Updating the Phase 1 habitat map of Wales, UK, using satellite sensor data

ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 81–102

Contents lists available at ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing

journal homepage: www.elsevier.com/locate/isprsjprs

Updating the Phase 1 habitat map of Wales, UK, using satellite sensor dataRichard Lucas a,∗, Katie Medcalf b, Alan Brown c, Peter Bunting a, Johanna Breyer a,b, Dan Clewley a,b,Steve Keyworth b, Philippa Blackmore a

a Institute of Geography and Earth Sciences, Aberystwyth University, Aberystwyth, Ceredigion, SY23 2EJ, United Kingdomb Environment Systems Ltd., 8G Cefn Llan Science Park, Aberystwyth, Ceredigion, SY23 3AH, United Kingdomc Countryside Council for Wales (CCW), Maes-y-Ffynnon, Penrhosgarnedd, Bangor, Gwynedd, LL57 2DW, United Kingdom

a r t i c l e i n f o

Article history:Received 5 April 2010Received in revised form10 September 2010Accepted 16 September 2010Available online 18 November 2010

Keywords:Land coverHabitat classificationSatelliteWalesObject-oriented eCognition

a b s t r a c t

The Phase 1 Survey is the most comprehensive and widely used national level map of semi-naturalhabitats inWales. However, the survey was based largely on field survey and was conducted over severaldecades, before being completed in 1997. Given that resources for a repeat surveywere limited, this studyhas used an object-orientated rule-based classification implementedwithin eCognition ofmulti-temporalsatellite sensor data acquired between 2003 and 2006 to map semi-natural habitats and agriculturalland across Wales, thereby allowing a progressive update of the Phase 1 Survey. The classification ofobjects to Phase 1 habitat classes was undertaken in two steps; firstly the landscape ofWales was dividedinto objects using orthorectified SPOT-5 High Resolution Geometric (HRG) reflectance data (10 m spatialresolution) and Land Parcel Information System (LPIS) boundaries. A rule-base was then developed toprogressively discriminate and map the distribution of 105 sub-habitats across Wales based on time-series of SPOT HRG, Terra-1 Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER)and Indian Remote Sensing Satellite (IRS) LISS-3 data, derived datasets (e.g., vegetation indices, fractionalimages) and ancillary information (e.g., topography). The rules coupled knowledge of ecology and theinformation content of these remote sensing data using a combination of thresholds, Boolean operationsand fuzzy membership functions. A second rule-base was then developed to translate the more detailedsub-habitat classification to Phase 1 habitat classes. Indicative accuracies of the revised Phase 1 mapping,based on comparisons with the later Phase 2 survey (for selected habitats), were >80% overall andtypically between 70% and 90% for many classes. Through this exercise, Wales has become the firstcountry in Europe to produce a national map of habitats (as opposed to land cover) through object-orientated classification of satellite sensor data. Furthermore, the approach can be adapted to allowcontinual monitoring of the extent and condition of habitats and agricultural land.

© 2010 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published byElsevier B.V. All rights reserved.

1. Introduction

The Phase 1Habitat Survey (Jones et al., 2003; JNCC, 2007) is theprimary spatial dataset showing the distribution of semi-naturalhabitats across Wales. This was initiated in 1979 as the UplandSurvey (Howe et al., 2005, completed in 1989) and followed by alowland Phase 1 survey (completed in 1997). The two surveysweresubsequently combined and reworked to provide a seamless Phase1 dataset for the entire country.

The maps were drawn county-by-county by trained fieldofficers who interpreted ground-observed habitats onto 1:10 000Ordnance Survey basemaps (Blackstock et al., 2007) using amix of

∗ Corresponding author. Tel.: +44 1970 622612; fax: +44 1970 622659.E-mail address: [email protected] (R. Lucas).

0924-2716/$ – see front matter© 2010 International Society for Photogrammetry anddoi:10.1016/j.isprsjprs.2010.09.004

pre-defined land use and land cover classes. Mapping was largelyfrom vantage points although unregistered aerial photographsviewed in stereo were used to support the ground survey,particularly in the uplands (JNCC, 2007). The level of detail was alsovariable between counties with, for example, hedgerows mappedin Powys but not in other counties. The Phase 1 Survey mapped103 distinct habitats (Table A.1) and was first released as a seriesof papermaps, then as a scanned raster layer, and finally as a digitallayer in the early 2000s. Today, the original Phase 1 Survey is stillbeing used as the primary habitat map for Wales as it has provendifficult and expensive to update. However, updates are needed assignificant changes in semi-natural vegetation have subsequentlyoccurred across theWelsh landscape, largely because of changes inagriculture and forestry activities, practices and policies. Furtherchanges are anticipated as a consequence of climate change andassociated mitigation strategies (Berry et al., 2008).

Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.

http://dx.doi.org/10.1016/j.isprsjprs.2010.09.004

http://www.elsevier.com/locate/isprsjprs

http://www.elsevier.com/locate/isprsjprs

mailto:[email protected]

http://dx.doi.org/10.1016/j.isprsjprs.2010.09.004

82 R. Lucas et al. / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 81–102

The primary purpose of the original Phase 1 Survey was toidentify candidate sites in Wales worth designating as Sites ofSpecial Scientific Interest (SSSI; Spash and Simpson, 1994), andeffort therefore focused on semi-natural habitats although all landwas mapped. Where the Phase 1 showed a lowland area to be ofhigh habitat quality, a Phase 2 Surveywas conducted subsequentlybut used detailed quadrat (1 × 1 m) survey methods (Rodwell,2006) to delineate and classify stands of vegetation accordingto species assemblages. For description, the National VegetationClassification (NVC) categories were used (Rodwell, 2006). Thesurveys recorded many species that were infrequent or small insize (JNCC, 2007) and hence some classes of habitats distinguishedon Phase 2 maps did not correspond with the distribution ofland cover observed in aerial photographs. The Phase 2 Survey,which is still ongoing, covers 1079 grassland sites, heathlands andlowland peatlands at many locations spatially dispersed acrossWales. Unlike the Phase 1 Survey, wall-to-wall coverage acrossWales will not be provided.

Neither the Phase 1 or Phase 2 Surveys utilised satellite sensordata. However, these data have been used formore timelymappingof land cover and largely for reporting at the national (UK) andEuropean level. Using 30 m spatial resolution Landsat ThematicMapper (TM) data for the period 1988–1989, the first satellite-derived map of land cover of Wales was created as part of theUnited Kingdom (UK) Land Cover Map (Fuller et al., 1994). Amaximum likelihood classification of 25 classes representing broadand nationally consistent vegetation categories was applied ata pixel-level, with mapping undertaken in conjunction with asample-based countryside survey (CS1990). A follow-up map for2000 (LCM2000) again used Landsat-sensor data acquired duringthe winter and summer months but these were segmented tobetter identify relatively homogeneous areas (e.g., fields, waterbodies; Fuller et al., 2002, 2005). Knowledge-based algorithms thatintegrated ancillary data, such as soils and elevation, were alsoapplied. The mapping identified 72 classes across the UK, whichwere subsequently grouped into 25 thematic subclasses. A furtherupdate is expected using a generalisation of the UK OrdnanceSurvey ‘mastermap’ GIS objects (Smith and Wyatt, 2007). Fulleret al. (2005) noted several limitations to the land cover maps.Many of the broad vegetation classes selected were not idealfor detection using remote sensing data as these had often beendefined by combinations of plant species, by the substratum or byelements of land use not discernible in images. Some vegetationcategories typically occurred in stands that were rare, localised orsmaller in area than the spatial resolution of the imagery. Whilelimitations on the separability of classes in the imagery couldbe partially solved by only mapping broad habitats, many usersrequired more detailed information on classes beyond these.

An alternative satellite-based approach to discriminate andmap broad habitats defined through the Biodiversity Action Plan(BAP; Jackson, 2000; JNCC, 2007) was suggested by Lucas et al.(2007) as part of a pilot study commissioned by the CountrysideCouncil for Wales (CCW) and the British National Space Centre(BNSC). In this study, which focused on the Berwyn Mountainsin mid north Wales, a time-series of Landsat TM/Enhanced TMdata from July 2001 and March, April and September 2002 wasused as input to a rule-based segmentation and classificationprocedure developed within eCognition (Definiens, 2008). Theapproach differed from more traditional methods of classificationin that the landscape was first segmented into objects of sizesthat varied from individual pixels to entire fields (as defined usingLand Parcel Information System (LPIS) boundaries). Numericalrules, designed to capture known correspondences between thedistribution of habitats in the landscape and their manifestationwithin remote sensing data, were then applied. The sequentialapplication of these rules allowed semi-natural habitats and

agricultural land cover classes to be discriminated, in some cases,to a level comparable to or better than those mapped previouslyusing Phase 1 Survey field methods.

This paper describes the extension of this approach to allcoastal, lowland and upland habitats across Wales based ondata acquired by several optical satellite sensors between 2003and 2006, for purposes of updating the Phase 1 Survey. Whilstsome Phase 1 habitats (e.g., A1.2.2 coniferous plantations andC1. Bracken; Appendix) were mapped directly, the majority weretranslated from a more detailed classification of sub-habitatsundertaken within eCognition (although generally with lessdetailed classes compared to the NVC used in Phase 2). The paperdescribes the approach to discriminating and mapping these sub-habitats using satellite sensor, presents the sub-habitat map forWales, and demonstrates a method for translating the sub-habitatclasses to Phase 1 habitat classes, using southeast Wales as anexample. The revised Phase 1 map for Wales will be presentedin a subsequent paper. The method of assessing the classificationaccuracy of the derived Phase 1 maps is described and indicativeaccuracies are provided.

The paper is structured as follows. Section 2 provides anoverview of the distribution and variability of semi-naturalhabitats in Wales and considers the factors complicating but alsofavouring the use of remote sensing data for habitat classification.Section 3 provides a detailed overview of the methods usedto discriminate and map vegetation communities (sub-habitats)across Wales and the approach used to subsequently translatethe classification to Phase 1 habitat classes. The detailed mapof sub-habitats across Wales is presented in Section 4, togetherwith the derived Phase 1 habitat map for Southeast Wales. Theapproach to accuracy assessment is described and indicativeaccuracies provided. The rule-based approach to classification andadaptability of the mapping is discussed in Section 5 and the studyis concluded in Section 6.

2. Study area and approaches to mapping

2.1. Physical environment and vegetation

Wales occupies an area of approximately 20 000 km2 and isprimarily a mountainous country (Fig. 1). At 1085 m, the highestmountain is Snowdon but over 20 distinct mountainous areas (e.g.,the Berwyn Mountains, Cambrian Mountains, Brecon Beacons,Black Mountains) have summits with elevations over 600 m.Lowlands are found east of the mountains, towards the coast andin broad valleys penetrating the uplands.

The climate of Wales is largely maritime. Weather is oftencloudy, wet, windy and mild but varies with altitude and distancefrom the sea. Average temperatures range from 6.5 °C in Januaryto 19.1 °C in July. Average rainfall increases from 600–1100 mmon the coastal margins and the southern and eastern lowlands to>3000 mm in the mountains of the northwest (MeteorologicalOffice, 2009).

Wales is renowned for its high diversity of landscapes (Black-stock et al., 2007). On the coastal margins, a range of grasslandsandheaths, sand dune complexes and estuarine habitats (includingsaltmarshes) occur. The lowland area (<∼300m) is used primarilyfor agriculture (sheep and cattle grazing but also arable crops) andmany of the extensive grasslands have been improved through re-seeding, liming and fertilisation. Semi-natural habitats (e.g., mires,marshy grasslands) still remain in the lowlands, although they arescattered and fragmented. Wales also has many forests; largelyconiferous plantations dominated by sitka spruce (Picea sitchensis)and European larch (Larix decidua) in the uplands and broadleavedand mixed forests in the lowlands. The unforested uplands sup-port low intensity agriculture (primarily for sheep production) onsemi-natural heaths, moors, bogs and open grasslands, with thelatter typically dominated byMolinia caerulea (purple moor grass),

R. Lucas et al. / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 81–102 83

Fig. 1. The location of Wales within the UK, showing major towns and cities and mountain ranges.

