Evidential reasoning with Landsat TM, DEM and GIS data for landcover classification in support of grizzly bear habitat mapping

int. j. remote sensing, 2002, vol. 23, no. 21, 4633–4652

Evidential reasoning with Landsat TM, DEM and GIS data forlandcover classi� cation in support of grizzly bear habitat mapping

S. E. FRANKLIN

Department of Geography, University of Calgary, 2500 University Dr., Calgary,Alberta, T2N 1N4, Canada

D. R. PEDDLE

Department of Geography, University of Lethbridge, 4401 University DriveW. Lethbridge, Alberta T1K 3M4, Canada

J. A. DECHKA

GeoAnalytic Inc., 300, 700—4th Avenue S.W. Calgary, Alberta T2P 3J4,Canada

and G. B. STENHOUSE

Foothills Model Forest, Box 6330, Hinton, Alberta T7V 1X6, Canada

Abstract. Multisource data consisting of satellite imagery, topographicdescriptors derived from DEMs, and GIS inventory information have been usedwith a detailed, � eld-based landcover classi� cation scheme to support a quantitat-ive analysis of the spatial distribution and con� guration of grizzly bear (Ursusarctos horribilis) habitat within the Alberta Yellowhead Ecosystem study area.The map is needed to determine if bear movement and habitat use patterns areaVected by changing landscape conditions and human activities. We compared amultisource Evidential Reasoning (ER) classi� cation algorithm, capable of hand-ling this large and diverse data set, to a more conventional maximum likelihooddecision rule which could only use a subset of the available data. The ER classi� erprovided an acceptable level of accuracy (ranging to 85% over 21 habitat classes)for a level 3 product, compared to 71% using a maximum likelihood classi� er.

1. IntroductionGrizzly bear (Ursus arctos horribilis) habitat mapping from satellite data has long

been of interest because of the potential that highly accurate and eVective maps canbe generated from the synoptic, repetitive and consistent re� ectance data over largeareas of wilderness. Several early studies (e.g. Butter� eld and Key 1985, Craigheadet al. 1985) emphasized the need to standardize remote sensing, image processingand � eld methods, and anticipated many of the problems that would be encounteredin a transition to an operational satellite grizzly bear habitat mapping procedure,for example:

This paper was presented at the 6th Circumpolar Symposium on Remote Sensing of PolarEnvironments held in Yellowknife, Northwest Territories, Canada, from 12–14 June 2000.

Internationa l Journal of Remote SensingISSN 0143-1161 print/ISSN 1366-5901 online © 2002 Taylor & Francis Ltd

http://www.tandf.co.uk/journalsDOI: 10.1080/01431160110113971

S. E. Franklin et al.4634

(1) the diVerence between spectrally-distinct landcover classes and interpretedhabitat classes;

(2) the use of imagery together with ancillary data such as digital elevationmodels (DEMs) and biophysical land classi� cations;

(3) the diYculty in � eld veri� cation of habitat suitability and use; and(4) the need to ensure consistency in mapping across ever larger management

units (ultimately encompassing whole mountain ranges and ecophysiographicregions).

As satellite remote sensing imagery are often only one of several diVerent sourcesof information that can be accessed by grizzly bear managers, one of the mostsigni� cant questions to emerge is: How best to use all of the available data in thegrizzly bear habitat mapping task? This question, and many of the larger proceduralissues, apply equally in the development of satellite-based habitat mapping protocolsfor other wide-ranging, wilderness species in North America and elsewhere, includingelk (Huber and Casler 1990), barren-ground caribou (Hillis et al. 1996), woodlandcaribou (Hansen et al. 2001) and wolverine (Whitman et al. 1986). A core contribu-tion from remote sensing will likely be an accurate and reliable landcover classi� -cation based on multisource digital data sets. Most statistical classi� ers can useinterval and ratio level data only; yet, many active wildlife management areas arewell-covered by nominal or ordinal data such as forest inventory maps, biophysicalclassi� cations and derivative maps containing measures of habitat eVectiveness orhabitat suitability ratings. These data are diYcult to incorporate into any newclassi� cation that is based on satellite data if only statistically-based decision rulesare available. One of the few options available to the image analyst is to attempt touse these types of data to help guide, stratify or augment training data collection(Franklin et al., 2001, Hansen et al. 2001) – often incorporated into a decision-treeapproach (Hansen et al. 1996 ).

One method that has been developed to handle the multisource digital data setsin a classi� cation task is the evidential reasoning (ER) classifer (Peddle 1993, 1995a,1995b) , which is based on the Dempster-Shafer theory of evidence (Shafer 1976).One important advantage of the ER classi� er is the ability to incorporate all fourtypes of geospatial input data (nominal, ordinal, interval, ratio). This classi� er pro-vides as output not only the classi� cation map (integrating satellite imagery andGIS data together in a single, complex decision rule) but it can also provide severalinterpretive measures (e.g. measures of support, con� ict, plausibility, consensus) thatcan be used to assess the relationship between the landcover class and the interpretedhabitat class. In the task of grizzly bear habitat mapping there may be signi� cantadvantages in using the ER approach rather than a statistically-based classi� erbecause of this ability to provide ‘soft’ (information on several key classes) in additionto ‘hard’ (one per-pixel label only) classi� cation maps (Foody 1999). The ‘hard’classi� cation approach may not be optimal when the ‘soft’ continuum of habitatconditions is the desired map output; for example, the strength of the relationshipbetween that location and the various habitat classes may be of interest, in additionto the likely class membership of a given parcel of land. This classi� cation approachwas valuable in previous work where the ER algorithm has provided high classi� ca-tion accuracies in subarctic and alpine vegetation studies (Duguay and Peddle 1996),using multisource remote sensing, GIS and climatic data (Peddle and Duguay 1998)as well as for permafrost active layer depth mapping in northern Canada (Peddleand Franklin 1992, 1993).

Sixth Circumpolar Symposium on Remote Sensing of Polar Environment 4635

In this study, we applied an ER classi� er in the task of landcover classi� cationto support grizzly bear habitat mapping in Alberta. We compare the results to thoseavailable from a conventional maximum likelihood classi� cation procedure used inthe Alberta Yellowhead ecosystem of the Canadian Rocky Mountains (� gure 1). Thelargest diVerence in these two approaches lies with the type of data that can beincorporated into the decision rule, and hence, the type and amount of informationthat can be used to de� ne the landcover classes.

2. Background2.1. Yellowhead ecosystem bear management units

A portion of the Yellowhead ecosystem study area within the Foothills ModelForest (FMF) southwest of Hinton, Alberta (� gure 1) was selected for this study.This area corresponds to the 16 prede� ned bear management units (BMUs) identi� edby the Alberta Government in their planning strategy in west-central Alberta. Theapproximately 5500 km2 study area lies primarily within the Eastern Main Rangesand Front Ranges of the Rocky Mountain Thrust Belt and consists of � ve naturalecosystem subregions, namely; Alpine, Subalpine, Montane, Upper Foothills/Sub-Boreal Spruce and Lower Foothills (AchuV 1994) (table 1). In an east to westdirection, the area includes rolling hills, rising to the foothills and east slopes of theRocky Mountains and � nally rising to the Rocky Mountain front ranges(GEOWEST 1997). It has a variety of landscape features including broad valleys,rugged mountains, glaciers and alpine meadows. The region includes foothills andstrongly dissected uplands covered by thin discontinuous loamy glacial till, someorganics, and clay lacustine and sandy glacio� uvial deposits. This results in soilsthat are moderately to poorly drained (GEOWEST 1997). The topographic elevationranges from approximately 945 m to 3405 m. Drainage is dictated by the topographyand generally � ows to the northeast.

