculating how much the variance of HSI was reduced by entering a particular finding (i.e. SIr

value) for one of the habitat variables. Greater variance reduction means that HSI was more

sensitive to a change in the state of the habitat variable (Marcot 2012; Smith et al. 2012).

4.3.3.2 Predictive performance

Predictive model performance was validated against four data sets of feral pig presence per

seasonal scenario (Table 4.3). Most data were independently collected by external agencies in

conjunction with aerial management activities. As aerial survey data were only available for the

eastern state of Queensland (Fig 4.1), we also obtained presence records (search term “Sus scrofa”)

from the national Atlas of Living Australia. This database contained only one dated record in the

Western Australian portion of our study area. Hence, we confined the downloaded ‘ALA’ data set

to records from the Northern Territory, which appeared adequately sampled (748 unfiltered

records). Where possible, we matched data to model assumptions, using only presence records that

corresponded to breeding herds (identified as female or with a group count greater than two) and

were dated in the late wet and late dry season respectively. To reduce unwanted noise from spatial

error in validation data or spatial proxies used for modelling (Hunsacker et al. 2001; Boyce et al.

2002), we upscaled both predicted HSI and feral pig presence records to a 1 km resolution. We

subsequently thinned data to ensure independence among data points, allowing only one data point

collected on the same day within a given 1 km pixel.

For each data set, we also defined a validation background, which served to contrast presences

with areas from which feral pips could be considered ‘absent’ (Fig 4.1). In doing so, we aimed to

strike a compromise between (a) evaluating performance across sufficiently large areas to justify

inferences about the models’ discriminatory power and (b) restricting evaluation to similar

environmental gradients as contained in the presence data so that validation metrics are not

artificially inflated (VanDerWal et al. 2009). Where possible, we defined backgrounds from

existing management units in which surveys were conducted (National Park boundaries for

‘Lakefield’ and ‘Oyala Thumotang’). Otherwise, we arbitrarily applied a 15 km buffer to data

points (for ‘Balkanu’, ‘NAQS’ and ‘ALA’), which corresponded to five times the highest expert-

elicited mobility threshold (home range boundary) in this study.

90

Table 4.3 Validation data sets with ancillary information.

Name Source No. of

records

Date of collection Method and purpose

of collection

Background

size (km2)

Typical habitat types

Balkanu Balkanu Cape

York Development

Corporation

⋅ dry: 181

⋅ wet: 67

⋅ Sep-Nov 2013-14

⋅ May 2015

Systematic aerial

survey and

management

(shooting)

⋅ dry: 3,954

⋅ wet: 3,089

Eucalyptus woodlands

& coastal wetlands

Lakefield Queensland Parks

& Wildlife Service

⋅ dry: 350

⋅ wet: 124

⋅ Oct-Dec 2009-13

⋅ Feb-May 2010-13

Systematic aerial

management

(shooting)

⋅ dry: 5,788

⋅ wet: 5,788

Eucalyptus / Melaleuca

woodlands, coastal

wetlands & grasslands

Oyala

Thumotang

Queensland Parks

& Wildlife Service

wet: 263 Apr-May 2010-13 Systematic aerial

management

(shooting)

wet: 3,819 Eucalyptus woodlands,

riparian Melaleuca

forests & rainforests

ALA Atlas of Living

Australia

⋅ dry: 111

⋅ wet: 144

⋅ Sep-Dec 98-2012

⋅ Feb-May 91-2012

Surveys and

sightings (purpose

unknown)

⋅ dry: 36,024

⋅ wet: 41,511

Eucalyptus woodlands,

floodplains, Melaleuca

forests & mangroves

NAQS Northern Australia

Quarantine

Strategy

dry: 103 Sep-Nov 2007-10 Opportunistic aerial

survey and disease

sampling (shooting)

dry: 11,630 Eucalyptus woodlands,

floodplain grasslands

& chenopod scrublands

We used the Continuous Boyce Index (CBI) to evaluate model performance (Appendix 4.4).

This method was developed specifically for evaluation against presence-only observations (Boyce

et al. 2002; Hirzel et al. 2006). A predicted-to-expected (P/E) ratio was computed as the (predicted)

proportion of presence records coinciding with a specified range of HSI values (bin) divided by the

(expected) proportion of the validation background covered by that bin. The CBI measures the

Spearman rank correlation coefficient of P/E against HSI and varies from 1 (correct model, P/E

steadily increases as HSI increases) to –1 (false model, P/E steadily decreases with increasing HSI).

Values close to zero indicate a random prediction (Hirzel et al. 2006). We computed the P/E ratio

and CBI for each expert model / validation data combination. We also computed the proportion of

each validation background expected to be highly or very highly suitable (HSI ≥ 60). This resulted

in 28 sets of validation metrics (P/E ratio + CBI + HSI ≥ 60) per seasonal scenario (6 expert models

x 4 validation data + 1 expert-averaged model x 4 validation data; Appendix 4.4). This allowed us

to evaluate uncertainty from diverging expert knowledge and test the expectation that an expert-

averaged model performs best (McBride et al. 2012b).

91

4.4 Results

4.4.1 Sensitivity analysis

Sensitivity analysis revealed that habitat suitability in the Bayesian network model was most

strongly influenced by water and food resources (24% and 26% variance reduction respectively)

and to a lesser degree by protection from heat (6%) and protection from disturbance (3%) (Fig 4.4).

Habitat suitability was least sensitive to expert opinion (< 1% variance reduction). Hence, experts

were mostly in agreement about the relative importance of the four habitat variables and quantified

model relationships similarly. However, one expert weighted the heat protection requirement as

highly as water and food resources (Expert 1: 15% variance reduction; Fig 4.4).

Fig 4.4 Sensitivity of habitat suitability to four habitat variables and expert opinion. Sensitivity to

findings was calculated as % variance reduction for each individual expert model and an averaged model

(red bar and percentages).

4.4.2 Predictive performance

In general, expert-elicited habitat suitability models performed well against the validation data

(Table 4.4 and Fig 4.5). For some seasonal scenario / validation data combinations, all expert

models performed well (e.g. wet or dry season model / Lakefield data). For others, there were

considerable differences (e.g. wet or dry season model / ALA data) (Table 4.4). As the HSI

Bayesian network model was not sensitive to expert opinion, model uncertainty stemmed largely

92

from disagreement about feral pigs’ resource-seeking home range movements, i.e. expert-elicited

response-to-pattern curves (Fig 4.3). Expert models that assumed high mobility thresholds (3 km for

experts 2, 4 and 6; Fig 4.3) predicted the highest proportions of suitable habitat in all validation

backgrounds. Average HSI ≥ 60 for these three expert models ranged between 71-78% in the wet

season and 36-42% in the dry season. Model accuracy was also generally highest for these experts,

with average CBI ranging between 0.58-0.85 in the wet season and 0.69-0.86 in the dry season.

Both metrics were lowest for the expert model that assumed the least mobility (1 km for expert 1;

Fig 4.3). Here, average HSI ≥ 60 was 47% (wet) and 15% (dry) and average CBI was 0.33 (wet) and

0.44 (dry).

The expert-averaged model performed similar to, or better than, the best individual expert

models for most validation data, except for the dry season model / NAQS data and dry season

model / ALA data combinations. Its average CBI across all validation data sets was 0.85 in the wet

season and 0.67 in the dry season. The predicted proportion of suitable habitat was modest

compared to individual expert models (except for expert 1), with an average HSI ≥ 60 of 60% (wet)

and 26% (dry). While averaging did not increase model accuracy as expected, it produced

consistently accurate results for all validation data (Table 4.4, Fig 4.5 and Appendix Fig 4.5).

Table 4.4 Validation metrics for individual expert and averaged seasonal habitat suitability models.

Habitat

suitability

model

Validation metrics per model scenario / validation data combination

Wet season scenario Dry season scenario

Balkanu Lakefield Oyala Thum ALA Balkanu Lakefield NAQS ALA

CBI HSI60 CBI HSI60 CBI HSI60 CBI HSI60 CBI HSI60 CBI HSI60 CBI HSI60 CBI HSI60

Expert 1 -0.22 59% 0.69 60% 0.66 41% 0.17 26% 0.59 28% 0.66 15% 0.27 12% 0.25 3%

Expert 2 0.94 83% 0.94 91% 0.61 77% 0.9 58% 0.94 64% 0.9 38% 0.94 44% 0.66 20%

Expert 3 0.73 80% 0.97 85% 0.64 68% 0.7 54% 0.94 58% 0.88 30% 0.75 44% 0.75 15%

Expert 4 0.83 82% 0.94 83% 0.7 68% 0.86 49% 0.98 56% 0.97 33% 0.72 42% 0.76 12%

Expert 5 0.38 74% 0.63 56% -0.24 69% 0 42% 0.89 50% 0.96 34% 0.74 30% 0.56 9%

Expert 6 0.44 83% 0.94 91% 0.43 77% 0.52 60% 0.94 63% 0.96 38% 0.35 44% 0.5 22%

Averaged 0.87 71% 0.92 70% 0.88 59% 0.71 39% 0.85 44% 0.97 26% 0.59 25% 0.27 7%

Validation was performed against four data sets per seasonal scenario (Table 4.3). We show the Continuous Boyce

Index (CBI) and proportion of validation background expected to be highly or very highly suitable habitat (HSI ≥ 60,

here shortened to HSI60). A CBI = 1 would indicate a perfectly accurate, a CBI ~ 0 a random, and a CBI < 0 a false

model.

93

Fig 4.5 Validation plots for individual expert and averaged seasonal habitat suitability models.

Validation was performed against four validation data sets per seasonal scenario (Table 4.3). The predicted-

to-expected (P/E) ratio (y axis) measures the proportion of feral pig presences relative to the proportion of

background pixels on a continuous scale of predicted HSI (x axis).

94

4.4.3 Seasonal habitat suitability

We present and discuss seasonal results only for the expert-averaged consensus model, which

produced consistently accurate results across the study region. Predicted habitat suitability varied

considerably between seasonal scenarios (Table 4.5 and Fig 4.6). Overall the model predicted

suitable habitat (HSI ≥ 40) in 36% of the study region (~640,000 km2) during the wet season and <

10% (~170,000 km2) during the dry season. Of this, about one quarter was highly or very highly

suitable habitat (HSI ≥ 60, 8% during the wet season and < 3% during the dry season, Table 4.5).

Fig 4.6 Seasonal habitat suitability for feral pig breeding in northern Australia. Habitat suitability

indices were averaged across all expert models and mapped for a wet (March / April – A) and dry (October /

November – B) season scenario.

95

Table 4.5 Amount of feral pig habitat in each habitat suitability index class per seasonal scenario.

Seasonal scenario Total area covered by each habitat suitability index class (km2 and %)

Very high High Moderate Low Very low

Wet season 16,726 (1%) 130,112 (~7%) 489,371 (~28%) 239,948 (~14%) 880,902 (~50%)

Dry season 11,147 (< 1%) 39,512 (~2%) 116,273 (~7%) 71,532 (~4%) 1,518,597 (~86%)

Habitat suitability also varied spatially between administrative units, vegetation types and land

use classes (Fig 4.6, Fig 4.7 and Appendix 4.6).

When analysed by states (shown in Fig 4.1), highly and very highly suitable feral pig habitat

was located mostly in Queensland, especially during the dry season. It was largely absent from

Western Australia in either scenario. The Northern Territory’s share increased more than three-fold

during the wet season. Moderately suitable habitat (40 ≤ HSI < 60) was somewhat more evenly

distributed across the study region, especially during the wetter months (covering about 40% of the

study region in Queensland, 20% in the Northern Territory and 10% in Western Australia,

Appendix Fig 4.6.1). Within each state, suitable habitat was concentrated in coastal environments

during the dry season (except for some large inland riverine and wetland systems), and expanded

widely during the wet season (Fig 4.6).

By broad vegetation types (Department of the Environment and Water Resources 2007; Fig

4.7), highly suitable habitat (HSI ≥ 60) was disproportionately found in rainforests, wetlands,

mangroves and modified (agricultural) vegetation, especially during the dry season (23%, 7%, 4%

and 17% share of total suitable habitat respectively). Yet, these vegetation types covered only 6.2%

of the study region combined. A large share of suitable habitat – 42% (dry) and 64% (wet) – was

also contained in the region’s dominant vegetation type, savanna woodlands, which was broadly

proportional to its overall cover (68% of the study region). Very highly suitable habitat was even

more concentrated in rainforests (68% (dry) and 54% (wet) share of HSI ≥ 80) and less frequently

found in savanna woodlands (22% (dry) and 29% (wet) share of HSI ≥ 80; Fig 4.7). During the dry

season, the vast majority of grasslands (98%), shrublands (93%) and woodlands (91%) were

modelled as unsuitable for feral pig breeding (HSI ≤ 40; Appendix Fig 4.6.2). During the wet

season, habitat concentration was weaker and suitable habitat was somewhat more evenly

distributed among vegetation types (Fig 4.7).

Habitat suitability was also unevenly distributed between land use classes (Australian Bureau

of Agricultural and Resource Economics and Sciences 2011; Fig 4.7). During the dry season, high

value land used for water resources, irrigated or dryland production together contained 17% of

96

suitable habitat on less than 2% of the study region. Interestingly, the 26% of the study region set

apart for nature conservation also contained a disproportionate amount of suitable habitat (52%

(dry) and 43% (wet) share of HSI ≥ 60). The dominant land use type “grazing natural vegetation”

contained a greater share of suitable habitat during the wet season (42%) than the dry season (24%)

on 71% of the study region. As with vegetation types, habitat concentration was further increased

for very highly suitable habitat (HSI ≥ 80) and weaker in the wet than in the dry season (Fig 4.7).

Fig 4.7 Share of modelled suitable habitat found in different vegetation types and land use classes for

each seasonal scenario. For each vegetation (panel A) or land use (panel B) category, we show its percent

share of highly (HSI ≥ 60) and very highly (HSI ≥ 80) suitable habitat during the dry (top bars) or wet

(bottom bars) season, compared to its percent share of the total study region (central bars). Spatial analyses

in panel (A) were based on Present Major Vegetation Groups (MVG V.4.1) and analyses in panel (B) on

Australian Land Use and Management classes (ALUM V.7).

4.5 Discussion

Effective management of invasive wildlife requires spatially-explicit knowledge of their

distribution and habitat use, especially for wide-ranging species (Venette et al. 2010). Yet,

continuous empirical information is rarely available over large areas (Stephens et al. 2015). Habitat

models can fill this knowledge gap, but their utility may be compromised by small study areas and

limited integration of species ecology or temporal variability (Franklin 2010). Here, we modelled,

for the first time, seasonally-specific habitat suitability for feral pigs, a widespread and harmful

97

invader, in northern Australia. Rigorous evaluation and validation showed that our resource-based,

expert-elicited model, which integrated important ecological factors such as home range movements

and breeding requirements, accurately predicted feral pig presence across the study region. Results

suggest that suitable feral pig habitat is more constrained in northern Australia than previously

thought, especially during the dry season. Mapped results may be used by land managers to

quantify impacts, assess risks, justify investments and target control activities. Our transparent

methodology could be applied to other wide-ranging species, especially in data-poor situations.

4.5.1 Seasonal habitat suitability

Our habitat model confirmed previous site-scale findings that the distribution and habitat use of

feral pigs in northern Australia is highly seasonal (Mitchell et al. 1982; Hone 1990a; Caley 1993;

Choquenot et al. 1996) for the entire region. Modelled habitat suitability for feral pig breeding and

persistence was mainly driven by seasonal availability of food and water resources, both of which

ultimately vary with northern Australia’s annual rainfall cycle. Our model indicated a four-fold

increase in suitable habitat during the wet season. Inter-annual climatic variability, which has been

shown to greatly affect feral pig populations in drier parts of Australia (Mitchell et al. 1982;

Choquenot 1998; Dexter 1998), was not investigated in this study. Our scenario approach could be

usefully extended to model cyclical, as well as seasonal, variability in feral pig distribution.

Seasonal fluctuations in habitat suitability were expressed at different spatial scales. At the

regional scale, habitat suitability varied along an east-west gradient. Suitable feral pig habitat was

concentrated in the eastern study region throughout the year. During the dry season, suitable habitat

contracted more in the west than in the east (eleven-fold in Western Australia, five-fold in the

Northern Territory and three-fold in Queensland). These patterns correspond well with prevailing

rainfall gradients and harsher dry season conditions in the west (Guerschman et al. 2009; Jones et

al. 2009). At the landscape scale, contiguous patches of suitable habitat were located predominantly

along the coastline during the dry season and expanded widely across the study region during the

wet season. If all suitable wet season habitat was to be used by feral pigs, this points to long

distance seasonal migration. However, such migratory behaviour in Australian feral pigs has been

rejected by previous research (Caley 1997; Mitchell et al. 2009; Hone 2012). Rather, empirical

findings suggest that feral pigs may expand and contract their home range in response to changing

conditions (Dexter 1999; Mitchell et al. 2009), or shift it entirely if adverse conditions persist

(Caley 1997; Choquenot et al. 1996). Hence, not all suitable wet season habitat may be within reach

of feral pig breeding herds dispersing from dry season refuges (Caley 1997). At the local scale, dry

season habitat was concentrated in productive rainforest, wetland and mangrove ecosystems as well

as high value agricultural lands, where resources remain abundant. While these dry season refuges

98

continued to provide suitable habitat during the wet season, habitat was more evenly distributed

among vegetation types and land use classes. This suggests that feral pigs forage widely into

grassland floodplains, savanna woodlands and coastal shrublands when conditions permit. Our

regional model showed that rainforests are a key year-round habitat for feral pig breeding (Fig 4.7).

This was partly due to model assumptions. For example, as actual freshwater presence was

inadequately mapped in densely vegetated rainforests, we assumed that water was uniformly

available (Table 4.2). Site-scale studies have already provided a more detailed understanding of the

fine-grained variations in actual habitat use within this broadly suitable habitat type (Mitchell 2002;

Mitchel et al. 2009; Wurster et al. 2012; Elledge et al. 2013).

While all parts of northern Australia contained at least some suitable feral pig habitat, our

model suggests that feral pigs are less widely distributed in the region than previously thought. For

example, moderately suitable habitat in Western Australia totalled between 3000 km2 (dry) and

32,000 km2 (wet) while Cowled et al.’s (2009) model predicted 89,000 km2 of suitable habitat in

the same area. Similarly, West (2008) reported that feral pigs are widespread throughout

Queensland while our model found that only 16% (dry) to 54% (wet) of this area contained all

required resources for persistent feral pig breeding. We note that (a) such area estimates are highly

dependent on habitat thresholds and therefore are unreliable and difficult to compare between

studies, and (b) our figures are possibly overestimates – they refer to a threshold of HSI ≥ 40, yet

validation plots (Fig 4.5) suggested that a higher, more restrictive threshold (e.g. HSI ≥ 60) may

better discriminate between suitable and unsuitable habitat in most environments. We suggest our

lower estimates are defensible when considering a number of methodological improvements in our

study. First, we specifically modelled resource requirements of feral pig breeding herds, which are

more limiting than those of solitary boars (Caley 1997; Choquenot et al. 1996; Mitchell et al. 2009).

Previous statistical models (Cowled et al. 2009) and mapping studies (Mitchell et al. 1982; West

2008) did not make this distinction, although long-term occupancy critically relies on breeding.

Second, previous work did not distinguish between seasonal scenarios but included any location

where feral pigs have occurred or may occur. This may have approximated wet season habitat and

likely led to overestimations by failing to consider the dry season as a limiting factor for feral pigs

in Australia’s north (Hone 1990a; Caley 1993). Finally, we used a finer spatial resolution (1 ha)

than previous studies (25 km2 in Cowled et al. 2009 or ~250 km2 in West 2008), resulting in more

detailed predictions and less upscaling error (e.g. one “suitable” pixel equalled 25 km2 in Cowled et

al. 2009, even if only a fraction of this area actually contained suitable habitat).

99


Model results were robust to both expert opinion and a range of independently collected

validation data sets of feral pig presence. All six experts contributing to the final consensus model

provided similar parameter estimates for the Bayesian network model of habitat suitability. That is,

they all agreed that water and food resources are more important to persistent feral pig breeding in

northern Australia than protection from heat and protection from disturbance. However, validation

revealed inconsistent performance between expert models in some instances. We attributed this

model uncertainty mainly to different expert assumptions about feral pigs’ resource-seeking home

range movements, described in elicited response-to-pattern curves (Fig 4.3)10. Our simple approach

to reducing uncertainty, equal-weighted expert averaging (Martin et al. 2012; McBride et al.

2012b), yielded similar, or more accurate, results than the best expert models for most validation

data.

While the averaged habitat suitability model always performed well, there were differences

between seasonal scenarios and validation data. This points to some limitations of our study. First,

most validation data were collected in environments with an above-average proportion of suitable

feral pig habitat. Second, model parameters were elicited from experts with field knowledge mostly

from resource-abundant environments in the eastern study region and applied to all of northern

Australia. Model accuracy in resource-poor inland environments and in those portions of the study

region for which no validation data were available and that were outside the expertise of our panel

of experts (especially Western Australia) needs further investigation (Murray et al. 2009). For

example, feral pigs may also sustain themselves in ‘unsuitable’ habitat during the dry season from

resources not included in our model (e.g. carrion and other animal matter). Third, reporting bias in

the data points used for validation (Phillips et al. 2009) and our definition of validation

backgrounds may have affected validation results. For example, most data sets were biased towards

highly suitable sites as management-focused survey efforts mostly concentrated on sites known to

be impacted by feral pigs. Less suitable sites, which nevertheless support feral pig breeding, may

therefore have few or no data points recorded, resulting in somewhat inflated CBI values. Further,

different buffering choices for defining the ALA, NAQS or Balkanu validation backgrounds may

have yielded poorer (likely for backgrounds that are more narrowly defined around presence

records than our 15 km buffer) or enhanced (likely for larger backgrounds that encompass high

proportions of unsuitable habitat especially in the spatially disjunct ALA data set; Appendix Fig

4.5) performance results. Fourth, the Continuous Boyce Index validation method is well suited to

10 In Chapter 3 I showed that such assumptions affect the amount and distribution of suitable habitat as well as model performance.

100

presence-only validation data but cannot evaluate model specificity, i.e. its ability to correctly

predict absences and minimize false positives (Boyce et al. 2002; Hirzel et al. 2006). Finally, we

applied the same expert-elicited mobility thresholds and distance-dependent response-to-pattern

curves to the entire study region and both seasonal scenarios. Yet, feral pigs may adjust their

resource-seeking home range movements to environmental conditions (Caley 1997; Dexter 1999;

Mitchell et al. 2009) and in fact respond to aspects of landscape structure other than distance such

as resource composition or heterogeneity (Turner & Gardner 2015).

A listing of study limitations is most useful for guiding future research. We refer to van

Klinken et al. (2015) who suggest methods to evaluate possible errors in expert-elicited models. In

particular, our study may be usefully improved by: (1) systematically collecting presence / absence

data, also from resource-poor environments, to eliminate reporting bias in validation data and

enable evaluation of both model sensitivity and specificity; (2) field-validating the accuracy of

spatial data proxies and whether they match the states of model explanatory variables; (3) revising

model parameters with experts from the Northern Territory and Western Australia; and (4)

conducting a multi-scale study (Jackson & Fahrig 2015; McGarigal et al. 2016) to better understand

feral pigs’ response to spatial patterns of resources with varying quality in different types of

environments.

4.5.3 Management implications

Our regional-scale, seasonally-specific and rigorously validated results can serve to better

manage the impacts of feral pigs in northern Australia. For example, we estimated habitat suitability

per broad vegetation and land use types. When combined with information on impacts or costs in

each category (Barrios-Garcia & Ballari 2012; Mitchell 2010; Krull et al. 2016), this may help to

more accurately quantify environmental or economic impacts across any area of interest. Habitat

suitability could also be analysed in other management units to justify investments in population

control or, if verified by stratified field surveys, to serve as a monitoring baseline in adaptive

management programs. Further, because habitat suitability was explicitly referenced to feral pig

breeding herds, it is a useful indicator of establishment risk for infectious animal diseases, which

often depend on a persistent supply of young susceptible animals (Kramer-Schadt et al. 2007;

Cowled & Garner 2008). Habitat connectivity for feral pigs is also a critical parameter in

understanding disease spread, but has not been explicitly studied in northern Australia. Future

models may integrate validation results from this study to derive habitat quality thresholds for patch

delineation (Hirzel et al. 2006) and reverse habitat suitability to define resistance to movement, i.e.

the cost of traversing habitat of different quality (Zeller et al. 2012). Finally, our resource-based

model captures the relative importance of four habitat requirements as well as spatial interactions

101

such as landscape complementation or supplementation (Dunning et al. 1992). This knowledge may

be used to manipulate resource access at strategic locations (e.g. by exclusion fencing or local

eradication) and model the effect on habitat suitability and connectivity within the broader

landscape.

4.6 Appendices

Appendix 4.1 Bayesian network conditional probability tables (supplements Fig 4.2).

Appendix 4.2 Classification of categorical spatial data attributes to match the states of model

explanatory variables (supplements Table 4.2).

Appendix 4.3 Methodology for spatial pattern suitability analysis (annotated R code).

Appendix 4.4 Methodology for model validation (annotated R code).

Appendix 4.5 Maps of seasonal habitat suitability per validation area.

Appendix 4.6 Seasonal habitat suitability per state and per broad vegetation type.

102

Chapter 5 Modelling habitat connectivity for biosecurity: the risk of infectious

disease spread in feral pigs in northern Australia

Summary

In this chapter I model and map seasonal habitat connectivity for feral pigs in northern

Australia. I apply the results from Chapter 4 to delineate contiguous patches of feral pig breeding

habitat, parameterize matrix resistance to inter-patch dispersal, and model seasonal connectivity in

the landscape matrix using a circuit-theoretic approach. Then, I estimate connectivity between

patches and evaluate its sensitivity to a range of model assumptions that are linked to gender-

specific differences in feral pig dispersal ability. This chapter addresses research question 3,

providing an indicative measure of the seasonal spread risk of directly transmitted infectious

diseases within feral pigs in northern Australia. It is being prepared for submission to Journal of

Applied Ecology.

5.1 Abstract

Directly transmitted diseases in wildlife hosts can spread between infected and susceptible

populations via host dispersal. Risk-based disease management requires spatial knowledge of the

functional links between populations, often over areas too large for empirical studies. Yet, habitat

connectivity models have not been widely applied to address this problem. We modelled seasonal

habitat connectivity for feral pigs, an abundant invader that is a susceptible host to numerous

infectious diseases, in northern Australia. We used a circuit-theoretic approach to model potential

dispersal paths in the landscape matrix and measured connectivity between patches of breeding

habitat as an indicator of disease spread risk. To account for gender-specific differences in dispersal

ability, we evaluated a range of model assumptions on the level of matrix resistance and on two

parameters for delineating dispersal paths. Habitat connectivity was significantly affected by

changes in these assumptions. Specifically, results indicated that dispersal paths of feral pigs, and

connectivity between habitat patches, differ considerably with climatic conditions and between

breeding herds and solitary boars. We identified three broad types of patches: some were always

isolated; some were always connected to large habitat components; and some were mostly isolated

but became connected to large habitat components only during the wet season or for wide-ranging

boars. Our results can directly inform risk-based management of infectious animal diseases at all

stages of invasion, including early detection surveillance, preparedness and incursion response. The

resource-based approach could readily be applied to other wildlife hosts, especially in regional-

scale and data-constrained studies.

103

5.2 Introduction

Infectious animal diseases are a major biosecurity threat in an increasingly connected world

(Jones et al. 2008). New incursions may seriously impact human, livestock and wildlife health and

incur large costs on industry and government (Daszak et al. 2000). Wildlife hosts are a well-

recognized risk factor for disease introduction, establishment and spread (Daszak et al. 2000; Jones

et al. 2008; Reisen 2010). For example, bats are important reservoir hosts of several emerging

pathogens, including the rabies, Nipah or Hendra viruses (Calisher et al. 2006); migratory birds

may contribute to global spread of highly pathogenic avian influenza H5N1, with devastating

consequences to human health and the poultry industry (Olsen et al. 2006); wild boar infected with

classical swine fever caused repeated outbreaks in domestic herds in Germany during the 1990s,

necessitating costly interventions (Fritzemeier et al. 2000; Kramer-Schadt et al. 2007). Invasive

wildlife species can be particularly problematic, introducing disease to susceptible hosts (Daszak et

al. 2000; Macpherson et al. 2016) or complicating disease management in widespread and

expanding populations. However, realistic spatially-explicit descriptions of wildlife hosts’

distribution and habitat use are rarely integrated into risk-based analyses of disease establishment

and spread (Ostfeld et al. 2005; Ward et al. 2011; Rees et al. 2013; Macpherson et al. 2016). Here,

we apply a resource-based approach (Hartemink et al. 2015) to model habitat connectivity for feral

pig (Sus scrofa) populations in northern Australia. The species is an abundant invader that is

susceptible to numerous emergency diseases such as foot-and-mouth disease, classical swine fever

and Japanese encephalitis (Animal Health Australia 2011).

We built on the theoretical ‘patch-corridor-matrix’ (Forman 1995) and metapopulation (Hanski

1998) frameworks, where local wildlife populations are viewed as persisting in patches of suitable

breeding habitat that are embedded in a matrix of unsuitable habitat. Some areas in the matrix

function as dispersal corridors between patches. Metapopulation persistence depends on local

population dynamics and inter-patch movements (Hanski 1998; Hastings 2014). A metapopulation

perspective on feral pigs in Australia is supported by landscape-genetic evidence of population

structuring, especially in heterogeneous landscapes (Hampton et al. 2004; Cowled et al. 2008;

Lopez et al. 2014). Patches and corridors also affect the risk of invasion and persistence by directly

transmitted diseases in a wildlife host metapopulation (Anderson et al. 1986; Riley 2007). Initial

establishment is contingent on locally dense breeding populations with high contact rates that let an

infectious pathogen’s basic reproductive rate R0 exceed 1. Subsequent disease spread requires host

dispersal between infected and susceptible subpopulations. Disease persistence depends on a lasting

supply of susceptible individuals within a regionally connected host metapopulation (Anderson et

al. 1986; Kramer-Schadt et al. 2007). Risk-based management of infectious diseases with wildlife

104

hosts would benefit from spatially-explicit knowledge of habitat connectivity at the regional scale

(Ostfeld et al. 2005; Riley 2007; Cowled & Garner 2008; East et al. 2013).

Connectivity research, which broadly aims at identifying links between patches, has

proliferated in recent decades (Tischendorf & Fahrig 2000; Galpern et al. 2011; Zeller et al. 2012;

Kool et al. 2013). Applications have mostly examined metapopulation dynamics (Moilanen &

Hanski 2001; Urban et al. 2009) or conservation benefits such as reserve networks (Beier et al.

2011; Doerr et al. 2011; Rayfield et al. 2011). More recently, research has also focused on the

management of invasive species (Glen et al. 2013; Stewart-Koster et al. 2015) or infectious diseases

(Riley 2007; Rioux Paquette et al. 2014; Macpherson et al. 2016). Modelling methods vary broadly

according to research aims and definitions of connectivity. Tischendorf & Fahrig (2000) defined

connectivity as “the degree to which the landscape facilitates or impedes movement” (p. 7), while

Kool et al. (2013) distinguished between structural, functional and genetic connectivity. Functional

approaches that consider structural heterogeneity in the landscape as well as species-, gender-,

scale- and time-specific factors have been increasingly recommended (Hamilton & Mather 2009;

Doerr et al. 2011; Rayfield et al. 2016). When modelling habitat connectivity for disease

management, identifying multiple dispersal paths between infected and susceptible patches is of

greater interest than optimal (least-cost) dispersal paths or genetic connectivity over long

timeframes (McRae et al. 2008; Kool et al. 2013). A variety of connectivity metrics have been

developed (Moilanen & Nieminen 2002; Kindlmann & Burel 2008; Rayfield et al. 2011), broadly

describing connectivity either as an attribute of the matrix, or the patch (Moilanen & Hanski 2001;

Calabrese & Fagan 2004). For the management of infectious diseases both interpretations are highly

relevant. Here, we emphasize this conceptual distinction by referring to ‘matrix connectivity’

(Where in the matrix is disease spread via host dispersal likely to occur?) and ‘patch connectivity’

(To which other patches is an infected patch connected?). We use the term ‘habitat connectivity’ as

an overarching concept that integrates both perspectives in the patch-corridor-matrix framework.

Here, we modelled the risk of spread of directly transmitted diseases within feral pigs in

northern Australia. We focused on understanding habitat connectivity for the wildlife host, as such

knowledge gaps have impeded epidemiological models and disease management in Australia and

elsewhere (Ostfeld et al. 2005; Cowled & Garner 2008; Rees et al. 2013; Macpherson et al. 2016).

Previously, we modelled habitat suitability for feral pig breeding in northern Australia (Chapter 4).

Our first objective was to apply the resource-based, empirically validated results of this earlier

study to model ‘matrix connectivity’ across the entire region. Our second objective was to measure

‘patch connectivity’ (Rayfield et al. 2011) as an indicator of disease spread risk. Previous

simulation models of disease spread in feral pigs have suggested that seasonal (distributional

105

patterns and dispersal behaviour vary according to conditions) and gender-specific (dispersal by

mostly solitary boars appears much less resource-constrained than that of breeding herds) factors

may affect disease transmission (Doran & Laffan 2005; Milne et al. 2008). Hence, our third

objective was to apply the model to two seasonal scenarios and to evaluate the effect of changes in

model parameters that may be reflective of gender-specific differences in dispersal ability.

5.3 Materials and methods

5.3.1 Study region and species

Our study region (1.76 million km2) in northern Australia spanned three states: Queensland,

Western Australia, and the Northern Territory (Fig. 5.1). The region has a tropical climate with a

seasonal rainfall cycle, alternating between a wet (November / December to April / May) and a dry

(May / June to October / November) season. Dominant monsoonal savanna woodlands and semi-

arid grasslands are interspersed with riverine channels, coastal wetlands, fragments of wet tropical

rainforest and seasonally inundated floodplains (Fox et al. 2001). Feral pigs (Sus scrofa) have

invaded most of the region. They are now reported to be widespread in the east and localised in the

north and west, achieving highest local densities in resource-abundant wetlands and floodplains

(West 2008). Savanna woodlands are not a preferred habitat (Choquenot et al. 1996).

Fig 5.1 Study region in northern Australia (shaded in grey).

5.3.2 Habitat connectivity model

5.3.2.1 Overview

We followed four broad methodological steps to model patch connectivity (Fig 5.2). Each step

was applied to two seasonal scenarios: the late wet season (March to April), when resources

required by feral pigs are abundant and widely distributed, and the late dry season (October to

November), when resources are scarce and scattered across the landscape. First, we parameterized

matrix resistance to dispersal from a previous resource-based habitat suitability model (Chapter 4).

106

Second, we delineated habitat patches large enough to sustain local breeding populations based on

patch quality and size thresholds also derived from the habitat suitability model. Third, we modelled

matrix connectivity using the methodology developed by Pelletier et al. (2014) for regional-scale

applications of the circuit-theoretic approach (McRae et al. 2008). Fourth, we estimated patch

connectivity, by computing, for each habitat patch, the total area the patch is connected to via

dispersal corridors. For each seasonal scenario, we computed 27 alternative patch connectivity

values based on three-way changes in ‘matrix resistance to dispersal’ (ResDisp), ‘inter-patch

dispersal distance’ (DispDist) and ‘corridor delineation threshold’ (CDThresh). Modelling

procedures and parameter estimates are explained in detail below.

5.3.2.2 Matrix resistance

Matrix resistance refers to the permeability of landscape elements to movement or the costs

(energy expenditure, mortality risk) associated with traversing them (Kindlmann & Burel 2008;

Zeller et al. 2012). Resistance surfaces, which assign species-specific cost values to each matrix

pixel, can be parameterized in several ways (Zeller et al. 2012). Here, we adopted a resource-based

approach (Hartemink et al. 2015) using a previous seasonal habitat suitability model (Chapter 4).

The model integrated expert-elicited Bayesian networks and spatial pattern suitability analysis to

characterize habitat quality for feral pigs as a function of complementary access to four key

resource requirements (water, food, heat refuge and disturbance refuge) within heterogeneous home

ranges. Mapped ‘habitat suitability indices (HSI)’ were empirically validated (Chapter 4).

Here, we computed ‘matrix resistance to dispersal (ResDisp)’ at a fine spatial resolution of 1 ha

by reversing the modelled HSI (Chapter 4; see Beier et al. 2007 for a similar approach). The

resource-based model specifically characterized habitat quality for feral pig breeding. To account

for gender-specific differences between dispersal-constrained breeding herds and wide-ranging

boars (Choquenot et al. 1996; Gabor et al. 1999; Milne et al. 2008), we parameterized three

alternative resistance levels. We linearly converted HSI to range between either 1-5 (‘low’

ResDisp), 1-21 (‘moderate’ ResDisp) or 1-100 (‘high’ ResDisp), while reversing values so that

higher HSI obtained lower resistance values using the formula:

ResDisp = ResDispmax – (((HSI – HSImin) / (HSImax – HSImin)) * (ResDispmax – ResDispmin))

Narrower ranges imply greater ease of movement through low quality matrix habitat, which may

approximate dispersal by boars. Wider ranges place higher penalties on such movements, reflecting

more constrained movements by breeding herds. An intermediate scenario was applied to reflect

uncertainty and behavioural variability.

107

Fig 5.2 Methodology for modelling habitat connectivity for feral pigs. For each of four broad

methodological steps, we show methods used, workflow and outputs generated. Methods for the habitat

suitability model (step 1a) are detailed in Chapter 4.

108

5.3.2.3 Habitat patches

First, contiguous habitat patches of sufficient quality and size to allow breeding populations to

persist without needing to interact with other patches were delineated. We used Gnarly Landscape

Utilities: Core Mapper software (Shirk & McRae 2013) with parameters specific to feral pigs.

First, we derived seasonal ‘habitat quality thresholds’ from HSI model validation using the

Continuous Boyce Index method (Chapter 4; Hirzel et al. 2006). We plotted the predicted-to-

expected (P/E) ratio, which compares, for each HSI, the proportion of observed feral pig presences

to the proportion of all pixels in a validation background (Fig 5.3). HSI values may be pooled when

the P/E ratio shows a flat or negative slope (Hirzel et al. 2006). We specified a threshold value for

‘high quality breeding habitat’ at HSI ≥ 60 and HSI ≥ 65 for the dry and wet season respectively

(Fig 5.3). Core Mapper applied thresholds to individual pixels and neighbourhood averages in order

to exclude small suitable areas embedded in low quality habitat (Beier et al. 2007; Shirk & McRae

2013). We calculated neighbourhood averages within a 200 m radius and using a slightly reduced

threshold (HSI ≥ 59 (dry) and HSI ≥ 64 (wet)) so that linear features such as rivers, which are an

important feral pig habitat especially in semi-arid areas (Cowled et al. 2008), were not removed.

109

Fig 5.3 Validation plots for seasonal habitat suitability models. Predicted-to-expected (P/E) ratios are

plotted for each of four independent validation data sets (thin labelled lines) and as a smoothed average of all

data sets (thick line). Vertical dotted lines indicate HSI thresholds used for patch delineation.

Next, Core Mapper ‘patch expansion’ was used to merge suitable areas separated by narrow

gaps of low quality habitat (Girvetz & Greco 2007; Shirk & McRae 2013). We assumed that feral

pigs can cross these gaps because of their home range mobility. We derived a movement radius of

2.33 km by averaging expert estimates on home range boundaries (Chapter 4). In Core Mapper,

these movements were cost-weighted according to matrix resistance (Shirk & McRae 2013).

Finally, merged patches below a specified ‘minimum patch size’ were considered insufficient to

support independent breeding populations and removed (Beier et al. 2007; Shirk & McRae 2013).

We defined this at 17.1 km2 by averaging the home range sizes suggested by experts in Chapter 4,

assuming that herds can persist within one resource-rich home range neighbourhood (Choquenot et

al. 1996). Patch delineation in Core Mapper was affected by matrix resistance, resulting in networks

of patches differing in shape, distribution and overall number for each resistance level.

5.3.2.4 Matrix connectivity

We used a circuit-theoretic approach, implemented in the Circuitscape model (McRae & Beier

2007; McRae et al. 2008; McRae et al. 2013) to model ‘matrix connectivity’ across the region.

Circuitscape treats the landscape analogous to an electrical circuit, passing current between pairs of

nodes (patches) through an intervening network of resistors (matrix pixels). Circuitscape computes

the current density for each pixel, which can be interpreted as the probability of a species moving

through that pixel via random walk theory (McRae et al. 2008). We used a recent extension of

Circuitscape described by Pelletier et al. (2014), which models movement probabilities across

entire landscapes rather than between nodes. It does not require delineated habitat patches

(Moilanen 2011; Koen et al. 2014) and overcomes computational limitations for regional scale

applications (see Leonard et al. 2016 for a high-performance computing solution).

We implemented the methodology for each level of matrix resistance using the R 3.2.2 and

Circuitscape 4.0 software (McRae et al. 2013; R Core Team 2015; Appendix 5.1). It involved (a)

dividing the 175 million study region pixels into 127 square tiles (150 x 150 km); (b) adding a 150

km buffer around each tile, filled with values drawn randomly from the resistance surface to

minimize border effects (Koen et al. 2010); (c) running the Circuitscape model in pairwise mode,

passing directional current between two parallel nodes placed as thin (one pixel-wide) strips along

the outer (west-east or north-south) borders of each buffered tile; (d) removing buffers and

reassembling tiles into directional mosaics spanning the entire study region and (e) computing

110

‘omnidirectional current density (OCD)’ by multiplication (Pelletier et al. 2014). Finally, we log10

transformed OCD (OCDlog10) for better visualisation and identification of consistent ‘corridor

delineation thresholds (CDThresh)’ for all three levels of matrix resistance. The CDThresh

parameter was used to distinguish between dispersal paths and those parts of the matrix where no

dispersal occurs. We selected three arbitrary threshold levels reflecting different levels of habitat

quality required for dispersal by breeding herds (‘strict’ CDThresh where dispersal is possible only

in high quality habitat with an OCDlog10 ≥ -7.25) and solitary boars (‘loose’ CDThresh where

dispersal can occur in most habitats with an OCDlog10 ≥ -7.75). An intermediate scenario

(‘moderate’ CDThresh with -7.25 > OCDlog10 ≥ -7.75) was again applied to reflect uncertainty and

behavioural variability.

