Characterizing the Spatial Distribution of Giant Pandas in China using MODIS Data and

Landscape Metrics

Xinping Ye March, 2008

Course Title: Geo-Information Science and Earth Observation

for Environmental Modelling and Management Level: Master of Science (Msc) Course Duration: September 2006 - March 2008 Consortium partners: University of Southampton (UK)

Lund University (Sweden) University of Warsaw (Poland) International Institute for Geo-Information Science and Earth Observation (ITC) (The Netherlands)

GEM thesis number: 2006-30

Characterizing the Spatial Distribution of Giant Pandas in China using MODIS Data

and Landscape Metrics

by

Xinping Ye Thesis submitted to the International Institute for Geo-information Science and Earth Observation in partial fulfilment of the requirements for the degree of Master of Science in Geo-information Science and Earth Observation for Environmental Modelling and Management Thesis Assessment Board Name Examiner 1 (Chair) Dr. J. (Jan) de Leeuw (ITC) Name Examiner 2 Dr. Ir. T.A. (Thomas) Groen (ITC) External Examiner Prof. Terry Dawson (University of Southampton) Primary Supervisor Prof. Dr. A.K. (Andrew) Skidmore (ITC) Secondary Supervisor Dr. A.G. (Bert) Toxopeus (ITC)

International Institute for Geo-Information Science and Earth Observation Enschede, The Netherlands

Disclaimer This document describes work undertaken as part of a programme of study at the International Institute for Geo-information Science and Earth Observation. All views and opinions expressed therein remain the sole responsibility of the author, and do not necessarily represent those of the institute.

Abstract

Although forest fragmentation has been recognized as one of major threats to wild panda population, little is known about the relationship between panda distribution and forest fragmentation. This study is unique as it presents a fruit attempt at understanding the role of forest fragmentation on panda distribution for the entire wild panda population. A remote sensing based approach was proposed for characterizing the distribution of giant panda habitat using the MODIS EVI time series and landscape metrics. A five-class land cover map was generated from a complete year of uninterrupted MODIS 250m EVI data for 2001 by combining ISODATA and a Neural Net classifier. Fifty-two class-level landscape metrics were calculated (for two forest classes) in FRAGSTAT 3.3 using a 3km×3km moving window; and eight metrics that best captured the variation in forest configuration among landscapes were selected as representative metrics by a landscape metrics reduction procedure. These eight metrics measure three aspects of the forests: patch area/density/edge, patch proximity/connectivity, and patch contagion/interspersion. All representative metrics were significantly different (P < 0.05) between panda presence and absence; and the spatial configurations of the forests among five mountain regions were also heterogeneous. The forest I class occupied by the giant panda in Qinling were less fragmented, while the forests occupied by the giant panda in Xiangling and Liangshan were most fragmented among five regions.

Using a forward stepwise logistic regression procedure, four landscape metrics were significant at P < 0.01 and included into the final logistic regression model, i.e. largest patch index, edge density, clumpy index of the forest I class, and mean patch area of the forest II class. These metrics indicate that the giant panda appear sensitive to patch size and isolation effects associated with forest fragmentation, and the forest I plays an important role in distribution of giant pandas. The performance of logistic regression model was significantly improved by applying a knowledge-based control to the modelling (P < 0.05). However, a low R-square value (0.451) of the model indicated that the landscape metrics alone can not effectively explain the distribution of the giant panda. The findings of this study have implications for the design of effective conservation plans for wild panda population limited by forest fragmentation. The presented approach may be applied to routine habitat monitoring and habitat evaluation for the giant panda.

Keywords: Ailuropoda melanoleuca, giant panda, spatial distribution, forest spatial configuration, MODIS 250 m EVI, landscape metrics, logistic regression.

i

Acknowledgements

I would like to express my sincere appreciation to all those who helped and supported me in one or another way during this course of the study.

Special Thanks to the Erasmus Mundus GEM-MSc Consortium (University of Southampton, UK; Lund University, Sweden; Warsaw University, Poland; and ITC, The Netherlands) and the European Commission for fully sponsoring my MSc studies. It has been a wonderful and rewarding experience to study in four European countries.

I would like to thank my supervisors, Prof. A. K. Skidmore (Natural Resources Department, ITC) and Dr. A.G. Toxopeus (Natural Resources Department, ITC), for their time, patience, criticisms, and suggestions. I would have been lost without their guidance.

I wish to also thank Mr. Tiejun Wang (PhD candidate, ITC) and his wife Ms. Wen Xue (Foping Nature Reserve, China), for their invaluable knowledge sharing and suggestions, concern and encouragement while pursuing this study.

I would also like to express my sincere appreciation to: Dr. Changqing Yu (Tsinghua University, China) for constructive suggestions and technical supporting; Mr. Xuelin Jin (Forestry Department of Shaanxi Province, China) for technical supporting and knowledge sharing; Dr. Lars Eklundh (Lund University) for helping me to manage the TIMESAT program; Dr. Xuehua Liu (Tsinghua University, China) for suggestions and encouragement; Mr. Zhanqiang Wen (State Forestry Administration of China) and Mr. Yange Yong (Foping Nature Reserve, China) for providing helps during the fieldwork in China.

Thanks also go to Prof. Peter Atkinson, Prof. Petter Pilesjo, Prof. Katarzyna Dabrowska, Andre Kooiman, Stef Webb, Karin Larsson, and Jorien Terlouw, for their support during the course of the study.

To my family and friends, thanks for your love, supports and encouragement in all my endeavors. To my fellow GEM students, thanks for the wonderful time together.

ii

Table of contents

1. Introduction ............................................................................................ 1 1.1. Background and Significance.................................................................. 1

1.1.1. The giant panda .................................................................................. 1 1.2. Research problem .................................................................................... 4 1.3. Research objectives ................................................................................. 5

1.3.1. General objective ................................................................................ 5 1.3.2. Specific objectives .............................................................................. 5

1.4. Research questions .................................................................................. 5 1.5. Hypotheses .............................................................................................. 5 1.6. Research approach................................................................................... 6

2. Materials and methods............................................................................ 7 2.1. Study area ................................................................................................ 7 2.2. Data description....................................................................................... 8

2.2.1. Time-series MODIS 250m EVI data .................................................. 8 2.2.2. Giant panda distribution data.............................................................. 9 2.2.3. Ancillary data ..................................................................................... 9

2.3. Reconstruction of cleaned MODIS EVI time series.............................. 10 2.4. Land cover classification....................................................................... 11

2.4.1. Land cover categories ....................................................................... 12 2.4.2. Reference data extraction.................................................................. 13 2.4.3. Principal component transformation of EVI time series................... 13 2.4.4. Classification procedure ................................................................... 14 2.4.5. Accuracy assessment ........................................................................ 16

2.5. Quantifing the spatial configuration of the forests ................................ 16 2.5.1. Landscape metrics computation........................................................ 16 2.5.2. Landscape metrics extraction............................................................ 18 2.5.3. Metric reduction analysis.................................................................. 18

2.6. Linking the forest spatial configuration to the giant panda ................... 19 2.6.1. Extracting presence-absence data for the giant panda ...................... 19 2.6.2. Significance testing........................................................................... 21 2.6.3. Mapping the presence-absence of giant pandas................................ 21

3. Results .................................................................................................. 24 3.1. Land cover classification....................................................................... 24 3.2. Reduction of Landscape metrics............................................................ 25 3.3. Comparisons of the forest spatial configuration .................................... 28

3.3.1. Between presence and absence of giant pandas................................ 28

iii

3.3.2. Panda presence in five mountain regions.......................................... 31 3.4. Logistic regression analysis of panda presence-absence data................ 35 3.5. Mapping the presence-absence of the giant panda ................................ 36

4. Discussion ............................................................................................ 40 4.1. Land cover classification from MODIS 250 m EVI time series............ 40 4.2. Quantification of the spatial configuration of forest landscape ............. 41 4.3. Spatial configuration of the forests occupied by giant pandas............... 42 4.4. Can the forest configuration predict the panda’s presence-absence?..... 44

5. Conclusion............................................................................................ 45

References .................................................................................................... 47

Appendices ................................................................................................... 54

iv

List of figures

Figure 1-1 Historic distribution of the giant panda and its present-day distribution ..1

Figure 1-2 A brief conceptual framework of this study..............................................6

Figure 2-1 Map of the study area delineated in red polygon. .....................................7

Figure 2-2 Sample EVI profiles showing noise removal during filtering and cleaning. ........................................................................................................11

Figure 2-3 A flowchart showing the hybrid method of land cover classification. ....12

Figure 2-4 Plots of the eigenvalues and accumulated percentage of variance against the corresponding principal component number for the study area...............14

Figure 2-5 Moving window computation .................................................................16

Figure 2-6 A flow chart showing the processes of linking the forest spatial configuration to the giant panda and mapping the presence-absence for giant pandas. ..................................................................................................20

Figure 3-1 Output map of five-class land cover classification of the study area. .....24

Figure 3-2 Boxplots of the eight metrics between panda presence and absence. .....30

Figure 3-3 Boxplots of eight metrics among the panda presence areas in five mountain regions. ..........................................................................................33

Figure 3-4 Suitability for giant pandas predicted by logistic regression model using landscape metrics statistic....................................................................37

Figure 3-5 Presence-absence of the giant panda predicted by logistic regression model (threshold = 0.5) .................................................................................38

Figure 3-6 Changes of overall mapping accuracy, sensitivity and specificity with different thresholds........................................................................................39

v

vi

List of tables

Table 2-1 Description of the land cover classification categories used in this study and extracted reference data for each land cover class. .................................13

Table 2-2 The Jeffries-Matusita distance and transform divergence separability matrix of pairwise land cover classes. ...........................................................15

Table 2-3 Class-level landscape metrics used in the study. ......................................17

Table 2-4 Statistics for extracted presence-absence data of the giant panda in five regions. ..........................................................................................................20

Table 2-5 Conditional probabilities of terrain factors in five mountain regions .......23

Table 3-1 Confusion matrix for five-class land cover classification of the study area. ...............................................................................................................25

Table 3-2 Factor analysis for 16 landscape metrics measuring the forest I class and varimax rotation of the first four factors........................................................26

Table 3-3 Factor analysis for 18 landscape metrics measuring the forest II class and varimax rotation of the first four factors........................................................27

Table 3-4 Representative metrics for characterizing spatial configurations of forest I and forest II at the home-range scale...........................................................28

Table 3-5 Collinearity statistics for eight representative metrics of presence-absence data of the giant panda. ....................................................................28

Table 3-6 Summary statistics and results of Brown-Forsythe's F test and Mann-Whitney test for metrics between presence and absence of the giant panda..29

Table 3-7 Summary statistics and Brown-Forsythe's F test for eight metrics among the panda presence in five mountain regions.................................................32

Table 3-8 Mean difference of each of eight metrics from pairwise multiple comparisons of the panda presence in five mountain regions. ......................34

Table 3-9 Statistics for the metrics with significant difference in pairwise mountain regions. ..........................................................................................................34

Table 3-10 Parameter estimates of the final logistic regression model.....................35

Table 3-11 Confusion matrices of the logistic regression modelling........................36

Table 3-12 Overall accuracy, sensitivity and specificity, kappa and kappa variance of the logistic regression modelling with and without knowledge-based control............................................................................................................36

1. Introduction

1.1. Background and Significance

1.1.1. The giant panda

The giant panda (Ailuropoda melanoleuca), one of the world’s most endangered mammals, is a robust animal with a distinctive black and white coat. The panda’s diet is very specific, consisting almost entirely of various bamboo species found in high mountain areas. The panda’s unique eating habits have earned it the epithet ‘bamboo bear’ among local people. Low in nutrients, bamboo provides enough nourishment for survival but little extra. However, giant pandas have developed adaptations to this diet. Usually, there are two or more bamboo species growing in any one area of panda habitat, enabling pandas to shift to another species when one flowers and dies, as bamboos do every 30 to 120 years. Giant pandas are solitary animals. Each adult has a well-defined home range, usually large, can up to 30km2.

Figure 1-1 Historic distribution of the giant panda and its present-day distribution delineated in red patches. (Adapted from Servheen et al., 1999)

Fossil evidence suggests ancestors of the giant panda were widely distributed over much of eastern and southern China as far north as Beijing (Schaller, 1994). Fossils were often found at the elevation of 500-700m in warm temperate or subtropical forests. Remarkable changes in the panda’s range may have happened rather recently. Much of the habitat loss occurred over the past several hundred years due to a dramatic increase in China’s human population and encroachment into the panda’s historical range (Sedano et al., 2005). Today this range is restricted to the temperate montane forests across five separate ranges (Figure 1-1), with a height of

1

1000 to 3500m where bamboo is the dominant understory forest plant (Schaller, 1994; Hu, 2001). According to the Third National Giant Panda Census accomplished by 2003, the total distribution area was estimated at about 23,000 km2, and around 45% of it is protected with a protected area network of 40 nature reserves (State Forestry Administration of China, 2006). Although the number of giant panda individuals seems to have increased, their distribution is discontinuous, with 24 isolated populations, and population density varies in different ranges (State Forestry Administration of China, 2006). If population isolation and habitat loss continue, the long-term viability of this species will be heavily threatened.

Since 1980s, numerous panda related scientific researches have focused mainly on the panda biology and behavior, such as reproduction, feeding behaviour, habitat utilization and selection (Hu, 2001). These biological studies are necessary but insufficient for effective conservation. Although habitat fragmentation and loss have already been recognized as major threats that pose a great danger for panda population (Hu, 2001; Pan, 2001), no quantitative and systematic research has been undertaken due to the low effectiveness of conventional methods in generalization of relationships between fragmentation and panda population. Benefited from some advanced technologies such as remote sensing and Geographic Information Systems (GIS), a few studies have been undertaken to spatially analyze, model and map panda habitat at the site-based level (Liu et al., 2000; Liu, 2001; Wang, 2003; Lindburg and Baragona, 2004). But little information is provided for clear insights into population-level implications of habitat fragmentation at the regional scale. Several studies attempted to design comprehensive conceptual frameworks for national-level panda conservation (Liu et al., 1999; Loucks et al., 2003; Lu et al., 2003; Xu et al., 2006). However, these frameworks seem too difficult to implement due in part to the lack of information about the response of the giant panda to habitat loss and fragmentation. An increasing need for effective conservation calls for more effective approaches to help understanding the relationship between the giant panda and forest landscape structure and pattern.