There are also diverse vegetation communities which have spatial patterns linked

Figure 1. Location of the Study Area in the Alberta Yellowhead ecosystem.


Table 1. Natural ecosystem subregions in the greater Alberta Yellowhead ecosystem (source:GEOWEST 1997) .

Ecosystem Subregion Description

Alpine Includes all areas above the treeline including vegetatedareas, rock, snow and glaciers.

Subalpine Uppermost forested zone which is steep and rugged.Dominated by confer forests.

Montane Characterized by open forests and grasslands and foundalong some major west-east river valleys.

Upper foothills Occurs on gently to strongly rolling topography along theeastern edge of the Rocky Mountains. Dominated by coniferforests.

Lower foothills Rolling topgraphy with mixed forests, absense of mixeddeciduous-coniferous forests and the presence of nearly pureconiferous forest cover.

to terrain variability and altitudinal zonation. At mid elevations, closed coniferforests are dominated by Engelmann spruce (Picea engelmanni ) and subalpine � r(Abies lasiocarpa) species. At lower elevations, Douglas � r (Pseudotsuga menziesii ),trembling aspen (Populus tremuloides) and open grasslands dominate. The easternand northern portions of the study area have mixedwood forests consisting oftrembling aspen, birch (Betula papyrifera) and white spruce (Picea glauca). Lodgepolepine (Pinus contorta) and balsam poplar (Populus balsamifera) also occur near thefoothills. The deciduous forests are primarily aspen interspersed with birch andpoplar species. A variety of shrubs and grasses occur in the area. Wildlife is diverse,with some of the larger mammals including grizzly bears, bighorn sheep (Oviscanadensis) and elk (Cervus elaphus).

Approximately 43% of the research study area falls within Jasper National Park.Historic grizzly bear data exist for portions of this area, and there are preliminarygrizzly bear habitat map products for this area generated by manual/analoguetechniques. The area selected provides an opportunity to contrast conditions withminimal human impact (i.e. Jasper National Park backcountry) against a land basewith higher levels of human activity including industrial forestry, mining, oil and gasexploration and development, and a variety of recreational activities.

2.2. Grizzly bear research in the Alberta Yellowhead ecosystemIn Alberta the grizzly bear population is estimated at approximately 850 animals

(Alberta Environmental Protection, unpublished data) in an area that is considerablyreduced compared to estimated historical grizzly bear range. The Yellowhead eco-system provides approximately 30% of the available habitat for grizzly bears inAlberta and is thought to contain at least 30% of the resident grizzly bear population.A reduction and fragmentation in bear distribution has been primarily attributed tounsustainable mortality rates combined with incremental habitat loss and habitatalienation (McLellan et al. 1999). Increased human activities and development aregenerally considered harmful to grizzly bear populations (Servheen 1990). If we areto sustain both human use activities and grizzly bears, intensive management basedon detailed biological and geographical information is required.

In 1999, 19 grizzly bears were radio-collared and tracked by global positioning


system (GPS) satellites throughout the summer and fall seasons in 16 prede� nedbear management units of the Alberta Yellowhead ecosystem. In addition to col-lecting data on movement and mortality, cub production and survivorship weremonitored. DNA � ngerprinting techniques were also used to identify individualbears in the area as well as to trace family lineages. With appropriate samplingtechniques and with adequate ‘captures’, statistical techniques have evolved to pro-vide estimates of population size within research study areas (Mysterud and Ims1999). However, quantifying bear populations and modelling bear behaviour hasproven diYcult. In the Yellowhead ecosystem, spatially-explicit models – habitat usemaps – are created as the probability of occurrence of bears on the landscape. Thesemodels are advantageous because both ecological and human use data can beincorporated into the prediction. In addition, such models provide a quantitativelink between populations and landscape conditions that can be compared to cumulat-ive eVects model (Purves and Doering 1998) results. Such comparative data, coupledwith mortality and reproductive rates as well as bear responses to human activities,can be used to assess and predict individual grizzly bear population resiliency withinthe study area (Manly et al. 1993). These models require an accurate and detailedmap of landscape conditions relevant to grizzly bear resource use.

2.3. Mapping grizzly bear habitatThe use of satellite imagery as an essential information source in the habitat

mapping task has been suggested by many scientists and managers; for example,Mace et al. (1999: 376) suggested that ‘conversion of satellite imagery to a validatedmap with vegetal and physiognomic descriptions should be the goal of habitatmanagers’. In essence the ‘validated map with vegetal and physiognomic descriptions’should be the product of (1 ) an optimal decision rule applied to (2 ) veri� ed � eldsamples and (3 ) extended with known con� dence to the imagery of a large area with(4) a logical classi� cation scheme (Lillesand 1996). Habitat class relationships aremultidimensional, scale-dependent and spatially complex. Typically, a hierarchicalapproach to habitat class de� nition is used that is consistent with land cover classesor covertypes, similar to those proposed by Anderson et al. (1976). But habitatclasses are not equivalent to land cover classes. Habitat classes are interpretationsof land conditions and often include land covertypes as a component, but could relyon predictive or assumed characteristics which accompany the various covertypes,such as understory conditions and presence/absence of certain food plants. The valueof this hierarchical approach is that classes identi� ed by the end user can bemaintained through their association with the mappable units, thereby helping toensure a reasonable level of classi� cation accuracy.

In this study, a set of hierarchical landcover classes was established by grizzlybear researchers as being representative of the landcover throughout this area, asdescribed earlier, and also functionally related to known associations with grizzlybear habitat and movement (table 2). Quantitative relations between habitat andlandcover were not established, but three levels of classes were established within ahierarchy thought to be useful in understanding bear habitat quality and use. Theclassi� cation analysis focused on the set of classes at the highest level of detail (LevelIII). At this time it is not yet known whether such classes are important to grizzlybears beyond the ‘traditional’ understanding of bear managers; presumably, withaccurate classi� cations additional insight into the importance of these classes forbears and bear management will emerge.


Table 2. Landcover classes used in grizzly bear habitat mapping of the greater AlbertaYellowhead ecosystem.

Level I Level II Level III

1. Ice/snow IceSnow

2. Shadow

3. Wetland Wetland Wetland—open bogWetland—treed bog

4. Rock Rock—alpine zone Rock—alpine zoneRock—other Rock—avalanche chutes

Rock—other

5. Water

6. Shrub/herbaceous Herbaceous Alpine/subalpine herbaceous(>1800 m)

Herbaceous reclamationHerbaceous (<1800 m)

Shrubland Shrubland (<1800 m)Alpine/subalpine—predominantly bryoid,willows, grasses (>1800 m)

7. Forest Conifer Conifer—open crown closureConifer—closed crown closure

Deciduous Deciduous—open crownclosure

Deciduous—closed crownclosure

Mixedwood Mixedwood—coniferdominated

Mixedwood—deciduousdominated

8. Disturbance Regeneration Young regeneration—burn(3–12 years)

Young regeneration—cutblock(3–12 years)

Cut Recent cut (0–2 years)Burn Recent burn (0–2 years)

9. Cultural Cultural features—urban Cultural features—urbanCultural features—gravel pits, Cultural features—gravel pits,roads, mine, etc. roads, mine, etc.