5.3.2.5 Patch connectivity

After pairing delineated patches and modelled OCD based on the level of matrix resistance

(high, moderate or low), we computed patch connectivity for each delineated habitat patch and for

the entire patch network (Rayfield et al. 2011). At the patch level we measured the total habitat area

(including breeding habitat within patches and dispersal habitat within the matrix) that each patch is

connected to (Moilanen & Nieminen 2002; Rayfield et al. 2011). First, we allocated delineated

dispersal habitat (see above) to the nearest habitat patch. We defined three alternative ‘inter-patch

dispersal distances (DispDist)’ at 5, 10 and 15 km, again reflecting differences for breeding herds

and solitary boars (Choquenot et al. 1996; Cowled & Garner 2008). These were implemented as

accumulative cost distances within uniformly weighted corridors (each pixel having a cost of ‘100

m’) to allow for non-linear corridor dispersal (Moilanen & Nieminen 2002; Galpern et al. 2011).

Next, we merged habitat patches that were connected via cost-weighted dispersal into ‘habitat

components’ (Rayfield et al. 2011), and assigned each individual habitat patch a value reflecting the

total size of its habitat component. Finally, we computed network-level patch connectivity by

aggregating patch-level values across the study region (Galpern et al. 2011). Analyses were

implemented in the ESRI ArcGIS 10.2 Spatial Analyst and ModelBuilder software (Appendix 5.2).

5.3.3 Model evaluation

To account for a lack of empirical data about model parameters, and intra-specific differences

between breeding herds and solitary boars, we evaluated patch connectivity under a range of model

assumptions. For each season, we computed 27 alternative patch connectivity scenarios in response

to three-way changes in three key parameters affecting various modelling steps (Fig 5.2; Galpern et

al. 2011, Moilanen 2011): ‘matrix resistance to dispersal’ (ResDisp, levels: high, moderate, low)

influenced patch delineation and matrix connectivity; ‘corridor delineation threshold’ (CDThresh,

111

levels: strict, moderate, loose) and ‘inter-patch dispersal distance’ (DispDist, levels: 5 km, 10 km,

15 km) influenced patch connectivity.

We evaluated network-level effects analogous to a factorial experiment. We focused on

differences in median, rather than mean, size of habitat components because we anticipated non-

Gaussian distributions dominated by many small components. First, we used the Kruskal-Wallis

rank sum test to probe for significant effects of changes in ResDisp irrespective of DispDist and

CDThresh. Where significant differences in the distribution of component sizes were found, we

performed two-way comparisons between ResDisp levels using the Mann-Whitney U test. Next we

applied the same non-parametric tests of significance to the parameters CDThresh and DispDist at

each ResDisp level. Separate evaluation was necessary because the shape, location and number of

patches to which component sizes were assigned differed between levels. We also investigated

possible interactions between CDThresh and DispDist. To visualise changes in patch connectivity

due to interaction effects, we linearly transformed absolute component sizes into relative values

(range 0-1: minimum to maximum area each patch is connected to). All analyses were implemented

in the R 3.2.2 software (R Core Team 2015).

5.4 Results

5.4.1 Matrix connectivity

There were large seasonal differences in matrix connectivity for feral pig dispersal. According

to variations in resource availability, large tracts of additional dispersal habitat became available in

inland areas during the wet season (Fig 5.4). The level of matrix resistance (ResDisp) assumed

during modelling also affected the amount of dispersal habitat (Table 5.1). As matrix resistance

influenced the ease of movement through low quality habitat rather than the spatial distribution of

resources, lower resistance levels mostly widened dispersal paths without adding dispersal habitat

in different regions. A notable exception occurred when matrix resistance was ‘low’ and the

corridor delineation threshold was ‘loose’. With this parameter combination, all matrix habitat was

considered suitable for feral pig dispersal in both seasons (Table 5.1). Spatial data files showing

seasonal matrix connectivity for each ResDisp level will be openly available upon publication of

this Chapter (to be submitted to Data Dryad).

112

Fig 5.4 Seasonal matrix connectivity in northern Australia (for moderate ‘matrix resistance

(ResDisp)’). Dispersal habitat is shown for three ‘corridor delineation thresholds (CDThresh)’. A

methodological artefact increased ‘omnidirectional current density (OCDlog10)’ at some tile corners. This

occurred mostly in low quality habitat (white areas) and had minimal impact on patch connectivity.

113

Table 5.1 Amount of dispersal habitat in the study region.

Wet season scenario

ResDisp CDThresh (% of study region)

strict mod loose no disp.

high (1-100) 4.7 19.4 44.8 55.2

mod (1-21) 6.3 27.5 57.6 42.4

low (1-5) 12 41 100 0

Dry season scenario

ResDisp CDThresh (% of study region)

strict mod loose no disp.

high (1-100) 1.8 4.3 12.5 87.5

mod (1-21) 2.2 5.9 20.1 79.9

low (1-5) 3.1 11.2 100 0

For each level of ‘matrix resistance (ResDisp)’, we compare the proportion of dispersal habitat for each ‘corridor

delineation threshold (CDThresh)’, compared to the proportion of non-dispersal habitat (shortened to ‘no disp.’).

5.4.2 Patch connectivity

5.4.2.1 Network-level connectivity

At the network level, patch connectivity also varied strongly between seasons (Fig 5.5 and

Appendix Table 5.3.1). The median and mean size of connected habitat components ranged

between 341-411 km2 and 7,820-45,174 km2 respectively during the dry season. During the wet

season, habitat component sizes increased considerably to 2,697-6,932 km2 (median) and 32,463-

86,139 km2 (mean). Sizes were distributed unevenly at all levels of matrix resistance, with many

small (isolated), some intermediate and few large (well-connected) habitat components (Fig 5.5).

Interestingly, while mean size increased steadily with decreasing resistance, median size remained

near constant (Appendix Table 5.3.1). This reflects a large increase in size of few well-connected

habitat components, and the addition of many small and isolated patches exceeding the specified

‘minimum patch size’ (doubling the overall patch count). When matrix resistance was ‘high’ or

‘moderate’, resource-poor environments (e.g. inland areas and all of Western Australia; Fig 5.1)

contained few or no patches. When testing for statistical significance, we found that the distribution

of component sizes was significantly different for ‘low’ resistance compared to both other levels in

both seasons (p < .001). There was no significant difference between ‘moderate’ and ‘high’ matrix

resistance in either the dry (p = .704) or the wet (p = .098) season (Appendix Table 5.3.1).

114

At each level of matrix resistance, we plotted and analysed median component size in response

to two-way changes in the ‘corridor delineation threshold (CDThresh)’ and ‘inter-patch dispersal

distance (DispDist)’ parameters (Fig 5.6 and Appendix Table 5.3.2). Using the Mann-Whitney U

test, we found significant effects for both parameters, as well as interaction effects between them,

but not for each factorial combination. For example, during the dry season, DispDist significantly

affected patch connectivity only when CDThresh was loose or when CDThresh was moderate and

ResDisp was low or moderate. Otherwise DispDist was not significant. Conversely, CDThresh had

significant effects at each factorial combination of ResDisp and DispDist. Yet, a major increase in

median component size was only observed when CDThresh changed to loose (Fig 5.6a-c and

Appendix Table 5.3.2a). During the wet season, DispDist and CDThresh interacted significantly at

all factorial combinations except when CDThresh was strict and ResDisp was high or moderate (no

significant effects of DispDist) or when CDThresh was strict and ResDisp was low (significant

decrease in patch connectivity only for DispDist = 5 km) (Fig 5.6d-f and Appendix Table 5.3.2b).

5.4.2.2 Patch-level connectivity

Individual habitat patches were affected differently by changes in the three key parameters

ResDisp, CDThresh and DispDist. We show results for two target patches (Fig 5.7). Patch A was

relatively large and surrounded by numerous other patches in a resource-abundant region in the

north-eastern study region. Patch B was much smaller and isolated, but in the vicinity of seasonally

resource-abundant coastal floodplains in the northern study region (Fig 5.7).

The total area of breeding and dispersal habitat connected to patch A increased steadily during

the wet season (median range 77,444 to 207,276 km2) and somewhat more abruptly during the dry

season (median range 2,322 to 94,269 km2) as model parameters became more permissive to

dispersal (Fig 5.7 boxplots and map panels d-f). This was partly due to increasing patch size via

intra-patch gap crossing and partly due to greater connectivity via inter-patch dispersal. In the dry

season, patch B was below delineation thresholds when ResDisp was high. It remained small and

isolated except for a sharp increase in component size (from values < 400 km2 to 40,852 km2) when

all model parameters were at their most permissive levels (ResDisp low, CDThresh loose and

DispDist 10 or 15 km). During the wet season, component size also dramatically increased (from

values < 3,000 km2 to values > 30,000 km2) under moderate levels of ResDisp and CDThresh when

DispDist was at its highest level 15 km (Fig 5.7 boxplots and map panels a-c).

Other, more isolated, habitat patches in inland areas had limited patch connectivity irrespective

of model assumptions. Spatial data files showing seasonal patch connectivity for all patches and

each factorial combination (two seasonal scenarios and three ResDisp / CDThresh / DispDist levels)

will be openly available (to be submitted to Data Dryad upon publication).

115

Fig 5.5 Seasonal patch connectivity for feral pigs in northern Australia. Median, interquartile range,

outlier and mean size of connected habitat components are shown for each ‘matrix resistance’ level.

Fig 5.6 Interactions between ‘corridor delineation threshold (CDThresh)’ and ‘inter-patch dispersal

distance (DispDist)’ at each level of ‘matrix resistance (ResDisp)’ for the dry (a-c) and wet (d-f) season.

We plot relative changes to patch connectivity, where 0 and 1 corresponded to the minimum and maximum

component size assigned to each patch. While this allowed comparing interactions across components of

different absolute size (~20-330,000 km2), it obscured the magnitude of effects.

118

Fig 5.7 Seasonal patch connectivity for two selected habitat patches of feral pigs in northern Australia.

For each ‘matrix resistance (ResDisp)’ level, we report median, interquartile range, outlier and mean size of

habitat components connected to the target patch (note different y-axis scale). Map panels a-f show the target

patch (blue) in the context of surrounding patches (colour-coded by the median size of their habitat

component) and dispersal habitat (shown for one ‘corridor delineation threshold (CDThresh)’ per panel and

colour-coded in shades of grey for different ‘inter-patch dispersal distances (DispDist)’).

5.5 Discussion

Directly transmitted animal diseases in wildlife can spread between infected and susceptible

populations via host dispersal (Anderson et al. 1986). Risk-based disease management requires

better spatial knowledge of the links (dispersal paths or ‘corridors’) between populations (Riley

2007; Cowled & Garner 2008; Rioux Paquette et al. 2014). Habitat connectivity models are well

suited to deliver such knowledge, but have not been widely applied to investigate risk of disease

spread (Kool et al. 2013). Here, we modelled habitat connectivity for feral pigs, an abundant

invasive species that is susceptible host to numerous infectious diseases, in northern Australia. Our

resource-based (Hartemink et al. 2015) modelling approach could readily be applied to other

wildlife hosts with a metapopulation structure, especially when empirical data is limited and risk is

investigated over large extents. Results highlighted seasonal variability in ‘patch connectivity’, an

indicative measure of disease spread risk. Model evaluation further suggested that wide-ranging

boars are a greater risk factor for disease spread than resource-constrained breeding herds. First, we

discuss results on ‘matrix’ and ‘patch’ connectivity, focusing on ecological interpretation and

methodological evaluation. Second, we discuss implications for risk-based disease management.

5.5.1 Matrix connectivity

We modelled matrix connectivity for two seasonal scenarios using the circuit-theoretic

approach (McRae & Beier 2007; McRae et al. 2008). This framework was suited to our study aims

as it (a) modelled movement probabilities along all possible dispersal paths (which is indicative of

disease spread risk) and (b) did not rely on strict distinctions between nodes (patches), edges

(corridors) and the matrix (which is likely not how generalist omnivores such as feral pigs perceive

and use the landscape; Moilanen 2011). The approach may be less useful for other systems, e.g.

when modelling movement along established migration routes (McClure et al. 2016).

Circuit-theoretic models can be computationally prohibitive when applied over large areas

(McRae et al. 2013). We used a recent extension (Pelletier et al. 2014) that models matrix

connectivity (as ‘omnidirectional current density’) on buffered tiles. These can join to form

arbitrarily large study areas (here 175 million pixels). Our automated implementation (Appendix

119

5.1) may be useful to future regional-scale connectivity models. We detected increased current

density at some tile corners. This artefact of tile buffering (Koen et al. 2010; Pelletier et al. 2014)

occurred mostly in low quality habitat, when resistance in the tile greatly exceeded resistance in its

randomly filled buffer. While this had minimal impact on our analyses of patch connectivity

(affected tiles contained no or few patches), it may disturb other applications. In heterogeneous

study areas, we suggest filling buffers with resistance values drawn randomly from each tile rather

than the entire study area (Koen et al. 2010) to ensure matching value ranges and eliminate border

effects.

Matrix connectivity was modelled independently of habitat patches (Koen et al. 2014), based

solely on resistance to dispersal. However, we did not have empirical data from movement or

landscape genetic studies (Kool et al. 2013) to parameterize matrix resistance. Instead, we used

converted results from a previously developed habitat suitability model (Chapter 4). While this is

not unusual (Beier et al. 2007; Zeller et al. 2012), and justified for feral pigs in Australia where

inter-patch dispersal does not appear to be limited by anything but habitat quality (Choquenot et al.

1996; Cowled & Garner 2008), it is arguably subjective. Moreover, modelled ‘omnidirectional

current density’ is a relative measure which has not yet been directly linked to ecological factors

that distinguish between dispersal corridors and non-dispersal matrix (Pelletier et al. 2014).

Our scenario approach evaluated the effects of a range of parameter estimates. We explicitly

linked these scenarios to different dispersal abilities between two distinct social entities in feral pig

populations (Choquenot et al. 1996; Gabor et al. 1999): ‘high’ resistance and a ‘strict’ corridor

delineation threshold may approximate dispersal by resource-constrained breeding herds; ‘low’

resistance and a ‘loose’ corridor delineation threshold may reflect dispersal by wide-ranging boars;

intermediate scenarios reflected empirical uncertainty and behavioural variability. Results showed

that season, matrix resistance and corridor delineation threshold all had a major effect on matrix

connectivity. This suggests that dispersal patterns in northern Australia may differ considerably

with climatic conditions and between feral pig breeding herds and solitary boars. Importantly,

matrix connectivity should be validated against movement data, or to test the assumed link to

disease spread, against epidemiological observations. However, at the time of this study such data

was unavailable or incomplete. Our results highlight that future data collection must consider

seasonal and intra-specific differences and should be stratified by broad environmental gradients to

capture variability in feral pig dispersal between resource-poor and resource-abundant landscapes.

5.5.2 Patch connectivity

We used modelled matrix connectivity to compute and evaluate connectivity between habitat

patches. Patch delineation in Core Mapper required definition of multiple thresholds for ‘habitat

120

quality’ and ‘minimum patch size’ (Shirk & McRae 2013). For feral pigs, we aimed to maintain

narrow patches (e.g. rivers in semi-arid environments) and allow for considerable gap crossing

abilities. Nevertheless, large inland tracts and all of Western Australia contained no or very few

delineated patches. This is inconsistent with previous research (West 2008; Cowled et al. 2009).

Our quality thresholds were derived from a model that was validated mostly in resource-abundant

environments in the eastern and northern parts of the study region (Chapter 4). Different quality or

size thresholds may in fact apply to habitat patches in inland environments or Western Australia.

Alternatively, previous research may have overestimated distribution in drier landscapes that, at

best, support small and isolated populations. Empirical research is needed to better understand

habitat use and movements (both intra-patch gap crossing movements and inter-patch dispersal) of

feral pigs in resource-poor regions of northern Australia.

For connecting habitat patches, we introduced a third parameter, inter-patch dispersal distance.

All three evaluated model parameters significantly affected patch aggregation into habitat

components. Interestingly, during the dry season, patch connectivity increased most as each

parameter changed to its most permissive level (‘low’ matrix resistance, ‘loose’ corridor delineation

threshold, or ‘15 km’ inter-patch dispersal distance). During the wet season, however, patches

became steadily more connected as parameters placed fewer constraints on dispersal. This suggests

that, during the dry season, patch connectivity for dispersal-constrained breeding herds of feral pigs

is low, even when allowing for uncertainty about their dispersal ability. However, unconstrained

dispersal by solitary boars in all habitat types and over long distances may dramatically increase

connectivity between habitat patches for these individuals. During the wet season, when resources

are more abundant and widely distributed, patch connectivity may still be greatest for boars, but less

different to that of breeding herds.

These network-level generalizations omit differences between patches, depending on their

location in the patch-network. We distinguished three broad types of patches: First, all modelled

scenarios contained many small isolated habitat patches. These were often located in inland areas

not connected to other patches irrespective of assumptions. Second, many small patches interacted

with surrounding patches to form larger habitat patches (via intra-patch gap crossing) or connected

habitat components (via inter-patch dispersal). These were mostly located in resource-abundant

areas and became increasingly connected to additional breeding habitat as parameters became more

permissive to dispersal (e.g. patch A in Fig 5.7). Third, some habitat patches were isolated under

most dispersal scenarios but became connected to large habitat components during the wet season

or for wide-ranging boars. These were usually located in marginal habitat adjacent to resource-

abundant regions and could act as “stepping stones” between components (e.g. patch B in Fig 5.7).

121

Our interpretations were based on measured patch connectivity in response to three-way

changes in key parameters. However, it is uncertain whether evaluated levels reflect the full

parameter space of feral pig dispersal. More empirical research is needed to improve model

parameterization (Kool et al. 2013), especially for the wet season when any difference in dispersal

ability had a significant effect on patch connectivity. Further, recasting delineated patches and

modelled matrix connectivity as a network graph with strictly defined nodes and links (Pelletier et

al. 2014) may allow further analyses: link thresholding could determine critical dispersal thresholds

at which network connectivity changes most; patch prioritisation could help to rank patches based

on their contribution to connectivity (Galpern et al. 2011).

5.5.3 Implications for disease management

Our results can directly inform risk-based management of infectious animal diseases at all

stages of invasion (Cowled & Garner 2008; Milne et al. 2008).

First, effective early detection surveillance should focus on patches with high risk of disease

occurrence (East et al. 2013). Spatial data files showing seasonal patch connectivity across northern

Australia for a range of assumptions about feral pigs’ dispersal ability will be openly available upon

publication of this chapter (to be submitted to Data Dryad). From this data, connected habitat

components can be identified. When combined with information on pathways for disease

introduction and, ideally, with estimates of potential feral pig abundance within each component,

surveillance programs could be designed so that each habitat component is adequately sampled for

disease occurrence. Our data could also be further analysed using a graph-theoretic approach to

prioritize individual patches that contribute most to connectivity within and between habitat

components for surveillance (Galpern et al. 2011). Finally, surveillance activities could be

optimized across time and space for such patch networks according to mathematical efficiency rules

(Chades et al. 2011).

Second, we modelled patch connectivity, which can be interpreted as an indicator of disease

spread risk. Deeper, dynamic insights into disease-host interactions and the spread of epidemics for

better incursion preparedness could be gained through epidemiological simulation models (Ostfeld

et al. 2005; Riley 2007; Milne et al. 2008). Our data on patch and matrix connectivity could be

integrated to realistically constrain pathways of spread or the extent of epidemiologically connected

zones (Cowled & Garner 2008; Rees et al. 2013, Macpherson et al. 2016). Epidemiological models

could also elucidate the relationship between connectivity and disease transmission or persistence in

wildlife hosts. While we implicitly assumed that connectivity increases risk, this is not necessarily

the case (Plowright et al. 2011; Huang et al. 2015).

122

Third, our results could guide management responses to disease incursions. For example, patch

connectivity data could be used to identify the type of infected patch and derive suitable

interventions. In isolated patches, disease spread could be easily contained and risk is minimal.

Patches that are mostly isolated but connected to large components under dispersal-permissive

assumptions could act as “stepping stones”. Management should aim to isolate the patch and

constrain boar and wet season dispersal, which may spread disease to susceptible nearby

components. Large patches or habitat components consisting of several highly connected patches

should be treated as contiguous management units with limited opportunities for containment. The

focus should be on efficient disease management within the infected component. Our matrix

connectivity data could help to identify transmission pathways between infected and susceptible

patches of any type, and design interventions that reduce connectivity. This could include

population control or vaccination in dispersal corridors, fencing as a dispersal barrier or to prevent

access to key resources. Importantly, effective control of connectivity corridors around an infected

habitat component may help justify declaration of disease-free compartments (Scott et al. 2006).

Functional connectivity models can yield useful insights on the spread risk of infectious

diseases with wildlife hosts and should be in the toolbox of epidemiologists, ecologists and

practitioners interested in combating this growing biosecurity threat.

5.6 Appendices

Appendix 5.1 Methodology for modelling omnidirectional current density (annotated R code).

Appendix 5.2 Methodology for computing patch connectivity. (ModelBuilder process)

Appendix 5.3 Two-way tests of significance between levels of ResDisp, CDThresh and DispDist

123

Chapter 6 General discussion and conclusions

Summary

In this chapter I relate the results from Chapters 3 to 5 to the aims of this thesis. For each of the

three research questions, I discuss the main findings of my research as well as their methodological,

ecological and applied significance11. Next, I synthesize how the adopted multi-level modelling

approach and its outputs can help to inform risk-based management of infectious animal diseases

for which feral pigs are a wildlife host. I conclude by summarizing limitations and giving

recommendations for future research that may further improve on this thesis’ findings.

6.1 Research question 1 – resource selection by individual breeding herds

How can habitat suitability for mobile species such as feral pigs be reliably modelled at the

regional scale, given uncertainty about the processes of habitat selection, in particular the

resource-seeking home range movements by individual breeding herds?

The purpose of the first research question was to lay the foundation for modelling habitat

suitability for feral pigs at the regional scale under empirical uncertainty. Like other mobile

animals, feral pigs are able to access required resources at different locations within heterogeneous

home ranges (Powell & Mitchell 2012). The suitability of a focal site to function as habitat thus

depends on landscape structure – that is, the availability and patterning of complementary and

supplementary resources (Section 2.1.3; Addicott et al. 1987; Dunning et al. 1992).

6.1.1 Main findings

In Chapter 3 I developed an integrated methodology that can be used to model habitat

suitability for mobile animals at regional extents, while also capturing resource-seeking home range

movements by individuals. I adapted a resource-based framework (Hartemink et al. 2015) using

Bayesian networks that has increasingly been applied to habitat models (Section 2.4.2). However,

Bayesian networks are inherently spatially ‘neutral’ or ‘implicit’. Few studies have accounted for

spatial interactions between variables (Section 2.4.3). Here, I developed a novel combined

methodology, spatial pattern suitability analysis, for capturing feral pigs’ selection of

supplementary resources within home ranges. This involved: (a) measuring structural patterns of

‘resource quality’ within home range-sized moving windows around each study region pixel, and

(b) using expert-elicited ‘response-to-pattern’ curves to relate these structural measures to

11 Due to the stand-alone nature of journal submissions, most of the key findings and discussions of their significance have similarly been stated in Chapters 3 to 5 (submitted or being prepared for submission). Here, I attempt to reiterate their relevance with regard to answering the research questions of this thesis.

124

functional ‘resource suitability’ indices that reflect the availability of resources in the landscape.

After applying this method to each resource required for sustained feral pig breeding (food, water,

protection from heat and protection from disturbance), I modelled habitat suitability as a function of

all complementary resource suitability indices.

For the purpose of this thesis, I investigated whether this integrated methodology improved

model accuracy in predicting habitat suitability for feral pig breeding and, if so, which aspects of

landscape structure and measurement scales are best suited to describing resource-seeking home

range movements by individual feral pig breeding herds in northern Australia. I explored this

question through comprehensive scenario analysis and model validation against three independent

distributional data sets from a subsection of the study region in northern Queensland. ‘Landscape

models’ that integrated landscape-scale ‘resource suitability indices’ based on either distance-,

composition- or combined distance / composition-dependent measures explained observed feral pig

presences better than a ‘control model’ that used site-scale ‘resource quality indices’ in 90% of

evaluated scenarios. However, there were differences between landscape models and validation data

sets. Overall, models that measured resource suitability as a function of distance (i.e. assigned

greater value to adjacent than distant resources) performed consistently best against all validation

data. Validation statistics were similar at all measurement scales (moving window radius = 1, 2 or 3

km), but most consistent at the intermediate scale (r = 2 km). Models that measured resource

suitability as a function of landscape composition or a combination of distance and composition (i.e.

the averaged quality or distance-weighted averaged quality of all resources in the home range)

displayed mixed results, ranging from very high to worse than random accuracy.

6.1.2 Significance and advances

Methodologically, the application of Bayesian network habitat models to mobile animals was

improved by accounting for resource-seeking home range movements. The Bayesian network

approach was preferred over more widely used correlative or mechanistic methods because: (a) it

could be applied at the regional extent, under empirical uncertainty and independent of

distributional data by utilizing unpublished expert knowledge; (b) it provided a robust statistical

framework for modelling habitat suitability as a function of complementary resources; and (c)

uncertainty in expert-elicited parameter estimates was explicitly accounted for (Uusitalo 2007;

Wilhere 2012; Landuyt et al. 2013). Spatial pattern suitability analysis essentially converted site-

scale into landscape-scale variables that could be used to model habitat suitability consistent with

landscape ecological principles. Compared to previous statistical or expert-based approaches for

integrating ‘landscape variables’ into habitat models (Section 2.4.3), the methodology offered three

main advantages: First, landscape-scale resource suitability indices were computed from response-

125

to-pattern curves, which provided ecological meaning to structural metrics and enhanced

transferability across large study areas. Second, response-to-pattern curves were elicited from

experts using a structured method that was easily implemented and understood by expert

practitioners and may be useful to future studies that are unable to utilize or generate empirical data.

Third, landscape variables were derived from modelled site-scale indices of resource quality that

reflected the species-specific value of supplementary resource items (Dunning et al. 1992).

Through comprehensive scenario analysis and validation, the methodologies’ application to

feral pigs also yielded ecologically significant results. The finding that habitat suitability for feral

pig breeding may be reliably described by measuring the distance to a suite of supplementary and

complementary resource patches confirms previous empirical results (Caley 1993; Choquenot &

Ruscoe 2003). However, it has also been reported that feral pig densities respond to the amount of

available food (Caley 1993). In this study, observed feral pig presences in some validation data sets

were also accurately explained by models based on composition-dependent landscape variables.

Yet, results were inconsistent and scale-dependent. Appropriate measurement scales for describing

the mobility limits of feral pig breeding herds in their regular resource-seeking activities were

elicited from experts at either 1 km (n = 1), 2 km (n = 2) or 3 km (n = 3). While evaluation did not

indicate selection of an ‘optimal’ scale, these a priori assumptions (Miguet et al. 2016) are broadly

consistent with the range of empirical home range estimates from Australia (Section 2.2.3). Overall,

findings may not allow conclusive inferences on the mechanisms of resource selection by feral pig

breeding herds. Yet, they do suggest that simple measures of distance are less vulnerable to site-

specific differences in behaviour or data and can be reliably used for integrating resource-seeking

home range movements into regional-scale habitat suitability models.

6.2 Research question 2 – seasonal habitat patches for subpopulations

How are patches of suitable feral pig breeding habitat that potentially support locally dense

populations distributed in northern Australia, and is their spatial distribution influenced by

seasonal effects?

The purpose of the second research question was to fill an existing knowledge gap about the

spatial distribution of feral pigs in northern Australia at spatial and temporal scales (resolution and

extent) that are useful for on-ground applications (Section 2.2.1), in particular the risk-based

management of infectious animal diseases. There are likely seasonal and gender-specific differences

in distribution and habitat use (Caley 1997; Dexter 1999; Mitchell et al. 2009). The research

question focused on those habitat patches that support persistently breeding local subpopulations.

126

6.2.1 Main findings

In Chapter 4 I applied the integrated methodology from Chapter 3 to model and map, for the

first time, seasonal habitat suitability for feral pig breeding and persistence in northern Australia. I

evaluated model sensitivity to habitat variables as well as expert opinion and validated the

seasonally-specific predictions from six individual expert models as well as an expert-averaged

‘consensus’ model against four independent distributional data sets per seasonal scenario. Finally, I

compared the mapped wet and dry season predictions from the consensus model, which produced

consistently accurate results across the study region, at different spatial scales.

Habitat suitability for persistent feral pig breeding was modelled as a function of four

complementary resource requirements – food, water, protection from heat, and protection from

disturbance – each measured as a distance-dependent landscape variable. The consensus model was

least sensitive to expert opinion. All experts parameterized the model so that water and food

resources were more important than heat or disturbance refuges. Overall, model predictions were

robust to validation data. However, there were some differences between expert models, seasonal

scenarios and validation data sets. Model uncertainty was largely attributed to different expert

assumptions about feral pigs’ resource-seeking home range movements as described in elicited

response-to-pattern curves. By averaging across experts, I was able to reduce uncertainty and

produce consistently accurate results for all validation data. As validation data sets were biased

towards the eastern portion of the study region and environments with high proportions of suitable

feral pig habitat, I caution that model accuracy in resource-poor inland environments and in the

western portion of the study region needs further evaluation.

Predicted habitat suitability from the expert-averaged model fluctuated considerably between

seasons. Overall, models revealed a four-fold increase in suitable breeding habitat during the wet

season (36% of the study region), when resources are abundant and widely distributed, compared to

the dry season (9.5% of the study region), when resources are scarce and scattered. Of this, about

one quarter was considered highly or very highly suitable habitat. At the regional scale, patches of

suitable breeding habitat were concentrated in the eastern study region, where they also contracted

less dramatically during the dry season than in the western study region. At the landscape scale,

habitat patches were located predominantly along the coastline and interior riverine and wetland

systems during the dry season and expanded widely across the study region during the wet season. I

suggest that not all suitable wet season habitat may be realized by feral pig subpopulations due to

dispersal constraints. At the local scale, patches of suitable habitat were highly concentrated in

resource-abundant rainforest, wetland and agricultural refuge areas during the dry season. During

the wet season, habitat patches were more evenly distributed among vegetation types and land use

127

classes. This suggests that feral pigs may forage widely into grassland floodplains, savanna

woodlands and coastal shrublands when conditions permit.


Methodologically, I demonstrated that the integrated modelling approach using Bayesian

networks and spatial pattern suitability analysis can be applied to accurately model habitat

suitability for feral pigs at the regional scale. The resource-based approach (Hartemink et al. 2015)

could readily be adapted to the study of other wide-ranging mobile species, especially in data-poor

situations. Previous habitat models using Bayesian networks have usually been applied to overcome

empirical uncertainty and lack of data. Consequently, models have frequently utilized expert

knowledge, but rarely validated results against independent data (Marcot et al. 2001; Smith et al.

2007; Murray et al. 2012; Tantipisanuh et al. 2014; van Klinken et al. 2015). The present study

developed and implemented a range of innovative techniques for the structured and repeatable

elicitation of expert knowledge (Section 2.6), including: (a) combining the CPT calculator (Cain

2001) and interval judgement (Speirs-Bridge et al. 2010; McBride et al. 2012a) approaches to elicit

conditional probabilities; and (b) using elicited response-to-pattern curves to derive ecologically

meaningful landscape variables. Nevertheless, some ecologists continue to question the value of

expert models. To my knowledge, this study is the first to apply the Continuous Boyce Index

method (Boyce et al. 2002, Hirzel et al. 2006) to empirically validate spatial predictions from a

Bayesian network model against truly independent species observations. Further, I evaluated the

performance of a range of models based on the opinions of individual experts as well as an average

of all expert opinions. Validation added considerably to the expert-based modelling approach,

allowing an evaluation of model uncertainty and providing confidence in the findings.

This study was ecologically significant as it provided the first seasonally-specific estimate of

feral pig breeding habitat in northern Australia. Although seasonal variability in distribution,

density and habitat use has been widely documented in site- or landscape-scale studies (Mitchell et

al. 1982; Hone 1990a; Caley 1993; Choquenot et al. 1996), validated model results confirmed these

findings at the regional scale. Moreover, spatial analyses of mapped results highlighted specific

vegetation types and land use classes that contribute most to seasonal shifts in suitable habitat

patches. Importantly, the model suggested that suitable feral pig habitat is less widely available in

northern Australia than previously thought (e.g. Mitchell et al. 1982; West 2008; Cowled et al.

2009). I suggest that previous overestimates could be confidently corrected because this study (a)

accounted for ecological constraints arising from the dry season and the resource requirements for

breeding herds, and (b) improved model resolution to avoid upscaling error.

128

Finally, the study’s findings have applied significance. The model was specifically designed to

be useful for feral pig management in northern Australia, featuring a regional analysis extent, a

sufficiently fine resolution to capture local variability in conditions, seasonally-specific predictions,

and comprehensive accuracy assessment. Mapped results may thus be confidently used to quantify

impacts, assess risks, justify management investments and target control activities. Further, because

habitat suitability was explicitly referenced to local breeding populations, it is a useful indicator of

the seasonal establishment risk for infectious animal diseases within feral pigs, which often depend

on a persistent supply of young susceptible animals. Results could help to target disease

management activities at high-risk areas (Section 6.4.2).

6.3 Research question 3 – seasonal patch connectivity for metapopulations

How are patches of suitable feral pig breeding habitat connected by potential dispersal

pathways to form a regionally connected metapopulation, and is their connectivity influenced

by seasonal or gender-specific effects?

The purpose of the third research question was to fill a second knowledge gap in our

understanding of feral pig populations in northern Australia, namely the dispersal between local

feral pig subpopulations within larger metapopulations. While population structuring in feral pigs

has been extensively documented, the emergence of population patterns from dispersal processes

and landscape heterogeneity in between patches of suitable breeding habitat has received little

attention (Section 2.2.4). In particular, there are likely seasonal and gender-specific differences in

inter-patch dispersal (Choquenot et al.1996; Gabor et al. 1999; Cowled et al. 2008; Hone 2012).

6.3.1 Main findings

In Chapter 5 I modelled and mapped, for the first time, seasonal habitat connectivity for feral

pigs in northern Australia using a functional, resource-based modelling approach (Hartemink et al.

2015). I applied the empirically validated habitat suitability indices from Chapter 4 to delineate

contiguous patches of feral pig breeding habitat (by identifying suitability thresholds; Hirzel et al.

2006), and to parameterize matrix resistance to inter-patch dispersal, i.e. the cost of traversing

habitat of different quality (by reversing suitability indices; Zeller et al. 2012). Then, I modelled

seasonal ‘matrix connectivity’ using a circuit-theoretic approach, which identified relative

movement probabilities along all possible dispersal paths in the entire landscape (McRae & Beier

2007; McRae et al. 2008). Finally, I computed seasonal ‘patch connectivity’ as the total size of the

habitat component (including breeding and dispersal habitat) that each patch is connected to

(Rayfield et al. 2011). To account for gender-specific differences in dispersal ability as well as

empirical uncertainty in parameter estimates, a range of model assumptions relating to ‘matrix

129

resistance’, ‘dispersal distance’ and a ‘corridor delineation threshold’ (which arbitrarily described

the boundary between dispersal paths and non-dispersal matrix) were explored.

Both matrix and patch connectivity were significantly affected by season as well as model

assumptions. Large tracts of additional dispersal habitat became available in inland areas during the

wet season, when resources were widely distributed across the study region. The assumed level of

matrix resistance mostly affected the width of dispersal paths, but not their spatial distribution

across the region. Patch connectivity was analysed for each delineated habitat patch and in the patch

network by aggregating patch-level values across the study region. At the network level, the median

size of connected habitat components increased from 341-411 km2 in the dry season to 2,697-6,932

km2 in the wet season. Within each seasonal scenario, there were interesting differences in the

effect of model assumptions. During the dry season, patch connectivity increased significantly only

under the most permissive dispersal parameters. During the wet season, however, any change in

model assumptions affected patch connectivity. Overall, there were many small (isolated), some

intermediate and few large (well-connected) habitat components. Although habitat patches were

delineated with the aim of maintaining small, narrow patches (e.g. inland rivers), large tracts of the

interior study region and all of Western Australia contained no or very few delineated habitat

patches in both seasonal scenarios. At the patch level, I identified three broad types of habitat

patches that were affected similarly by model assumptions, depending on their location in the patch-

network: some were always isolated; some were always connected to very large habitat

components; and some were mostly isolated but became connected to large habitat components

only during the wet season or under the most permissive parameter estimates.


Methodologically, matrix connectivity was modelled using a recently developed extension of

the circuit-theoretic approach that computes “omnidirectional current density” on buffered tiles

(McRae et al. 2008; Pelletier et al. 2014). This methodology overcomes computational limitations

when applied over large areas (McRae et al. 2013). The automated implementation in this study

(provided as Appendix 5.1) may be useful to future regional-scale connectivity models.

Ecologically significant inferences could be made by linking model assumptions to the gender-

specific differences in dispersal ability between two distinct social entities in feral pig populations

(Choquenot et al.1996; Gabor et al. 1999). I assumed that ‘high’ matrix resistance, a ‘strict’

corridor delineation threshold and a ‘low’ maximum dispersal distance may approximate dispersal

by resource-constrained breeding herds; ‘low’ matrix resistance, a ‘loose’ corridor delineation

threshold and a ‘high’ maximum dispersal distance may reflect dispersal by wide-ranging boars;

intermediate scenarios reflected empirical uncertainty and behavioural variability. Given this, the

130

study’s findings suggest that many habitat patches are disconnected for dispersal-constrained

breeding herds during the dry season, even when allowing for uncertainty. However, unconstrained

dispersal by solitary boars may dramatically increase connectivity between patches for these

individuals. During the wet season, connectivity between patches may still be greatest for boars, but

less different to that of breeding herds. Lastly, the paucity of habitat patches in the interior and

western study region is inconsistent with previous research (West 2008; Cowled et al. 2009). This

may reflect errors in patch delineation in this thesis or previous overestimations regarding the

distribution of feral pigs in resource-poor environments that, at best, support small and isolated

populations.

This study is also of applied significance. To my knowledge, it is one of the first to investigate

functional connectivity for a widely distributed invasive species in Australia. Yet, biosecurity and

invasive species management could benefit greatly from such analyses and aim to reduce habitat

connectivity for harmful species (Glen et al. 2013). In this thesis, seasonal patch connectivity for

breeding herds and solitary boars could be interpreted as an indicator of disease spread risk between

infected and susceptible subpopulations (Kramer-Schadt et al. 2007; Cowled & Garner 2008).

Results can inform risk-based management of infectious animal diseases at all stages of invasion,

including early detection surveillance, preparedness and incursion response (Section 6.4.2).

6.4 Synthesis and applications

6.4.1 Contributions to the thesis aim

In the following, I synthesize how the research approach, conceptual framework and presented

findings contributed to achieving the thesis aim: to generate spatially-explicit, seasonally-specific

and regional-scale knowledge of feral pig populations in northern Australia to inform risk-based

management of directly transmitted infectious animal diseases with feral pigs as a wildlife host.

6.4.1.1 Synthesis of the research approach

In summary, this thesis generated new spatial knowledge of feral pig populations in northern

Australia at multiple levels of biological organisation and associated scales (resolutions and extents)

of analysis. The multi-level research approach was specifically designed to be useful for post-border

management of directly transmitted infectious animal diseases with feral pigs as a wildlife host.

(A) Analyses were conducted at the regional extent, covering all of tropical Australia north of

the Tropic of Capricorn. This study region was selected because it is vulnerable to disease

introduction from neighbouring countries and contains widespread feral pig populations.

Consequently, there is a need for effective post-border disease management in feral pigs, yet

131

spatially-explicit knowledge with sufficient detail to inform risk-based management has so far been

limited (Choquenot et al. 1996; Australian National Audit Office 2012).

(B) I focused on improving three types of detail. First, analyses were conducted at a

sufficiently fine spatial resolution (1 ha pixel size) to capture local variability and enable on-ground

applications. Second, analyses introduced a temporal resolution, which has often been missing from

regional-scale studies (West 2008; Cowled et al. 2009). For northern Australia, a scenario approach

encompassing the dominant climatic wet / dry season cycle was deemed adequate, while inter-

annual climatic variability was not considered. Third, analyses contained an ecological resolution

based on feral pigs’ social organisation. Gender-specific differences in feral pig habitat use, home

range movements and dispersal ability were explicitly considered at all levels of analysis

(Choquenot et al. 1996; Gabor et al. 1999; Mitchell 2008).

(C) Due to empirical knowledge gaps at the regional extent, I adopted a resource-based

approach to modelling habitat suitability and connectivity based largely on expert knowledge.

Although such an approach is arguably better suited to making regional inferences than

extrapolations from statistical models (Hartemink et al. 2015), uncertainty is inherent to expert

models. Thus, I attempted to rigorously evaluate model accuracy and assumptions where possible.

In conclusion, efforts were made to balance generality against detail, and maintain

transparency for decision-makers who may ultimately use results to inform management. Specific

contributions to the field of habitat modelling include: extending previous modelling approaches

using Bayesian networks and circuit theory in scope and application; and developing a novel

methodology for integrating home range movements into resource-based habitat models, which may

be usefully applied to other mobile animals in data-poor situations.

6.4.1.2 Synthesis of the conceptual framework

The multi-level perspective on feral pig populations was explicitly linked to a multi-level

perspective on infectious diseases. I conceptualized feral pigs in northern Australia as a

metapopulation and the landscape as displaying a patch-corridor-matrix structure (Forman 1995;

Hanski 1998). Individual breeding herds select supplementary and complementary resources within

home ranges (Dunning et al. 1992). Several herds with adjacent or overlapping home ranges

interact frequently to form contiguous, persistent subpopulations within patches of suitable breeding

habitat. Certain areas in the matrix function as dispersal corridors between patches. However,

regional connectivity between habitat patches varies by gender, with resource-constrained female

breeders having a lower dispersal ability than often solitary males (Choquenot et al. 1996; Gabor et

al. 1999). Movements and distributional patterns at each level may fluctuate seasonally. This

132

structure in the feral pig host population also affects the risk of invasion and persistence by directly

transmitted diseases such as classical swine fever: following introduction, successful disease

establishment is contingent on a locally dense host subpopulation; subsequent disease spread

requires host dispersal between infected and susceptible subpopulations; long-term disease

persistence depends on a persistent supply of susceptible individuals within a regionally connected

host metapopulation (Anderson et al. 1986; Kramer-Schadt et al. 2007; Riley 2007).