Recent innovations in the theories of landscape ecology and the technologies of remote sensing and GIS offer a promising approach for quantitative characterization of relations between the giant panda and landscape structure and pattern. Landscape ecology is an interdisciplinary branch of science that addresses how spatial variation in the landscape affects ecological processes such as the distribution and flow of energy, materials and individuals in the environment (Forman and Godron, 1986; Turner, 1989b; Gustafson, 1998). The basic concepts in landscape ecology are structure, function, and change. The structure is further divided into composition, configuration, and connectivity (Taylor et al., 1993; Merriam, 1995; Bennett, 1998).

2

Composition describes the proportion of different landscape elements without any reference to their location, while configuration explicitly locates landscape elements with respect to other elements in a spatial context (Dunning et al., 1992). The principle aspects of configuration include patch size and shape, patches distribution and density, isolation, and contagion. Connectivity characterizes processes in the landscape that contribute to a functional demographic unit (Merriam, 1984). For quantitative research, landscape metrics were developed to describe composition, configuration, and connectivity. These indices allow the interpretation of landscapes, and further assist land managers in applying land management plans to solve environmental problems (Jones, 2002).

With the development of landscape ecology, a large number of landscape metrics have been proposed to quantify specific spatial characteristics of patches, classes of patches, or entire landscape. Simultaneously, extensive applications of landscape metrics have been made to quantify landscape structure as well as the relationships with ecological processes (Gustafson, 1998). Many studies have been undertaken to link landscape structure with species to predict dispersal, abundance, distribution, and survival probability of species or population in heterogeneous landscapes (Taylor et al., 1993; With and King, 1997; Corsi et al., 2000; Griffiths and Mather, 2000; Kupfer and Kneitz, 2000; With, 2002). Most of these studies revealed significant statistical relations between some landscape indices and response variables such as species distribution and dispersal, suggesting the potential of landscape indices to assess the interactions amongst species, habitat, and landscape (Fortin et al., 2003; Jordán et al., 2003; Li et al., 2005; Koper et al., 2007).

The landscape metrics can be extracted from a land use/land cover (LULC) map derived from remotely sensed data. Because most landscape metrics are scale-dependent and landscape elements are species/process-specific (Cain et al., 1997; Saura, 2004), an appropriate spatial resolution of remotely sensed data and subsequent land cover units are important to the application of landscape metrics. The remotely sensed data should be carefully selected with the consideration of target species/process.

Over the past decades, medium resolution data from Landsat TM and ETM+ (30 m), and coarse resolution data from AVHRR (1 and 8 km), have been broadly used for classifying LULC from local to global levels (Reybenayas and Pope, 1995; Barbosa et al., 1999; DeFries and Chan, 2000; Giannetti et al., 2001; Wilson and Sader, 2002; Stehman et al., 2003; Jiang et al., 2004; Lo and Choi, 2004). However, the land cover classification, with the purpose of mapping and monitoring for a regional-distributed species, requires the remotely sensed data have wide geographic

3

coverage and appropriate spatial resolution. Remotely sensed data from Landsat TM and ETM+ and AVHRR data are limited for such classification due to the coarse spatial/temporal resolution, data availability and cost.

The Moderate Resolution Imaging Spectroradiometer (MODIS) offers new options for large area LULC classification by providing broad geographical coverage (swath width 2330 km), intermediate spatial resolution (250m, 500m, 1km), high temporal resolution, and cost-free data (Friedl et al., 2002; Justice et al., 2002; Yang and Lo, 2002; Price, 2003). Time-series MODIS 250 m Vegetation Index (VI) datasets, including NDVI and EVI, have been confirmed to be effective in land cover and vegetation mapping across a large area (Tarnavsky et al., ; Wessels et al., 2004; Liu and Kafatos, 2005; Knight et al., 2006; Houborg et al., 2007; Tottrup et al., 2007; Wardlow et al., 2007). During plant growth periods, different vegetation types can be distinguished by time-series VI data acquired at each individual phase (Boles et al., 2004). However, NDVI has several limitations that affect the accuracy of classification, including sensitivity to atmospheric conditions and the soil background, and a tendency to saturate at high biomass levels (Huete et al., 1997; Huete et al., 2002). The EVI is designed to minimize the effects of the atmosphere and soil background, and to enhance the green vegetation signal (Huete et al., 2002). Gao et al. (Gao et al., 2000)found that EVI was more responsive to canopy structure (e.g., LAI, plant physiognomy, and canopy type) variations than NDVI. Several efforts have illustrated the usefulness of MODIS time-series EVI data in regional-scale LULC classification (Bagan et al., 2005; Liu and Kafatos, 2005; Xavier et al., 2006; Wardlow et al., 2007).

1.2. Research problem

Knowledge of the giant panda distribution is fundamental to conserve this endangered species and ensure its long-term survival. However, little is known about the distribution pattern of the giant panda at a regional scale, as previous studies have focused on the relationship of panda occurrence and micro-environmental factors at the site-based level. In other words, to date no studies have attempted to address the panda distribution with relation to the spatial configuration of forested landscapes, e.g. forest patch size, patch isolation and aggregation. There is a clear need to uncover the underlying relationship between the heterogeneity of the forest landscape and distribution pattern of giant pandas for more effective conservation of the wild panda population.

4

1.3. Research objectives

1.3.1. General objective

The general objective of this research is to characterize current spatial distribution of giant pandas in China in relation to the spatial configuration of forested landscape using MODIS EVI time-series data and landscape metrics.

1.3.2. Specific objectives

• To identify and quantify the spatial characteristics of the forest and forest fragments within and surrounding the present-day distribution area of giant pandas.

• To explore the relationship between the giant panda distribution and various metrics measuring the forest fragmentation over the giant panda’s current-day range.

1.4. Research questions

• What are the main spatial configuration of the forested landscapes in the study area, with respect to patch size, shape and connectivity?

• What is the relationship between the giant panda distribution and the patch size of the forest landscapes?

• What is the relationship between the giant panda distribution and the patch shape of the forest landscapes?

• What is the relationship between the giant panda distribution and the patch connectivity of the forest landscapes?

1.5. Hypotheses

The research hypotheses are:

• The distribution of the giant panda is significantly and positively related to forest patch size and connectivity.

• Giant pandas are absent from isolated forest patches (nearest-neighbor distance >1500 m) that are smaller than 3 km2.

• A single large forest patch is more favourable for the giant panda than an aggregation of several small patches.

5

1.6. Research approach

The conceptual framework of this research is shown in Figure 1-2.

Figure 1-2 A brief conceptual framework of this study.

6

2. Materials and methods

2.1. Study area

The study area includes 45 administrative counties of three provinces in China, which covers the entire panda distribution area (roughly 28°00′- 34°30′ N and 101°00′- 109°00′ E). The area cross five regions along the eastern edge of the Tibetan Plateau: Qinling in Shaanxi Province, Minshan in Gansu and Sichuan Provinces as well as Qionglai, Xiangling and Liangshan in Sichuan Province (Figure 2-1). This area is also the investigation area of the 3rd National Panda Survey (State Forestry Administration of China, 2006). The total area is around 160000 km2. The elevation area ranges from 560 m to 6500 m. The mountain regions are described in detail as below.

Figure 2-1 Map of the study area delineated in red polygon. The image presented is MODIS 250m EVI 16-day composite during Julian day 193-209 of 2001.

7

The Qinling region, which runs from east to west, is the northernmost area of the present-day distribution of the giant panda. The climate in this region is mild and moist, and terrain is relatively flat and covered with evergreen broadleaf and upland conifer forests. The bamboo species Bashania fargesii and Fargesia qinlingensis thrive in this region (Ren, 1998). Giant pandas are distributed mainly on the southern slopes, but they are sometimes found on northern slopes. The density of giant pandas in the Qinling is the highest among five main distribution areas (State Forestry Administration of China, 2006).

The Minshan region is located to the southwest of the Qinling and extends from north to south. The west side is steep and dry, whereas the eastern slopes are gentler and moister. The pandas are distributed over the area from the Baishuijiang watershed in the southern part of Gansu Province to the upper Pei River drainage in the central part of Sichuan Province. Dense coniferous forests grow in the middle and upper elevations with an understory of arrow bamboo (China Vegetation Compiling Committee, 1980). In all of these areas, pandas are found mainly on the climatically more favourable eastern slopes. The population is estimated to be the largest of any of the five regions (State Forestry Administration of China, 2006).

The Qionglai region is located to the southwest of the Minshan, extending from northeast to southwest. This area, together with Minshan region, flanks the eastern rim of the Tibetan Plateau, and has a steep terrain, and cool and humid climate. Although located at subtropical latitudes, significant snow accumulates above 3000 m in winter. Forest composition and zonation in this region are similar to that of Minshan region. Pandas are found on the flat slopes of the upper Minjiang River and farther west and south areas (Hu, 2001).

The Xiangling and Liangshan regions are the most southernly existing panda habitat between the Qionglai Mountains and in the upper Yangtze River. Throughout the regions, the valley floors are quite dry, and cultivation extends far up the sides of the mountains (Hu, 2001). Vegetation is mainly broadleaf evergreen forest, whereas upper shadow conifer forests dominate in the higher elevation (China Vegetation Compiling Committee, 1980). Both regions are reported to be heavily fragmented panda habitat due to human activity such as road construction, farming, mining and overuse of forest and wood products (State Forestry Administration of China, 2006).

2.2. Data description

2.2.1. Time-series MODIS 250m EVI data

The multi-temporal images used in this study were acquired from the NASA Terra satellite’s Moderate Resolution Imaging Spectroradiometer (MODIS) sensor. A

8

time-series of 16-day composite MODIS 250 m EVI data (MOD13Q1 V005) for calendar year 2001 through 2003 were downloaded from the NASA website (http://edcimswww.cr.usgs.gov/pub/imswelcome), consisting of 69 composite periods and four tiles (h26v05, h26v06, h27v05, and h27v06). The MODIS 250 m EVI data have high pixel geolocational accuracy (±50 m (1σ) at nadir), so the influence of the changes due to geometric inaccuracies between composite periods in the time series are minimal (Wolfe et al., 2002). For each composite period, the MODIS 250 m EVI data were extracted by tile, mosaicked, reprojected from the Sinusoidal to the Albers Equal Area Conic projection for China using a nearest neighbour operator, and then subset to the study area.

2.2.2. Giant panda distribution data

The giant panda distribution data for this study were derived from the Third National Panda Survey Database established by the State Forestry Administration of China. The database contains giant panda occurrence data and relevant habitat information collected in 1999-2002 (State Forestry Administration of China, 2006). All data have been digitized and stored in vector format. Data used in this study include the occurrence of giant pandas, investigation areas and habitat of the giant panda.

2.2.3. Ancillary data

The ancillary data used in this study include: 1) the National Land Cover Dataset of China, 2) geophysical data of study area, and 3) pixel reliability layer included in MODIS 250m EVI products.

1) The National Land Cover Dataset of China (NLCD-2000) was used as reference land cover data for signature training and accuracy assessment. The NLCD-2000 dataset was developed through the National Land Cover Project that was organized by the Chinese Academy of Sciences. Over 500 Landsat TM images acquired in 1999 and 2000 were used to generate 1:100 000 scale land cover maps in China (Liu et al., 2002). The NLCD-2000 dataset comprised six classes at level I and 25 classes at level II; data classes represent a mix of land use and land cover, with an overall accuracy of 81% for China (Liu et al., 2002; Boles et al., 2004). It is temporally and spatially adequate for signature training and accuracy assessment in this study.

2) The geophysical data involved in this study includes the Digital Elevation Model (DEM) and the Bio-Climatic Division Map of China. The DEM was taken from the Shuttle Radar Topography Mission (SRTM) seamless digital topographic data, with a ground resolution of 3 arc second (http://www2.jpl.nasa.gov/srtm/dataprod.htm). Data were mosaicked, reprojected from the Geographic to the Albers Equal Area Conic projection for China, and then subset to the study area. The Bio-Climatic

9

Division Map of China comprised nine bio-climatic regions defined by the criteria of temperature and moisture conditions (Liu et al., 2003). These data were used for land cover classification and spatial analyses of panda occurrences.

3) The V005 MOD13 products include a summary quality layer pertinent to vegetation indices, the pixel reliability layer. This layer contains ranked values describing the cloud status for each pixel in the corresponding composite (Didan and Huete, 2006). In this study, per-pixel reliability data were used to improve the fitting of the cloudy MODIS EVI values by weighting the original EVI value based on cloud status.

2.3. Reconstruction of cleaned MODIS EVI time series

The MODIS EVI products have been geometrically rectified and corrected for atmospheric scattering and absorption (Vermote et al., 2002). However, there is still a certain level of noise persisting in the EVI data, mainly caused by remnants of clouds and its shadows. Thus, the TIMESAT package (Jönsson and Eklundh, 2004) was employed to reconstruct a clean, smooth time series of EVI data from three complete years of EVI data (n=69). TIMESAT is a program for analyzing time series of satellite sensor data and provides an adaptive Savitzky–Golay smoothing filter using local polynomial functions in the fitting of upper envelope of the vegetation index data (Jönsson and Eklundh, 2004). The most important input parameters are the cut-off for spike and the window size for the Savitzky-Golay filter (Chen et al., 2004). After running several tests on sample pixels, the cut-off for spike was set to 1.5 (based on Chen et al (2004) as well as maximum per-pixel reliability), and the window size was set to 6, 7, and 8 for three fitted steps, respectively. Furthermore, the per-pixel reliability layer was used to estimate the uncertainty of each data point so that a higher weight was given to cloud-free pixels than to pixels with a mixed cloud cover, while fully cloudy pixels were excluded. Figure 2-2 shows the sample of EVI time series profiles reconstruction by the TIMESAT.

Additionally, three annual seasonal parameters were extracted from the resulting smoothed EVI time series to enhance land cover classifying: annual EVI maximum (MAX), the maximum value of the fitted function; annual EVI base (BASE), defined as the average value of the left and right minima; and the annual EVI amplitude (AMP), obtained as the difference between the maximum EVI and base EVI value. These three measures represent the maximum, minimum, and the range of phenological activity.

10

Figure 2-2 Sample EVI profiles showing noise removal during filtering and cleaning. The x-axes in the chart are time series, consists of 69 dimensions of 16-day composites from 2001 to 2003. The blue circles represent original EVI values; the red squares represent fitted EVI values.

2.4. Land cover classification

Although the three-year MODIS EVI time series were acquired by reconstructing a cloud-free EVI time series, only the data in 2001 (n = 23), which were closer to the time of national panda survey and NLCD-2000 dataset, were directly used for the land cover classification. Figure 2-3 is the flowchart of the land cover classification.

11

Figure 2-3 A flowchart showing the hybrid method of land cover classification.