Several authors have reported successful use of remote sensing imagery in map-ping useful to grizzly bear habitat analysis using this type of hierarchical approach.In one study in the Kananaskis region of southern Alberta, adjacent to the currentstudy area, Kansas (1999) used satellite imagery, digital Alberta vegetation inventory(AVI) data, and DEMs, to map four levels of bear habitat. The integration of remotesensing and GIS data was accomplished after a conventional maximum likelihoodLandsat Thematic Mapper (TM) image classi� cation was implemented; i.e. theapproach was to use the AVI data for training area data collection, and then tostratify broad image classes using the available ancillary data, principally the DEM.At the most detailed level (Level IV), more than 500 classes were created by stratifyingbroad covertypes by three or more slope conditions and up to � ve diVerent terrain


aspects (e.g. lodgepole pine on � at surfaces, on slopes between 0–5°, 6–15°, greaterthan 15°, with terrain aspect facing N, E, S, W, NE, SE, SW, NW). Even at LevelIII of the classi� cation hierarchy, more than 100 classes of habitat were possible.The resulting maps conveyed the spatial complexity of the landscape and were useful;unfortunately, no accuracy assessment was performed on the resulting map products.Further issues with this post-classi� cation sorting approach include: (1) the initialclassi� cation may be sub-optimal since only satellite image data were used; and(2) the end-result classes were numerous, further complicating the need to interpretclasses with respect to habitat associations.

In this study, we � rst identi� ed several key aspects of landcover mapping insupport of grizzly bear habitat mapping and analysis with respect to informationrequirements and processing needs. Our goal is to maximize the eYciency andaccuracy of classi� cation over large areas; we sought to assemble the methodologicaland analytical tools to best achieve that goal in this project, and within the broadercontext of habitat mapping. Four key aspects of the landcover/grizzly bear habitatclassi� cation problem were identi� ed:

(1) The need for multisource data – previous work has shown that individualdata sets alone (e.g. remote sensing imagery, topographic data, GIS basedland inventory and bear management data) may not be suYcient to charac-terise grizzly bear habitat, but when combined can provide accurate andspatially comprehensive information of value (Kansas 1999).

(2) The need for appropriate number and � exibility in class de� nitions – classesmust be more precise instead of general, broadly de� ned landcover classes(e.g. Level I), yet not so numerous as to require additional grouping andinterpretation (e.g. the post-classi� cation strati� cation approach) . The classstructure must match the required grizzly bear habitat associations directly,with any class or data speci� c rules imbedded in the classi� cation logic; ifthis step could be quanti� ed, additional utility in map products would follow.

(3) Flexibility for digital class interpretation, integration and assessment,to provide a higher level of meaningful grizzly bear habitat information.This would include capabilities such as hierarchical class de� nitions, mixedpixel identi� cation and dynamic labeling, uncertainty measures based onconsensus, con� ict, and class separability, and more sophisticated errorassessment and analyses.

(4) The need for an improved, more robust classi� cation algorithm to satisfyrequirements (1–3). Further, the algorithm must be appropriate for use inan operational environment and so must be straight-forward to use, requirelittle or no subjective input and be capable of processing large, diverse datasets with eYciency.

The evidential reasoning algorithm was identi� ed as a candidate classi� cationtool that could meet the needs in the landcover/grizzly bear habitat mappingapplication.

2.4. T he evidential reasoning approachThe evidential reasoning approach to spatial data classi� cation has been adapted

from the mathematical theory of evidence (Shafer 1976) and is a generalization ofthe Bayesian theory of statistical inference, as reviewed in Peddle (1995a). The powerof evidential reasoning stems both from the generality of its formulation, and also


from the fact that it was designed speci� cally to provide a rich set of calculus forcombining disparate, multisource data with diVerent properties and levels of informa-tion. It has been used in a variety of applications such as medicine (Gordon andShortliVe 1985), forestry (Goldberg et al. 1985), expert systems and arti� cial intelli-gence (Shafer and Logan 1987), route selection (Garvey 1987), geology (Moon 1990,1993), classi� cation (Srinivasan and Richards 1990, Wilkinson and Megier 1990),water resources (Caselton and Luo 1992) and target detection and reconnaissancestudies (Rey et al. 1993). In the classi� cation context, evidential reasoning regardsinformation from diVerent data sources as individual pieces of evidence over a setof prede� ned or hierarchical classes. Evidence is speci� ed or derived as magnitudesof support and plausibility corresponding to the evidence in favour of each givenclass, and that which fails to refute each class, respectively (Shafer 1976). Datauncertainty is quanti� ed explicitly as the degree of remaining evidence not assignedto any class. This explicit formulation of data uncertainty is particularly importantwith large, multisource data sets, which typically possess varying levels of relevance,contradictory information, missing or unde� ned entities or error – each of whichcreates uncertainty in decision making.

After the evidence from each data source has been derived for each class, thisinformation is compiled into a mass function or evidential vector (magnitude ofsupport and plausibility for each class). The mass functions are combined to identifythe class with the overall greatest magnitude of integrated evidence with respect tothe support, plausibility and uncertainty measures. This is achieved by source-speci� corthogonal summation (Dempster 1967). For source 1 (with mass m1 over a set oflabels a) and source 2 (with mass m2 over a set of labels b), the orthogonal sum(m1Cm2 ) to determine the mass mê assigned to a labeling proposition x is computedas

m ê (x)=[1 Sm1 (ai ) m2 (bj )]Õ 1 Sm1 (ai ) m2 (bj )

aim bj=w aim bj=x (1)

From this formulation, the extent of con� ict between the two sources can alsobe computed. This process is repeated sequentially for each source in the data set,after which all mass functions have been reduced to a singular evidential vector forwhich a decision rule can be invoked to determine a � nal pixel label classi� cation.

In this study, we used the Multisource Evidential Reasoning Classi� cation(MERCURYC) software system (Peddle 1995a). MERCURYC has been imple-mented as a non-parametric , supervised multisource data classi� er, with a keycomponent being its frequency-based approach to knowledge formulation of evidencefrom multisource training data (Peddle 1995b) . Unlike algorithms such as maximumlikelihood (ML), the evidential reasoning approach does not use measures of centraltendency, it is independent of the data scale of measurement, and is free of anyrequirements by statistical models or expected data distributions. This is particularlyimportant since it allows GIS data such as nominal-type forest inventory data ordirectional information such as terrain aspect or wind vectors to be integrated andclassi� ed together with ratio-type remote sensing imagery, regardless of the distribu-tion of these data (e.g. non-normal data, bi-modal or multi-modal distributions) . TheMERCURYC classi� er is particularly well suited for processing a larger number ofdata dimensions, which often characterize multisource GIS and remote sensingdata sets. In contrast, algorithms such as ML sometimes suVer from the ‘Hughes


phenomena’ (Kim and Swain 1990) in which classi� cation accuracy decreases asadditional data layers are introduced, even though these additional data layersprovide new information that should aid the analysis. This phenomena can beattributed to a sensitivity to greater levels of multi-dimensionality (e.g. hyperspectralairborne imagery, hypertemporal image sequences, large GIS inventory data), ancil-lary data channels which violate statistical assumptions of parametric classi� ers, orboth. The MERCURYC software system is compatible with training data generatedindependently or from existing image analysis systems, and can process full imageformat data as well as attribute table � les (e.g. data generated from statisticalpackages or database spreadsheets) . Its ability to explicitly deal with missing orunde� ned data was particularly important in this study in which many of the GISdata variables were associated with speci� c management units and therefore did notcover the entire study region.