Nardus stricta (Mat grass), Agrostis (Bent) and Festuca species(Sheep’s fescue).

2.2. Mapping of Phase 1 habitats: challenges for remote sensing

The Phase 1 Habitat Survey divides habitats in Wales into 10broad types (lettered A–J; Table 1), within which 103 specifictypes are recognized (Appendix 1; JNCC, 2007). While thesecan be readily differentiated and described in the field, theirappearance in remote sensing (spectral) data varies as a functionof regional climate, seasons, soils and topography along with morelocal effects relating to land use and management. In particular,variations in regional and local climate across Wales influencethe phenology of vegetation (i.e., the timing of leafing, floweringand senescence). For example, in the spring, fronds of bracken(Pteridium aquilinum) appear earlier in the warmer south than inthe north and on sunny south-facing slopes before shaded north-facing slopes. The overall structure and species composition ofsome habitats (e.g., scrub and broadleaved woodlands) also varieswith elevation, geology, soils and maritime influences (Grimeet al., 1988; Yeo and Blackstock, 2002; Rodwell, 2006). Whilstvariability in the seasonal and geographical response of vegetationcommunities associated with these habitats can complicate the

interpretation of remote sensing data, differences in reflectanceand derived data (e.g., vegetation indices) as well as managementpractices (e.g., ploughing, grazing) and the physical environment(e.g., topography) can conversely be exploited to facilitate theirdiscrimination within a rule-based classification.

Reflecting on LCM2000, Fuller et al. (2005) considered thatthe definitions of habitats can be too broad or variable to beextracted from remote sensing (namely Landsat sensor) data asa single class. In this context, some Phase 1 habitats (e.g., C1.1bracken) are well suited for classification as they are extensiveand dominated by a single species. However, a greater number arecomprised either of a variable mix of species sharing dominance(or with sub-dominance), or have alternative dominants givingthem the appearance of discrete and separate plant communities.As examples, Vaccinium myrtillus (bilberry) and Calluna vulgaris(heather) species occur together in upland mosaics where eitherone can be locally dominant, but nevertheless are constituents ofthe same Phase 1 class (Dry Dwarf Shrub Heath; D1.1). L. deciduaand Scots Pine (Pinus silvestris) are found separately (and oftenin discrete homogeneous stands) but are referred to jointly asconiferous woodland (A1.2.2). Any attempt at mapping Phase 1classes from remote sensing data must therefore either combinemaps of these different species/communities based on relative


Table 1Phase 1 habitats and their distribution across Wales. See Table A.1 for class names.

Woodlands Grasslands Tall herb &fen

Heathland Mires Swamp,marginal &inundated

Open water Coastland Rockexposureand waste

Miscellaneous

A (14.2%)1 B (62.2%) C (3.2%) D (5.2%) E (3.0%) F (0.1%) G (0.8%) H (2.8%) I (0.7%) J (7.9%)Class (%)2 Class (%) Class (%) Class (%) Class (%) Class (%) Class (%) Class (%) Class (%) Class (%)

A1.2.2 57.8 B4 78.8 C1.1 98.3 D1.1 61.3 E1.6.1 33.1 F1 90.1 G1 56.4 H1.1 58.7 I2.2 30.8 J3.6 47.7A1.1.1 29.9 B1.1 10.5 C3.1 1.6 D5 24.7 E1.7 24.8 F2.2 9.9 G2 43.6 H2.6 10.6 I2.1 26.4 J1.1 40.0A2.1 5.1 B5 3.5 Otherc 0.1 D2 9.5 E2.1 23.5 H1.2 6.0 I2.3 19.5 J1.2 6.5A1.3.2 3.6 B2.2 2.9 D6 4.1 E1.8 10.0 H6.8 5.1 I1.2.1 6.9 J3.7 1.6A1.1.2 1.9 B5.2 2.2 Otherd 0.4 E3.1 2.2 H6.5 4.5 I2.4 4.7 J1.5 1.4A4.2 1.6 B1.2 2.0 E3 1.7 H1.3 3.4 I1.4 3.9 J3.4 1.3Othera 0.1 Otherb 0.2 E1.6.2 1.6 H8.4 2.9 I1.2 3.7 Otherj 1.5

E3.1.1 1.4 H3.1 1.8 I1.4.1 1.7Othere 1.6 H8.5 1.6 Otheri 2.4

H8.1 1.5H6.4 1.2H3.2 1.1Otherh 1.5

1 Area of each Phase 1 habitat category (A–J) as a percentage of the total habitat in Wales.2 Percentage area of Phase 1 habitat category.a A4.1, A1.3.1, A4.3, A1.2.1.b B3.1, B3.2, B5.2, B1, B2.1.c C, C3.2.d D3, D7, D1.2.e E3.2, D4, E2.2, E2, E3.2.1, E3.3, E2.3, E3.3.h H4, H8.6, H8.2, H6.6.i I1.1.1, I1.1. I1.1, I1.4.2, I1.2.2, I1.3, I1.1.2, I1.5.j J4, J1.3, J1.4, J3.5.

proportions (e.g., Vaccinium, Calluna) or classify these separately(e.g., Pinus, Larix) and thenmerge to a Phase 1 class in a subsequentstep. For these reasons, a more detailed classification of sub-habitats is first needed which can then be translated to Phase 1habitat classes.

3. Methods

The following sections provide an overview of the availableremote sensing and ancillary data and pre-processing performed,the division of Wales based on biogeographical regions and imagefootprints, and subsequent segmentation of the landscape. Theapproach to generating a detailed classification of sub-habitatsusing decision rules implemented within eCognition and thesubsequent translation of these classes to Phase 1 habitat classesare also described. The sequence of processing, segmentation andclassification is indicated in Fig. 2.

3.1. Remote sensing and ancillary data

Previous mapping of Wales has relied primarily on the use ofLandsat sensor data but this imagery is limited by the relativelycoarse (∼30 m) spatial resolution and low frequency of cloud-freeobservations, particularly during the spring and summer periods(Lucas et al., 2007). The difficulty in acquiring sufficient Landsatscenes during these periods reduced the capacity to classifyhabitats as the seasonal variability in the spectral response ofvegetation could not be adequately captured. To overcome this,data acquired by other optical sensors of finer (∼10–25 m) spatialresolution — but operating in similar wavelength regions — wereexploited.

For most of Wales, SPOT-5 High Resolution Geometric (HRG)visible (400–700nm) to shortwave infrared (SWIR; 1400–2500nm)data (14 scenes with several on the same date) were acquired forthe period 2003–2006, with the majority being relatively free ofcloud. Most data were acquired in March, although, for several ar-eas, images were only available for July and/or September. TwoSPOTHRG scenes acquired on the samedate over Pembrokeshire in

southwest Wales were combined as a mosaic. The 10 m resolutionvisible and near infrared (NIR; 700–1400 nm) data provided thebasis for subsequent segmentation of the imagery, allowing moredetailed features in the landscape (e.g., hedgerows, buildings) to beresolved.

To capture the seasonal variation in the reflectance of vegeta-tion, additional data acquired by the Terra-1 Advanced SpaceborneThermal Emission and Reflection Radiometer (ASTER; 15–30 m)and Indian Remote Sensing Satellite (IRS) LISS-3 (24 m) were used.Four IRS scenes acquired in July 2006 provided near complete cov-erage of Wales, although some sections were partially cloud cov-ered. Several ASTER scenes extending from Swansea Bay in thesouth to the Berwyn Mountains in the north were acquired inMarch 2003 and were combined into a single strip mosaic. ASTERdata were also available during the spring and summer for muchof Wales. While some Landsat scenes were available, these werecompromised by cloud cover and were considered to be of a spa-tial resolution insufficient for detailed habitat mapping, especiallywithin lowland areas.

Digital aerial photography and thematic data were provided bythe Welsh Assembly Government (WAG) and CCW to inform theclassification of the remote sensing data. For Wales, a completeset of true colour and NIR digital (Vexcel ultracam D) aerialphotographs at 0.4 m spatial resolution had been acquired largelywithin the summer of 2006. Full aerial photographic coveragefor Wales was also available for 2001. Field ecologists, alreadyexperienced inmapping Phase 1 habitats using aerial photographs,contributed to the interpretation of this imagery.

To support the classification of habitats, ancillary datasetsincluding a 5 m spatial resolution NextMap Intermap DigitalTerrainModel (DEM), Land Parcel Information Service (LPIS) vectordata, OSMastermap urban andwater body data, and a coastalmask(generated from the previous Phase 1 mapping) were used. TheLPIS data contained the field boundaries of all land parcels wherefarmswere claiming subsidy payments. Phase 2 (JNCC, 2007)mapsfor lowland grasslands and upland sites were used for subsequentassessment of classification accuracy.


Fig. 2. Flow chart showing the sequence of procedures used in the preparation of data and the subsequent segmentation and classification of sub-habitats and Phase 1habitats.

a b

Fig. 3. Spatial distribution of (a) 15 biogeographic regions (A–O) and (b) these same regions integratedwith the satellite image footprints, creating 16 project (42 sub-project)areas.

3.2. Use of biogeographical regions

As the biogeographical and seasonal variability could not beaccommodated easily into only one realisation of a biophysicallyderived rule-base, Wales was divided into 15 biogeographicregions through reference to bioclimatic maps (Bendelow andHartnup, 1980; White and Perry, 1989). These reflected a rangeof environmental influences, topographic characteristics and thedistribution of habitats, as determined from the original Phase 1Survey (Fig. 3(a)). Several regions (A, D, F, I, K and O) encompassedthe coastal area, with habitats ranging from mudflats and hard

cliffs to coastal grasslands and heaths. Lowlands were observedin all regions, although were confined largely to A, D, F, H, I, J,K, N and O. Uplands mainly occurred in regions B, C, E, G, J, Land M. These 15 biogeographic regions were subdivided basedprimarily on the footprints of the available SPOT-5 HRG and IRSdata, with consideration given to the degree of overlap with thefootprints of ASTER scenes. Further subdivision took place duringthe early stages of analysis, both to reduce the variability alongsteep climatic gradients (e.g., in the highmountains of Snowdonia)or for logistical reasons where sub-regions were too large forcomputer processing. This resulted in 16 project areas collectively


Table 2Image acquisition periods by sensor for biogeographical regions 1–16.

Sensor Year Month Bioregion1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

SPOT HRG

2003 Feb • • • •

2003 Mar • • • •

2003 Apr •

2003 Sep • • • •

2004 Apr •

2005 Jul • • • •

2005 Sep • • •

2006 Nov •

Terra-1 ASTER

2003 Apr • • • • • • • • •

2003 Jul • • • •

2004 Apr •

2004 May •

2005 Jun •

2005 Sep •

IRS LISS 32005 May •

2006 Jul • • • • • • • • • • • • • • • •

Denotes spring imagery

a b

Fig. 4. SPOT-5 HRG image (NIR, SWIR and visible red in RGB) (a) before and (b) after topographic correction using ATCOR 3. (For interpretation of the references to colourin this figure legend, the reader is referred to the web version of this article.)

containing 42 individual project sub-areas (numbered 1.1 throughto 16.1; Fig. 3(b)). The image dates associated with each of theproject areas (i.e., 1–16) are given in Table 2. SPOT HRG data wereavailable for all project areas except 9. Spring and summer images(all sensors) were available for all project areas with the exceptionof 5 (2 dates in July) and 11 (July and September).