2.4.1. Land cover categories

The NLCD-2000 dataset comprised six classes at level I and 25 classes at level II (Appendix 1). For this study, the level II set of classes was too detailed for mapping from MODIS 250m EVI data, while the level I set was less species-specific. Since the giant panda is a forest inhabitant with a strong preference for high canopy cover (Hu, 2001), forest classes at level II and non-forest classes at level I in the categories of the NLCD-2000 dataset were adapted to produce a generalized land cover map depicting five land cover types that are species-specific and potentially could be classified using MODIS 250 m EVI data (Table 2-1).

12

Table 2-1 Description of the land cover classification categories used in this study and extracted reference data for each land cover class.

Number of pixels Class Description Training Test Forest I Natural or man-made forest with canopy

cover greater than 30% 1890 646

Forest II Lands covered by trees with canopy cover less than 30% or shrub

1587 529

Grassland Lands covered by herbaceous plant with coverage greater than 20%

901 330

Cropland Lands for agriculture 927 333

Nonvegeted Non-vegetated lands, including built-up, water body, bare land, ice/snow areas.

686 274

2.4.2. Reference data extraction

Reference data were derived from a reclassified NLCD-2000 map using a simple random sample stratified by land cover type by means of an add-ons sampling tool for ArcGIS 9.2 (http://arcscripts.esri.com/details.asp?dbid=15080). Considering the accuracy of NLCD-2000 map and 250 m spatial resolution of MODIS data, samples falling in land cover patches with a size less than 500 ha were discarded to ensure the reliability of reference data. In total, 8103 pixels were extracted, 5991 pixels for training and 2112 pixels for accuracy assessment (Table 2-1).

2.4.3. Principal component transformation of EVI time series

To reduce the data volume involved in the classification, the reconstructed EVI time series (n = 23) was transformed into principal components (PCs) using Principal Component Analysis (PCA). PCA aims at constructing a small set of derived variables (PCs) which summarize the original data, thereby reducing the dimension of the original data (Byrne et al., 1980; Richards, 1984). The derived variables are uncorrelated and are computed in decreasing order of importance; the first variable accounts for as much as possible of the variation in the original data, the second variable accounts for the second largest portion of the variation in the original data, and so on. Figure 2-4 shows the eigenvalues and accumulated percentage of variance against the number of PC bands. To avoid discarding important information contained in the EVI time series, the first five PCs were retained, accounting for 99.1% of variance in the cleaned EVI time series. The retained five PC bands and three annual seasonal parameters (AMP, MAX., and BASE) were stacked into one multilayer image as input for further land cover classification of study area.

13

99.1%

0.E+00

2.E+06

4.E+06

6.E+06

8.E+06

1.E+07

1 2 3 4 5 6 7 8 9 10PC band number

Eigr

nval

ue

60%

70%

80%

90%

100%

Acc

um. %Eigenvalue

Accum. %

Figure 2-4 Plots of the eigenvalues and accumulated percentage of variance against the corresponding principal component number for the study area, only the first 10 of 23 bands are presented in the figure.

2.4.4. Classification procedure

For the purpose of selecting an efficient classifier, spectral separability tests were applied to the signatures for each class to examine the statistical separability among five proposed land cover classes by using two widely-used measures, the Jeffries-Matusita distance and transform divergence. Values greater than 1.9 indicate that the class pairs have very good separability, whereas values less than 1 indicate low separability of the class pairs (Richards, 1999). The results of the separability tests (Table 2-2) indicated that that proposed five land cover categories do separate; however, there is some confusion between forest I and forest II. This confusion could be expected because the two classes are defined by the canopy density instead of forest compositions. In addition, the regional variations and altitudinal variations of land covers existed in the study area call for the aid of ancillary data. In this case, the Maximum Likelihood classifier would be inefficient to classify vegetated areas. Thus, a hybrid classification approach was developed by integrating unsupervised ISODATA with a (supervised) Neural Network classifier.

Firstly, an unsupervised Iterative Self-Organizing Data Analysis (ISODATA) was performed to separate vegetated areas and non-vegetated areas. ISODATA is used frequently in remote sensing analyses as the initial classification step. This method does not require training sets, merely grouping pixels based on spectral similarity, and providing the most comprehensive information on the spectral characteristics of an area. However, as an unsupervised classification method, ISODATA can have the risk of mismatching spectral signature clusters and thematic classes (Cihlar, 2000). Thus, a hyperclustering technique (Bauer et al., 1994) which generates a much higher number of clusters than classes was adopted to minimize the risk of

14

mismatch. The study area was clustered into 50 classes (with maximum 100 iterations) and then labelled as vegetated areas and non-vegetated areas.

Table 2-2 The Jeffries-Matusita distance and transform divergence separability matrix of pairwise land cover classes. Values in italic are Jeffries-Matusita distances.

Land cover class Forest I Forest II Grassland Cropland Nonvegetated

Forest I – 1.485 1.957 1.996 1.984

Forest II 1.595 – 1.687 1.961 1.746

Grassland 1.983 1.844 – 2.000 1.895

Cropland 2.000 2.000 2.000 – 1.975

Nonvegetated 2.000 1.847 2.000 2.000 –

For vegetated areas, especially for forest land cover, the ISODATA clustering technique alone was inadequate because of the similarity of their VI responses. Therefore, a supervised standard backpropagation Neural Net classifier (NNC) was employed to classify vegetated areas. Standard backpropagation NN is a popular learning method with some advantages including being nonparametric, easily adapted to different types of data and input structures, capability of identifying subtle patterns in training data, and fuzzy output values (Paola and Schowengerdt, 1995; Atkinson and Tatnall, 1997; Skidmore et al., 1997). In this study, the parameters of training rate, momentum, training threshold contribution, number of hidden layers and number of training iterations, were set to 0.05, 0.01, 0.5, 1 and 2000, based on experimental testing (Skidmore et al., 1997). In addition, the Bio-Climatic Division Map of China (Liu et al., 2003) and two topographical layers (DEM and slope gradient) were integrated into the classification because topography is believed to pose a stable control on vegetation in mountainous regions (Franklin, 1995; Deng et al., 2007), and the integration of terrain data has been shown to improve the classification accuracy especially when distinguishing between different classes is difficult due to low spectral separability (Janssen et al., 1990; Harris and J., 1995). Finally, the resulting land cover classes from ISODATA and NN classification were combined together using a binary union to produce a complete land cover map with five classes for the study area. The output classification map was mode filtered using window sizes 3×3 in order to eliminate any single noise pixels. All classification processes were performed using ENVI 4.3 (ITT Industries, Inc.)

15

2.4.5. Accuracy assessment

Land cover classification accuracy was assessed using a confusion matrix as cross-tabulations of the mapped class vs. the reference class (Story and Congalton, 1986; Congalton, 1991; Congalton and Green, 1999). Overall accuracy, producer’s and user’s accuracies, and the kappa statistic were then derived from the confusion matrices. The kappa statistic incorporates the off-diagonal elements of the error matrices (i.e., classification errors) and represents agreement obtained after removing the proportion of agreement that could be expected to occur by chance (Congalton, 1991; Skidmore et al., 1996).

2.5. Quantifing the spatial configuration of the forests

2.5.1. Landscape metrics computation

In line with the objectives of this study, the landscape spatial configurations were quantified at class level, while forest I and forest II were considered as focal land cover categories based on previous scientific research (Hu et al., 1985; Hu, 2001) and the report of Third National Panda Survey (State Forestry Administration of China, 2006). By using FRAGSTATS 3.3 (McGarigal and Marks, 1995), 26 class-level metrics were computed for each focal class. Metrics were grouped by: i) area and density; ii) shape and edge; iii) core area; iv) proximity and connectivity; and v) contagion and interspersion (Table 2-3). The definitions and descriptions of these indices are given by the FRAGSTATS user’s guide (McGarigal and Marks, 1995).

FRAGSTATS is a program for quantifying landscape structure that was developed by the Forest Science Department of Oregon State University, U.S.A (McGarigal and Marks, 1995). There are two versions of FRAGSTATS, Vector (ARC/INFO) and Raster (image maps) versions. In this study, the raster version of FRAGSTATS was used to compute metrics with a constant spatial unit and to create a continuous landscape metric surface for statistical analysis, by using a square moving window computing technique. Within each window, each selected metric were computed and the value returned to the focal (centre) cell (Figure 2-5). The moving window is passed over the grid until every positively valued cell containing a full window is assessed in, and the output is a new grid for each selected metric.

Figure 2-5 Moving window computation

16

Table 2-3 Class-level landscape metrics used in the study.

Group Metric name Value range Largest Patch Index 0 ~ 100 Patch Density > 0 Percentage of Landscape 0 ~ 100 Mean Patch Area CDS* Landscape Shape Index ≥ 1, without limit Edge Density ≥ 0, without limit

I. Area/Density/ Edge

Radius of Gyration Distribution CDS*

Contiguity Index CDS* Fractal Dimension Index CDS* Perimeter Area Ratio CDS*

II. Shape

Shape Index CDS*

Core Percentage of Landscape 0 ~ 100 Disjunct Core Area Density ≥ 0, without limit Disjunct Core Area Distribution CDS* Core Area Index CDS*

III. Core area

Core Area CDS*

Patch Cohesion Index 0 ~ 100 Connectance Index 0 ~ 100 Euclidian Nearest Neighbor Index CDS*

IV. Proximity/ Connectivity

Proximity Index CDS*

Aggregation Index 0 ~ 100 Clumpy Index -1 ~ 1 Landscape Division Index 0 ~ 1 Interspersion Juxtaposition Index 0 ~ 100 Percentage of Like Adjacencies 0 ~ 100

V. Contagion/ Interspersion

Splitting Index 1 ~ no. of cells *Regarding class distribution statistics (CDS) metrics, only the mean were computed.

It is critical that the defined landscape scale represents the organism or ecological phenomenon under study, otherwise the landscape patterns detected will have little ecological meaning (Turner, 1989a; Jelinski and Wu, 1996; Qi and Wu, 1996). Previous studies indicated that the size of giant panda’s territory ranges from 5 km2 to 30 km2, with an average home range of 10 km2 (Hu, 2001; Pan, 2001), therefore, the size of the moving window was set to 3 km × 3 km, so as to have an extent close to the size of panda’s home range; patch neighbours were delineated by an 8-cell

17

rule; and a search distance of 2 km was used in the calculation of proximity index. Additionally, all boundary weight values are equal to 1 (i.e. both class and landscape boundaries are treated as maximum-contrast patch edges).

2.5.2. Landscape metrics extraction

For a given confidence level, the larger the sample size, the smaller the confidence interval, and the sample size will need to be larger if the greater the variance of variables (Cormack et al., 1979; Baiulescu et al., 1991), Therefore, 4000 random points were sampled within the forested areas (forest I and forest II) with the minimum distance of 3 km between each other and minimum distance of 3 km to the forest border. Values of landscape metrics were then extracted for each point using the extraction tool in Spatial Analyst Tools of AcrGIS 9.2 (ESRI Inc.).

Giant pandas rarely occur in areas above 3500 m or with a slope greater than 50º (Hu, 2001); thus, points beyond these two thresholds were discarded. In addition, an “outlier treatment” was conducted using the Cook's distance and Leverage statistic, because unusual outliers can increase sample variance and radically alter the outcome of analysis. Points with a Leverage value above 0.2 or the Cook's distance larger than 4/n (where n is the number of cases) were treated as outliers and eliminated from the dataset (Sokal and Rohlf, 1994). These processes eliminated 265 points, reducing the sample size from 4000 points to 3735 points. The remaining points were used in metrics reduction analysis and extracting the presence-absence data for the giant panda.

2.5.3. Metric reduction analysis

2.5.3.1. Partial correlation analysis

Because some of the landscape metrics are closely correlated, some of them can be immediately eliminated by examining pairwise correlation coefficients of metrics; and therefore reduce multicollinearity in further regression analysis. Many metrics were found also correlated with the gradient of elevation, thus a two-tailed partial correlation analysis was employed to measure the association within landscape metrics, simultaneously controlling for the effect of elevation. Of the pairs of landscape metrics with correlation coefficients ≥ |0.9|, only one metric was retained; this criterion is commonly used to reduce multicollinearity to an ‘acceptable’ level (Riitters et al., 1995; Griffith et al., 2000). The criteria for determining which of the highly correlated metrics to retain were as follow: 1) most commonly used in literatures; 2) density metrics and distribution statistics metrics were preferred to absolute metrics.

18

19

2.5.3.2. Factor analysis

A factor analysis (FA) was performed to derive representative metrics from the remaining metrics. FA technique has been commonly used to reduce redundancy in the information provided by sets of landscape indices (McGarigal and McComb, 1995; Riitters et al., 1995; Cain et al., 1997). Based on the test for independence of samples (Durbin-Watson coefficient1, d = 1.68), the principal components method was employed to extract factors; and an orthogonal (varimax) rotation was used to facilitate the interpretation of underlying factors. Factors were retained by using two criteria: the shape of the screen-plot and Kaiser rule that the eigenvalue of the factor should be greater than 1.0 (Bulmer, 1967). The retained factors were interpreted by examining the loadings of each metric on each of the factors. The metrics with the highest absolute loading on each of retained factors were then defined as the representative metrics for corresponding factor and to be used as landscape variables in further analyses. If two metrics have equal loadings for a same factor, the same selection criteria mentioned above were applied.

2.6. Linking the forest spatial configuration to the giant panda

2.6.1. Extracting presence-absence data for the giant panda

The presence-absence data were extracted from the outlier-treated point dataset (n = 3735) with the support of the 3rd National Panda Survey Dataset. Firstly a 3-km buffer zone was created for giant panda spoors points to represent the areas occupied by the giant panda, i.e. territory of the giant panda (see Section 2.5.1). By overlapping random points with the buffer zone, the 1124 points that completely lay within the buffer zone were categorized as the ‘presence’ data for the giant panda. Since these panda spoors data were collected by an exhausted survey throughout the study area (State Forestry Administration of China, 2006), these points are sufficient to infer panda absence in the study area. Two criteria were used in panda ‘absence’ data selection: 1) located outside the buffer zone; and 2) at a minimum distance of 3 km to the boundary of the buffer zone. In total, 1278 points were selected from the outlier-treated point dataset as samples for panda absence. Data were then labelled with their corresponding ranges to facilitate further analyses (Table 2-4).

1 The Durbin-Watson coefficient, d, tests for serial independence of observations. The value ranges from 0 to 4. Values close to 0 indicate extreme positive autocorrelation; close to 4 indicates extreme negative autocorrelation; and close to 2 indicates no serial autocorrelation. As a rule of thumb, d should be between 1.5 and 2.5 to indicate independence of observations (Shaw and Wheeler, 1994).

Table 2-4 Statistics for extracted presence-absence data of the giant panda in five regions.