In terms of performance evaluations, the MERCURYC classi� er has been com-pared with ML, linear discriminant analysis, and neural network approaches, whereit has been shown to provide increased classi� cation accuracy, faster run-times, andgreater ease of use owing to the small number of user inputs required (Peddle 1993,Duguay and Peddle 1996, 1998). In addition to advantages over other algorithms,MERCURYC represents a signi� cant improvement over previous implementationsof evidential reasoning, which were subjective, parametric, or required data speci� crules (Goldberg et al. 1985, Lee et al. 1987, Moon 1990, Wilkinson and Megier1990). It has recently been optimized (Peddle and Ferguson 2002), and used in adetailed forestry study (Hall et al. 2000) as well as being coupled with radiativetransfer models to obtain terrestrial biophysical information (Peddle et al. 2001 ),where it was shown to provide improved results compared to vegetation indexapproaches, while also being better suited for broad-scale, multi-sensor studies thatwill characterise the next generation of remote sensing (e.g. NASA Terra Satellite).With this background, the present study represents the � rst use of evidentialreasoning classi� cation for a detailed habitat mapping application involving ahierarchical complex of landcover classes.

3. Data collection3.1. Image and GIS data

A Landsat TM image was acquired on 29 August 1998. An earlier image, acquired30 August 1996, was used to perform a change detection analysis to identify recentclearcuts and burns. Little radiometric distortion was apparent in these images whichwere acquired during dry, cloud-free conditions. The images were corrected forpossible atmospheric eVects with reference to pseudo-invariant re� ectance targetswith known spectral properties (Richter 1990). Following radiometric correction theimagery were processed for texture measures using the spatial co-occurrenceapproach; homogeneity and entropy statistics were derived over small (5×5) pixelwindows.

The GIS data consisted of orthophotography (used to identify locations andapproximate class identity for each sample point described below) and where avail-able, contiguous DEM, Alberta vegetation inventory (AVI), and Jasper NationalPark biophysical inventory coverages. The DEM was spatially-interpolated to a25 m grid cell from contours originally produced from 1:20 000 black and whitestereopairs of metric aerial photography. AVI data are a fourth-generation vegetationcovertype product derived from polygons of homogeneous vegetation attributes


interpreted on aerial photographs and veri� ed with permanent � eld sample plots. Inforested areas the AVI maps are similar to standard forest inventory maps andcontain forest stand information such as species composition, height, age, tree dia-meter and density classes. In other vegetated areas, such as wetlands, the classesdisplayed on the map are based on species, hydrology, soils and vegetation cover.The Jasper biophysical inventory is an ecological land classi� cation product gener-ated in the 1970s through standard biophysical mapping methods in Canada’sNational Parks (Lacate 1969). The map displays integrated ecological units thatwere formed through aerial photointerpretation of homogeneous topographic,edaphic and vegetative conditions and veri� ed through a � eld sampling programmecomprised of more than 1000 plot locations.

3.2. Field dataAn interdisciplinary team of remote sensing scientists, foresters, wildlife biologists

and botanists collected various � eld data samples in support of this study and otherenvironmental impact studies in the area. These data consisted primarily of � eldobservations using standard forest inventory or ecosystematic classi� cation � eldprotocols (Kansas 1999, Strong 1999) in random and purposive sampling strategies(Justice and Townshend 1981). In addition to vegetation types, each site was locatedprecisely using GPS or a reliable map/orthophotograph/TM coordinate for futurereference. We acquired 700 data points distributed across the habitat mapping classesand used these data in two ways:

training data for the supervised class� ers;test data used for classi� cation veri� cation and accuracy assessment.

The points were used as a guide to locate sample sites after they were systematic-ally evaluated in relation to their suitability for classi� er training sites and accuracyassessment test sites. Randomly identi� ed sites that were not homogenous or whichstraddled two or more covertypes were either eliminated or shifted manually to anarea that had a unique, homogeneous covertype. For most classes, a sample of anarea approximately 4 ha surrounding the site was extracted; occassionally, for smallclasses, the site was sampled as an individual pixel. Every attempt was made to selecttraining areas from spectrally homogenous sites with a high degree of geometriccon� dence. The identi� ed sites were compared with known � eld samples, AVIdata and orthophotography to determine the con� dence that a site belongs to aparticular class.

4. MethodsThe ML classi� er (ER MAPPER 1995) and the MERCURYC ER classi� er were

used to classify each pixel in the test data sets, using the same training data as input.Four runs of the ML classi� er are reported: (1) TM data alone; (2) TM and DEMdata, where the DEM data consist of slope and aspect variables; (3 ) TM and TMtexture (TEX) variables, where the texture variables consist of entropy and homogen-eity derived by spatial co-occurrence; and (4) TM, DEM and TEX variables. Forthe ER classi� cation, we duplicated these four runs, and added two additionalclassi� cations: (1) the AVI and biophysical information; and (2) the full data set,consisting of TM, DEM, TEX and AVI variables. The training and test data consistedof over 15 000 pixel sites, divided randomly with two-thirds allocated to training


data, one-third to test data. The same pixel training and independent test data wereused for each classi� cation run.

Accuracy was computed as the percentage of correctly classi� ed pixels, bothoverall and in each class for the independent test data. Overall kappa coeYcientswere also computed for contingency matrices.

5. Results and analysis5.1. Maximum likelihood classi� cation

The results of the four maximum likelihood classi� cations are contained intable 3. The use of Landsat TM data alone in the decision rule provided 65.27%accuracy and a kappa coeYcient of 0.626. This was improved to 70.62% (kappacoeYcient=0.684) when the DEM data were added to the decision rule. This levelof accuracy and this trend is similar to those reported in earlier studies in whichsatellite data were augmented with DEM data in complex, mountainous terrain inAlberta (e.g. Franklin et al. 1994) and elsewhere (e.g. Bolstad and Lillesand 1992).The use of texture did not increase the accuracy overall, and often decreased accuracyof individual classes, although at least two classes (recent and older cuts) improvedsubstantially. The use of TM, DEM and texture variables together resulted in asmall (not statistically signi� cant) increase in accuracy (to 71.18% and kappa coeY-cient of 0.69). The lowest class accuracies were consistently reported in the mixedconifer, mixed deciduous, open deciduous and open conifer forest classes. The shrubclass, limited by elevation, was also poorly classi� ed in all four attempts using themaximum likelihood decision rule. The AVI data consisted of nominal level variables,and therefore could not be processed using ML since: (1) measures of central

Table 3. Accuracy results of ML classi� cation using four diVerent sets of discriminatingvariables in test sites (5013 pixels).