Three research questions were formulated to capitalize on these conceptual links by

investigating disease risk as an attribute of the wildlife host. The purpose of research question 1 was

to lay the foundation for modelling subpopulations in habitat patches by understanding how

individual breeding herds select resources within home ranges and what site-level conditions

determine habitat suitability. Research question 2 aimed to generate knowledge about the seasonal

distribution of persistent subpopulations within patches of suitable breeding habitat. Landscape-

level habitat suitability for breeding served as an indicator of disease establishment risk. The

purpose of research question 3 was to generate knowledge about the seasonal connectivity between

local breeding populations. Regional-level patch connectivity served as an indicator of disease

spread risk. Similar combinations of habitat suitability and connectivity models have been applied

to gain a more complete understanding of mobile species’ use of entire landscapes (Cianfrani et al.

2013; Dickson et al. 2013; Stewart-Koster et al. 2015). Resource-based habitat approaches such as

the one adopted in this thesis have recently been recommended for application to the holistic study

of pathogen-host interactions (Hartemink et al. 2015).

6.4.1.3 Synthesis of findings

This thesis showed that habitat suitability for persistent feral pig breeding is dependent on

spatial interactions between four key habitat requirements: water and food resources as well as

protection from heat and from disturbance. Individual breeding herds can access supplementary and

complementary resources within home ranges. Here I showed that habitat suitability of a focal site

is most reliably modelled as a function of distance to each of these complementary requirements.

The availability and quality of resources in northern Australia varies seasonally according to

climatic conditions. This was also reflected in model results. The spatial distribution of feral pig

habitat patches, which support locally dense populations and pose the highest risk of disease

establishment, varied markedly between seasons. Disease establishment risk may be locally highest

during the dry season, when breeding herds must congregate in resource-abundant refuge sites.

Habitat suitability analyses further showed that during the wet season, breeding subpopulations are

able to expand widely. These shifts in patch size and distribution may lead to less locally intense

(due to lower local densities), but more widely distributed (due to increased contact within merged

133

patches) risks of disease establishment. Importantly, findings suggest that dry season conditions

restrict habitat suitability for feral pig breeding much more than previously thought, with many

areas in the region’s interior and western portions containing no or only a few small habitat patches

under that scenario. Due to dispersal constraints, breeding herds may be unable to realize much of

the potentially suitable wet season habitat. Large tracts of the study region may thus present a

minimal risk to disease establishment.

Patch size and distribution also affect the connectivity between local subpopulations, and hence

the risk of disease spread across the region. Connectivity analyses showed that habitat patches were

smaller and less connected during the dry season, especially for breeding herds. Under this scenario,

less dispersal-constrained solitary boars are likely the main carrier of disease between infected and

susceptible local subpopulations. During the wet season, however, many isolated dry season patches

became connected to form larger habitat components. As patch connectivity was less constrained by

dispersal ability, both breeding herds and boars may contribute to disease spread, which greatly

increases overall risk. Nevertheless, many habitat patches remained isolated regardless of season or

gender-specific dispersal. Should a disease be introduced and establish in these local

subpopulations, there is a low risk of subsequent disease spread within the larger metapopulation.

In conclusion, the research contained in this thesis provides, for the first time, a complete and

coherent, spatially-explicit, seasonally-specific and regional-scale picture of areas most at risk of

disease establishment (via host habitat suitability) and spread (via host habitat connectivity) in feral

pigs in northern Australia. The resource-based modelling approach is transparent and flexible, and

could be applied to other invasive species and wildlife hosts of infectious animal diseases,

especially in data-constrained situations and for wide-ranging species.

6.4.2 Applications to risk-based disease management

In the following, I synthesize how these findings could be applied to inform the risk-based

management of directly transmitted infectious animal diseases with feral pigs as a wildlife host. I

use the specific example of classical swine fever to illustrate recommendations. Spatial data files

showing seasonal habitat suitability and patch connectivity to support these applications will be

openly available to decision-makers upon publication of the manuscripts contained in this thesis.

6.4.2.1 Early detection surveillance

Early detection surveillance for classical swine fever in feral pigs in northern Australia is

administered by the Northern Australia Quarantine Strategy (Australian National Audit Office

2012). The program’s risk-based surveillance framework targets broad areas that are deemed to

have a high likelihood of disease occurrence based on estimated risks of introduction, establishment

134

and spread (Cookson et al. 2012; East et al. 2013). Within these areas, surveys are conducted and

feral pigs representatively sampled for classical swine fever prevalence (Cookson pers. comm.).

Our findings could be incorporated into this framework as improved spatially-explicit parameters

describing the risks of disease establishment and spread in feral pigs. Firstly, efficient surveillance

could be more narrowly targeted at seasonally delineated habitat patches instead of broad risk areas,

as these patches have the highest likelihood of disease establishment and detection of an infected

individual. Secondly, likelihood of disease persistence as a function of establishment and spread is

highest in patches that are connected to larger habitat components throughout the year. Such

patches could be prioritised for sampling. Thirdly, when combined with information on the risks of

disease introduction (e.g. ship arrivals from infected zones overseas) and, ideally, with estimates of

feral pig abundance within each habitat component, surveillance programs could be designed so that

each connected metapopulation is adequately sampled for disease occurrence.

6.4.2.2 Incursion preparedness

Incursion preparedness aims at increasing our understanding of the course and impacts of a

disease incursion prior to its occurrence. Given that classical swine fever is currently not present in

Australia, the research contained in this thesis contributes to incursion preparedness by highlighting

those areas most at risk of disease establishment (via host availability) and spread (via host

dispersal) across northern Australia. Here, disease risk was investigated purely as an attribute of the

wildlife host. This view is justified for directly transmitted diseases such as classical swine fever

and allows generalizations to other diseases with feral pigs as a wildlife host. However, it omits

other important factors of disease risk. Epidemiological simulation models could combine host risk

with disease-specific introduction pathways, transmission rates and other factors to generate deeper,

dynamic insights into disease-host interactions and the spread of epidemics for better incursion

preparedness (Section 6.5.3.2; Ostfeld et al. 2005; Riley 2007; Milne et al. 2008). Epidemiological

models could also elucidate the relationship between connectivity and disease transmission or

persistence in wildlife hosts. While it was implicitly assumed in this thesis that connectivity

increases risk, this is not necessarily the case. In some circumstances increased connectivity could

actually reduce the risk of epidemics by encouraging host diversity and immunity or ‘migratory

escape’ (Plowright et al. 2011; Huang et al. 2015)

6.4.2.3 Incursion response

Response to an incursion of classical swine fever in Northern Australia is regulated by the

Emergency Animal Disease Response Agreement (2012), the AUSVETPLAN: Disease strategy

classical swine fever (Animal Health Australia 2012) and the Wild Animal Response Strategy

135

(Animal Health Australia 2011). These guidelines prescribe a range of measures (Animal Health

Australia 2011 & 2012) that can be informed by the findings in this thesis:

(A) Collecting information about feral pig host populations: This thesis provides spatially-

explicit information about feral pig populations in northern Australia at multiple biological levels

and scales of analysis. Data generation has been designed to be useful for disease management

(regional extent, fine spatial resolution, seasonal scenarios, ecological basis and accuracy

assessment) and data will be openly available upon publication of manuscripts.

(B) Carrying out disease surveillance: (see Section 6.4.2.1)

(C) Preventing spread and containing the disease: Patch connectivity could be used to identify

the type of infected patch and derive suitable interventions. In patches that are always isolated,

disease spread could be easily contained and risk is minimal. Patches that are mostly isolated but

connected to large habitat components under assumptions that place few constraints on dispersal

could act as “stepping stones”. Management should aim to isolate the patch and constrain boar and

wet season dispersal, which may enable disease spread into susceptible habitat components. Large

patches or habitat components consisting of several highly connected patches should be treated as

contiguous management units with limited opportunities for containment.

(D) Controlling susceptible populations to enable disease fadeout via culling or vaccination: If

infected individuals are found within large or well-connected patches, the focus should be on

efficient disease management within the habitat component. Further, matrix connectivity could help

to identify all possible pathways of spread between infected and susceptible patches of any type and

design interventions that reduce connectivity. This could include population control or vaccination

in dispersal corridors, and fencing as a direct dispersal barrier or to prevent access to key resources

required for dispersal.

(E) Demonstrating freedom from disease: Contiguous patches of suitable feral pig breeding

habitat should be regarded as single epidemiological units (Cowled & Garner 2008). Patch

connectivity could help to identify subpopulations with potentially different disease status and

design sampling strategies that demonstrate this. Further, effective control of connectivity corridors

around an infected habitat component may help justify declaration of disease-free compartments

(Scott et al. 2006).

136

6.5 Limitations and future research

6.5.1 Empirical research

Perhaps most importantly, the findings presented in this thesis were derived from resource-

based models that were applied to a large study region (Hartemink et al. 2015). Due to empirical

knowledge gaps about feral pigs’ resource use within habitat patches and dispersal between patches,

models were heavily reliant on expert-elicited knowledge. This was ultimately deemed most

appropriate for the aims of this thesis. While I attempted to rigorously evaluate model results

(Chapters 3 and 4) and assumptions (Chapters 3 and 5), uncertainty is inherent to any expert model.

Further empirical research is needed to test findings and reduce uncertainty.

6.5.1.1 Resource selection by individual breeding herds

While the methodology developed in Chapter 3, spatial pattern suitability analysis, usefully

improved the application of expert-based Bayesian networks to mobile animals, knowledge about

resource selection by individuals in home ranges is more robustly derived from empirical ‘multi-

scale’ studies (Jackson & Fahrig 2015; McGarigal et al. 2016). For example, here I evaluated feral

pigs’ response to three simple landscape variables (distance-weighted resource quality, averaged

resource quality or distance-weighted averaged resource quality) that were relatively easy to relate

to experts’ field knowledge. Yet, habitat suitability for feral pig breeding may depend on other

aspects of landscape structure such as fragmentation (Jackson & Fahrig 2016), edge proximity

(Dijak & Rittenhouse 2009) or heterogeneity (McClure et al. 2015). Secondly, an ‘optimal’ scale of

measurement for these landscape variables could not be identified. This may have been due to an

insufficient range of elicited scales (experts defined mobility thresholds of 1km, 2km or 3km) or

due to the fact that most models measured selection of all resources (water, food, heat refuge and

disturbance refuge) at the same scale (Jackson & Fahrig 2015; McGarigal et al. 2016). In fact,

individuals may select different resources at different scales (Wiens 1989; Miguet et al. 2016).

Finally, I was only able to determine the ‘most reliable’ landscape variable for describing resource-

seeking home range movements by feral pig breeding herds via validation against multiple data

sets. Without empirical validation data, the accuracy of expert estimates remains unclear. A

dedicated empirical multi-scale study could compare a response variable (for example the presence

of tagged female pigs) to a suite of landscape variables measuring different aspects of landscape

composition and configuration at scales differing in orders of magnitude. Empirical data could also

be collected from different types of environments (e.g. resource-poor vs. resource-abundant) and

across seasons and years, so that a differentiated, ‘scale-optimized’ understanding of feral pigs’

habitat use can be gained (McGarigal et al. 2016).

137

6.5.1.2 Seasonal habitat patches for subpopulations

Model-predicted seasonal habitat suitability for feral pig breeding was validated against

observed feral pig presences. Although efforts were made to match the data to model assumptions,

the relationship between model output (potential habitat suitability) and response variable (actual

herd presence) is, at best, indirect (Section 2.1.1). Furthermore, validation data were collected by

third parties, often for management purposes rather than science and in a non-systematic fashion.

Hence, there was considerable uncertainty about potential error (e.g. GPS recording or timing error,

limited ability to detect in closed vegetation, movement and scattering upon aerial detection) and

bias (e.g. ‘sample selection’ or ‘reporting’ bias (Phillips et al. 2009) because highly suitable,

densely populated sites were overrepresented in targeted surveys) in the data. These issues almost

certainly impacted validation results. Finally, all data were collected in environments with an

above-average proportion of suitable feral pig habitat. Model accuracy in resource-poor inland

environments needs further investigation. Systematically collected presence / absence data from a

range of environments could actively seek to minimize sampling bias and increase confidence in

model accuracy. Even more beneficially, data that are more directly related to habitat suitability

(e.g. home range movement trajectories of individual breeding herds acquired from telemetry

studies) would greatly enhance the quality of validation.

A related issue is the accuracy of spatial proxies linked to model explanatory variables. Here,

potential problems may pertain to mapping, resampling (all proxies required a common model

resolution), averaging (proxies for seasonally-specific model variables reflected average conditions

over a five-year time period and may have failed to capture resources which were actually available

at the time of a given detection in the validation data) or reclassification (proxies may have

inadequately reflected the defined states of model variables, Appendix 4.2) error (Hunsacker et al.

2001; Boyce et al. 2002; van Klinken et al. 2015). None of these issues could be quantified in this

study due to lacking information. I suggest that field-validating the mapping accuracy of resampled

spatial proxies and whether they match the states of model explanatory would be highly beneficial.

6.5.1.3 Seasonal patch connectivity for metapopulations:

No empirical data was available to inform the habitat connectivity model (Chapter 5). I used

the expert-based, but empirically validated habitat suitability model (Chapter 4) to parameterize

matrix resistance and delineate habitat patches. While somewhat arbitrary procedures are not

unusual in connectivity models due to data constraints (Beier et al. 2007; Girvetz & Greco 2007;

Moilanen 2011; Zeller et al. 2012), they are arguably subjective and non-desirable. Further, I

evaluated the effect of a range of model scenarios to address uncertainty in parameterization, but (a)

whether these assumptions reflect the full parameter space of feral pig dispersal remained unknown

138

and (b) model results were not validated. Kool et al. (2013) suggested a variety of techniques for

collecting empirical connectivity data that can be used for model parameterization or validation. For

the purpose of estimating disease spread, telemetry studies focusing on regular or seasonal inter-

patch dispersal movements would be most useful. The results in this thesis highlight that movement

studies must consider seasonal and intra-specific variability and be stratified by resource-poor and -

abundant environments. Finally, the assumed link between patch connectivity and disease spread

must be specifically tested against spatially-explicit epidemiological data collected during future

incursions.

6.5.2 Improved habitat suitability and connectivity models

Empirical research could yield improved data for both model parameterization and evaluation.

This could ultimately also allow adoption of more sophisticated modelling and validation methods.

6.5.2.1 Statistical habitat suitability models

The modelling approach in this thesis was reliant on expert-elicited knowledge. I made every

effort to maintain a structured approach to elicitation, carefully documented methods used (in

Chapters 3 and 4) and results obtained to ensure reproducibility (Bayesian network CPTs in in

Appendix 4.1 and response-to-pattern curves in Appendix 3.1) and evaluated model uncertainty (in

Chapters 3, 4 and 5). Nevertheless, due to the inherent uncertainty and biases in expert-based

models, many ecologists remain sceptical of their value and prefer repeatable statistical analyses

(Johnson et al. 2012b). Many techniques are available to model species’ habitat and distributions

statistically (Section 2.4.1). These ‘species distribution models’, and their own pitfalls, have been

discussed at length elsewhere (Guisan & Thuiller 2005; Austin 2007; Elith & Leathwick 2009; Elith

et al. 2010 McGarigal et al. 2016; Yackulic & Ginsberg 2016). Given a suitable data set on the

response variable (e.g. feral pig presence / absence), and inclusion of ecologically meaningful

‘scale-optimized’ landscape variables (Section 2.4.3), I suggest that statistical multi-scale models

could also be applied to feral pigs in northern Australia and yield powerful inferences. However,

these may pertain to actual distributions (as evident in the data) rather than potential (resource-

based) habitat quality, and must be interpreted as such (Section 2.1.1).

I used the Continuous Boyce Index to evaluate model performance (Boyce et al. 2002; Hirzel et

al. 2006). This method was well suited to the presence-only validation data that I was able to source

for this study. However, it provides no information on model specificity, that is, its ability to

correctly predict absences and minimize false positives (Hirzel et al. 2006; Jiménez-Valverde et al.

2011). Presence-only validation also necessitates definition of a ‘background’. In the CBI method,

these serve to evaluate model deviation from a random prediction (e.g. if 20% of the background

139

contains highly suitable habitat, a useful model should find a higher proportion of actual presences

in predicted highly suitable habitat). However, background definition may influence validation

results. Generally, “too small an area can produce spurious models while […] too large of an area

can lead to artificially inflated test statistics” (VanDerWal et al. 2009, p. 592). A systematically

collected presence / absence data set could help to overcome these limitations and pitfalls of

presence-only validation techniques and facilitate a confusion matrix-based validation of model

sensitivity as well as specificity and is suggested as a future research priority.

6.5.2.2 Improved Bayesian network / spatial analysis integration

A further area of improvement relates to the integration of spatial pattern suitability analysis

within the probabilistic Bayesian network modelling framework proposed in Chapter 3. Firstly, a

major theoretical advantage of the Bayesian network approach to habitat modelling lies in the

explicit representation and propagation of uncertainty throughout a model (Wilhere 2012). Here,

resource quality at a given site was modelled in Bayesian networks and therefore expressed as a

conditional probability distribution. However, spatial pattern suitability analysis required

deterministic site-scale input values (resource quality indices, computed as model expected value)

and converted them into equally deterministic landscape-scale output values (resource suitability

indices). Model uncertainty was then re-introduced when habitat suitability was modelled in another

Bayesian network as a function of all landscape-scale indices. Future research should work towards

an uninterrupted workflow and propagation of uncertainty throughout an integrated model (Wilhere

2012).

6.5.2.3 Improved habitat connectivity models

This study used the circuit-theoretic approach to connectivity modelling, which is well suited

to identifying multiple, rather than optimal, dispersal paths throughout the entire landscape (McRae

& Beier 2007; McRae et al. 2008; Pelletier et al. 2014). Yet, the model output, ‘omnidirectional

current density’, is not easily interpreted based on an underlying ecological mechanism (Pelletier et

al. 2014). Other methods for modelling connectivity (for example graph-theoretic network graphs;

Section 2.5.1) may provide a more robust platform for further analyses. Most importantly, model

validation is critical (Kool et al. 2013; Fletcher et al. 2016a), but relies on empirical knowledge that

was unavailable or incomplete at the time of this study. Procedures for patch delineation in this

study must be better supported by data about home range movements (e.g. intra-patch gap crossing

abilities), especially in resource-poor environments that contained no modelled patches. Matrix

connectivity must be validated against observed inter-patch dispersal paths. However, a likely

mismatch between predictions (potential connectivity) and empirical data (actual dispersal) applies

to connectivity models as it does to habitat suitability models (Section 2.1.5).

140

6.5.3 Spatiotemporal dynamics

6.5.3.1 Host dynamics

The models developed in this thesis generated new spatially-explicit ‘snapshot’ knowledge

about seasonal habitat suitability for feral pig breeding and habitat connectivity between patches of

breeding habitat. In line with the resource-based habitat definition in Section 2.1.1, modelled

indices can be used to infer potential distributions and dispersal and associated risks of disease

establishment and spread. Yet, these static indices, although seasonally-specific, may not always

correspond well to actual occupancy, and less likely to local abundance or population viability

(Pulliam 2000; Larson et al. 2009; Stephens et al. 2015). For example, the models in Chapter 4

showed a four-fold increase in suitable habitat during the wet season. Yet, how much of this can

actually be realized by feral pig breeding herds dispersing from dry season refuges remains unclear.

Similarly, the connectivity models in Chapter 5 showed seasonal and gender-specific differences.

Yet, actual dispersal between habitat patches may depend on the occupancy of each patch at a given

time, the directionality between source and sink habitats, and the density-dependent population

pressure in source habitats (Pulliam 2000; Calabrese & Fagan 2004; Fletcher et al. 2016a).

Ecological simulation models that integrate density-dependent, intra-specific or metapopulation

dynamics are needed to provide deeper insights on local abundance, population growth, dispersal,

and metapopulation persistence (Hastings 2014; O’Reilly-Nugent et al. 2016). While many

techniques are available for different purposes (Section 2.4.1), dynamic models are necessarily

complex and require a large amount of empirical data for parameterization (Gallien et al. 2010,

Franklin 2010, Kool et al. 2013). Our results on habitat suitability for breeding and connectivity

could be used to constrain and inform models (Gallien et al. 2010; O’Reilly-Nugent et al. 2016), for

example by defining spatially-explicit patches and functional distances between them in

metapopulation models (Guisan & Thuiller 2005). Graph-theoretic dynamic network models, which

integrate local population growth, dispersal of individuals between patches, and allow for temporal

stochasticity in network connections, may be a useful tool to integrate our seasonal scenarios into

one dynamic framework (Ferrari et al. 2014). Lastly, recent models incorporating the effects of

management interventions on feral pig population densities in New Zealand (Krull et al. 2016) or

invasive rabbit metapopulation persistence in Australia (Lurgi et al. 2016) could also be used to

investigate how management can reduce the risk of disease establishment and risk in wildlife hosts.

6.5.3.2 Disease dynamics

Similarly, inferences about the establishment and spread risk of directly transmitted infectious

diseases were made from habitat measures describing potential for breeding, and connectivity

measures describing potential pathways of dispersal between habitat patches, respectively.

141

However, these seasonally static measures may not equate to the actual establishment, spread and

persistence of a given disease at a given point in time. For example, actual disease establishment

may depend on local ‘threshold’ densities and contact rates rather than habitat suitability, and actual

disease spread on the directionality, timing, frequency and number of dispersal movements as well

as many disease-specific parameters (e.g. basic reproductive rate, infectiousness, infectious period,

host mortality, or effects on host behaviour; Doran & Laffan 2005; Cowled & Garner 2008; Milne

et al. 2008). To effectively and rapidly respond to an incursion, spatially-explicit dynamic

epidemiological simulation models must be formulated based on the on-ground conditions at the

time, or ideally, for multiple scenarios of on-ground conditions in advance (Cowled & Garner

2008). Yet, such models are even more complex than ecological simulation models as they require

specification of dynamic parameters related to both host persistence and spread, and disease

transmission between infected and susceptible hosts (McCallum et al. 2001). The results from this

thesis could be used to improve the spatially-explicit ‘host’ component of such investigations

(Doran & Laffan 2005). For example, assumptions of random mixing could be more realistically

constrained by spatially clustered habitat patches and dispersal corridors (Cowled & Garner 2008).

Our results could also be used to generate improved estimates of host densities (albeit with the

considerable uncertainties discussed above), which in turn could ‘seed’ spatially-explicit cellular

automata or individual-based models (Doran & Laffan 2005; Ward et al. 2011; Rees et al. 2013).

6.5.4 Optimization models

This thesis generated improved regional-scale data about feral pig populations and suggested a

number of options how these could be used to inform risk-based disease management in northern

Australia (Section 6.4.2). However, these recommendations remained suggestive and unevaluated.

A large field of research has emerged on structured decision-making and deriving mathematically

‘optimal’ management recommendations, often in the context of systematic conservation planning

(Januchowski-Hartley et al. 2011; Glen et al. 2013). For invasive species, Baxter & Possingham

(2011) traded off search effort and improved spatial knowledge about species occurrence to derive

optimal management investments into either widespread uninformed or focused well-informed

surveillance. By investigating patterns in connectivity networks, Chades et al. (2011) showed “how

to prioritize management and survey effort across time and space for networks of susceptible–

infected–susceptible subpopulations” (p. 8323). The results in this thesis could be used to spatially

inform such management-focused optimization models. For example, individual patches that

contribute most to connectivity within and between habitat components could be prioritized using a

graph-theoretic approach (Galpern et al. 2011). Surveillance activities could then be optimized

across time and space within such patch networks according to mathematical efficiency rules

142

(Chades et al. 2011). Tools for structured decision-making could also combine the disease risk

indices developed in this thesis (establishment and spread) with other important criteria such as

likely impact of disease occurrence or budgetary constraints (Glen et al. 2013).

6.5.5 Other study systems and applications

Finally, despite the limitations and potential improvements discussed above, the multi-level

conceptual framework and expert-based modelling approaches presented in this thesis provide a

useful framework for investigating potential distributions (via resource-based habitat suitability)

and dispersal (via resource-based habitat connectivity) of mobile animals that could be readily

applied to other wide-ranging species, especially in situations where empirical data is limited.

Moreover, spatially-explicit, seasonally-specific and ecologically justified model predictions were

explicitly linked to an applied management problem, the risk-based management of directly

transmitted infectious animal diseases with feral pigs as a wildlife host. The integrated study of

organisms at multiple behavioural and organisational levels and scales in space and time is a current

frontier in ecology (McGarigal et al. 2016; Yackulic & Ginsberg 2016). Similar combinations of

habitat suitability and connectivity models to the one presented in this thesis have increasingly been

applied to aid conservation decisions (Cianfrani et al. 2013; Dickson et al. 2013) and, more rarely,

invasive species management (Stewart-Koster et al. 2015). Yet, Glen et al. (2013) suggested that

‘integrated landscape approaches’ based on metapopulation theory, connectivity models and spatial

optimization are currently underutilized to inform applied biosecurity problems.

The approaches developed here could be adapted, for example, to other widespread vertebrate

pest animals in Australia (e.g. deer, horses, wild dogs, foxes, cats), for which regional-scale

landscape ecological knowledge may be equally limited. To my knowledge, few connectivity

models have been applied to such species in Australia, although many beneficial insights for

invasion management may be gained. For example, results could be used to establish connected

management units or target local population control activities in patches while minimizing

reinvasion along dispersal corridors. The resource-based approaches could also be applied to other

host-pathogen systems. As disease risk was conceptualized as an attribute of the host, the

framework may be applied to directly transmitted diseases in a range of species. Yet, it could also

be extended to study more complex interactions. For example, by modelling habitat patches for

wildlife hosts (and vectors) and combining this information with spatial patterns of domestic animal

or human use, hotspots of disease emergence may be identified and efficient sampling campaigns

designed (Hartemink et al. 2015). It is the hope of the author that the research presented in this

thesis may inspire other scientists and practitioners to continue the journey towards a spatially-

informed, risk-based management of infectious animal diseases.

143

References

Addicott, JF, Aho, JM, Antolin, MF, Padilla, DK, Richardson, JS et al. (1987). "Ecological

neighborhoods: scaling environmental patterns." Oikos 49(3): 340-346.

Alexander, NS, Massei, G & Wint, W (2016). "The European distribution of Sus scrofa. Model

outputs from the project described within the poster – where are all the boars? An attempt to

gain a continental perspective." Open Health Data 4: e1.

Allaire, JJ (2016). Package ‘rmarkdown’: dynamic documents for R. URL

http://rmarkdown.rstudio.com.

Altizer, S, Bartel, R & Han, BA (2011). "Animal migration and infectious disease risk." Science

331(6015): 296.

Anderson, RM, May, RM, Joysey, K, Mollison, D, Conway, GR et al. (1986). "The invasion,

persistence and spread of infectious diseases within animal and plant communities."

Philosophical Transactions of the Royal Society B: Biological Sciences 314(1167): 533-570.

Animal Health Australia (2011). Wild animal response strategy (Version 3.3). Australian

Veterinary Emergency Plan (AUSVETPLAN), Edition 3. Primary Industries Ministerial

Council, Canberra.

Animal Health Australia (2012). Disease strategy: classical swine fever (Version 4.0). Australian

Veterinary Emergency Plan (AUSVETPLAN), Edition 4. National Biosecurity Committee,

Canberra.

Arthur, SM, Manly, BFJ, McDonald, LL & Garner, GW (1996). "Assessing habitat selection when

availability changes." Ecology 77(1): 215-227.

Artois, M, Depner, K, Guberti, V, Hars, J, Rossi, S et al. (2002). "Classical swine fever (hog

cholera) in wild boar in Europe." Revue scientifique et technique (International Office of

Epizootics) 21(1): 287-304.

Atlas of Living Australia, ALA (2017). URL http://www.ala.org.au/.

Austin, M (2007). "Species distribution models and ecological theory: a critical assessment and

some possible new approaches." Ecological Modelling 200(1-2): 1-19.

Australian Bureau of Agricultural and Resource Economics and Sciences, ABARES (2011).

Guidelines for land use mapping in Australia: principles, procedures and definitions, fourth

edition. Commonwealth of Australia, Canberra.

http://rmarkdown.rstudio.com/

http://www.ala.org.au/

144

Australian National Audit Office (2012). Administration of the Northern Australia Quarantine

Strategy: ANAO audit report no. 46 2011–12. Commonwealth of Australia (ANAO), Canberra.

Bahn, V & McGill, BJ (2007). "Can niche-based distribution models outperform spatial

interpolation?" Global Ecology and Biogeography 16(6): 733-742.

Barrios-Garcia, MN & Ballari, SA (2012). "Impact of wild boar (Sus scrofa) in its introduced and

native range: a review." Biological Invasions 14(11): 2283-2300.

Bates, D & Maechler, M (2015). Package 'Matrix': sparse and dense matrix classes and methods.

URL http://Matrix.R-forge.R-project.org/.

Baxter, PWJ & Possingham, HP (2011). "Optimizing search strategies for invasive pests: learn

before you leap." Journal of Applied Ecology 48(1): 86-95.

Beale, R, Fairbrother, J, Inglis, A & Trebeck, D (2008). One biosecurity: a working partnership.

Commonwealth of Australia, Canberra.

Beck, JL & Suring, LH (2009). "Chapter 10: Wildlife habitat-relationships models: description and

evaluation of existing frameworks." in JJ Millspaugh & FR Thompson III (eds). Models for

planning wildlife conservation in large landscapes. Academic Press, San Diego: pp. 251-285.

Beier, P, Majka, D & Jenness, J (2007). Conceptual steps for designing wildlife corridors.

Retrieved June 2014 from http://corridordesign.org/designing_corridors.

Beier, P, Majka, DR & Spencer, WD (2008). "Forks in the road: choices in procedures for

designing wildland linkages." Conservation Biology 22(4): 836-851.

Beier, P, Spencer, W, Baldwin, RF & McRae, BH (2011). "Toward best practices for developing

regional connectivity maps." Conservation Biology 25(5): 879-892.

Bellamy, C & Altringham, J (2015). "Predicting species distributions using record centre data:

multi-scale modelling of habitat suitability for bat roosts." PLoS ONE 10(6): e0128440.

Bengsen, AJ, Gentle, MN, Mitchell, JL, Pearson, HE & Saunders, GR (2014). "Impacts and

management of wild pigs Sus scrofa in Australia." Mammal Review 44(2): 135-147.

Berry, JK (2013). Beyond mapping III: a compilation of Beyond Mapping columns from 1996 to

2013. Basis Press, Fort Collins.

Beutel, TS, Beeton, RJS & Baxter, GS (1999). "Building better wildlife-habitat models." Ecography

22(2): 219-219.

http://matrix.r-forge.r-project.org/

http://corridordesign.org/designing_corridors

145

Boets, P, Landuyt, D, Everaert, G, Broekx, S & Goethals, PLM (2015). "Evaluation and comparison

of data-driven and knowledge-supported Bayesian belief networks to assess the habitat

suitability for alien macroinvertebrates." Environmental Modelling & Software 74: 92-103.

Börger, L, Dalziel, BD & Fryxell, JM (2008). "Are there general mechanisms of animal home range

behaviour? A review and prospects for future research." Ecology Letters 11(6): 637-650.

Bosch, J, Mardones, F, Pérez, A & Muñoz, MJ (2014). "A maximum entropy model for predicting

wild boar distribution in Spain." Spanish Journal of Agricultural Research 12(4): 984-999.

Boyce, MS, Vernier, PR, Nielsen, SE & Schmiegelow, FKA (2002). "Evaluating resource selection

functions." Ecological Modelling 157(2-3): 281-300.

Boyce, MS, Johnson, CJ, Merrill, EH, Nielsen, SE, Solberg, EJ et al. (2016). "REVIEW: can habitat

selection predict abundance?" Journal of Animal Ecology 85(1): 11-20.

Broennimann, O (2015). Package 'ecospat': spatial ecology miscellaneous methods. URL

http://cran.r-project.org/web/packages/ecospat/.

Brookes, VJ, Hernández-Jover, M, Cowled, B, Holyoake, PK & Ward, MP (2014). "Building a

picture: prioritisation of exotic diseases for the pig industry in Australia using multi-criteria

decision analysis." Preventive Veterinary Medicine 113(1): 103-117.

Brooks, RP (1997). "Improving habitat suitability index models." Wildlife Society Bulletin 25(1):

163-167.

Bureau of Meteorology, BOM (2012). Australian hydrological geospatial fabric (geofabric) data

product specification: surface cartography Version 2.1. Bureau of Meteorology, Canberra.

Burt, WH (1943). "Territoriality and home range concepts as applied to mammals." Journal of

Mammalogy 24(3): 346-352.

Cain, J (2001). Planning improvement in natural resource management. Centre for Ecology and

Hydrology, Wallingford, Oxon.

Calabrese, JM & Fagan, WF (2004). "A comparison-shopper's guide to connectivity metrics."

Frontiers in Ecology and the Environment 2(10): 529-536.

Caley, P (1993). "Population dynamics of feral pigs (Sus Scrofa) in a tropical riverine habitat

complex." Wildlife Research 20(5): 625-636.

Caley, P (1997). "Movements, activity patterns and habitat use of feral pigs (Sus scrofa) in a

tropical habitat." Wildlife Research 24(1): 77-87.

http://cran.r-project.org/web/packages/ecospat/

146

Calisher, CH, Childs, JE, Field, HE, Holmes, KV & Schountz, T (2006). "Bats: important reservoir

hosts of emerging viruses." Clinical Microbiology Reviews 19(3): 531-545.

Chades, I, Martin, TG, Nicol, S, Burgman, MA, Possingham, HP et al. (2011). "General rules for

managing and surveying networks of pests, diseases, and endangered species." Proceedings of

the National Academy of Sciences 108(20): 8323–8328.

Chen, SH & Pollino, CA (2012). "Good practice in Bayesian network modelling." Environmental

Modelling & Software 37: 134-145.

Choquenot, D, McIlroy, J & Korn, T (1996). Managing vertebrate pests: feral pigs. Commonwealth

of Australia, Canberra.

Choquenot, D (1998). "Testing the relative influence of intrinsic and extrinsic variation in food

availability on feral pig populations in Australia's rangelands." Journal of Animal Ecology

67(6): 887-907.

Choquenot, D & Ruscoe, WA (2003). "Landscape complementation and food limitation of large

herbivores: habitat-related constraints on the foraging efficiency of wild pigs." Journal of

Animal Ecology 72(1): 14-26.

Cianfrani, C, Maiorano, L, Loy, A, Kranz, A, Lehmann, A et al. (2013). "There and back again?

Combining habitat suitability modelling and connectivity analyses to assess a potential return

of the otter to Switzerland." Animal Conservation 16(5): 584-594.

Cookson, B, Sergeant, ESG & Martin, PA (2012). "Risk-based prioritisation of surveillance for

exotic animal diseases in northern Australia." Proceedings of the 13th International

Symposium on Veterinary Epidemiology and Economics. Maastricht, Netherlands: p. 379.

Corbett, L (1995). "Does dingo predation or buffalo competition regulate feral pig populations in

the Australian wet-dry tropics? An experimental study." Wildlife Research 22(1): 65-74.

Cowled, BD, Lapidge, SJ, Hampton, JO & Spencer, PBS (2006). "Measuring the demographic and

genetic effects of pest control in a highly persecuted feral pig population." The Journal of

Wildlife Management 70(6): 1690-1697.

Cowled, BD & Giannini, F (2007). Spatial predictive modelling for feral pigs in northern Australia

(2.054R). Department of Agriculture, Fisheries and Forestry, Canberra.

Cowled, BD, Aldenhoven, J, Odeh, IOA, Garrett, T, Moran, C et al. (2008). "Feral pig population

structuring in the rangelands of eastern Australia: applications for designing adaptive

management units." Conservation Genetics 9(1): 211-224.

147

Cowled, B & Garner, G (2008). "A review of geospatial and ecological factors affecting disease

spread in wild pigs: considerations for models of foot-and-mouth disease spread." Preventive

Veterinary Medicine 87(3-4): 197-212.

Cowled, BD, Giannini, F, Beckett, SD, Woolnough, A, Barry, S et al. (2009). "Feral pigs:

predicting future distributions." Wildlife Research 36(3): 242-251.

Cowled, B, Garner, M, Negus, K & Ward, M (2012). "Controlling disease outbreaks in wildlife

using limited culling: modelling classical swine fever incursions in wild pigs in Australia."

Veterinary Research 43(3): 1-16.

Daszak, P, Cunningham, AA & Hyatt, AD (2000). "Emerging infectious diseases of wildlife:

threats to biodiversity and human health." Science 287(5452): 443.

DeCesare, NJ, Hebblewhite, M, Schmiegelow, F, Hervieux, D, McDermid, GJ et al. (2012).

"Transcending scale dependence in identifying habitat with resource selection functions."

Ecological Applications 22(4): 1068-1083.

Department of Agriculture and Fisheries, DAF (2015). Annual pest distribution survey 2013-14.

The State of Queensland.

Department of the Environment and Water Resources, DEWR (2007). Australia’s native

vegetation: a summary of Australia’s major vegetation groups, 2007. Commonwealth of

Australia, Canberra.

Department of the Environment (2012). Interim Biogeographic Regionalisation for Australia

(Subregions) v.7 (IBRA). URL https://www.environment.gov.au/land/nrs/science/ibra#ibra.

Department of the Environment (2015). Threat abatement plan for predation, habitat degradation,

competition and disease transmission by feral pigs (Sus scrofa). Commonwealth of Australia,

Canberra.

Department of Science, IT, Innovation and the Arts, DSITIA (2012). Land cover change in

Queensland 2009–10: a Statewide Landcover and Trees Study (SLATS) report. The State of

Queensland, Brisbane.

Dexter, N (1998). "The influence of pasture distribution and temperature on habitat selection by

feral pigs in a semi-arid environment." Wildlife Research 25(5): 547-559.

Dexter, N (1999). "The influence of pasture distribution, temperature and sex on home-range size of

feral pigs in a semi-arid environment." Wildlife Research 26(6): 755-762.

https://www.environment.gov.au/land/nrs/science/ibra%23ibra

148

Dexter, N & McLeod, SR (2015). "Modeling ecological traps for the control of feral pigs." Ecology

and Evolution 5(10): 2036-2047.

Dickson, BG, Roemer, GW, McRae, BH & Rundall, JM (2013). "Models of regional habitat quality

and connectivity for pumas (Puma concolor) in the southwestern United States." PLoS ONE

8(12): e81898.

Dijak, WD, Rittenhouse, CD, Larson, MA, Thompson, FR & Millspaugh, JJ (2007). "Landscape

habitat suitability index software." The Journal of Wildlife Management 71(2): 668-670.

Dijak, WD & Rittenhouse, CD (2009). "Chapter 14: Development and application of habitat

suitability models to large landscapes." in JJ Millspaugh & FR Thompson III (eds). Models for

planning wildlife conservation in large landscapes. Academic Press, San Diego: pp. 367-389.

Doerr, VAJ, Barrett, T & Doerr, ED (2011). "Connectivity, dispersal behaviour and conservation

under climate change: a response to Hodgson et al." Journal of Applied Ecology 48(1): 143-

147.

Doran, RJ & Laffan, SW (2005). "Simulating the spatial dynamics of foot and mouth disease

outbreaks in feral pigs and livestock in Queensland, Australia, using a susceptible-infected-

recovered cellular automata model." Preventive Veterinary Medicine 70(1-2): 133-152.

Dormann, CF, Schymanski, SJ, Cabral, J, Chuine, I, Graham, CH et al. (2012). "Correlation and

process in species distribution models: bridging a dichotomy." Journal of Biogeography

39(12): 2119-2131.

Douglas, SJ & Newton, AC (2014). "Evaluation of Bayesian networks for modelling habitat

suitability and management of a protected area." Journal for Nature Conservation 22(3): 235-

246.

Dowle, M, Srinivasan, A, Short, T & Lianoglou, S (2015). Package 'data.table': extension of

data.frame. URL https://github.com/Rdatatable/data.table/wiki/.

Ducci, L, Agnelli, P, Di Febbraro, M, Frate, L, Russo, D et al. (2015). "Different bat guilds perceive

their habitat in different ways: a multiscale landscape approach for variable selection in species

distribution modelling." Landscape Ecology 30(10): 2147-2159.

Dunning, JB, Danielson, BJ & Pulliam, HR (1992). "Ecological processes that affect populations in

complex landscapes." Oikos 65(1): 169-175.

East, IJ, Wicks, RM, Martin, PAJ, Sergeant, ESG, Randall, LA et al. (2013). "Use of a multi-criteria

analysis framework to inform the design of risk based general surveillance systems for animal

disease in Australia." Preventive Veterinary Medicine 112(3-4): 230-247.

https://github.com/Rdatatable/data.table/wiki/

149

Elith, J & Leathwick, JR (2009). "Species distribution models: ecological explanation and

prediction across space and time." Annual Review of Ecology, Evolution and Systematics 40:

677-697.

Elith, J, Kearney, M & Phillips, SJ (2010). "The art of modelling range-shifting species." Methods

in Ecology and Evolution 1(4): 330-342.

Elledge, AE, McAlpine, CA, Murray, PJ & Gordon, IJ (2013). "Modelling habitat preferences of

feral pigs for rooting in lowland rainforest." Biological Invasions 15(7): 1523-1535.

Emergency Animal Disease Response Agreement (2012). Government and livestock industry cost

sharing deed in respect of emergency animal disease responses (Variation No. 12/01 -

25/05/2012), Canberra.

Environmental Protection Agency, EPA (2005). Wetland mapping and classification methodology.

Overall framework: a method to provide baseline mapping and classification for wetlands in

Queensland, version 1.2. The State of Queensland, Brisbane.

Etherington, TR & Holland, EP (2013). "Least-cost path length versus accumulated-cost as

connectivity measures." Landscape Ecology 28(7): 1223-1229.

Etherington, TR, Perry, GLW, Cowan, PE & Clout, MN (2014). "Quantifying the direct transfer

costs of common brushtail possum dispersal using least-cost modelling: a combined cost-

surface and accumulated-cost dispersal kernel approach." PLoS ONE 9(2): e88293.

Farina, A (2006). Principles and methods in landscape ecology. Chapman & Hall, London.

Fenton, NE, Neil, M & Caballero, JG (2007). "Using ranked nodes to model qualitative judgments

in Bayesian networks." IEEE Transactions on Knowledge and Data Engineering 19(10): 1420-

1432.

Fenton, NE & Neil, M (2013). Risk assessment and decision analysis with Bayesian networks. CRC

Press, Boca Raton.

FeralPigScan (2016). URL http://www.feralscan.org.au/feralpigscan/default.aspx.