3.3. Pre-processing of satellite data

Prior to mapping, pre-processing was necessary to allow in-tegration of the available satellite sensor data. Separate proce-dures were implemented for the SPOT HRG, IRS LISS-3 and ASTERdata and focused primarily on orthorectification, radiance calibra-tion and atmospheric correction, and the correction of topographicshadowing. All imagery were orthorectified (within IDL ENVI orErdas Imagine) using the 5 m spatial resolution NextMap BritainDigital ElevationModel (DEM) as a topographic reference. Depend-ing upon the orthorectification procedure, between 1 (for ASTER)and 20 (SPOT, IRS) ground control points were used, with these ex-tracted from the true colour aerial photograph mosaic of Wales.All satellite sensor data, regardless of their original spatial res-olution, were resampled to 5 m resolution. Initially, resamplingto the finest spatial resolution available (i.e., the 10 m SPOT HRGdata) was evaluated but many linear features (e.g., hedgerows) orboundaries between discrete vegetation communities (e.g., forests

and grasslands) became less distinct. However, by resampling to5 m, such features were better represented in the resulting seg-mentations. A nearest neighbour resampling algorithm was usedas the data values of the original image are retained (Richards andJia, 2006). The imagery were atmospherically corrected to unitsof surface reflectance (%) using the ENVI Fast Line-of-sight Atmo-spheric Analysis of Spectral Hypercubes (FLAASH) extension andtopographically corrected using ATCOR 3 (Richter and Schläpfer,2002; Richter, 2009). Topographic shadowing in the mountainousareas reduced surface reflectance, particularly on steep slopeswitha northerly aspect and where solar elevations (from horizontal)were <∼ 40° (between early November and late February). Spec-tral separation between vegetation with typically low reflectance(e.g., coniferous forests, Calluna-dominated heaths) was also re-duced on northerly facing slopes, and only a few vegetation types(e.g., improved grasslands) could be discriminated. For this reason,only imagery acquired outside of this period were used and, forthese, topographic correction was applied using ATCOR 3 to min-imize differences in reflectance as a function of slope, aspect andsolar angle (Fig. 4). Both atmospheric and topographic correctionallowed comparison of images acquired on different dates, fromdifferent sensors and also over different regions.

Following atmospheric and topographic correction, a numberof ‘standard’ image products were derived including estimates ofthe relative amount (fractions) of shade/moisture, photosynthetic


Fig. 5. Hierarchical segmentation of the super-level (top), level 1 (centre) and sub-level (bottom).

(green) vegetation (PV) and non-photosynthetic (dead/senescent;NPV) vegetation (Adams et al., 1995) and vegetation indices (e.g.,the Normalised Difference Vegetation Index (NDVI); Tucker, 1979).Cloud and cloud shadow screening was undertaken using a semi-automated procedure developed by Definiens AG. The resultingcloud masks allowed cloud-free sections of the imagery to be usedeven though other parts of the image were cloud covered.

3.4. Segmentation and initial classification

3.4.1. Integrating existing data layers.For each of the 42 projects, segmentation and classification was

performed first on spatial subsets (e.g., 300× 300 pixels), then re-evaluated on other subsets prior to application across the wholeproject area. Segmentation was undertaken in several stages. First,objects representing urban areas andwater bodies were generatedfrom the OSMastermap thematic layers and classified accordingly.All remaining unclassified areas were then segmented to objects1–2 pixels in area using a multi-resolution segmentation of theSPOT-5 HRG data (Benz et al., 2003). Objects outside of the LPISboundarieswere then classified as either uplands or lowlands,withthe distinction based upon an elevation threshold approximatingthe upper boundaries of the enclosed fields. Within the lowlands,the coastal zone was defined by the inland limits of coastal classesidentified in the Phase 1 on the assumption that, collectively, thesehad not changed in their distribution since the original mapping.

3.4.2. Generating hierarchical layersFollowing segmentation and initial classification (i.e., into

urban, water, lowland, upland and coastal), the objects wereduplicated to form an identical layer above (referred to as thesuper-level) and below (sub-level) the original (referred to aslevel 1; Fig. 5). At the super-level and within the LPIS, the imagewas resegmented using the LPIS data to give one object perLPIS unit (typically a field). For areas outside of the LPIS, largersegments (that approximated the size of fields) were generatedfrom the spectral data (Lucas et al., 2007). At the sub-level,and for objects outside of the LPIS, the image was resegmentedusing a chessboard segmentation (one 5 m pixel per object). Aresegmentation (through merging) of objects in level 1 was alsoundertaken to account for spectral variability within, for example,the larger LPIS units generated in the super-level (e.g., associatedwith invasive scrub or marshy grassland). The sub-level retainedthe original object sizes (typically 1–2pixels in area) andultimatelycontained the sub-habitat classification (see later sections).

The use of multi-level hierarchical layers is a feature of eCog-nition (Definiens, 2008) and permits variable segmentation of thelandscape depending on the dimensions, shape and homogeneityof landscape units. Larger objects were created at the super-level

Table 3Primary habitats classified within different hierarchical levels.

Hierarchicallevel

Zone Example habitats

Super-level Primarilylowland

ArableImproved grasslands

Level 1 Upland/lowland Woodlands (including felled) and scrubHedgerowsUrban and waterBare ground and rock

Coastal Mudflats, sand, saltmarshSub-level Upland/lowland Heaths, mires, bogs

BrackenImproved grasslandUnimproved grasslandsSemi-improved grasslandsFlushes and fens

Coastal GrasslandsHeathsDunes

to best capture homogeneous objects of relatively large area (e.g.,arable and improved fields) whilst the size of objects was progres-sively decreased at level 1 and in the sub-level (typically to 1–2pixels) to allow, for example, better description of complex up-land and lowland (including coastal) mosaic habitats (e.g., by usingfuzzymembership functions for vegetation communities compris-ing the mosaic). Woodlands (including felled), scrub, urban, waterand bare ground/rock) were assigned as discrete classes to objectsat level 1, with these typically being of smaller and larger size thanthose at the super-level and sub-level respectively. By using thesehierarchical layers, the landscape could be classified with consid-eration given to the size and complexity of habitats. Table 3 givesan overview of the levels within which classification of the mainhabitats occurred.

3.5. Classification of objects and tiling

Following initial segmentation, analytical rules were progres-sively developed by comparing existing ecological knowledge ofvegetation distributionswithin the landscapewith the informationcontent of the remote sensing data. A combination of single thresh-olds, Boolean (logical) operations and fuzzy rules were used, withthese developed within the eCognition GUI environment. Valuesfor rules were determined from the ecologists’ own local knowl-edge of the site preferences and characteristics of plant speciesand vegetation communities and with reference to field observa-tions conducted during the course of the study. Rules were devel-oped and tested on subsets of the imagery, and the distributionof mapped classes was compared with those observed within the2006 aerial photography. All rules and data used also gave consid-eration to ecological or biophysical characteristics (e.g., reflectanceof different species and growth stages, photosynthetic activity,proportion of dead material, moisture content, roughness, slope).Over 200 rules were developed in a logical sequence to map habi-tats acrossWales, with these based on reflectance data, band ratiosand differences, derived indices (e.g., NDVI, endmember fractions),relative difference to neighbours and seasonal differences in thesevalues (Tables 4 and 5). Several indices developed were also founduseful for discriminating specific habitats (e.g., broadleavedwood-lands, heaths). Whilst the data used within the rules remainedsimilar, the values generally varied depending upon the imageryused (i.e., ASTER, SPOT HRG or IRS LISS-3) and the dates of acquisi-tion. Up to three rules were also used for classification of the sub-habitats, with these adjusted for use in region-growing or fuzzyclassification procedures and implemented within and betweenthe different hierarchical layers. Because of the complexity of therule-base,actual values and individual rules are not presented but


Table 4Satellite single channel data used within the rule-base for classifying sub-habitats.

Channel Key habitats

1. Green (summer) Broadleaved woodland, lowland wet heath, uplands bogs and dry heath (Calluna-dominated), sand2. Red (summer) Non-vegetation, ploughed fields, felled coniferous forest, marshy grasslands (Molinia-dominated) in lowlands and

uplands, gorse (coastal), dune slacks3. NIR (summer) Water, bracken, hedgerows, improved grassland, gorse (coastal)4. SWIR (summer) Bracken, coniferous forest, improved grasslands, lowland mires and fens, upland unimproved grasslands

(Festuca-dominated)5. Green (spring) Broadleaved woodland, hedgerows, lowland marshy grasslands (Juncus-dominated), rock, gorse (coastal),

saltmarsh6. Red (spring) Water, lowland mires and fens, lowland scrub, upland marshy grasslands (Juncus-dominated), upland heaths

(Calluna-dominated), grey dunes7. NIR (spring) Water, improved grasslands, upland unimproved grasslands (Festuca-dominated), coastal grasslands, rock, gorse

(coastal)8. SWIR (spring) Water, coniferous forests, lowland/upland marshy grasslands (Molinia-dominated), lowland unimproved

grasslands. Upland unimproved grasslands (Nardus/Festuca-dominated), coastal scrub9. Vegetation clusters Open water, gorse

10. Shade fraction (spring) Gorse and other scrub11. PV fraction (spring) Bare ground, lowland/upland semi-improved grasslands, grey dunes, yellow dunes12. NPV fraction (spring) Lowland wet heath, upland bogs and dry heaths (Calluna-dominated), coastal scrub, dune slacks13. Shade fraction (summer) Rock features, coniferous forest14. PV fraction (summer) Bracken, lowland/upland semi-improved grasslands15. NPV fraction (summer) Bare ground16. Elevation Bracken, coniferous forest, coastal habitats17. Slope Bracken, coniferous forest, lowland/uplandMolinia-dominated marshy grasslands, lowland mires and fens and wet

heath, bog (Eriophorum-dominated and unmodified), cliffs

rather the sequence of rules (as they are implemented) and thedata and indices used for the classification of sub-habitats are de-scribed.

3.5.1. Thresholds and Boolean operationsInitially, rules were developed to classify habitats that were

spectrally distinct, largely contiguous and relatively homogeneous.These included non-vegetated areas, coastal marginal habitats(e.g., mudflats), agricultural fields, bracken, woodlands and densescrub. Typically, rules classified the core area using either non-fuzzy thresholds or combinations of thresholds (using and/orBoolean operations). Region-growing procedures were then usedto extend the area initially classified to the edges of the habitat.The rules were applied sequentially and, once assigned to a sub-habitat class, objects were then not available for subsequentreclassification to another sub-habitat class. In the followingdescription, the numbers given in parentheses refer to the channelsand indices used (Tables 4 and 5).Non-vegetated areas: Non-vegetated areas included the pre-defined classes of water and urban. Remaining water bodies wereidentified as they exhibited low NIR and SWIR reflectance (7, 8)whilst other non-vegetated surfaces were associated with a lowNDVI. Large rock and spoil features were distinguished by theirhigh green reflectance (5) and low NDVI (20) and by using channelratios (23, 24).Coastal marginal habitats: As a starting point, areas of open water(marine or estuarine) were classified and then intertidal objectslocated at the edge of water were associated with, for example,mudflats, sand beaches or rock based on spectral thresholds(e.g., 1, 23). Bordering objects were then assigned to these sameclasses using iterative region-growing techniques (Baatz andSchäpe, 2000). Other coastal classes were then assigned based ontheir adjacency to these mapped classes and similarly expandedthrough region growing. For example, objects were identifiedas lower salt marsh only if they were adjacent to mudflats oropen water (depending upon the tidal state). Objects adjacentto lower salt marsh were then iteratively classified as such ifthey met specified criteria. Where these criteria were not met,objects were assigned instead to mid salt marsh (where present)and iteratively expanded. Upper salt marsh classes were similarlyclassified. Through this approach, the zonation patterns typically

associated with many coastal marginal habitats were captured.The extent of many habitats occurring on the coastal margins waslimited by elevation.Agricultural fields (including improved grasslands): At the super-level, and within the lowlands, the larger objects were identifiedas ploughed, harvested, mown or grazed fields (i.e., croplands orpasture) where they exhibited the reflectance characteristics ofgrowing vegetation in one observation period (season) but non- orsparse/short vegetation (including stubble) in another (Lucas et al.,2007). As the timing of the agricultural cycle varies though the year,additional imagery or agricultural surveys (already obtained aspart of the LPIS) would facilitate better classification of such areas.Fields under improved grasslands but with less change over theyear (e.g., because of no ploughing), exhibited a consistently high(relative to less improved fields) NIR reflectance, photosyntheticfraction and/or NDVI during the growing season. Smaller objectsin level 1 and the sub-level that were spatially nested within thelarger objects of the super-level were then assigned according totheir allocation in the super-level.Bracken: Bracken was discriminated from a combination of springand summer imagery, using rules that utilised region-growing pro-cedures and exploited the large difference in the photosyntheticfraction and NIR reflectance between dates (e.g., 38, 41). Separatemodified rules were also developed to account for differences inthe timing and appearance of bracken growing on north and steepfacing slopes. All bracken ruleswere applied to unclassified objectsin the lowlands and uplands.Woodlands and dense scrub: Coniferous forests (plantations) werelargely confined to areas outside of the LPIS and exhibited a lowSWIR reflectance in the spring and summermonths. Separate ruleswere developed for L. decidua (European Larch) because of thedeciduous nature of this tree species, with these exploiting thelarge differences in the NPV fraction (42) between the spring andsummer imagery. Felled woodlands exhibited large differencesin the red and SWIR bands (37, 39) and NDVI (40) betweendates of satellite data acquisition. Where woodlands had beenfelled previous to data acquisitions, these exhibited low summerNDVI (20) and high red reflectance (2) values and were oftenin proximity to existing woodlands. Broad-leaved woodlandstypically exhibited a low green reflectance (1, 5) in both spring and