Class Description Landsat TM TM,DEM TM,TEX All

1 Ice 98.3 98.7 99.5 98.32 Snow 99.5 99.0 100.0 99.53 Shadow 100.0 100.0 95.5 100.04 Wetland—open bog 57.9 56.9 45.9 57.95 Wetland—treed bog 50.3 51.2 41.0 50.36 Rock 100.0 98.9 99.5 100.07 Water 99.6 99.3 97.7 99.68 Herbaceous reclamation 98.3 90.6 95.0 98.39 Herbaceous <1800 m 35.7 32.1 33.0 35.7

10 Shrubland <1800 m 37.6 34.5 28.1 37.611 Alpine/subalpine—grasses 81.6 81.8 53.1 81.612 Open conifer 37.7 40.8 28.3 37.713 Closed conifer 78.0 74.8 78.1 78.014 Open deciduous 20.5 17.9 11.2 20.515 Closed deciduous 41.1 44.3 30.9 41.116 Mixed conifer 12.5 14.6 14.7 12.517 Mixed deciduous 20.8 18.6 14.1 20.818 Young regeneration—cut (3–12 years) 13.6 30.7 43.7 13.619 Recent cut (0–2 years) 63.1 60.9 74.8 63.120 Recent burn (0–2 years) 94.9 96.6 77.4 94.921 Cultural features—mine 75.2 76.1 55.8 75.2

Overall (%) 65.2 70.6 65.3 71.1


tendency are not appropriate for use with nominal level data; and (2) these data didnot conform to a normal (Gaussian) distribution, and therefore violated a funda-mental statistical requirement of the ML algorithm. However, these nominal level,non-normal data were suitable for processing by the evidential reasoning algorithm,for which results are described next.

5.2. Evidential reasoning classi� cationUsing evidential reasoning, it was possible to include all available 37 variables

from multi-date satellite remote sensing (Landsat TM) imagery and image texture(TEX) variables, DEM variables such as slope, aspect and other derived terrainvariables, and the set of GIS variables from the AVI and other site descriptors, suchas soil conditions.

Six diVerent sets of variables were classi� ed using the MERCURYC evidentialreasoning software to assess the viability of diVerent digital data sources for grizzlybear habitat classi� cation, with results summarized in table 4. The best overall resultobtained was 85.9%, using all 37 variables from the combined, multisource data set(TM, texture, DEM, AVI) (table 5). Use of individual data sets alone yielded substan-tially lower accuracies (45–50%), although we note that less attention was given tooptimising ER results for these data subsets. The highest accuracy achieved withoutuse of the GIS AVI data was 55.9%, using the TM, texture and DEM data. Althoughthese overall accuracies were similar, the pattern of individual class accuracies wasrather diVerent. For example, the satellite imagery performed reasonably well fordisturbance classes and non-vegetated classes, however, there was considerable confu-sion amongst forest density classes and also amongst herbaceous, shrub and grass-lands, with image texture providing small improvements in certain classes. Thetopographic data set (DEM) showed moderate results in some of the non-vegetatedclasses, however, topographic information alone was insuYcient to discriminateforest density classes and lower vegetation classes. The AVI data alone performedwell for forest density and wetland classes, however, low accuracies were found withherbaceous, shrub and grassland classes. One possible explanation for this is the factthat AVI is primarily a forest inventory product and the level of detail and theaccuracy of some of the other vegetation polygons on the map might be questionable.A number of classes also had no associated AVI data (e.g. classes 1, 2: ice, snow)while other classes had small sample sizes.

However, when all available data sets were combined, the individual and overallaccuracies improved substantially (table 5), with an overall accuracy of 85%achieved. This means that the information content of individual data sets is notcompletely redundant, and instead provides unique and new information for classi-� cation that is not contained in other data sets and which is captured more compre-hensively using the evidential reasoning approach. This is important, since it providesa fundamental rationale for assembling and applying multiple data sets to a complexclassi� cation task.

6. Discussion6.1. Classi� er performance

The main objective of this work was to determine the diVerences in landcovermap and classi� cation accuracy that could be attributed to the use of a classi� erthat can handle large, diverse data sets such as those contained in the AVI andJasper National Park biophysical inventories in the grizzly bear habitat mapping


Tab

le4.

Ind

ivid

ual

and

over

all

class

accu

raci

es(%

)fr

om

evid

enti

alcl

ass

i�ca

tio

nof

diV

eren

tvar

iab

lese

ts.

Cla

ssD

escr

ipti

on

Land

sat

TM

TM

,DE

MT

M,T

exT

M,D

EM

,T

exA

VI

All

1Ic

e50.2

66.

571

.381.

3–

83.

32

Sno

w87.2

87.

686

.486.

8–

86.

43

Shado

w92.5

92.

592

.592.

5–

93.

74

Wet

lan

d—

op

enbo

g61.5

78.

654

.978.

286.

290.

95

Wet

lan

d—

tree

dbo

g24.

943.

328

.047.

175.

886.

26

Rock

89.9

92.

488

.489.

2–

91.

77

Wate

r64.5

78.

064

.379.

290.

894.

88

Her

bace

ous

recl

amati

on

59.2

51.

063

.359.

20.

063.

39

Her

bace

ous

<180

0m

4.1

2.4

3.3

0.8

40.

769.

910

Shru

bla

nd

<18

00

m18.0

20.

428

.630.

152.

987.

411

Alp

ine/

subalp

ine—

gra

sses

1.3

1.3

3.8

3.8

27.

944.

312

Open

con

ifer

0.3

0.3

0.3

0.3

54.

169.

513

Clo

sed

conif

er53.8

49.

751

.048.

525.

886.

614

Open

dec

idu

ou

s4.

29.

24.9

9.9

85.

292.

315

Clo

sed

dec

iduo

us

49.4

40.

647

.243.

378.

984.

416

Mix

edco

nifer

7.4

21.

214.

325.

861.

387.

617

Mix

edd

ecid

uo

us

21.7

31.

320

.931.

389.

681.

718

You

ng

rege

ner

atio

n—

cut

(3–12

yea

rs)

67.3

78.

566

.476.

687.

991.

619

Rec

ent

cut

(0–2

year

s)75

.676.

980

.080.

662.

580.

620

Rec

ent

burn

(0–2

yea

rs)

45.2

43.

749

.649.

6100.

010

0.0

21

Cult

ura

lfe

atu

res—

min

e70.5

73.

877

.177.

16.

678.

7

Ove

rall

(%)

49.

354.

150.

655.

947.

885.

9


Table

5.C

on

tingen

cym

atr

ixw

ith

ind

ivid

ual

and

ove

rall

clas

sacc

ura

cies

(per

cen

tage

and

kap

pa

co-e

Yic

ent)

fro

mev

iden

tial

reaso

nin

gcl

ass

i�ca

tion

of

all

vari

able

s(c

lass

eslist

edin

tab

le4

).

Cla

ssi�

cati

on

Label

Cla

ss1

23

45

67

89

10

1112

13

14

1516

17

18

1920

21

N%

Kc

1174

014

11

68

10

00

01

00

01

20

00

209

83.

25

0.83

214

229

21

05

14

12

00

10

10

21

00

126

586.

42

0.86

30

023

61

11

00

00

00

60

40

20

00

125

293.

65

0.93

40

00

250

190

01

30

01

00

00

01

00

027

590.

91

0.90

50

00

23

249

00

00

13

38

00

11

00

00

289

86.

16

0.85

65

00

10

364

31

20

00

00

00

013

00

839

791.

69

0.91

70

03

22

040

10

04

02

80

00

10

00

042

394.