Ferrari, JR, Preisser, EL & Fitzpatrick, MC (2014). "Modeling the spread of invasive species using

dynamic network models." Biological Invasions 16(4): 949-960.

Fletcher, RJ, Burrell, NS, Reichert, BE, Vasudev, D & Austin, JD (2016a). "Divergent perspectives

on landscape connectivity reveal consistent effects from genes to communities." Current

Landscape Ecology Reports 1(2): 67-79.

http://www.feralscan.org.au/feralpigscan/default.aspx

150

Fletcher, RJ, McCleery, RA, Greene, DU & Tye, CA (2016b). "Integrated models that unite local

and regional data reveal larger-scale environmental relationships and improve predictions of

species distributions." Landscape Ecology 31(6): 1369-1382.

Forman, RT (1995). "Some general principles of landscape and regional ecology." Landscape

Ecology 10(3): 133-142.

Fox, ID, Neldner, VJ, Wilson, GW & Bannink, PJ (2001). The vegetation of the Australian tropical

savannas. Queensland Herbarium, Environmental Protection Agency, Brisbane.

Franklin, J (2010). "Moving beyond static species distribution models in support of conservation

biogeography." Diversity and Distributions 16(3): 321-330.

Fritzemeier, J, Teuffert, J, Greiser-Wilke, I, Staubach, C, Schluter, H et al. (2000). "Epidemiology

of classical swine fever in Germany in the 1990s." Veterinary Microbiology 77(1-2): 29-41.

Froese, JG, Smith, CS, McAlpine, CA, Durr, PA & van Klinken, RD (2015). "Moving window

analysis links landscape-scale resource utilization to habitat suitability models of feral pigs in

northern Australia." in T Weber, MJ McPhee & RS Anderssen (eds). MODSIM2015, 21st

International Congress on Modelling and Simulation. Modelling and Simulation Society of

Australia and New Zealand, Broadbeach: pp. 1352-1358.

Gabor, TM, Hellgren, EC, Van Den Bussche, RA & Silvy, NJ (1999). "Demography, sociospatial

behaviour and genetics of feral pigs (Sus scrofa) in a semi-arid environment." Journal of

Zoology 247(3): 311-322.

Gallant, JC, Dowling, TI, Read, AM, Wilson, N, Tickle, P et al. (2011). 1 second SRTM derived

products user guide, Version 1.0.4. Geoscience Australia, Canberra.

Gallien, L, Münkemüller, T, Albert, CH, Boulangeat, I & Thuiller, W (2010). "Predicting potential

distributions of invasive species: where to go from here?" Diversity and Distributions 16(3):

331-342.

Galpern, P, Manseau, M & Fall, A (2011). "Patch-based graphs of landscape connectivity: a guide

to construction, analysis and application for conservation." Biological Conservation 144(1):

44-55.

Giles, JR (1980). The ecology of the feral pig in western New South Wales. A thesis submitted for

the degree of Doctor of Philosophy, University of Sydney.

Gill, T, Heidenreich, S & Guerschman, JP (2014). MODIS monthly fractional cover: product

creation and distribution. Joint Remote Sensing Research Program Publication Series.

151

Girvetz, E & Greco, S (2007). "How to define a patch: a spatial model for hierarchically delineating

organism-specific habitat patches." Landscape Ecology 22(8): 1131-1142.

Glen, A, Pech, R & Byrom, A (2013). "Connectivity and invasive species management: towards an

integrated landscape approach." Biological Invasions 15(10): 2127-2138.

Gormley, AM, Forsyth, DM, Griffioen, P, Lindeman, M, Ramsey, DSL et al. (2011). "Using

presence-only and presence–absence data to estimate the current and potential distributions of

established invasive species." Journal of Applied Ecology 48(1): 25-34.

Greer, EB (1983). "Heat stress and reproduction in pigs: its role in seasonal infertility." RAAN

Conference Proceedings. pp. 216-230.

Grinnell, J (1917). "Field tests of theories concerning distributional control." The American

Naturalist 51(602): 115-128.

Guerschman, JP, Hill, MJ, Renzullo, LJ, Barrett, DJ, Marks, AS et al. (2009). "Estimating fractional

cover of photosynthetic vegetation, non-photosynthetic vegetation and bare soil in the

Australian tropical savanna region upscaling the EO-1 Hyperion and MODIS sensors." Remote

Sensing of Environment 113(5): 928-945.

Guisan, A & Thuiller, W (2005). "Predicting species distribution: offering more than simple habitat

models." Ecology Letters 8(9): 993-1009.

Guisan, A, Tingley, R, Baumgartner, JB, Naujokaitis-Lewis, I, Sutcliffe, PR et al. (2013).

"Predicting species distributions for conservation decisions." Ecology Letters 16(12): 1424-

1435.

Hall, LS, Krausman, PR & Morrison, ML (1997). "The habitat concept and a plea for standard

terminology." Wildlife Society Bulletin 25(1): 173-182.

Hamilton, GS, Mather, PB & Wilson, JC (2006). "Habitat heterogeneity influences connectivity in a

spatially structured pest population." Journal of Applied Ecology 43(2): 219-226.

Hamilton, GS & Mather, PB (2009). "Differential impacts of habitat heterogeneity on male and

female connectivity in a spatially structured pest system." Austral Ecology 34(1): 41-49.

Hamilton, SH, Pollino, CA & Jakeman, AJ (2015). "Habitat suitability modelling of rare species

using Bayesian networks: model evaluation under limited data." Ecological Modelling 299: 64-

78.

152

Hampton, JO, Spencer, PBS, Alpers, DL, Twigg, LE, Woolnough, AP et al. (2004). "Molecular

techniques, wildlife management and the importance of genetic population structure and

dispersal: a case study with feral pigs." Journal of Applied Ecology 41(4): 735-743.

Hampton, JO, Spencer, PBS, Elliot, AD & Thompson, RCA (2006). "Prevalence of zoonotic

pathogens from feral pigs in major public drinking water catchments in Western Australia."

EcoHealth 3(2): 103-108.

Hanski, I (1998). "Metapopulation dynamics." Nature 396(6706): 41-49.

Hartemink, N, Vanwambeke, SO, Purse, BV, Gilbert, M & Van Dyck, H (2015). "Towards a

resource-based habitat approach for spatial modelling of vector-borne disease risks." Biological

Reviews 90(4): 1151-1162.

Hastings, A (2014). "Persistence and management of spatially distributed populations." Population

Ecology 56(1): 21-26.

Henderson, WR (2008). Research on wildlife disease preparedness in Australia. Invasive Animals

Cooperative Research Centre, Canberra.

Hengl, T, Sierdsema, H, Radović, A & Dilo, A (2009). "Spatial prediction of species’ distributions

from occurrence-only records: combining point pattern analysis, ENFA and regression-

kriging." Ecological Modelling 220(24): 3499-3511.

Hijmans, RJ (2015). Package 'raster': geographic data analysis and modeling. URL http://cran.r-

project.org/web/packages/raster/.

Hirzel, AH, Le Lay, G, Helfer, V, Randin, C & Guisan, A (2006). "Evaluating the ability of habitat

suitability models to predict species presences." Ecological Modelling 199(2): 142-152.

Hirzel, AH & Le Lay, G (2008). "Habitat suitability modelling and niche theory." Journal of

Applied Ecology 45(5): 1372-1381.

Holland, EP, Aegerter, JN & Smith, GC (2007). "Spatial sensitivity of a generic population model,

using wild boar (Sus scrofa) as a test case." Ecological Modelling 205(1-2): 146-158.

Holland, JD, Bert, DG & Fahrig, L (2004). "Determining the spatial scale of species' response to

habitat." BioScience 54(3): 227-233.

Holland, JD & Yang, S (2016). "Multi-scale studies and the ecological neighborhood." Current

Landscape Ecology Reports 1(4): 135-145.

Hone, J (1990a). "Note on seasonal changes in population density of feral pigs in three tropical

habitats." Australian Wildlife Research 17(2): 131-134.

http://cran.r-project.org/web/packages/raster/

http://cran.r-project.org/web/packages/raster/

153

Hone, J (1990b). "How many feral pigs in Australia?" Australian Wildlife Research 17(6): 571-572.

Hone, J, Pech, R & Yip, P (1992). "Estimation of the dynamics and rate of transmission of classical

swine fever (hog cholera) in wild pigs." Epidemiology and Infection 108(2): 377-386.

Hone, J (2007). Wildlife damage control. CSIRO Publishing, Collingwood.

Hone, J (2012). Applied population and community ecology: the case of feral pigs in Australia.

John Wiley & Sons.

Huang, ZYX, van Langevelde, F, Prins, HHT & de Boer, WF (2015). "Dilution versus facilitation:

Impact of connectivity on disease risk in metapopulations." Journal of Theoretical Biology

376: 66-73.

Hunsaker, CT, Goodchild, MF, Friedl, MA & Case, TJ (2001). Spatial uncertainty in ecology:

implications for remote sensing and GIS applications. Springer, New York.

Hutchinson, GE (1957). "Concluding remarks." Cold Spring Harbor Symposia on Quantitative

Biology 22: 415-427.

Hutchinson, MF, McIntyre, S, Hobbs, RJ, Stein, JL, Garnett, S et al. (2005). "Integrating a global

agro-climatic classification with bioregional boundaries in Australia." Global Ecology and

Biogeography 14(3): 197-212.

Jackson, HB & Fahrig, L (2015). "Are ecologists conducting research at the optimal scale?" Global

Ecology and Biogeography 24(1): 52-63.

Jackson, ND & Fahrig, L (2014). "Landscape context affects genetic diversity at a much larger

spatial extent than population abundance." Ecology 95(4): 871-881.

Jackson, ND & Fahrig, L (2016). "Habitat amount, not habitat configuration, best predicts

population genetic structure in fragmented landscapes." Landscape Ecology 31(5): 951-968.

Januchowski-Hartley, SR, Visconti, P & Pressey, RL (2011). "A systematic approach for

prioritizing multiple management actions for invasive species." Biological Invasions 13(5):

1241-1253.

Jiménez-Valverde, A, Peterson, A, Soberón, J, Overton, J, Aragón, P et al. (2011). "Use of niche

models in invasive species risk assessments." Biological Invasions 13(12): 2785-2797.

Johansen, K, Gill, T, Trevithick, R, Armston, J, Scarth, P et al. (2012). "Validation of Landsat based

time-series of persistent green-vegetation fraction for Australia." Proceedings of the XXII

Congress of the International Photogrammetry and Remote Sensing Society 2012. Melbourne.

154

Johnson, DH (1980). "The comparison of usage and availability measurements for evaluating

resource preference." Ecology 61(1): 65-71.

Johnson, S, Low-Choy, S & Mengersen, K (2012a). "Integrating Bayesian networks and geographic

information systems: Good practice examples." Integrated Environmental Assessment and

Management 8(3): 473-479.

Johnson, CJ, Hurley, M, Rapaport, E & Pullinger, M (2012b). "Using expert knowledge effectively:

lessons from species distribution models for wildlife conservation and management." in AH

Perera, CA Drew & CJ Johnson (eds). Expert knowledge and its application in landscape

ecology. Springer, New York: pp. 153-171.

Jones, DA, Wang, W & Fawcett, R (2009). "High-quality spatial climate data-sets for Australia."

Australian Meteorological and Oceanographic Journal 58(4): 233-248.

Jones, KE, Patel, NG, Levy, MA, Storeygard, A, Balk, D et al. (2008). "Global trends in emerging

infectious diseases." Nature 451(7181): 990-993.

Kanagaraj, R, Wiegand, T, Kramer-Schadt, S & Goyal, SP (2013). "Using individual-based

movement models to assess inter-patch connectivity for large carnivores in fragmented

landscapes." Biological Conservation 167: 298-309.

Kearney, M (2006). "Habitat, environment and niche: what are we modelling?" Oikos 115(1): 186-

191.

Kearney, M & Porter, W (2009). "Mechanistic niche modelling: combining physiological and

spatial data to predict species' ranges." Ecology Letters 12(4): 334-350.

Kearney, M, Porter, WP, Williams, C, Ritchie, S & Hoffmann, AA (2009). "Integrating biophysical

models and evolutionary theory to predict climatic impacts on species' ranges: the dengue

mosquito Aedes aegypti in Australia." Functional Ecology 23(3): 528-538.

Kie, JG, Matthiopoulos, J, Fieberg, J, Powell, RA, Cagnacci, F et al. (2010). "The home-range

concept: are traditional estimators still relevant with modern telemetry technology?"

Philosophical Transactions of the Royal Society of London B: Biological Sciences 365(1550):

2221-2231.

Kindlmann, P & Burel, F (2008). "Connectivity measures: a review." Landscape Ecology 23(8):

879-890.

Koen, EL, Garroway, CJ, Wilson, PJ & Bowman, J (2010). "The effect of map boundary on

estimates of landscape resistance to animal movement." PLoS ONE 5(7): e11785.

155

Koen, EL, Bowman, J, Sadowski, C & Walpole, AA (2014). "Landscape connectivity for wildlife:

development and validation of multispecies linkage maps." Methods in Ecology and Evolution

5(7): 626-633.

Kool, JT, Moilanen, A & Treml, EA (2013). "Population connectivity: recent advances and new

perspectives." Landscape Ecology 28(2): 165-185.

Kramer-Schadt, S, Fernandez, N & Thulke, HH (2007). "Potential ecological and epidemiological

factors affecting the persistence of classical swine fever in wild boar Sus scrofa populations."

Mammal Review 37(1): 1-20.

Krull, CR, Stanley, MC, Burns, BR, Choquenot, D & Etherington, TR (2016). "Reducing wildlife

damage with cost-effective management programmes." PLoS ONE 11(1): e0146765.

Kuhnert, PM, Martin, TG & Griffiths, SP (2010). "A guide to eliciting and using expert knowledge

in Bayesian ecological models." Ecology Letters 13(7): 900-914.

Kupfer, JA (2012). "Landscape ecology and biogeography: rethinking landscape metrics in a post-

FRAGSTATS landscape." Progress in Physical Geography 36(3): 400-420.

Laforge, MP, Brook, RK, Beest, FM, Bayne, EM & McLoughlin, PD (2016). "Grain-dependent

functional responses in habitat selection." Landscape Ecology 31(4): 855-863.

Landuyt, D, Broekx, S, D'Hondt, R, Engelen, G, Aertsens, J et al. (2013). "A review of Bayesian

belief networks in ecosystem service modelling." Environmental Modelling & Software 46: 1-

11.

Landuyt, D, Van der Biest, K, Broekx, S, Staes, J, Meire, P et al. (2015). "A GIS plug-in for

Bayesian belief networks: towards a transparent software framework to assess and visualise

uncertainties in ecosystem service mapping." Environmental Modelling & Software 71: 30-38.

Larson, MA, Millspaugh, JJ & Thompson III, FR (2009). "Chapter 9: A review of methods for

quantifying wildlife habitat in large landscapes." in JJ Millspaugh & FR Thompson III (eds).

Models for planning wildlife conservation in large landscapes. Academic Press, San Diego:

pp. 225-250.

Leibold, MA (1995). "The niche concept revisited: mechanistic models and community context."

Ecology 76(5): 1371-1382.

Leonard, PB, Duffy, EB, Baldwin, RF, McRae, BH, Shah, VB et al. (2017). "gflow: software for

modelling circuit theory-based connectivity at any scale." Methods in Ecology and Evolution

8(4): 519-526.

156

Leslie, E, Cowled, B, Graeme Garner, M, Toribio, J-ALML & Ward, MP (2014). "Effective

surveillance strategies following a potential classical swine fever incursion in a remote wild pig

population in north-western Australia." Transboundary and Emerging Diseases 61(5): 432-

442.

Levin, SA (1992). "The problem of pattern and scale in ecology: the Robert H. MacArthur Award

Lecture." Ecology 73(6): 1943-1967.

Lookingbill, TR, Gardner, RH, Ferrari, JR & Keller, CE (2010). "Combining a dispersal model with

network theory to assess habitat connectivity." Ecological Applications 20(2): 427-441.

Lopez, J, Hurwood, D, Dryden, B & Fuller, S (2014). "Feral pig populations are structured at fine

spatial scales in tropical Queensland, Australia." PLoS ONE 9(3): e91657.

Lorena, AC, Jacintho, LFO, Siqueira, MF, Giovanni, RD, Lohmann, LG et al. (2011). "Comparing

machine learning classifiers in potential distribution modelling." Expert Systems with

Applications 38(5): 5268-5275.

Lurgi, M, Wells, K, Kennedy, M, Campbell, S & Fordham, DA (2016). "A landscape approach to

invasive species management." PLoS ONE 11(7): e0160417.

MacArthur, RH & Wilson, EO (1967). The theory of island biogeography. Princeton University

Press.

Macpherson, MF, Davidson, RS, Duncan, DB, Lurz, PW, Jarrott, A et al. (2016). "Incorporating

habitat distribution in wildlife disease models: conservation implications for the threat of

squirrelpox on the Isle of Arran." Animal Conservation 19(1): 3-14.

Manly, B, McDonald, L, Thomas, D, McDonald, TL & Erickson, WP (2002). Resource selection by

animals: statistical design and analysis for field studies. Springer Netherlands.

Marcot, BG, Holthausen, RS, Raphael, MG, Rowland, MM & Wisdom, MJ (2001). "Using

Bayesian belief networks to evaluate fish and wildlife population viability under land

management alternatives from an environmental impact statement." Forest Ecology and

Management 153(1-3): 29-42.

Marcot, BG (2006). "Characterizing species at risk I: modeling rare species under the Northwest

Forest Plan." Ecology and Society 11(2): 10.

Marcot, BG (2012). "Metrics for evaluating performance and uncertainty of Bayesian network

models." Ecological Modelling 230: 50-62.

157

Margosian, ML, Garrett, KA, Hutchinson, JMS & With, KA (2009). "Connectivity of the American

agricultural landscape: assessing the national risk of crop pest and disease spread." BioScience

59(2): 141-151.

Martin, AE & Fahrig, L (2012). "Measuring and selecting scales of effect for landscape predictors

in species–habitat models." Ecological Applications 22(8): 2277-2292.

Martin, TG, Burgman, MA, Fidler, F, Kuhnert, PM, Low-Choy, S et al. (2012). "Eliciting expert

knowledge in conservation science." Conservation Biology 26(1): 29-38.

Mayor, SJ, Schneider, DC, Schaefer, JA & Mahoney, SP (2009). "Habitat selection at multiple

scales." Écoscience 16(2): 238-247.

McBride, MF, Garnett, ST, Szabo, JK, Burbidge, AH, Butchart, SHM et al. (2012a). "Structured

elicitation of expert judgments for threatened species assessment: a case study on a continental

scale using email." Methods in Ecology and Evolution 3(5): 906-920.

McBride, MF, Fidler, F & Burgman, MA (2012b). "Evaluating the accuracy and calibration of

expert predictions under uncertainty: predicting the outcomes of ecological research." Diversity

and Distributions 18(8): 782-794.

McBride, M (2013). Delivering expert knowledge for conservation: tools for enhancing the quality

of expert judgement. A thesis submitted for the degree of Doctor of Philosophy, The University

of Melbourne.

McCallum, H, Barlow, N & Hone, J (2001). "How should pathogen transmission be modelled?"

Trends in Ecology & Evolution 16(6): 295-300.

McCann, RK, Marcot, BG & Ellis, R (2006). "Bayesian belief networks: applications in ecology

and natural resource management." Canadian Journal of Forest Research 36(12): 3053-3062.

McClure, ML, Burdett, CL, Farnsworth, ML, Lutman, MW, Theobald, DM et al. (2015). "Modeling

and mapping the probability of occurrence of invasive wild pigs across the contiguous United

States." PLoS ONE 10(8): e0133771.

McClure, ML, Hansen, AJ & Inman, RM (2016). "Connecting models to movements: testing

connectivity model predictions against empirical migration and dispersal data." Landscape

Ecology 31(7): 1419-1432.

McGarigal, K & Marks, BJ (1995). FRAGSTATS: spatial pattern analysis program for quantifying

landscape structure. US Department of Agriculture, Forest Service, Pacific Northwest

Research Station, Oregon.

158

McGarigal, K, Cushman, SE & Ene, E (2012). FRAGSTATS v4: spatial pattern analysis program

for categorical and continuous maps. University of Massachusetts, Amherst.

McGarigal, K, Wan, HY, Zeller, KA, Timm, BC & Cushman, SA (2016). "Multi-scale habitat

selection modeling: a review and outlook." Landscape Ecology 31(6): 1161-1175.

McNay, RS, Marcot, BG, Brumovsky, V & Ellis, R (2006). "A Bayesian approach to evaluating

habitat for woodland caribou in north-central British Columbia." Canadian Journal of Forest

Research 36(12): 3117-3133.

McRae, BH (2006). "Isolation by resistance." Evolution 60(8): 1551-1561.

McRae, BH & Beier, P (2007). "Circuit theory predicts gene flow in plant and animal populations."

Proceedings of the National Academy of Sciences 104(50): 19885-19890.

McRae, BH, Dickson, BG, Keitt, TH & Shah, VB (2008). "Using circuit theory to model

connectivity in ecology, evolution, and conservation." Ecology 89(10): 2712-2724.

McRae, BH, Shah, VB & Mohapatra, TK (2013). Circuitscape 4 user guide. The Nature

Conservancy.

Miguet, P, Jackson, HB, Jackson, ND, Martin, AE & Fahrig, L (2016). "What determines the spatial

extent of landscape effects on species?" Landscape Ecology 31(6): 1177-1194.

Millspaugh, JJ & Thompson III, FR (eds) (2009). Models for planning wildlife conservation in

large landscapes. Academic Press, San Diego.

Milne, G, Fermanis, C & Johnston, P (2008). "A mobility model for classical swine fever in feral

pig populations." Veterinary Research 39(6): 53.

Mitchell, J, Allen, L & Merrell, P (1982). Vertebrate pests of Queensland. Stock Routes and Rural

Lands Protection Board, Queensland Department of Lands, Brisbane.

Mitchell, J (2002). Ecology and management of feral pigs (Sus scrofa) in rainforests A thesis

submitted for the degree of Doctor of Philosophy, James Cook University.

Mitchell, J (2008). Feral pig control: a practical guide to pig control in Queensland. The State of

Queensland (DPI&F), Charters Towers.

Mitchell, J, Dorney, W, Mayer, R & McIlroy, J (2009). "Migration of feral pigs (Sus scrofa) in

rainforests of north Queensland: fact or fiction?" Wildlife Research 36(2): 110-116.

Mitchell, J (2010). Experimental research to quantify the environmental impact of feral pigs within

tropical freshwater ecosystems. Final Report to the Department of the Environment, Water,

Heritage and the Arts, Canberra.

159

Mitchell, MS, Zimmerman, JW & Powell, RA (2002). "Test of a habitat suitability index for black

bears in the southern Appalachians." Wildlife Society Bulletin 30(3): 794-808.

Mitchell, MS & Powell, RA (2004). "A mechanistic home range model for optimal use of spatially

distributed resources." Ecological Modelling 177(1-2): 209-232.

Mitchell, MS & Powell, RA (2012). "Foraging optimally for home ranges." Journal of Mammalogy

93(4): 917-928.

Moilanen, A & Hanski, I (2001). "On the use of connectivity measures in spatial ecology." Oikos

95(1): 147-151.

Moilanen, A & Nieminen, M (2002). "Simple connectivity measures in spatial ecology." Ecology

83(4): 1131-1145.

Moilanen, A (2011). "On the limitations of graph-theoretic connectivity in spatial ecology and

conservation." Journal of Applied Ecology 48(6): 1543-1547.

Moorcroft, PR (2012). "Mechanistic approaches to understanding and predicting mammalian space

use: recent advances, future directions." Journal of Mammalogy 93(4): 903-916.

Morelle, K, Lehaire, F & Lejeune, P (2014). "Is wild boar heading towards movement ecology? A

review of trends and gaps." Wildlife Biology 20(4): 196-205.

Morelle, K & Lejeune, P (2015). "Seasonal variations of wild boar Sus scrofa distribution in

agricultural landscapes: a species distribution modelling approach." European Journal of

Wildlife Research 61(1): 45-56.

Morrison, ML, Marcot, BG & Mannan, RW (2006). Wildlife-habitat relationships: concepts and

applications. Island Press, Washington.

Murray, JV, Goldizen, AW, O’Leary, RA, McAlpine, CA, Possingham, HP et al. (2009). "How

useful is expert opinion for predicting the distribution of a species within and beyond the

region of expertise? A case study using brush-tailed rock-wallabies Petrogale penicillata."

Journal of Applied Ecology 46(4): 842-851.

Murray, JV, Stokes, KE & van Klinken, RD (2012). "Predicting the potential distribution of a

riparian invasive plant: the effects of changing climate, flood regimes and land-use patterns."

Global Change Biology 18(5): 1738–1753.

Murray, JV, Berman, DM & van Klinken, RD (2014). "Predictive modelling to aid the regional-

scale management of a vertebrate pest." Biological Invasions 16(11): 2403-2425.

160

Murray, J, Froese, J, Perry, J, Navarro Garcia, J & van Klinken, R (2015). Impact modelling for

rabbits and feral pigs in the QMDB. CSIRO Biosecurity, Brisbane.

National Audit Office (2002). The 2001 outbreak of foot and mouth disease. Stationary Office,

London.

National Committee on Soil and Terrain, NCST (2009). Australian soil and land survey field

handbook. CSIRO Publishing, Collingwood.

O’Reilly-Nugent, A, Palit, R, Lopez-Aldana, A, Medina-Romero, M, Wandrag, E et al. (2016).

"Landscape effects on the spread of invasive species." Current Landscape Ecology Reports

1(3): 107-114.

Olsen, B, Munster, VJ, Wallensten, A, Waldenström, J, Osterhaus, ADME et al. (2006). "Global

patterns of Influenza A virus in wild birds." Science 312(5772): 384.

Ostfeld, RS, Glass, GE & Keesing, F (2005). "Spatial epidemiology: an emerging (or re-emerging)

discipline." Trends in Ecology & Evolution 20(6): 328-336.

Paton, RS & Matthiopoulos, J (2016). "Defining the scale of habitat availability for models of

habitat selection." Ecology 97(5): 1113-1122.

Pearson, HE (2012). Understanding and mitigating the risk of pathogen transmission from wild

animals to domestic pigs in Australia. A thesis submitted for the degree of Doctor of

Philosophy, The University of Sydney.

Pearson, RG (2007). Species’ distribution modeling for conservation educators and practitioners:

synthesis. American Museum of Natural History.

Pech, RP & Hone, J (1988). "A model of the dynamics and control of an outbreak of foot and

mouth disease in feral pigs in Australia." Journal of Applied Ecology 25(1): 63-77.

Pelletier, D, Clark, M, Anderson, MG, Rayfield, B, Wulder, MA et al. (2014). "Applying circuit

theory for corridor expansion and management at regional scales: tiling, pinch points, and

omnidirectional connectivity." PLoS ONE 9(1): e84135.

Perera, AH, Drew, CA & Johnson, CJ (eds) (2012). Expert knowledge and its application in

landscape ecology. Springer, New York.

Phillips, BL, Chipperfield, JD & Kearney, MR (2008). "The toad ahead: challenges of modelling

the range and spread of an invasive species." Wildlife Research 35(3): 222-234.

Phillips, SJ, Anderson, RP & Schapire, RE (2006). "Maximum entropy modeling of species

geographic distributions." Ecological Modelling 190(3–4): 231-259.

161

Phillips, SJ, Dudík, M, Elith, J, Graham, CH, Lehmann, A et al. (2009). "Sample selection bias and

presence-only distribution models: implications for background and pseudo-absence data."

Ecological Applications 19(1): 181-197.

Plowright, RK, Foley, P, Field, HE, Dobson, AP, Foley, JE et al. (2011). "Urban habituation,

ecological connectivity and epidemic dampening: the emergence of Hendra virus from flying

foxes." Proceedings of the Royal Society B: Biological Sciences 278(1725): 3703.

Powell, RA & Mitchell, MS (2012). "What is a home range?" Journal of Mammalogy 93(4): 948-

958.

Pulliam, HR & Danielson, BJ (1991). "Sources, sinks, and habitat selection: a landscape perspective

on population dynamics." The American Naturalist 137(Supplement): S50-S66.

Pulliam, HR (2000). "On the relationship between niche and distribution." Ecology Letters 3(4):

349-361.

Raupach, M, Briggs, P, Haverd, V, King, E, Paget, M et al. (2009). Australian Water Availability

Project (AWAP), CSIRO Marine and Atmospheric Research component: final report for phase

3. Technical Report No. 013. Centre for Australian Weather and Climate Research (Bureau of

Meteorology and CSIRO), Melbourne.

Rayfield, B, Fortin, M-J & Fall, A (2011). "Connectivity for conservation: a framework to classify

network measures." Ecology 92(4): 847-858.

Rayfield, B, Pelletier, D, Dumitru, M, Cardille, JA & Gonzalez, A (2016). "Multipurpose habitat

networks for short-range and long-range connectivity: a new method combining graph and

circuit connectivity." Methods in Ecology and Evolution 7(2): 222-231.

RCoreTeam (2015). R: a language and environment for statistical computing. R Foundation for

Statistical Computing, Vienna, Austria. URL http://www.R-project.org/.

Rees, EE, Pond, BA, Tinline, RR & Bélanger, D (2013). "Modelling the effect of landscape

heterogeneity on the efficacy of vaccination for wildlife infectious disease control." Journal of


Reisen, WK (2010). "Landscape epidemiology of vector-borne diseases." Annual Review of

Entomology 55(1): 461-483.

Rhodes, JR, McAlpine, CA, Lunney, D & Possingham, HP (2005). "A spatially explicit habitat

selection model incorporating home range behaviour." Ecology 86(5): 1199-1205.

http://www.r-project.org/

162

Riley, S (2007). "Large-scale spatial-transmission models of infectious disease." Science

316(5829): 1298.

Rioux Paquette, S, Talbot, B, Garant, D, Mainguy, J & Pelletier, F (2014). "Modelling the dispersal

of the two main hosts of the raccoon rabies variant in heterogeneous environments with

landscape genetics." Evolutionary Applications 7(7): 734-749.

Ross, B (2009). Diet selectivity and feeding ecology of feral pigs (Sus scrofa) in Lakefield National

Park, Cape York Peninsula. A thesis submitted for the degree of BSc with Honours, James

Cook University.

Scott, A, Zepeda, C, Garber, L, Smith, J, Swayne, D et al. (2006). "The concept of

compartmentalisation." Revue scientifique et technique (International Office of Epizootics)

25(3): 873-907.

Shirk, AJ & McRae, BH (2013). Gnarly Landscape Utilities: Core Mapper user guide. The Nature

Conservancy, Fort Collins.

Smith, CS, Howes, AL, Price, B & McAlpine, CA (2007). "Using a Bayesian belief network to

predict suitable habitat of an endangered mammal: the Julia Creek dunnart (Sminthopsis

douglasi)." Biological Conservation 139(3-4): 333-347.

Smith, CS, van Klinken, RD, Seabrook, L & McAlpine, C (2012). "Estimating the influence of land

management change on weed invasion potential using expert knowledge." Diversity and

Distributions 18(8): 818-831.

Speirs-Bridge, A, Fidler, F, McBride, M, Flander, L, Cumming, G et al. (2010). "Reducing

overconfidence in the interval judgments of experts." Risk Analysis 30(3): 512-523.

Spencer, PBS, Lapidge, SJ, Hampton, JO & Pluske, JR (2005). "The sociogenetic structure of a

controlled feral pig population." Wildlife Research 32(4): 297-304.

Stankus, T (2012). "Razorbacks: feral pigs as agricultural pests, disruptors of ecosystems, reservoirs

of contagion, and favored game for sport and subsistence hunters: a review of the literature,

2005-2011." Journal of Agricultural & Food Information 13(3): 283-301.

Stephens, PA, Pettorelli, N, Barlow, J, Whittingham, MJ & Cadotte, MW (2015). "Management by

proxy? The use of indices in applied ecology." Journal of Applied Ecology 52(1): 1-6.

Stewart-Koster, B, Olden, JD & Johnson, PTJ (2015). "Integrating landscape connectivity and

habitat suitability to guide offensive and defensive invasive species management." Journal of


163

Tantipisanuh, N, Gale, GA & Pollino, C (2014). "Bayesian networks for habitat suitability

modeling: a potential tool for conservation planning with scarce resources." Ecological

Applications 24(7): 1705-1718.

Tischendorf, L & Fahrig, L (2000). "On the usage and measurement of landscape connectivity."

Oikos 90(1): 7-19.

Turner, MG & Garder, RH (2015). Landscape ecology in theory and practice: pattern and process.

Springer, New York.

Urban, DL, Minor, ES, Treml, EA & Schick, RS (2009). "Graph models of habitat mosaics."

Ecology Letters 12(3): 260-273.

Uuemaa, E, Antrop, M, Roosaare, J, Marja, R & Mander, Ü (2009). "Landscape metrics and

indices: an overview of their use in landscape research." Living Reviews in Landscape

Research 3(1): 1-28.

Uusitalo, L (2007). "Advantages and challenges of Bayesian networks in environmental modelling."

Ecological Modelling 203(3-4): 312-318.

van Klinken, RD, Murray, JV & Smith, C (2015). "Process-based pest risk mapping using Bayesian

networks." in RC Venette (ed). Pest risk modelling and mapping for invasive alien species.

CABI International: pp. 171-188.

VanDerWal, J, Shoo, LP, Graham, C & Williams, SE (2009). "Selecting pseudo-absence data for

presence-only distribution modeling: how far should you stray from what you know?"

Ecological Modelling 220(4): 589-594.

Venette, RC, Kriticos, DJ, Magarey, RD, Koch, FH, Baker, RHA et al. (2010). "Pest risk maps for

invasive alien species: a roadmap for improvement." BioScience 60(5): 349-362.

von Humboldt, A & Bonpland, A (1807). Ideen zu einer Geographie der Pflanzen : nebst einem

Naturgemälde der Tropenländer : auf Beobachtungen und Messungen gegründet, welche vom

10ten Grade nördlicher bis zum 10ten Grade südlicher Breite, in den Jahren 1799, 1800, 1801,

1802 und 1803 angestellt worden sind. Bey F.G. Cotta/Bey F. Schoell, Tübingen/Paris.

Ward, MP, Laffan, SW & Highfield, LD (2011). "Disease spread models in wild and feral animal

populations: application of artificial life models." Revue scientifique et technique (International

Office of Epizootics) 30(2): 437-446.

Weber, MM, Stevens, RD, Diniz-Filho, JAF & Grelle, CEV (in press). "Is there a correlation

between abundance and environmental suitability derived from ecological niche modelling? A

meta-analysis." Ecography.

164

West, P (2008). Assessing invasive animals in Australia 2008. Invasive Animals Cooperative

Research Centre, Canberra.

Wiens, JA (1989). "Spatial scaling in ecology." Functional Ecology 3(4): 385-397.

Wilhere, GF (2012). "Using Bayesian networks to incorporate uncertainty in habitat suitability

index models." The Journal of Wildlife Management 76(6): 1298-1309.

Wilson, G, Dexter, N, O'Brien, P & Bomford, M (1992). Pest animals in Australia: a survey of

introduced wild mammals. Kangaroo Press, Kenthurst.

Wilson, MFJ, O’Connell, B, Brown, C, Guinan, JC & Grehan, AJ (2007). "Multiscale terrain

analysis of multibeam bathymetry data for habitat mapping on the continental slope." Marine

Geodesy 30(1-2): 3-35.

World Organisation for Animal Health, OIE (2016). Terrestrial animal health code, 25th edition.

URL http://www.oie.int/international-standard-setting/terrestrial-code/access-online/.

Wu, J (2013). "Landscape ecology." in R Leemans (ed). Ecological systems. Springer New York:

pp. 179-200.

Wurster, CM, Robertson, J, Westcott, DA, Dryden, B, Zazzo, A et al. (2012). "Utilization of

sugarcane habitat by feral pig (Sus scrofa) in northern tropical Queensland: evidence from the

stable isotope composition of hair." PLoS ONE 7(9): e43538.

Xie, Y (2016). Package 'knitr': a general-purpose package for dynamic report generation in R.

URL http://yihui.name/knitr/.

Yackulic, CB & Ginsberg, JR (2016). "The scaling of geographic ranges: implications for species

distribution models." Landscape Ecology 31(6): 1195-1208.

Zanardi, G, Macchi, C, Sacchi, C & Rutili, D (2003). "Classical swine fever in wild boar in the

Lombardy region of Italy from 1997 to 2002." Veterinary Record 152(15): 461-465.

Zeller, KA, McGarigal, K & Whiteley, AR (2012). "Estimating landscape resistance to movement:

a review." Landscape Ecology 27(6): 777-797.

http://www.oie.int/international-standard-setting/terrestrial-code/access-online/

http://yihui.name/knitr/

165

Appendices

Appendix 1 Human ethics approval

Letter 1: The University of Queensland, School of Agriculture and Food Sciences Ethics Committee

9 October 2013

Dear Jens

Re: Ethical Research Application

On behalf of the SAFS Ethics Committee this letter is to formally advise that your application for ethical

research involved with your PhD research project titled ‘Regional-scale spatial modeling of feral pig

distribution and abundance in northern Australia’ has been approved.

We wish you every success with your research.

Yours sincerely

Kaelene Matts | Academic Administration Officer

School of Agriculture & Food Sciences | Faculty of Science

The University of Queensland | Gatton Queensland 4343 | Australia

T: +61 7 5460 1471 F: +61 7 5460 1324

E: [email protected] W: http://www.uq.edu.au/agriculture/

166

Letter 2: CSIRO Social Science Human Research Ethics Committee

27 August 2013

Subject: Ethics Clearance 075/13 Assessing Regional-scale spatial modelling of feral pig distribution and

abundance in northern Australia

Dear Jens

We have now undertaken an initial ethics review and risk assessment of your project proposal entitled

“Assessing Regional-scale spatial modelling of feral pig distribution and abundance in northern Australia"

(075/13).

Please be advised that based on the information you have provided your project has been assessed against the

requirements of the National Statement as posing a low risk to participants. As such ethical clearance has

been granted for you to undertake the research and for the project to be managed within CSIRO's existing

project management processes. Ethical clearance has been granted for the period 01/09/13 – 31/01/16 subject

to the following condition:

A copy of the participant information / consent materials that will be used for the project should be

provided for Executive review and clearance prior to use.

In granting this clearance we would like to remind you of the need to adhere to the requirements of the

National Statement at all times during the life of the project. Should any adverse events occur to participants

during or resulting from the research or any ethically relevant variations be needed regarding the project's

implementation or completion you are required to notify us immediately for further advice or amended

clearance. You will also be required to complete a brief report about the project upon its completion. A

template for this report will be forwarded to you in due course.

We wish you all the best with the research.

Sincerely

Cathy Pitkin

Manager, Social Responsibility and Ethics

CSIRO

EcoSciences Precinct, 41 Boggo Rd, Dutton Park, QLD 4102

GPO Box 2583, Brisbane QLD 4001

Ph: +61-7-3833 5693 M: +61-409-441-055 Fax: +61-7-3833 5504 Web: http://www.ces.csiro.au/

167

Appendix 3.1 Model variables – parameters

(Appendix Table 3.1.1 included as Table 4.1 in this thesis.)

(Appendix Table 3.1.2 included as Table 4.2 in this thesis.)

(Appendix Table 3.1.3 included as Appendix Table 4.1 in this thesis.)

Appendix Table 3.1.4 Expert-elicited response-to-pattern curves. For each habitat variable – water (A),

food (B), protection from heat stress (C) and protection from disturbance (D) – tables show resource

suitability indices for distance-dependent measurement (SIDr) and composition-dependent measurement

(SICr) as elicited from each individual expert.

168

(A) Expert Water suitability index (SID water) per distance band

Very close Close Medium Far Very far

Expert 1 100 50 30 30 10

Expert 2 100 100 100 50 10

Expert 3 100 100 100 70 50

Expert 4 100 70 50 30 10

Expert 5 100 100 100 70 50

Expert 6 100 100 100 70 50

Expert Water suitability index (SIC water) per average quality class (to % of max)


(-100) (-80) (-60) (-40) (-30) (-25) (-20) (-15) (-10) (-5)

Expert 1 90 90 90 90 90 70 50 30 10 10

Expert 2 90 90 90 90 90 90 70 70 50 10

Expert 3 90 90 90 90 90 90 70 70 50 10

Expert 4 90 90 90 90 90 90 70 50 30 10

Expert 5 90 90 90 90 90 70 50 30 30 10

Expert 6 90 90 90 90 90 90 70 50 30 10

(B) Expert Food suitability index (SID food) per distance band


Expert 1 100 70 50 30 10

Expert 2 100 70 50 30 10

Expert 3 100 70 50 30 10

Expert 4 100 70 50 30 10

Expert 5 100 100 100 70 50

Expert 6 100 70 50 30 10

Expert Food suitability index (SIC food) per average quality class (to % of max)


(-100) (-80) (-60) (-40) (-30) (-25) (-20) (-15) (-10) (-5)

Expert 1 90 90 90 70 50 50 30 30 30 10

Expert 2 90 90 90 70 50 50 30 30 10 10

Expert 3 90 90 70 50 50 50 30 30 30 10

Expert 4 90 90 90 70 70 70 50 50 30 10

Expert 5 90 90 70 70 70 70 50 30 10 10

Expert 6 90 70 50 30 30 30 10 10 10 10

169

(C) Expert Heat protection suitability index (SID heat) per distance band


Expert 1 100 30 10 10 10

Expert 2 100 50 10 10 10

Expert 3 100 100 100 70 30

Expert 4 100 70 70 70 30

Expert 5 100 100 100 100 70

Expert 6 Not elicited

Expert Heat protection suitability index (SIC heat) per average quality class (to % of max)


(-100) (-80) (-60) (-40) (-30) (-25) (-20) (-15) (-10) (-5)

Expert 1 90 90 70 50 50 50 30 30 10 10

Expert 2 90 90 90 70 70 70 50 30 30 10

Expert 3 90 90 90 90 90 90 70 70 50 10

Expert 4 90 90 90 90 90 90 70 50 30 10

Expert 5 90 90 90 90 90 90 70 70 50 10


(D) Expert Disturbance protection suitability index (SID disturbance) per distance band


Expert 1 100 30 10 10 10

Expert 2 100 70 50 30 10


Expert 4 100 30 10 10 10

Expert 5 100 50 10 10 10

Expert 6 100 50 30 10 10

Expert Disturbance protection suitability index (SIC disturbance) per average quality class (to % of max)


(-100) (-80) (-60) (-40) (-30) (-25) (-20) (-15) (-10) (-5)

Expert 1 90 70 50 30 30 30 10 10 10 10

Expert 2 90 90 90 70 70 70 50 30 30 10


Expert 4 90 90 70 50 50 50 30 30 10 10

Expert 5 90 70 50 30 30 30 10 10 10 10

Expert 6 90 90 90 70 70 70 50 50 30 10

170

Appendix 3.2 Methodology – spatial pattern suitability analysis

This page is intentionally blank.