Table 5Indices used for classifying sub-habitats within the rule-based classification and based on reflectance (ρ) in the green (g), red (r), NIR (n) or SWIR (s) wavelength regions.

Index Formula Key habitats

Standard indices

18. NDVI (spring) ρn−ρrρn+ρr

Lowland scrub, upland marshy grasslands (Juncus-dominated), improvedgrasslands (uplands)

19. NDVI (summer)1 ρn−ρrρn+ρr

Lowland flushes and marshy grasslands, gorse and unimproved/improvedgrasslands. Broad-leaved woodland and larch plantations. Acid flushes,non-vegetation, unmodified bogs (Eriophorum-dominated), dry heath(Calluna-dominated)

20. NDVI (summer)2 ρn−ρrρn+ρr

Lowland mires and improved grassland, broadleaved woodland, felled coniferousforest, lowland mires and fens, saltmarsh

Band ratios

21. Green:SWIR (summer) ρg/ρs Broadleaved woodland, rock, yellow dunes22. NIR:Red (summer) ρn/ρr Upland Juncus and Nardus-dominated grasslands23. NIR:Green (summer) ρn/ρg Rock features24. Red:SWIR (summer) ρr/ρs Rock features

Band differences and products

25. NIR-SWIR (spring) ρn − ρs Lowland unimproved grasslands26. NIR-SWIR (summer) ρn − ρs Upland unimproved grasslands (Nardus and Festuca-dominated)27. SWIR-NIR (summer) ρs − ρn Wet bog with Eriophorum spp28. SWIR-NIR (spring)3 ρs − ρn Sphagnum bog, coniferous forest29. SWIR-Green (spring)3 ρs − ρg Coniferous forest30. NIR x Green ρnρg Gorse and other isolated or contrasting patches of vegetation, hedgerows31. NPV-PV fraction (spring) NPVspring − PVspring Hedgerows

32. Normalised seasonal red difference ρr,spring−ρr,summerρr,spring+ρr,summer

Lowland/upland marshy grasslands (Juncus-dominated), dry heath(Vaccinium-dominated), unmodified bogs (Eriophorum vaginatum-dominated)

Relative difference to neighbours

33. NIR reflectance (summer)4,5 ρ ′n − ρni Gorse and other isolated contrasting patches of vegetation, hedgerows

34. Green reflectance (summer)4,5 ρ ′g − ρgi Hedgerows

35. SWIR (spring)5 ρ ′s − ρs Juncus-dominated upland marshy grasslands

Seasonal difference images

36. Green difference ρg,spring − ρg,summer Flushes (Juncus-dominated)37. Red difference ρr,spring − ρr,summer Upland unimproved grasslands (Festuca-dominated), gorse scrub, felled woodland38. NIR difference ρn,summer − ρn,spring Grasslands (Festuca-dominated) with bracken, hedgerows39. SWIR difference ρs,summer − ρs,spring Lowland mires and fens, felled woodlands40. NDVI difference NDVIsummer − NDVIspring Grasslands (Nardus and Festuca-dominated), unimproved grasslands, felled

woodland41. PV fraction difference PVsummer − PVspring Lowland/upland grasslands (Molinea-dominated), bracken, mature Calluna42. NPV fraction difference NPVsummer − NPVspring Larch plantations43. Green difference: SWIR difference ρg,spring−ρg,summer

ρs,summer−ρs,springHedgerows Grasslands (Juncus or Nardus-dominated), gorse

44. NIR (spring) & green (summer) ρn,summer + ρg,summer Upland improved grasslands45. NIR (spring) & SWIR (summer) ρn,spring − ρs,summer Upland unimproved grasslands (Nardus-dominated)

Other indices

46. Based on spring imagery ρr−ρsρr+ρsρr

Hedgerows, Larch plantations, broadleaved woodlands and dense scrub47. Based on summer imagery ρr−ρs

ρr+ρsρrLowland mires and fens and Festuca ovina grasslands (steep slopes, lowlands anduplands). Unmodified bogs (Eriophorum-dominated), dense scrub andbroadleaved woodland. Dry heath (Vaccinium-dominated)

48. Based on spring imagery ρr−ρsρr+ρ2

sLowland semi-improved grasslands

49. Based on summer imagery ρr−ρsρr+ρ2

sLowland semi-improved grasslands

50. Based on summer imagery ρr−ρsρr+ρsρg

Vaccinium-dominated dry heath51. Based on spring imagery ρnρg Hedgerows, broadleaved woodland, gorse. Lowland/upland marshy grasslands

(Juncus-dominated), upland dry heath (Calluna-dominated), upland unimprovedgrasslands (Festuca-dominated)

52. Heath detection ρs − ρrρn Lowland heath and scrub53. Gorse detection ρg,spring−ρg,summer

ρs,summer−ρs,springGorse

1 Indices based primarily on SPOT-5 HRG summer (March) imagery.2 IRS or SPOT summer (July) imagery.3 ASTER data only.4 SPOT High Resolution Geometric (HRG).5 ρ ′ represents reflectance of the neighbouring object.

summer and occurred primarily outside of the LPIS. Index 46 wasparticularly useful for allowing consistentmapping of broadleavedwoodlands within and between projects, even within shadowedareas associated with steeply sloping terrain. Similar rules, butwith lower thresholds of Index 46, were applied to map densescrub.

Scattered scrub and hedgerows: Areas of scattered scrub (e.g.,primarily dominated by gorse and/or blackthorn) were associatedwith isolated objects (e.g., within fields)and were identified usingrules similar to those applied to woodlands. Objects associatedwith hedges were identified initially by exploiting the relativecontrast in visible green (in spring; 5) or NIR (in summer; 3)


reflectance with surrounding open fields. Subsequent rules, basedlargely on adjacency, then allowed these objects to be ‘grown’along the field boundaries such that the hedgerow was mappedin its entirety. The mapping benefitted from the resampling of theSPOT HRG and all other imagery to 5 m spatial resolution.

3.5.2. Classification of objects (fuzzy)For complex mosaic habitats, a series of rules were applied

with these based on user-defined fuzzy membership functionsdeveloped within eCognition (Benz et al., 2003). For each classof sub-habitat (e.g., Calluna-dominated heath), up to threefunctions were defined, with these using reflectance data and/orderived products and, in some cases, topographic information(e.g., slope, aspect). Within eCognition, the functions wereapplied together and for all occurring classes (e.g., Calluna-dominated heath,Vaccinium-dominated heath,Molinia-dominatedgrasslands) rather than hierarchically, giving each object a fuzzymembership value for each class. The ruleswere applied separatelyfor habitats occurring in both the lowlands and uplands (e.g.,heaths) and on the coast (e.g., sand dune complexes, coastalheaths).Lowland habitats: All objects in the lowlands which had not yetbeen allocated to a class were collectively assigned to a ‘lowlandhabitats’ class, within which wet grasslands (Juncus and Molinia-dominated), semi-improved and unimproved grasslands, gorse,scrub, wet heath and both mires and fens were classified. Wetgrasslands dominated by Juncus species were identified primarilybecause of their low visible green (5) and red (6) reflectance, whichcontrasted with many surrounding habitats. These grasslands alsoexhibited lower NIR (7) reflectance and NDVI (18) in the springmonths compared to those that were more productive. Largedifferences in the photosynthetic fraction between the spring andthe summer months (41) as well as a high red (summer; 2) andlow SWIR (spring; 8) reflectance (associated with a combination ofhigh non-photosynthetic cover andmoisture content respectively)were used primarily to distinguish Molinia-dominated grasslandsfrom adjoining habitats (e.g., bracken). These grasslands werealso associated with areas of relatively level or gently slopingterrain (17). The drier unimproved grasslands were generally lessproductive, giving a lower SWIR reflectance (particularly in thespring; 8) and NDVI compared to other grassland types (excludingmarshy grasslands; 19). Semi-improved grasslands were generallymore productive, as indicated by a greater photosynthetic fractionin the spring and summer (11, 14), but less so than improvedgrasslands. Vegetation indices (e.g., 48, 49) incorporating the SWIRand red reflectance in the spring and summer were also used todiscriminate semi-improved grasslands. Lowland scrub associatedwith gorse (Ulex europaeus), hawthorn (Crataegus monogyna) andblackthorn (Prunus spinosa) supported a low NDVI (18) but highshade fraction in the spring (10) and also low and small seasonaldifferences in red reflectance (37). Indices 47 and 52were also usedto assist discrimination. Whilst lowland wet heath was relativelyuncommon, this sub-habitat was discriminated by the relativelylow green reflectance (summer; 1) and high NPV fraction (spring;12). Wet heath was also assumed not to occur on slopes toosteep for water to accumulate (17). Fens and mires were similarlyassociated with relatively level terrain and the classification usedindicators of wet areas, including low SWIR reflectance (summer;4). All lowland habitat not assigned to a class was reassigned tolowland improved grassland.Upland habitats: The uplands were classified into wet (marshy),unimproved and semi-improved grasslands using modificationsof the membership functions applied to the lowlands. Fordiscriminating Juncus-dominated grasslands and flushes, a numberof additional indices were used including 51. Improved grasslands

were also associated with a relatively high spring NIR reflectance(7) and NDVI (18). Unimproved grasslands dominated by Nardusspecies exhibited low differences in the NDVI between the springand summer (40) and within-scene differences in the NIR (7)and SWIR (8) reflectance relative to adjoining upland habitats,particularly in the spring. Festuca-dominated grasslands (withshort turf) typically exhibited a high spring SWIR and NIRreflectance (7, 8), with additional indices (e.g., 26, 51) providingfurther discrimination. Ulex (gorse)-dominated scrub was mappedusing rules developed for lowland habitats. All bog habitats wereassociated with shallow slopes (17). Calluna-dominated bogs,which are typically modified by drainage and are an ecotone intowet heath, exhibited a high NPV fraction (spring; 12) and alsolow green reflectance compared to adjoining habitats (summer; 1).Bogs dominated by cotton grass (Eriophorum vaginatum) exhibiteda relatively low NDVI (19) and several indices (27, 32 and 47)provided additional discrimination.Coastal habitats: Remaining coastal habitats were classified intograsslands, heathlands (with bracken, scrub and heath vegeta-tion) and dunes (open, grey and yellow). Coastal grasslands typ-ically exhibited a higher NIR reflectance in the spring (7), whichdistinguished them from most other communities. Scrub commu-nities were associatedwith a high NPV (spring; 12) and low red re-flectance (summer; 2) compared to adjoining coastal habitats. Theclassification of dunes (e.g., grey, yellow) and dune slacks utilisedthe red reflectance (spring and summer; 2), the photosynthetic andnon-photosynthetic fractions (spring, 11, 12) and the ratio of greento SWIR reflectance (summer; 21). Bracken and heathland (typ-ically Calluna/Erica-dominated) was classified using similar rulesapplied in the lowlands and uplands.