80

0.94

80

00

90

00

310

30

00

00

00

60

00

49

63.

27

0.63

90

00

34

74

186

10

50

00

10

11

00

012

369.

92

0.69

100

00

00

03

02

180

50

015

00

01

00

020

687.

38

0.87

110

02

40

60

110

15

352

01

00

03

00

079

44.

30

0.43

120

01

96

05

05

325

221

114

12

141

10

00

318

69.

50

0.68

131

024

01

010

01

00

2504

126

57

00

00

582

86.

60

0.85

140

00

00

00

00

00

00

131

12

80

00

014

292.

25

0.92

150

00

60

04

00

00

01

015

23

140

00

018

084.

44

0.84

160

00

00

00

00

10

36

06

190

110

00

021

787.

56

0.87

170

00

00

00

00

00

12

68

394

01

00

115

81.

74

0.81

180

00

60

00

12

00

00

00

00

98

00

010

791.

59

0.91

190

00

00

00

01

00

00

00

00

21

129

90

160

80.

62

0.80

200

00

00

00

00

00

00

00

00

00

135

013

510

0.0

1.0

210

00

10

60

01

10

00

00

00

30

148

61

78.

69

0.78

N194

229

282

317

283

395

439

41114

211

68240

548

158

210

219

142

161

130

145

58

458

485.

90.

85

Note

:O

vera

ll%

corr

ect:

85.

89%

;O

ver

all

acc

ura

cy:±

1.03%

;(8

4.86%

to86.9

1%

at95

%co

n�d

ence

level

);O

ver

all

kapp

aco

eYci

ent:

0.85.


application. In comparing landcover classi� cation results from the ML and ERclassi� ers, two main trends are evident:

(1) when dealing with lower dimensional data sets which adhere to the statisticalrequirements of ML, the ML classi� er provided higher accuracies than ER,with results ranging to 71%; and

(2) The ability of ER to integrate diverse multisource data sets, including thosewhich cannot be processed using ML, represents an important capability asevidenced by the highest overall accuracy of 85% obtained.

The GIS AVI data clearly provided new information to the classi� cation process,however, these data by themselves still had the lowest ER accuracy (47%) relativeto other ER results. This is important, since it demonstrates the importance ofproviding an algorithm capable of integrating data from diVerent data sets. Individualdata sources may provide unacceptable levels of accuracy, but when combined andprocessed by an appropriate classi� cation algorithm, provide levels of accuracy thatare both signi� cantly higher relative to the individual data sets, and also to a levelacceptable for vegetation classi� cation, habitat mapping, and ecological analyses.The key is using an algorithm with suYcient generality which can handle the diverseproperties which characterise multisource data sets.

Overall the ER classi� er outperformed the ML classi� er when the ancillary datawere employed in the decision. The highest accuracy possible using the ML algorithmwas 71.18%, partly because the ML algorithm was restricted to only a subset of theavailable input variables; with none of the existing map variables included, exceptin the form of information of possible use to the analyst collecting training data.The ER classi� er appeared to optimize the available information, providing thehighest overall classi� cation accuracy of 85.9%. The remaining errors in the classi-� cation appeared in some instances to be a result of the inaccuracies contained inthe ancillary data and GIS maps which the ER classi� er employed; for example, thealpine/subalpine grass class remained poorly classi� ed since these areas do nottypically receive a great deal of attention in the compilation of forest inventory maps.The Jasper biophysical inventory data similarly was not detailed; polygons in theseareas were typically undersampled and poorly understood; hence, the continuedmisclassi� cation of these areas is perhaps not surprising.

The main conclusion here is that the algorithms operate optimally in situationsfor which they were designed. ML was designed for processing lower dimensionaldata sets which conform to certain statistical properties (ratio-type data, normally-distributed, equal class variances, as discussed earlier). When these properties weremet in the data set we used, ML was a powerful classi� er which extracted theavailable information for classi� cation. However, ML is not intended or well-designed for processing higher dimensional, statistically diverse data which mayviolate the required input properties. A new problem arises, however, when thecomplexity of the environmental application or the level of detail within a classstructure exceeds that which can be provided by individual data sets using the MLapproach. This is the case in this study. Here, although ML provided higher accurac-ies than ER when dealing with TM and DEM data, this level of accuracy (71%overall, with more than one-third of individual landcover class accuracies <50%)was still unacceptably low for the grizzly bear habitat application. For this type ofclassi� cation involving complex habitat-landcover associations, the existing GIS mapdata were important. Yet these data could not be processed using ML. ER, on the


other hand, was designed speci� cally to provide � exibility of input data types anddiverse statistical data properties.

The diVerence between the two classi� cation algorithms was examined spatiallyfor a representative sample areas (not shown here). A sense of the diVerent classi� ca-tion accuracies was obtained by visually assessing these products; what was quiteinteresting was the diVerent spatial con� gurations of the classes that resulted fromthe use of the two diVerent algorithms. Both ML and ER maps appeared to providea reasonable generalization of the terrain useful in portraying and interpreting broadpatterns of grizzly bear habitat.

6.2. Future researchBased on these results obtained using evidential reasoning, we have identi� ed a

number of areas for further investigation which draw more fully on the power ofevidential reasoning for information provision beyond the ‘hard’ classi� cation labelsderived in this � rst analysis:

(1) Evidential reasoning has an explicit mechanism for handling and quantifyinginformation uncertainty, both on input and as part of the output stream. Forexample, it is possible to output measures of con� ict for a given label,indicating strength of con� dence in a given assignment, as well as providingsecondary or tertiary class labels. This latter point can be particularly import-ant in more re� ned class structures, such as the grizzly bear habitat classesused in this study. For example, as class precision is increased, the probabilityof encountering ‘mixed pixels’ with respect to those classes increases substan-tially. At the spatial resolution of Landsat TM and other, more coarseresolution sensors suitable for regional scale studies, the occurrence of mixedpixels is also greater. Therefore, consideration of mixed pixels is importantboth from a class structure and from a sensor selection perspective. Theseissues could be particularly important in quantifying habitat fragmentationmeasures as a function of spatial variability in multisource data.

(2) One approach to dealing with mixed pixels explicitly is to use spectral mixtureanalysis in which the fraction of occurrence of scene elements is quanti� edat sub-pixel scales (e.g. fraction of tree canopy, shadow, ground vegetation –see Peddle et al. 1999). However, in the evidential reasoning classi� cationframework, another approach would be to use internal evidential measuresdirectly to establish evidence thresholds, above which qualifying classes wouldbe � agged for membership in a mixed pixel. This would provide an establishedand quantitative rank-order as well as associated measures of data and classspeci� c evidential con� ict for identifying class components at sub-pixel scales.This would be an important improvement with respect to current accuracyassessments and error reporting involving ‘hard’ classi� cation products.Indeed, in the case of mixed pixels, this approach represents a more realisticand therefore accurate representation of the natural landscape. The abilityof evidential reasoning to handle a classi� cation hierarchy could also be morefully explored, particularly given the existing three-level hierarchy of classesalready in place for this study. With appropriate validation information, it isquite possible that this approach could be shown to provide a higher levelof individual and overall class accuracy compared to the restrictive natureof ‘hard’ class labels, in which a given class may have only a marginally


higher occurrence than another class, yet it is � agged as being dominant withthe implicit assumption of homogeneity over the entire pixel area.