R code - spatial pattern suitability analysis(PATTSI)Jens G. Froese

10 August 2016

This document provides a detailed, reproducible description of the spatial pattern suitability analysis methodology. It isSupporting Information to the manuscript:

Froese JG, Smith CS, McAlpine CA, Durr PA, van Klinken RD. Integrating landscapestructure improves habitat models of mobile animals: feral pigs in Australia.

It is written in R Markdown ([1]) and knitr ([2]), two R ([3]) packages for writing dynamic, reproducible reports. A

.zip file containing data inputs to reproduce analyses can be requested from the authors. Some parts of the code

used to print this document have been suppressed to enhance readability. A generalized version of the code isavailable at URL https://github.com/jgfroese/PATTSI (https://github.com/jgfroese/PATTSI).

Load required R packagesR packages raster ([4]) and Matrix ([5]) and their dependencies are required for spatial pattern suitability

analysis. Session information incl. package versions are listed at the bottom of this document.

require(raster) # for all analyses of raster objects incl. moving w indow analysis

require(Matrix) # for function `nnzero`

1. Expert elicitationOverview

For each habitat variable (i.e. its modelled resource quality indices xr ), we elicited two response-to-pattern curves

from experts, relating structural metrics of resource patterns to functional suitability indices SIr . We focused on a

distance-dependent and a composition-dependent response curve. The information below supplements thedescription of elicitation procedures in Chapter 3. Elicitation results from each indivudal expert are provided inAppendix Table 3.1.4. Expert-averaged response-to-pattern curves are shown in Fig 3.5.

Step 1: Distance-dependent response-to-pattern curve

Distance-dependent curves followed a step-wise pattern, because we discretised both:

distance into five equal distance bands (“very close”, “close”, “medium”, “far” and “very far”), relative to eachexpert’s defined mobility threshold (i.e. 1km, 2km or 3km).resource suitability indices ( SIr ) into five equal classes (“very good (80-100)”, “good (60-80)”, “moderate

(40-60)”, “poor (20-40)” and “very poor (0-20)”)

We asked experts to relate each distance band to a corresponding suitability class under the assumption that othervariables do not constrain suitability.

To derive distance weights for computation of 2. Distance-dependent resource suitability indices, Step 4, we usedthe mid-points of elicited suitability index classes divided by 100 (e.g. class “moderate (40-60)” = SIr 50 = weight

0.5). For class “very good (80-100)” we did not use the mid-point SIr = 90 but assigned SIr = 100 (= weight 1.0)

to avoid unintended distance penalties (i.e. an adjacent resource of quality xr = 60 should compute as distance-

weighted suitability SIr = 60 (if weight is 1.0) and not SIr = 54 (if weight is 0.9)).

Figure 1 Step-wise distance-dependent response curve for habitat variable “food” elicited from expert 4 (A) anddistance weights derived for computation of food suitability indices (B)

Step 2: Composition-dependent response-to-pattern-curve

Composition-dependent curves also followed a step-wise pattern, because we discretised both:

averagen quality into five equal classes (“very high (80-100%)”, “high (60-80%)”, “moderate (40-60%)”, “low(20-40%)” and “very low (0-20%)”), relative relative to 100% cover with highest quality resources.resource suitability indices ( SIr ) into five equal classes (“very good (80-100)”, “good (60-80)”, “moderate

(40-60)”, “poor (20-40)” and “very poor (0-20)”)

We asked experts to relate each average quality class to a corresponding suitability class under the assumption thatother variables do not constrain suitability.

Prior to computation of 3. Composition-dependent resource suitability indices, Step 2, we slightly adjustedresponse curves to avoid unintended results. For example, some experts related “very low” (= 0-20%) average qualityto “moderate” suitability to highlight feral pigs’ ability to effectively utilize small quantities of a high quality resource. Assuch a function would also describe absent resources as “moderately” suitable, we adjusted it to 0-5% average quality= “very poor” suitability, 5-10% average quality = “poor” suitability and 10-20% average quality = “moderate” suitability.

Figure 2 Step-wise composition-dependent response curve for habitat variable “food” elicited from expert 4 (A) andadjusted curve used for computation of food suitability indices (B)]

2. Distance-dependent resource suitability indicesGoal

Focal pixel resource suitability index SIr depends on the distance of a (numerical) habitat variable.

Method:

Generate a circular moving window where each position is weighted by its distance from the focal pixel(radius/weights derived from 1. Expert elicitation).compute the focal pixel SIr as the highest weighted value ( xr ) of a habitat variable within this moving

window.

Step 1

Define a function that returns a circular matrix of given radius and resolution and assigns value 1 if matrix position <=

radius and value NA if matrix position > radius (Source: [6])

make_circ_filter <- function(radius, res){

circ_filter <- matrix(NA, nrow=1+(2*radius/res), ncol =1+(2*radius/res))

dimnames(circ_filter)[[1]] <- seq(-radius, radius, by =res)


sweeper <- function(mat){

for(row in 1:nrow(mat)){

for(col in 1:ncol(mat)){

dist <- sqrt((as.numeric(dimnames(mat)[[1]])[row])^2 +

(as.numeric(dimnames(mat)[[1]])[col])^2)

if(dist<=radius) {mat[row, col]<-1}

}

}

return(mat)

}

out <- sweeper(circ_filter)

return(out)

}

Step 2

Apply function to generate five matrices with different radii (= distance bands), relative to each expert’s definedmobility threshold (i.e. 1km for Expert1, 2km for Experts 3/5 and 3km for Experts 2/4/6).

res <- 1 # resolution (= pixel size, e.g. 100m)

mr <- 10 # matrix radius (= mobility threshold, must be mult iple of res, e.g 1km = 10 x 100

m)

m.vf <- make_circ_filter(mr, res) # distance band 'very far' (= mobility threshold)

m.f <- make_circ_filter((mr/5)*4, res) # distance band 'far'

m.m <- make_circ_filter((mr/5)*3, res) # distance band 'medium'

m.c <- make_circ_filter((mr/5)*2, res) # distance band 'close'

m.vc <- make_circ_filter((mr/5), res) # distance band 'very close'

Replace value==1 with unique temp value in ascending order from largest to smallest matrix

m.vf[m.vf == 1] <- 1

m.f[m.f == 1] <- 2

m.m[m.m == 1] <- 3

m.c[m.c == 1] <- 4

m.vc[m.vc == 1] <- 5

Step 3

Combine the five matrices into one (two at a time starting with the smallest):

a.1 <- array(NA, dim(m.c), dimnames(m.c)) # create temp array of size = larger matrix

a.1[rownames(m.vc), colnames(m.vc)] <- m.vc # ... with values = smaller matrix

m.c <- pmax(m.c, a.1, na.rm = TRUE) # combine values: larger matrix + temp array

a.2 <- array(NA, dim(m.m), dimnames(m.m)) # repeat with: output + next-larger matrix

a.2[rownames(m.c), colnames(m.c)] <- m.c

m.m <- pmax(m.m, a.2, na.rm = TRUE)

a.3 <- array(NA, dim(m.f), dimnames(m.f))

a.3[rownames(m.m), colnames(m.m)] <- m.m

m.f <- pmax(m.f, a.3, na.rm = TRUE)

a.4 <- array(NA, dim(m.vf), dimnames(m.vf))

a.4[rownames(m.f), colnames(m.f)] <- m.f

m.band <- pmax(m.vf, a.4, na.rm = TRUE)

m.band

## -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10

## -10 NA NA NA NA NA NA NA NA NA NA 1 NA NA NA NA NA NA NA NA NA NA

## -9 NA NA NA NA NA NA 1 1 1 1 1 1 1 1 1 NA NA NA NA NA NA

## -8 NA NA NA NA 1 1 1 1 1 1 2 1 1 1 1 1 1 NA NA NA NA

## -7 NA NA NA 1 1 1 1 2 2 2 2 2 2 2 1 1 1 1 NA NA NA

## -6 NA NA 1 1 1 2 2 2 2 2 3 2 2 2 2 2 1 1 1 NA NA

## -5 NA NA 1 1 2 2 2 3 3 3 3 3 3 3 2 2 2 1 1 NA NA

## -4 NA 1 1 1 2 2 3 3 3 3 4 3 3 3 3 2 2 1 1 1 NA

## -3 NA 1 1 2 2 3 3 3 4 4 4 4 4 3 3 3 2 2 1 1 NA

## -2 NA 1 1 2 2 3 3 4 4 4 5 4 4 4 3 3 2 2 1 1 NA

## -1 NA 1 1 2 2 3 3 4 4 5 5 5 4 4 3 3 2 2 1 1 NA

## 0 1 1 2 2 3 3 4 4 5 5 5 5 5 4 4 3 3 2 2 1 1

## 1 NA 1 1 2 2 3 3 4 4 5 5 5 4 4 3 3 2 2 1 1 NA

## 2 NA 1 1 2 2 3 3 4 4 4 5 4 4 4 3 3 2 2 1 1 NA

## 3 NA 1 1 2 2 3 3 3 4 4 4 4 4 3 3 3 2 2 1 1 NA

## 4 NA 1 1 1 2 2 3 3 3 3 4 3 3 3 3 2 2 1 1 1 NA

## 5 NA NA 1 1 2 2 2 3 3 3 3 3 3 3 2 2 2 1 1 NA NA

## 6 NA NA 1 1 1 2 2 2 2 2 3 2 2 2 2 2 1 1 1 NA NA

## 7 NA NA NA 1 1 1 1 2 2 2 2 2 2 2 1 1 1 1 NA NA NA

## 8 NA NA NA NA 1 1 1 1 1 1 2 1 1 1 1 1 1 NA NA NA NA

## 9 NA NA NA NA NA NA 1 1 1 1 1 1 1 1 1 NA NA NA NA NA NA

## 10 NA NA NA NA NA NA NA NA NA NA 1 NA NA NA NA NA NA NA NA NA NA

Step 4

Replace temp values with averaged expert-elicited weight for each distance band ( ESMdata.zip file containing

f_DFood.csv and other weights derived from 1. Expert elicitation, Step 1 can be requested from the authors)

csv.DFood = read.csv("ESMdata/MWA/f/f_DFood.csv")

m.band[m.band == 1] <- mean(csv.DFood$very.far)

m.band[m.band == 2] <- mean(csv.DFood$far)

m.band[m.band == 3] <- mean(csv.DFood$medium)

m.band[m.band == 4] <- mean(csv.DFood$close)

m.band[m.band == 5] <- mean(csv.DFood$very.close)

## very.close close medium far ve ry.far

## 1.00 0.75 0.58 0.37 0.17

Step 5

Perform moving window analysis using function focal {raster} with parameters ( ESMdata.zip file containing

raster layers with resource quality indices can be requested from the authors):

r = raster("ESMdata/MWA/GIS/Food-quality.tif") # raster layer with numerical resource quali

ty index

w = m.band # moving window is banded weights matrix

fun = max # focal pixel takes highest weighted resource quali ty index within moving window

WARNING! The following process may take several hours depending on the size of r and w

r.D <- focal(r, w, fun, na.rm = TRUE, pad = FALSE, padVal ue = NA) # na.rm = TRUE ignores No

Data

r.SI.D <- mask(r.D, r) # extract by r to remove padded edges

writeRaster(r.SI.D, filename = paste("ESMdata/MWA/out /Food_SI_D1k.tif", sep="")) # save oup

ut raster

3. Composition-dependent resource suitability indicesOverview

For each focal pixel, we computed the average resource quality index within a moving window and reclassified thisaverage value to a numerical suitability index from a reclassification matrix (here: circular window, five equal averagequality classes, radius/reclassifcation matrix elicited from experts in 1. Expert elicitation, Step 2)

Goal

Focal pixel suitability index SIr depends on the amount of a (numerical) habitat variable

Method:

For each focal pixel, compute the average value ( xr ) of a habitat variable within a moving window

reclassify the average value to a numerical focal pixel SIr from a reclassification matrix (here: circular window,

five equal average value classes, radius/reclassifcation matrix elicited from experts in 1. Expert elicitation,Step 2)

Step 1

Perform moving window analysis using the ESRI ArcGIS Focal Statistics tool with parameters:

in_raster = “ESMdata/MWA/GIS/Food-quality.tif”: raster layer with numerical resource quality index( ESMdata.zip file containing raster layers with resource quality indices can be requested from the authors)

neighborhood = NbrCircle, radius = mr: circular moving window with radius = mobility threshold (see above)statistics_type = mean: focal pixel takes average resource quality index within moving windowignore_nodata = TRUE: ignores NoData

## Currently not functioning: implementation using function "focal {raster}" with parameter

s:

r = raster("ESMdata/MWA/GIS/Food-quality.tif") # ra ster layer with numerical resource quali

ty index

w = m.vf # moving window is circular matrix with ra dius = moblity threshold

fun = mean # focal pixel takes average resource qua lity index within moving window

r.C <- focal(r, w, fun, na.rm = TRUE, pad = FALSE, padValue = NA) # na.rm = TRUE ignores No

Data

## returns empty result due to `NA` values at mat rix positions > radius

## changing `NA` to `0` values distorts computati ons of mean value

r.C = raster("ESMdata/MWA/GIS/Food_C1k.tif") # load ArcGIS output raster with average resou

rce quality indices

r.C <- mask(r.C, r) # extract by r to remove padded edges

Step 2

Determine minimum and maximum averaged resource quality index values

fm.min <- minValue(r) # all resource quality indices within the moving wi ndow take minimum

value

fm.max <- maxValue(r) # all resource quality indices within moving window take maximum valu

e

Step 3

Replace averaged resource quality indices with averaged expert-elicited SIr for each average quality class

( ESMdata.zip file containing f_CFood.csv and other SIr derived from 1. Expert elicitation, Step 2 can be

requested from the authors)

csv.CFood = read.csv("ESMdata/MWA/f/f_CFood.csv")

rcl.C <- c(minValue(r.C), fm.min + (fm.max - fm.min) * 0.05 , mean(csv.CFood$X5),

fm.min + (fm.max-fm.min) * 0.05, fm.min + (fm.max-fm.min) * 0.10, mean(csv.CFood$X10),








fm.min + (fm.max-fm.min) * 0.80, fm.min + (fm.max-fm.min) * 1.00, mean(csv.CFood$X100))

m.rcl.C <- matrix(rcl.C, ncol=3, byrow=TRUE) # reclassification matrix (3 columns: "from" /

"to" average quality, "SI" )

## X100 X80 X60 X40 X30 X25 X20 X15 X10 X5

## 90.00 86.67 76.67 60.00 53.33 53.33 33.33 30.00 20.00 10.00

r.SI.C <- reclassify(r.C, m.rcl.C) # apply function "reclassify {raster}"

writeRaster(r.SI.C, filename = paste("ESMdata/MWA/out /Food_SI_C1k.tif", sep="")) # save oup

ut raster

4. Combined composition/distance-dependentresource suitability indicesGoal

Focal pixel suitability index SIr depends on both distance and amount of a (numerical) habitat variable

Method:

Generate a circular moving window where each position is weighted by its distance from the focal pixel(radius/weights derived from 1. Expert elicitation).For each focal pixel, compute the average value ( xr ) of a habitat variable within this weighted moving window

reclassify the average value to a numerical focal pixel SIr from a reclassification matrix (here: circular window,

five equal average value classes, radius/reclassifcation matrix elicited from experts in 1. Expert elicitation,Step 2)

Step 1

Save banded weights matrix (as TXT, add header with number of rows/columns after export (e.g. 21 for matrix with 1km radius)

m.exp <- m.band

m.exp[is.na(m.exp)] <- 0 # change NAs to 0s

write.table(m.exp, file = "ESMdata/MWA/out/m_Food_D1k .txt", row.names = F, col.names = F)

Step 2

Perform moving window analysis using the ESRI ArcGIS Focal Statistics tool with parameters:

in_raster = “ESMdata/MWA/GIS/Food-quality.tif”: raster layer with numerical resource quality index( ESMdata.zip file containing raster layers with resource quality indices can be requested from the authors)

neighborhood = NbrWeight: select exported banded weights matrix .txt

statistics_type = mean: focal pixel takes weighted average resource quality index within moving windowignore_nodata = TRUE: ignores NoData

## Currently not functioning: implementation using function "focal {raster}" with parameter

s:

r = raster("ESMdata/MWA/GIS/Food-quality.tif") # ra ster layer with numerical resource quali

ty index

w = m.band # moving window is banded weights matrix

fun = mean # focal pixel takes average resource qua lity index within moving window

r.DC <- focal(r, w, fun, na.rm = TRUE, pad = FALSE, padValue = NA) # na.rm = TRUE ignores N

oData

## returns empty result due to `NA` values at mat rix positions > radius

## changing `NA` to `0` values distorts computati ons of mean value

r.DC = raster("ESMdata/MWA/GIS/Food_DC1k.tif") # load ArcGIS output raster with distance-we

ighted average resource quality indices

r.DC <- mask(r.DC, r) # extract by r to remove padded edges

Step 3

Determine minimum and maximum distance-weighted averaged resource quality index values as: (number of pixels ineach distance band * (minValue(r)/maxValue(r) * weight for distance band)) / total number of pixels in moving window

m.t <- table(m.vf) # number of pixels in each distance band

fwm.min <- (m.t[5] * minValue(r) * mean(csv.DFood$very.cl ose) + m.t[4] * minValue(r) * mean

(csv.DFood$close) + m.t[3] * minValue(r) * mean(csv.DFood $medium) + m.t[2] * minValue(r) *

mean(csv.DFood$far) + m.t[1] * minValue(r) * mean(csv.DFo od$very.far)) / nnzero(m.band, na.

counted = FALSE)

fwm.max <- (m.t[5] * maxValue(r) * mean(csv.DFood$very.cl ose) + m.t[4] * maxValue(r) * mean

(csv.DFood$close) + m.t[3] * maxValue(r) * mean(csv.DFood $medium) + m.t[2] * maxValue(r) *

mean(csv.DFood$far) + m.t[1] * maxValue(r) * mean(csv.DFo od$very.far)) / nnzero(m.band, na.

counted = FALSE)

as.vector(round(fwm.min, digits = 1)); as.vector(roun d(fwm.max, digits = 1)) # print as rou

nded vector

Here, we computed distance-weighted average minima/maxima as the average across all moving window sizes(radius = 1/2/3 km). The following values were used for analysis:

Water: minimum = 6.1 (xrmin = 11); maximum = 49.5 (xrmax = 89)Food: minimum = 4.5 (xrmin = 11); maximum = 35.9 (xrmax = 88)Heat protection: minimum = 7.4 (xrmin = 15); maximum = 43.6 (xrmax = 89)Disturbance protection: minimum = 3.6 (xrmin = 17); maximum = 18.9 (xrmax = 89)

Step 4

Replace distance-weighted averaged resource quality indices with averaged expert-elicited SIr for each average

quality class ( ESMdata.zip file containing f_CFood.csv and other SIr derived from 1. Expert elicitation, Step 2

can be requested from the authors)

csv.CFood = read.csv("ESMdata/MWA/f/f_CFood.csv")

rcl.DC <- c(minValue(r.DC), fwm.min + (fwm.max - fwm.min) * 0.05, mean(csv.CFood$X5),

fwm.min + (fwm.max - fwm.min) * 0.05, fwm.min + (fwm.max - fw m.min) * 0.10, mean(csv.CFoo

d$X10),


d$X15),


d$X20),


d$X25),


d$X30),


d$X40),


d$X60),


d$X80),


d$X100))

m.rcl.DC <- matrix(rcl.DC, ncol=3, byrow=TRUE) # reclassification matrix (3 columns: "from"

/ "to" distance-weighted average quality, "SI" )

r.SI.DC <- reclassify(r.DC, m.rcl.DC)

writeRaster(r.SI.DC, filename = paste("ESMdata/MWA/ou t/Food_SI_DC1k.tif", sep="")) # save o

uput raster

5. Example plots

References[1] Allaire, J.J. et al. 2016. Package ‘rmarkdown’: dynamic documents for R. URL http://rmarkdown.rstudio.com/(http://rmarkdown.rstudio.com/).

[2] Xie, Y. 2016. Package ‘knitr’: a general-purpose package for dynamic report generation in R. URL http://yihui.name/knitr/ (http://yihui.name/knitr/).

[3] RCoreTeam 2015. R: a language and environment for statistical computing. R Foundation for StatisticalComputing, Vienna, Austria. URL http://www.R-project.org/ (http://www.R-project.org/).

[4] Hijmans, R.J. 2015. Package ‘raster’: geographic data analysis and modeling. URL http://cran.r-project.org/web/packages/raster/ (http://cran.r-project.org/web/packages/raster/).

[5] Bates, D. and Maechler, M. 2015. Package ‘Matrix’: sparse and dense matrix classes and methods. URLhttp://Matrix.R-forge.R-project.org/ (http://Matrix.R-forge.R-project.org/).

[6] Scroggie, M. 2012. Applying a circular moving window filter to raster data in R. URLhttps://scrogster.wordpress.com/2012/10/05/applying-a-circular-moving-window-filter-to-raster-data-in-r/(https://scrogster.wordpress.com/2012/10/05/applying-a-circular-moving-window-filter-to-raster-data-in-r/).

Session information

## Session info ----------------------------------- ---------------------------

## setting value

## version R version 3.1.3 (2015-03-09)

## system x86_64, mingw32

## ui RTerm

## language (EN)

## collate English_Australia.1252

## tz Australia/Brisbane

## date 2017-04-17

## Packages --------------------------------------- ---------------------------

## package * version date source

## devtools 1.10.0 2016-01-23 CRAN (R 3.1.3)

## digest 0.6.8 2014-12-31 CRAN (R 3.1.3)

## evaluate 0.8 2015-09-18 CRAN (R 3.1.3)

## formatR 1.2.1 2015-09-18 CRAN (R 3.1.3)

## htmltools 0.3 2015-12-29 CRAN (R 3.1.3)

## knitr 1.12.3 2016-01-22 CRAN (R 3.1.3)

## lattice 0.20-30 2015-02-22 CRAN (R 3.1.3)

## magrittr 1.5 2014-11-22 CRAN (R 3.1.3)

## Matrix * 1.2-3 2015-11-28 CRAN (R 3.1.3)

## memoise 1.0.0 2016-01-29 CRAN (R 3.1.3)

## raster * 2.4-20 2015-09-08 CRAN (R 3.1.3)

## Rcpp 0.12.1 2015-09-10 CRAN (R 3.1.3)

## rgdal 1.1-1 2015-11-02 CRAN (R 3.1.3)

## rmarkdown 0.9.2 2016-01-01 CRAN (R 3.1.3)

## sp * 1.2-1 2015-10-18 CRAN (R 3.1.3)

## stringi 1.0-1 2015-10-22 CRAN (R 3.1.3)

## stringr 1.0.0 2015-04-30 CRAN (R 3.1.3)

## yaml 2.1.13 2014-06-12 CRAN (R 3.1.3)

180

Appendix 3.3 Methodology – validation


R code - model evaluation and validationJens G. Froese

10 August 2016

This document provides a detailed, reproducible description of the methodology used to evaluate and validateperformance of habitat suitability models using the Continuous Boyce Index. It is Supporting Information to themanuscript:

Froese JG, Smith CS, McAlpine CA, Durr PA, van Klinken RD. Integrating landscapestructure improves habitat models of mobile animals: feral pigs in Australia.

The document is written in R Markdown ([1]) and knitr ([2]), two R ([3]) packages for writing dynamic, reproducible

reports. A .zip file containing data inputs to reproduce analyses can be requested from the authors. Some parts of

the code used to print this document have been surpressed to enhance readability. A generalized version of thesource code is available at URL https://github.com/jgfroese/HSI-CBI-validation (https://github.com/jgfroese/HSI-CBI-validation).

Load required R packagesR packages ecospat ([4]), data.table ([5]) and zoo ([6]) and their dependencies are required for HSI-CBI-

validation. Session information incl. package versions are listed at the bottom of this document.

require(data.table) # for function `setnames`

require(ecospat) # for function `ecospat.boyce`

1. Data preparationStep 1

This R script requires two .TXT files for each model / validation data combination, which have to be prepared in

package {raster} or alternative GIS software as described below. These files can be requested from the authors.

Expected HSI across validation background:define three validation backgrounds ( Balkanu = BLKW , Lakefield = LKNP , NAQS - see Table 3.1)

mask raster layers of the eleven alternative habitat suitability models ( 1 - 10 with PATTSI, X without

PATTSI - see manuscript Table 3.2) by each validation backgroundexport raster attribute tables to .TXT with 3 columns: [ID], [HSI], [pixel count]

1.

Predicted HSI at species presence records:convert presence records of three validation data sets into raster layercombine masked raster layers of the eleven alternative models with each set of presence recordsexport raster attribute tables to .TXT with 5 columns: [ID], [Value], [pixel count], [HSI], [number of

presence records per pixel]

2.

Step 2

Compute model-predicted HSI at feral pig presence records

First, read .TXT files for each model (m1, m2, …) / validation presences (BLKW, LKNP, NAQS) combination as data

frame,

pred.m1.BLKW = read.csv("ESMdata/Validation/BLKW/Pred icted/SiPred_D1_BLKW1000.txt")

...

pred.m10.BLKW = read.csv("ESMdata/Validation/BLKW/Pre dicted/SiPred_BK_BLKW1000.txt")

pred.mX.BLKW = read.csv("ESMdata/Validation/BLKW/Pred icted/SiPred_X_BLKW1000.txt")

...

pred.m10.LKNP = read.csv("ESMdata/Validation/LKNP/Pre dicted/SiPred_BK_LKNP1000.txt")

pred.mX.LKNP = read.csv("ESMdata/Validation/LKNP/Pred icted/SiPred_X_LKNP1000.txt")

...

pred.m10.NAQS = read.csv("ESMdata/Validation/NAQS/Pre dicted/SiPred_BK_NAQS1000.txt")

pred.mX.NAQS = read.csv("ESMdata/Validation/NAQS/Pred icted/SiPred_X_NAQS1000.txt")

and combine all data frames in a list for faster analysis.

pred.list <- list (pred.m1.BLKW, pred.m2.BLKW, pred.m3 .BLKW, pred.m4.BLKW, pred.m5.BLKW, pr

ed.m6.BLKW, pred.m7.BLKW, pred.m8.BLKW, pred.m9.BLK W, pred.m10.BLKW, pred.mX.BLKW, pred.m1.

LKNP, pred.m2.LKNP, pred.m3.LKNP, pred.m4.LKNP, pre d.m5.LKNP, pred.m6.LKNP, pred.m7.LKNP, p

red.m8.LKNP, pred.m9.LKNP, pred.m10.LKNP, pred.mX.L KNP, pred.m1.NAQS, pred.m2.NAQS, pred.m3

.NAQS, pred.m4.NAQS, pred.m5.NAQS, pred.m6.NAQS, pr ed.m7.NAQS, pred.m8.NAQS, pred.m9.NAQS,

pred.m10.NAQS, pred.mX.NAQS)

n.list <- 33 # the number of data frames in your list

Then, homogenise the five column names for all data frames,

for (i in seq_along(pred.list)) {

setnames(pred.list[[i]], c("ID", "Value", "Pixelcou nt", "HSI", "Presences"))

}

and calculate the total number of presence records per HSI value (one pixel may contain multiple records)

pred.sum.list <- vector("list", n.list)


pred.sum.list[[i]] <- aggregate(cbind(Pixelcount*Pres ences)~HSI, data = pred.list[[i]], s

um)

}

Finally, add descriptive column names to the new list of data frames

for (i in seq_along(pred.sum.list)) {

setnames(pred.sum.list[[i]], c("HSI", "Presences"))

}

and convert it into a list of vectors (= HSI at feral pig presence records)

pred.v.list = vector("list", n.list)


pred.v = vector()

for (j in 1:length(pred.sum.list[[i]]$Presences)) {

for (k in 1:pred.sum.list[[i]][j, 2]) {

pred.v <-append(pred.v, pred.sum.list[[i]][j, 1])

}

}

pred.v.list[[i]] <- append(pred.v.list[[i]], pred.v)

}

e.g. HSI of model 1 at Balkanu presence records (first in list)

## [1] 11 11 11 11 11 11 11 11 11 35 40 43 43 43 44 44 44 46 46 46 46 46 48

## [24] 48 48 48 48 48 48 49 49 49 49 49 49 49 49 49 49 49 49 52 53 57 58 58

## [47] 58 58 58 58 58 60 61 61 61 61 61 61 61 61 66 66 66 66 66 66 66 66 66

## [70] 66 67 67 67 67 67 67 67 67 71 71 71 71 71 71 71 71 71 71 71 71 71 71

## [93] 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71

## [116] 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71 71

## [139] 71 71 71 71 71 71 71 71 71 71 71 72 72 75 75 75 75 75 75 75 75 75 75

## [162] 75 75 75 75 75 75 75 75 75 75 75 75 75 75 75 75 75 75 77 77

Step 3

Compute model-predicted HSI across validation backgrounds

First, read .TXT files for each model (m1, m2, …) / validation background (BLKW, LKNP, NAQS) combination as data

frame

exp.m1.BLKW = read.csv("ESMdata/Validation/BLKW/Expec ted/SiExp_D1_BLKW1000.txt")

...

exp.m10.BLKW = read.csv("ESMdata/Validation/BLKW/Expe cted/SiExp_BK_BLKW1000.txt")

exp.mX.BLKW = read.csv("ESMdata/Validation/BLKW/Expec ted/SiExp_X_BLKW1000.txt")

...

exp.m10.LKNP = read.csv("ESMdata/Validation/LKNP/Expe cted/SiExp_BK_LKNP1000.txt")

exp.mX.LKNP = read.csv("ESMdata/Validation/LKNP/Expec ted/SiExp_X_LKNP1000.txt")

...

exp.m10.NAQS = read.csv("ESMdata/Validation/NAQS/Expe cted/SiExp_BK_NAQS1000.txt")

exp.mX.NAQS = read.csv("ESMdata/Validation/NAQS/Expec ted/SiExp_X_NAQS1000.txt")


exp.list <- list (exp.m1.BLKW, exp.m2.BLKW, exp.m3.BLK W, exp.m4.BLKW, exp.m5.BLKW, exp.m6.B

LKW, exp.m7.BLKW, exp.m8.BLKW, exp.m9.BLKW, exp.m10 .BLKW, exp.mX.BLKW, exp.m1.LKNP, exp.m2.

LKNP, exp.m3.LKNP, exp.m4.LKNP, exp.m5.LKNP, exp.m6 .LKNP, exp.m7.LKNP, exp.m8.LKNP, exp.m9.

LKNP, exp.m10.LKNP, exp.mX.LKNP, exp.m1.NAQS, exp.m 2.NAQS, exp.m3.NAQS, exp.m4.NAQS, exp.m5

.NAQS, exp.m6.NAQS, exp.m7.NAQS, exp.m8.NAQS, exp.m 9.NAQS, exp.m10.NAQS, exp.mX.NAQS)

Then, homogenise the three column names for all data frames

for (i in seq_along(exp.list)) {

setnames(exp.list[[i]], c("ID", "HSI", "Pixelcount" ))

}

and convert it into a list of vectors (= HSI across validation backgrounds [potentially very large])

exp.v.list = vector("list", n.list)


exp.v = vector()

for (j in 1:length(exp.list[[i]]$Pixelcount)) {

for (k in 1:exp.list[[i]][j, 3]) {

exp.v <-append(exp.v, exp.list[[i]][j, 2])

}

}

exp.v.list[[i]] <- append(exp.v.list[[i]], exp.v)

}

2. Data analysisStep 4

Apply function boyce {ecospat} to each model (m1, m2, …) / validation data (BLKW, LKNP, NAQS) combination

with parameters:

exp.v.list[[i]] # Expected HSI across background ([[1]] = m1.BLKW, ... , [[33]] = mX.NAQS)

pred.v.list[[i]] # Predicted HSI at presences ([[1]] = m1.BLKW, ... , [[33]] = mX.NAQS)

nclass = 0 # defaults to moving window (continuous, classifica tion-independent) computation

with arguments

window.w = 10 # moving window width (i.e. 10 adjacent HSI values are considered in each com

putation)

res = 100 # resolution factor (i.e. 100 computations across t he total range of HSI)

PEplot = F # no PEplot is generated (customised plots - see Fi g 3.6 and Appendix 3.4)

boyce.m1.BLKW <- ecospat.boyce(exp.v.list[[1]], pred. v.list[[1]], nclass, window.w, res, PE

plot)

...

boyce.m10.BLKW <- ecospat.boyce(exp.v.list[[10]], pre d.v.list[[10]], nclass, window.w, res,

PEplot)

boyce.mX.BLKW <- ecospat.boyce(exp.v.list[[11]], pred .v.list[[11]], nclass, window.w, res,

PEplot)

...

boyce.m10.LKNP <- ecospat.boyce(exp.v.list[[21]], pre d.v.list[[21]], nclass, window.w, res,

PEplot)

boyce.mX.LKNP <- ecospat.boyce(exp.v.list[[22]], pred .v.list[[22]], nclass, window.w, res,

PEplot)

...

boyce.m10.NAQS <- ecospat.boyce(exp.v.list[[32]], pre d.v.list[[32]], nclass, window.w, res,

PEplot)

boyce.mX.NAQS <- ecospat.boyce(exp.v.list[[33]], pred .v.list[[33]], nclass, window.w, res,

PEplot)

Step 5

Investigate results of CBI analysis

Combine all results in list for faster analysis,

boyce.list <- list (boyce.m1.BLKW, boyce.m2.BLKW, boyc e.m3.BLKW, boyce.m4.BLKW, boyce.m5.BL

KW, boyce.m6.BLKW, boyce.m7.BLKW, boyce.m8.BLKW, bo yce.m9.BLKW, boyce.m10.BLKW, boyce.mX.BL

KW, boyce.m1.LKNP, boyce.m2.LKNP, boyce.m3.LKNP, bo yce.m4.LKNP, boyce.m5.LKNP, boyce.m6.LKN

P, boyce.m7.LKNP, boyce.m8.LKNP, boyce.m9.LKNP, boy ce.m10.LKNP, boyce.mX.LKNP, boyce.m1.NAQ

S, boyce.m2.NAQS, boyce.m3.NAQS, boyce.m4.NAQS, boy ce.m5.NAQS, boyce.m6.NAQS, boyce.m7.NAQS

, boyce.m8.NAQS, boyce.m9.NAQS, boyce.m10.NAQS, boy ce.mX.NAQS)

and print CBI ($Spearman.cor) for all model / validation data combinations (see Table 3.2)

CBI.list = vector("list", n.list)

for (i in seq_along(boyce.list)) {

CBI.list[[i]] <- append(CBI.list[[i]], round(boyce.li st[[i]]$Spearman.cor, digits = 2))

}

CBI.list

e.g. CBI of model 1 at Balkanu presence records (first in list)

## [1] 0.86

Step 6

Compute proportion of validation background expected to be highly or very highly suitable habitat (HSI >= 60) for allmodel / validation data combinations (see Table 3.2)

t.HSI <- 59.99 # HSI threshold

HSI.60 <- vector("list", n.list)

for (i in seq_along(HSI.60)) {

HSI.60[[i]] <- aggregate(Pixelcount~HSI > t.HSI, data = e xp.list[[i]], sum) / sum(exp.lis

t[[i]]$Pixelcount)

}

HSI.60.list <- vector("list", n.list)

for (i in seq_along(HSI.60.list)) {

HSI.60.list[[i]] <- append(HSI.60.list[[i]], (round(H SI.60[[i]][2, "Pixelcount"] * 100, d

igits = 0)))

}

HSI.60.list

e.g. HSI >= 60 of model 1 at Balkanu presence records (first in list)

## [1] 34

Plots comparing the P/E ratio of the three validation data sets for each of the eleven alternative habitat suitabilitymodels are provided as Appendix 3.4.

Additional analysis

Fit linear regression model between CBI and HSI >= 60

dat <- read.csv("ESMdata/Validation/lmHSI-CBI.csv")

str(dat)

## 'data.frame': 33 obs. of 2 variables:

## $ HSI: int 2 4 4 5 5 5 5 5 5 5 ...

## $ CBI: num -0.88 -0.14 0.17 0.71 0.47 -0.01 0. 16 -0.25 0.05 0.31 ...

plot(CBI ~ HSI, data = dat) # plot CBI against HSI >= 60

lmodel <- lm(CBI ~ log10(HSI), data = dat) # log transform HSI >= 60

plot(CBI ~ log10(HSI), data = dat); abline(lmodel) # plot again with regression line

summary(lmodel) # summary statistics

##

## Call:

## lm(formula = CBI ~ log10(HSI), data = dat)

##

## Residuals:

## Min 1Q Median 3Q Max

## -0.81480 -0.14624 0.05406 0.19734 0.51376

##

## Coefficients:

## Estimate Std. Error t value Pr(>|t|)

## (Intercept) -0.3829 0.1718 -2.229 0.0332 *

## log10(HSI) 0.8286 0.1487 5.573 4.15e-06 ***

## ---

## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

##

## Residual standard error: 0.332 on 31 degrees of freedom

## Multiple R-squared: 0.5005, Adjusted R-squared: 0.4844

## F-statistic: 31.06 on 1 and 31 DF, p-value: 4.1 5e-06

par(mfrow = c(2,2))

plot(lmodel) # plot model diagnostics



[3] RCoreTeam 2015. R: a language and environment for statistical computing. R Foundation for StatisticalComputing, Vienna, Austria. URL http://www.R-project.org/ (http://www.R-project.org/).

[4] Broennimann, O. 2015. Package ‘ecospat’: spatial ecology miscellaneous methods. URL http://cran.r-project.org/web/packages/ecospat/ (http://cran.r-project.org/web/packages/ecospat/).

[5] Dowle, M. et al. 2015. Package ‘data.table’: extension of data.frame. URL https://github.com/Rdatatable/data.table/wiki/ (https://github.com/Rdatatable/data.table/wiki/).

[6] Zeileis, A. et al. 2015. Package ‘zoo’: S3 infrastructure for regular and irregular time series. URL http://zoo.R-Forge.R-project.org/ (http://zoo.R-Forge.R-project.org/).

Session information

## Session info ----------------------------------- ---------------------------

## setting value



## ui RTerm

## language (EN)



## date 2017-04-17

## Packages --------------------------------------- ---------------------------


## abind 1.4-3 2015-03-13 CRAN (R 3.1.3 )

## acepack 1.3-3.3 2013-05-03 CRAN (R 3.1.3 )

## ade4 * 1.7-2 2015-04-14 CRAN (R 3.1.3 )

## adehabitatHR 0.4.14 2015-07-22 CRAN (R 3.1.3 )

## adehabitatLT 0.3.20 2015-07-22 CRAN (R 3.1.3 )

## adehabitatMA 0.3.10 2015-07-22 CRAN (R 3.1.3 )

## ape * 3.3 2015-05-29 CRAN (R 3.1.3 )

## biomod2 3.1-64 2014-12-10 CRAN (R 3.1.3 )

## boot 1.3-15 2015-02-24 CRAN (R 3.1.3 )

## chron 2.3-47 2015-06-24 CRAN (R 3.1.3 )

## CircStats 0.2-4 2012-10-29 CRAN (R 3.1.3 )

## class 7.3-12 2015-02-11 CRAN (R 3.1.3 )

## cluster 2.0.1 2015-01-31 CRAN (R 3.1.3 )

## codetools 0.2-10 2015-01-17 CRAN (R 3.1.3 )

## colorspace 1.2-6 2015-03-11 CRAN (R 3.1.3 )

## data.table * 1.9.6 2015-09-19 CRAN (R 3.1.3 )

## deldir 0.1-9 2015-03-09 CRAN (R 3.1.3 )

## devtools 1.10.0 2016-01-23 CRAN (R 3.1.3 )

## digest 0.6.8 2014-12-31 CRAN (R 3.1.3 )

## dismo 1.0-12 2015-03-15 CRAN (R 3.1.3 )

## ecodist 1.2.9 2013-12-03 CRAN (R 3.1.3 )

## ecospat * 1.1 2015-03-06 CRAN (R 3.1.3 )

## evaluate 0.8 2015-09-18 CRAN (R 3.1.3 )

## foreach * 1.4.3 2015-10-13 CRAN (R 3.1.3 )

## foreign 0.8-63 2015-02-20 CRAN (R 3.1.3 )

## formatR 1.2.1 2015-09-18 CRAN (R 3.1.3 )

## Formula 1.2-1 2015-04-07 CRAN (R 3.1.3 )

## gam * 1.12 2015-05-13 CRAN (R 3.1.3 )

## gbm * 2.1.1 2015-03-11 CRAN (R 3.1.3 )

## ggplot2 2.1.0 2016-03-01 CRAN (R 3.1.3 )

## goftest 1.0-3 2015-07-03 CRAN (R 3.1.3 )

## gridExtra 2.0.0 2015-07-14 CRAN (R 3.1.3 )

## gtable 0.1.2 2012-12-05 CRAN (R 3.1.3 )

## hexbin 1.27.1 2015-08-19 CRAN (R 3.1.3 )

## Hmisc 3.17-0 2015-09-21 CRAN (R 3.1.3 )

## htmltools 0.3 2015-12-29 CRAN (R 3.1.3 )

## iterators 1.0.8 2015-10-13 CRAN (R 3.1.3 )

## knitr 1.12.3 2016-01-22 CRAN (R 3.1.3 )

## lattice * 0.20-30 2015-02-22 CRAN (R 3.1.3 )

## latticeExtra 0.6-26 2013-08-15 CRAN (R 3.1.3 )

## magrittr 1.5 2014-11-22 CRAN (R 3.1.3 )

## maptools 0.8-37 2015-09-29 CRAN (R 3.1.3 )

## MASS 7.3-39 2015-02-24 CRAN (R 3.1.3 )

## Matrix 1.2-3 2015-11-28 CRAN (R 3.1.3 )

## MatrixModels 0.4-1 2015-08-22 CRAN (R 3.1.3 )

## mda 0.4-7 2015-05-25 CRAN (R 3.1.3 )

## memoise 1.0.0 2016-01-29 CRAN (R 3.1.3 )

## mgcv 1.8-4 2014-11-27 CRAN (R 3.1.3 )

## multcomp 1.4-1 2015-07-23 CRAN (R 3.1.3 )

## munsell 0.4.2 2013-07-11 CRAN (R 3.1.3 )

## mvtnorm 1.0-3 2015-07-22 CRAN (R 3.1.3 )

## nlme 3.1-120 2015-02-20 CRAN (R 3.1.3 )

## nnet 7.3-9 2015-02-11 CRAN (R 3.1.3 )

## plyr 1.8.3 2015-06-12 CRAN (R 3.1.3 )

## polspline 1.1.12 2015-07-14 CRAN (R 3.1.3 )

## polyclip 1.3-2 2015-05-27 CRAN (R 3.1.3 )

## pROC 1.8 2015-05-05 CRAN (R 3.1.3 )

## proto 0.3-10 2012-12-22 CRAN (R 3.1.3 )

## quantreg 5.19 2015-08-31 CRAN (R 3.1.3 )

## randomForest 4.6-12 2015-10-07 CRAN (R 3.1.3 )

## raster 2.4-20 2015-09-08 CRAN (R 3.1.3 )

## rasterVis 0.37 2015-09-06 CRAN (R 3.1.3 )

## RColorBrewer 1.1-2 2014-12-07 CRAN (R 3.1.3 )

## Rcpp 0.12.1 2015-09-10 CRAN (R 3.1.3 )

## reshape 0.8.5 2014-04-23 CRAN (R 3.1.3 )

## rmarkdown 0.9.2 2016-01-01 CRAN (R 3.1.3 )

## rms 4.4-0 2015-09-28 CRAN (R 3.1.3 )

## rpart 4.1-9 2015-02-24 CRAN (R 3.1.3 )

## sandwich 2.3-4 2015-09-24 CRAN (R 3.1.3 )

## scales 0.3.0 2015-08-25 CRAN (R 3.1.3 )

## sp * 1.2-1 2015-10-18 CRAN (R 3.1.3 )

## SparseM 1.7 2015-08-15 CRAN (R 3.1.3 )

## spatstat 1.41-1 2015-02-27 CRAN (R 3.1.3 )

## stringi 1.0-1 2015-10-22 CRAN (R 3.1.3 )

## stringr 1.0.0 2015-04-30 CRAN (R 3.1.3 )

## survival * 2.38-1 2015-02-24 CRAN (R 3.1.3 )

## tensor 1.5 2012-05-05 CRAN (R 3.1.3 )

## TH.data 1.0-6 2015-01-05 CRAN (R 3.1.3 )

## yaml 2.1.13 2014-06-12 CRAN (R 3.1.3 )

## zoo 1.7-12 2015-03-16 CRAN (R 3.1.3 )

190

Appendix 3.4 Validation plots – model scenarios

Appendix Fig 3.4 Validation plots for eleven habitat suitability models (panels A-K, refer to Table 3.2

for corresponding landscape metrics and scales of measurement). The predicted-to-expected (P/E) ratio is

plotted for each validation data set (black lines) and as a smoothed average of all data sets (red line).