3.6. Translation to Phase 1 classes

3.6.1. Transfer of classes to the sub-levelFor each project, the classification of habitats was maintained

in three layers (i.e., the super-level, level 1 and sub-level), with ob-jects progressively decreasing in size from the super-level to thesub-level. To emulate the spatial scale of habitats mapped withinthe original Phase 1, larger objects were needed and generated asfollows. In the super-level, large objects already existed and in-cluded fields with arable crops and uniform improved grasslands.In the level below (i.e., level 1), objects associatedwith broadleavedwoodlands, coniferous woodlands, dense scrub, bracken and acidflusheswere simplymerged separately into large, continuous poly-gons. All class assignments in the super-level and level 1were thentransferred to objects in the sub-level. In this sub-level, all remain-ing objectswere generally smaller (1–2 pixels) and associatedwithcomplex habitats on the coast and in the uplands and lowlands,with each containing fuzzymembership values. The combined areaassociated with these objects was therefore resegmented usingSPOT spectral data (particularly the NIR and SWIR bands) and acoarse scale factor to create larger objects representing broad pat-terns of habitats across the landscape. All three sets of these largerobjects were then combined in a new separate level, with thistermed the ‘Phase 1 level’. This level was separate from, but spa-tially linked to, the original sub-habitat classification maintainedin the sub-level.

3.6.2. Object assignment to Phase 1 classesTo assign objects associatedwith sub-habitats to a Phase 1 class,

twomethodswere employed.Where a one-to-one correspondencebetween the sub-habitat class and a Phase 1 class occurred, theseobjects were simply reassigned (e.g., the class open water wasassigned to the Phase 1 class Open Water; G1). Other examplesof directly corresponding classes included broadleaved woodlands


Fig. 6. Map of 105 sub-habitat classes, Wales (see text for description of colours).

(A1.1), bracken (C1.1), dense scrub (A2.1) and urban (J3.4). Insome cases, several classes were combined (e.g., Larch and otherconiferous plantations were assigned to coniferous plantations;A1.1.1). However, where complex mosaics occurred, with theserepresented by the remaining large unclassified objects in thePhase 1 level, the fuzzy membership values of the nested smallersub-level objects were used to assign a Phase 1 habitat class.From these values, a multi-layered grid (with cells at 5 m spatialresolution) was generated for each project in turn, with eachlayer representing the fuzzy membership value (0–1) of the sub-habitat classes. The number of grid cells within each Phase 1 objectexceeding a fuzzy membership threshold of 0.2 (referred to as analpha cut; Nguyen and Walker (2006) was counted for the cellsbelonging to each Phase 1 object, allowing the sub-classes to beranked by number. The number of counts per class as a percentageof the top three classes in each Phase 1 object was then calculatedand passed on to a second rule base.

The second rule base, running in a C++ program (Bunting andClewley, 2010), translated these ‘top three’ percentage values intoPhase 1 habitats based on expert knowledge. The ruleswere run se-quentially, allowing objects assigned an initial Phase 1 class to bereassigned to another Phase 1 class by a later rule. As an example,

the percentage count for an object in the Phase 1 layer might havebeen greatest for the sub-habitat class ‘‘lowland mire’’, with thisdominated by Juncus species. As such, the object was initially clas-sified as the Phase 1 classMire (E2). However, if lowland grasslandswere also included with Juncus-dominated lowland mire in thetop three, then the object was reassigned subsequently to MarshyGrassland (B5). In this case, the encoded expert knowledge sug-gested that the marshy grassland might be developing into a mire(through invasion) or degrading from a mire (through drainage orgrazing) into a comparatively drier marshy grassland. The rule-base also considered other information including elevation, slope,convexity of slope and distance to water to assist in the translationof sub-level habitats to Phase 1 classes. In this second rule set, everycombination of sub-level classes could be associatedwith a Phase 1class. A number of Phase 1 classes had no distinct sub-habitat class,either because they were types of land use within the same landcover class (e.g., amenity grasslands which are identical to agricul-tural improved grasslands), were temporary (ephemeralwetlands)or were too similar to other classes (e.g., broadleaved plantationsas opposed to woodlands).


Marshy grasslands

Unimproved acid grasslands

Semi-improved acid grasslands

Improved grasslands

Dry heath

Bogs

Felled woodland

Broadleaved woodland

Bracken

Open water

Coniferous forest

a b

c

Fig. 7. (a) Sub-habitat classification and (b) Vexcel aerial photography of the Cambrian Mountains. An enlarged part of the classification is provided in (c).

4. Results and accuracy assessment

4.1. Sub-habitat map for Wales

The map of sub-habitats across Wales is given in Fig. 6 where105 sub-habitat classes were mapped across the coastal areas,

lowlands and uplands. At a broad scale, noticeable features inthe classification are the extensive areas of improved grasslands(pale green) throughout the lowlands and lower slopes of theupland regions. Coniferous forests (dark green) are distributedmainly along the spine of mountains extending from Snowdoniain the north to the Welsh Valleys in the south, whilst areas of


broadleaved woodlands (mid green) occupy many of the steeperslopes in the numerous valleys distributed throughout Wales.Coastal saltmarshes (pink) and beaches (yellow) are particularlyprominent around the Gower North Peninsula and in the estuariesof the west coast (including the Dyfi and Mawddach Estuaries andthe Menai Straits, Anglesey). Bracken (orange) is most extensivetowards the east and south of the Cambrian Mountains whilst theupper slopes and tops of this range are dominated by unimproved(pale yellow) and marshy grasslands (blue/green). In much ofSnowdonia and the Berywn Mountains, bog and mosaics are morecommonplace.

An enlarged part of the mapping is provided in Fig. 7 for theCambrian (west of a north–south line between Mynydd Bach andNant-y-Arian) and exemplifies the level of detail achieved. Here,extensive areas of marshy grassland dominated by M. caeruleaoccur in the higher altitude areas with these interspersed withdry and wet heath, occasional bogs and unimproved acid grass-lands. On the lower slopes and particularly on the eastern flanks,bracken (P. aquilinum) and flushes (Juncus-dominated) exist alongwith grasslands at various stages of improvement. Coniferous plan-tations (including felled areas) are also widespread. Broadleavedwoodlands and scrub are scattered in fragments throughout thearea. A more detailed view of the sub-habitat mapping is providedin Fig. 7(c).

As an example of the mapping of coniferous forests, broad-leaved woodlands and hedgerows (including clusters of trees),these classes are overlain onto a 2006 aerial photograph mosaicof Monmouthshire in southeast Wales. The majority of hedgerowsare mapped, with only those that are limited in width/height orwith gaps omitted. The mapping also includes isolated patchesof trees or scrub (Fig. 8). Similar detail in the mapping hasbeen achieved for much of Wales, although the delineation ofhedgerows was compromised when SPOT-HRG data were notavailable for segmentation (e.g., Project 9).

Compared to the original Phase 1 Survey, the sub-habitat clas-sification provides a more detailed breakdown of the communitiesoccurring. For mosaic communities on the coast and within bothlowlands anduplands, the fuzzymembership values provide an ap-proximate relativemeasure of the proportions of dominant speciesor genera associated with each object; hence even greater detail inthe classification is provided. As an example, Fig. 9 illustrates fuzzymembership images for Calluna- andVaccinium-dominated heaths,combinations of whichwere used to define the extent of Dry Heath(D1.1).

4.2. Revised Phase 1 maps

The generation of the revised Phase 1 map for Wales is ongoingand will be the subject of a forthcoming paper. However, anexample of the maps generated from the sub-habitat classificationis given for southeast Wales (Fig. 10 with key in Fig. 10b; Projects15.1, 15.2 (west) and 15.3). The revised Phase 1 classification is lessdetailed than the sub-habitat classification but for this wider area,shows the dominance of dry heaths (D1.1), bracken (C1.1), marshygrassland (B5) and unimproved acid grassland (B1.1) in the uplandregions of the Black Mountains and eastern edge of the BreconBeacons. Broadleaved woodlands (A1.1.1) are most prevalent inthe Wye Valley whilst coniferous plantations (A1.2.2) are morewidespread across the region. Much of the eastern area is occupiedby improved grasslands.

A more detailed example of the revised Phase 1 mapping, thatincludes the Blaenavon Industrial World Heritage site (focusedon Torfaen), is given in Fig. 10b. Here, the distribution of urbanareas is influenced by the former coal mining industries andsettlements follow the valley running north to south. The lowlandareas towards the south are dominated by improved agriculture,

Fig. 8. Sub-level classification of agricultural land west of the Wye Valley inMonmouthshire showing coniferous plantations (dark green; A1.2.2), semi-naturalbroadleaved woodlands (mid green; A1.1) and intact hedgerows (pale green;J2.1) which are directly translated into Phase 1 classes. Non-vegetated surfaces(primarily buildings and infrastructure) are shown in black. Isolated clusters of treesare alsomapped. (For interpretation of the references to colour in this figure legend,the reader is referred to the web version of this article.)

although a range of grassland types (unimproved to improved)occur. The steep valley sides support coniferous plantations,broadleaved woodlands and dense/scattered scrub. The valleys inthis region are separated by upland areas dominated by dry heathand unimproved acid grasslands. Hedgerows were not mapped inthe original Phase 1 Survey (with the exception of Powys) but areincluded in the revision.

The translation of sub-level habitats to Phase 1 habitatsinvolved some loss of spatial and thematic detail. Nevertheless, theresulting maps conveyed the general layout of the landscape anda high level of detail for each of the habitats. A particular benefitof the classification was that many features (e.g., hedgerows,water bodies) were easily recognisable, which better placed theclassification into the context of the landscape.

4.3. Accuracy assessment of individual projects

Within each project, three levels of information were provided;the sub-habitat classification of objects, the fuzzy membershipvalues associated with each object (in areas with complexmosaics), and the Phase 1 habitat classification (derived fromthe sub-habitat object classes or fuzzy membership values). Ateach stage, errors in the classification could be contributed bya number of factors, including the input data used, the rulesthemselves, and themethods for translating the fuzzymembership


Membership

N

1 Kilometres0 0.25 0.5

a b c

0 - 0.10.1 - 0.20.2 - 0.40.4 - 0.80.8 - 1

Fig. 9. (a) Vexcel aerial photograph of semi-natural habitats near Blaenavon and fuzzy membership values mapped for (b) Calluna and (c) Vaccinium spp. with the samearea of dry heath (D1.1).

N

20 Kilometres0 5 10

Fig. 10a. Revised Phase 1 classification of semi- natural habitats, southeast Wales. The red box indicates the area illustrated in Fig. 10b (For interpretation of the referencesto colour in this figure legend, the reader is referred to the web version of this article.)

values into the Phase 1 classes. Assessing classification accuracyat these different levels was therefore complex, particularly giventhe range of habitats considered. The nature of fuzzy membershipfunctions also meant that they could not be assessed for accuracyusing a traditional error model. Such models require crisp classes

and an underlying assumption that variability in data values canbe modeled using probability. However, fuzzy membership (orpossibility measures, Dubois et al., 1993) models expert opinionon the interpretation of data values, which cannot be observeddirectly in the field. On this basis, the accuracy was collapsed


Fig. 10b. Detailed classification of Phase 1 habitats focusing on the Blanaevon World Heritage Site (For Phase 1 codes, see Table A.1).

into one evaluation, that of the derived Phase 1 map. The basisfor undertaking assessments of accuracy is typically to provide anindication of how well the classification performed. However, inthis case, the approach to accuracy assessment was also to identifywhere confusion occurred such that adjustments to the rule-base could be made and the imagery reclassified subsequently.This is one of the benefits of the mapping approach and onethat lends itself to the development of an ongoing mapping andmonitoring system for Wales. Furthermore, given the diversity ofPhase 1 habitats occurring across Wales, this paper sought onlyto demonstrate the approach to assessment, as this is an ongoingprocess, and present indicative accuracies.