Region name Presence Absence Total Qinling 187 298 485 Minshan 482 429 911 Qionglai 336 260 596 Xiangling 47 118 165 Liangshan 72 173 245 Sum of samples 1124 1278 2402

Outlier-treated point dataset

Panda presence/absence

data

Panda occurrences dataset

Overlay

Representative metrics grids

Significance testing

Logistic regression

DEM & Slopelayers

validationMapping in

ERDAS

Panda occurrence probability map

Occurrence pattern of giant pandas

3 kmBuffering

Training datan=2000

Reference datan=402

Presence vs. Absence&

Pairwise of five ranges

Panda presence-absence

map

Knowledge-based rules

Figure 2-6 A flow chart showing the processes of linking the forest spatial configuration to the giant panda and mapping the presence-absence for giant pandas.

Furthermore, the spatial autocorrelation in the dataset was investigated by the Moran’s I statistic in ArcGIS 9.2 (ERSI, 2007). In addition, a diagnostic statistic -

20

the Tolerance and Variance-Inflation Factors (VIF) - was performed to examine whether the multicollinearity is problematic. The rule is that the multicollinearity is a problem when tolerance < 0.10 and/or VIF > 10 (Menard, 2002).

2.6.2. Significance testing

2.6.2.1. Between the presence and absence of the giant panda

The representative metrics of the panda presence/absence were compared to investigate which forest spatial patterns account for the presence/absence of giant pandas. Because some metrics did not meet the assumption of homogeneity of variances and some were non-normally distributed, a Brown-Forsythe's F test was employed to test whether landscape variables are significantly different between the presence and absence of the giant panda. The Brown-Forsythe's F test is more robust when groups are unequal in size and the absolute deviation scores (deviations from the group means) are highly skewed, causing a violation of the normality assumption. The Brown-Forsythe F test does not assume homogeneity of variances (Rutherford, 2001). In addition, a nonparametric Mann-Whitney U test was also performed for the presence-absence of the giant panda.

2.6.2.2. Panda presence in five mountain regions

Considering the extent of the study area, the significance of differences of panda presence in five mountain regions was examined using the Brown-Forsythe's F test. Because the sample sizes for the panda presence in five mountain regions were unequal (Table 2-4), the Games-Howell test was employed to determine how the metrics of panda presences differ among different mountain regions. The Games-Howell test is a post-hoc test designed for pairwise multiple comparisons with unequal variances. It is based on Welch’s correction to df with the t-test and uses the studentized range statistic. This test is recommended for the situation of unequal (or equal) sample sizes and unequal or unknown variances (Toothacker, 1993).

2.6.3. Mapping the presence-absence of giant pandas

2.6.3.1. Binomial logistic regression

In this study, binomial logistic regression was employed for delineating the relation between panda presence/absence and landscape variables (landscape metrics). Binomial logistic regression is a common statistical method used to estimate occurrence probabilities in relation to predictors (Cowley et al., 2000). It uses a logit link to describe the relationship between the response and the linear sum of the predictor variables, and estimate the probability of a certain event occurring (Shaw and Wheeler, 1994).

21

eep xaxaxaa

xaxaxaa

ii

ii

i +++++

++++=

L

L

22110

22110

1

where pi is the probability; xi is explanatory variable; ai is the constant.

By applying the regression equation, presence-absence of the giant panda were transformed into a continuous probability pi ranging from 0 to 1; pi close to 1 represents high probability of presence, and pi close to 0 represents high probability of absence. The pi, in this case, can also be regarded as an indicator of the suitability of forest spatial configuration for the giant panda at the home-range scale.

2.6.3.2. Fitting the logistic regression model

Stepwise model-fitting with forward selection was used to help construct a model with good fit to the data, in which the variable with the most significant change in deviance at each stage was incorporated into the model until no other variables were significant at the P < 0.05 level. The best model was selected mainly based on Nagelkerke R Square and Hosmer-Lemeshow goodness of fit test (Davis, 2002; O'Connel, 2006). The Nagelkerke R Square summarizes how much of the variability in the data is successfully explained by the model. Larger values of R-square (maximum value of 1) indicate that the model captures more of the data variability. Hosmer-Lemeshow goodness of fit examines how closely the observed and predicted probabilities match. The null hypothesis is “the model fits” and a P value > 0.05 is expected (Hosmer and Lemeshow, 2000). The presence-absence dataset (n = 2402) was randomly split into two parts, one (n = 2000) for model building, another (n = 402) for model evaluation. All statistical analyses were conducted in SPSS 15.0 (SPSS Inc. 2006)

2.6.3.3. Knowledge-based control for model implementation

Because the logistic regression model was only using landscape metrics, whereas the distribution of the giant panda was limited by a range of environmental conditions, the model may overpredict the landscape suitability regardless of the environmental tolerance of the giant panda. To mitigate the risk of overestimation, a knowledge-based control was developed by integrating the logistic regression model with terrain information (elevation and slope) in mapping the landscape suitability for the giant panda, described as below:

Pi' = pi × cele × cslp

where Pi' is adjusted probability; pi is the probability estimated by logistic regression model; cele is the conditional probability with respect to elevation; cslp is the conditional probability with respect to slope.

22

The knowledge-based rules for the control were formulated based on the integration of knowledge from several sources: 1) literatures (Hu, 2001; Pan, 2001); 2) detailed discussion with several specialists; 3) knowledge acquired from field observations; and 4) analyses of the 3rd national panda survey data. Table 2-5 shows the detailed conditional probability in terms of mountain region.

The spatial implementation of the logistic regression model was achieved in ERDAS IMAGINE 9.1 software (LLC, 2006).

Table 2-5 Conditional probabilities of terrain factors in five mountain regions

Conditional probability Terrain factors

Qinling Minshan Qionglai Xiangling Liangshan Elevation < 1200 m 0.10 0.01 0.01 0.01 0.01 1200 m ~ 2000 m 1.00 0.50 0.50 0.10 0.10 2000 m ~ 2500 m 1.00 1.00 1.00 1.00 1.00 2500 m ~ 3000 m 0.80 1.00 1.00 1.00 1.00 3000 m ~ 3500 m 0.01 0.80 0.80 0.80 0.80 > 3500 m 0.01 0.10 0.10 0.50 0.50 Slope < 10º 0.80 0.70 0.70 0.40 0.60 10º ~ 40º 1.00 1.00 1.00 1.00 1.00 40º ~ 50º 0.20 0.60 0.60 0.10 0.10 > 50º 0.01 0.10 0.10 0.01 0.01

2.6.3.4. Model evaluation

The performance of final logistic regression model was assessed by overall accuracy, sensitivity and specificity, kappa coefficient and Z-test using an independent presence-absence data (n = 402). Sensitivity is defined as the proportion of correctly predicted presence to the total number of presence in testing samples; and specificity is the proportion of correctly predicted absence to the total number of absence in testing samples (Fielding and Bell, 1997). The kappa coefficient and its variance (Cohen, 1960; Congalton, 1991; Skidmore et al., 1996) were computed and the effect of the knowledge-based control was examined through a z-statistic using kappa coefficients (Cohen, 1960; Congalton, 1991). The output of the logistic regression model is a continuous probability surface, whereas the evaluation is based on presence-absence. Thus a threshold of 0.5 was arbitrarily selected to convert the continuous probability surface to a discrete absence-presence map. A probability greater than or equal to 0.5 was coded as presence, and less than 0.5 was absence. However, the 0.5 value may not be optimal in all cases (Manel et al., 1999); therefore, a sensitivity analysis was conducted to consider thresholds from 0.3 to 0.7.

23

3. Results

3.1. Land cover classification

The final product of land cover classification for the study area is illustrated in Figure 3-1. Table 3-1 provides the complete confusion matrix of the classification and Table 3-2 shows the producer’s accuracy, user’s accuracy, overall accuracy and kappa coefficient estimates. The overall accuracy of the classification was 83.9%, and the kappa coefficient was 0.79. The forest I class was mainly confused with the forest II class, and to some extent with grassland. The forest II class was confused not only with the forest I class and grassland, but slightly with cropland. The most accurately mapped classes were cropland and nonvegetated. Both were mainly confused with grassland. No confusion was detected between the forest and nonvegetated lands.

Figure 3-1 Output map of five-class land cover classification of the study area.

24

Table 3-1 Confusion matrix for five-class land cover classification of the study area. Bold type indicates the numbers of correctly classified testing pixels.

Reference data Classified data Forest I Forest II Grassland Cropland Nonveg Total

Forest I 537 69 14 620 Forest II 94 424 25 6 549 Grassland 15 24 265 23 18 345 Cropland 12 9 296 7 324 Nonvegetated 17 8 249 274 Total 646 529 330 333 274 2112 Producer’s acc. (%) 83.1 80.2 80.3 88.9 90.9 User’s acc. (%) 86.6 77.2 76.8 91.4 90.9 Overall acc. (%) 83.9 Kappa 0.79

3.2. Reduction of Landscape metrics

The process of two-tailed partial correlation analysis with a threshold of r < |0.9| eliminated 18 metrics from the original 52 landscape metrics, leaving 34 metrics for subsequent factor analysis. A few remaining metrics were non-normally distributed (Skewness > |2.0|).

Table 3-2 shows the result of the factor analysis for 16 metrics measuring the forest I class. After factoring the correlation matrix by the principal components method, the first four factors are seen to have eigenvalues > 1.0 and together explained about 77.8% of the variation in the 16 landscape metrics. The factors were interpreted by examining the loadings of each metric on each of the first four factors after orthogonal rotation. The first factor was termed patch size/density of the forest I class because it is most correlated with Largest patch index, Mean contiguity index and Patch density. The second factor was termed patch compaction of the forest I class, because it is most correlated with Edge density and Landscape shape index which measuring the patch compaction. The third factor, termed patch aggregation, is highly correlated with Clumpy index and Splitting index. The fourth factor was termed as patch connectivity because metrics Mean proximity index and Connectance index are most associated with this factor.

25

26

Table 3-2 Factor analysis for 16 landscape metrics measuring the forest I class and varimax rotation of the first four factors. Factor loadings above |0.8| are underlined and factor loadings less than |0.2| are not presented.

Component Metrics

1st 2nd 3rd 4th Communality

Edge density 0.923 0.208 0.918 Large patch index 0.895 0.293 0.884 Landscape shape index -0.459 0.839 0.928 Patch density -0.801 0.218 -0.267 0.775 Mean contiguity index 0.881 -0.246 0.891 Mean shape index 0.728 0.449 0.761 Disjunct core area density 0.226 0.701 0.312 0.643 Mean disjunct core area 0.612 -0.464 0.612 Nearest neighbour index -0.665 -0.233 0.266 0.568 Mean proximity index 0.279 0.887 0.872 Splitting index -0.219 -0.867 0.806 Interspersion index 0.529 0.299 Clumpy index -0.292 0.870 0.845 Aggregation index 0.748 0.527 0.293 0.934 Connectance index 0.886 0.796 Patch cohesion index 0.563 0.348 0.632 0.285 0.919 Eigenvalue 5.191 3.535 2.096 1.630

% of Variance 32.45 22.10 13.10 10.19

Cumulative % 32.45 54.54 67.64 77.83

Extraction Method: Principal Components Method. Rotation Method: Varimax with Kaiser Normalization.

The result of factor analysis for 18 metrics measuring the forest II class is shown in Table 3-3. Again, the first four factors are seen to have eigenvalues > 1.0 and together accounted for about 78.3% of the variation in the 18 landscape metrics of the forest II class. The first factor is most correlated with mean Mean patch area, Mean contiguity index, and Mean shape index, which mainly measure the patch size and shape complexity of the forest II class. The second factor was termed patch connectivity of the forest II class because it is most associated with the Mean proximity index and Connectance index. The third factor is highly associated with Landscape shape index and Patch density and leading to the name of patch compaction of the forest II class. The fourth factor was termed patch aggregation of the forest II class because it is highly correlated with metrics Splitting index and Clumpy index.

27

Table 3-3 Factor analysis for 18 landscape metrics measuring the forest II class and varimax rotation of the first four factors. Factor loadings above |0.8| are underlined and factor loadings less than |0.2| are not presented.

Component Metrics

1st 2nd 3rd 4th Communality

Mean patch size 0.959 0.936 Edge density 0.492 0.805 0.950 Landscape shape index -0.209 0.949 0.947 Patch density -0.488 -0.524 0.528 0.791 Percentage of landscape 0.681 0.629 0.203 0.936 Mean contiguity index 0.873 0.223 0.843 Mean shape index 0.809 0.286 0.774 Core percentage of landscape 0.768 0.446 0.848 Disjunct core area density 0.295 0.678 0.375 0.703 Mean disjunct core area 0.624 -0.309 0.206 0.562 Nearest neighbour index -0.597 -0.512 0.621 Mean proximity index 0.899 0.867 Splitting index -0.285 -0.834 0.776 Interspersion index 0.208 0.098 Clumpy index 0.255 -0.258 0.857 0.896 Aggregation index 0.542 0.593 0.526 0.926 Connectance index 0.827 0.714 Patch cohesion index 0.433 0.598 0.284 0.519 0.895 Eigenvalue 8.178 2.821 1.698 1.388

% of Variance 45.43 15.67 9.44 7.71

Cumulative % 45.43 61.11 70.54 78.25

Extraction Method: Principal Components Method. Rotation Method: Varimax with Kaiser Normalization.

Based on the loadings of each metric on each of the first four factors, four metrics measuring the forest I class and four metrics measuring the forest II class that holding the highest absolute loading were selected as representative metrics to delineate the heterogeneity of the spatial configurations of forest I and forest II at the home-range scale (Table 3-4). In general, these eight representative metrics measure three aspects of the focal classes: patch area/density/edge, patch proximity/ connectivity, and patch contagion/ interspersion. For a better understanding these eight metrics, the detailed descriptions were attached in Appendix 3.

Partial correlation coefficients were examined among eight representative metrics by controlling for the effect of elevation. Overall, there were low pairwise correlations

(<|0.4|) for all pairs of eight representative metrics (see Appendix 4), indicating the lack of strong interaction or linear relationship between each other.

The evaluation for spatial autocorrelation indicated that the spatial autocorrelation was not problematic in the presence-absence dataset (Moran’s I = 0.03, Z-score = 1.91, P > 0.05). The result of the test for multicollinearity (Table 3-5) also indicated that eight metrics of the dataset were not affected by the multicollinearity (tolerance < 0.10 and/or VIF > 10 indicate the Collinearity may be a problem (Menard, 2002)).

Table 3-4 Representative metrics for characterizing spatial configurations of forest I and forest II at the home-range scale.

