Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone...

25
Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental data Chris Wright , Alisa Gallant U.S. Geological Survey Center for Earth Resources Observation and Science (EROS), 47914 252nd Street, Sioux Falls, SD 57198-0001, United States Received 27 March 2006; received in revised form 19 October 2006; accepted 21 October 2006 Abstract The U.S. Fish and Wildlife Service uses the term palustrine wetland to describe vegetated wetlands traditionally identified as marsh, bog, fen, swamp, or wet meadow. Landsat TM imagery was combined with image texture and ancillary environmental data to model probabilities of palustrine wetland occurrence in Yellowstone National Park using classification trees. Model training and test locations were identified from National Wetlands Inventory maps, and classification trees were built for seven years spanning a range of annual precipitation. At a coarse level, palustrine wetland was separated from upland. At a finer level, five palustrine wetland types were discriminated: aquatic bed (PAB), emergent (PEM), forested (PFO), scrubshrub (PSS), and unconsolidated shore (PUS). TM-derived variables alone were relatively accurate at separating wetland from upland, but model error rates dropped incrementally as image texture, DEM-derived terrain variables, and other ancillary GIS layers were added. For classification trees making use of all available predictors, average overall test error rates were 7.8% for palustrine wetland/upland models and 17.0% for palustrine wetland type models, with consistent accuracies across years. However, models were prone to wetland over- prediction. While the predominant PEM class was classified with omission and commission error rates less than 14%, we had difficulty identifying the PAB and PSS classes. Ancillary vegetation information greatly improved PSS classification and moderately improved PFO discrimination. Association with geothermal areas distinguished PUS wetlands. Wetland over-prediction was exacerbated by class imbalance in likely combination with spatial and spectral limitations of the TM sensor. Wetland probability surfaces may be more informative than hard classification, and appear to respond to climate-driven wetland variability. The developed method is portable, relatively easy to implement, and should be applicable in other settings and over larger extents. © 2006 Elsevier Inc. All rights reserved. Keywords: Wetland mapping; Palustrine wetlands; Landsat Thematic Mapper; Yellowstone National Park; Classification trees; Ancillary data; National Wetlands Inventory 1. Introduction The U.S. Geological Survey has a history of providing national and global land cover products (e.g., Loveland et al., 2000; Vogelmann et al., 2001), but wetlands have proved difficult to map with satellite remote sensing (Stehman et al., 2003; Wickham et al., 2004) because they may be rare in occurrence and their spectral and spatial characteristics are highly context-dependent. Our work was motivated by a need for reliable and relatively automated methods for mapping palustrine wetlands, vegetated wetlands traditionally identified as marsh, bog, fen, swamp, or wet meadow, across large areas. We were interested in the extent to which image texture, terrain data, and other ancillary environmental information could be used to improve wetland detection and classification. We selected Yellowstone National Park as a study area because of the availability of high-quality ancillary data for the park, and high interest by researchers there for wetland maps better re- flecting climate-driven wetland variability. This research was conceived as a first step in developing a satellite-based capabil- ity for monitoring wetlands annually in Yellowstone National Park. Additionally, we tested the application of classification trees for purposes of data fusion, and as an approach that might be extended over larger areas and other regional settings. Remote Sensing of Environment 107 (2007) 582 605 www.elsevier.com/locate/rse Corresponding author. Tel.: +1 605 594 2553. E-mail addresses: [email protected] (C. Wright), [email protected] (A. Gallant). 0034-4257/$ - see front matter © 2006 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2006.10.019

Transcript of Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone...

Page 1: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

t 107 (2007) 582–605www.elsevier.com/locate/rse

Remote Sensing of Environmen

Improved wetland remote sensing in Yellowstone National Park usingclassification trees to combine TM imagery and ancillary environmental data

Chris Wright ⁎, Alisa Gallant

U.S. Geological Survey Center for Earth Resources Observation and Science (EROS),47914 252nd Street, Sioux Falls, SD 57198-0001, United States

Received 27 March 2006; received in revised form 19 October 2006; accepted 21 October 2006

Abstract

The U.S. Fish and Wildlife Service uses the term palustrine wetland to describe vegetated wetlands traditionally identified as marsh, bog, fen,swamp, or wet meadow. Landsat TM imagery was combined with image texture and ancillary environmental data to model probabilities ofpalustrine wetland occurrence in Yellowstone National Park using classification trees. Model training and test locations were identified fromNational Wetlands Inventory maps, and classification trees were built for seven years spanning a range of annual precipitation. At a coarse level,palustrine wetland was separated from upland. At a finer level, five palustrine wetland types were discriminated: aquatic bed (PAB), emergent(PEM), forested (PFO), scrub–shrub (PSS), and unconsolidated shore (PUS). TM-derived variables alone were relatively accurate at separatingwetland from upland, but model error rates dropped incrementally as image texture, DEM-derived terrain variables, and other ancillary GIS layerswere added. For classification trees making use of all available predictors, average overall test error rates were 7.8% for palustrine wetland/uplandmodels and 17.0% for palustrine wetland type models, with consistent accuracies across years. However, models were prone to wetland over-prediction. While the predominant PEM class was classified with omission and commission error rates less than 14%, we had difficulty identifyingthe PAB and PSS classes. Ancillary vegetation information greatly improved PSS classification and moderately improved PFO discrimination.Association with geothermal areas distinguished PUS wetlands. Wetland over-prediction was exacerbated by class imbalance in likelycombination with spatial and spectral limitations of the TM sensor. Wetland probability surfaces may be more informative than hard classification,and appear to respond to climate-driven wetland variability. The developed method is portable, relatively easy to implement, and should beapplicable in other settings and over larger extents.© 2006 Elsevier Inc. All rights reserved.

Keywords: Wetland mapping; Palustrine wetlands; Landsat Thematic Mapper; Yellowstone National Park; Classification trees; Ancillary data; National WetlandsInventory

1. Introduction

The U.S. Geological Survey has a history of providingnational and global land cover products (e.g., Loveland et al.,2000; Vogelmann et al., 2001), but wetlands have proveddifficult to map with satellite remote sensing (Stehman et al.,2003; Wickham et al., 2004) because they may be rare inoccurrence and their spectral and spatial characteristics arehighly context-dependent. Our work was motivated by a needfor reliable and relatively automated methods for mapping

⁎ Corresponding author. Tel.: +1 605 594 2553.E-mail addresses: [email protected] (C. Wright), [email protected]

(A. Gallant).

0034-4257/$ - see front matter © 2006 Elsevier Inc. All rights reserved.doi:10.1016/j.rse.2006.10.019

palustrine wetlands, vegetated wetlands traditionally identifiedas marsh, bog, fen, swamp, or wet meadow, across large areas.We were interested in the extent to which image texture, terraindata, and other ancillary environmental information could beused to improve wetland detection and classification. Weselected Yellowstone National Park as a study area because ofthe availability of high-quality ancillary data for the park, andhigh interest by researchers there for wetland maps better re-flecting climate-driven wetland variability. This research wasconceived as a first step in developing a satellite-based capabil-ity for monitoring wetlands annually in Yellowstone NationalPark. Additionally, we tested the application of classificationtrees for purposes of data fusion, and as an approach that mightbe extended over larger areas and other regional settings.

Page 2: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

583C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

1.1. Importance of wetlands

Wetlands foster high levels of biodiversity (Gibbs, 2000) andperform vital ecosystem services including carbon sequestration(Gorham, 1991), flood mitigation, and water quality protection(Carter, 1999). Not surprisingly, wetlands may be particularlyvulnerable to climate change (Johnson et al., 2005; Poiani et al.,1996). Nearly one-half of plant and animal species listed asendangered by the U.S. Fish and Wildlife Service are wetlanddependent (U.S. Fish and Wildlife Service, 2002), and wetlandloss is arguably the most important cause of global amphibiandeclines (Blaustein et al., 1994). Although continental U.S. andglobal wetland losses are estimated to be on the order of 50%since the early 1700s (Dahl, 2000; Finlayson & Davidson, 1999),the importance of wetland conservation is now well establishedas a matter of U.S. and international public policy (NationalResearch Council, 2001; Ramsar Convention Bureau, 2004).

1.2. The U.S. National Wetlands Inventory

Inventory and monitoring of wetlands is fundamental toconservation efforts. In the United States, a nationwide effort tomap wetlands by the U.S. Fish and Wildlife Service, the NationalWetlands Inventory (NWI), began in 1975. The NWI is manuallyinterpreted from 1:58,000-scale color-infrared aerial photographyin combination with some 1:80,000-scale black-and-whitephotography (Tiner, 1990). Wetland locations are describedusing the Cowardin system (Cowardin et al., 1979), and mustsatisfy at least one of the following conditions: (1) at leastperiodically, the site supports predominantly hydrophytic vege-tation; (2) the substrate is predominantly undrained hydric soil;(3) the substrate is not soil and is saturated or covered by shallowwater at some time during the growing season of each year.

With respect to classification, the Cowardin system is de-fined hierarchically. At the highest level, freshwater wetlandsare divided into palustrine, riverine, and lacustrine systems(Cowardin et al., 1979). In addition to vegetated wetlands, thepalustrine system includes small, shallow, open water bodies, orponds. Riverine wetlands are defined very narrowly under theCowardin scheme, and are limited to river channels containingopen water or nonpersistent vegetation. Lacustrine wetlandsinclude both permanent and intermittent lakes, but, like riverinewetlands, are restricted to open water containing nonpersistentvegetation. Given these narrow definitions of riverine andlacustrine wetlands, and the prevalence of palustrine wetlands inour study area, we chose to limit this study to palustrine wet-lands. However, note that persistent wetland vegetation withinfloodplains adjacent to river channels, or adjacent to open waterwithin lake basins, is classified as palustrine wetland within theCowardin system (Cowardin et al., 1979).

Palustrine wetlands are further classified by dominant vege-tative life form and/or substrate composition (Cowardin et al.,1979). In Yellowstone National Park, these include aquatic bed(PAB), emergent (PEM), forested (PFO), scrub–shrub (PSS),and unconsolidated shore (PUS) classes. The emergent class ischaracterized by erect, rooted herbaceous hydrophytes, whilethe aquatic bed class contains vegetation growing on or below

the water surface. Scrub–shrub wetlands are dominated bywoody vegetation less than 6 m tall, including both true shrubsand young trees, while forested wetlands are composed ofwoody vegetation taller than 6 m. By definition, the uncon-solidated shore class contains less than 30% cover of plantsother than pioneering species (Cowardin et al., 1979).

Limited evaluation of the NWI has revealed varying accu-racy. Very low omission and commission error rates, less than5%, were reported in Massachusetts (Swarthout et al., 1981) andMichigan (Kudray & Gale, 2000). However, Kuzila et al.(1991) found that only 50% of hydric soils were also mapped aswetland by the NWI in southeast Nebraska. In the Blue RidgeMountains of Virginia, all palustrine wetlands delineated on two2.5-minute NWI maps were correctly identified, but with omis-sion error rates greater than 85% (Stolt & Baker, 1995). Ingeneral, the NWI appears to be biased toward errors of omission(Tiner, 1997).