To assess the accuracy of the revised Phase 1 maps, the onlyexisting maps available were the original Phase 1 Survey(1970s–1980s), the Upland Survey and the Phase 2 Survey (1990s).However, these maps were outdated and the most recent timelyreference data were the aerial photography acquired in 2001 and2006. Through interpretation of the aerial photography data, in-dicative accuracies for common upland and lowland Phase 1 habi-tats in the Cambrian Mountains (Project 8.3) and Torfaen (Project15.2) were generated based on standard confusion matrices(Congalton andGreen, 1999). For both projects, 5×5 kmareaswereselected containing 5877 and 7502 polygons respectively (basedon the revised Phase 1 classes). From each set, 100 circular poly-gons (25 m in radius) were located randomly (using tools avail-able at www.random.org) and intersected with the revised Phase1 classification. These polygons were then intersected with theoriginal Phase 1 polygons resulting in 190 polygons for the Cam-brian Mountains (15 habitat classes) and 193 for Torfaen respec-tively. Each polygon was then overlain onto the 2001 and 2006photographs and the assignment of habitats to a Phase 1 cate-gory was confirmed (or otherwise) independently and by experts

in aerial photograph interpretation. These assignments were thencompared against the revised and also original Phase 1 classifica-tion. For Project 8.2, overall accuracies for the revised Phase 1were87%, with these being least for dense scrub (A2.1) and acid dryheath (D1.1). This compared to 74% for the original Phase 1. ForProject 15.2, overall accuracies for the revised and original Phase 1were 83% and 74% respectively.

A limitation of using confusionmatrices is that they consider anobject to belong to only one class, whereas several classesmay stillbe regarded as acceptable. For example, areas observed as brackenmay be assigned directly to the Phase 1 class dense (C1.1) orscattered bracken (C1.2) or to open broadleaved woodland (A1.1)or scattered scrub (A2.2), as these can have a bracken understory.Hence, an alternative approach to assessing accuracy based on thePhase 2 Surveywas adopted. Focus on the Phase 2 Survey locationsgave the following benefits:

(a) Whilst the survey was conducted in the 1990s, most habitatsin survey areas were unlikely to have changed subsequentlyas the sites were protected and any change that might haveoccurred (e.g., increases in woodlands, scrub and bracken as aresult of neglect) could be identified through reference to 2006aerial photography.

(b) Habitats that were of high conservation interest but whichwere relatively scarceweremore likely to appear in a sample ofpoints selected by randomwhereas if selected across thewiderlandscape, these would be less well represented. Such habitatswere more common on protected sites mapped in the Phase 2survey.

The approach to assessing errors followed that of Stehman et al.(2008) who advocated the advantages of combining stratification(e.g., based on locations of Phase 2 surveys) and cluster sampling

http://www.random.org


Table 6Categories used in the comparison between the Phase 2 and revised Phase 1 classification.

Category Explanation

1 Disagree The two maps show a different habitat class2 Possible The habitat shown on the Phase 2 map might have become that shown on the revised Phase 1 map3 Plausible Habitat shown on the Phase 2 map is very similar to that on the revised Phase 1 map4 Agree The two maps show the same habitat class

Table 7Error categories for checks of polygonsin Category 1 (disagree) and 2 (possible).

Phase 2 Revised Phase 1

A Correct CorrectB Correct WrongC Wrong CorrectD Wrong WrongU Undecided

(based on groups of polygons) where estimation of class-specificaccuracy was a high priority objective. As demonstration of theapproach, the example of protected lowland habitats near Torfaen(Project 15.2) is presented. The Phase 2mapping indicated that thearea was dominated by grasslands, woodlands and hedges and in-cluded unimproved grasslands, mires and heaths.Within this area,the Phase 2 map was intersected with the revised Phase 1 out-put from the classification. This dataset was then again intersectedwith a set of 296 randomly located samples (25mdiameter circles).Intersected polygons <10 m2 in area were subsequently removedas many were very small slivers or fell outside of the boundary ofthe Phase 2 survey. On this basis, a total of 2137 polygons repre-senting an area of 51 ha were extracted, with each attributed withthe Phase 2 and new Phase 1 classification.

The assessment of accuracywas then undertaken in twophases.In the first, the Phase 2 habitat classes were compared against therevised Phase 1 class and the level of agreement and disagree-ment between the two classifications was captured in linguisticcategories (Table 6). It should be noted that the descriptions ofthe Phase 2 classification followed that of the NVC rather than theoriginal Phase 1 and hence there was not a one-to-one correspon-dence between these. However, the NVC classes were placed intothe most appropriate Phase 1 category.

Where the Phase 2 and revised Phase 1 were in agreement (i.e.,associated with classes 3 and 4), a selection of polygons was com-pared against the 2006 aerial photography to establish whethererrors in the classification of either the Phase 2 or revised Phase1 occurred. To confirm any disagreement (i.e., classes 1 and 2), allpolygons were similarly checked against the aerial photography.

From these comparisons, the accuracy of classification was as-sessed as follows. All polygons in Categories 3 and 4 were as-sumed to be correctly classified, since the probability of both beingwrong in the same way (i.e., the conditional probability of the sameclass, given the wrong class) is very small. This assumption wasconfirmed as being correct by checking a random sample of 100polygons in these categories against the 2006 aerial photography.Through reference to both the 2001 and 2006 aerial photographs,all remaining polygons in classes 1 and 2 were classified accord-ing to whether they were or might have been correctly assigned inthe original Phase 2 map, the revised Phase 1 map, in both or nei-ther. The analysis also considered that some Phase 1 classes (e.g.,B1.5 Marshy grasslands) were associated with a much wider rangeof Phase 2 classes. For this reason, confirming the allocation of thePhase 1 rather than the Phase 2 class was prioritized, particularlyas the latter was often unable to be identified from aerial photog-raphy. In a small number of cases, the correct class could not bedecided and these were rejected from the comparison. The break-down of polygons into error categories (A–U) is given in Table 7.

As the comparison was undertaken using polygons rather thanpoints, the two maps were able to disagree and yet both couldbe correct. As an example, a part of the object (polygon) mightbe classified as B1.5 marshy grassland whilst the other as B1.1improved grassland. Furthermore, a single point may lie on anindividual tree canopy but if a polygon is used, then the areaof grassland beneath and surrounding the tree can be observedand considered. This information cannot be shown conventionallyon the diagonal of an error (confusion) matrix showing singletonhabitat classes as, for this, the matrix would need to be the scalarproduct of the power set of classes. Therefore, the accuracies ofthe categories 3 and 4 (plausible or with full agreement) and Aand C (correct in the revised Phase 1), with these representing the‘on-diagonal’ classes, were summarized separately (Table 8) fromthe ‘off-diagonal’ confusions (Table 9) associated with categories1 and 2 (possible or no agreement) and B and D (incorrect inthe revised Phase 2). The comparison indicated that accuracies inthe classification of Phase 1 categories in the revised map weregenerally greater than 70%withmany exceeding 80% (broadleavedand coniferouswoodlands, improved and somemarshy grasslands,bracken water and arable). Lower classification accuracies wereassociated with scrub (39%) and semi-improved grasslands (34%),largely because of the relatively small areas observed. An overallaccuracy of 80.7% was obtained. The areas (in m2) that werecorrectly or incorrectly classified are also provided. However, theestimates of accuracy do not take into account the areas of habitatthat were not mapped or present. The breakdown of classes whereconfusions occur (Table 9) is based on the total area of incorrectlyclassified habitat rather than on individual habitats. As illustration,6.5% of all the polygons that were incorrectly classified occurredbecause of confusion between A1.1 broadleaved woodland andthe Phase 2 category M. caerulea/P. erecta mire. Similarly, 15.6%of errors arise because of confusion between semi-improvedgrasslands and category M. caerulea/P. erecta mires. However, asindicated above, the areamapped as B1.2 semi-improved grasslandis small relative to the total area. In effect, Table 9 provides abreakdown of where confusion occurs between the Phase 2 andrevised Phase 1 maps. The advantage of this approach is that suchconfusions can be identified allowing the rules to be revisited andrefinements of the mapping to be undertaken.

5. Discussion

Object-orientated methods are distinguished from object-based methods as the objects are not simply created and thenclassified, but are modified from object primitives during theclassification to create the units in the finalmap (Baatz and Schäpe,2000). Object-based rule-based classifications of habitats havebeen applied previously (e.g., reviewed in Blaschke et al., 2009).However, this study presents the first implementation of an object-orientated rule-based classification of detailed habitats (ratherthan broader land cover classes) at a national level. Rules weredeveloped by effectively capturing the knowledge and expertise offield ecologists and remote sensing scientists through the softwareinterface of eCognition (Lang and Tiede, 2007).

For Wales, a single rule-base was formulated that was appliedprogressively to 16 projects (42 sub-projects) to classify habitatsfound in the lowlands, then uplands and finally the coastal


Table8

Accu

racy

intheclassific

ationof

Phase1ha

bitats

(rev

ised

map

)based

onco

mpa

riso

nwith

thePh

ase2classific

ation,

Torfae

n,So

uthW

ales.(Notethat

nota

llPh

ase1ha

bitats

arefoun

din

area

).W

oodlan

dGrassland

andmarsh

A1.1

broa

dlea

f(sem

i-na

tural)

A1.2

conif-

erou

splan

tatio

n

A.2

scru

bA.4

felle

dB1

.1un

im-

prov

ed(acid)

B1.2

semi-

impr

oved

B2.1

neutral

B4 impr

oved

Marsh

ygrasslan

dB6 po

orsemi-

impr

oved

C1.1

bracke

nE2 flu

sh& sp

ring

G1

water

I1 rock

and

waste

land

J1.1

arab

leHed

ges

(with

tree

s)a

J3.6

urba

nJ4 ba

regrou

nd

B5.1

Molinia

B5.2

Junc

us

Correc

t(area

,m2 )

10909

026

3963

971

6228

779

792

7573

8138

470

7572

4733

1175

70

2248

736

3133

3155

021

9920

77Inco

rrec

t(area

,m2 )

1899

832

398

291

1960

028

767

4533

1101

222

7142

563

2299

1349

720

00

8543

470

0Co

rrec

t(%)

85.2

89.1

39.3

76.9

76.1

73.5

62.6

88.1

75.7

94.5

34.1

83.6

0.0

81.9

78.6

100.0

78.7

82.4

100.0

Ove

rallaccu

racy

(%)

80.71

aNoco

debe

caus

eno

tmap

pedas

aha

bitatintheoriginal

Phase1su

rvey

.


Table9

Confus

ionbe

twee

nPh

ase1(rev

ised

map

)and

Phase2classes,ex

pressedas

ape

rcen

tage

ofthetotala

reathat

was

inco

rrec

tlyclassifie

d(9840

6m

2 ).For

clarity

,value

sbe

low

0.1%

areom

itted

,value

sab

ove1%

show

nin

bold.