Component Forest I Forest II 1st Largest patch index Mean patch area 2nd Edge density Mean proximity index 3rd Clumpy index Landscape shape index 4th Mean proximity index Clumpy index

Table 3-5 Collinearity statistics for eight representative metrics of presence-absence data of the giant panda. Tolerance < 0.10 and/or VIF > 10 indicate the Collinearity may be problematic.

Metrics Tolerance VIF Of forest I Edge density 0.550 1.818 Largest patch index 0.553 1.807 Mean proximity index 0.666 1.500 Clumpy index 0.894 1.118 Of forest II Mean patch area 0.701 1.426 Landscape shape index 0.651 1.536 Mean proximity index 0.658 1.520 Clumpy index 0.780 1.282

3.3. Comparisons of the forest spatial configuration

3.3.1. Between presence and absence of giant pandas

The summary statistics and results of Brown-Forsythe's F test and Mann-Whitney U test for eight metrics between the presence and absence of the giant panda are shown in Table 3-6. All metrics differed significantly between the presence and absence of giant pandas (P < 0.05). The result of the nonparametric Mann-Whitney U test is

28

also evident that all metrics were significantly different between the panda presence and absence area.

With respect to the forest I, Edge density, Largest patch index, and Mean proximity index in the panda presence area were higher than the absences, indicating these features may be required by the giant panda. Since the metrics were computed with a landscape extent of 36 km2 (see Section 2.5.1), the largest patch size in the panda presence was on average 19.2 ± 7.8 km2, in contrast to the panda absence in which the largest patch size was 12.1 ± 9.2 km2. As depicted by the metrics Mean proximity index with a search radius of 2 km, the neighbourhood of forest I was more occupied by the patches of forest I in the panda presence, and those patches became closer and more contiguous in distribution. The patches of the forest I tended to aggregated (Clumpy index > 0) in both panda presence and absence areas, but patches occupied by the giant panda were less aggregated.

For the forest II class, different patterns were observed. The Mean patch area, Mean proximity index, and Clumpy index of the panda presence were lower than the absence’s, while the Landscape shape index had an inverse pattern. The mean patch size of forest II in the panda absence was significantly higher than in the panda presence. In this case, however, the standard deviation is much larger than the mean of Mean patch area, pointing to over-dispersion of the mean patch size.

Table 3-6 Summary statistics and results of Brown-Forsythe's F test and Mann-Whitney test for metrics between presence and absence of the giant panda.

Presence (n=1124) Absence (n=1278) Metrics

Mean S.D. Mean S.D. F Z

Of forest I Edge density 22.57 5.45 16.75 7.18 507.02* -20.24* Largest patch index 53.39 21.75 33.74 25.44 416.30* -19.10* Mean proximity index 16.10 10.89 9.94 9.36 217.95* -12.82* Clumpy index 0.38 0.12 0.44 0.26 62.07* -14.48* Of forest II Mean patch area 75.89 124.1 184.3 280.3 156.30* -14.38* Landscape shape index 4.01 0.79 3.69 0.82 97.94* -10.28* Mean proximity index 6.10 6.29 9.46 8.58 121.54* -8.81* Clumpy index 0.35 0.15 0.42 0.19 89.28* -13.12*

* Difference is significant at the 0.05 level.

The differences in range of the eight metrics between the panda presence and absence were depicted using boxplots, as shown in Figure 3-2. The median of the values indicate that metrics Mean patch area and Mean proximity index are slightly skewed, while other metrics are normally distributed or nearly normal-distributed.

29

Presence Absence0

25

50

75

100La

rges

t pat

ch in

dex

of fo

rest

I

Presence Absence

0

100

200

300

400

Mea

n pa

tch

area

of

fore

st II

Presence Absence0

10

20

30

Edge

den

sity

of f

ores

t I

Presence Absence1

2

3

4

5

Land

scap

e sh

ape

inde

x of

fore

st II

Presence Absence0

10

20

30

40

Mea

n pr

oxim

ity in

dex

of fo

rest

I

Presence Absence0

10

20

30

Mea

n pr

oxim

ity in

dex

of fo

rest

II

Presence Absence

0.2

0.4

0.6

0.8

1.0

Clu

mpy

inde

x of

fore

st I

Presence Absence

0.2

0.4

0.6

0.8

1.0

Clu

mpy

inde

x of

fore

st II

Figure 3-2 Boxplots of the eight metrics between panda presence and absence. The lines within the box are medians of the data; the upper and lower ends of the box are upper quartile and lower quartile; the upper and low whiskers are highest quartile and lowermost quartile.

30

3.3.2. Panda presence in five mountain regions

Summary statistics and results for the Brown-Forsythe's F test for eight metrics among the panda presence in the five mountain regions are shown in Table 3-7. Figure 3-3 details the range of the eight metrics among the panda presence in five ranges depicted using boxplots. All representative metrics were significantly different (P < 0.05) among the panda presence for the five mountain regions. The average largest patch size of the forest I class in Qinling was the highest (21.5 ± 8.1 km2), followed by Minshan (20 ± 7.1 km2), Qionglai (17.8 ± 8.1 km2), Xiangling (16.5 ± 6.3 km2) and Liangshan (15.5 ± 8.3 km2) in descending sequence. In contrast, the Edge density of forest I in Qinling was the lowest, significantly different (P < 0.05) to the edge density of the other four regions. The Mean proximity index of the forest I class in Minshan and Xiangling were higher than that for the other regions; but as for the forest II class, Qinling and Liangshan held a higher mean patch proximity. Both the forest I and forest II in Qinling were more aggregated (Clumpy index) than in other regions.

Table 3-8 shows the mean differences of each of representative metrics from pairwise multiple comparisons of the panda presence in five ranges by running the Games-Howell test. The significance of the mean difference for each metric varied among different pairs of ranges. In most case, as we can see, Qinling was significantly different (P < 0.05) from the other four regions; and in the cases of Edge density and Clumpy index of the forest I class, no significant mean differences were detected for pairs of regions except for Qinling. The metric Largest patch index of the forest I class was found no significant differences (P > 0.05) in two pairs, Qinling-Minshan and Xiangling-Liangshan. The Mean proximity index of forest I show the least variation among five mountain regions, with significant differences (P < 0.05) in only two pairs - Qinling-Minshan and Liangshan-Minshan. Of the eight metrics, only the Mean patch area and Mean proximity index of the forest II class were found to be significantly different (P < 0.05) for the pair of Xiangling and Liangshan. Table 3-9 details the statistics for the metrics with significant mean difference in pairwise mountain regions, as summarized from Table 3-8. Various metrics with significant mean difference in different pairs of ranges illustrates the subtle distinction of forest spatial configurations among the panda presence in the five mountain regions, however, only the spatial configurations in Qinling was distinct from other regions; and the mean differences of the most of eight representative metrics were insignificant between the panda presence in Xiangling and in Liangshan regions.

31

F

78.7

1* 13

.77*

54.3

6* 7

.916

* 27

.05*

23.7

6* 15

.43*

56.0

1*

S.D

. 4.

85

22.9

9 0.

11

8.92

73

.98

0.76

6.

79

0.11

Lian

gsha

n

Mea

n 24

.87

43.0

5 0.

34

13.3

7 87

.21

4.23

9.

19

0.39

S.D

. 4.

32

17.5

7 0.

10

9.39

30

.67

0.57

3.

62

0.11

Xia

nglin

g

Mea

n 25

.85

45.8

2 0.

34

18.5

2 40

.94

4.23

3.

71

0.32

S.D

. 4.

45

22.5

9 0.

10

10.9

4 11

5.7

0.68

6.

40

0.12

Qio

ngla

i

Mea

n 23

.54

49.5

8 0.

36

15.6

6 72

.32

4.07

6.

03

0.35

S.D

. 4.

85

19.6

2 0.

10

11.0

8 50

.92

0.85

5.

23

0.15

Min

shan

Mea

n 23

.49

55.6

7 0.

36

17.6

9 48

.23

4.10

4.

99

0.31

S.D

. 5.

50

22.6

0 0.

13

10.5

2 22

3.5

0.72

7.

62

0.13

Qin

ling

Mea

n 16

.97

59.8

1 0.

49

13.5

5 15

4.7

3.56

8.

23

0.47

Tab

le 3

-7 S

umm

ary

stat

istic

s and

Bro

wn-

Fors

ythe

's F

test

for

eigh

t met

rics

am

ong

the

pand

a pr

esen

ce in

five

m

ount

ain

regi

ons.

Met

rics

Of f

ores

t I

Ed

ge d

ensi

ty

La

rges

t pat

ch in

dex

C

lum

py in

dex

M

ean

prox

imity

inde

x O

f for

est I

I

Mea

n pa

tch

area

Land

scap

e sh

ape

inde

x

Mea

n pr

oxim

ity in

dex

C

lum

py in

dex

* Diff

eren

ce is

sign

ifica

nt a

t the

0.0

5 le

vel

32

Qinling Minshan Qionglai Xiangling Liangshan0

25

50

75

100La

rges

t pat

ch in

dex

of fo

rest

I

Qinling Minshan Qionglai Xiangling Liangshan0

100

200

300

400

Mea

n pa

tch

area

of f

ores

t II

Qinling Minshan Qionglai Xiangling Liangshan0

10

20

30

Edge

den

sity

of f

ores

t I

Qinling Minshan Qionglai Xiangling Liangshan1

2

3

4

5

Land

scap

e sh

ape

inde

x of

fore

st II

Qinling Minshan Qionglai Xiangling Liangshan0

10

20

30

40

Mea

n pr

oxim

ity in

dex

of fo

rest

I

Qinling Minshan Qionglai Xiangling Liangshan0

10

20

30

Mea

n pr

oxim

ity in

dex

of fo

rest

II

Qinling Minshan Qionglai Xiangling Liangshan

0.2

0.4

0.6

0.8

1.0

Clu

mpy

inde

x of

fore

st I

Qinling Minshan Qionglai Xiangling Liangshan

0.2

0.4

0.6

0.8

1.0

Clu

mpy

inde

x of

fore

st II

Figure 3-3 Boxplots of eight metrics among the panda presence areas in five mountain regions. The lines within the box are medians of the data; the upper and lower ends of the box are upper quartile and lower quartile; the upper and low whiskers are highest quartile and lowermost quartile.

33

34

Table 3-8 Mean difference of each of eight metrics from pairwise multiple comparisons of the panda presence in five mountain regions.

Largest patch index of forest I M1 M2 M3 M4 M5M1 M2 4.13 M3 10.2* 6.10* M4 14.0* 9.86* 3.76 M5 16.8* 12.6* 6.53* 2.77

Edge density of forest I M1 M2 M3 M4 M5

M1 M2 -6.52* M3 -6.56* -0.05 M4 -8.87* -2.35* -2.31 M5 -7.90* -1.38 -1.34 0.97

Mean proximity index of forest I M1 M2 M3 M4 M5

M1 M2 -4.15* M3 -2.12 2.03 M4 -4.98 -0.83 -2.86M5 0.18 4.32* 2.29 5.15

Clumpy index of forest I M1 M2 M3 M4 M5

M1 M2 0.13* M3 0.13* 0 M4 0.15* 0.02 0.02 M5 0.15* 0.02 0.02 0

Mean patch area of forest II M1 M2 M3 M4 M5 M1 M2 106* M3 82* -24* M4 114* 7.3 31* M5 67* -39* -15 -46*

Landscape shape index of forest II M1 M2 M3 M4 M5

M1 M2 -0.54* M3 -0.51* 0.03 M4 -0.67* -0.13 -0.16 M5 -0.67* -0.13 -0.16 0

Mean proximity index of forest II M1 M2 M3 M4 M5

M1 M2 3.24* M3 2.19* -1.05 M4 4.51* 1.27 2.32* M5 -0.96 -4.20* -3.16* -5.48*

Clumpy index of forest II M1 M2 M3 M4 M5

M1 M2 0.16* M3 0.12* -0.04* M4 0.14* -0.01 0.02 M5 0.08* -0.07* -0.04 -0.06

*Games-Howell test, the mean difference is significant at the 0.05 level. M1: Qinling; M2: Minshan; M3: Qionglai; M4: Xiangling; M5: Liangshan.

Table 3-9 Statistics for the metrics with significant difference in pairwise mountain regions.

Range Qinling Minshan Qionglai Xiangling Liangshan Qinling - Minshan 7 - Qionglai 7 3 - Xiangling 7 2 2 - Liangshan 6 5 2 2 -

3.4. Logistic regression analysis of panda presence-absence data

Eight representative metrics were used as explanatory variables to determine the best model to explain the presence and absence of the giant panda in a forward stepwise logistic regression procedure. Of the eight representative metrics, four metrics were significant at P < 0.01 (Table 3-10) and the rest were not included into the final model (at P < 0.05). The Nagelkerke R Square was 0.451, indicating that the final model explains around 45% of the overall variance of the response variable in the training dataset. The Hosmer-Lemeshow statistic was 15.864 (df = 8, P = 0.044), indicating that the model might not fit the data well.

As for the metrics included in the final model, three out of four metrics are measuring the configuration of the forest I class, indicating that the configuration of the forest I has more contribution to panda presence than the forest II class. Furthermore, three metrics belong to the group I metrics which measuring patch size/density/edge (see Table 2-3), leading to the conclusion that patch size/density/edge played more significant roles in panda presence-absence than other aspects.

Table 3-10 Parameter estimates of the final logistic regression model. The significance of the coefficients was assessed using the Wald statistic

Parameter Coefficient S.E. P

Edge density of forest I 0.160 0.010 < 0.001

Largest patch index of forest I 0.042 0.003 < 0.001

Clumpy index of forest I -1.475 0.335 < 0.001

Mean patch area of forest II 0.001 0.000 < 0.01

Constant -4.736 0.363 < 0.001

The equation of the final logistic regression model is:

AREA)*0.001CI*475.1LPI*0.042ED*0.16EXP(-4.7361AREA)*0.001CI*475.1LPI*0.042ED*0.16EXP(-4.736

+−++++−++

=pi

where pi is the probability of panda occurrence, ED is Edge density of forest I, LPI is Largest patch index of forest I, CI is Clumpy index of forest I, and AREA is Mean patch area of forest II.

35

36

3.5. Mapping the presence-absence of the giant panda

The probability of the occurrence of the giant panda was predicted by a logistic regression model with knowledge-based control (see Section 2.6.3.3). Compared with the results from the logistic regression model without knowledge-based control, the area with high probability of the panda occurrence shrank, which mainly happened in Qionglai, Xiangling, and Liangshan ranges, as shown in Figure 3-4. This was shown by comparing two presence-absence maps for giant pandas (Figure 3-5), where we can see the obvious reduction in panda habitat for these three mountain regions.