Single-date aerial photographs used to generate the NWI arenot informative with respect to wetland dynamics, an importantshortcoming as wetland water levels can vary widely both withinand among years (Cowardin & Golet, 1995). Precipitationvariability in the Prairie Pothole region of North America causesinter-annual fluctuation in the extent of open water within wet-land basins (Poiani et al., 1996), which in turn drives succes-sional dynamics of wetland vegetation (Van der Valk, 1981). Inmountainous landscapes like Yellowstone National Park, lowerelevation wetlands receive water primarily from headwaterstreams and groundwater flow systems where the up-gradientwatershed is relatively small, while higher elevation wetlandsare nearly entirely dependent on precipitation inputs. As a result,wetlands in these landscapes are expected to respond toprecipitation variability fairly rapidly (Winter, 2000). At present,the magnitude and ecological importance of climate-drivenwetland variability in Yellowstone National Park is unknown.

Another shortcoming of the NWI is that large gaps exist inthe completion status and digital availability of wetland mapsacross the United States. In fact, it appears that the U.S. Fish andWildlife Service has backed away from generating a nationalproduct altogether. Rather than continuing wetland mapping inunfinished regions, the U.S. Fish and Wildlife Service recentlyannounced its intention to concentrate resources on updatingand digitizing existing wetland maps in areas with higherconservation priority (U.S. Fish and Wildlife Service, 2002).

While manual interpretation of aerial photographs is a veryaccurate method of land cover mapping (Sohl et al., 2004), itappears that a U.S. wetland mapping and monitoring programrelying solely on aerial photography will not prove feasible overthe long term. Given insufficient resources to map the entirenation once, let alone to monitor wetlands over time, it isimportant to augment the NWI with more automated and less-expensive methods.

1.3. Satellite remote sensing of wetlands

Large-area satellite remote sensing of wetlands has a historyin North America beginning in the mid-1970s (Work & Gilmer,1976). The U.S. Fish and Wildlife Service actually considered

Page 3: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

Fig. 1. Location of Yellowstone National Park within the United States. Scalebar applies to inset.

584 C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

using Landsat MSS data for the NWI but concluded that theinstrument could not meet classification detail and accuracyrequirements (Wilen & Bates, 1995). Changes in wetland com-position were documented in the Florida Everglades (Jensenet al., 1995) and on the Georgia Coastal Plain (Houhoulis &Michener, 2000) using SPOT imagery. The Canadian WetlandInventory, which will eventually encompass approximatelyone-quarter of the Earth's freshwater wetlands, combines TMand RADARSAT data (Leahy, 2003). Using TM imagery, theU.S. Coastal Change Analysis Program maps land cover alongthe U.S. coastline (including the Great Lakes), with a number ofwetland classes largely analogous to the Cowardin system(Brenner & Palmer, 2005). Also derived from TM imagery, theNational Land-Cover Data (NLCD) for 1992 (Vogelmann et al.,2001) includes woody and emergent wetland classes (AndersonLevel II, Anderson et al., 1976) which, in general, are classifiedwith very low accuracies, e.g., over the RockyMountain region,estimated producer's accuracy for a combined woody andemergent wetland class (Anderson Level I) was 27% with auser's accuracy of 14% (Wickham et al., 2004).

A key issue in large-area wetland remote sensing is themoderate spatial and spectral resolution of multispectral instru-ments like the TM sensor. With respect to wetland detection, theminimum mapping unit of TM imagery appears to be on theorder of 0.8–1.0 ha (Federal Geographic Data Committee, 1992;Johnston & Barston, 1993; Lunetta & Balogh, 1999). Bycontrast, the minimum mapping unit for 1:58,000-scale color-infrared photography is 0.5–1.2 ha for forested wetlands, andless than 0.5 ha for non-forested ponds and pothole marshes(Tiner, 1990). Spectrally, it can be difficult to discriminatedeciduous forested upland from deciduous forested wetland(Sader et al., 1995), especially following spring leaf emergence(Lunetta & Balogh, 1999). Ernst-Dottavio et al. (1981) foundthat shallow marsh, shrub swamp, and deciduous forestedswamp were spectrally similar to each other and to uplanddeciduous forest. Other authors have found that aggregatingsimilar wetland vegetation classes into more general super-classes was necessary to achieve acceptable classificationaccuracies (Harvey & Hill, 2001; Rutchey & Vilcheck, 1994).

1.4. Ancillary environmental information

While ancillary information has been used in a number ofremote sensing studies (e.g., Peddle & Duguay, 1995;Vogelmann et al., 1998), its application to wetland remotesensing has been limited, and restricted to small study areas(Huang & Jensen, 1997; Sader et al., 1995). A substantial bodyof ancillary data is available for Yellowstone National Park,including vegetation, soil, landform, and bedrock geologyclassifications, and terrain and climate data. These resourcespresented a unique opportunity for researching enhanced remotesensing of wetlands over a large area (nearly 900,000 ha).

1.5. Classification trees

Classification trees have been successfully used to classifyland cover from multiple sources of data (e.g., Friedl & Brodley,

1997; Hansen et al., 1996), including in the Greater YellowstoneRegion (Lawrence & Wright, 2001). Classification trees easilyaccommodate data from all measurement scales (i.e., nominal,ordinal, interval, and ratio scales) and make no distributionalassumptions (Breiman et al., 1984). As a class of machine-learning algorithms, classification trees automatically selectvariables and their hierarchical structure is capable of detectingnon-additive interactions between variables without explicitspecification (Clark & Pregibon, 1992).

Classification trees were used in this study in an effort to bringall available information to bear on the problem of detecting andclassifying palustrine wetlands in Yellowstone National Park.Concurrently, we evaluated the importance of different types ofwetland predictors as a logical progression in the degree of datafusion one might choose, or be able to conduct.

2. Methods

2.1. Study area

Straddling the Continental Divide, Yellowstone NationalPark is located principally in the northwest corner of Wyoming,U.S.A. and covers nearly 900,000 ha of steep mountains andhigh plateaus (Fig. 1).Much of the area is covered in Tertiary andQuaternary volcanic deposits that have been glaciated severaltimes. Elevations vary from about 1550 m to well over 3000 m.Long, cold winters and short, cool summers are characteristic ofthe climate. Total annual precipitation is greatest (around200 cm) in the southwest corner of the park and in the mountainranges in the north and east; elsewhere, the landscape is in a rainshadow. Most of Yellowstone National Park receives between75 cm and 125 cm of annual precipitation, although drier areasreceive only about 25 cm. Much of the annual precipitationoccurs as snow, and spring snowmelt is an important source ofmoisture for both surface waters and soils (Despain, 1990).

Emergent wetlands in Yellowstone National Park are typi-cally composed of sedges (Carex spp.), while scrub–shrubwetlands are dominated by willows (Salix spp.). Two coniferous

Page 4: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

585C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

tree species commonly found in forested wetlands areEngelmann spruce (Picea engellmannii) and subalpine fir(Abies lasiocarpa). Other common tree species include lodge-pole pine (Pinus contorta), whitebark pine (Pinus albicaulis),and Douglas-fir (Pseudotsuga menziesii). Aspen (Populustremuloides), while widely distributed, is relatively rare withrespect to its overall cover (Despain, 1990).

2.2. Data

Cloudless, terrain-corrected Landsat-5 TM images of Yel-lowstone National Park from single dates in 1988, 1997, and2003 were acquired from the USGS Center for Earth ResourcesObservation and Science (Table 1). From 1999–2002, cloudless,terrain-corrected Landsat-7 ETM+ images were available fortwo dates during the growing season (Table 1). Assembledimagery spanned a range of annual precipitation. YellowstoneNational Park was unusually dry in 1988, the year of catastro-phic wildfires (28 cm precipitation recorded at Mammoth HotSprings, WY), while 1997 was unusually wet (43 cm precip-itation recorded at the same station) (Western Regional ClimateCenter, 2004). From 1999 through 2003, Yellowstone NationalPark experienced a multi-year drought (Swanson et al., 2004).

Digital numbers were converted to at-satellite reflectance andtasseled-cap transformed to brightness, greenness, and wetnessbands using the coefficients of Huang et al. (2002). In yearswhere two TM scenes were available, per-pixel differences werecalculated for the six reflectance bands and the three tasseled-capbands. By subtracting late-summer values from mid-summervalues, we hoped to capture phenological differences betweenwetlands and uplands (e.g., late-summer senescence in uplandsdue to moisture limitation) and between different wetland types.

While we expected tasseled-cap bands to be useful predictors,we also wanted to explore other dimension-reducing lineartransformations potentially more specific to the problem ofdiscriminating wetland from upland. We chose to conduct acanonical discriminant analysis, where dimension reduction isachieved in a way that maximizes differences between groupmeans (Zhao & Maclean, 2000). In developing wetland/uplandmodels, canonical discriminant analyses were first conducted onwetland/upland training data using the S-PLUS ‘cancor’function (S-PLUS 6.1.2, Insightful Corporation, Seattle,Washington). In years where one TM image was available, theobservation matrix consisted of at-satellite reflectance in TMbands 1–5 and band 7. When two image dates were available,the observation matrix contained twelve variables, i.e. six TM

Table 1Acquisition dates for TM imagery

Year Date(s) Landsat-

1988 22 July 51997 15 July 51999 14 August, 15 September 72000 15 July, 16 August 72001 2 July, 20 September 72002 5 July, 23 September 72003 1 August 5

bands at two dates. As there were two classes in the wetland/upland analysis, six- and twelve-dimensional reflectance datawere reduced to one-dimensional canonical discriminant scores.

Image texture was quantified using a variance filter, a non-linear, non-directional edge detector effective in settings wheresudden changes in brightness occur without directional bias(Wilson, 1997). Visually, Yellowstone wetlands tend to exhibitabrupt boundaries, say the transition from senescent uplandvegetation to actively transpiring (green) wetland vegetation inlate summer. Palustrine wetlands in Yellowstone National Parkare typically small, and we expected within-wetland texture tobe dominated by edge effects. Image texture was calculated in3×3, 5×5, and 7×7 pixel moving windows in the brightness,greenness, and wetness tasseled-cap bands and in TM bands 4and 5. Preliminary results indicated that 5×5 windows wereoptimal for wetland/upland discrimination and 3×3 windowswere optimal for distinguishing among palustrine wetlandclasses (Wright, 2004). As a result, only 5×5 image textureswere used as predictors in wetland/upland classification trees,while only 3×3 image textures were used in palustrine wetlandtypes models.

A 30-m DEM (Yellowstone National Park Spatial AnalysisCenter, unpublished metadata) was used to derive predictorsindicative of water-collecting terrain. Undrained sinks in theDEM were filled using the algorithm of Planchon and Darboux(2002), and sink fill-depths were kept as an ancillary variable.TAPES-G terrain analysis software (Wilson & Gallant, 2000)was used to calculate slope and a steady-state topographicwetness index, log(upstream contributing area / tan slope).While we expected wetlands to be more likely in pixels wherethe slope gradient is minimal, the topographic wetness indexalso takes into account the size of the upslope area potentiallycontributing water to any given pixel (Wilson & Gallant, 2000).