Woo

dlan

dGrassland

andmarsh

A1.1

broa

dlea

f(sem

i-na

tural)

A1.2

conif-

erou

splan

tatio

nA.2

scru

bA.4

felle

dB1

.1un

im-

prov

ed(acid)

B1.2

semi-

impr

oved

B2.1

neutral

B4 impr

oved

Marsh

ygrasslan

dB6 po

orsemi-

impr

oved

C1.1

bracke

nE2 flu

sh& sp

ring

G1

water

I1rock

and

waste

land

J1.1

arab

leHed

ges

(with

tree

s)a

J3.6

urba

nJ4 ba

regrou

nd

B5.1

Molinia

B5.2

Junc

us

Amen

itygrasslan

d–

––

––

––

–0.3

––

––

––

––

––

Dam

psemi-im

prov

edgrasslan

d–

––

––

––

––

––

––

––

––

––

Den

sebracke

n–

0.1

––

0.8

0.2

0.1

–0.8

–0.1

––

––

–0.1

––

Den

sescru

b/woo

dlan

d–

––

–7.7

2.3

0.2

––

0.5

––

–0.5

––

––

–

Inun

datio

nve

getatio

n–

––

––

––

0.1

––

––

––

––

––

–Scirp

uscespito

sus/Erica

tetralix

wet

heath

––

––

1.6

3.6

–0.7

––

––

––

––

––

–

Erica

tetralix/Sph

agnu

mcompa

ctum

––

––

–0.1

––

––

––

––

––

––

–

Junc

usef-

fusus/acutiflorus/G

alium

palustre

rush

pastur

e

3.7

––

–3.7

5.1

2.2

––

––

––

––

–2.0

––

M.cae

rulea/Cirsium

dissectum

fen

mea

dow

1.7

––

–0.3

3.8

––

––

–0.3

––

––

0.1

––

M.cae

rulea/Po

tentilla

erecta

mire

6.5

––

–3.5

15.6

2.6

4.0

––

–1.8

––

––

3.5

0.2

–

Carex

echina

ta/Sph

agrum

recu

rvum

/auriculatum

mire

0.4

0.1

––

––

––

––

––

––

––

0.5

––

Holcu

slan

atus/Ju

ncus

effususr

ushpa

stur

e0.3

––

––

––

––

––

––

––

–0.1

––

Cyno

surus

cristatus/Ce

ntau

rea

nigragrasslan

d

2.2

––

––

––

–0.3

––

0.6

––

––

0.1

––

Loliu

mpe

renn

e/Cy

nosurus

cristatusg

rassland

0.2

––

––

––

1.8

––

––

––

––

0.4

––

Vege

tatio

nmos

aics

3.9

––

––

0.1

–2.8

––

––

––

––

1.3

––

Other

ombrog

enou

smire

––

––

–0.6

–0.1

––

––

––

––

––

–

Other

semi-im

prov

edgrasslan

d–

––

––

––

––

––

––

––

–0.1

––

Other

wet

heath

––

––

––

–2.1

––

––

––

––

––

–Ro

ad–

––

––

––

0.2

––

––

––

––

––

–


Table9(con

tinue

d)

Woo

dlan

dGrassland

andmarsh

A1.1

broa

dlea

f(sem

i-na

tural)

A1.2

conif-

erou

splan

tatio

nA.2

scru

bA.4

felle

dB1

.1un

im-

prov

ed(acid)

B1.2

semi-

impr

oved

B2.1

neutral

B4 impr

oved

Marsh

ygrasslan

dB6 po

orsemi-

impr

oved

C1.1

bracke

nE2 flu

sh& sp

ring

G1

water

I1rock

and

waste

land

J1.1

arab

leHed

ges

(with

tree

s)a

J3.6

urba

nJ4 ba

regrou

nd

B5.1

Molinia

B5.2

Junc

us

Semi-im

prov

edgrasslan

d0.3

0.1

––

––

––

––

––

––

––

––

–

Track

––

––

0.2

––

––

––

––

––

––

––

F.ov

ina/Ag

rostis

capilla

ris/G

alium

saxa

tilegrasslan

d

0.1

––

––

––

––

––

––

––

––

––

F.ov

ina/Ag

rostis

capilla

ris/G

alium

saxa

tilegrasslan

d

0.2

––

––

––

––

––

––

––

–0.4

––

aNoco

debe

caus

eno

tmap

pedas

aha

bitatintheoriginal

Phase1su

rvey

.


regions. In total, 105 different sub-habitatswere discriminated andmapped. Aswithmore traditional supervised approaches, the rulesutilised spectral information but also considered other knowledgerelating to topography (slope, aspect) and context (adjacency,enclosure). The sequence of rules was logical and the numberand content remained the same for all projects. The rules alsosuccessfully omitted habitats that did not occur and no habitatswere mapped in entirely the wrong place (e.g., coastal heaths inupland areas). Such omissions are often not considered in standardaccuracy assessments, which only consider successfully mappedpresences but not absences and therefore progressively penaliseschemes that attempt to map a large number of classes. Slightadjustments to the values used in the rules (e.g., for the NDVI,NIR reflectance) were necessary between projects to compensatefor differences in the timing of seasonal changes in phenologybetween biogeographical regions and also agricultural practices.This adjustment allowed for the variability in natural phenologicalevents (e.g., the spring flush of vegetation communities) that occurin the same order though the timing may vary as a function ofaltitude, latitude or distance from the coast. For the classification,data from three different optical sensors acquired at differenttimes during the spring and summer months were necessarilyused. Values were therefore also amended to accommodate fordifferences between sensors and image acquisition dates. As anexample, SPOT HRG data acquired in late February (e.g., Projects3.1 and 15.2) showed significantly less topographic shadowingcompared to those acquired in late March. For some projects (e.g.,13.2), SPOT-HRG datawere only available for July and observationsfor March were necessarily provided by ASTER or IRS data. Despiteapplication to different biogeographical regions and the use ofdifferent sources and timings of imagery, the approach developedhere allowed a consistent rule-base to be maintained for allprojects.

The high level of detail obtained in the classification was at-tributable largely to the use of the 10 m SPOT HRG data (resam-pled to 5 m) for segmenting the landscape and the application offuzzy membership functions for describing complex communities(e.g., heathmosaics). The landscapewas segmented into units (e.g.,fields and associated hedgerows) thatwere recognisable and of rel-evance to a wide range of users (e.g., conservationists, land man-agers). The use of optical data acquired prior to and during thegrowing season allowed the seasonal phenology of habitats to bebetter captured, thereby increasing opportunities for their discrim-ination. Whilst a greater number of satellite acquisitions duringthe year would lead to further enhancement of the classification,those acquired prior to the spring flush (late March) and followingthe flush (July) were considered sufficient. Imagery acquired early(before mid February) or late (after mid November) were of lim-ited use, particularly in mountainous areas and primarily becauseof the low sun angle.

The integration of SPOT HRG, ASTER and IRS data increasedthe number of cloud-free observations for a region with frequentcloud cover. The use of rules to isolate areas associated withcloud also allowed a greater number of scenes to be includedin the classification. For example, the cloud-free coastal sectionscould be used even when much of the inland area was covered incloud. The use of all available imagery, including those with partialcloud cover, therefore maximized opportunities for classifyinghabitats. The different sensors acquired data on different datesbut in similar wavelength regions. These data were comparablebecause of the rigorous geometric, radiometric, atmospheric andtopographic correction procedures applied. This also facilitatedgreater confidence in the use of derived products (e.g., vegetationindices, endmember fractions) between images and different datesof observation.

In contrast to many habitat classifications, where a single mapis produced, the hierarchical approach developed here resulted in

three different levels of information; the Phase 1 habitat map, thesub-habitat map, and fuzzy membership values associated withcomplex mosaic habitats. As such, a comprehensive description ofhabitats across Wales was generated with direct application to,for example, agricultural planning, forestry-based activities andconservation. Even though a diversity of habitats were consideredand sensor data used, accuracies of classification were in excessof 70% for most Phase 1 habitats, with those associated withmore contiguous and homogeneous habitats (e.g., broadleaved andconiferous forest, improved grasslands) exceeding 85%. For twoprojects (8.3 and 15.2), the overall accuracy levels were greaterthan those obtained for the original Phase 1 Survey. A uniquefeature of the mapping is that rules and hence the classificationcan be refined subsequently (e.g., for confused classes) based onthe outcomes of the accuracy assessment.

Assessing the classification accuracy for such a diverse range ofhabitats and at a national scale is complex, especially given thecontrast between what can be observed in the field and in theremote sensing data. For example, no assessment can be madeof the correct values of fuzzy membership (apart from 0 or 1) inthe field: fuzzy values capture expert opinion over membership ofclasses with vague boundaries (Nguyen and Walker, 2006) ratherthan an observable metric such as proportion or cover. Since thefuzzy system is required to generate crisp products such as thePhase 1 map, accuracy was only assessed for the derived Phase1 habitat classes, not the sub-habitat classes nor associated fuzzymembership values. This approach to assessment was consideredparticularly appropriate given the ambiguity in the assignmentof vegetation to a Phase 1 class, even at ground level. Accuracyassessmentwas also based onobjects not points such that thematicaccuracy could be measured (Radoux et al., 2008).

Accuracy in the classification of semi-natural habitats will varybetween the 42 projects and assessment is complicated by theuse of different sensors, environmental variability within andbetween biogeographical zones (Yeo and Blackstock, 2002), thegeneration of different levels of information, and the use of fuzzymembership values. Changes in land use and cover were alsoevident in the observation period because of different agriculturalpractices (e.g., grazing, cropping) and other activities relatingto land use and management (e.g., burning, bracken control).Values for rules were adjusted to accommodate for the observedvariability in manifestation of habitats within the remotely senseddata attributable to these factors. Even so, confidence in theclassification was lower for some projects, primarily becauseof low sun angles and the associated increase in topographicshadowing (e.g., for Project 3.4, only a February SPOT HRG imagewas available), cloud or cloud-cover obscuring some of the landsurface (e.g., Project 2.1), or the necessary use of ‘patches’ ofimagery from different dates to obtain a cloud-free mosaic (e.g.,Project 14.5). The complexity of the classification has neverthelessgiven insights into the capture of variable, vague habitat classesand their translation into artificial choropleth maps at a very largescale and has been one of the achievements of the research. Fromthis experience, a robust and consistent approach to assessingthe accuracy has been developed to provide greater confidencein the outputs at a national level and facilitate refinement of themapping.

6. Summary and conclusions

Through development and implementation of a comprehensiverule-base and a logical set of processes within eCognition, amap representing 105 sub-habitats across Wales (with a nominalresolution of 5m) has been generatedwhich is being progressivelytranslated into a revised national Phase 1 map. The high level ofdetail achieved in the mapping is attributable primarily to:


Table A.1Phase 1 habitat classes (for Wales).