Table 3-11 shows the confusion matrices for the panda presence-absence mapping with and without knowledge-based control. The overall accuracy, sensitivity and specificity, kappa and kappa variance of the mapping are summarized in Table 3-12. By applying the knowledge-based control to the mapping, the overall accuracy increased by 5% and the specificity increased from 58.5% to 71.6%; however, the sensitivity declined slightly from 79.5% to 77.6%. This was expected because the role of knowledge-based control is to decrease the risk of overprediction of presence of the regression model. As shown in Table 3-11, by applying the knowledge-based control, the number of correctly predicted absence samples increased notably, while there was a slight decrease in the number of correctly predicted presence samples. The Z-test for kappa coefficients shows that the accuracy of the mapping was significantly improved by the knowledge-based control (P < 0.05).

Table 3-11 Confusion matrices of the logistic regression modelling. Bold type indicates the numbers of correctly predicted testing samples. (a) Without knowledge-based control

Reference Predicted

Presence Absence Total

Presence 174 76 234 Absence 45 107 168 Total 219 183 402

(b) With knowledge-based control Reference

Predicted Presence Absence

Total

Presence 170 52 222 Absence 49 131 180 Total 219 183 402

Table 3-12 Overall accuracy, sensitivity and specificity, kappa and kappa variance of the logistic regression modelling with and without knowledge-based control.

Modelling Overall accuracy Sensitivity Specificity Kappa Kappa

variance Z

Without knowledge-based control 69.9% 79.5% 58.5% 0.389 0.00048 3.42*

With knowledge-based control 74.9% 77.6% 71.6% 0.493 0.00044 * Difference is significant at the 0.05 level.

Figure 3-4 Suitability for giant pandas predicted by logistic regression model using landscape metrics statistic; (a) mapping without knowledge-based control; (b) mapping with knowledge-based control for elevation and slope.

37

Figure 3-5 Presence-absence of the giant panda predicted by logistic regression model (threshold = 0.5); (a) mapping without knowledge-based control; (b) mapping with knowledge-based control for elevation and slope.

38

The results of the sensitivity analysis are shown in Figure 3-6. As we can see, the overall accuracy at a threshold of 0.5 is the highest. With the increase of the threshold from 0.3 to 0.7, the specificity increases steadily, while the sensitivity keeps decreasing, and this decline becomes sharply when the threshold is larger than 0.5. The sensitivity-specificity difference (Liu et al., 2005) is minimum (zero) at a threshold around 0.52. Thus, it appears that selecting a threshold of 0.5 is appropriate for transforming the results of logistic regression modelling from probabilities of panda occurrence to presences/absences.

20.0%

30.0%

40.0%

50.0%

60.0%

70.0%

80.0%

90.0%

100.0%

0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70

Threshold

Perc

enta

ge o

f acc

urac

y

Overall acc.SensitivitySpecificity

Figure 3-6 Changes of overall mapping accuracy, sensitivity and specificity with different thresholds.

39

4. Discussion

4.1. Land cover classification from MODIS 250 m EVI time series

When choosing a data source for land cover classification based on change in vegetation phenology, data availability and spatial resolution are two important issues. High spatial resolution sensors have a relatively narrow swath width and long revisit times, their acquisition and interpretation are expensive and time-consuming. It is also difficult to acquire sufficient images for a large area. In contrast, lower spatial resolution sensors such as MODIS provide wide swath width and daily coverage of the earth, and so weather events are much less of an obstacle. In this study, it would require 30 scenes of ETM+ images to cover all panda distribution areas, while only 4 tiles of MODIS data is needed. MODIS products provide significant refinements in spectral, radiometric, and geometric properties compared to previously available data sets with similar spatial resolution (Justice et al., 2002; Zhang et al., 2003). Of the two MODIS VI products, EVI is less sensitive to soil and atmospheric effects than NDVI and simultaneously remain sensitive to increases in canopy density beyond where NDVI becomes saturated because it incorporates blue spectral wavelengths (Huete et al., 2002). These advantages let multi-temporal MODIS 250 m EVI data stand out as the optimal data source for both phenology-based land cover classification and routine vegetation monitoring of large area. In this study, time series of the 16-day composite MODIS 250 m EVI data was proven to have sufficient spectral, temporal, and radiometric resolutions to discriminate the region's major land cover types.

Several aspects need to be considered when mapping land cover form MODIS EVI time series. Firstly, the reconstruction of cleaned EVI time series is necessary because there is a level of noise persisting in the EVI data, mainly caused by remnants of clouds and its shadows. A number of methods for reducing noise and constructing high-quality NDVI time series have been developed in the last two decades, e.g. Fourier-based fitting methods (Cihlar, 1996), asymmetric Gaussian function fitting approach (Jönsson and Eklundh, 2002) and adaptive Savitzky–Golay smoothing approach (Jönsson and Eklundh, 2004). These methods are thought to be applicable to EVI data because of its similar nature to NDVI. In this study a simple but robust method based on the Savitzky–Golay filter was employed to smooth out noise in EVI time series, and the result indicates that this method is effective in obtaining smoothed EVI time series.

40

Secondly, regional intra-class variations caused by climate and elevation gradient in the MODIS EVI time series have to be considered during the selection of training samples prior to classification to ensure that both the climate and elevation-related sub-classes within each land cover class are represented in the training set. In addition to integrating the bio-climatic and topographical information into the classification, the delineation of relatively homogeneous mapping units and the use of non-parametric Neural Net classifier are illustrated to be effective approaches handling intra-class variability. However, due to the 250 m spatial resolution of MODIS data, the ‘mixed’ pixels of multiple land covers may be an issue for some locations or classes, especially the areas with intensive human modification.

4.2. Quantification of the spatial configuration of forests

Due to the strong influence of scale on the behavior of landscape metrics, the choice of appropriate scales is fundamental in landscape analysis (Gustafson, 1998; Turner et al., 2001; Wu and Hobbs, 2002). In general, three domains of scale are discriminated in animal ecological studies: home range, local, and regional scales. Single individuals operate at the level of a home range, whereas dispersal and population processes occur at the local scale, and patterns within a species geographic range are distinguished at regional scales of observation (Kotliar and Wiens, 1990). A species’ presence-absence data characterizes essentially the qualitative occupancy status of the habitat patch, and so is the response of the individuals to habitat characteristics within the extent of a particular scale, to many mammals, if not most, the home range. The giant panda has its exclusive home range and a daily movement distance of less than 1 km during migration periods (Hu, 2001; Liu, 2001). Thus, the size of giant panda’s home range can be regarded as an appropriate scale to adequately quantify the spatial configuration of forests related to the presence-absence of the giant panda.

A large number of landscape structure metrics are currently available and many of them are highly correlated (McGarigal and Marks, 1995). It is necessary to use several representative metrics to characterize a particular landscape because the use of many highly correlated indices does not yield new information and leads to problems in the interpretation of the results (Jones, 2002; Li and Wu, 2004), such as multicollinearity and autocorrelation. The selected metrics should be relatively independent of one another, providing a unique and ecologically meaningful contribution to the understanding of landscape structure (Hargis et al., 1998). By using a factor analysis approach, eight metrics of two focal classes were selected as representative (from 52 metrics) to characterize the spatial configuration of forests at a given scale. Of these metrics, four measure the patch area/density/edge of the focal

41

classes; one metric measures patch proximity/connectivity, and two metrics measure patch contagion/interspersion of the focal classes. It’s notable that the metrics of two focal classes with high loading are coincident; only differ in the sequence of the metrics formed by the component number.

From the literature review, these eight selected metrics were also commonly used as the measures of habitat fragmentation (Hargis et al., 1998; Young and Jarvis, 2001; McAlpine and Eyre, 2002). Edge density is an effective tool for evaluating the effects of patch shape and area on the abundance of habitat edge, and is not sensitive to spatial pattern (Hargis et al., 1998). Largest patch index is a simple measure of dominance of the focal class by quantifying the percentage of total landscape area comprised by the largest patch at the class level (McGarigal and Marks, 1995). The proximity index is an effective measure of patch connectivity, but sensitive to the distance between patches rather than patch size (Hargis et al., 1998). It should be noticed that these representative metrics were selected at a scale of the home range of giant pandas, and may not applicable at other scales (e.g. regional scale landscape analysis). Behaviours of these metrics at multiple scales need to be tested.

Landscape metrics may be sensitive to the level of detail in categorical map data that is determined by the schemes used for map classification (Turner et al., 2001). Thus, a species-centred categorical map is essential for analyzing the relationship between the species and landscape pattern. In this study, a five-class classification scheme was adopted based on the biology of the giant panda, and the forests were classified into two categories by the canopy cover of 30%. Though the criterion is to some extent arbitrary, the division is necessary and ecologically meaningful, because giant pandas have been proven that have a strong preference for forest patches with a high canopy cover rather than its composition (Hu, 2001). The test for the sensitivity of landscape metrics towards different division of forests may help the understanding of the relationship between forest spatial patterns and the response of the giant panda; but this is beyond the objective of this study.

4.3. Spatial configuration of the forests occupied by giant pandas

From an ecological point of view, every species has its particular position in the ecosystem, which is termed ‘niche’, the range of environmental conditions, both biological and physical, and from geographical to microenvironmental (Elton, 2001). The giant panda is a forest inhabitant with exclusive home range; hence the spatial configuration of the forest plays an important role in the distribution of the giant panda. In this analysis, eight metrics that best captured the variation in forest configuration among landscapes were selected as representative metrics for landscape pattern analysis. Both the Brown-Forsythe's F test and nonparametric

42

Mann-Whitney U test showed that eight representative metrics were significantly different between the presence and absence of the giant panda, and the high F-ratio of the metrics was evidence the between-groups variances were larger than the within-groups variances (Table 3-6), indicating the heterogeneity of forest spatial configurations had a significant contribution to the presence-absence of giant pandas. At the panda’s home-range scale, the largest patch size and mean proximity of the forest I in the panda presence areas were higher than the absence areas, indicating that patches of the forest I occupied by giant pandas tended to be larger, closer and more contiguous (or less fragmented) in distribution. However, the difference in metric Clumpy index between panda presence and absence indicated that giant pandas prefer less aggregated forest.

Previous studies documented that Xianglin and Liangshan were two regions in which panda habitats were heavily fragmented (Hu, 2001). This is supported by the findings of this study. The results from the comparisons of the spatial configuration of the forests in five mountain regions (Table 3-7) illustrate that the forests in Qinling were less fragmented than the other four regions. On the contrary, the forests in Xiangling and Liangshan were the most fragmented among the five regions. The pairwise multiple comparisons of the panda presence in five regions (Table 3-8) showed the delicate variations of the forests spatial configuration crossing the pandas’ distribution areas.

Using stepwise logistic regression analysis, four landscape metrics were observed that had significant contributions to the presence-absence of giant pandas. Among these four metrics, three metrics measure the patch area/edge, pointing out that the giant panda appear sensitive to area and isolation effects associated with forest fragmentation. Additionally, three out of the four metrics measure spatial configuration of the forest I, indicating that the forest I plays an important role in distribution of giant pandas. This is easy to understand because of the giant panda’s preference for forests with high canopy cover, which has been reported in previous research (Hu, 2001).

Note that the spatial configuration of the forests involved in this study was only on a single temporal dimension. However, it’s difficult to interpret underlying causal relationship between the spatial configuration of forests and presence-absence of giant pandas just by a single investigation of the landscape. A multi-temporal spatial pattern analysis is necessary to detect structural changes in forested landscapes and predict how this dynamic influence the distribution of giant pandas. Nevertheless, this remote sensing based approach was demonstrated to be an effective method for routine habitat monitoring and habitat evaluation for the giant panda. It can also

43

facilitate reasonable landscape conservation planning and design of corridor for the giant panda.

4.4. Can the forest configuration predict the panda’s presence-absence?

The logistic regression model with knowledge-based control is capable of predicting the spatial distribution of the giant panda using spatial configuration of the forest at the home-range scale (Table 3-11). The outputs (Figure 3-4(b) and Figure 3-5(b)) match the giant panda’s distribution maps generated by the 3rd National Panda Survey (see Appendix 5). The accuracy of the mapping was significantly improved by applying a knowledge-based control to the logistic regression model, and the map outputs are apparently different, the area of the predicted panda presence shrank, particularly in Qionglai, Xiangling, and Liangshan regions.

As shown in Table 3-12, the knowledge-based control had little effect on the sensitivity of the modelling, but increased the specificity. This is because the role of knowledge-based control is to decrease the risk of overprediction of presence of the logistic regression model by integrating the topographical constraints into the modelling. Logistic regression itself is a transformed linear regression which merely depends on explanatory variables included in the model. As mentioned before, the distribution of the giant panda also limited by the range of environmental conditions. In the absence of consideration of environmental constraints, the model could not exclude the areas which beyond the environmental thresholds of the giant panda, leading to the overprediction of the presence of giant pandas. The utilization of knowledge-based control can diminish this problem. However, to design an ecologically meaningful control needs adequate relevant knowledge and well-understanding of the relationship between the species and the environmental factors.

The low R-square value of the model indicates that landscape metrics alone could not fully explain the presence-absence of giant pandas. Indeed, there are many factors influencing the occurrence of the giant panda, such as food, shelter, water, and predator; and the heterogeneity of the forests occupied by the giant panda in different regions may also decrease the power of the model. A fundamental assumption of this study is that bamboo is sufficient and thus not a constraint of panda distribution. However, bamboo resources are unevenly distributed across five mountain regions (State Forestry Administration of China, 2006). To accurately map the distribution of giant pandas or design corridors, the spatial pattern and quality of bamboo forest are required. In addition to the knowledge of bamboo, other environmental explanatory factors are also needed, such as forest compositions and road network.

44

5. Conclusion

This study demonstrated a successful approach for characterizing the spatial distribution of giant pandas using MODIS 250 m EVI time series and landscape metrics. A five-class land cover map of the study area was produced from cleaned and smoothed time series of the MODIS EVI 250m in 2001 using the combination of the ISODATA and Neural Net classifier (overall accuracy was 83.9%, and the kappa was 0.79). 52 class-level landscape metrics were then calculated using a 3 km×3 km moving window in FRAGSTAT program. By applying a landscape metrics reduction procedure, eight metrics that best captured the variation in forest configuration were selected as representative metrics at the panda’s home range scale. These eight metrics measure three aspects of the forests: patch area/density/edge, patch proximity/connectivity, and patch contagion/interspersion. All representative metrics were significantly different (P < 0.05) between the presence and absence of the giant panda. At a given landscape extent of 36 km2 (as defined by the moving window size), the largest patch size in panda presence areas was on average 19.2 ± 7.8 km2, in contrast to the panda absence areas in which the largest patch size was 12.1 ± 9.2 km2. The spatial configurations of the forests among the five mountain regions were heterogeneous. The forest I occupied by the giant panda in Qinling were less fragmented than in the other four regions, and the forest in Xiangling and Liangshan were more fragmented. Of the eight class-level landscape metrics, only four metrics contributed significantly to the final logistic regression model, the largest patch index, edge density, clumpy index of the forest I, and the mean patch area of the forest II. From the regression analysis, the spatial configuration of the forest I contributed high to panda spatial distribution. The size/density/edge of forest patches played significant roles in panda presence-absence than other metrics. These findings have implications for the design of conservation plans for wild panda population limited by fragmentation.