191

Appendix 4.1 Model variables – parameters (supplements Fig 4.2)

Appendix Table 4.1 Bayesian network conditional probability tables. Used to model resource quality

indices for water (A), food (B), protection from heat stress (C) and protection from disturbance (D1 and D2)

as well as an index of habitat suitability for feral pig breeding and persistence (E*) as shown in Fig 4.2.

* Due to size constraints in this Appendix, the CPT for the habitat suitability index is shown for only one expert (Expert

1) for illustrative purposes. Full CPTs can be requested from the author in .xlsx format.

(A) Expert Freshwater presence

Terrain ruggedness

Water quality index (% probability) Very good Good Moderate Poor Very poor

Expert1 Yes Low 97 3 0 0 0 Expert1 Yes Moderate 50 40 10 0 0 Expert1 Yes High 3 23 48 23 3 Expert1 No Low 0 0 0 16 84 Expert1 No Moderate 0 0 0 16 84 Expert1 No High 0 0 0 16 84 Expert2 Yes Low 97 3 0 0 0 Expert2 Yes Moderate 97 3 0 0 0 Expert2 Yes High 33 33 23 9 2 Expert2 No Low 0 0 0 3 97 Expert2 No Moderate 0 0 0 3 97 Expert2 No High 0 0 0 3 97 Expert3 Yes Low 97 3 0 0 0 Expert3 Yes Moderate 97 3 0 0 0 Expert3 Yes High 33 33 23 9 2 Expert3 No Low 0 0 0 3 97 Expert3 No Moderate 0 0 0 3 97 Expert3 No High 0 0 0 3 97 Expert4 Yes Low 97 3 0 0 0 Expert4 Yes Moderate 97 3 0 0 0 Expert4 Yes High 3 47 47 3 0 Expert4 No Low 0 0 0 3 97 Expert4 No Moderate 0 0 0 3 97 Expert4 No High 0 0 0 3 97

192

(B) Expert Food quality

Food quantity

Food quality index (% probability) Very good Good Moderate Poor Very poor

Expert1 High High 68 29 3 0 0 Expert1 High Moderate 43 46 11 0 0 Expert1 High Low 29 50 20 1 0 Expert1 Moderate High 33 52 15 0 0 Expert1 Moderate Moderate 2 31 51 15 1 Expert1 Moderate Low 0 15 47 34 4 Expert1 Low High 13 50 34 3 0 Expert1 Low Moderate 0 8 44 42 6 Expert1 Low Low 0 0 0 3 97 Expert2 High High 97 3 0 0 0 Expert2 High Moderate 36 50 14 0 0 Expert2 High Low 17 51 30 2 0 Expert2 Moderate High 15 54 29 2 0 Expert2 Moderate Moderate 1 26 54 18 1 Expert2 Moderate Low 0 9 45 40 6 Expert2 Low High 0 0 0 26 74 Expert2 Low Moderate 0 0 0 15 85 Expert2 Low Low 0 0 0 3 97 Expert3 High High 97 3 0 0 0 Expert3 High Moderate 36 50 14 0 0 Expert3 High Low 17 51 30 2 0 Expert3 Moderate High 15 54 29 2 0 Expert3 Moderate Moderate 1 26 54 18 1 Expert3 Moderate Low 0 9 45 40 6 Expert3 Low High 0 3 23 48 26 Expert3 Low Moderate 0 0 12 26 62 Expert3 Low Low 0 0 0 3 97 Expert4 High High 97 3 0 0 0 Expert4 High Moderate 33 37 22 7 1 Expert4 High Low 23 35 28 12 2 Expert4 Moderate High 23 35 28 12 2 Expert4 Moderate Moderate 10 26 33 23 8 Expert4 Moderate Low 5 18 32 30 15 Expert4 Low High 9 24 34 24 9 Expert4 Low Moderate 2 12 28 35 23 Expert4 Low Low 0 0 0 3 97 Expert5 High High 97 3 0 0 0 Expert5 High Moderate 50 47 3 0 0 Expert5 High Low 3 47 47 3 0 Expert5 Moderate High 3 47 47 3 0 Expert5 Moderate Moderate 1 27 53 18 1 Expert5 Moderate Low 0 12 47 36 5 Expert5 Low High 0 0 3 47 50 Expert5 Low Moderate 0 0 2 25 73 Expert5 Low Low 0 0 0 3 97

193

(C) Expert Heat stress Shady vegetation cover

Heat protection quality index (% probability) Very good Good Moderate Poor Very poor

Expert1 Low Good 97 3 0 0 0 Expert1 Low Moderate 97 3 0 0 0 Expert1 Low Poor 97 3 0 0 0 Expert1 Moderate Good 65 33 2 0 0 Expert1 Moderate Moderate 26 48 23 3 0 Expert1 Moderate Poor 3 23 48 23 3 Expert1 High Good 26 47 24 3 0 Expert1 High Moderate 20 35 40 4 1 Expert1 High Poor 0 0 0 3 97 Expert2 Low Good 97 3 0 0 0 Expert2 Low Moderate 97 3 0 0 0 Expert2 Low Poor 97 3 0 0 0 Expert2 Moderate Good 97 3 0 0 0 Expert2 Moderate Moderate 3 47 47 3 0 Expert2 Moderate Poor 0 3 47 47 3 Expert2 High Good 97 3 0 0 0 Expert2 High Moderate 0 16 68 16 0 Expert2 High Poor 0 0 0 3 97 Expert3 Low Good 97 3 0 0 0 Expert3 Low Moderate 97 3 0 0 0 Expert3 Low Poor 97 3 0 0 0 Expert3 Moderate Good 97 3 0 0 0 Expert3 Moderate Moderate 62 26 11 1 0 Expert3 Moderate Poor 26 48 23 3 0 Expert3 High Good 97 3 0 0 0 Expert3 High Moderate 26 48 23 3 0 Expert3 High Poor 0 0 0 3 97 Expert4 Low Good 97 3 0 0 0 Expert4 Low Moderate 97 3 0 0 0 Expert4 Low Poor 97 3 0 0 0 Expert4 Moderate Good 97 3 0 0 0 Expert4 Moderate Moderate 3 47 47 3 0 Expert4 Moderate Poor 0 3 47 47 3 Expert4 High Good 97 3 0 0 0 Expert4 High Moderate 3 23 48 23 3 Expert4 High Poor 0 3 23 48 26 Expert5 Low Good 97 3 0 0 0 Expert5 Low Moderate 97 3 0 0 0 Expert5 Low Poor 97 3 0 0 0 Expert5 Moderate Good 97 3 0 0 0 Expert5 Moderate Moderate 16 68 16 0 0 Expert5 Moderate Poor 0 16 68 16 0 Expert5 High Good 97 3 0 0 0 Expert5 High Moderate 3 47 47 3 0 Expert5 High Poor 0 0 0 3 97

194

(D1) Expert Disturbance stress

Dense vegetation cover

Disturbance protection quality index (% probability) Very good Good Moderate Poor Very poor

Expert1 Low Good 97 3 0 0 0 Expert1 Low Moderate 97 3 0 0 0 Expert1 Low Poor 97 3 0 0 0 Expert1 Moderate Good 74 26 0 0 0 Expert1 Moderate Moderate 26 71 3 0 0 Expert1 Moderate Poor 0 16 68 16 0 Expert1 High Good 74 26 0 0 0 Expert1 High Moderate 0 1 53 45 1 Expert1 High Poor 0 0 0 3 97 Expert2 Low Good 97 3 0 0 0 Expert2 Low Moderate 97 3 0 0 0 Expert2 Low Poor 97 3 0 0 0 Expert2 Moderate Good 97 3 0 0 0 Expert2 Moderate Moderate 3 47 47 3 0 Expert2 Moderate Poor 0 16 68 16 0 Expert2 High Good 16 68 16 0 0 Expert2 High Moderate 0 3 47 47 3 Expert2 High Poor 0 0 0 3 97 Expert4 Low Good 97 3 0 0 0 Expert4 Low Moderate 97 3 0 0 0 Expert4 Low Poor 97 3 0 0 0 Expert4 Moderate Good 97 3 0 0 0 Expert4 Moderate Moderate 50 47 3 0 0 Expert4 Moderate Poor 0 3 47 47 3 Expert4 High Good 74 26 0 0 0 Expert4 High Moderate 0 50 50 0 0 Expert4 High Poor 0 0 3 95 2 Expert5 Low Good 97 3 0 0 0 Expert5 Low Moderate 97 3 0 0 0 Expert5 Low Poor 97 3 0 0 0 Expert5 Moderate Good 97 3 0 0 0 Expert5 Moderate Moderate 26 48 23 3 0 Expert5 Moderate Poor 3 23 48 23 3 Expert5 High Good 3 47 47 3 0 Expert5 High Moderate 0 3 47 47 3 Expert5 High Poor 0 0 0 3 97 Expert6 Low Good 97 3 0 0 0 Expert6 Low Moderate 97 3 0 0 0 Expert6 Low Poor 97 3 0 0 0 Expert6 Moderate Good 97 3 0 0 0 Expert6 Moderate Moderate 86 14 0 0 0 Expert6 Moderate Poor 74 26 0 0 0 Expert6 High Good 74 26 0 0 0 Expert6 High Moderate 3 47 47 3 0 Expert6 High Poor 0 0 0 3 97

195

(D2) Expert Intensity of control

Frequency of control

Predator presence

Disturbance stress (% probability) Low Moderate High

Expert1 Low Low Low 79 21 0 Expert1 Low Low High 56 44 0 Expert1 Low High Low 3 86 11 Expert1 Low High High 0 71 29 Expert1 High Low Low 33 67 0 Expert1 High Low High 13 85 2 Expert1 High High Low 0 24 76 Expert1 High High High 0 24 76 Expert2 Low Low Low 100 0 0 Expert2 Low Low High 50 48 2 Expert2 Low High Low 0 50 50 Expert2 Low High High 0 49 51 Expert2 High Low Low 50 50 0 Expert2 High Low High 25 73 2 Expert2 High High Low 0 0 100 Expert2 High High High 0 0 100 Expert4 Low Low Low 97 3 0 Expert4 Low Low High 97 3 0 Expert4 Low High Low 50 48 2 Expert4 Low High High 50 48 2 Expert4 High Low Low 2 48 50 Expert4 High Low High 2 48 50 Expert4 High High Low 0 3 97 Expert4 High High High 0 3 97 Expert5 Low Low Low 97 3 0 Expert5 Low Low High 97 3 0 Expert5 Low High Low 0 50 50 Expert5 Low High High 0 50 50 Expert5 High Low Low 50 50 0 Expert5 High Low High 50 50 0 Expert5 High High Low 0 3 97 Expert5 High High High 0 3 97 Expert6 Low Low Low 97 3 0 Expert6 Low Low High 50 48 2 Expert6 Low High Low 2 48 50 Expert6 Low High High 1 48 51 Expert6 High Low Low 97 3 0 Expert6 High Low High 50 48 2 Expert6 High High Low 0 3 97 Expert6 High High High 0 3 97

197

Expert1 Very good Moderate Very good Very good 12 62 25 1 0 Expert1 Very good Moderate Very good Good 7 58 34 1 0 Expert1 Very good Moderate Very good Moderate 5 51 42 2 0 Expert1 Very good Moderate Very good Poor 3 42 51 4 0 Expert1 Very good Moderate Very good Very poor 0 3 23 49 25 Expert1 Very good Moderate Good Very good 7 58 34 1 0 Expert1 Very good Moderate Good Good 5 51 42 2 0 Expert1 Very good Moderate Good Moderate 3 42 51 4 0 Expert1 Very good Moderate Good Poor 1 33 58 8 0 Expert1 Very good Moderate Good Very poor 0 3 23 49 25 Expert1 Very good Moderate Moderate Very good 3 46 48 3 0 Expert1 Very good Moderate Moderate Good 2 37 56 5 0 Expert1 Very good Moderate Moderate Moderate 1 28 61 10 0 Expert1 Very good Moderate Moderate Poor 0 21 64 15 0 Expert1 Very good Moderate Moderate Very poor 0 3 23 49 25 Expert1 Very good Moderate Poor Very good 0 18 64 18 0 Expert1 Very good Moderate Poor Good 0 12 62 25 1 Expert1 Very good Moderate Poor Moderate 0 8 58 33 1 Expert1 Very good Moderate Poor Poor 0 5 51 42 2 Expert1 Very good Moderate Poor Very poor 0 3 23 49 25 Expert1 Very good Moderate Very poor Very good 0 0 0 26 74 Expert1 Very good Moderate Very poor Good 0 0 0 26 74 Expert1 Very good Moderate Very poor Moderate 0 0 0 26 74 Expert1 Very good Moderate Very poor Poor 0 0 0 26 74 Expert1 Very good Moderate Very poor Very poor 0 0 0 26 74 Expert1 Very good Poor Very good Very good 2 33 58 7 0 Expert1 Very good Poor Very good Good 1 25 62 12 0 Expert1 Very good Poor Very good Moderate 0 18 64 18 0 Expert1 Very good Poor Very good Poor 0 12 62 25 1 Expert1 Very good Poor Very good Very poor 0 3 23 49 25 Expert1 Very good Poor Good Very good 1 25 62 12 0 Expert1 Very good Poor Good Good 0 18 64 18 0 Expert1 Very good Poor Good Moderate 0 12 62 25 1 Expert1 Very good Poor Good Poor 0 8 58 33 1 Expert1 Very good Poor Good Very poor 0 3 23 49 25 Expert1 Very good Poor Moderate Very good 0 18 64 18 0 Expert1 Very good Poor Moderate Good 0 12 62 25 1 Expert1 Very good Poor Moderate Moderate 0 8 58 33 1 Expert1 Very good Poor Moderate Poor 0 5 51 42 2 Expert1 Very good Poor Moderate Very poor 0 3 23 49 25 Expert1 Very good Poor Poor Very good 0 9 60 30 1 Expert1 Very good Poor Poor Good 0 5 54 39 2 Expert1 Very good Poor Poor Moderate 0 3 46 48 3 Expert1 Very good Poor Poor Poor 0 2 37 56 5 Expert1 Very good Poor Poor Very poor 0 3 23 49 25 Expert1 Very good Poor Very poor Very good 0 0 0 26 74 Expert1 Very good Poor Very poor Good 0 0 0 26 74 Expert1 Very good Poor Very poor Moderate 0 0 0 26 74 Expert1 Very good Poor Very poor Poor 0 0 0 26 74 Expert1 Very good Poor Very poor Very poor 0 0 0 26 74

198

Expert1 Very good Very poor Very good Very good 0 0 0 26 74 Expert1 Very good Very poor Very good Good 0 0 0 26 74 Expert1 Very good Very poor Very good Moderate 0 0 0 26 74 Expert1 Very good Very poor Very good Poor 0 0 0 26 74 Expert1 Very good Very poor Very good Very poor 0 0 0 26 74 Expert1 Very good Very poor Good Very good 0 0 0 26 74 Expert1 Very good Very poor Good Good 0 0 0 26 74 Expert1 Very good Very poor Good Moderate 0 0 0 26 74 Expert1 Very good Very poor Good Poor 0 0 0 26 74 Expert1 Very good Very poor Good Very poor 0 0 0 26 74 Expert1 Very good Very poor Moderate Very good 0 0 0 26 74 Expert1 Very good Very poor Moderate Good 0 0 0 26 74 Expert1 Very good Very poor Moderate Moderate 0 0 0 26 74 Expert1 Very good Very poor Moderate Poor 0 0 0 26 74 Expert1 Very good Very poor Moderate Very poor 0 0 0 26 74 Expert1 Very good Very poor Poor Very good 0 0 0 26 74 Expert1 Very good Very poor Poor Good 0 0 0 26 74 Expert1 Very good Very poor Poor Moderate 0 0 0 26 74 Expert1 Very good Very poor Poor Poor 0 0 0 26 74 Expert1 Very good Very poor Poor Very poor 0 0 0 26 74 Expert1 Very good Very poor Very poor Very good 0 0 0 26 74 Expert1 Very good Very poor Very poor Good 0 0 0 26 74 Expert1 Very good Very poor Very poor Moderate 0 0 0 26 74 Expert1 Very good Very poor Very poor Poor 0 0 0 26 74 Expert1 Very good Very poor Very poor Very poor 0 0 0 26 74 Expert1 Good Very good Very good Very good 44 52 4 0 0 Expert1 Good Very good Very good Good 34 58 8 0 0 Expert1 Good Very good Very good Moderate 25 63 12 0 0 Expert1 Good Very good Very good Poor 15 65 20 0 0 Expert1 Good Very good Very good Very poor 0 3 23 49 25 Expert1 Good Very good Good Very good 28 62 10 0 0 Expert1 Good Very good Good Good 20 65 15 0 0 Expert1 Good Very good Good Moderate 14 64 22 0 0 Expert1 Good Very good Good Poor 7 59 33 1 0 Expert1 Good Very good Good Very poor 0 3 23 49 25 Expert1 Good Very good Moderate Very good 7 58 34 1 0 Expert1 Good Very good Moderate Good 5 51 42 2 0 Expert1 Good Very good Moderate Moderate 3 42 51 4 0 Expert1 Good Very good Moderate Poor 1 33 58 8 0 Expert1 Good Very good Moderate Very poor 0 3 23 49 25 Expert1 Good Very good Poor Very good 1 25 62 12 0 Expert1 Good Very good Poor Good 0 18 64 18 0 Expert1 Good Very good Poor Moderate 0 12 62 25 1 Expert1 Good Very good Poor Poor 0 8 58 33 1 Expert1 Good Very good Poor Very poor 0 3 23 49 25 Expert1 Good Very good Very poor Very good 0 0 0 26 74 Expert1 Good Very good Very poor Good 0 0 0 26 74 Expert1 Good Very good Very poor Moderate 0 0 0 26 74 Expert1 Good Very good Very poor Poor 0 0 0 26 74 Expert1 Good Very good Very poor Very poor 0 0 0 26 74

199

Expert1 Good Good Very good Very good 28 62 10 0 0 Expert1 Good Good Very good Good 20 65 15 0 0 Expert1 Good Good Very good Moderate 14 64 22 0 0 Expert1 Good Good Very good Poor 7 59 33 1 0 Expert1 Good Good Very good Very poor 0 3 23 49 25 Expert1 Good Good Good Very good 18 65 17 0 0 Expert1 Good Good Good Good 12 64 24 0 0 Expert1 Good Good Good Moderate 7 58 34 1 0 Expert1 Good Good Good Poor 3 47 47 3 0 Expert1 Good Good Good Very poor 0 3 23 49 25 Expert1 Good Good Moderate Very good 5 51 42 2 0 Expert1 Good Good Moderate Good 3 42 51 4 0 Expert1 Good Good Moderate Moderate 1 33 58 8 0 Expert1 Good Good Moderate Poor 1 24 63 12 0 Expert1 Good Good Moderate Very poor 0 3 23 49 25 Expert1 Good Good Poor Very good 0 18 64 18 0 Expert1 Good Good Poor Good 0 12 62 25 1 Expert1 Good Good Poor Moderate 0 8 58 33 1 Expert1 Good Good Poor Poor 0 5 51 42 2 Expert1 Good Good Poor Very poor 0 3 23 49 25 Expert1 Good Good Very poor Very good 0 0 0 26 74 Expert1 Good Good Very poor Good 0 0 0 26 74 Expert1 Good Good Very poor Moderate 0 0 0 26 74 Expert1 Good Good Very poor Poor 0 0 0 26 74 Expert1 Good Good Very poor Very poor 0 0 0 26 74 Expert1 Good Moderate Very good Very good 7 58 34 1 0 Expert1 Good Moderate Very good Good 5 51 42 2 0 Expert1 Good Moderate Very good Moderate 3 42 51 4 0 Expert1 Good Moderate Very good Poor 1 33 58 8 0 Expert1 Good Moderate Very good Very poor 0 3 23 49 25 Expert1 Good Moderate Good Very good 5 51 42 2 0 Expert1 Good Moderate Good Good 3 42 51 4 0 Expert1 Good Moderate Good Moderate 1 33 58 8 0 Expert1 Good Moderate Good Poor 1 24 63 12 0 Expert1 Good Moderate Good Very poor 0 3 23 49 25 Expert1 Good Moderate Moderate Very good 2 37 56 5 0 Expert1 Good Moderate Moderate Good 1 28 61 10 0 Expert1 Good Moderate Moderate Moderate 0 21 64 15 0 Expert1 Good Moderate Moderate Poor 0 14 64 22 0 Expert1 Good Moderate Moderate Very poor 0 3 23 49 25 Expert1 Good Moderate Poor Very good 0 12 62 25 1 Expert1 Good Moderate Poor Good 0 8 58 33 1 Expert1 Good Moderate Poor Moderate 0 5 51 42 2 Expert1 Good Moderate Poor Poor 0 3 42 51 4 Expert1 Good Moderate Poor Very poor 0 3 23 49 25 Expert1 Good Moderate Very poor Very good 0 0 0 26 74 Expert1 Good Moderate Very poor Good 0 0 0 26 74 Expert1 Good Moderate Very poor Moderate 0 0 0 26 74 Expert1 Good Moderate Very poor Poor 0 0 0 26 74 Expert1 Good Moderate Very poor Very poor 0 0 0 26 74

200

Expert1 Good Poor Very good Very good 1 25 62 12 0 Expert1 Good Poor Very good Good 0 18 64 18 0 Expert1 Good Poor Very good Moderate 0 12 62 25 1 Expert1 Good Poor Very good Poor 0 8 58 33 1 Expert1 Good Poor Very good Very poor 0 3 23 49 25 Expert1 Good Poor Good Very good 0 18 64 18 0 Expert1 Good Poor Good Good 0 12 62 25 1 Expert1 Good Poor Good Moderate 0 8 58 33 1 Expert1 Good Poor Good Poor 0 5 51 42 2 Expert1 Good Poor Good Very poor 0 3 23 49 25 Expert1 Good Poor Moderate Very good 0 12 62 25 1 Expert1 Good Poor Moderate Good 0 8 58 33 1 Expert1 Good Poor Moderate Moderate 0 5 51 42 2 Expert1 Good Poor Moderate Poor 0 3 42 51 4 Expert1 Good Poor Moderate Very poor 0 3 23 49 25 Expert1 Good Poor Poor Very good 0 5 54 39 2 Expert1 Good Poor Poor Good 0 3 46 48 3 Expert1 Good Poor Poor Moderate 0 2 37 56 5 Expert1 Good Poor Poor Poor 0 1 28 62 9 Expert1 Good Poor Poor Very poor 0 3 23 49 25 Expert1 Good Poor Very poor Very good 0 0 0 26 74 Expert1 Good Poor Very poor Good 0 0 0 26 74 Expert1 Good Poor Very poor Moderate 0 0 0 26 74 Expert1 Good Poor Very poor Poor 0 0 0 26 74 Expert1 Good Poor Very poor Very poor 0 0 0 26 74 Expert1 Good Very poor Very good Very good 0 0 0 26 74 Expert1 Good Very poor Very good Good 0 0 0 26 74 Expert1 Good Very poor Very good Moderate 0 0 0 26 74 Expert1 Good Very poor Very good Poor 0 0 0 26 74 Expert1 Good Very poor Very good Very poor 0 0 0 26 74 Expert1 Good Very poor Good Very good 0 0 0 26 74 Expert1 Good Very poor Good Good 0 0 0 26 74 Expert1 Good Very poor Good Moderate 0 0 0 26 74 Expert1 Good Very poor Good Poor 0 0 0 26 74 Expert1 Good Very poor Good Very poor 0 0 0 26 74 Expert1 Good Very poor Moderate Very good 0 0 0 26 74 Expert1 Good Very poor Moderate Good 0 0 0 26 74 Expert1 Good Very poor Moderate Moderate 0 0 0 26 74 Expert1 Good Very poor Moderate Poor 0 0 0 26 74 Expert1 Good Very poor Moderate Very poor 0 0 0 26 74 Expert1 Good Very poor Poor Very good 0 0 0 26 74 Expert1 Good Very poor Poor Good 0 0 0 26 74 Expert1 Good Very poor Poor Moderate 0 0 0 26 74 Expert1 Good Very poor Poor Poor 0 0 0 26 74 Expert1 Good Very poor Poor Very poor 0 0 0 26 74 Expert1 Good Very poor Very poor Very good 0 0 0 26 74 Expert1 Good Very poor Very poor Good 0 0 0 26 74 Expert1 Good Very poor Very poor Moderate 0 0 0 26 74 Expert1 Good Very poor Very poor Poor 0 0 0 26 74 Expert1 Good Very poor Very poor Very poor 0 0 0 26 74

201

Expert1 Moderate Very good Very good Very good 12 62 25 1 0 Expert1 Moderate Very good Very good Good 7 58 34 1 0 Expert1 Moderate Very good Very good Moderate 5 51 42 2 0 Expert1 Moderate Very good Very good Poor 3 42 51 4 0 Expert1 Moderate Very good Very good Very poor 0 3 23 49 25 Expert1 Moderate Very good Good Very good 7 58 34 1 0 Expert1 Moderate Very good Good Good 5 51 42 2 0 Expert1 Moderate Very good Good Moderate 3 42 51 4 0 Expert1 Moderate Very good Good Poor 1 33 58 8 0 Expert1 Moderate Very good Good Very poor 0 3 23 49 25 Expert1 Moderate Very good Moderate Very good 3 46 48 3 0 Expert1 Moderate Very good Moderate Good 2 37 56 5 0 Expert1 Moderate Very good Moderate Moderate 1 28 61 10 0 Expert1 Moderate Very good Moderate Poor 0 21 64 15 0 Expert1 Moderate Very good Moderate Very poor 0 3 23 49 25 Expert1 Moderate Very good Poor Very good 0 18 64 18 0 Expert1 Moderate Very good Poor Good 0 12 62 25 1 Expert1 Moderate Very good Poor Moderate 0 8 58 33 1 Expert1 Moderate Very good Poor Poor 0 5 51 42 2 Expert1 Moderate Very good Poor Very poor 0 3 23 49 25 Expert1 Moderate Very good Very poor Very good 0 0 0 26 74 Expert1 Moderate Very good Very poor Good 0 0 0 26 74 Expert1 Moderate Very good Very poor Moderate 0 0 0 26 74 Expert1 Moderate Very good Very poor Poor 0 0 0 26 74 Expert1 Moderate Very good Very poor Very poor 0 0 0 26 74 Expert1 Moderate Good Very good Very good 8 58 34 0 0 Expert1 Moderate Good Very good Good 5 51 42 2 0 Expert1 Moderate Good Very good Moderate 3 42 51 4 0 Expert1 Moderate Good Very good Poor 1 33 58 8 0 Expert1 Moderate Good Very good Very poor 0 3 23 49 25 Expert1 Moderate Good Good Very good 5 51 42 2 0 Expert1 Moderate Good Good Good 3 42 51 4 0 Expert1 Moderate Good Good Moderate 1 33 58 8 0 Expert1 Moderate Good Good Poor 1 24 63 12 0 Expert1 Moderate Good Good Very poor 0 3 23 49 25 Expert1 Moderate Good Moderate Very good 2 37 56 5 0 Expert1 Moderate Good Moderate Good 1 28 61 10 0 Expert1 Moderate Good Moderate Moderate 0 21 64 15 0 Expert1 Moderate Good Moderate Poor 0 14 64 22 0 Expert1 Moderate Good Moderate Very poor 0 3 23 49 25 Expert1 Moderate Good Poor Very good 0 12 62 25 1 Expert1 Moderate Good Poor Good 0 8 58 33 1 Expert1 Moderate Good Poor Moderate 0 5 51 42 2 Expert1 Moderate Good Poor Poor 0 3 42 51 4 Expert1 Moderate Good Poor Very poor 0 3 23 49 25 Expert1 Moderate Good Very poor Very good 0 0 0 26 74 Expert1 Moderate Good Very poor Good 0 0 0 26 74 Expert1 Moderate Good Very poor Moderate 0 0 0 26 74 Expert1 Moderate Good Very poor Poor 0 0 0 26 74 Expert1 Moderate Good Very poor Very poor 0 0 0 26 74

202

Expert1 Moderate Moderate Very good Very good 3 46 48 3 0 Expert1 Moderate Moderate Very good Good 2 37 56 5 0 Expert1 Moderate Moderate Very good Moderate 1 28 61 10 0 Expert1 Moderate Moderate Very good Poor 0 21 64 15 0 Expert1 Moderate Moderate Very good Very poor 0 3 23 49 25 Expert1 Moderate Moderate Good Very good 2 37 56 5 0 Expert1 Moderate Moderate Good Good 1 28 61 10 0 Expert1 Moderate Moderate Good Moderate 0 21 64 15 0 Expert1 Moderate Moderate Good Poor 0 14 64 22 0 Expert1 Moderate Moderate Good Very poor 0 3 23 49 25 Expert1 Moderate Moderate Moderate Very good 1 25 63 11 0 Expert1 Moderate Moderate Moderate Good 0 18 65 17 0 Expert1 Moderate Moderate Moderate Moderate 0 12 64 24 0 Expert1 Moderate Moderate Moderate Poor 0 7 58 34 1 Expert1 Moderate Moderate Moderate Very poor 0 3 23 49 25 Expert1 Moderate Moderate Poor Very good 0 8 58 33 1 Expert1 Moderate Moderate Poor Good 0 5 51 42 2 Expert1 Moderate Moderate Poor Moderate 0 3 42 51 4 Expert1 Moderate Moderate Poor Poor 0 1 33 58 8 Expert1 Moderate Moderate Poor Very poor 0 3 23 49 25 Expert1 Moderate Moderate Very poor Very good 0 0 0 26 74 Expert1 Moderate Moderate Very poor Good 0 0 0 26 74 Expert1 Moderate Moderate Very poor Moderate 0 0 0 26 74 Expert1 Moderate Moderate Very poor Poor 0 0 0 26 74 Expert1 Moderate Moderate Very poor Very poor 0 0 0 26 74 Expert1 Moderate Poor Very good Very good 0 18 64 18 0 Expert1 Moderate Poor Very good Good 0 12 62 25 1 Expert1 Moderate Poor Very good Moderate 0 8 58 33 1 Expert1 Moderate Poor Very good Poor 0 5 51 42 2 Expert1 Moderate Poor Very good Very poor 0 3 23 49 25 Expert1 Moderate Poor Good Very good 0 12 62 25 1 Expert1 Moderate Poor Good Good 0 8 58 33 1 Expert1 Moderate Poor Good Moderate 0 5 51 42 2 Expert1 Moderate Poor Good Poor 0 3 42 51 4 Expert1 Moderate Poor Good Very poor 0 3 23 49 25 Expert1 Moderate Poor Moderate Very good 0 8 58 33 1 Expert1 Moderate Poor Moderate Good 0 5 51 42 2 Expert1 Moderate Poor Moderate Moderate 0 3 42 51 4 Expert1 Moderate Poor Moderate Poor 0 1 33 58 8 Expert1 Moderate Poor Moderate Very poor 0 3 23 49 25 Expert1 Moderate Poor Poor Very good 0 3 46 48 3 Expert1 Moderate Poor Poor Good 0 2 37 56 5 Expert1 Moderate Poor Poor Moderate 0 1 28 62 9 Expert1 Moderate Poor Poor Poor 0 0 20 65 15 Expert1 Moderate Poor Poor Very poor 0 0 14 64 22 Expert1 Moderate Poor Very poor Very good 0 0 0 26 74 Expert1 Moderate Poor Very poor Good 0 0 0 26 74 Expert1 Moderate Poor Very poor Moderate 0 0 0 26 74 Expert1 Moderate Poor Very poor Poor 0 0 0 26 74 Expert1 Moderate Poor Very poor Very poor 0 0 0 26 74

203

Expert1 Moderate Very poor Very good Very good 0 0 0 26 74 Expert1 Moderate Very poor Very good Good 0 0 0 26 74 Expert1 Moderate Very poor Very good Moderate 0 0 0 26 74 Expert1 Moderate Very poor Very good Poor 0 0 0 26 74 Expert1 Moderate Very poor Very good Very poor 0 0 0 26 74 Expert1 Moderate Very poor Good Very good 0 0 0 26 74 Expert1 Moderate Very poor Good Good 0 0 0 26 74 Expert1 Moderate Very poor Good Moderate 0 0 0 26 74 Expert1 Moderate Very poor Good Poor 0 0 0 26 74 Expert1 Moderate Very poor Good Very poor 0 0 0 26 74 Expert1 Moderate Very poor Moderate Very good 0 0 0 26 74 Expert1 Moderate Very poor Moderate Good 0 0 0 26 74 Expert1 Moderate Very poor Moderate Moderate 0 0 0 26 74 Expert1 Moderate Very poor Moderate Poor 0 0 0 26 74 Expert1 Moderate Very poor Moderate Very poor 0 0 0 26 74 Expert1 Moderate Very poor Poor Very good 0 0 0 26 74 Expert1 Moderate Very poor Poor Good 0 0 0 26 74 Expert1 Moderate Very poor Poor Moderate 0 0 0 26 74 Expert1 Moderate Very poor Poor Poor 0 0 0 26 74 Expert1 Moderate Very poor Poor Very poor 0 0 0 26 74 Expert1 Moderate Very poor Very poor Very good 0 0 0 26 74 Expert1 Moderate Very poor Very poor Good 0 0 0 26 74 Expert1 Moderate Very poor Very poor Moderate 0 0 0 26 74 Expert1 Moderate Very poor Very poor Poor 0 0 0 26 74 Expert1 Moderate Very poor Very poor Very poor 0 0 0 26 74 Expert1 Poor Very good Very good Very good 2 33 58 7 0 Expert1 Poor Very good Very good Good 1 25 62 12 0 Expert1 Poor Very good Very good Moderate 0 18 64 18 0 Expert1 Poor Very good Very good Poor 0 12 62 25 1 Expert1 Poor Very good Very good Very poor 0 3 23 49 25 Expert1 Poor Very good Good Very good 1 25 62 12 0 Expert1 Poor Very good Good Good 0 18 64 18 0 Expert1 Poor Very good Good Moderate 0 12 62 25 1 Expert1 Poor Very good Good Poor 0 8 58 33 1 Expert1 Poor Very good Good Very poor 0 3 23 49 25 Expert1 Poor Very good Moderate Very good 0 18 64 18 0 Expert1 Poor Very good Moderate Good 0 12 62 25 1 Expert1 Poor Very good Moderate Moderate 0 8 58 33 1 Expert1 Poor Very good Moderate Poor 0 5 51 42 2 Expert1 Poor Very good Moderate Very poor 0 3 23 49 25 Expert1 Poor Very good Poor Very good 0 9 60 30 1 Expert1 Poor Very good Poor Good 0 5 54 39 2 Expert1 Poor Very good Poor Moderate 0 3 46 48 3 Expert1 Poor Very good Poor Poor 0 2 37 56 5 Expert1 Poor Very good Poor Very poor 0 3 23 49 25 Expert1 Poor Very good Very poor Very good 0 0 0 26 74 Expert1 Poor Very good Very poor Good 0 0 0 26 74 Expert1 Poor Very good Very poor Moderate 0 0 0 26 74 Expert1 Poor Very good Very poor Poor 0 0 0 26 74 Expert1 Poor Very good Very poor Very poor 0 0 0 26 74

204

Expert1 Poor Good Very good Very good 1 25 62 12 0 Expert1 Poor Good Very good Good 0 18 64 18 0 Expert1 Poor Good Very good Moderate 0 12 62 25 1 Expert1 Poor Good Very good Poor 0 8 58 33 1 Expert1 Poor Good Very good Very poor 0 3 23 49 25 Expert1 Poor Good Good Very good 0 18 64 18 0 Expert1 Poor Good Good Good 0 12 62 25 1 Expert1 Poor Good Good Moderate 0 8 58 33 1 Expert1 Poor Good Good Poor 0 5 51 42 2 Expert1 Poor Good Good Very poor 0 3 23 49 25 Expert1 Poor Good Moderate Very good 0 12 62 25 1 Expert1 Poor Good Moderate Good 0 8 58 33 1 Expert1 Poor Good Moderate Moderate 0 5 51 42 2 Expert1 Poor Good Moderate Poor 0 3 42 51 4 Expert1 Poor Good Moderate Very poor 0 3 23 49 25 Expert1 Poor Good Poor Very good 0 5 54 39 2 Expert1 Poor Good Poor Good 0 3 46 48 3 Expert1 Poor Good Poor Moderate 0 2 37 56 5 Expert1 Poor Good Poor Poor 0 1 28 62 9 Expert1 Poor Good Poor Very poor 0 3 23 49 25 Expert1 Poor Good Very poor Very good 0 0 0 26 74 Expert1 Poor Good Very poor Good 0 0 0 26 74 Expert1 Poor Good Very poor Moderate 0 0 0 26 74 Expert1 Poor Good Very poor Poor 0 0 0 26 74 Expert1 Poor Good Very poor Very poor 0 0 0 26 74 Expert1 Poor Moderate Very good Very good 0 18 64 18 0 Expert1 Poor Moderate Very good Good 0 12 62 25 1 Expert1 Poor Moderate Very good Moderate 0 8 58 33 1 Expert1 Poor Moderate Very good Poor 0 5 51 42 2 Expert1 Poor Moderate Very good Very poor 0 3 23 49 25 Expert1 Poor Moderate Good Very good 0 12 62 25 1 Expert1 Poor Moderate Good Good 0 8 58 33 1 Expert1 Poor Moderate Good Moderate 0 5 51 42 2 Expert1 Poor Moderate Good Poor 0 3 42 51 4 Expert1 Poor Moderate Good Very poor 0 3 23 49 25 Expert1 Poor Moderate Moderate Very good 0 8 58 33 1 Expert1 Poor Moderate Moderate Good 0 5 51 42 2 Expert1 Poor Moderate Moderate Moderate 0 3 42 51 4 Expert1 Poor Moderate Moderate Poor 0 1 33 58 8 Expert1 Poor Moderate Moderate Very poor 0 3 23 49 25 Expert1 Poor Moderate Poor Very good 0 3 46 48 3 Expert1 Poor Moderate Poor Good 0 2 37 56 5 Expert1 Poor Moderate Poor Moderate 0 1 28 62 9 Expert1 Poor Moderate Poor Poor 0 0 20 65 15 Expert1 Poor Moderate Poor Very poor 0 0 14 64 22 Expert1 Poor Moderate Very poor Very good 0 0 0 26 74 Expert1 Poor Moderate Very poor Good 0 0 0 26 74 Expert1 Poor Moderate Very poor Moderate 0 0 0 26 74 Expert1 Poor Moderate Very poor Poor 0 0 0 26 74 Expert1 Poor Moderate Very poor Very poor 0 0 0 26 74