Woodland and scrub Heathland Open water Rock exposure and wasteA1.1.1 Semi-natural broadleaved

woodlandD1.1 Dry acid heath G1 Standing water I1.3 Limestone pavement

A1.1.2 Planted broadleavedwoodland

D1.2 Dry basic heath G2 Running water I1.4 Other rock exposure

A1.2.1 Semi-natural coniferouswoodland

D1.3 Scattered dry heath Coastland I.1.4.1 Acid/neutral rock

A1.2.2 Planted coniferous woodland D2 Wet heath H1.1 Intertidal mud/sand I1.4.2 Basic rockA1.3.1 Semi-natural mixed

woodlandD3 Lichen/bryophyte heath H1.2 Intertidal cobbles/shingle I1.5 Cave

A1.3.2 Planted mixed woodland D4 Dry heath/acid grasslandmosaic

H1.3 Intertidal rocks/boulders I2.1 Quarry

A2.1 Dense scrub D5 Wet heath/acid grasslandmosaic

H2.4 Scattered salt marsh plants I2.2 Spoil

A2.2 Scattered scrub D6 Basic dry heath/calcareousgrassland mosaic

H2.6 Salt marsh I2.3 Mine

A3.1 Scattered broadleaved trees Mire H.3.1 Mud/sand above mhw I.2.4 Refuse-tipA3.2 Scattered coniferous trees E1.6.1 Blanket bog H3.2 Shingle/gravel above mhw MiscellaneousA3.3 Scattered mixed trees E1.6.2 Raised bog H4 Rocks/boulders above mhw J1.1 ArableA4.1 Felled broadleaved woodland E1.7 Wet modified bog H6.4 Dune slack J1.2 Amenity grasslandA4.2 Felled coniferous woodland E1.8 Dry modified bog H6.5 Dune grassland J1.3 Ephemeral/short perennialA4.3 Felled mixed woodland E2 Flush and spring H6.6 Dune heath J1.4 Introduced scrubGrassland and marsh E2.1 Acid/neutral flush H.6.7 Dune scrub J.1.5 GardensB1.1 Unimproved acid grassland E2.2 Basic flush H6.8 Open dune J3.4 Caravan siteB1.2 Semi-improved acid

grasslandE2.3 Bryophyte-dominated spring H8.1 Hard cliff J3.5 Sea-wall

B2.1 Unimproved neutralgrassland

E3 Fen H8.2 Soft cliff J3.6 Buildings

B2.2 Semi-improved neutralgrassland

E3.1 Valley mire H8.4 Coastal grassland J3.7 Track

B3.1 Unimproved calcareousgrassland

E3.1.1 Modified valley mire H8.5 Coastal heath J4 Bare ground

B3.2 Semi-improved calcareousgrassland

E3.2 Basin mire H8.6 Coastal heath/coastalgrassland mosaic

NA Not accessed land

B4 Improved grassland E3.2.1 Modified basin mire Rock exposure and wasteB5 Marshy grassland E3.3 Flood-plain mire I1 Natural rock exposureB5.1 Marshy grassland

Juncus-dominatedE3.3.1 Modified flood plain mire I1.1 Inland cliff

B5.2 Marshy grasslandMolinia-dominated

E4 Bare peat I1.1.1 Acid/neutral inland cliff

Tall herb and fern F1 Swamp I1.1.2 Basic inland cliffC1.1 Bracken F1.1 Scattered swamp I1.2 ScreeC1.2 Scattered bracken F2.2 Inundation vegetation I1.2.1 Acid/neutral screeC2 Upland species rich ledges I1.2.2 Basic screeC3.1 Tall ruderal herbC3.2 Non-ruderal herb and fern

(a) The use of SPOT-5 HRG data and vector layers (e.g., LPISboundaries) as the primary sources for segmenting thelandscapes into objects, the dimensions of which are linkedto patterns within the landscape (e.g., vegetation mosaics,hedgerows and fields).

(b) Ensuring comparability between SPOT, ASTER and IRS dataacquired in different seasons/years through rigorous orthorec-tification and atmospheric and topographic correction proce-dures.

(c) The application of a logical sequence of rules, based on thresh-olds, Boolean operations and/or fuzzy membership, to classifyhabitats, with each formulated by combining knowledge of thepreference and characteristics of communities with that of theinformation content of remote sensing data.

A high level of accuracy in the classification of Phase 1 habi-tats, as translated from the sub-habitat classification, has beenachieved. Indicative overall accuracies were >80%, with these be-ing greater formore contiguous and homogeneous vegetation (e.g.,coniferous forests, improved grasslands). The maps are unique inthat they provide the detail necessary to make informed decisionsby a wide range of users including conservationists, policy makersand landholders involved in, for example, agriculture and forestryactivities. The mapping also provides a baseline against which toquantify and monitor changes associated with human activities

and processes (e.g., degradation of bogs and heaths) linked withchanges in management and climate.

Whilst many countries in Europe have generated maps of landcover from satellite sensor data, Wales is the first to produce adetailedmap of habitats. The classification has been achieved usinga single rule-base which requires only minor adjustments withthese associated with differences in image acquisition dates andtypes of satellite sensor data used. As such, the mapping approachhas application in other countries of the UK and indeedworldwide.Whilst maps have been generated for the period 2003–2006,these can be updated and refined as and when new knowledgeand/or satellite sensor data become available and released asnew versions of the same map. The mapping approach is alsoadaptable and provides capacity and impetus for the developmentand implementation of a satellite-based monitoring system forWales.

Acknowledgements

The authors would like to thank the Countryside Councilfor Wales (CCW) for supporting this work, the United KingdomSpace Agency (UKSA) and particularly Dr. Ian Thomas and IanDowney. The research could not have been undertaken withoutthe enormous support of Definiens AG, who provided software,expertise and advice. Claire Horton of the Welsh AssemblyGovernment is thanked for her contribution to the project.


Appendix. Phase 1 habitat classes (for Wales)

See Table A.1.

References

Adams, J.B., Sabol, D.E., Kapos, V., Filho, R.A., Roberts, D.A., Smith, M.O.,Gillespie, A.R., 1995. Classification of multispectral images based on fractionendmembers: application to land-cover in the Brazilian Amazon. RemoteSensing of Environment 52 (2), 137–154.

Baatz, M., Schäpe, A., 2000. Multiresolution segmentation—an optimizationapproach for high quality multi-scale image segmentation. In: Strobl, J.,et al. (Eds.), Angewandte Geographische Informationsverarbeitung XII. Beiträgezum AGIT-Symposium. Salzburg 2000. Herbert Wichmann Verlag, Karlsruhe,pp. 12–23.

Bendelow, V.C., Hartnup, R., 1980. Climate classification of England and Wales.Technical Monograph—Soil Survey of England and Wales, No. 15.

Benz, U.C., Hofmann, P., Willhauck, G., Lingenfelder, I., Heynen, M., 2003. Multi-resolution, object-orientated fuzzy analysis of remote sensing data for GIS-ready information. ISPRS Journal of Photogrammetry and Remote Sensing 58(3), 239–258.

Berry, P.M., Dawson, T.P., Harrison, P.A., Pearson, R.G., 2008. Modeling potentialimpacts of climate change on the bioclimatic envelope of species in Britain andIreland. Global Ecology and Biogeography 11 (6), 453–462.

Blackstock, T.H., Burrows, C.R., Howe, E.A., Stevens, D.P., Stevens, J.P., 2007. Habitatinventory at a regional scale: a comparison of estimates of terrestrial broadhabitat cover from stratified sample field survey and full census field surveyfor Wales, UK. Journal of Environmental Management 85 (1), 224–231.

Blaschke, T., Lang, S., Hay, G., 2009. Object-Based Image Analysis: Spatial Conceptsfor Knowledge-Driven Remote Sensing Applications. Springer Verlag.

Bunting, P., Clewley, D., 2010. The Remote Sensing and GIS Software Library(RSGISLib). In: Institute of Geography and Earth Sciences. AberystwythUniversity, UK. http://www.rsgislib.org.

Congalton, R.G., Green, K., 1999. Assessing the Accuracy of Remotely Sensed Data:Principles and Practices. Lewis Publishers.

Definiens, 2008. http://www.definiens.com/ (accessed 16.09.10).Dubois, D., Prade, H., Sandri, S., 1993. On possibility/probability transformations.

In: Lowen, R., Roubens, M. (Eds.), Fuzzy Logic: State of the Art. Kluwer AcademicPublishers, Dordrecht.

Fuller, R.M., Cox, R., Clarke, R.T., Rothery, P., Hill, R.A., Smith, G.M., Thomson, A.G.,Brown, N.J., Howard, D.C., Stott, A.P., 2005. The UK land cover map 2000:planning, construction and calibration of a remote sensing, user-orientatedmap of broad habitats. International Journal of Applied Earth Observation andGeoinformation 7 (3), 202–216.

Fuller, R.M., Groom, G.B., Jones, A.R., 1994. The land-cover map of great Britain:an automated classification of landsat thematic mapper data. PhotogrammetricEngineering and Remote Sensing 60 (5), 553–562.

Fuller, R.M., Smith, G.M., Sanderson, J.M., Hill, R.A., Thomson, A.G., 2002. The UKland cover map 2000: construction of a parcel-based vector map from satelliteimages. Cartographic Journal 39 (1), 15–25.

Grime, J.P., Hodgson, J.G., Hunt, R., 1988. Comparative Plant Ecology. A FunctionalApproach to Common British Species. Unwin Hyman, London.

Howe, L., Blackstock, T., Burrows, C., Stevens, J., 2005. The habitat survey of Wales.British Wildlife 16 (3), 153–162.

Jackson, D.L., 2000. Guidance on the interpretation of the biodiversity broad habitatclassification (terrestrial and freshwater types): definitions and the relationshipwith other classifications. Report 307. Joint Nature Conservation Committee,JNCC, Report Series.

Joint Nature Conservation Committee, 2007. Handbook for phase 1 habitat survey:a technique for environmental audit. Nature Conservancy Council.

Jones, P.S., Stevens, D.P., Blackstock, T.H., Burrows, C.R., Howe, E.A., 2003. Priorityhabitats of Wales: a technical guide. Countryside Council for Wales, Bangor.

Lang, S., Tiede, D., 2007. Definiens developer. GIS Business 9/2007, pp. 34–37.Lucas, R.M., Rowlands, A., Keyworth, S., Brown, A., Bunting, P., 2007. Mapping

habitats and agricultural land covers in Wales using multi-temporal remotesensing data. ISPRS Journal of Photogrammetry and Remote Sensing 62 (3),165–185.

Meteorological Office, 2009. Regional mapped climate averages. http://www.metoffice.gov.uk/climate/uk/averages/ (accessed 10.11.09).

Nguyen, H.T., Walker, E.A., 2006. A First Course in Fuzzy Logic, 4th ed. Chapman andHall, CRC Press, London.

Radoux, J., Defourny, P., Bogaert, P., 2008. Comparison of pixel—and object-basedsampling strategies for thematic accuracy assessment, 6711. In: Proc. Geobia2008: Pixels, Objects, Intelligence—Geographic Object-Based Image Analysis forthe 21st Century, Calgary, Canada, 6–7 August (on CD-ROM).

Richards, J.A., Jia, X., 2006. Remote Sensing Digital Image Analysis: An Introduction,4th ed. Springer-Verlag, Heidelberg, Germany.

Richter, R., 2009. Atmospheric/topographic correction for satellite imagery. ATCOR-2/3 User Guide. Version 7.0. DLRGermanAerospace Centre,Wessling, Germany.DLR-1B 565-01/09.

Richter, R., Schläpfer, D., 2002. Geo-atmospheric processing of airborne imagingspectrometry data. Part 2: atmospheric/topographic correction. InternationalJournal of Remote Sensing 23 (13), 2631–2649.

Rodwell, J.S., 2006. NVC Users’ Handbook. Joint Nature Conservation Committee.JNCC.

Smith, G.M., Wyatt, B.K., 2007. Multi-scale survey by sample-based field methodsand remote sensing: a comparison of UK experience with European environ-mental assessments. Landscape and Urban Planning 79 (2), 170–176.

Spash, C.L., Simpson, I.A., 1994. Utilitarian and rights-based alternatives forprotecting sites of special scientific interest. Journal of Agricultural Economics45 (1), 15–26.

Stehman, S.V., Wickham, J.D., Wade, T.G., Smith, J.H., 2008. Designing a multi-objective, multi-support accuracy assessment of the 2001 National LandCover Data (NLCD, 2001) of the conterminous United States. PhotogrammetricEngineering & Remote Sensing 74 (12), 1561–1571.

Tucker, C.J., 1979. Red and photographic infrared linear combinations formonitoring vegetation. Remote Sensing of Environment 8 (2), 127–150.

White, E.J., Perry, A.H., 1989. Classification of the climate of England and Walesbased on agroclimatic data. International Journal of Climatology 9 (3), 271–292.

Yeo, M.J.M., Blackstock, T.H., 2002. A vegetation analysis of the pastoral landscapesof upland Wales, UK. Journal of Vegetation Science 13 (6), 803–816.

http://www.rsgislib.org

http://www.definiens.com/

http://www.metoffice.gov.uk/climate/uk/averages/








Updating the Phase 1 habitat map of Wales, UK, using satellite sensor data

Documents

Transcript of Updating the Phase 1 habitat map of Wales, UK, using satellite sensor data