The presence-absence of the giant panda can be predicted with an acceptable accuracy, the final logistic regression model yielded an overall prediction accuracy of 69.9%, and this performance was improved 5% by applying the knowledge-based control to the modelling. The Z-test for kappa coefficients shows that the mapping accuracy was significantly improved by the knowledge-based control (P < 0.05). The sensitive analysis shows that the threshold of 0.5 is appropriate for dividing the presence-absence of the giant panda in the regression modelling. However, the low

45

R-square value of the model indicates that the landscape metrics alone can not fully explain the distribution of the giant panda.

The remote sensing based approach proposed in this study can be applied to routine habitat monitoring and habitat evaluation for the giant panda at national level. It is also an effective method for corridor design and landscape conservation planning for the giant panda. Further work can address the underlying relationship between the structural changes in forested landscapes and the dynamic of the giant panda by multi-temporal spatial pattern analysis.

46

References

Atkinson, P.M. and Tatnall, A.R.L., 1997. Introduction Neural networks in remote sensing. International Journal of Remote Sensing, 18(4): 699 - 709.

Bagan, H., Wang, Q., Watanabe, M., Yang, Y. and Ma, J., 2005. Land cover classification from MODIS EVI times-series data using SOM neural network. International Journal of Remote Sensing, 26(22): 4999 - 5012.

Baiulescu, G.E., Dumitrescu, P. and Zugravescu, P.G., 1991. Sampling. (Ellis Horwood series in analytical chemistry). Ellis Horwood, New York etc., 184 pp.

Barbosa, P.M., Gregoire, J.-M. and Pereira, J.M.C., 1999. An Algorithm for Extracting Burned Areas from Time Series of AVHRR GAC Data Applied at a Continental Scale. Remote Sensing of Environment, 69(3): 253-263.

Bauer, M.E. et al., 1994. Satellite inventory of Minnesota forest resources. Photogrammetric Engineering and Remote Sensing, 60: 287-298.

Bennett, A., 1998. Linkages in the landscape – the role of the connectivity in wildlife conservation. IUCN, , Gland, Switzerland and Cambridge.

Boles, S.H. et al., 2004. Land cover characterization of Temperate East Asia using multi-temporal VEGETATION sensor data. Remote Sensing of Environment, 90(4): 477-489.

Bulmer, M.G., 1967. Principles of statistics. Dover, New York, 252 pp. Byrne, G.F., Crapper, P.F. and Mayo, K.K., 1980. Monitoring land-cover change by

principal component analysis of multitemporal Landsat data. Remote Sensing of Environment, 10: 175-184.

Cain, D.H., Riitters, K. and Orvis, K., 1997. A multi-scale analysis of landscape statistics. Landscape Ecology, 12(4): 199-212.

Chen, J. et al., 2004. A simple method for reconstructing a high-quality NDVI time-series data set based on the Savitzky-Golay filter. Remote Sensing of Environment, 91(3-4): 332-344.

China Vegetation Compiling Committee, C., 1980. China vegetation. Science Press, Beijing.

Cihlar, J., 1996. Identification of contaminated pixels in AVHRR composite images for studies of land biosphere. Remote Sensing of Environment, 56(3): 149-163.

Cihlar, J., 2000. Land cover mapping of large areas from satellites: status and research priorities. International Journal of Remote Sensing, 21(6): 1093 - 1114.

Cohen, J., 1960. A coefficient of agreement for nominal scales. Edu. Psychol. Measure, 20: 37-46.

Congalton, R. and Green, K., 1999. Assessing the Accuracy of Remotely Sensed Data: Principles and Practices. CRC/Lewis Press, Boca Raton, 137 pp.

Congalton, R.G., 1991. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sensing of Environment, 37(1): 35-46.

47

Cormack, R.M.e., Patil, G.P.e. and Robson, D.S.e., 1979. Sampling biological populations. Statistical ecology series;5. International Co-operative Publishing House, Fairland, 392 pp.

Corsi, F., de Leeuw, J. and Skidmore, A.K., 2000. Modeling species distribution with GIS. In: Research techniques in animal ecology : controversies and consequences / Boitani, L. and Fuller T.K. ed. - 2000. pp. 389-434.

Cowley, M.J.R. et al., 2000. Habitat-based statistical models for predicting the spatial distribution of butterflies and the day-flying moths in a fragmented landscape. Journal of Applied Ecology, 37(Suppl. 1): 60-72.

Davis, J.C., 2002. Statistics and data analysis in geology. Wiley & Sons, New York etc., 638 pp.

DeFries, R.S. and Chan, J.C.-W., 2000. Multiple Criteria for Evaluating Machine Learning Algorithms for Land Cover Classification from Satellite Data. Remote Sensing of Environment, 74(3): 503-515.

Deng, Y., Chen, X., Chuvieco, E., Warner, T. and Wilson, J.P., 2007. Multi-scale linkages between topographic attributes and vegetation indices in a mountainous landscape. Remote Sensing of Environment, 111(1): 122-134.

Didan, K. and Huete, A., 2006. MODIS Vegetation Index Product Series Collection 5 Change Summary.

Dunning, J.B., Danielson, B.J. and Pulliam, H.R., 1992. Ecological Processes That Affect Populations in Complex Landscapes. Oikos, 65(1): 169-175.

Elton, C.S., 2001. Animal ecology. University of Chicago Press, Chicago. Fielding, A.H. and Bell, J.F., 1997. A review of methods for the assessment of

prediction errors in conservation presence/absence models. Environmental Conservation, 24(1): 38-49.

Forman, R.T.T. and Godron, M., 1986. Landscape Ecology. John Wiley and Sons, Inc, New York.

Fortin, M.J., Boots, B., Csillag, F. and Remmel, T.K., 2003. On the role of spatial stochastic models in understanding landscape indices in ecology. Oikos, 102(1): 203-212.

Franklin, J., 1995. Predictive vegetation mapping: Geographic modeling of biospatial patterns in relation to environmental gradients. Progress in Physical Geography, 19: 474-499.

Friedl, M.A. et al., 2002. Global land cover mapping from MODIS: algorithms and early results. Remote Sensing of Environment, 83(1-2): 287-302.

Gao, X., Huete, A.R., Ni, W.G. and Miura, T., 2000. Optical-biophysical relationships of vegetation spectra without background contamination. Remote Sensing of Environment, 74(3): 609-620.

Giannetti, F., Montanarella, L. and Salandin, R., 2001. Integrated use of satellite images, DEMs, soil and substrate data in studying mountainous lands. International Journal of Applied Earth Observation and Geoinformation, 3(1): 25-29.

Griffith, J.A., Martinko, E.A. and Price, K.P., 2000. Landscape structure analysis of Kansas at three scales. Landscape and Urban Planning, 52(1): 45-61.

48

Griffiths, G.H. and Mather, P.M., 2000. Remote sensing and landscape ecology: landscape patterns and landscape change. International Journal of Remote Sensing, 21(13-14): 2537-2539.

Gustafson, E.J., 1998. Quantifying landscape spatial pattern: What is the state of the art? Ecosystems, 1(2): 143-156.

Hargis, C.D., Bissonette, J.A. and David, J.L., 1998. The behavior of landscape metrics commonly used in the study of habitat fragmentation. Landscape Ecology, 13(3): 167-186.

Harris, P.M. and J., V.S., 1995. The Integration of Geographic Data with Remotely Sensed Imagery to Improve Classification in an Urban Area. Photogrammetric Engineering and Remote Sensing 61(8): 993-998.

Hosmer, D.W. and Lemeshow, S., 2000. Applied logistic regression, 2nd Edition. Wiley, New York.

Houborg, R., Soegaard, H. and Boegh, E., 2007. Combining vegetation index and model inversion methods for the extraction of key vegetation biophysical parameters using Terra and Aqua MODIS reflectance data. Remote Sensing of Environment, 106(1): 39-58.

Hu, J., 2001. Research on the giant panda. Shanghai Scientific & Technological Education Publishers, Shanghai, 402 pp.

Hu, J., Schaller, G.B., Pan, W. and Zhu, J., 1985. The giant panda in Wolong. Sichuan Science and Technology Press, Chengdu, 225 pp.

Huete, A. et al., 2002. Overview of the radiometric and biophysical performance of the MODIS vegetation indices. Remote Sensing of Environment, 83(1-2): 195-213.

Huete, A.R., Liu, H.Q., Batchily, K. and vanLeeuwen, W., 1997. A comparison of vegetation indices global set of TM images for EOS-MODIS. Remote Sensing of Environment, 59(3): 440-451.

Janssen, L.L.F., Jaarsma, M.N. and Linden, E.T.M., 1990. Integrating Topographic Data with Remote Sensing for Land Cover Classification. Photogrammetric Engineering and Remote Sensing 56(11): 1503-1506.

Jelinski, D. and Wu, J., 1996. The modifiable areal unit problem and implications for landscape ecology. Landscape Ecology, 11(3): 129-140.

Jiang, H., Strittholt, J.R., Frost, P.A. and Slosser, N.C., 2004. The classification of late seral forests in the Pacific Northwest, USA using Landsat ETM+ imagery. Remote Sensing of Environment, 91(3-4): 320-331.

Jones, R., 2002. Landscape ecology: In theory and practice, pattern and process. Geography, 87: 369-370.

Jönsson, P. and Eklundh, L., 2002. Seasonality extraction by function fitting to time-series of satellite sensor data. IEEE Transactions on Geoscience and Remote Sensing, 40(8): 1824– 1832.

Jönsson, P. and Eklundh, L., 2004. TIMESAT - a program for analyzing time-series of satellite sensor dat. Computers and Geosciences, 30: 833–845.

Jordán, F., Báldi, A., Orci, K.M., Rácz, I. and Varga, Z., 2003. Characterizing the importance of habitat patches and corridors in maintaining the landscape connectivity of a Pholidoptera transsylvanica (Orthoptera) metapopulation. Landscape Ecology, 18(1): 83-92.

49

Justice, C.O. et al., 2002. An overview of MODIS Land data processing and product status. Remote Sensing of Environment, 83(1-2): 3-15.

Knight, J.F., Lunetta, R.S., Ediriwickrema, J. and Khorram, S., 2006. Regional Scale Land Cover Characterization Using MODIS-NDVI 250 m Multi-Temporal Imagery: A Phenology-Based Approach. GIScience and Remote Sensing, 43: 1-23.

Koper, N., Schmiegelow, F. and Merrill, E., 2007. Residuals cannot distinguish between ecological effects of habitat amount and fragmentation: implications for the debate. Landscape Ecology, 22(6): 811-820.

Kotliar, N. and Wiens, J., 1990. Multiple scales of patchiness and patch structure: a hierarchical framework for the study of heterogeneity. Oikos, 59: 250-260.

Kupfer, A. and Kneitz, S., 2000. Population ecology of the great crested newt (Triturus cristatus) in an agricultural landscape: Dynamics, pond fidelity and dispersal. Herpetological Journal, 10(4): 165-171.

Li, H. and Wu, J., 2004. Use and misuse of landscape indices. Landscape Ecology, 19(4): 389-399.

Li, X. et al., 2005. The adequacy of different landscape metrics for various landscape patterns. Pattern Recognition, 38(12): 2626-2638.

Lindburg, D.e. and Baragona, K.e., 2004. Giant panda's : biology and conservation. University of California Press, Berkeley etc., 308 pp.

Liu, C., Berry, P.M., Dawson, T.P. and Pearson, R.G., 2005. Selecting thresholds of occurrence in the prediction of species distributions. Ecography, 28: 385-393.

Liu, J. et al., 2002. The Land-use and land-cover change database and its relative studies in China. Journal of Geographical Sciences 12(3): 275-282.

Liu, J.G. et al., 1999. A framework for evaluating the effects of human factors on wildlife habitat: the case of giant pandas. Conservation Biology, 13(6): 1360-1370.

Liu, J.Y., Zhuang, D.F., Luo, D. and Xiao, X., 2003. Land-cover classification of China: integrated analysis of AVHRR imagery and geophysical data. International Journal of Remote Sensing, 24(12): 2485 - 2500.

Liu, X., 2001. Mapping and modelling the habitat of giant panda's in Foping nature reserve, China, Wageningen University, Wageningen, 155 pp.

Liu, X. et al., 2000. Mapping habitat suitability for giant pandas on the Foping nature reserve, China. In: Panda 2000 : conservation priorities for the new millennium : proceedings of an international conference, San Diego, Calilfornia, USA, October 15-18, 2000 / ed. by. D. Lindberg and K. Baragona. San Diego: The Shape of Enrichment, 2000. pp. 87-94.

Liu, X. and Kafatos, M., 2005. Land-cover mixing and spectral vegetation indices. International Journal of Remote Sensing, 26(15): 3321-3327.

Lo, C.P. and Choi, J., 2004. A hybrid approach to urban land usecover mapping using Landsat 7 Enhanced Thematic Mapper Plus ETM+ images. International Journal of Remote Sensing, 25: 2687-2700.

Loucks, C.J. et al., 2003. The giant pandas of the Qinling Mountains, China: A case study in designing conservation landscapes for elevational migrants. Conservation Biology, 17(2): 558-565.

50

Lu, Y.H., Chen, L.D., Fu, B.J. and Liu, S.L., 2003. A framework for evaluating the effectiveness of protected areas: the case of Wolong Biosphere Reserve. Landscape and Urban Planning, 63(4): 213-223.

Manel, S., Dias, J.-M. and Ormerod, S.J., 1999. Comparing discriminant analysis, neural networks and logistic regression for predicting species distributions: a case study with a Himalayan river bird. Ecological Modelling, 120(2/3): 337-347.

McAlpine, C.A. and Eyre, T.J., 2002. Testing landscape metrics as indicators of habitat loss and fragmentation in continuous eucalypt forests (Queensland, Australia). Landscape Ecology, 17(8): 711-728.

McGarigal, K. and Marks, B.J., 1995. Fragstats: spatial pattern analysis program for quantifying landscape structure.U. S. Forest Service General Technical Report, Portland, OR, USA.