205

Expert1 Poor Poor Very good Very good 0 9 60 30 1 Expert1 Poor Poor Very good Good 0 5 54 39 2 Expert1 Poor Poor Very good Moderate 0 3 46 48 3 Expert1 Poor Poor Very good Poor 0 2 37 56 5 Expert1 Poor Poor Very good Very poor 0 3 23 49 25 Expert1 Poor Poor Good Very good 0 5 54 39 2 Expert1 Poor Poor Good Good 0 3 46 48 3 Expert1 Poor Poor Good Moderate 0 2 37 56 5 Expert1 Poor Poor Good Poor 0 1 28 62 9 Expert1 Poor Poor Good Very poor 0 3 23 49 25 Expert1 Poor Poor Moderate Very good 0 3 46 48 3 Expert1 Poor Poor Moderate Good 0 2 37 56 5 Expert1 Poor Poor Moderate Moderate 0 1 28 62 9 Expert1 Poor Poor Moderate Poor 0 0 20 65 15 Expert1 Poor Poor Moderate Very poor 0 0 14 64 22 Expert1 Poor Poor Poor Very good 0 1 34 58 7 Expert1 Poor Poor Poor Good 0 1 25 63 11 Expert1 Poor Poor Poor Moderate 0 0 18 65 17 Expert1 Poor Poor Poor Poor 0 0 12 64 24 Expert1 Poor Poor Poor Very poor 0 0 7 59 34 Expert1 Poor Poor Very poor Very good 0 0 0 26 74 Expert1 Poor Poor Very poor Good 0 0 0 26 74 Expert1 Poor Poor Very poor Moderate 0 0 0 26 74 Expert1 Poor Poor Very poor Poor 0 0 0 26 74 Expert1 Poor Poor Very poor Very poor 0 0 0 26 74 Expert1 Poor Very poor Very good Very good 0 0 0 26 74 Expert1 Poor Very poor Very good Good 0 0 0 26 74 Expert1 Poor Very poor Very good Moderate 0 0 0 26 74 Expert1 Poor Very poor Very good Poor 0 0 0 26 74 Expert1 Poor Very poor Very good Very poor 0 0 0 26 74 Expert1 Poor Very poor Good Very good 0 0 0 26 74 Expert1 Poor Very poor Good Good 0 0 0 26 74 Expert1 Poor Very poor Good Moderate 0 0 0 26 74 Expert1 Poor Very poor Good Poor 0 0 0 26 74 Expert1 Poor Very poor Good Very poor 0 0 0 26 74 Expert1 Poor Very poor Moderate Very good 0 0 0 26 74 Expert1 Poor Very poor Moderate Good 0 0 0 26 74 Expert1 Poor Very poor Moderate Moderate 0 0 0 26 74 Expert1 Poor Very poor Moderate Poor 0 0 0 26 74 Expert1 Poor Very poor Moderate Very poor 0 0 0 26 74 Expert1 Poor Very poor Poor Very good 0 0 0 26 74 Expert1 Poor Very poor Poor Good 0 0 0 26 74 Expert1 Poor Very poor Poor Moderate 0 0 0 26 74 Expert1 Poor Very poor Poor Poor 0 0 0 26 74 Expert1 Poor Very poor Poor Very poor 0 0 0 26 74 Expert1 Poor Very poor Very poor Very good 0 0 0 26 74 Expert1 Poor Very poor Very poor Good 0 0 0 26 74 Expert1 Poor Very poor Very poor Moderate 0 0 0 26 74 Expert1 Poor Very poor Very poor Poor 0 0 0 26 74 Expert1 Poor Very poor Very poor Very poor 0 0 0 23 77

206

Expert1 Very poor Very good Very good Very good 0 0 0 26 74 Expert1 Very poor Very good Very good Good 0 0 0 26 74 Expert1 Very poor Very good Very good Moderate 0 0 0 26 74 Expert1 Very poor Very good Very good Poor 0 0 0 26 74 Expert1 Very poor Very good Very good Very poor 0 0 0 26 74 Expert1 Very poor Very good Good Very good 0 0 0 26 74 Expert1 Very poor Very good Good Good 0 0 0 26 74 Expert1 Very poor Very good Good Moderate 0 0 0 26 74 Expert1 Very poor Very good Good Poor 0 0 0 26 74 Expert1 Very poor Very good Good Very poor 0 0 0 26 74 Expert1 Very poor Very good Moderate Very good 0 0 0 26 74 Expert1 Very poor Very good Moderate Good 0 0 0 26 74 Expert1 Very poor Very good Moderate Moderate 0 0 0 26 74 Expert1 Very poor Very good Moderate Poor 0 0 0 26 74 Expert1 Very poor Very good Moderate Very poor 0 0 0 26 74 Expert1 Very poor Very good Poor Very good 0 0 0 26 74 Expert1 Very poor Very good Poor Good 0 0 0 26 74 Expert1 Very poor Very good Poor Moderate 0 0 0 26 74 Expert1 Very poor Very good Poor Poor 0 0 0 26 74 Expert1 Very poor Very good Poor Very poor 0 0 0 26 74 Expert1 Very poor Very good Very poor Very good 0 0 0 26 74 Expert1 Very poor Very good Very poor Good 0 0 0 26 74 Expert1 Very poor Very good Very poor Moderate 0 0 0 26 74 Expert1 Very poor Very good Very poor Poor 0 0 0 26 74 Expert1 Very poor Very good Very poor Very poor 0 0 0 26 74 Expert1 Very poor Good Very good Very good 0 0 0 26 74 Expert1 Very poor Good Very good Good 0 0 0 26 74 Expert1 Very poor Good Very good Moderate 0 0 0 26 74 Expert1 Very poor Good Very good Poor 0 0 0 26 74 Expert1 Very poor Good Very good Very poor 0 0 0 26 74 Expert1 Very poor Good Good Very good 0 0 0 26 74 Expert1 Very poor Good Good Good 0 0 0 26 74 Expert1 Very poor Good Good Moderate 0 0 0 26 74 Expert1 Very poor Good Good Poor 0 0 0 26 74 Expert1 Very poor Good Good Very poor 0 0 0 26 74 Expert1 Very poor Good Moderate Very good 0 0 0 26 74 Expert1 Very poor Good Moderate Good 0 0 0 26 74 Expert1 Very poor Good Moderate Moderate 0 0 0 26 74 Expert1 Very poor Good Moderate Poor 0 0 0 26 74 Expert1 Very poor Good Moderate Very poor 0 0 0 26 74 Expert1 Very poor Good Poor Very good 0 0 0 26 74 Expert1 Very poor Good Poor Good 0 0 0 26 74 Expert1 Very poor Good Poor Moderate 0 0 0 26 74 Expert1 Very poor Good Poor Poor 0 0 0 26 74 Expert1 Very poor Good Poor Very poor 0 0 0 26 74 Expert1 Very poor Good Very poor Very good 0 0 0 26 74 Expert1 Very poor Good Very poor Good 0 0 0 26 74 Expert1 Very poor Good Very poor Moderate 0 0 0 26 74 Expert1 Very poor Good Very poor Poor 0 0 0 26 74 Expert1 Very poor Good Very poor Very poor 0 0 0 26 74

207

Expert1 Very poor Moderate Very good Very good 0 0 0 26 74 Expert1 Very poor Moderate Very good Good 0 0 0 26 74 Expert1 Very poor Moderate Very good Moderate 0 0 0 26 74 Expert1 Very poor Moderate Very good Poor 0 0 0 26 74 Expert1 Very poor Moderate Very good Very poor 0 0 0 26 74 Expert1 Very poor Moderate Good Very good 0 0 0 26 74 Expert1 Very poor Moderate Good Good 0 0 0 26 74 Expert1 Very poor Moderate Good Moderate 0 0 0 26 74 Expert1 Very poor Moderate Good Poor 0 0 0 26 74 Expert1 Very poor Moderate Good Very poor 0 0 0 26 74 Expert1 Very poor Moderate Moderate Very good 0 0 0 26 74 Expert1 Very poor Moderate Moderate Good 0 0 0 26 74 Expert1 Very poor Moderate Moderate Moderate 0 0 0 26 74 Expert1 Very poor Moderate Moderate Poor 0 0 0 26 74 Expert1 Very poor Moderate Moderate Very poor 0 0 0 26 74 Expert1 Very poor Moderate Poor Very good 0 0 0 26 74 Expert1 Very poor Moderate Poor Good 0 0 0 26 74 Expert1 Very poor Moderate Poor Moderate 0 0 0 26 74 Expert1 Very poor Moderate Poor Poor 0 0 0 26 74 Expert1 Very poor Moderate Poor Very poor 0 0 0 26 74 Expert1 Very poor Moderate Very poor Very good 0 0 0 26 74 Expert1 Very poor Moderate Very poor Good 0 0 0 26 74 Expert1 Very poor Moderate Very poor Moderate 0 0 0 26 74 Expert1 Very poor Moderate Very poor Poor 0 0 0 26 74 Expert1 Very poor Moderate Very poor Very poor 0 0 0 26 74 Expert1 Very poor Poor Very good Very good 0 0 0 26 74 Expert1 Very poor Poor Very good Good 0 0 0 26 74 Expert1 Very poor Poor Very good Moderate 0 0 0 26 74 Expert1 Very poor Poor Very good Poor 0 0 0 26 74 Expert1 Very poor Poor Very good Very poor 0 0 0 26 74 Expert1 Very poor Poor Good Very good 0 0 0 26 74 Expert1 Very poor Poor Good Good 0 0 0 26 74 Expert1 Very poor Poor Good Moderate 0 0 0 26 74 Expert1 Very poor Poor Good Poor 0 0 0 26 74 Expert1 Very poor Poor Good Very poor 0 0 0 26 74 Expert1 Very poor Poor Moderate Very good 0 0 0 26 74 Expert1 Very poor Poor Moderate Good 0 0 0 26 74 Expert1 Very poor Poor Moderate Moderate 0 0 0 26 74 Expert1 Very poor Poor Moderate Poor 0 0 0 26 74 Expert1 Very poor Poor Moderate Very poor 0 0 0 26 74 Expert1 Very poor Poor Poor Very good 0 0 0 26 74 Expert1 Very poor Poor Poor Good 0 0 0 26 74 Expert1 Very poor Poor Poor Moderate 0 0 0 26 74 Expert1 Very poor Poor Poor Poor 0 0 0 26 74 Expert1 Very poor Poor Poor Very poor 0 0 0 26 74 Expert1 Very poor Poor Very poor Very good 0 0 0 26 74 Expert1 Very poor Poor Very poor Good 0 0 0 26 74 Expert1 Very poor Poor Very poor Moderate 0 0 0 26 74 Expert1 Very poor Poor Very poor Poor 0 0 0 26 74 Expert1 Very poor Poor Very poor Very poor 0 0 0 23 77

208

Expert1 Very poor Very poor Very good Very good 0 0 0 26 74 Expert1 Very poor Very poor Very good Good 0 0 0 26 74 Expert1 Very poor Very poor Very good Moderate 0 0 0 26 74 Expert1 Very poor Very poor Very good Poor 0 0 0 26 74 Expert1 Very poor Very poor Very good Very poor 0 0 0 26 74 Expert1 Very poor Very poor Good Very good 0 0 0 26 74 Expert1 Very poor Very poor Good Good 0 0 0 26 74 Expert1 Very poor Very poor Good Moderate 0 0 0 26 74 Expert1 Very poor Very poor Good Poor 0 0 0 26 74 Expert1 Very poor Very poor Good Very poor 0 0 0 26 74 Expert1 Very poor Very poor Moderate Very good 0 0 0 26 74 Expert1 Very poor Very poor Moderate Good 0 0 0 26 74 Expert1 Very poor Very poor Moderate Moderate 0 0 0 26 74 Expert1 Very poor Very poor Moderate Poor 0 0 0 26 74 Expert1 Very poor Very poor Moderate Very poor 0 0 0 26 74 Expert1 Very poor Very poor Poor Very good 0 0 0 26 74 Expert1 Very poor Very poor Poor Good 0 0 0 26 74 Expert1 Very poor Very poor Poor Moderate 0 0 0 26 74 Expert1 Very poor Very poor Poor Poor 0 0 0 26 74 Expert1 Very poor Very poor Poor Very poor 0 0 0 23 77 Expert1 Very poor Very poor Very poor Very good 0 0 0 26 74 Expert1 Very poor Very poor Very poor Good 0 0 0 26 74 Expert1 Very poor Very poor Very poor Moderate 0 0 0 26 74 Expert1 Very poor Very poor Very poor Poor 0 0 0 21 79 Expert1 Very poor Very poor Very poor Very poor 0 0 0 15 85

209

Appendix 4.2 Model variables – spatial data (supplements Table 4.2)

Appendix Table 4.2.1 Classification of Present Major Vegetation Subgroups (MVS V.4.1 *) into food

quality under good and poor accessibility of below-ground protein sources.

MVS No. MVS Name Type

Food quality (access)

(good) (poor)

2 Tropical or sub-tropical rainforest Rainforest High High

4 Eucalyptus open forests, shrubby understorey Woodland Low Low

5 Eucalyptus open forests, grassy understorey Woodland Moderate Low

6 Warm Temperate Rainforest Rainforest Moderate Moderate

7 Tropical Eucalyptus forest & woodlands, tall annual grassy understorey Woodland Low Low

8 Eucalyptus woodlands, shrubby understorey Woodland Low Low

9 Eucalyptus woodlands, tussock grass understorey Woodland Low Low

10 Eucalyptus woodlands, hummock grass understorey Woodland Low Low

11 Tropical mixed spp forests & woodlands Woodland Moderate Low

12 Callitris forests & woodlands Woodland Low Low

13 Brigalow forests & woodlands Woodland Low Low

14 Other Acacia forests & woodlands Woodland Low Low

15 Melaleuca open forests & woodlands Woodland Moderate Low

16 Other forests & woodlands Woodland Low Low

18 Eucalyptus low open woodlands with hummock grass Woodland Low Low

19 Eucalyptus low open woodlands with tussock grass Woodland Moderate Low

20 Mulga woodlands +/- tussock grass +/- forbs Shrubland Low Low

21 Other Acacia tall open shrublands & [tall] shrublands Shrubland Low Low

23 Acacia open woodlands & shrublands with hummock grass Shrubland Low Low

24 Acacia open woodlands & shrublands +/- tussock grass Shrubland Low Low

25 Acacia open woodlands & sparse shrublands, shrubby understorey Shrubland Low Low

26 Casuarina & Allocasuarina forests & woodlands Woodland Low Low

27 Mallee with hummock grass Shrubland Low Low

28 Low closed forest or tall closed shrublands Shrubland Low Low

30 Heath Shrubland Low Low

31 Saltbush & Bluebush shrublands Shrubland Low Low

32 Other shrublands Shrubland Low Low

33 Hummock grasslands Grassland Low Low

34 Mitchell grass tussock grasslands Grassland Moderate Low

35 Blue grass & tall bunch grass tussock grasslands Grassland High Low

37 Other tussock grasslands Grassland Moderate Low

38 Wet tussock grassland with herbs, sedges or rushes, herblands or ferns Wetland High High

39 Mixed chenopod, samphire +/- forbs Shrubland Low Low

40 Mangroves Mangroves High High

41 Saline or brackish sedgelands or grasslands Wetland High High

42 Naturally bare, sand, rock, claypan, mudflat None Low Low

210

* Department of the Environment and Water Resources (2007). Australia’s native vegetation: a summary of Australia’s

major vegetation groups, 2007. Commonwealth of Australia, Canberra.

43 Salt lakes & lagoons Wetland Moderate Moderate

44 Freshwater, dams, lakes, lagoons or aquatic plants Wetland Moderate Moderate

45 Mulga open woodlands & sparse shrublands +/- tussock grass Shrubland Low Low

46 Sea, estuaries (includes seagrass) None Low Low

47 Eucalyptus open woodlands with shrubby understorey Woodland Low Low

48 Eucalyptus open woodlands, grassy understorey Woodland Moderate Low

49 Melaleuca shrublands & open shrublands Shrubland Low Low

51 Mulga woodlands & shrublands with hummock grass Shrubland Low Low

52 Mulga open woodlands & sparse shrublands with hummock grass Shrubland Low Low

53 Eucalyptus low open woodlands, shrubby understorey Woodland Low Low

56 Eucalyptus open woodlands, chenopod or samphire understorey Woodland Low Low

57 Lignum shrublands & wetlands Wetland Moderate Moderate

59 Eucalyptus woodlands, ferns, herbs, sedges, rushes, wet tussock grasses Woodland High Moderate

60 Eucalyptus open forests, ferns, herbs, sedges, rushes, wet tussock grasses Woodland High Moderate

62 Dry rainforest or vine thickets Rainforest Moderate Low

63 Sedgelands, rushs or reeds Wetland High High

64 Other grasslands Grassland Moderate Low

70 Callitris open woodlands Woodland Low Low

71 Casuarina & Allocasuarina open woodlands, tussock grass understorey Woodland Low Low

74 Casuarina & Allocasuarina open woodlands, shrubby understorey Woodland Low Low

75 Melaleuca open woodlands Woodland Moderate Low

79 Other open Woodlands Woodland Low Low

80 Other sparse shrublands & sparse heathlands Shrubland Low Low

90 Regrowth or modified forests & woodlands Unknown Low Low

92 Regrowth or modified graminoids Unknown Low Low

96 Unclassified Forest Woodland Low Low

97 Unclassified native vegetation Unknown Low Low

98 Cleared, non-native vegetation, buildings None Low Low

99 Unknown/No data Unknown Low Low

211

Appendix Table 4.2.2 Classification of Australian Land Use and Management (ALUM V.7 *) primary

(and secondary / tertiary where applicable) classes into food quality under good and poor accessibility

of below-ground protein sources.

* Australian Bureau of Agricultural and Resource Economics and Sciences (2011). Guidelines for land use mapping in

Australia: principles, procedures and definitions, fourth edition. Commonwealth of Australia, Canberra.

Primary (secondary, tertiary) ALUM class Food quality (access)

(good) (poor)

1 Conservation and natural environments Low Low

2 Production from relatively natural environments Low Low

3 Production from dryland agriculture and plantations 3.1 Plantation forestry, 3.6 Land in transition

Low Low

3 Production from dryland agriculture and plantations 3.2 Grazing modified pastures

Moderate Low

3 Production from dryland agriculture and plantations 3.3 Cropping

High Moderate

3 Production from dryland agriculture and plantations 3.4 Perennial horticulture, 3.5 Seasonal horticulture

High High

4 Production from irrigated agriculture and plantations 4.1 Irrigated plantation forestry, 4.5.5 Irrigated turf farming & 4.6 Irrigated land in transition

Low Low

4 Production from irrigated agriculture and plantations 4.2 Grazing irrigated modified pastures

Moderate Moderate

4 Production from irrigated agriculture and plantations 4.3 Irrigated cropping, 4.4 Irrigated perennial horticulture, 4.5 Irrigated seasonal horticulture (except 4.5.5)

High High

5 Intensive uses Low Low

6 Water (except 6.3.0 & 6.3.1) Low Low

6 Water 6.3.0 River, 6.3.1 River - conservation

Moderate Moderate

212

Appendix Table 4.2.3 Classification of Present Major Vegetation Groups (MVG V.4.1 *) into dense

vegetation cover.

* Department of the Environment and Water Resources (2007). Australia’s native vegetation: a summary of Australia’s

major vegetation groups, 2007. Commonwealth of Australia, Canberra.

MVG No. MVG Name Type Cover

1 Rainforests and Vine Thickets Rainforest Good

2 Eucalypt Tall Open Forests Woodland Moderate

3 Eucalypt Open Forests Woodland Moderate

4 Eucalypt Low Open Forests Woodland Moderate

5 Eucalypt Woodlands Woodland Moderate

6 Acacia Forests and Woodlands Woodland Moderate

7 Callitris Forests and Woodlands Woodland Moderate

8 Casuarina Forests and Woodlands Woodland Moderate

9 Melaleuca Forests and Woodlands Woodland Moderate

10 Other Forests and Woodlands Woodland Moderate

11 Eucalypt Open Woodlands Woodland Poor

12 Tropical Eucalypt Woodlands/Grasslands Woodland Moderate

13 Acacia Open Woodlands Shrubland Poor

14 Mallee Woodlands and Shrublands Shrubland Moderate

15 Low Closed Forests and Tall Closed Shrublands Shrubland Good

16 Acacia Shrublands Shrubland Moderate

17 Other Shrublands Shrubland Moderate

18 Heathlands Shrubland Good

19 Tussock Grasslands Grassland Poor

20 Hummock Grasslands Grassland Poor

21 Other Grasslands, Herblands, Sedgelands and Rushlands Wetland Poor

22 Chenopod Shrublands, Samphire Shrublands and Forblands Shrubland Poor

23 Mangroves Mangroves Good

24 Inland aquatic - freshwater, salt lakes, lagoons Wetland Poor

25 Cleared, non-native vegetation, buildings None/modified Poor

26 Unclassified native vegetation None/modified Poor

27 Naturally bare - sand, rock, claypan, mudflat None/modified Poor

28 Sea and estuaries None/modified Poor

29 Regrowth, modified native vegetation None/modified Poor

30 Unclassified Forest Woodland Moderate

31 Other Open Woodlands Woodland Poor

32 Mallee Open Woodlands and Sparse Mallee Shrublands Shrubland Poor

99 Unknown/no data None/modified Poor

213

Appendix Table 4.2.4 Classification of Australian Land Use and Management (ALUM V.7 *) primary

(and secondary / tertiary where applicable) classes into dense vegetation cover.

* Australian Bureau of Agricultural and Resource Economics and Sciences (2011). Guidelines for land use mapping in

Australia: principles, procedures and definitions, fourth edition. Commonwealth of Australia, Canberra.

Primary (secondary, tertiary) ALUM class Cover

1 Conservation and natural environments Poor

2 Production from relatively natural environments 2.1 Grazing natural vegetation

Poor

2 Production from relatively natural environments 2.2 Production forestry

Moderate

3 Production from dryland agriculture and plantations 3.1 Plantation forestry, 3.3.5 Sugar, 3.4 Perennial horticulture (except 3.4.0)

Moderate

3 Production from dryland agriculture and plantations 3.2 Grazing modified pastures, 3.3 Cropping (except 3.3.5), 3.4.0 Perennial horticulture, 3.5 Seasonal horticulture, 3.6 Land in transition

Poor

4 Production from irrigated agriculture and plantations 4.1 Irrigated plantation forestry, 4.3.5 Irrigated sugar, 4.4 Irrigated perennial horticulture (except 4.4.0 & 4.4.7), 4.5.1 Irrigated seasonal fruits

Moderate

4 Production from irrigated agriculture and plantations 4.2 Grazing irrigated modified pastures, 4.3 Irrigated cropping (except 4.3.5), 4.4.0 Irrigated perennial horticulture, 4.4.7 Irrigated perennial vegetables & herbs, 4.5 Irrigated seasonal horticulture (except 4.5.1), 4.6 Irrigated land in transition

Poor

5 Intensive uses Poor

6 Water Poor

214

Appendix 4.3 Methodology – spatial pattern suitability analysis


S3 Appendix. Rcode spatial pattern suitabilityanalysis (PDF)Jens G. Froese

23 February 2017

This document provides a detailed, reproducible description of the spatial pattern suitability analysis methodology. It isSupporting Information (S3) to the manuscript:

Froese JG, Smith CS, Durr PA, McAlpine CA, van Klinken RD. Modelling seasonal habitatsuitability for wide-ranging species: invasive wild pigs in northern Australia. Submitted toPLoS ONE.

It is written in R Markdown ([1]) and knitr ([2]), two R ([3]) packages for writing dynamic, reproducible reports. A

.zip file containing data inputs to reproduce analyses can be downloaded from Dryad (http://dx.doi.org/10.5061

/dryad.v103v (http://dx.doi.org/10.5061/dryad.v103v)). Some parts of the code used to print this document have beensurpressed to enhance readability. A generalized version of the code is available at URL https://github.com/jgfroese/PATTSI (https://github.com/jgfroese/PATTSI).

Load required R packagesR package raster ([4]) and its dependencies are required for spatial pattern suitability analysis. Session information

incl. package versions are listed at the bottom of this document.

require(raster) # for all analyses of raster objects incl. moving window analysis

1. Expert elicitationFor each habitat variable (i.e. its modelled resource quality indices xr ), we elicited a distance-dependent response-

to-pattern curve ( fDr ) from each individual expert (see manuscript Fig 3). These curves followed a step-wise pattern,

because we discretised both:

distance into five equal distance bands (“very close”, “close”, “medium”, “far” and “very far”), relative to eachexpert’s defined mobility threshold (i.e. 1km, 2km or 3km).resource suitability indices ( SIr ) into five equal classes (“very good (80-100)”, “good (60-80)”, “moderate

(40-60)”, “poor (20-40)” and “very poor (0-20)”), see manuscript S2.1 Table.

We asked experts to relate each distance band to a corresponding suitability class under the assumption that othervariables do not constrain suitability. To fill two elicitation gaps (expert 3 did not define fDDisturbance and expert 6

did not define fDHeat ) we

applied all other experts’ fDr to the missing expert’s defined mobility threshold (Expert3 = 2km, Expert6 = 3km)

computed the average fDr and used it for the missing expert’s model

To derive distance weights for computation of 2. Resource suitability indices, Step 4, we used the mid-points ofelicited suitability index classes divided by 100 (e.g. class “moderate (40-60)” = SIr 50 = weight 0.5). For class “very

good (80-100)” we did not use the mid-point SIr = 90 but assigned SIr = 100 (= weight 1.0) to avoid unintended

distance penalties (i.e. an adjacent resource of quality xr = 60 should compute as distance-weighted suitability

SIr = 60 (if weight is 1.0) and not SIr = 54 (if weight is 0.9)).

Figure S3.1 Step-wise distance-dependent response-to-pattern curve for habitat variable “food” elicited from expert 4(A) and distance weights derived for computation of food suitability indices (B).

2. Resource suitability indicesGoal

Focal pixel resource suitability index ( SIr ) depends on the distance of a (numerical) habitat variable.

Method:

Generate a circular moving window where each position is weighted by its distance from the focal pixel(radius/weights derived from 1. Expert elicitation).compute the focal pixel SIr as the highest weighted value ( xr ) of a habitat variable within this moving window.

Step 1

Define a function that returns a circular matrix of given radius and resolution and assigns value 1 if matrix position <=

radius and value NA if matrix position > radius (Source: [5]).

make_circ_filter <- function(radius, res){

circ_filter <- matrix(NA, nrow=1+(2*radius/res), ncol =1+(2*radius/res))



sweeper <- function(mat){

for(row in 1:nrow(mat)){

for(col in 1:ncol(mat)){

dist <- sqrt((as.numeric(dimnames(mat)[[1]])[row])^2 +

(as.numeric(dimnames(mat)[[1]])[col])^2)

if(dist<=radius) {mat[row, col]<-1}

}

}

return(mat)

}

out <- sweeper(circ_filter)

return(out)

}

Step 2

Apply function to generate five matrices with different radii (= distance bands), relative to each expert’s defined mobilitythreshold (i.e. 1km for Expert1, 2km for Experts 3/5 and 3km for Experts 2/4/6).

res <- 1 # resolution (= pixel size, e.g. 100m)

mr.1 <- 10 # matrix radius (= mobility threshold, must be multiple of res, e.g 1km = 10 x 10

0m)

m.1 <- make_circ_filter(mr.1, res)

m.2 <- make_circ_filter((mr.1/5)*4, res)



m.5 <- make_circ_filter((mr.1/5), res)

Replace value==1 with unique temp value in ascending order from largest to smallest matrix.

m.1[m.1 == 1] <- 1

m.2[m.2 == 1] <- 2

m.3[m.3 == 1] <- 3

m.4[m.4 == 1] <- 4

m.5[m.5 == 1] <- 5

Step 3

Combine the five matrices into one (two at a time starting with the smallest):

a.5 <- array(NA, dim(m.4), dimnames(m.4)) # create temp array of size = larger matrix

a.5[rownames(m.5), colnames(m.5)] <- m.5 # ... with values = smaller matrix

m.4 <- pmax(m.4, a.5, na.rm = TRUE) # combine values: larger matrix + temp array

a.4 <- array(NA, dim(m.3), dimnames(m.3)) # repeat with: output + next-larger matrix

a.4[rownames(m.4), colnames(m.4)] <- m.4

m.3 <- pmax(m.3, a.4, na.rm = TRUE)

a.3 <- array(NA, dim(m.2), dimnames(m.2))


m.2 <- pmax(m.2, a.3, na.rm = TRUE)

a.2 <- array(NA, dim(m.1), dimnames(m.1))


m.band.1 <- pmax(m.1, a.2, na.rm = TRUE)

## -10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10

## -10 NA NA NA NA NA NA NA NA NA NA 1 NA NA NA NA NA NA NA NA NA NA

## -9 NA NA NA NA NA NA 1 1 1 1 1 1 1 1 1 NA NA NA NA NA NA

## -8 NA NA NA NA 1 1 1 1 1 1 2 1 1 1 1 1 1 NA NA NA NA

## -7 NA NA NA 1 1 1 1 2 2 2 2 2 2 2 1 1 1 1 NA NA NA

## -6 NA NA 1 1 1 2 2 2 2 2 3 2 2 2 2 2 1 1 1 NA NA

## -5 NA NA 1 1 2 2 2 3 3 3 3 3 3 3 2 2 2 1 1 NA NA

## -4 NA 1 1 1 2 2 3 3 3 3 4 3 3 3 3 2 2 1 1 1 NA

## -3 NA 1 1 2 2 3 3 3 4 4 4 4 4 3 3 3 2 2 1 1 NA

## -2 NA 1 1 2 2 3 3 4 4 4 5 4 4 4 3 3 2 2 1 1 NA

## -1 NA 1 1 2 2 3 3 4 4 5 5 5 4 4 3 3 2 2 1 1 NA

## 0 1 1 2 2 3 3 4 4 5 5 5 5 5 4 4 3 3 2 2 1 1

## 1 NA 1 1 2 2 3 3 4 4 5 5 5 4 4 3 3 2 2 1 1 NA

## 2 NA 1 1 2 2 3 3 4 4 4 5 4 4 4 3 3 2 2 1 1 NA

## 3 NA 1 1 2 2 3 3 3 4 4 4 4 4 3 3 3 2 2 1 1 NA

## 4 NA 1 1 1 2 2 3 3 3 3 4 3 3 3 3 2 2 1 1 1 NA

## 5 NA NA 1 1 2 2 2 3 3 3 3 3 3 3 2 2 2 1 1 NA NA

## 6 NA NA 1 1 1 2 2 2 2 2 3 2 2 2 2 2 1 1 1 NA NA

## 7 NA NA NA 1 1 1 1 2 2 2 2 2 2 2 1 1 1 1 NA NA NA

## 8 NA NA NA NA 1 1 1 1 1 1 2 1 1 1 1 1 1 NA NA NA NA

## 9 NA NA NA NA NA NA 1 1 1 1 1 1 1 1 1 NA NA NA NA NA NA

## 10 NA NA NA NA NA NA NA NA NA NA 1 NA NA NA NA NA NA NA NA NA NA

Repeat steps 2 and 3 for mobility thresholds 2km and 3km.

mr.2 <- 20 # moving window radius 2km (= 20 x 100m)

mr.3 <- 30 # moving window radius 3km (= 30 x 100m)

...

m.band.2

m.band.3

Step 4

Replace temp values with expert-elicited weight for each distance band ( SIdata.zip file containing f_DFood.csv

and other weights derived from 1. Expert elicitation can be downloaded from Dryad (http://dx.doi.org/10.5061/dryad.v103v (http://dx.doi.org/10.5061/dryad.v103v))).

csv.Food = read.csv("SIdata/S3Appendix/fD/f_DFood.csv ")

csv.Food.E1 <- subset(csv.Food, Expert == 'Expert1') # for Expert1 use m.band.1

m.band.1[m.band.1 == 1] <- csv.Food.E1$X1

m.band.1[m.band.1 == 2] <- csv.Food.E1$X0.8




m.band.Food.E1 <- m.band.1















Step 5

Perform moving window analysis using function focal {raster} with parameters ( SIdata.zip file containing raster

layers with resource quality indices xr can be downloaded from Dryad (http://dx.doi.org/10.5061/dryad.v103v

(http://dx.doi.org/10.5061/dryad.v103v))).

r = raster("SIdata/S3Appendix/GIS/Food-quality-dry.ti f") # raster layer with numerical resou

rce quality index, e.g. Food quality in dry season scenario

w = m.band.Food.E1 # moving window is banded weights matrix, e.g. Expert 1 f_Dfood

fun = max # focal pixel takes highest weighted resource quality index within moving window

WARNING! The following process may take several hours depending on the size of r and w.

r.f <- focal(r, w, fun, na.rm = TRUE, pad = FALSE, padVal ue = NA) # na.rm = TRUE ignores NoD

ata

r.m <- mask(r.f, r) # extract by r to remove padded edges

writeRaster(r.m, filename = paste("SIdata/S3Appendix/ out/Food-SI-dry_E1.tif", sep="")) # sav

e ouput raster

Repeat steps 4 and 5 for all four habitat variables and six experts in two seasonal scenarios:

Water wet/dry = 12 runsFood wet/dry = 12 runsHeat wet/dry = 12 runsDisturbance global scenario = 6 runs



[3] RCoreTeam 2015. R: a language and environment for statistical computing. R Foundation for Statistical Computing,Vienna, Austria. URL http://www.R-project.org/ (http://www.R-project.org/).

[4] Hijmans, R.J. 2015. Package ‘raster’: geographic data analysis and modeling. URL http://cran.r-project.org/web/packages/raster/ (http://cran.r-project.org/web/packages/raster/).

[5] Scroggie, M. 2012. Applying a circular moving window filter to raster data in R. URL https://scrogster.wordpress.com/2012/10/05/applying-a-circular-moving-window-filter-to-raster-data-in-r/ (https://scrogster.wordpress.com/2012/10/05/applying-a-circular-moving-window-filter-to-raster-data-in-r/).

Session information

## Session info ----------------------------------- ---------------------------

## setting value



## ui RTerm

## language (EN)



## date 2017-02-23

## Packages --------------------------------------- ---------------------------


## devtools 1.10.0 2016-01-23 CRAN (R 3.1.3)

## digest 0.6.8 2014-12-31 CRAN (R 3.1.3)

## evaluate 0.8 2015-09-18 CRAN (R 3.1.3)

## formatR 1.2.1 2015-09-18 CRAN (R 3.1.3)

## htmltools 0.3 2015-12-29 CRAN (R 3.1.3)

## knitr 1.12.3 2016-01-22 CRAN (R 3.1.3)

## lattice 0.20-30 2015-02-22 CRAN (R 3.1.3)

## magrittr 1.5 2014-11-22 CRAN (R 3.1.3)

## memoise 1.0.0 2016-01-29 CRAN (R 3.1.3)

## raster * 2.4-20 2015-09-08 CRAN (R 3.1.3)

## Rcpp 0.12.1 2015-09-10 CRAN (R 3.1.3)

## rgdal 1.1-1 2015-11-02 CRAN (R 3.1.3)

## rmarkdown 0.9.2 2016-01-01 CRAN (R 3.1.3)

## sp * 1.2-1 2015-10-18 CRAN (R 3.1.3)

## stringi 1.0-1 2015-10-22 CRAN (R 3.1.3)

## stringr 1.0.0 2015-04-30 CRAN (R 3.1.3)

## yaml 2.1.13 2014-06-12 CRAN (R 3.1.3)

220

Appendix 4.4 Methodology – validation


S4 Appendix. Rcode validation (PDF)Jens G. Froese

17 April 2017

This document provides a detailed, reproducible description of the methodology used to evaluate and validate performance of habitatsuitability models using the Continuous Boyce Index (CBI). It is Supporting Information (S4 Appendix) to the manuscript:

Froese JG, Smith CS, Durr PA, McAlpine CA, van Klinken RD. Modelling seasonal habitat suitabilityfor wide-ranging species: invasive wild pigs in northern Australia. Submitted to PLoS ONE.

The document is written in R Markdown ([1]) and knitr ([2]), two R ([3]) packages for writing dynamic, reproducible reports. A

.zip file containing data inputs to reproduce analyses can be downloaded from Dryad

( http://dx.doi.org/10.5061/dryad.v103v ). Some parts of the code used to print this document have been suppressed to

enhance readability. A generalized version of the code is available at URL https://github.com/jgfroese/HSI-CBI-validation .

Load required R packagesR packages ecospat ([4]), data.table ([5]) and zoo ([6]) and their dependencies are required for HSI-CBI-validation. Session

information incl. package versions are listed at the bottom of this document.

require(data.table) # for function `setnames`

require(ecospat) # for function `ecospat.boyce`

require(zoo) # for plotting function `na.fill`

1. Data preparationStep 1

This R script requires two .TXT files for each model / validation data combination, which have to be prepared in package {raster}

or alternative GIS software as described below. These files can be downloaded from Dryad( http://dx.doi.org/10.5061/dryad.v103v ).

Expected HSI across validation background:define validation backgrounds ( Balkanu = BLKW , Lakefield = LKNP , Oyala Thumotang = OTNP , NAQS, ALA -

see manuscript Table 2)mask raster layers of individual expert models and an average model by each validation backgroundexport raster attribute tables to .TXT with 3 columns: [ID], [HSI], [pixel count]

1.

Predicted HSI at wild pig presence records:convert presence records of each validation data set into raster layercombine masked raster layer of each expert/average model with each set of presence recordsexport raster attribute tables to .TXT with 5 columns: [ID], [Value], [pixel count], [HSI], [number of presence records per

pixel]

2.

Step 2

Compute model-predicted HSI at wild pig presence records.

First, read .TXT files for each model (Expert1, Expert2, …, averaged) / validation presences (BLKW, LKNP, OTNP, NAQS, ALA)

combination as data frame,

pred.E1.wetBLKW = read.csv("SIdata/S4Appendix/wet/Pre dicted/SiPred_wet_E1_BLKW1000.txt")

...

pred.E6.wetBLKW = read.csv("SIdata/S4Appendix/wet/Pre dicted/SiPred_wet_E6_BLKW1000.txt")

pred.av.wetBLKW = read.csv("SIdata/S4Appendix/wet/Pre dicted/SiPred_wet_Eall_BLKW1000.txt")

pred.E1.dryBLKW = read.csv("SIdata/S4Appendix/dry/Pre dicted/SiPred_dry_E1_BLKW1000.txt")

...

pred.E6.dryBLKW = read.csv("SIdata/S4Appendix/dry/Pre dicted/SiPred_dry_E6_BLKW1000.txt")

pred.av.dryBLKW = read.csv("SIdata/S4Appendix/dry/Pre dicted/SiPred_dry_Eall_BLKW1000.txt")

...

pred.av.wetLKNP = read.csv("SIdata/S4Appendix/wet/Pre dicted/SiPred_wet_Eall_LKNP1000.txt")

...

pred.av.dryLKNP = read.csv("SIdata/S4Appendix/dry/Pre dicted/SiPred_dry_Eall_LKNP1000.txt")

...

pred.av.wetOTNP = read.csv("SIdata/S4Appendix/wet/Pre dicted/SiPred_wet_Eall_OTNP1000.txt")

...

pred.av.dryNAQS = read.csv("SIdata/S4Appendix/dry/Pre dicted/SiPred_dry_Eall_NAQS1000.txt")

...

pred.av.wetALA = read.csv("SIdata/S4Appendix/wet/Pred icted/SiPred_wet_Eall_ALA1000.txt")

...

pred.av.dryALA = read.csv("SIdata/S4Appendix/dry/Pred icted/SiPred_dry_Eall_ALA1000.txt")


pred.list <- list (pred.E1.wetBLKW, pred.E2.wetBLKW, p red.E3.wetBLKW, pred.E4.wetBLKW, pred.E5.wetBLKW,

pred.E6.wetBLKW, pred.av.wetBLKW, pred.E1.dryBLKW, pred.E2.dryBLKW, pred.E3.dryBLKW, pred.E4.dryBLKW, p

red.E5.dryBLKW, pred.E6.dryBLKW, pred.av.dryBLKW, p red.E1.wetLKNP, pred.E2.wetLKNP, pred.E3.wetLKNP, p r

ed.E4.wetLKNP, pred.E5.wetLKNP, pred.E6.wetLKNP, pr ed.av.wetLKNP, pred.E1.dryLKNP, pred.E2.dryLKNP, pr e

d.E3.dryLKNP, pred.E4.dryLKNP, pred.E5.dryLKNP, pre d.E6.dryLKNP, pred.av.dryLKNP, pred.E1.wetOTNP, pr e

d.E2.wetOTNP, pred.E3.wetOTNP, pred.E4.wetOTNP,pred .E5.wetOTNP, pred.E6.wetOTNP, pred.av.wetOTNP, pred .

E1.dryNAQS, pred.E2.dryNAQS, pred.E3.dryNAQS, pred. E4.dryNAQS, pred.E5.dryNAQS, pred.E6.dryNAQS, pred. a

v.dryNAQS, pred.E1.wetALA, pred.E2.wetALA, pred.E3. wetALA, pred.E4.wetALA, pred.E5.wetALA, pred.E6.wet A

LA, pred.av.wetALA, pred.E1.dryALA, pred.E2.dryALA, pred.E3.dryALA, pred.E4.dryALA, pred.E5.dryALA, pr e

d.E6.dryALA, pred.av.dryALA)

n.list <- 56 # the number of data frames in your list

Then, homogenise the five column names for all data frames,


setnames(pred.list[[i]], c("ID", "Value", "Pixelcou nt", "HSI", "Presences"))

}

and calculate the total number of presence records per HSI value (one pixel may contain multiple records).

pred.sum.list <- vector("list", n.list)


pred.sum.list[[i]] <- aggregate(cbind(Pixelcount*Pres ences)~HSI, data = pred.list[[i]], sum)

}

Finally, add descriptive column names to the new list of data frames,


setnames(pred.sum.list[[i]], c("HSI", "Presences"))

}

and convert it into a list of vectors (= HSI at feral pig presence records),

pred.v.list = vector("list", n.list)


pred.v = vector()

for (j in 1:length(pred.sum.list[[i]]$Presences)) {

for (k in 1:pred.sum.list[[i]][j, 2]) {

pred.v <-append(pred.v, pred.sum.list[[i]][j, 1])

}

}

pred.v.list[[i]] <- append(pred.v.list[[i]], pred.v)

}

e.g. HSI of model 1 at Balkanu presence records (first in list).

Step 3

Compute model-predicted HSI across validation backgrounds.