Ancillary GIS data compiled for Yellowstone National Parkwere obtained from the U.S. National Park Service. Availablevegetation layers included habitat type and cover type classifi-cations. Habitat types indicate potential climax vegetation andare determined by a set of environmental factors (e.g., tempera-ture and moisture regime, light availability, mineral nutrientavailability, and disturbance frequency) and expert knowledgeof the Yellowstone flora (Despain, 1990). Cover types aredefined by the successional stage of the plant communityfollowing the most recent major disturbance, typically fire(Despain, 1990), and indicate actual vegetation. Yellowstonecover types were delineated for forested vegetation circa 1991,i.e., following the catastrophic fires of 1988 (YellowstoneNational Park Spatial Analysis Center, unpublished metadata).

Yellowstone landforms have been classified as ‘meso-scale’elements at extents of 1–100 ha (Shovic, 1996). The classifi-cation is based on both visible and inferred characteristicsincluding type and degree of stream drainage dissection, type ofsurficial materials, slope gradient distribution, slope curvature,relief, and proportion and shape of bedrock exposure (Shovic,1996). This list is likely to include factors associated withwetland formation and, as such, we expected the landformclassification to integrate factors influencing wetland presenceat scales of 1–100 ha.

Page 5: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

586 C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

The bedrock geology classification was derived from the1972 USGS Geologic Map of Yellowstone National Park(Yellowstone National Park Spatial Analysis Center, unpub-lished metadata). Yellowstone National Park experiencedextensive glaciation from 2 million to 14,000 years ago, andthe resulting glacial tills constitute the parent materials of mostpresent-day Yellowstone soils. Because the Yellowstone ice capwas largely stationary, these glacial tills were generally notmoved, and most present-day soils formed from the underlyingbedrock (Despain, 1990). This may have important implicationswith respect to wetland formation. Coarse-textured rhyoliticsoils have higher hydraulic conductivities and lower water-holding capacities than finer-textured andesitic soils. Despain(1990) noted that wet forests in Yellowstone National Park tendto form on the margins of Quaternary rhyolitic flows.

The Yellowstone soil classification (Rodman et al., 1996)consists of eighty soil complexes, too many classes to be used ina classification tree analysis. Generalization of the soil classifi-cation is complicated by the fact that individual soil polygonsare reported as mixtures of component soil families (identifiedby their percent cover). From attribute summaries for each soilfamily (Rodman et al., 1996), we calculated weighted averages(by percent cover) for percent clay content and percent rockfragments (N2 mm in diameter), expecting these two attributesto influence soil texture and hydraulic conductivity. Soil poly-gons were then assigned these weighted averages.

Available climate data for Yellowstone National Park in-cludes 30-year average annual precipitation (1961–1990) and15-year average annual snowfall (1958–1972) interpolatedfrom snow course, snow survey telemetry, and climate stationdata (Yellowstone National Park Spatial Analysis Center,unpublished metadata). Values within individual precipitationand snowfall polygons are reported as intervals, e.g. the lowestinterval for average annual precipitation is 25.4–30.5 cm. Forpurposes of analysis, these intervals were treated as ordinalvariables, with interval midpoints assigned to each precipitationor snowfall polygon.

The Yellowstone NWI is derived primarily from 1:58,000-scale color infrared aerial photography acquired in August andSeptember of 1982–84 (Elliot & Hektner, 2000), although some

Table 2Summary of wetland predictors available in Yellowstone National Park

Rationale for inclusion Variable

Water-collecting terrain – Landform type– Slope– Topographic wetness index– Sink fill depth

Substrates that influence sub-surface hydrology – Soil texture– Bedrock geology

Precipitation inputs to wetlands – Average annual precipitatio– Average annual snowfall

Potential climax vegetation as a responseto physical environment

– Habitat type

Coarse classification of existing vegetation – Cover type

Seasonal conditions – Satellite data

portions were interpreted from 1970s-era 1:80,000-scale black-and-white aerial photographs (Yellowstone National ParkSpatial Analysis Center, unpublished metadata). The averagesize of palustrine wetland polygons in the Yellowstone NWI is1.6 ha, with a median size of 0.26 ha. Polygons smaller thanapproximately 12 ha account for 50% of the total palustrinewetland area in the Yellowstone NWI.

The 30-m DEM and derived terrain elements were posi-tioned to align with TM pixels by nearest-neighbor sampling.All vector data, and all raster data with resolutions coarser than30 m, was converted to 30-m raster data using the TM grid andnearest-neighbor sampling.

Assembled wetland predictors span processes operating ontimescales ranging from seasonal (satellite data) to the moderateterm (cover type, habitat type, 30-year average precipitation),and over the long term (soil texture, slope, topographic wetnessindex, bedrock geology, and landforms). Rationales for in-cluding different types of wetland predictors are reviewed inTable 2. Habitat type, cover type, landform, and bedrockgeology class names are summarized in Table 3.

2.3. Model building and wetland mapping

Following the Cowardin scheme, we developed wetlandclassification trees hierarchically. At the most general level,classification trees distinguished palustrine wetland from up-land. At the next level, classification trees discriminated amongthe five palustrine wetland types: PAB, PEM, PFO, PSS, andPUS.

We generated a common set of training locations to buildwetland/upland classification trees for each of the seven years ofTM imagery (Fig. 2). This allowed us to assess the consistencywith which classification trees made use of identical ancillaryinformation given TM data from different years. Training loca-tions were selected using the Yellowstone NWI. Anticipatingthat the minimum mapping unit for TM-based wetland remotesensing in Yellowstone National Park would be on the order of0.8–1.0 ha (Federal Geographic Data Committee, 1992;Johnston & Barston, 1993; Lunetta & Balogh, 1999), randomsampling of wetlands was restricted to pixels falling inside

Timescale Reference

Long term Shovic (1996)Yellowstone National Park Spatial Analysis Center(unpublished metadata)

Rodman et al. (1996)Yellowstone National Park Spatial Analysis Center(unpublished metadata)

n Medium term Yellowstone National Park Spatial Analysis Center(unpublished metadata)Despain (1990)

Yellowstone National Park Spatial Analysis Center(unpublished metadata)

Short term

Page 6: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

Table 3Habitat type, cover type, landform, and bedrock geology classes

Habitat type Cover type Landform Bedrock geology

Douglas-fir Aspen Alluvial landforms (al) Cambrian sedimentsLodgepole pine Douglas-fir climax Breaklands and colluvial slopes (bcs) Cretaceous sedimentsNon-forested alpine Douglas-fir post-disturbance Fluvial plateaus (fp) Devonian sedimentsNon-forested dry Douglas-fir successional Fluvial rolling uplands (fru) Mississippian sedimentsNon-forested moist Engelmann spruce and subalpine fir climax Glacial troughs and cirques (gtc) Ordovician sedimentsNon-forested sagebrush Krummholz Glaciated plateaus (gp) Pennsylvanian sedimentsNon-forested wet Lodgepole pine climax Glaciated rolling uplands (gru) Permian sedimentsNon-forested willow Lodgepole pine post-disturbance Glaciofluvial terraces and plains (gtp) Precambrian metamorphicsPitchstone plateau complex Lodgepole pine successional Hydrothermal landforms (hl) Quaternary basalt flowsSubalpine fir, lodgepole pine Non-forest Landslides (ls) Quaternary rhyolite flowsSubalpine fir moist Pygmy lodgepole pine Other glaciofluvial landforms (ogl) Quaternary rhyolite tuffsSubalpine fir wet Water Water (w) Quaternary hydrothermal depositsSubalpine fir, whitebark pine Whitebark pine climax Quaternary sedimentsTalus Whitebark pine post-disturbance Tertiary andesitesThermal vegetation Whitebark pine successional Tertiary andesites, basaltsWater Tertiary basaltsWet forest Tertiary intrusivesWhitebark pine Tertiary sediments

Triassic sediments

587C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

palustrine polygons larger than 1 ha (approximately 11 TMpixels). Stratified random sampling of upland training pixels wasoriented relative to NWI polygons. To acquire a representative

Fig. 2. Flowchart of steps leading to palustrine wetland m

sample from what we expected to be areas of wetland-to-uplandtransition, one stratum was limited to pixels within 60–120 mdistance from any NWI polygon edge. The choice of an

odel predictions across Yellowstone National Park.

Page 7: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

Table 4Error rates for TM+TXT+DEM+GIS palustrine wetland/upland classificationtrees by year

Omission error(%)

Commission error(%)

Overall error(%)

Wetland Upland Wetland Upland

1988 8.85 6.71 55.44 0.56 6.831997 9.97 8.65 61.89 0.64 8.721999⁎ 7.59 8.17 59.91 0.49 8.142000⁎ 7.57 8.00 59.39 0.48 7.982001⁎ 9.80 6.82 56.10 0.62 6.992002⁎ 8.68 8.68 61.63 0.56 8.682003 8.06 6.96 56.13 0.51 7.02Mean 8.65 7.71 58.64 0.55 7.76s.e.m. 0.37 0.33 1.03 0.02 0.31

⁎Years for which two images were available.

588 C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

approximately 2-pixel buffer around NWI polygons wasarbitrary, but was intended to reduce mixed-pixel samplingandmislabeling due to positional errors, given that the positionalaccuracy of the Yellowstone NWI is unknown. A second uplandstratum was sampled from all pixels greater than 120 m distancefrom any NWI polygon to ensure representative coverage ofuplands across the entire park. The number of pixels sampled inthe two upland strata was proportional to their relative areas.

Palustrine wetlands are relatively rare in YellowstoneNational Park, occupying approximately 6% of the total landarea (Elliot & Hektner, 2000). This rarity poses a problem fordiscrimination between wetland and upland, since classificationtrees trained on unbalanced data tend to be highly accurateidentifying dominant classes and much less accurate identifyingrare classes. When classification trees are trained with equal-sized samples from each class, prediction accuracies tend to bemore similar (Breiman et al., 1984). To best support the needs of

Fig. 3. Average omission and commission error rates for 1988, 1997, and 1999–2003 palustrine wetland/upland classification trees across four predictivemodels (+/−s.e.m.).

field researchers in Yellowstone National Park, we focused oureffort on limiting errors of omission for the palustrine wetlandclass, or failures to identify wetland. Thus we chose to sampleequal numbers of wetland and upland training pixels, in thiscase 50,000 pixels from each class.

A second, common set of training locations was sampled tobuild palustrine wetland type classification trees for each ofthe seven years of TM imagery. Palustrine wetlands in theYellowstone NWI are dominated by the PEM class (nearly 69%of the total palustrine wetland area) and PFO class (approxi-mately 22%), while PSS, PAB, and PUS wetlands are rare, withless than 6% cover in each (Elliot & Hektner, 2000). Wesampled palustrine wetlands in proportion to the area of eachclass in the Yellowstone NWI, recognizing that this wouldlikely favor classification of the predominant PEM type (emer-gent wetlands are an important habitat for biological monitoringin Yellowstone National Park). A total of 100,000 traininglocations were randomly sampled from NWI polygons largerthan 1 ha, including 68,774 PEM pixels, 22,242 PFO pixels,5232 PSS pixels, 2596 PUS pixels, and 1156 PAB pixels.