McGarigal, K. and McComb, W.C., 1995. Relationships between landscape structure and breeding birds in the oregon coast range. Ecol Mono, 65: 235–260.

Menard, S., 2002. Applied Logistic Regression Analysis. CA: Sage Publications, Thousand Oaks.

Merriam, G., 1984. Connectivity: a fundamental ecological characteristic of landscape pattern. In: J. Brandt and P. Agger (Editors), Methodology in landscape ecological research and planning. Roskilde Universitetsforlag GeuRuc, Roskilde, pp. 5-15.

Merriam, G., 1995. Movement in spatially divided populations: responses to landscape structure. In: W.J. Lidicker (Editor), Landscape approaches in mammalian ecology and conservation. University of Minnesota Press, Minneapolis, pp. p 64-77.

O'Connel, A.A., 2006. Logistic regression models for ordinal response variables. Sage, Thousand Oaks.

Pan, W., 2001. The opportunity of survival. Beijing University Press, Beijing, 522 pp.

Paola, J.D. and Schowengerdt, R.A., 1995. A review and analysis of backpropagation neural networks for classification of remotelysensed multispectral imagery. International Journal of Remote Sensing 16(16): 3033-3058.

Price, J.C., 2003. Comparing MODIS and ETM+ data for regional and global land classification. Remote Sensing of Environment, 86(4): 491-499.

Qi, Y. and Wu, J., 1996. Effects of changing spatial resolution on the results of landscape pattern analysis using spatial autocorrelation indices. Landscape Ecology, 11(1): 39-49.

Ren, Y., 1998. Vegetation within the Giant Panda's habitat in Qinling Mountains. Shaanxi Science and Technology Press, Xi'an, 488 pp.

Reybenayas, J.M. and Pope, K.O., 1995. Landscape Ecology and Diversity Patterns in the Seasonal Tropics from Landsat Tm Imagery. Ecological Applications, 5(2): 386-394.

Richards, J.A., 1984. Thematic mapping from multitemporal image data using

51

the principal components transformation. Remote Sensing of Environment, 16: 25-46.

Richards, J.A., 1999. Remote Sensing Digital Image Analysis. Springer-Verlag, Berlin, 240 pp.

Riitters, K.H. et al., 1995. A factor analysis of landscape pattern and structure metrics. Landscape Ecology, 10(1): 23-39.

Rutherford, A., 2001. Introducing ANOVA and ANCOVA: A GLM approach. CA: Sage Publications, Thousand Oaks.

Saura, S., 2004. Effects of remote sensor spatial resolution and data aggregation on selected fragmentation indices. Landscape Ecology, 19(2): 197-209.

Schaller, G.B., 1994. The last panda. University of Chicago Press, Chicago ; London, xx, 299 [16] of plates pp.

Sedano, F., Gong, P. and Ferrao, M., 2005. Land cover assessment with MODIS imagery in southern African Miombo ecosystems. Remote Sensing of Environment, 98(4): 429-441.

Shaw, G. and Wheeler, D., 1994. Statistical techniques in geographical analysis. David Fulton, London, 359 pp.

Skidmore, A.K., Turner, B.J., Brinkhof, W. and Knowles, E., 1997. Performance of a neural network: mapping forests using GIS and remotely sensed data. Photogrammetric Engineering and Remote Sensing 63(5): 501– 514.

Skidmore, A.K., Watford, F., Luckananurug, P. and Ryan, P.J., 1996. An operational GIS expert system for mapping forest soils. . Photogrammetric Engineering and Remote Sensing, 62(5): 501– 511.

Sokal, R.R. and Rohlf, F.J., 1994. Biometry : the principles and practice of statistics in biological research. W.H. Freeman, New York, 887 pp.

State Forestry Administration of China, S., 2006. Report of the third giant panda census in China. Science Press, Beijing, 280 pp.

Stehman, S.V., Sohl, T.L. and Loveland, T.R., 2003. Statistical sampling to characterize recent United States land-cover change. Remote Sensing of Environment, 86(4): 517-529.

Story, M. and Congalton, R.G., 1986. Accuracy assessment: A user's perspective. Photogrammetric Engineering and Remote Sensing, 52: 397-399.

Tarnavsky, E., Garrigues, S. and Brown, M.E., Multiscale geostatistical analysis of AVHRR, SPOT-VGT, and MODIS global NDVI products. Remote Sensing of Environment, In Press, Corrected Proof.

Taylor, P.D., Fahrig, L., Henein, K. and Merriam, G., 1993. Connectivity Is a Vital Element of Landscape Structure. Oikos, 68(3): 571-573.

Toothacker, L.E., 1993. Multiple comparisons procedures. . CA: Sage Publications. , Thousand Oaks.

Tottrup, C., Rasmussen, M.S., Eklundh, L., ouml and nsson, P., 2007. Mapping fractional forest cover across the highlands of mainland Southeast Asia using MODIS data and regression tree modelling. International Journal of Remote Sensing, 28(1): 23 - 46.

Turner, M.G., 1989a. Landscape Ecology - the Effect of Pattern on Process. Annual Review of Ecology and Systematics, 20: 171-197.

52

Turner, M.G., 1989b. Landscape ecology: the effect of pattern on process. Annual Review of Ecology and Systematics(20): 171-197.

Turner, M.G., Gardner, R.H. and O'Neill, R.V., 2001. Landscape ecology in theory and practice : patterns and process. Springer, Berlin etc., 401 pp.

Vermote, E.F., El Saleous, N.Z. and Justice, C.O., 2002. Atmospheric correction of MODIS data in the visible to middle infrared: first results. Remote Sensing of Environment, 83(1-2): 97-111.

Wang, T., 2003. Habitat analysis for giant panda in Laoxiancheng nature reserve in the Qinling mountains China, ITC, Enschede, 70 pp.

Wardlow, B.D., Egbert, S.L. and Kastens, J.H., 2007. Analysis of time-series MODIS 250 m vegetation index data for crop classification in the U.S. Central Great Plains. Remote Sensing of Environment, 108(3): 290-310.

Wessels, K.J. et al., 2004. Mapping regional land cover with MODIS data for biological conservation: Examples from the Greater Yellowstone Ecosystem, USA and Para State, Brazil. Remote Sensing of Environment, 92(1): 67-83.

Wilson, E.H. and Sader, S.A., 2002. Detection of forest harvest type using multiple dates of Landsat TM imagery. Remote Sensing of Environment, 80(3): 385-396.

With, K.A., 2002. The landscape ecology of invasive spread. Conservation Biology, 16(5): 1192-1203.

With, K.A. and King, A.W., 1997. The use and misuse of neutral landscape models in ecology. Oikos, 79(2): 219-229.

Wolfe, R.E. et al., 2002. Achieving sub-pixel geolocation accuracy in support of MODIS land science. Remote Sensing of Environment, 83(1-2): 31-49.

Wu, J.G. and Hobbs, R., 2002. Key issues and research priorities in landscape ecology: An idiosyncratic synthesis. Landscape Ecology, 17(4): 355-365.

Xavier, A.C. et al., 2006. Multi-temporal analysis of MODIS data to classify sugarcane crop. International Journal of Remote Sensing, 27(4): 755 - 768.

Xu, W. et al., 2006. Designing a conservation plan for protecting the habitat for giant pandas in the Qionglai mountain range, China. Diversity and Distributions, 12(5): 610-619.

Yang, X. and Lo, C.P., 2002. Using a time series of satellite imagery to detect land use and land cover changes in the Atlanta, Georgia metropolitan area. International Journal of Remote Sensing, 23(9): 1775 - 1798.

Young, C.H. and Jarvis, P.J., 2001. Measuring urban habitat fragmentation: an example from the Black Country, UK. Landscape Ecology, 16(7): 643-658.

Zhang, X.Y. et al., 2003. Monitoring vegetation phenology using MODIS. Remote Sensing of Environment, 84(3): 471-475.

53

Appendices

Appendix 1

The NLCD-2000 classification system of and its definition

Level I Level II Description Paddy Farmlands with water resource guarantee and irrigating

facilities using for rice growing Farmland

Dry farming Lands for cultivating without irrigating facilities; dry croplands and lands growing vegetables

Forest Natural or man-made forest with canopy cover greater than 30%

Shrub Lands covered by trees less than 2 meters high, the canopy cover >40%

Woods Lands covered by trees with canopy cover between 10-30%

Forest

Others Lands such as tea-garden, orchid and non-grownup forest Dense grass Grassland with canopy cover greater than 50% Moderate Grass Grassland with canopy cover between 20-50%

Grassland

Sparse grass Grassland with canopy cover between 5-20% Rivers Lands covered by rivers including canals Lakes Lands covered by lakes Reservoir and ponds

Man-made facilities for water reservation

Permanent Ice and Snow

Lands covered by ice and snow all the year

Beach and Shore Lands between high tide level and low tide level

Water body

Bottomland Lands between normal water level and flood level City built-up Lands used for cities and counties. Rural settlements Lands used for settlements in villages

Man-made Built-up

others Lands used for factories, quarries, mining, oil-field slattern outside cities and lands for special uses such as transportation and airport.

Sand Lands expose sand and never have more than 5% vegetated cover.

Gobi Lands expose gravel and never have more than 5% vegetated cover.

Salina Lands with salina accumulation and sparse vegetation Wetland Lands with a permanent mixture of water and herbaceous

or woody vegetation that cover extensive areas Bare Soil Lands expose soil and never have more than 5% vegetated

cover. Bare Rock Lands expose rocks and never have more than 5%

vegetated cover.

Unused

Others Other lands such as alpine desert and tundra

54

Appendix 2

Summary statistics (n = 3735) for 34 landscape metrics remaining from the partial correlation analysis. Skew within the range +2 to -2 indicate the data are normally distributed.

Metrics Min. Max. Mean S.E.a S.D.b Skewness Of forest I Edge density 0.55 35.71 19.47 0.14 7.06 -0.538 Large patch index 0.35 100 42.93 0.52 25.72 0.112 Landscape shape index 1 6.05 3.54 0.02 0.95 -0.218 Patch density 0.06 1.11 0.28 0.00 0.19 0.942 Mean contiguity index 0 0.92 0.35 0.00 0.21 1.133 Mean shape index 1 5 1.73 0.01 0.59 1.827 Disjunct core area density 0 0.72 0.19 0.00 0.12 0.586 Mean disjunct core area 0 1400 129 4.40 216 3.299 Nearest neighbour index 0 4272 529 6.46 316 2.546 Mean proximity index 0 41.56 12.82 0.22 10.56 0.405 Splitting index 1.01 83521 658 120 5855 12.553 Interspersion index 0 100 52.82 0.54 26.57 -0.355 Clumpy index -1 1 0.41 0.00 0.21 -3.251 Aggregation index 0 100 70.80 0.33 16.32 -1.241 Connectance index 0 100 45.77 0.66 32.41 0.324 Patch cohesion index 0 100 89.36 0.32 15.59 -3.234 Of forest II Mean patch size 6.25 1681 134 4.65 228 3.847 Edge density 0.42 33.08 17.85 0.13 6.60 -0.403 Landscape shape index 1 5.76 3.84 0.02 0.82 -0.494 Patch density 0.06 1.27 0.44 0.00 0.23 0.356 Percentage of landscape 0.35 93.08 30.32 0.38 18.58 0.615 Mean contiguity index 0 0.88 0.22 0.00 0.13 1.998 Mean shape index 1 4.19 1.43 0.01 0.39 2.621 Core percentage of landscape 0 63.67 6.38 0.18 8.85 2.138 Disjunct core area density 0 0.66 0.13 0.00 0.11 0.686 Mean disjunct core area 0 1075 43.66 1.78 87.24 5.498 Nearest neighbour index 0 3400 628 4.43 217 2.121 Mean proximity index 0 38.25 7.88 0.16 7.78 1.180 Splitting index 1.15 83521 648 88 4332 15.751 Interspersion index 0 100 55.61 0.51 24.87 -0.465 Clumpy index -1 1 0.39 0.00 0.17 -2.775 Aggregation index 0 100 56.89 0.37 18.03 -0.413 Connectance index 0 100 34.59 0.46 22.72 1.344 Patch cohesion index 0 100 77.37 0.41 19.98 -1.273

a Standard Error. b Standard Deviation.

55

Appendix 3

Detailed descriptions of the representative metrics selected by the metric reduction analysis.

Metrics Description

Largest Patch Index Quantifying the percentage of total landscape area comprised by the largest patch. Value ranges from 0 to 100. Value approaches 0 when the largest patch of the corresponding patch type is increasingly small. Value equals 100 when the largest patch comprises 100% of the landscape.

Landscape Shape Index Measuring class aggregation or clumpiness. LSI ≥ 1, without limit. Value equals 1 when the landscape consists of a single square or maximally compact patch of focal class. Value increases without limit as the patch type becomes more disaggregated.

Edge Density Equals the total edge length of focal class divided by the total landscape area. Value ≥ 0 (meters per hectare) without limit. Value equals 0 when there is no class edge in the landscape.

Mean Patch Area Equals the sum of all area of patches of the focal class divided by the number of patches of the focal class (Hectares).

Proximity Index Designed to measure the isolation of a patch within a complex of patches, given a specified search radius. Index equals 0 if a patch has no neighbors of the same patch type within the specified search radius. Value increases as the neighbourhood is increasingly occupied by patches of the same type and as those patches become closer and more contiguous (or less fragmented) in distribution.

Clumpy Index Accounting for the proportion of like adjacencies will equal the proportion of the landscape occupied by focal class for a completely random distribution. Value ranges from -1 to 1. Index equals -1 when the focal patch type is maximally disaggregated; equals 0 when the focal patch type is distributed randomly, and approaches 1 when the patch type is maximally aggregated.

56

Appendix 4

Partial correlation coefficients (n = 3735) among the eight representative metrics, controlling for the elevation.

Metrics 1 2 3 4 5 6 7 8 1. Edge density of forest I - 2. Largest patch index of forest I 0.07 - 3. Mean proximity index of forest I 0.33 0.24 - 4. Clumpy index of forest I -0.07 0.10 0.08 - 5. Mean patch area of forest II -0.19 -0.33 -0.23 -0.03 - 6. Landscape shape index of forest II 0.33 0.00 0.24 -0.16 -0.27 - 7. Mean proximity index of forest II 0.10 -0.32 -0.16 -0.01 0.18 0.10 - 8. Clumpy index of forest II -0.22 -0.18 -0.05 0.23 0.15 -0.27 0.21 -

57

Appendix 5

Map of the giant panda’s habitat in China. Green areas in the map are habitat of the giant panda. This map is cited from the 3rd National Panda Survey Report (State Forestry Administration of China, 2006).

58