First, read .TXT files for each model (Expert1, Expert2, …, averaged) / validation background (BLKW, LKNP, OTNP, NAQS, ALA)

combination as data frame,

exp.E1.wetBLKW = read.csv("SIdata/S4Appendix/wet/Expe cted/SiExp_wet_E1_BLKW1000.txt")

...

exp.E6.wetBLKW = read.csv("SIdata/S4Appendix/wet/Expe cted/SiExp_wet_E6_BLKW1000.txt")

exp.av.wetBLKW = read.csv("SIdata/S4Appendix/wet/Expe cted/SiExp_wet_Eall_BLKW1000.txt")

exp.E1.dryBLKW = read.csv("SIdata/S4Appendix/dry/Expe cted/SiExp_dry_E1_BLKW1000.txt")

...

exp.E6.dryBLKW = read.csv("SIdata/S4Appendix/dry/Expe cted/SiExp_dry_E6_BLKW1000.txt")

exp.av.dryBLKW = read.csv("SIdata/S4Appendix/dry/Expe cted/SiExp_dry_Eall_BLKW1000.txt")

...

exp.av.wetLKNP = read.csv("SIdata/S4Appendix/wet/Expe cted/SiExp_wet_Eall_LKNP1000.txt")

...

exp.av.dryLKNP = read.csv("SIdata/S4Appendix/dry/Expe cted/SiExp_dry_Eall_LKNP1000.txt")

...

exp.av.wetOTNP = read.csv("SIdata/S4Appendix/wet/Expe cted/SiExp_wet_Eall_OTNP1000.txt")

...

exp.av.dryNAQS = read.csv("SIdata/S4Appendix/dry/Expe cted/SiExp_dry_Eall_NAQS1000.txt")

...

exp.av.wetALA = read.csv("SIdata/S4Appendix/wet/Expec ted/SiPred_wet_Eall_ALA1000.txt")

...

exp.av.dryALA = read.csv("SIdata/S4Appendix/dry/Expec ted/SiPred_dry_Eall_ALA1000.txt")


exp.list <- list (exp.E1.wetBLKW, exp.E2.wetBLKW, exp. E3.wetBLKW, exp.E4.wetBLKW, exp.E5.wetBLKW, exp.E

6.wetBLKW, exp.av.wetBLKW, exp.E1.dryBLKW, exp.E2.d ryBLKW, exp.E3.dryBLKW, exp.E4.dryBLKW, exp.E5.dryB L

KW, exp.E6.dryBLKW, exp.av.dryBLKW, exp.E1.wetLKNP, exp.E2.wetLKNP, exp.E3.wetLKNP, exp.E4.wetLKNP, ex p

.E5.wetLKNP, exp.E6.wetLKNP, exp.av.wetLKNP, exp.E1 .dryLKNP, exp.E2.dryLKNP, exp.E3.dryLKNP, exp.E4.dr y

LKNP, exp.E5.dryLKNP, exp.E6.dryLKNP, exp.av.dryLKN P, exp.E1.wetOTNP, exp.E2.wetOTNP, exp.E3.wetOTNP, e

xp.E4.wetOTNP, exp.E5.wetOTNP, exp.E6.wetOTNP, exp. av.wetOTNP, exp.E1.dryNAQS, exp.E2.dryNAQS, exp.E3. d

ryNAQS, exp.E4.dryNAQS, exp.E5.dryNAQS, exp.E6.dryN AQS, exp.av.dryNAQS, exp.E1.wetALA, exp.E2.wetALA, e

xp.E3.wetALA, exp.E4.wetALA, exp.E5.wetALA, exp.E6. wetALA, exp.av.wetALA, exp.E1.dryALA, exp.E2.dryALA ,

exp.E3.dryALA, exp.E4.dryALA, exp.E5.dryALA, exp.E6 .dryALA, exp.av.dryALA)

Then, homogenise the three column names for all data frames,


setnames(exp.list[[i]], c("ID", "HSI", "Pixelcount" ))

}

and convert it into a list of vectors (= HSI across validation backgrounds [potentially very large - see manuscript Table 1]).

exp.v.list = vector("list", n.list)


exp.v = vector()

for (j in 1:length(exp.list[[i]]$Pixelcount)) {

for (k in 1:exp.list[[i]][j, 3]) {

exp.v <-append(exp.v, exp.list[[i]][j, 2])

}

}

exp.v.list[[i]] <- append(exp.v.list[[i]], exp.v)

}

2. Data analysisStep 4

Apply function boyce {ecospat} to each model (Expert1, Expert2, …, averaged) / validation data (BLKW, LKNP, OTNP, NAQS,

ALA) combination with parameters:

exp.v.list[[i]] # Expected HSI across background ([[1]] = E1.wetBLK W, ... , [[56]] = av.dryALA)

pred.v.list[[i]] # Predicted HSI at presences ([[1]] = E1.wetBLKW, . .. , [[56]] = av.dryALA)

nclass = 0 # defaults to moving window (continuous, classifica tion-independent) computation with argume

nts:

window.w = 10 # moving window width (i.e. 10 adjacent HSI values are considered in each computation)

res = 100 # resolution factor (i.e. 100 computations across t he total range of HSI)

PEplot = F # no PEplot is generated (customised plots - see ma nuscript Figure 6)

boyce.E1.wetBLKW <- ecospat.boyce(exp.v.list[[1]], pr ed.v.list[[1]], nclass, window.w, res, PEplot)

...

boyce.E6.wetBLKW <- ecospat.boyce(exp.v.list[[6]], pr ed.v.list[[6]], nclass, window.w, res, PEplot)

boyce.av.wetBLKW <- ecospat.boyce(exp.v.list[[7]], pr ed.v.list[[7]], nclass, window.w, res, PEplot)

boyce.E1.dryBLKW <- ecospat.boyce(exp.v.list[[8]], pr ed.v.list[[8]], nclass, window.w, res, PEplot)

...

boyce.E6.dryBLKW <- ecospat.boyce(exp.v.list[[13]], p red.v.list[[13]], nclass, window.w, res, PEplot)

boyce.av.dryBLKW <- ecospat.boyce(exp.v.list[[14]], p red.v.list[[14]], nclass, window.w, res, PEplot)

...

boyce.av.wetLKNP <- ecospat.boyce(exp.v.list[[21]], p red.v.list[[21]], nclass, window.w, res, PEplot)

...

boyce.av.dryLKNP <- ecospat.boyce(exp.v.list[[28]], p red.v.list[[28]], nclass, window.w, res, PEplot)

...

boyce.av.wetOTNP <- ecospat.boyce(exp.v.list[[35]], p red.v.list[[35]], nclass, window.w, res, PEplot)

...

boyce.av.dryNAQS <- ecospat.boyce(exp.v.list[[42]], p red.v.list[[42]], nclass, window.w, res, PEplot)

...

boyce.av.wetALA <- ecospat.boyce(exp.v.list[[49]], pr ed.v.list[[49]], nclass, window.w, res, PEplot)

...

boyce.av.dryALA <- ecospat.boyce(exp.v.list[[56]], pr ed.v.list[[56]], nclass, window.w, res, PEplot)

Step 5

Investigate results of CBI analysis.

Combine all results in list for faster analysis,

boyce.list <- list (boyce.E1.wetBLKW, boyce.E2.wetBLKW , boyce.E3.wetBLKW, boyce.E4.wetBLKW, boyce.E5.we

tBLKW, boyce.E6.wetBLKW, boyce.av.wetBLKW, boyce.E1 .dryBLKW, boyce.E2.dryBLKW, boyce.E3.dryBLKW, boyce .

E4.dryBLKW, boyce.E5.dryBLKW, boyce.E6.dryBLKW, boy ce.av.dryBLKW, boyce.E1.wetLKNP, boyce.E2.wetLKNP, b

oyce.E3.wetLKNP, boyce.E4.wetLKNP, boyce.E5.wetLKNP , boyce.E6.wetLKNP, boyce.av.wetLKNP, boyce.E1.dryL K

NP, boyce.E2.dryLKNP, boyce.E3.dryLKNP, boyce.E4.dr yLKNP, boyce.E5.dryLKNP, boyce.E6.dryLKNP, boyce.av .

dryLKNP, boyce.E1.wetOTNP, boyce.E2.wetOTNP, boyce. E3.wetOTNP, boyce.E4.wetOTNP, boyce.E5.wetOTNP, boy c

e.E6.wetOTNP, boyce.av.wetOTNP, boyce.E1.dryNAQS, b oyce.E2.dryNAQS, boyce.E3.dryNAQS, boyce.E4.dryNAQS ,

boyce.E5.dryNAQS, boyce.E6.dryNAQS, boyce.av.dryNAQ S, boyce.E1.wetALA, boyce.E2.wetALA, boyce.E3.wetAL A

, boyce.E4.wetALA, boyce.E5.wetALA, boyce.E6.wetALA , boyce.av.wetALA, boyce.E1.dryALA, boyce.E2.dryALA ,

boyce.E3.dryALA, boyce.E4.dryALA, boyce.E5.dryALA, boyce.E6.dryALA, boyce.av.dryALA)

and print CBI ( $Spearman.cor ) for all model / validation data combinations (see manuscript Table 3),

CBI.list = vector("list", n.list)

for (i in seq_along(boyce.list)) {

CBI.list[[i]] <- append(CBI.list[[i]], round(boyce.li st[[i]]$Spearman.cor, digits = 2))

}

CBI.list

e.g. CBI of expert model 1 validated against wet season Balkanu presence records (first in list).

Step 6

Compute proportion of validation background expected to be highly or very highly suitable habitat (HSI >= 60) for all model /validation data combinations (see manuscript Table 3),

t.HSI <- 59.99 # HSI threshold

HSI.60 <- vector("list", n.list)

for (i in seq_along(HSI.60)) {

HSI.60[[i]] <- aggregate(Pixelcount~HSI > t.HSI, data = e xp.list[[i]], sum) / sum(exp.list[[i]]$Pixel

count)

}

HSI.60.list <- vector("list", n.list)

for (i in seq_along(HSI.60.list)) {

HSI.60.list[[i]] <- append(HSI.60.list[[i]], (round(H SI.60[[i]][2, "Pixelcount"] * 100, digits = 0)))

}

HSI.60.list

e.g. HSI 60 of expert model 1 in wet season Balkanu validation background (first in list).

Step 7

Compare P/E ratio between individual expert models and an averaged model for each validation data set.

Create a nested list (first list expert/average models per validation data set, then list validation data sets),

boyce.wetBLKW <- list (boyce.E1.wetBLKW, boyce.E2.wetB LKW, boyce.E3.wetBLKW, boyce.E4.wetBLKW, boyce.E5

.wetBLKW, boyce.E6.wetBLKW, boyce.av.wetBLKW)

boyce.dryBLKW <- list (boyce.E1.dryBLKW, boyce.E2.dryB LKW, boyce.E3.dryBLKW, boyce.E4.dryBLKW, boyce.E5

.dryBLKW, boyce.E6.dryBLKW, boyce.av.dryBLKW)

boyce.wetLKNP <- list (boyce.E1.wetLKNP, boyce.E2.wetL KNP, boyce.E3.wetLKNP, boyce.E4.wetLKNP, boyce.E5

.wetLKNP, boyce.E6.wetLKNP, boyce.av.wetLKNP)

boyce.dryLKNP <- list (boyce.E1.dryLKNP, boyce.E2.dryL KNP, boyce.E3.dryLKNP, boyce.E4.dryLKNP, boyce.E5

.dryLKNP, boyce.E6.dryLKNP, boyce.av.dryLKNP)

boyce.wetOTNP <- list (boyce.E1.wetOTNP, boyce.E2.wetO TNP, boyce.E3.wetOTNP, boyce.E4.wetOTNP, boyce.E5

.wetOTNP, boyce.E6.wetOTNP, boyce.av.wetOTNP)

boyce.dryNAQS <- list (boyce.E1.dryNAQS, boyce.E2.dryN AQS, boyce.E3.dryNAQS, boyce.E4.dryNAQS, boyce.E5

.dryNAQS, boyce.E6.dryNAQS, boyce.av.dryNAQS)

boyce.wetALA <- list (boyce.E1.wetALA, boyce.E2.wetALA , boyce.E3.wetALA, boyce.E4.wetALA, boyce.E5.wetA

LA, boyce.E6.wetALA, boyce.av.wetALA)

boyce.dryALA <- list (boyce.E1.dryALA, boyce.E2.dryALA , boyce.E3.dryALA, boyce.E4.dryALA, boyce.E5.dryA

LA, boyce.E6.dryALA, boyce.av.dryALA)

plot.v.list <- list ("Balkanu (wet season)" = boyce.wetB LKW, "Balkanu (dry season)" = boyce.dryBLKW, "L

akefield (wet season)" = boyce.wetLKNP, "Lakefield (d ry season)" = boyce.dryLKNP, "Oyala Thumotang (wet

season)" = boyce.wetOTNP, "NAQS (dry season)" = boyce. dryNAQS, "ALA NT (wet season)" = boyce.wetALA, "

ALA NT (dry season)" = boyce.dryALA)

and plot from nested list (see manuscript Fig 5).

graphics.off()

par(mfrow = c(4, 4), mar = c(2, 2, 3, 0), oma = c(4, 3, 0 , 0))

for (i in seq_along(plot.v.list)) {

plot(plot.v.list[[i]][[1]]$HS, plot.v.list[[i]][[1] ]$F.ratio, type = "n", xlab = '', ylab = '',

xlim = c(15, 80), ylim = c(0, 3), main = paste(names(plot .v.list[i])))

lines(plot.v.list[[i]][[1]]$HS, na.fill(plot.v.list [[i]][[1]]$F.ratio, 0), col = "black", lty = 1)






lines(plot.v.list[[i]][[7]]$HS, na.fill(plot.v.list [[i]][[7]]$F.ratio, 0), col = "red", lty = 1, lwd

= 2)

mtext(LETTERS[i], side = 3, line = -2, adj = 0.05)

plot.new()

legend("left", c((paste("Expert 1, CBI =", round(pl ot.v.list[[i]][[1]]$Spearman.cor, digits = 2))),

(paste("Expert 2, CBI =", round(plot.v.list[[i]][[2 ]]$Spearman.cor, digits = 2))),





(paste("Averaged, CBI =", round(plot.v.list[[i]][[7 ]]$Spearman.cor, digits = 2))))

,

bty = "n", lty = c(1,2,3,4,5,6,1), col = c("black", "blac k", "black", "black", "black", "black

", "red"), lwd = c(1, 1, 1, 1, 1, 1, 2), y.intersp = 1. 3, title = expression(bold(Legend)), title.adj =

0.05)

}

mtext(expression(italic(Habitat~suitability~index~( HSI))), side = 1, outer = TRUE, cex = 1.2, line = 2.

2)

mtext(expression(italic(Predicted-to-expected~(P/E) ~ratio)), side = 2, outer = TRUE, cex = 1.2, line =

0.8)

References[1] Allaire, J.J. et al. 2016. Package ‘rmarkdown’: dynamic documents for R. URL http://rmarkdown.rstudio.com/ .

[2] Xie, Y. 2016. Package ‘knitr’: a general-purpose package for dynamic report generation in R. URL http://yihui.name/knitr/ .

[3] RCoreTeam 2015. R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna,Austria. URL http://www.R-project.org/ .

[4] Broennimann, O. 2015. Package ‘ecospat’: spatial ecology miscellaneous methods. URLhttp://cran.r-project.org/web/packages/ecospat/ .

[5] Dowle, M. et al. 2015. Package ‘data.table’: extension of data.frame. URLhttps://github.com/Rdatatable/data.table/wiki/ .

[6] Zeileis, A. et al. 2015. Package ‘zoo’: S3 infrastructure for regular and irregular time series. URLhttp://zoo.R-Forge.R-project.org/ .

Session information

## Session info ----------------------------------- ---------------------------

## setting value



## ui RTerm

## language (EN)



## date 2017-04-18

## Packages --------------------------------------- ---------------------------


## abind 1.4-3 2015-03-13 CRAN (R 3.1.3 )

## acepack 1.3-3.3 2013-05-03 CRAN (R 3.1.3 )

## ade4 * 1.7-2 2015-04-14 CRAN (R 3.1.3 )

## adehabitatHR 0.4.14 2015-07-22 CRAN (R 3.1.3 )

## adehabitatLT 0.3.20 2015-07-22 CRAN (R 3.1.3 )

## adehabitatMA 0.3.10 2015-07-22 CRAN (R 3.1.3 )

## ape * 3.3 2015-05-29 CRAN (R 3.1.3 )

## biomod2 3.1-64 2014-12-10 CRAN (R 3.1.3 )

## boot 1.3-15 2015-02-24 CRAN (R 3.1.3 )

## chron 2.3-47 2015-06-24 CRAN (R 3.1.3 )

## CircStats 0.2-4 2012-10-29 CRAN (R 3.1.3 )

## class 7.3-12 2015-02-11 CRAN (R 3.1.3 )

## cluster 2.0.1 2015-01-31 CRAN (R 3.1.3 )

## codetools 0.2-10 2015-01-17 CRAN (R 3.1.3 )

## colorspace 1.2-6 2015-03-11 CRAN (R 3.1.3 )

## data.table * 1.9.6 2015-09-19 CRAN (R 3.1.3 )

## deldir 0.1-9 2015-03-09 CRAN (R 3.1.3 )

## devtools 1.10.0 2016-01-23 CRAN (R 3.1.3 )

## digest 0.6.8 2014-12-31 CRAN (R 3.1.3 )

## dismo 1.0-12 2015-03-15 CRAN (R 3.1.3 )

## ecodist 1.2.9 2013-12-03 CRAN (R 3.1.3 )

## ecospat * 1.1 2015-03-06 CRAN (R 3.1.3 )

## evaluate 0.8 2015-09-18 CRAN (R 3.1.3 )

## foreach * 1.4.3 2015-10-13 CRAN (R 3.1.3 )

## foreign 0.8-63 2015-02-20 CRAN (R 3.1.3 )

## formatR 1.2.1 2015-09-18 CRAN (R 3.1.3 )

## Formula 1.2-1 2015-04-07 CRAN (R 3.1.3 )

## gam * 1.12 2015-05-13 CRAN (R 3.1.3 )

## gbm * 2.1.1 2015-03-11 CRAN (R 3.1.3 )

## ggplot2 2.1.0 2016-03-01 CRAN (R 3.1.3 )

## goftest 1.0-3 2015-07-03 CRAN (R 3.1.3 )

## gridExtra 2.0.0 2015-07-14 CRAN (R 3.1.3 )

## gtable 0.1.2 2012-12-05 CRAN (R 3.1.3 )

## hexbin 1.27.1 2015-08-19 CRAN (R 3.1.3 )

## Hmisc 3.17-0 2015-09-21 CRAN (R 3.1.3 )

## htmltools 0.3 2015-12-29 CRAN (R 3.1.3 )

## iterators 1.0.8 2015-10-13 CRAN (R 3.1.3 )

## knitr 1.12.3 2016-01-22 CRAN (R 3.1.3 )

## lattice * 0.20-30 2015-02-22 CRAN (R 3.1.3 )

## latticeExtra 0.6-26 2013-08-15 CRAN (R 3.1.3 )

## magrittr 1.5 2014-11-22 CRAN (R 3.1.3 )

## maptools 0.8-37 2015-09-29 CRAN (R 3.1.3 )

## MASS 7.3-39 2015-02-24 CRAN (R 3.1.3 )

## Matrix 1.2-3 2015-11-28 CRAN (R 3.1.3 )

## MatrixModels 0.4-1 2015-08-22 CRAN (R 3.1.3 )

## mda 0.4-7 2015-05-25 CRAN (R 3.1.3 )

## memoise 1.0.0 2016-01-29 CRAN (R 3.1.3 )

## mgcv 1.8-4 2014-11-27 CRAN (R 3.1.3 )

## multcomp 1.4-1 2015-07-23 CRAN (R 3.1.3 )

## munsell 0.4.2 2013-07-11 CRAN (R 3.1.3 )

## mvtnorm 1.0-3 2015-07-22 CRAN (R 3.1.3 )

## nlme 3.1-120 2015-02-20 CRAN (R 3.1.3 )

## nnet 7.3-9 2015-02-11 CRAN (R 3.1.3 )

## plyr 1.8.3 2015-06-12 CRAN (R 3.1.3 )

## polspline 1.1.12 2015-07-14 CRAN (R 3.1.3 )

## polyclip 1.3-2 2015-05-27 CRAN (R 3.1.3 )

## pROC 1.8 2015-05-05 CRAN (R 3.1.3 )

## proto 0.3-10 2012-12-22 CRAN (R 3.1.3 )

## quantreg 5.19 2015-08-31 CRAN (R 3.1.3 )

## randomForest 4.6-12 2015-10-07 CRAN (R 3.1.3 )

## raster 2.4-20 2015-09-08 CRAN (R 3.1.3 )

## rasterVis 0.37 2015-09-06 CRAN (R 3.1.3 )

## RColorBrewer 1.1-2 2014-12-07 CRAN (R 3.1.3 )

## Rcpp 0.12.1 2015-09-10 CRAN (R 3.1.3 )

## reshape 0.8.5 2014-04-23 CRAN (R 3.1.3 )

## rmarkdown 0.9.2 2016-01-01 CRAN (R 3.1.3 )

## rms 4.4-0 2015-09-28 CRAN (R 3.1.3 )

## rpart 4.1-9 2015-02-24 CRAN (R 3.1.3 )

## sandwich 2.3-4 2015-09-24 CRAN (R 3.1.3 )

## scales 0.3.0 2015-08-25 CRAN (R 3.1.3 )

## sp * 1.2-1 2015-10-18 CRAN (R 3.1.3 )

## SparseM 1.7 2015-08-15 CRAN (R 3.1.3 )

## spatstat 1.41-1 2015-02-27 CRAN (R 3.1.3 )

## stringi 1.0-1 2015-10-22 CRAN (R 3.1.3 )

## stringr 1.0.0 2015-04-30 CRAN (R 3.1.3 )

## survival * 2.38-1 2015-02-24 CRAN (R 3.1.3 )

## tensor 1.5 2012-05-05 CRAN (R 3.1.3 )

## TH.data 1.0-6 2015-01-05 CRAN (R 3.1.3 )

## yaml 2.1.13 2014-06-12 CRAN (R 3.1.3 )

## zoo * 1.7-12 2015-03-16 CRAN (R 3.1.3 )

229

Appendix 4.5 Validation maps – seasonal habitat suitability

230

Appendix Fig 4.5 Seasonal habitat suitability for feral pig breeding in the four validation

backgrounds. Presence records used for validation are shown for: Balkanu wet (A) and dry season (B),

Lakefield wet (C) and dry (D) season, Oyala Thumotang wet season (E), NAQS dry season (F) and ALA wet

(G) and dry (H) season. Descriptions of data sets are in Table 4.3. Validation backgrounds were defined from

existing management units or by buffering data points.

231

Appendix 4.6 Additional analyses – seasonal habitat suitability

App

endi

x Fi

g 4.

6.1

Shar

e of

seas

onal

hab

itat p

er st

ate.

Per

cent

ages

wer

e ca

lcul

ated

for e

ach

suita

bilit

y cl

ass s

epar

atel

y. L

ocat

ion

of st

ates

is sh

own

in

Fig

4.1.

232

App

endi

x Fi

g 4.

6.2

Dis

trib

utio

n of

mod

elle

d ha

bita

t sui

tabi

lity

for

each

bro

ad v

eget

atio

n ty

pe. H

abita

t sui

tabi

lity

clas

ses w

ere

take

n fr

om th

e B

ayes

ian

netw

ork

mod

el (T

able

4.1

). Pe

rcen

tage

s wer

e ca

lcul

ated

sepa

rate

ly fo

r eac

h ve

geta

tion

type

(fro

m P

rese

nt M

ajor

Veg

etat

ion

Gro

ups (

MV

G V

.4.1

).

233

Appendix 5.1 Methodology – omnidirectional current density


Rcode - modelling omnidirectional currentdensityJens G. Froese

18 April 2017

This document provides a detailed, reproducible description of the omnidirectional current density methodology. It isSupporting Information to the manuscript:

Froese JG, Smith CS, Durr PA, McAlpine CA & van Klinken RD (in prep). “Modelling habitatconnectivity for biosecurity: the risk of infectious disease spread in wild pigs in northernAustralia.”

It is written in R Markdown ([1]) and knitr ([2]), two R ([3]) packages for writing dynamic, reproducible reports. Some

parts of the code used to print this document have been suppressed to enhance readability.

The methodoology was adapted from:

Pelletier D, Clark M, Anderson, MG, Rayfield, B, Wulder, MA et al. (2014). “Applying circuittheory for corridor expansion and management at regional scales: tiling, pinch points, andomnidirectional connectivity.” PLoS ONE 9(1): e84135.

It extends the circuit-theoretic Circuitscape (CS) model to regional-scale applications (millions of study area pixels).

While access to a high-performance computing cluster is not required, running Circuitscape is computationally demanding,a high powered computer with as much RAM as possible is recommended (we had 40GB).

Step 1: PreparationPrepare your workspace and load required packages

rm(list=ls()) # clear workspace

dev.off() # clear graphics

require(raster) # load packages

require(rgdal)

Set raster options (incl. tmpdir for easily deleting temporary files that otherwise clog up yourC:/ drive)

rasterOptions (tmpdir = "[your-temp-dir]", progress = "t ext", timer = TRUE)

Load study area wide resistance and random raster

Prepare input raster data in GIS software of your choice (we used .tif format, must have same spatial reference, extent,

resolution)

matrix resistance for entire study area res

matrix resistance tiles (divide study area into suitably sized tiles, collated in one directory, seePelletier et al. 2014 for guidance)

random resistance values for entire study area ran (must have same value range as resistance raster, WARNING!

for heterogeneous study areas, random values should reflect the range of resistance values in each individual tile toavoid border effects)

res = raster("[your-resistance-raster.tif]")

res; #plot(res)

ran = raster("[your-random-raster.tif]")

ran; #plot(ran)

Create buffered resistance and current source tiles

Set wd to your tiles dir (if you choose base wd, then for loop print(i) produces error) and load resistance .tif tiles

into list

setwd("[your-tiles-dir]")

l.rtl = dir(getwd(), pattern = ".tif")

Loop through each tile in list and write out buffered resistance and current source (two directions WE and NS) .tif files

to tifdir

tifdir = "[your-tif-dir]"

for (i in l.rtl) {

print(i)

rtl = raster(i) # read tile

rtl.temp <- extend(rtl, c(1500,1500), value=NA) # extend by 100% buffer (our tiles were 1500

*1500 pixels, need to adjust value), buffer is NA

ran.buf <- crop(ran, rtl.temp) # crop random raster by buffered tile

rtl.buf <- cover(rtl.temp, ran.buf) # replace NA in buffer with random values

we.buf <- rtl.buf

we.buf[] <- NA; we.buf[,1] <- 1; we.buf[,4500] <- 2 # current source tile (size = buffered t

ile, values = NA except thin strip west/east)

ns.buf <- rtl.buf

ns.buf[] <- NA; ns.buf[1,] <- 1; ns.buf[4500,] <- 2 # current source tile (size = buffered t

ile, values = NA except thin strip north/south)

writeRaster(rtl.buf, datatype='INT4S', filename = pas te(tifdir, "Tiles_buf/", i, ".asc", sep

= "")) # write buffered resistance tile as .asc

writeRaster(we.buf, datatype='INT4S', filename = past e(tifdir, "SourceWE_buf/SourceWE_", i,

".asc", sep = "")) # write current source WE tile as .asc

writeRaster(ns.buf, datatype='INT4S', filename = past e(tifdir, "SourceNS_buf/SourceNS_", i,

".asc", sep = "")) # write current source NS tile as .asc

}

Clean up your temp files (to avoid memory issues)

rtemp = "[your-temp-dir]"

setwd(rtemp)

file.remove(dir(rtemp))

gc()

Step 2: Directional Circuitscape (CS) runCode was adapted from

Marrotte 2015: http://robbymarrotte.weebly.com/blog/running-circuitscape-in-r-windows-os(http://robbymarrotte.weebly.com/blog/running-circuitscape-in-r-windows-os)

Prepare CS inputs from Step 1 in lists

Set wd to your tiles dir

setwd("[your-tif-dir]")

Create two lists for resistance and current source (here WE direction, repeat for NS) .asc files

l.tile = dir("[your-tif-dir]/Tiles_buf/", pattern = ".a sc")

l.source = dir("[your-tif-dir]/SourceWE_buf/", patter n = ".asc")

Prepare CS .ini files for each tile

For each listed resistance tile, write lines with CS options, input .asc and output .out … (WE or NS)

l.tile.ini = list() # same as: vector("list")

for (i in l.tile) {

tile.ini <- c("[circuitscape options]",

"data_type = raster",

"scenario = pairwise",

"write_cur_maps = True",

"write_cum_cur_map_only = True",

paste("habitat_file = ", getwd(), "/Tiles_buf/", i, sep = ""),

paste("output_file = ", getwd(), "/CS_outWE/", i, " .out", sep = ""))

l.tile.ini[[i]] <- append(l.tile.ini[[i]], tile.ini)

}

For each listed current source tile (WE or NS), write line with input .asc

l.source.ini = list()

for (i in l.source) {

source.ini <- c(paste("point_file = ", getwd(), "/SourceWE_ buf/", i, sep = ""))

l.source.ini[[i]] <- append(l.source.ini[[i]], source.ini)

}

For each list element, combine lines and write lines out to .ini file (to CS_runWE or CS_runNS dir), then combine all

written out .ini files in a run list

l.run <- list()

for (i in 1:length(l.tile.ini)) {

l.tile.ini[[i]] <- c(l.tile.ini[[i]], l.source.ini[[i ]])

writeLines(l.tile.ini[[i]], paste(getwd(), "/CS_run WE/Tile", i, ".ini", sep = ""))

l.run[i] <- paste(getwd(), "/CS_runWE/Tile", i, ".ini ", sep = "")

}

Run the Circuitscape model

For instructions on software installation see

McRae BH, Shah VB, Mohapatra TK (2013). Circuitscape 4 user guide: The NatureConservancy. http://www.circuitscape.org (http://www.circuitscape.org)

Make a CS run CMD and run command for each .ini file in the run list (this runs the model)

CS_exe <- 'C:/"Program Files"/Circuitscape/cs_run.exe ' # Don't forget quotation marks in "Prog

ram Files"

sapply(l.run, function(x) system(paste(CS_exe, x)))

Step 3: Directional current density mosaicPrepare workspace and load packages

rm(list=ls()) # clear workspace

tifdir = "[your-tif-dir]/MosaicWE/"

setwd("[your-tif-dir]/CS_outWE/")

require(raster)

require(rgdal)

projection <- "+proj=aea +lat_1=-18 +lat_2=-36 +lat_0 =0 +lon_0=132 +x_0=0 +y_0=0 +ellps=GRS80

+units=m +no_defs" # define a projection system (should match your inp ut rasters, we used Aust

ralian Albers prjection)

Load cumulative current .asc tiles (WE or NS), combine in a list and crop buffer by rows and columns

l.in = dir(getwd(), pattern = "curmap.asc")

l.cur = list()

for (i in l.in) {

cur = raster(i)

cur.crop <- crop(cur, extent(cur, 1501, 3000, 1501, 3 000)) # this extracts the orginal tile

area (here 1500*1500 pixels), from the larger buffe r tile (here 4500*4500 pixels)

crs(cur.crop) <- projection

writeRaster(cur.crop, filename = paste(tifdir, i, ".t if", sep = ""))

l.cur[[i]] <- cur.crop

}

Mosaic (use merge because tiles do not overlap) cropped cumulative current tiles into one raster and write out as .tif

x <- list()

for (i in 1:length(l.cur)){

x[i] <- l.cur[i]

}

x$filename <- '[your-[WE/NS]mosaic-raster.tif]'

x$overwrite <- TRUE

cur.mos <- do.call(merge, x)

writeRaster(cur.mos, filename = paste(tifdir, "[your- [WE/NS]mosaic-raster.tif]", sep = ""))

Step 4: Omnidirectional current density mosaicLoad resistance and directional cumulative current mosaic rasters (WE and NS)

res = raster("[your-resistance-raster.tif]")

res; #plot(res)

setwd("[your-tif-dir]")

tifdir = "[your-tif-dir]/Current/"

mos.WE = raster(paste(getwd(), "/MosaicWE/[your-NSmos aic-raster.tif]", sep = ""))

mos.NS = raster(paste(getwd(), "/MosaicNS/[your-WEmos aic-raster.tif]", sep = ""))

Mask mosaic rasters (WE and NS) by resistance raster

res <- crop(res, mos.WE)

mos.WE <- mask(mos.WE, res)

mos.WE; #plot(mos.WE)

writeRaster(mos.WE, filename = paste(tifdir, "[your-m asked-WEmosaic-raster.tif]", sep = ""))

mos.NS <- mask(mos.NS, res)

mos.NS; #plot(mos.NS)

writeRaster(mos.NS, filename = paste(tifdir, "[your-m asked-NSmosaic-raster.tif]", sep = ""))

Create omnidirectional current density mosaic by multiplication and log10 transformation

mos <- overlay(mos.NS, mos.WE, fun = function(x, y) { return(log10(x * y))} )

writeRaster(mos, filename = paste(tifdir, "[your-OCDl og10-raster.tif", sep = ""))

Optional: explore mosaic histogram

hist(mos,

col="springgreen4",

main="Histogram of omnidirectional current density" ,

ylab="Number of Pixels",

xlab="current density")

Clean up your temp files (to avoid memory issues)

rtemp = "[your-temp-dir]"

setwd(rtemp)

file.remove(dir(rtemp))

gc()

Session information

## Session info ----------------------------------- ---------------------------

## setting value



## ui RTerm

## language (EN)



## date 2017-04-18

## Packages --------------------------------------- ---------------------------


## devtools 1.10.0 2016-01-23 CRAN (R 3.1.3)

## digest 0.6.8 2014-12-31 CRAN (R 3.1.3)

## evaluate 0.8 2015-09-18 CRAN (R 3.1.3)

## formatR 1.2.1 2015-09-18 CRAN (R 3.1.3)

## htmltools 0.3 2015-12-29 CRAN (R 3.1.3)

## knitr 1.12.3 2016-01-22 CRAN (R 3.1.3)

## magrittr 1.5 2014-11-22 CRAN (R 3.1.3)

## memoise 1.0.0 2016-01-29 CRAN (R 3.1.3)

## rmarkdown 0.9.2 2016-01-01 CRAN (R 3.1.3)

## stringi 1.0-1 2015-10-22 CRAN (R 3.1.3)

## stringr 1.0.0 2015-04-30 CRAN (R 3.1.3)

## yaml 2.1.13 2014-06-12 CRAN (R 3.1.3)

239

Appendix 5.2 Methodology – patch connectivity


241

Appendix 5.3 Analysis – tests of significance

Appendix Table 5.3.1 Two-way tests of significance between levels of ResDisp for the dry (a) and wet (b)

season.

Median and mean patch connectivity within the network of feral pig habitat patches in northern Australia is shown for

each level (size of connected habitat components in km2). Pairwise p values have been calculated using the Mann-

Whitney U-test of significance. *p < .05, **p < .01, ***p < .001.

(a) Dry season scenario

Matrix resistance (ResDisp)

high moderate low p value

Median (km2)

393 341 - 0.704

- 341 411 < .001

393 - 411 < .001

Mean (km2) 7,820 15,282 45,174

(b) Wet season scenario

Matrix resistance (ResDisp)

high moderate low p value

Median (km2)

5,640 2,697 - 0.098

- 2,697 6,932 < .001

5,640 - 6,932 < .001

Mean (km2) 32,463 44,961 86,139

242 (a

)

Dry

seas

on sc

enar

io /

ResD

isp

high

(1-1

00)

Dis

pDis

t

CD

Thre

sh

5k

10k

15

k

st

rict

m

oder

ate

lo

ose

CD

Th

resh

m

od

(279

) st

rict

(53)

C

D

Thre

sh

mod

(3

69)

stric

t (5

3)

CD

Th

resh

m

od

(580

) st

rict

(53)

Dis

p D

ist

10k

(53)

5k

(5

3)

Dis

p D

ist

10k

(369

) 5k

(2

79)

Dis

p D

ist

10k

(5,8

43)

5k

(1,4

72)

loos

e (1

,472

) **

* **

* lo

ose

(5,8

43)

***

***

loos

e (3

3,39

4)

***

***

15

k (5

3)

p =

0.95

3

15k

(580

) p

= 0.

142

15k

(33,

394)

**

* **

*

mod

(2

79)

**

* m

od

(369

)

***

mod

(5

80)

**

*

10k

(53)

10

k (3

69)

10k

(5,8

43)

**

*

Dry

seas

on sc

enar

io /

ResD

isp

mod

erat

e (1

-21)

Dis

pDis

t

CD

Thre

sh

5k

10k

15

k

st

rict

m

oder

ate

lo

ose

CD

Th

resh

m

od

(356

) st

rict

(43)

C

D

Thre

sh

mod

(5

43)

stric

t (4

3)

CD

Th

resh

m

od

(735

) st

rict

(43)

Dis

p D

ist

10k

(43)

5k

(4

3)

Dis

p D

ist

10k

(543

) 5k

(3

56)

Dis

p D

ist

10k

(47,

172)

5k

(1

,901

)

loos

e (1

,901

) **

* **

* lo

ose

(47,

172)

**

* **

* lo

ose

(112

,943

) **

* **

*

15k

(43)

p

= 0.

988

15k

(735

) *

***

15k

(112

,943

) **

* **

*

mod

(3

56)

**

* m

od

(543

)

***

mod

(7

35)

**

*

10k

(43)

10

k (5

43)

*

10k

(47,

172)

***

Dry

seas

on sc

enar

io /

ResD

isp

low

(1-5

)

Dis

pDis

t

CD

Thre

sh

5k

10k

15

k

st

rict

m

oder

ate

lo

ose

CD

Th

resh

m

od

(299

) st

rict

(38)

C

D

Thre

sh

mod

(8

74)

stric

t (3

8)

CD

Th

resh

m

od

(1,9

69)

stric

t (3

8)

D

isp

Dis

t 10

k (3

8)

5k

(38)

10k

(874

) 5k

(2

99)

Dis

p D

ist

10k

(29,

681)

5k

(4

,065

)

loos

e (4

,065

) **

* **

* lo

ose

(29,

681)

**

* **

* lo

ose

(275

,592

) **

* **

*

15k

(38)

p

= 0.

998

15k

(1,9

69)

***

***

15k

(275

,592

) **

* **

*

mod

(2

99)

**

* m

od

(874

)

***

mod

(1

,969

)

***

10

k (3

8)

10k

(874

)

***

10k

(29,

681)

***

App

endi

x T

able

5.3

.2 T

wo-

way

test

s of s

igni

fican

ce b

etw

een

leve

ls o

f CD

Thre

sh a

nd D

ispD

ist a

t eac

h le

vel o

f Res

Dis

p fo

r the

dry

(tab

le a

) and

wet

(tab

le b

)

seas

on.

243 (b

)

W

et se

ason

scen

ario

/ Re

sDisp

hig

h (1

-100

)

Disp

Dis

t

CD

Thre

sh

5k

10k

15

k

st

rict

m

oder

ate

lo

ose

CD

Th

resh

m

od

(5,6

40)

stric

t (9

7)

CD

Th

resh

m

od

(10,

737)

st

rict

(100

) C

D

Thre

sh

mod

(1

2,54

5)

stric

t (1

01)

D

isp

Dist

10

k (1

00)

5k

(97)

10k

(10,

737)

5k

(5

,640

) D

isp

Dist

10

k (1

5,56

3)

5k

(8,8

13)

loos

e (8

,813

) **

* **

* lo

ose

(15,

563)

**

* **

* lo

ose

(18,

455)

**

* **

*

15k

(101

) p

= 0.

443

15k

(12,

545)

**

* **

* 15

k (1

8,45

5)

***

***

mod

( 5

,640

)

***

mod

(1

0,73

7)

**

* m

od

(12,

545 )

***

10

k (1

00)

10k

(10,

737)

***

10k

(15,

563)

***

Wet

seas

on sc

enar

io /

ResD

isp m

oder

ate

(1-2

1)

Disp

Dis

t

CD

Thre

sh

5k

10k

15

k

st

rict

m

oder

ate

lo

ose

CD

Th

resh

m

od

(4,2

38)

stric

t (6

4)

CD

Th

resh

m

od

(13,

588)

st

rict

(71)

C

D

Thre

sh

mod

(1

39,2

46)

stric

t (7

2)

D

isp

Dist

10

k (7

1)

5k

(64)

10k

(13,

588)

5k

(4

,238

) D

isp

Dist

10

k (1

8,81

4)

5k

(6,2

66)

loos

e (6

,266

) **

* **

* lo

ose

(18,

814)

**

* **

* lo

ose

(208

,590

) **

* **

*

15k

(72)

p

= 0.

353

15k

(139

,246

) **

* **

* 15

k (2

08,5

90)

***

***

mod

(4

,238

)

***

mod

(1

3,58

8)

**

* m

od

(139

,246

)

***

10

k (7

1)

10k

(13,

588)

***

10k

(18,

814)

***

Wet

seas

on sc

enar

io /

ResD

isp lo

w (1

-5)

Disp

Dis

t

CD

Thre

sh

5k

10k

15

k

st

rict

m

oder

ate

lo

ose

CD

Th

resh

m

od

(4,8

11)

stric

t (7

9)

CD

Th

resh

m

od

(36,

017)

st

rict

(95)

C

D

Thre

sh

mod

(4

0,02

7)

stric

t (1

00)

D

isp

Dist

10

k (9

5)

5k

(79)

10k

(36,

017)

5k

(4

,811

) D

isp

Dist

10

k (7

6,94

2)

5k

(19,

191)

loos

e (1

9,19

1)

***

***

loos

e (7

6,94

2)

***

***

loos

e (3

34,6

44)

***

***

15

k (1

00)

p =

0.18

6 **

15

k (4

0,02

7)

***

***

15k

(334

,644

) **

* **

*

mod

(4

,811

)

***

mod

(3

6,01

7)

**

* m

od

(40,

027)

***

10

k (9

5)

*

10k

(36,

017)

***

10k

(76,

942)

***

Due

to o

bser

ved

inte

ract

ion

effe

cts (

Fig.

6),

we

test

ed c

hang

es to

eith

er v

aria

ble

for e

ach

leve

l of t

he o

ther

two

varia

bles

. Med

ian

patc

h co

nnec

tivity

for e

ach

fact

oria

l co

mbi

natio

n is

repo

rted

in b

rack

ets (

size

of c

onne

cted

hab

itat c

ompo

nent

s in

km2 )

. Pai

rwis

e p

valu

es h

ave

been

cal

cula

ted

usin

g th

e M

ann-

Whi

tney

U-te

st o

f sig

nific

ance

. Th

ese

are

only

repo

rted

if a

thre

e-w

ay K

rusk

a-W

allis

rank

sum

test

indi

cate

d a

sign

ifica

nt e

ffect

with

in a

fact

oria

l sub

set.

Oth

erw

ise,

the

p va

lue

of th

is te

st is

repo

rted.

*p

< .0

5, *

*p <

.01,

***

p <

.001

.

Modelling seasonal habitat suitability and connectivity ...

Documents

Transcript of Modelling seasonal habitat suitability and connectivity ...