We built classification trees using S-PLUS software (S-PLUS6.1.2, Insightful Corporation, Seattle, Washington), and thenpruned those trees by the automated cost-complexity method ofBreiman et al. (1984), in which more extensive tree pruningoccurs as the cost-complexity parameter is increased. Error rateswere determined for successively pruned classification trees usingpruning datasets assembled for each year of TM imagery (Fig. 2).Pruning datasets were compiled by stratified random samplingidentical to that described above for training data and bothwetland/upland and palustrine wetland type pruning datasetscontained 100,000 pixels randomly drawn from locations notpreviously selected for training. Optimally pruned classificationtreeswere identified by the cost-complexity parameter resulting inthe smallest overall classification error rate for the pruning data.

We constructed classification trees for each of the seven yearsof satellite imagery using four different combinations of predic-tors: (1) TM imagery alone (including canonical discriminantscores in wetland/upland trees only), (2) TM imagery plusimage texture (TM+TXT model), where image texture wascomputed for tasseled-cap brightness, greenness, and wetness,and TM bands 4 and 5, (3) TM imagery plus texture and terrain

Page 8: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

Table 5Sample error matrix and estimated population error matrix for the 1988 TM+TXT+DEM+GIS palustrine wetland/upland classification tree

(a.) Sample error matrix

Predicted class

WET UPL Producer's accuracy (%) Omission error (%)

Actual class WET (n=50,000) 45,573 4427 91.15 8.85UPL (n=50,000) 3357 46,643 93.29 6.71

(b.) Estimated population error matrix

Predicted class

WET UPL

Actual class WET (n=462,844) (0.9115)(462,844)=421,882 (0.0885)(462,855)=40,962UPL (n=7,822,088) (0.0671)(7,822,088)=524,862 (0.9329)(7,822,088)=7,297,226

Overall error (%)Commission error (%) 55.45 0.56 6.83

Producer's accuracies and omission error rates estimated from the sample error matrix were used to fill the population error matrix given known numbers of palustrinewetland and upland pixels in the study area. Commission and overall error rates were subsequently calculated from these estimates.

589C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

information derived from the Yellowstone DEM (TM+TXT+DEM model), including elevation, sink fill-depth, slope andtopographic wetness index, and (4) all predictors including TMimagery, image texture, terrain information, and all additionalancillary GIS information (TM+TXT+DEM+GIS model).

Palustrine wetland/upland classification trees using the fullsuite of available predictors (TM+TXT+DEM+GIS models)were used to predict probabilities of palustrine wetland occur-rence across the Yellowstone landscape. For mapping purposes,pixels with probabilities of palustrine wetland occurrence ≥0.5were identified as palustrine wetland (Fig. 2). We then reclas-sified these pixels according to palustrine wetland type, assign-ing labels from the highest-probability palustrine class.

2.4. Error assessment

We evaluated optimally pruned classification trees by apply-ing them to test data assembled for each year of TM imageryusing common locations across the seven years and stratifiedrandom sampling identical to that used for training and pruningdata (Fig. 2). Both wetland/upland and palustrine wetland typestest data sets contained 100,000 pixels sampled from locations

Table 6Kappa coefficients for individual palustrine wetland/upland models by year and Z-stamodels in any given year

Model

TM TM+TXT TM+TXT+DEM TMDEM

Kappa Standarderror

Kappa Standarderror

Kappa Standarderror

Kap

1988 0.3271 2.75E-03 0.3669 2.97E-03 0.4901 3.79E-03 0.561997 0.2505 2.17E-03 0.2894 2.45E-03 0.3749 3.05E-03 0.491999 0.3600 2.97E-03 0.3906 3.14E-03 0.4926 3.80E-03 0.522000 0.4057 3.29E-03 0.4416 3.50E-03 0.4695 3.63E-03 0.522001 0.3646 3.02E-03 0.3232 2.69E-03 0.4009 3.22E-03 0.552002 0.3464 2.88E-03 0.3860 3.16E-03 0.4592 3.62E-03 0.502003 0.3605 3.00E-03 0.4218 3.37E-03 0.4816 3.72E-03 0.56

⁎Increase in Kappa coefficient statistically significant at the p=0.01 level.

not previously used for training or pruning purposes, andwetland pixels were again limited to NWI polygons larger than1 ha. Population error matrices were estimated using omis-sion error rates from sample error matrices and the Horvitz–Thompson estimator for stratified random sampling (Thomp-son, 1992). We then used population error matrices to estimatecommission and overall error rates. Maximum-likelihoodestimators for stratified designs were used to calculate Kappacoefficients of agreement and their variances (Stehman, 1999).In any given year, we evaluated the statistical significance ofincreased Kappa coefficients following the addition of morepredictors (e.g., from the TM-only model to the TM+TXTmodel) by calculating Z-statistics for pairwise differences ofKappa coefficients (Congalton & Green, 1999).

3. Results

3.1. Palustrine wetland/upland models

Test data error rates decreased incrementally as morepredictors were added to palustrine wetland/upland classifica-tion trees (Fig. 3). Adding image texture reduced omission

tistics for increases in Kappa coefficients when additional predictors are added to

Z-statistic for Kappa differences

+TXT++GIS

(TM+TXT) (TM+TXT+DEM) (TM+TXT+DEM+GIS)

pa Standarderror

– (TM) – (TM+TXT) – (TM+TXT+DEM)

60 4.15E-03 9.83⁎ 25.57⁎ 13.50⁎

60 3.82E-03 11.88⁎ 21.86⁎ 24.77⁎

20 3.93E-03 7.09⁎ 20.71⁎ 5.38⁎

76 3.96E-03 7.49⁎ 5.53⁎ 10.80⁎

73 4.13E-03 18.50⁎ 29.89⁎

11 3.83E-03 9.26⁎ 15.23⁎ 7.94⁎

08 4.12E-03 13.58⁎ 11.90⁎ 14.26⁎

Page 9: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

Fig. 4. Average deviance reduction by variable for 1988, 1997, and 1999–2003TM+TXT+DEM+GIS palustrinewetland/upland classification trees (+/−s.e.m.).

590 C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

errors more in the wetland class than in the upland class, whileadding DEM-derived variables improved classification of up-land more than wetland (Fig 3a). When all predictors were madeavailable in the full TM+TXT+DEM+GIS model, the averagerate of omission errors dropped to 8.7% in the wetland classand 7.7% in the upland class. Low overall and omission errorrates for full models were consistent across years (Table 4).There was no evidence that drought from 1999–2003 affectedaccuracy, and in years where two dates of imagery wereavailable (1999–2002) model performance did not improve.Average commission error rates in the upland class wereextremely low for all four models, less than 1% (Fig. 3b). In thewetland class, however, average commission error rates (i.e.,wetland false-positives) were very high, ranging from nearly75% for the TM-only model to nearly 60% for the full model.High commission error rates in the wetland class were, in part, aresult of the overwhelming preponderance of the upland classon the Yellowstone landscape (Table 5). Within individualyears, adding more predictors to wetland/upland models result-ed in statistically significant increases in Kappa coefficients(pb0.01), except from the TMmodel to the TM+TXTmodel in2001, where Kappa decreased (Table 6).

In TM+TXT+DEM+GIS models, 27 predictors werepresented to the classification tree algorithm in years where asingle TM image was available, increasing to 50 variables in 2-image models. The classifier selected 15 variables or fewer inall years but 1997, where, in an apparent response to greaterclassification difficulty in a very wet year (Table 4), the algo-rithm used 19 predictors.

One way to assess variable importance in classification treesis to rank predictors by deviance reduction. Deviance is ameasure of class purity, where smaller values correspond tomixtures dominated by one class (scaled by the number of casesin each class). Classification tree rules that sort a heterogeneousmixture of classes into more pure subsets cause a reduction indeviance (Venables & Ripley, 1994), analogous to reduction ofvariance in regression. For purposes of comparison, by-variabledeviance reduction was averaged over the seven TM+DEM+TXT+GIS classification trees (Fig. 4). Canonical discriminantscores were overwhelmingly important, followed by habitat andlandform types. Image texture (in both TM band 4 and withrespect to tasseled-cap greenness) and slope were of interme-diate importance. Smaller deviance reduction was associatedwith bedrock geology, percent clay content, tasseled-capgreenness, and cover type. A limited number of other predictorsselected by classification trees, but not included in Fig. 4,accounted for negligible deviance reduction.

Building classification trees for different years allowed us toassess the consistency with which TM+TXT+DEM+GISmodels used wetland predictors. For purposes of illustration, welimit these comparisons to 1988, 1997, and 2001 trees andhigher levels of tree structure (Fig. 5). Despite their simplicity,these trees accounted for over 80% of the total deviancereduction by full classification trees. And although 1988 was avery dry year, 1997 was a very wet year, and two dates ofimagery were available in 2001, the resulting classification treeswere structurally very similar.

The first split in all three trees is made on canonicaldiscriminant scores, with larger values generally associated withpalustrine wetland and smaller values generally associated withupland (Fig. 5). These initial splits were also relatively accurate,with omission error rates less than 20% for both classes basedsolely on these rules. Following these initial splits, each treecontains an upland main branch to the left and a wetland mainbranch to the right. The ensuing challenge for the classificationtree algorithm is isolating the rare class in each main branch,i.e., upland pixels within the wetland main branch and wetlandpixels within the upland main branch.

Within the upland main branch, each tree makes an initialsplit on habitat type (Fig. 5). In wetter habitat types (non-forestwet, non-forest willow, subalpine fir wet, thermal vegetation,water, and wet forest), wetland pixels are isolated by greaterimage texture with respect to TM band 4 or tasseled-capgreenness. In drier habitat types (including dry to moist non-forested classes, a number of forested classes, etc.), wetlandpixels are identified by greater image texture, lesser slope, andassociation with alluvial and hydrothermal landforms.

Down the wetland main branch, an initial split on habitattype is also made in each tree (Fig. 5). In general, upland pixelsfrom more mesic habitat types are separated by association withgreater slope gradients and upslope landform types (breaklandsand colluvial slopes, fluvial plateaus, fluvial rolling uplands,glaciated plateaus, glaciated rolling uplands, glacial troughs andcirques, and landslides).

3.2. Canonical discriminant analysis

As noted above, canonical discriminant scores were strongpredictors of palustrine wetland occurrence in YellowstoneNational Park. In years where only a single date of TM imagerywas available, standardized canonical weights appear tocontrast TM bands 1, 2, 4, and 5 with TM bands 3 and 7(Table 7). However, the dominant contrast appears to bebetween TM bands 5 and 7. Canonical weights for band 7 arerelatively large and negative across both single- and two-imageyears, while canonical weights for band 5 are relatively largeand positive.

Page 10: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

591C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

3.3. Palustrine wetland type models

Overall error rates of palustrine wetland typemodels declinedincrementally as more predictors were added, although terraindata had a negligible effect (Fig. 6a). In the predominant PEMclass, average omission error rates were low, less than 10% in thelast three models. But in the PSS class, model performance wasvery poor until thematic GIS information was made available,

Fig. 5. Simple palustrine wetland/upland classification trees for 1988 (dry year), 19indicated rule are sent to the left; otherwise, pixels are sent to the right. Refer to Ta

approximately halving omission errors. Similarly, GIS datareduced the average rate of omission errors for PFO wetlands toapproximately 30%. In the PUS class, omission error rateswere relatively constant across models (increasing slightly in theTM+TXT model), although there was a downward trend withrespect to errors of commission (Fig. 6b). Similar downwardtrends in commission error rates occurred in the PEM, PFO, andPSS classes as more predictors were added. Both omission and

97 (wet year), and 2001 (two TM images). At each node, pixels satisfying theble 3 for definitions of landform abbreviations.