The implications of this approach could be quite important to the analysis ofresults in this study, and to follow-on grizzly bear habitat studies in the AlbertaYellowhead ecosystem. For example, in a more complex class structure such as inthis study (e.g. nine forest classes, four low vegetation classes), there will be diVerentlevels of error severity which could be formalized as an error severity index (ESI)and analysed with respect to landcover associations identi� ed for grizzly bear habitat.An ESI could be interpreted and used in several ways. For example, from an imageprocessing classi� cation standpoint, confusion between classes such as wetlands invalley bottoms and high elevation alpine snow may be considered more severe thanconfusion amongst ‘open conifer’ and ‘closed conifer’ classes, since in the latter casethis may represent spatial association within a mixed pixel. From a habitat mappingperspective, the ability to tolerate error amongst two levels of forest density de� ningopen and closed conifer classes may be somewhat diminished (i.e. higher ESI), ifthese classes characterized diVerent habitat associations.

Regardless of interpretation, however, the ability to provide an explicit mixed-class capability through evidential thresholds in a ‘soft’ classi� cation frameworkcould represent a more appropriate representation of land cover and habitat classes,with error assessments having increased meaning and providing a greater level ofimage understanding. This may also lead to a better way of dealing with ‘mixedforest’ classes, in which arbitrary (i.e. ‘hard’) mixed classes are not speci� ed a prioribut instead would be de� ned as more precise combinations of particular classes.This process could be constrained to only permit class combinations which mayoccur in nature, and further, could be extended to provide a posteriori class groupingsin which frequently occurring class combinations are allocated to a newly de� nedclass. This also gives rise to the notion of partial correctness, in which one or more,but not all of the classes identi� ed as exceeding the evidential threshold are inagreement with � eld or ground validation data or observations. In this case, it wouldbe possible and indeed appropriate to compute a more re� ned estimate of classi� ca-tion ‘accuracy’. Issues dealing with classi� cation accuracy versus agreement withrespect to the quality of ground data used in the validation of remotely sensed mapproducts would be particularly important in these cases.

7. ConclusionValidated grizzly bear habitat maps for the Alberta Yellowhead ecosystem study

area will serve as important new management tools for land managers facing issuesrelated to grizzly bear conservation and sustainable development. We have demon-strated that a complex set of landcover classes, currently qualitatively related togrizzly bear habitat classes, can be classi� ed with an evidential reasoning classi� cationapproach. An acceptable level of accuracy was achieved, ranging to 85% overall fora Level III classi� cation map product. This was an improvement over the resultsthat could be achieved using a maximum likelihood classi� er (71%), principallybecause the evidential reasoning approach is more powerful in the way existingpolygonal ancillary data, such as forest inventory data, categorical biophysicalvegetation maps, and DEM data, can be used in the classi� cation process. Theevidential reasoning approach is also suitable for more re� ned classi� cation outputssuch as fuzzy or soft labels, class speci� c indicators of con� dence, uncertainty and


information con� ict, as well as providing image processing mechanisms for dealingwith mixed pixels at various scales, all of which are thought to be important indeveloping methods of landcover and wildlife habitat mapping in circumpolarenvironments and elsewhere.

AcknowledgmentsThis work was made possible through the ongoing work and support of the

Northern East Slopes Environmental Resources Committee. We acknowledge � nan-cial support through the Remote Sensing Data Development Program of the CanadaCentre for Remote Sensing (through Research Contract #23413-9-D220-01/SQ toJ. A. Dechka, GeoAnalytic Inc.), the Natural Sciences and Engineering ResearchCouncil of Canada (through Research Grants to Dr Peddle and Dr Franklin), andthe Foothills Model Forest (through the Grizzly Bear Research Program under thesupervision of G. B. Stenhouse) .

ReferencesAchuff, P. L., 1994, Natural Regions, Subregions and Natural History T hemes of Alberta; A

Classi� cation for Protected Areas Management (Edmonton: Alberta EnvironmentalProtection).

Anderson, J. R., Hardy, E. E., Roach, J. T., and Witmer, R. E., 1976, A Land Use andLand Cover Classi� cation System for Use with Remote Sensor Data. US GeologicalSurvey Professional Paper 964 (Washington), 28 pp.

Bolstad, P. V., and Lillesand, T. M., 1992, Improved classi� cation of forest vegetation innorthern Wisconsin through a rule-based combination of soils, terrain, and LandsatThematic Mapper data. Forest Science, 38, 5–20.

Butterfield, B. A., and Key, C. W., 1985, Mapping grizzly bear habitat in Glacier NationalPark using a strati� ed Landsat classi� cation: a pilot study. Proceedings, Grizzly BearHabitat Symposium (Ogden, Utah: USDA Forest Service Intermountain ResearchStation), pp. 58–66.

Caselton, W. F., and Luo, W., 1992, Decision making with imprecise probabilities—Dempster–Shafer theory and application. Water Resources Research, 28, 3071–3083.

Craighead, J. J., Craighead, F. L., and Craighead, D. J., 1985, Using satellites to evaluateecosystems as grizzly bear habitat. Proceedings, Grizzly Bear Habitat Symposium(Ogden, Utah: USDA Forest Service Intermountain Research Station), pp. 101–112.

Dempster, A. P., 1967, Upper and lower probabilities induced by a multivalued Mapping.Annals of Mathematical Statistics, 38, 325–339.

Duguay, C. R., and Peddle, D. R., 1996, Comparison of evidential reasoning and neuralnetwork approaches in a multisource classi� cation of alpine tundra vegetation.Canadian Journal of Remote Sensing, 22, 433–440.

ER MAPPER, 1995. ER MAPPER Reference Manual 5.0 (San Diego: Earth ResourceMapping Pty. Ltd).

Foody, G., 1999, The continuum of classi� cation fuzziness in thematic mapping.Photogrammetric Engineering and Remote Sensing, 65, 443–451.

Franklin, S. E., Stenhouse, G. B., Hansen, M. J., Popplewell, C. C., Dechka, J. A., andPeddle, D. R., 2001, An Integrated Decision Tree Approach (IDTA) to mappinglandcover using satellite remote sensing in support of grizzly bear habitat analysis inthe Alberta Yellowhead Ecosystem. Canadian Journal of Remote Sensing, 27, 579–592.

Garvey, T. D., Lowrance, J. D., and Fischler, M. A., 1981, An inference technique forintegrating knowledge from disparate sources. Proceedings, Seventh InternationalConference on Arti� cial Intelligence, Vancouver, Canada, 319–325.

Garvey, T. D., 1987, Evidential reasoning for geographic evaluation for helicopter routeplanning. IEEE T ransactions on Geoscience and Remote Sensing, 25, 294–304.

GEOWEST, 1997, Common Ecological L and Classi� cation and Associated Attribute Databasefor the Yellowhead Ecosystem (Edmonton, Alberta: GEOWEST EnvironmentalConsultants Ltd).


Goldberg, M., Goodenough, D. G., Alvo, M., and Karam, G., 1985, A hierarchical expertsystem for updating forestry maps with Landsat data. Proceedings of the IEEE, 73,1054–1063.

Gordon, J., and Shortliffe, E. H., 1985, A method for managing evidential reasoning in ahierarchical hypothesis space. Arti� cial Intelligence, 26, 323–357.