Page 11: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

Fig. 5 (continued ).

592 C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

commission error rates were high for PABwetlands, and actuallyincreased across the four models.

Error rates for TM+TXT+DEM+GISmodels were relativelyconsistent across years, and availability of an additional TMimage in 1999–2002 did not improve performance (Table 8). Inthe 1988 sample error matrix (Table 9), most misidentified pixelsare classified as PEM. Within each year, Z-statistics for increasesin Kappa coefficients were significant at the p=0.01 level(Table 10), except for differences between the TM+TXT andTM+TXT+DEMmodels in 1988 (pb0.10) and 1999 (pb0.05).

TM+TXT+DEM+GIS models for discriminating amongpalustrine wetland types used more predictors than their wet-land/upland counterparts: 18–20 variables in years with one TMimage and 24–31 variables in years with two images. Withrespect to deviance reduction, cover type was the most impor-

Table 7Palustrine wetland/upland standardized canonical weights by year

Standardized canonical weights (×10−2)

1988 1997 2003

1st image Band 1 −0.0154 0.0943 0.06Band 2 0.0563 0.1566 0.01Band 3 −0.0476 −0.2618 −0.01Band 4 0.2417 0.0916 0.14Band 5 0.2596 0.3748 0.42Band 7 −0.2853 −0.4723 −0.51

2nd image Band 1Band 2Band 3Band 4Band 5Band 7

tant variable (Fig. 7), in contrast with its relative unimportancein wetland/upland models (Fig. 4). On average, habitat typeresolved the 2nd most amount of deviance (with a very smallstandard error), followed by TM band 4 texture, TM band 3reflectance, and a number of other TM-derived predictors.Elevation was ranked within the middle of this group. Largestandard errors of mean deviance reduction for the tasseled-capbrightness, TM 5 difference, and TM 3 difference variablesreflect the fact that these were important predictors only insingle years. Landform type and percent clay content, twovariables relatively important in wetland/upland models(Fig. 4), play a minor role. Minimal deviance reduction bymean annual snowfall and precipitation suggests a limited influ-ence of climate in determining wetland type, although climatemight have been better represented by a covariate like elevation.

1999 2000 2001 2002

02 0.1188 0.0692 −0.0516 0.045911 0.1412 0.0966 0.1076 0.134660 −0.2703 −0.0413 0.0142 −0.182380 0.1022 −0.1360 0.0581 −0.020544 −0.0206 0.1468 0.1617 0.200515 −0.0849 −0.2506 −0.2676 −0.2666

−0.0451 −0.0436 −0.0978 −0.03530.1632 0.1348 0.1782 0.05310.0066 −0.1707 −0.0455 0.0765

−0.0015 0.2306 −0.0346 0.05190.4370 0.3169 0.5331 0.4417

−0.4236 −0.2872 −0.4520 −0.4191

Page 12: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

Fig. 6. Average omission and commission error rates for 1988, 1997, and 1999–2003 palustrine wetland type classification trees across four predictive models(+/−s.e.m.).

593C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

Simple TM+TXT+DEM+GIS wetland type classificationtrees for 1988, 1997, and 2001 (Fig. 8) also achieved substantialdeviance reduction using a small number of high-level rules (onthe order of 60–70%), although they were not as consistent aswetland/upland models in selecting predictors. For example,three different variables are used in initial splits sending mostPFO pixels down the right main branch of each tree: tasseled-cap brightness in 1988, cover type in 1997, and TM band 3differences in 2001. Subsequently, PFO wetlands are distin-guished from PEM wetlands by less image texture (in TM band4 or with respect to tasseled-cap brightness) and either asso-ciation with forested cover types (1988), less tasseled-capgreenness (1997), or less reflectance in TM band 3 (2001). ThePSS class is separated by its association with non-forest willow,Douglas-fir, and lodgepole pine habitat types, but non-forestwillow is the key habitat type in this rule. Higher reflectance in

TM bands 1 and 3, or, equivalently, less tasseled-cap greenness,are used to differentiate the PUS class. Finally, note that in fullclassification trees PAB wetlands are discriminated at levelsbelow those illustrated in Fig. 8.

3.4. Wetland maps

In an approximately 280-km2 area, largely southeast of theGrand Canyon of the Yellowstone River (Fig. 9), the 2003palustrine wetland probability surface shows that high-proba-bility pixels tend to coincide with NWI palustrine wetlandpolygons (Fig. 10). The majority of pixels classified as wetlandwere subsequently assigned to the PEM class. Apparent wetlandover-prediction tends to occur along the margins of NWIpolygons, and many of these pixels are classified as PFO. Anumber of active hydrothermal features visible in the raw image(Fig. 9) are correctly classified as PUS wetland (Fig. 10), withthe exception of extensive geothermal deposits on canyon wallsabove the Yellowstone River. Within a handful of lakes seen inthe raw image, pixels along shorelines are typically classified asPAB, while lake interiors (presumably deeper water with nofloating vegetation) are not identified as palustrine wetland.Apparent failures to detect wetlands are nearly always limited tosmaller NWI polygons and narrow, linear wetland features.However, the palustrine wetland/upland model does appear tobe responding to many smaller wetlands in this area.

4. Discussion

We found that image texture, terrain data, and ancillary GISinformation greatly improved palustrine wetland detection,nearly halving the average overall error rate of palustrinewetland/upland models (Fig. 3). Classification trees were veryeffective in discriminating the predominant PEM class fromother palustrine wetland types (Fig. 6). Our results acrossYellowstone National Park were approximately three timesmore accurate than those achieved by the 1992 NLCD over theentire Rocky Mountain region (Wickham et al., 2004). Modelperformance was stable from year-to-year (Tables 4 and 8), eventhough TM imagery was acquired from varied dates within thegrowing season, over a wide range of annual precipitation, andfrom two different, but comparable, sensors. Lastly, classifi-cation trees used different wetland predictors consistently(Figs. 4, 5, 7, and 8), allowing ready interpretation of theirroles and relative importance in this geographic setting.

4.1. Important wetland predictors

The ability of canonical discriminant scores to discriminatepalustrine wetland from upland with omission error rates lessthan 20% was unexpected. Apparent importance of the contrastbetween TM bands 5 and 7 (Table 7) may indicate an associationbetween palustrine wetlands and hydrothermally altered miner-als. Hellman and Ramsey (2004) found that ASTER band 4/6ratio (equivalent to TM bands 5/7) greater than one wereassociated with hydrothermal areas in Yellowstone NationalPark.While Yellowstone is notable for its hydrothermal features,

Page 13: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

Table 9Sample error matrix for the 1988 TM+TXT+DEM+GIS palustrine wetlandtype classification tree

Predicted class

PAB PEM PFO PSS PUS Omissionerror (%)

PAB 471 497 125 59 4 59.26PEM 134 63,182 3885 956 617 8.13PFO 84 5288 16,728 439 3 24.79PSS 34 2047 133 3008 10 42.51PUS 3 695 8 0 1890 27.20Commissionerror (%)

35.12 11.89 19.88 27.72 25.12 Overallerror (%)14.72

Table 8Error rates for TM+TXT+DEM+GIS palustrine wetland type classification trees by year

Omission error (%) Commission error (%) Overall error (%)

PAB PEM PFO PSS PUS PAB PEM PFO PSS PUS

1988 59.26 8.13 24.79 42.51 27.20 35.12 11.89 19.88 27.73 25.12 14.721997 75.35 10.12 29.29 54.30 39.25 49.82 14.74 26.75 30.66 21.85 18.211999⁎ 77.77 8.62 34.66 39.05 27.85 54.19 14.38 22.50 34.29 22.83 17.302000⁎ 55.02 10.47 31.58 39.43 24.65 42.86 13.54 25.36 32.83 29.28 17.562001⁎ 82.00 8.93 29.77 38.84 26.89 51.29 13.42 22.87 31.33 18.33 16.442002⁎ 71.28 9.90 25.79 37.29 32.20 41.55 12.14 24.50 34.98 12.13 16.162003 61.51 10.59 32.83 45.01 27.31 41.22 14.35 27.11 31.47 31.53 18.36Mean 68.88 9.54 29.82 42.35 29.34 45.15 13.49 24.14 31.90 23.01 16.96s.e.m. 3.90 0.37 1.36 2.22 1.86 2.56 0.42 0.97 0.92 2.48 0.49

⁎Years for which two dates of imagery were available.

594 C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

they occupy only a small portion of the landscape (approxi-mately 1%). Perhaps hydrothermally-altered minerals are dis-tributed widely enough beyond these areas to generate a wetlandsignal in TM bands 5 and 7 across the park. It is also interestingthat TM bands 5 and 7 were selected as individual predictorsonly in 1997 and 1999, and at very low levels of classificationtrees. Rules generated by the S-PLUS classification treealgorithm operate on only one variable at a time. In this case, alinear combination of TM bands 5 and 7, as manifested incanonical discriminant scores, was an important predictor, ratherthan either band individually.

Although classification trees made use of per-pixel differ-ences and other spectral information from two dates whenavailable, mostly in palustrine wetland type models (e.g.,Fig. 8c), availability of a second TM image did not improvepalustrine wetland detection or classification (Tables 4 and 8).Apparently, other covariates were just as effective predictors asany temporal signal between mid- and late summer.

TM band 4 texture and tasseled-cap greenness texture werethe dominant texture measures used to discriminate palustrinewetland from upland (Figs. 4 and 5). Discontinuities in bothvegetation type and condition across wetland/upland boundariesresulted in greater variances within moving windows straddlingthose boundaries. Image texture was also useful in classifyingpalustrine wetland types (Figs. 7 and 8), with the PFO classgenerally exhibiting less texture. The effectiveness of the simplevariance filter demonstrates the importance of spatial context inwetland remote sensing. Further wetland mapping efforts inYellowstone National Park should explore the question ofcontext more deeply, either in terms of more sophisticatedtexture measures derived from grey-level co-occurrence matri-ces (e.g., Chiu & Couloigner, 2004; Dechka et al., 2002), or inmodels where spatial context is modeled explicitly, e.g., Markovrandom fields (Tso & Mather, 2001).

Habitat type, cover type, and landform type were the mosteffective thematic GIS predictors (Figs. 4, 5, 7, and 8). Inwetland/upland classification trees, habitat types were broadlystratified in terms of wetness, with non-forest wet, non-forestwillow, subalpine fir wet, thermal vegetation, water, and wetforest typically separated frommore mesic habitat types (Fig. 5).Landscape hydrologic processes are reflected in the associationof palustrine wetlands with alluvial, hydrothermal, and glacio-

fluvial landforms. Alluvial landforms include alluvial fans,basins, and floodplains, while glaciofluvial landforms wereformed by glacial meltwaters (Shovic, 1996). The habitat typevariable was particularly useful for classifying PSS wetlands, asit includes a non-forest willow class with Salix species as thepredicted climax vegetation (Fig. 8). Cover type informationimproved classification of PFO wetlands by separating forestfrom non-forest.

One potential hazard in using ancillary vegetation informa-tion to classify palustrine wetland types is the possibility thatclassification trees will simply replicate existing vegetationlayers, say, assigning most pixels to the PSS class within thenon-forest willow habitat type. In an example along theYellowstone River, an extensive area of non-forest willowhabitat is identified, with a large patch of wet forest to the west(Fig. 11). Within wet forest, relatively equal proportions areclassified as PFO or PEM pixels, interspersed with upland. PSSpixels occur largely along river channels, while mostly PEMwetland is predicted over the rest of the non-forest willowhabitat type. Thus, while we found that ancillary vegetation dataimproved classification of PFO and PSS wetlands (Fig. 6),classification trees did not simply replicate that information. It isimportant to recall that the habitat type layer indicates potentialclimax vegetation, not existing vegetation. Although non-forestwillow and wet forest habitat types have greater probabilities ofcontaining PSS and PFO wetlands, respectively, classificationtrees used other predictors in combination with this information.

Page 14: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

Table 10Kappa coefficients for individual palustrine wetland type models by year and Z-statistics for increases in Kappa coefficients when additional predictors are added tomodels in any given year

Model Z-statistic for Kappa differences

TM TM+TXT TM+TXT+DEM TM+TXT+DEM+GIS

(TM+TXT) (TM+TXT+DEM) (TM+TXT+DEM+GIS)

Kappa Standarderror

Kappa Standarderror

Kappa Standarderror

Kappa Standarderror

– (TM) – (TM+TXT) – (TM+TXT+DEM)

1988 0.5241 2.25E-03 0.6045 2.11E-03 0.6113 2.21E-03 0.6782 2.11E-03 26.12⁎ 2.20⁎⁎⁎ 21.89⁎

1997 0.4158 2.46E-03 0.4322 2.49E-03 0.4504 2.46E-03 0.5966 2.30E-03 4.69⁎ 5.20⁎ 43.45⁎

1999 0.4564 2.43E-03 0.5138 2.35E-03 0.5227 2.36E-03 0.6157 2.27E-03 16.98⁎ 2.67⁎⁎ 28.43⁎

2000 0.4638 2.43E-03 0.5082 2.39E-03 0.5263 2.39E-03 0.6195 2.25E-03 13.06⁎ 5.35⁎ 28.37⁎

2001 0.5046 2.42E-03 0.5679 2.35E-03 0.5843 2.26E-03 0.6381 2.21E-03 18.77⁎ 5.02⁎ 17.04⁎

2002 0.5159 2.36E-03 0.5621 2.32E-03 0.5908 2.27E-03 0.6511 2.16E-03 13.97⁎ 8.84⁎ 19.27⁎

2003 0.4067 2.53E-03 0.4589 2.53E-03 0.4764 2.50E-03 0.5989 2.30E-03 14.59⁎ 4.90⁎ 36.12⁎

⁎Increase in Kappa coefficient statistically significant at the p=0.01 level, ⁎⁎p=0.05 level, ⁎⁎⁎p=0.10 level.

595C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

Terrain information, specifically slope, improved identifica-tion of upland pixels (Fig. 3), but was not useful fordiscriminating among palustrine wetland types (Fig. 6). Notsurprisingly, higher slope gradients identified elements of thelandscape where wetland occurrence was highly unlikely. Withrespect to soil properties, palustrine wetlands were typicallyassociated with lower clay content and rhyolitic parent materi-als. This result supports the observation of Despain (1990) thatwetlands in Yellowstone National Park tend to form in coarse-textured soils with limited water-holding capacities.

We were somewhat disappointed that the topographicwetness index was not an important wetland predictor (selectedby only the 1988 wetland/upland classification tree), given thecomputational effort required to generate this layer (Planchon &Darboux, 2002; Wilson & Gallant, 2000). However, thetopographic wetness index is based on a steady-state assump-tion that the entire upstream area contributes to water flowreaching any arbitrary point on the landscape (Wilson &

Fig. 7. Average deviance reduction by variable for 1988, 1997, and 1999–2003TM+TXT+DEM+GIS palustrine wetland type classification trees (+/−s.e.m.).

Gallant, 2000). In drier environments, subsurface flow may bevery slow, with individual points receiving water from only asmall proportion of their upstream contributing area. In thiscase, the steady-state assumption would not hold, and highervalues of the topographic wetness index would not necessarilycorrespond to saturated soils (Wilson & Gallant, 2000).Despite the fact that much of Yellowstone National Park isrelatively arid, we had hoped that classification trees wouldstratify the landscape in a way that identified areas where thetopographic wetness index applied. Apparently, this was notthe case.

With respect to applying the developed approach to large-area wetland mapping in other settings, some ancillary data thatwere useful in Yellowstone National Park may not be widelyavailable, particularly vegetation and landform classifications.Our results suggest that unavailability of thematic GIS layerswould negatively impact separation of different wetland types(Fig. 6) more than wetland detection (Fig. 3). Additionally,ancillary data availability will likely be an even greater concernin situations where TM-based predictors, alone, are not aseffective as we found in Yellowstone National Park.

4.2. Wetland over-prediction and class imbalance

High commission error rates for palustrine wetland pixels(Fig. 3b), and apparent wetland over-prediction, are a concern.Spectral confusion and limited spatial resolution of the TMsensor have been documented in other wetland studies (FederalGeographic Data Committee, 1992; Johnston & Barston, 1993;Lunetta & Balogh, 1999) and are likely factors contributing towetland over-prediction. However, the relative rarity ofpalustrine wetlands in Yellowstone National Park (less than6% cover) also plays a role. In 1988, for example, wetland andupland pixels were classified with low omission error rates, lessthan 9% in the wetland class and less than 7% in the uplandclass. But the overwhelming prevalence of upland in the studyarea led to a large estimate of upland pixels erroneously clas-sified as wetland relative to correctly identified wetland pixels(Table 5). In order to achieve a more acceptable wetland com-mission error rate in 1988, say 20%, while maintaining existing

Page 15: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

596 C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

wetland detectability, upland omission error rates would have tobe reduced to 1.3%, an accuracy likely not achievable withmoderate spatial- and spectral-resolution TM data, regardless ofancillary data availability or classifier sophistication. Whilehigh spatial-resolution satellite data, e.g., sub-5 m, mightachieve a better tradeoff between wetland detection and over-prediction, it has been applied to wetland remote sensing onlyover small areas (Dechka et al., 2002 and Lawrence et al.,2006). Across large extents, wetland mapping with high-resolution satellite imagery would likely be cost-prohibitive and

Fig. 8. Simple palustrine wetland type classification trees for 1988 (dry year), 1997 (wrule are sent to the left; otherwise, pixels are sent to the right.

logistically difficult (particularly with respect to data volume ifancillary data is also used).

The negative effect of class imbalance that we observe is acommon problem in the field of data mining (Weiss, 2004),where an event of interest is often quite rare (e.g., credit cardfraud, or an unusual medical diagnosis), there is a high valueassociated with detecting these rare instances, but also asubstantial cost incurred when reacting to false-positives. Oneapproach to this dilemma is directed over-sampling or under-sampling of training data such that errors of omission and

et year), and 2001 (two TM images). At each node, pixels satisfying the indicated

Page 16: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

Fig. 8 (continued ).

597C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

commission are adjusted to desired levels (Weiss, 2004). In ourcase, one way to reduce wetland over-prediction would be todecrease the proportion of wetland pixels in training samples,with an expectation of increased accuracy in the upland classand fewer wetland false-positives. However, this would also

Fig. 9. 1 August 2003 Landsat-5 TM bands 7, 4, 2 in the vicinity of the Grand Canyofires; more recent fire scars appear red. Forest is brownish/darker green, non-forestedlight blue.

increase the wetland omission error rate. For example, using thesame 2003 TM image used in this study and similar ancillarydata, we found that decreasing the proportion of wetland pixelsto 0.175 of the training sample reduced the wetland commissionerror rate to approximately 40%, while simultaneously tuning

n of the Yellowstone River. Pink areas are fire scars from the 1988 catastrophicareas are light green, open water is dark blue, and geothermal areas appear white/

Page 17: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

Fig. 10. Palustrine wetland probability surface and predicted palustrine wetland classes in the vicinity of the Grand Canyon of the Yellowstone River for 1 August2003.

598 C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

the wetland omission error rate to a target value of 20% (Wright,2004).

Other authors, to the best of our knowledge, have notaddressed the problem of class imbalance in wetland remotesensing. A number of studies report low omission error rates forboth wetland and upland classes (typically less than 10%), but donot report commission errors (Harvey & Hill, 2001; Töyrä et al.,2001;Wang et al., 1998). As we have shown, significant wetlandover-prediction can occur even if high producer's accuracies areachieved in all classes. In two studies where more balanced errorrates were found (Lunetta & Balogh, 1999; Sader et al., 1995),wetlands occupied a greater proportion of the landscape (more

than 20%), and class imbalance may have been less problematic.In general, however, we would recommend that potentialwetland over-prediction resulting from class imbalance beexplicitly addressed in all wetland remote-sensing studies.

InYellowstoneNational Park, we collaboratedwith amphibianresearchers interested in locating potential wetland breedinghabitat, both to assist amphibian field surveys (Peterson et al.,2005) and in modeling the geographic distribution of amphibianpopulations (Spear et al., 2005). These researchers were lessconcerned with accurate delineation of wetland boundaries thanin identifying areas where errors of omission in the YellowstoneNWI weremost likely. From this perspective, wetland probability

Page 18: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

Fig. 11. Habitat types and predicted palustrine wetland classes for 1 August 2003 along a stretch of the Yellowstone River on the southern border of YellowstoneNational Park.

599C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

surfaces were more informative than hard classification ofindividual pixels. Field crews conducted amphibian surveys inareas where classification trees predicted elevated probabilities ofpalustrine wetland occurrence but the NWI did not identifywetlands (Charles Peterson, personal communication). Spearet al. (2005) used probabilities ofwetland occurrence as an index

of the physiological cost of amphibian movement across thelandscape (high wetland probability equals low movement cost,low probability equals high cost). Least-cost paths determinedusing this index explained genetic relatedness of spatially distinctsub-populations of amphibians, and included pixels withprobabilities of wetland occurrence both greater- and less-than

Page 19: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

600 C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

the 0.5 threshold for hard-classification as wetland, i.e. it wasinformative to know how ‘like’ a palustrine wetland a pixel was(in a fuzzy-classification sense).

Comparing classification tree predictions with the Yellow-stone NWI near the Grand Canyon of the Yellowstone River(Fig. 10), we see that the 2003 model does a good jobgenerating a broad-stroke approximation of the NWI, largelydiscriminating PEM wetland from burned and intact forest(Fig. 9). Beyond the NWI, the model predicts additional wet-land, mostly contiguous with NWI polygons, but also in isolat-ed, smaller clusters of pixels. Forested wetlands are difficult toidentify by both aerial photography (Tiner, 1990) and satelliteremote sensing (Lunetta & Balogh, 1999; Sader et al., 1995). Asubstantial number of PFO pixels are predicted around themargins of PEM wetlands (Fig. 10), a likely place to findforested wetlands not previously identified by the NWI.