Hall, R. J., Peddle, D. R., and Klita, D. L., 2000, Mapping conifer understory within borealmixedwoods from Landsat TM satellite imagery and forest inventory information.T he Forestry Chronicle, 76, 75–90.

Hansen, M., Dubayah, R., and DeFries, R., 1996, Classi� cation trees: an alternative totraditional land cover classi� ers. International Journal of Remote Sensing, 17,1075–1081.

Hansen, M. J., Franklin, S. E., Woudsma, C., and Peterson, M., 2001, Caribou habitatmapping and landscape structure analysis in the North Columbia Mountains, BritishColumbia, Remote Sensing of Environment, 77, 50–65.

Hillis, T. L., Mallory, F. F., Dalton, W. J., and Smiegielski, A. J., 1996, Preliminaryanalysis of habitat utilization by woodland caribou in northwestern Ontario usingsatellite telemetry. Rangifer, 10, 1998.

Huber, T. P., and Casler, K. E., 1990, Initial analysis of Landsat TM data for elk habitatmonitoring. International Journal of Remote Sensing, 11, 907–812.

Justice C. O., and Townshend, J. R. G., 1981. Integrating ground data with remote sensing.In T errain Analysis and Remote Sensing, edited by J. R. G. Townshend (Boston: GeorgeAllen & Unwin), pp. 38–58.

Kansas, J., 1999, EVects of mapping scale, disturbance coeYcients and seasonal habitat ongrizzly bear habitat eVectiveness modeling in Kananaskis Country, Alberta. MScthesis, University of Calgary.

Kim, H. and Swain, P. H., 1990, A method for classi� cation of multisource data using interval-valued probabilities and its applications to Hiris data. Proceedings of a Workshop onMultisource Data Integration in Remote Sensing. NASA Conference Publication 3099(Washington), 75–81.

Lacate, D. S., 1969, Guidelines for Biophysical L and Classi� cation (Ottawa: Canadian ForestService).

Lee, T., Richards, J. A. and Swain, P. H., 1987, Probabilistic and evidential approaches formulti-source data analysis. IEEE T ransactions on Geoscience and Remote Sensing,25, 283–292.

Lillesand, T. M., 1996, A protocol for satellite-based land cover classi� cation in the upperMidwest. 103–118. In Gap Analysis: A L andscape Approach to Biodiversity Planning,edited by J. M. Scott, T. H. Tear and F. W. Davis (Bethesda, MA: American Societyfor Photogrammetry and Remote Sensing).

Mace, R. D., Waller, J. S., Manley, T. L., Ake, K., and Wittinger, W. T., 1999, Landscapeevaluation of grizzly bear habitat in western Montana. Conservation Biology, 13,366–376.

Manly, B. F. J., McDonald, L. L., and Thomas, D. L., 1993, Resource selection by animals(London: Chapman & Hall ).

McLellan, B. N., Hovey, F. W., Mace, R. D., Woods, J. G., Carney, D. W., Gibeau, M. L.,Wakkinen, W. L., and Kasworm, W. F., 1999, Rates and causes of grizzly bearmortality in the interior mountains of British Columbia, Alberta, Montana,Washington, and Idaho. J. W ildl. Manage, 63, 911–920.

Moon, W. M., 1990, Integration of geophysical and geological data using evidential belieffunction.IEEE T ransactions on Geoscience and Remote Sensing, 28, 711–720.

Moon, W. M., 1993. On mathematical representation and integration of multiple spatialgeoscience data sets. Canadian Journal of Remote Sensing, 19, 63–67.

Mysterud, A., and Ims, R. A., 1999, Relating populations to habitats. T rends Ecol. Evol.,14, 489–490.

Peddle, D. R., 1993, An empirical comparison of evidential reasoning, linear discriminantanalysis and maximum likelihood algorithms for alpine land cover classi� cation.Canadian Journal of Remote Sensing, 19, 31–44.

Peddle, D. R., 1995a, MERCURYC: An Evidential Reasoning Image Classi� er. Computers& Geosciences, 21, 1163–1176.

Sixth Circumpolar Symposium on Remote Sensing of Polar Environment4652

Peddle, D. R., 1995b, Knowledge formulation for supervised evidential classi� cation.Photogrammetric Engineering & Remote Sensing, 61, 409–417.

Peddle, D. R., and Duguay, C. R., 1998, Mountain terrain analysis using a knowledge basedinterface to a GIS. Geomatica, 52, 145–163.

Peddle, D. R., and Ferguson, D. T., 2002, Optimisation of multisource data analysis: anexample using evidential reasoning for GIS data classi� cation. Computers &Geosciences, 28, 45–52.

Peddle, D. R., and Franklin, S. E., 1992, Multi-source evidential classi� cation of surfacecover and frozen ground. International Journal of Remote Sensing, 13, 3375–3380.

Peddle, D. R., and Franklin, S. E., 1993, Classi� cation of permafrost active layer depth fromremotely sensed and topographic evidence. Remote Sensing of Environment, 44, 67–80.

Peddle, D. R., Hall, F. G., and LeDrew, E. F., 1999, Spectral mixture analysis and geometricoptical re� ectance modeling of boreal forest biophysical structure. Remote Sensing ofEnvironment, 67, 288–297.

Peddle, D. R., Hall, F. G., and LeDrew, E. F., 2001, An integrated re� ectance model—evidential reasoning approach for physically-based satellite image analysis of forestbiophysical structure. Remote Sensing of Environment.

Purves, H., and Doering, C., 1998, Grizzly bear habitat eVectiveness: assessing cumulativeeVects of human use in Jasper National Park. Paper presented at the 18th AnnualESRI User Group Meeting, San Diego, CA.

Rey, M., Tunaley, K. E., and Sibbald, T., 1993, Use of the Dempster–Shafer algorithm forthe detection of SAR ship wakes. IEEE T ransactions on Geoscience and Remote Sensing,31, 1114–1118.

Richter, R., 1990. A fast atmospheric correction algorithm applied to Landsat TM images.International Journal of Remote Sensing, 11, 159–166.

Servheen, C., 1990, The status and conservation of the bears of the world. Intl. Conf . BearRes. and Manage. Monogr. Ser., 2, 38.

Shafer, G., 1976, A Mathematical T heory of Evidence (Princeton: Princeton University Press).Shafer, G., and Logan, R., 1987, Implementing Dempster’s rule for hierarchical evidence.

Arti� cial Intelligence, 33, 271–298.Srinivasan, A., and Richards, J. A., 1990, Knowledge-based techniques for multi-source

classi� cation. International Journal of Remote Sensing, 11, 505–525.Strong, W. L., 1999, Cumulative EVects Assessment for Vegetation and Botanical Resources

(Calgary, Alberta: Ecological Lands Surveys Ltd).Whitman, J. S., Ballard, W. B., and Gardner, C. L., 1986, Home range and habitat use by

wolverines in southwestern Alaska. Journal of W ildlife Management, 50, 440–462.Wilkinson, G. G., and Me’gier, J., 1990, Evidential reasoning in a pixel classi� cation hier-

archy—a potential method for integrating image classi� ers and expert system rulesbased on geographic context. International Journal of Remote Sensing, 11, 1963–1968.

Evidential reasoning with Landsat TM, DEM and GIS data for landcover classification in support of grizzly bear habitat mapping

Documents