In an another example from the Northern Range of Yellow-stone National Park (Figs. 12 and 13), apparent wetland over-prediction is more extensive across a mostly non-forestedlandscape of alluvial bottomlands surrounded by glaciatedrolling uplands and plateaus. Spectrally, we see that most lowerand middle-elevation vegetation was quite dry as of 1 August2003 (Fig. 12). Valley-bottom wetlands, like the large wetlandcomplex along Slough Creek, stand out as green patches andlargely coincide with NWI-delineated wetlands. Outside of NWIpolygons, elevated probabilities of palustrine wetland occur-rence are predicted in pixels that were still green as of mid-summer, including various-sized patches within valley bottomsand along minor drainage courses descending from surrounding

Fig. 12. 1 August 2003 Landsat-5 TM bands 7, 4, 2 on the Northern Range of Yellowand/or senescent grasses and forbs) and light green (photosynthetically active grassegeothermal areas are white/light blue.

uplands (Figs. 12 and 13). An extensive area of apparent wetlandover-prediction can be seen in the upper right of Fig. 13, in thiscase, across a glaciated plateau landform at higher elevation,where we would expect more favorable soil moisture conditionsas of mid-summer. However, spectral data was not usedindiscriminately. The majority of pixels appearing to containphotosynthetically active, non-forested vegetation in the rawimage (Fig. 12) were subsequently classified as upland (Fig. 13).

These two examples illustrate a key shortcoming of using thenow 20-year old Yellowstone NWI to evaluate model accuracy.According to the NWI, classification trees badly over-predictpalustrine wetlands on a per-pixel basis. But these same predic-tions also appear to capture essential aspects of landscapehydrology, i.e., identifying areas where one might logicallyexpect to find wetlands not included in the NWI. The extent towhich wetland/upland models may actually be detecting previ-ously unidentified wetlands, or wetlands that formed followingacquisition of aerial photos in the 1980s, is obviously unknown.However, field-verification of a different wetland model forYellowstone National Park showed that nearly 30% ofpalustrine wetland pixels predicted outside of NWI polygonsactually were wetland (Wright, 2004), demonstrating a robustability to detect wetlands not included in the Yellowstone NWI,albeit while still paying a penalty in terms of over-prediction.

Looking forward, balancing wetland over-prediction againstfailures to detect wetlands will be a key issue in operationalwetland mapping using TM imagery. Data mining algorithmsemploying ensembles of classification trees, like random forests(Breiman, 2001) or stochastic gradient descent (Friedman,

stone National Park. Non-forested areas appear pink (sagebrush Artemesia spp.s, forbs, and sedges), forest is darker green, open water appears dark blue, and

Page 20: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

Fig. 13. Palustrine wetland probability surface and predicted palustrine wetland classes across the Northern Range for 1 August 2003.

601C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

2001), may be less susceptible to class imbalance than singleclassification trees (Chen et al., 2004; Lawrence et al., 2004).However, ensemble-based classifiers lack the interpretability ofsingle trees. As this study was largely exploratory, the ability todirectly interpret models was important. Future application ofless-transparent, ensemble-based methods might improvewetland-mapping efforts, and this is certainly an avenue forfurther research.

4.3. Discrimination of palustrine wetland types

High error rates in the PSS class for models using only TM-and DEM-derived predictors (Fig. 6) were comparable to poor

classification of shrub wetlands found by other researchers(Lunetta & Balogh, 1999). And while ancillary vegetation datagreatly enhanced discrimination of the PSS class it was oftenconfused with the predominant PEM class (Table 9). Results forPFO wetlands composed of coniferous tree species were similarto classification accuracies for forested wetlands containingdeciduous/coniferous mixtures in the eastern United States(Lunetta & Balogh, 1999; Sader et al., 1995), but, again,confusion with the PEM class was common (Table 9). Otherauthors report improved classification of woody wetlands usingimagery from more than one date, e.g., adding an early spring,leaf-off image to identify inundated soils (Lunetta & Balogh,1999). However, with the exception of scrub–shrub wetlands,

Page 21: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

602 C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

deciduous species are not common in Yellowstone wetlands.And in the PSS class, it would be difficult to acquire early-spring, leaf-off imagery that did not contain extensive snowcover.

The small size of aquatic bed wetlands in YellowstoneNational Park likely contributed to high error rates in the PABclass (Fig. 6), as we would anticipate a large proportion ofmixed PAB pixels within training and test datasets (median sizeof PAB polygons in the Yellowstone NWI is approximately1 ha, with a maximum of only 8.7 ha). Relative to the PEM andPFO classes, PAB pixels were also quite rare in trainingdatasets. However, the PUS class was classified with muchlower error rates (Fig. 6), even though relatively few trainingpixels were allocated to it. In this case, unique reflectancecharacteristics of unconsolidated shore wetlands were compen-satory. These wetlands are typically found near active hydro-

Fig. 14. Time series of 1999, 2001, and 2003 palustrine

thermal features, where siliceous sinters reflect strongly in thevisible spectra (Hellman & Ramsey, 2004).

4.4. Palustrine wetland monitoring

With respect to wetland monitoring in Yellowstone NationalPark, we are most interested in identifying spatially coherent,multi-pixel discrepancies between classification tree predictionsand the Yellowstone NWI. In an important way, the developedmethod nicely complements the existing NWI layer. If a baseNWI map is digitally available, it can be used to rapidlyassemble large training datasets. Over the course of a wetland-monitoring program, it may be necessary to acquire morecurrent data to develop new base maps, perhaps from aerialphotography or high-resolution satellite data, as the signal-to-noise ratio of training data derived from older maps declines.

wetland probability surfaces in the Lamar Valley.

Page 22: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

603C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

We envision large-area wetland monitoring proceeding in a two-step fashion, with frequent, low-cost, and moderate-resolutionwetland mapping interspersed with infrequent and relativelyexpensive finer-resolution efforts.

Amphibian researchers in Yellowstone National Park areinterested in relating amphibian population dynamics toclimate-driven wetland variability (Charles Peterson, personalcommunication). In an example from the Lamar Valley, declin-ing probabilities of wetland occurrence are observed over the1999–2003 drought, both within and surrounding existing NWIpolygons (Fig. 14). From an amphibian perspective (and alsowith respect to other wetland-dependent wildlife), detection ofchanging wetland condition is more important than delineatingexact wetland boundaries at any given time. A substantialchallenge in working with wetlands as a land cover type is that,by definition, a location might be a wetland (i.e., have at leastone characteristic of wetland under the Cowardin system), whilenot actually being wet, or functioning like a wetland, at the timeof image acquisition. This subtlety is ignored when evaluatingmodels on the basis of per-pixel error rates, but we clearly seein Fig. 14 that a number of NWI polygons are no longerdistinguishable from upland by 2003.

Continued development of a wetland monitoring capabilityin Yellowstone National Park should focus on field assessmentof the ability to detect both changes in wetland extent andchanges in wetland condition. Establishment of permanentwetland plots in Yellowstone would aid this effort, and providea source of field-verified training data for future models, asopposed to relying solely on the dated NWI. Field efforts shouldalso evaluate the ability of models to detect wetlands smallerthan 1 ha.

5. Conclusions

Could a classification tree approach combining TM imageryand ancillary data be applied to palustrine wetland mapping andmonitoring in other settings and over larger extents? Highlysignificant increases in Kappa coefficients (Tables 5 and 10)demonstrate the value of using classification trees to combinevery different types of information. Model consistency and lowomission error rates in the combined palustrine wetland andPEM classes are encouraging. Our results across YellowstoneNational Park suggest that accuracies of wetland classes in acontinent-wide effort like the NLCD (Stehman et al., 2003;Wickham et al., 2004) could be improved by using our approachto develop wetland/upland masks. However, wetland over-prediction and difficulty discriminating palustrine wetland typesraise legitimate concerns about the spatial and spectral resolu-tion of the TM instrument. Given these limitations, we wouldnot expect our TM-based method to provide crisp delineationbetween wetland and upland over large areas. Wetland prob-ability surfaces will likely prove to be more informative thanper-pixel classification, particularly as a means to better capturespatial and temporal variability of palustrine wetlands. Wheremore precise results are required, say in legally defining wetlandboundaries for regulatory purposes (Federal InteragencyCommittee for Wetland Delineation, 1989), more precise data

and methods of wetland delineation will be required. Effects ofclass imbalance should always be considered explicitly. Withinthese bounds, the classification tree approach we tested inYellowstone National Park is portable, relatively easy toimplement, and could prove widely applicable to a demonstra-ble need for affordable, large-area wetland mapping.

Acknowledgments

This work was funded by the USGS Amphibian Researchand Monitoring Initiative. We thank Daniel Goodman (MontanaState University) for serving as the first author's dissertationadvisor. Charles Peterson (Idaho State University), RobertKlaver (USGS EROS), and Paul Bartelt (USGS EROS andWaldorf College) provided invaluable assistance over thecourse of this work. Two anonymous reviewers contributedhelpful reviews of the original manuscript.

The use of firm, trade, and brand names in this article is foridentification purposes only, and does not constitute endorse-ment by the U.S. Government.

References

Anderson, J. R., Hardy, E. E., Roach, J. T., & Witmer, R. E. (1976). A land useand land cover classification system for use with remote sensor data. U.S.Geological Survey Circular, 671 (Washington, DC).

Blaustein, A. R., Wake, D. B., & Sousa, W. P. (1994). Amphibian declines:Judging stability, persistence, and susceptibility of populations to local andglobal extinctions. Conservation Biology, 8, 60−71.

Brenner, A., & Palmer, M. (2005). Change analysis program monitors impact ofcoastal development. Earth Imaging Journal, 2, 16−20.

Breiman, L. (2001). Random forests. Machine Learning, 45, 5−32.Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classifi-

cation and regression trees. Boca Raton, FL: Chapman and Hall.Carter, V. (1999). Technical aspects of wetlands — Wetland hydrology, water

quality, and associated functions. U.S. Geological Survey Water-SupplyPaper, 2425.

Chen, C., Liaw, A., & Breiman, L. (2004). Using random forest to learnimbalanced data. Technical report, Vol. 666. Berkeley, California: StatisticsDepartment, University of California at Berkeley.

Chiu,W., &Couloigner, I. (2004). Evaluation of incorporating texture into wetlandmapping from multispectral images. EARSeL Proceedings, 3, 363−371.

Clark, L. A., & Pregibon, D. (1992). Tree-based models. In J. M. Chambers, &T. J. Hastie (Eds.), Statistical models in S (pp. 377−419). Pacific Grove, CA:Wadsworth and Brooks/Cole.

Congalton, R. G., & Green, K. (1999). Assessing the accuracy of remotelysensed data: Principles and practices. Boca Raton, FL: CRC Press.

Cowardin, L. M., Carter, V., Golet, F. C., & Laroe, E. T. (1979). Classification ofwetlands and deepwater habitats of the United States.Washington, DC: U.S.Fish and Wildlife Service.

Cowardin, L. M., & Golet, F. C. (1995). U.S. Fish and Wildlife Service 1979wetland classification: A review. Vegetatio, 118, 134−15.

Dahl, T. E. (2000). Status and trends of wetlands in the conterminous UnitedStates 1986 to 1997. Fish and Wildlife Service, Washington, DC: U.S.Department of the Interior.

Dechka, J. A., Franklin, S. E., Watmough, M. D., Bennett, R. P., & Ingstrup,D. W. (2002). Classification of wetland habitat and vegetation commu-nities using multi-temporal Ikonos imagery in southern Saskatchewan.Canadian Journal of Remote Sensing, 28, 679−685.

Despain, D. (1990). Yellowstone vegetation: Consequences of environment andhistory in a natural setting. Boulder, CO: Roberts Rinehart.

Elliot, C. R., & Hektner, M. M. (2000). Wetland resources of YellowstoneNational Park. Mammoth, WY: Yellowstone National Park, Center forResources.

Page 23: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

604 C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

Ernst-Dottavio, C. L., Hoffer, R. M., & Mrocynski, R. P. (1981). Spectralcharacteristics of wetland habitats. Photogrammetric Engineering andRemote Sensing, 47, 223−227.

Federal Geographic Data Committee (1992). Application of satellite data formapping and monitoring wetlands: Fact finding report. Reston, VA:Wetlands Subcommittee.

Federal Interagency Committee for Wetland Delineation. (1989). FederalManual for Identifying and Delineating Jurisdictional Wetlands. Washing-ton, D.C.: U.S. Army Corps of Engineers, U.S. Environmental ProtectionAgency, U.S. Fish and Wildlife Service, and U.S.D.A. Soil ConservationService.

Finlayson, C.M.,&Davidson,N.C. (1999).Global reviewofwetland resources andpriorities for wetland inventory: Summary report. In C. M. Finlayson & A. G.Spiers (Eds.), Global review of wetland resources and priorities for wetlandinventory. Canberra, Australia: Supervising Scientist.

Friedman, J. (2001). Greedy function approximation: A gradient boostingmachine. Annals of Statistics, 29, 1189−1232.

Friedl, M. A., & Brodley, C. E. (1997). Decision tree classification of land coverfrom remotely sensed data. Remote Sensing of Environment, 61, 399−409.

Gibbs, J. P. (2000). Wetland loss and biodiversity conservation. ConservationBiology, 14, 314−317.

Gorham, E. (1991). Northern peatlands: Role in the carbon cycle and probableresponses to climatic warming. Ecological Applications, 1, 182−195.

Hansen, M., Dubayah, R., & Defries, R. (1996). Classification trees: Analternative to traditional land cover classifiers. International Journal ofRemote Sensing, 17, 1075−1081.

Harvey, K. R., & Hill, G. J. E. (2001). Vegetation mapping of a tropicalfreshwater swamp in the Northern Territory, Australia: A comparison ofaerial photography, Landsat TM and SPOT imagery. International Journalof Remote Sensing, 22, 2911−2925.

Hellman, M. J., & Ramsey, M. S. (2004). Analysis of hot springs and associateddeposits in Yellowstone National Park using ASTER and AVIRIS remotesensing. Journal of Volcanology and Geothermal Research, 135, 195−219.

Houhoulis, P. F., & Michener, W. K. (2000). Detecting wetland change: A rule-based approach using NWI and SPOT-XS data. PhotogrammetricEngineering and Remote Sensing, 66, 205−211.

Huang, X., & Jensen, J. R. (1997). A machine-learning approach to automatedknowledge-base building for remote sensing image analysis with GIS data.Photogrammetric Engineering and Remote Sensing, 63, 1185−1194.

Huang, C., Wylie, B., Yang, L., Homer, C., & Zylstra, G. (2002). Derivation of atasseled-cap transformation based on the Landsat 7 at-satellite reflectance.International Journal of Remote Sensing, 23, 1741−1748.

Jensen, J. R., Rutchey, K., Koch, M. S., & Narumalani, S. (1995). Inlandwetland change detection in the Everglades water conservation area 2Ausing a time series of normalized remotely sensed data. PhotogrammetricEngineering and Remote Sensing, 61, 199−209.

Johnston, R. M., & Barston, M. M. (1993). Remote sensing of Australianwetlands: An evaluation of Landsat TM data for inventory and classification.Australian Journal of Marine Freshwater Resources, 44, 235−252.

Johnson,W.C.,Millett, B.V.,Gilmanov, T., Voldseth,R.A.,Guntenspergen,G. R.,& Naugle, D. E. (2005). Vulnerability of northern prairie wetlands to climatechange. BioScience, 55, 863−872.

Kudray, G. M., & Gale, M. R. (2000). Evaluation of National Wetland Inventorymaps in a heavily forested region in the upper Great Lakes. Wetlands, 20,581−587.

Kuzila, M. S., Rundquist, D. C., & Green, J. A. (1991). Methods for estimatingwetland loss: The rainbasin region of Nebraska, 1927–1981. Journal of Soiland Water Conservation, 44, 441−445.

Lawrence, R., Bunn, A., Powell, S., & Zambon, M. (2004). Classification ofremotely sensed imagery using stochastic gradient boosting as a refinementof classification tree analysis. Remote Sensing of Environment, 90,331−336.

Lawrence, R. L., Hurst, R., Weaver, T., & Aspinall, R. (2006). Mapping PrairiePothole communities with multitemporal IKONOS satellite imagery. Pho-togrammetric Engineering and Remote Sensing, 72, 169−174.

Lawrence, R. L., & Wright, A. (2001). Rule-based classification systems usingclassification and regression tree (CART) analysis. PhotogrammetricEngineering and Remote Sensing, 67, 1137−1142.

Leahy, S. (2003). Wetlands from space: The national wetland inventory. Con-servator, 24, 13−16.

Loveland, T. R., Reed, B. C., Brown, J. F., Ohlen, D. O., Zhu, J., Yang, L., et al.(2000). Development of a global land cover characteristics database andIGBP DISCover from 1-km AVHRR data. International Journal of RemoteSensing, 21, 1303−1330.

Lunetta, R. S., & Balogh, M. E. (1999). Application of multi-temporal Landsat 5TM imagery for wetland identification. Photogrammetric Engingeering andRemote Sensing, 65, 1303−1310.

National Research Council (2001). Compensating for wetland losses under theclean water act. Washington, DC: National Academy Press.

Peddle, D. R., & Duguay, C. R. (1995). Incorporating topographic and climaticGIS data into satellite image analysis of an alpine tundra ecosystem, frontrange, Colorado Rocky Mountains. Geocarto International, 10, 43−60.

Peterson, C. R., Burton, S. R., & Patla, D. A. (2005). Geographical informationsystems and survey designs. In Michael Lannoo (Ed.), Amphibian declines:The conservation status of United States species (pp. 320−325). Berkeley,CA: University of California Press.

Planchon, O., & Darboux, F. (2002). A fast, simple and versatile algorithm to fillthe depressions of digital elevation models. Catena, 46, 159−176.

Poiani, K. A., Johnson, W. C., Swanson, G. A., & Winter, T. C. (1996). Climatechange and northern prairie wetlands: Simulations of long-term dynamics.Limnology and Oceanography, 41, 871−881.

Ramsar Convention Bureau (2004). The list of wetlands of internationalimportance: 16 September 2004. Obtained October 27, 2004 from http://ramsar.org/sitelist.pdf

Rodman, A., Shovic, H. F., & Thoma, D. (1996). Soils of Yellowstone NationalPark. Mammoth, WY: Yellowstone National Park, Yellowstone Center forResources.

Rutchey, K., & Vilcheck, L. (1994). Development of an everglades vegetationmap using a SPOT image and the global positioning system. Photogram-metric Engineering and Remote Sensing, 60, 767−775.

Sader, S. A., Ahl, D., & Liou, W. -S. (1995). Accuracy of Landsat-TM and GISrule-based methods for forest wetland classification in Maine. RemoteSensing of Environment, 53, 133−144.

Shovic, H. F. (1996). Landforms and associated surficial materials ofYellowstone National Park. Mammoth, WY: Yellowstone National Park,Yellowstone Center for Resources.

Sohl, T. L., Gallant, A. L., & Loveland, T. R. (2004). The characteristics andinterpretability of land surface change and implications for project design.Photogrammetric Engineering and Remote Sensing, 70, 439−448.

Spear, S. F., Peterson, C. R., Matocq, M. D., & Storfer, A. (2005). Landscapegenetics of the blotched tiger salamander (Ambystoma tigrinum melanos-tictum). Molecular Ecology, 14, 2553−2564.

Stehman, S. V. (1999). Estimating the Kappa coefficient and its variance understratified random sampling. Photogrammetric Engineering and RemoteSensing, 62, 401−407.

Stehman, S. V., Wickham, J. D., Smith, J. H., & Yang, L. (2003). Thematicaccuracy of the 1992 National Land-Cover Data for the eastern UnitedStates: Statistical methodology and regional results. Remote Sensing ofEnvironment, 86, 500−516.

Stolt, M. H., & Baker, J. C. (1995). Evaluation of National Wetland Inventorymaps to inventory wetlands in the southern Blue Ridge of Virginia. Wet-lands, 15, 346−353.

Swanson, R. B., Miller, K. A., Woodruff, R. E., Laidlaw, G. A., Watson, K. R.,& Clark, M. L. (2004). Summary of hydrologic conditions. Water resourcesdata for Wyoming, water year 2003: Volume 1. Surface-water dataCheyenne, WY: U.S. Geological Survey, Water Resources Division.

Swarthout, D. J., MacConnell, W. P., & Finn, J. T. (1981, Aug 9–14). Anevaluation of the National Wetland Inventory in Massachusetts. In-placeresource inventories: Principles and practices, proceedings of a nationalworkshop (pp. 685−691). Orono, ME: University of Maine.

Thompson, S. K. (1992). Sampling. New York: Wiley.Tiner, R. W. (1990). Use of high-altitude aerial photography for inventorying

forested wetlands in the United States. Forest Ecology and Management,33/34, 593−604.

Tiner, R. W. (1997). NWI maps: What they tell us. National WetlandsNewsletter, 19, 7−12.

Page 24: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

605C. Wright, A. Gallant / Remote Sensing of Environment 107 (2007) 582–605

Töyrä, J., Pietroniro, A., & Martz, L. M. (2001). Multisensor hydrologicassessment of a freshwater wetland. Remote Sensing of Environment, 75,162−173.

Tso, B., & Mather, P. M. (2001). Modelling context using Markov randomfields. Classification methods for remotely sensed data (pp. 230−270).London: Taylor and Francis.

U.S. Fish and Wildlife Service (2002). National wetlands inventory: A strategyfor the 21st century. Washington, DC: U.S. Department of the Interior, Fishand Wildlife Service.

Van der Valk, A. G. (1981). Succession in wetlands: A Gleasonian approach.Ecology, 62, 688−696.

Venables, W. N., & Ripley, B. D. (1994). Tree-based methods. Modern appliedstatistics with S-Plus (pp. 329−347). New York: Springer-Verlag.

Vogelmann, J. E., Howard, S. M., Yang, L., Larson, C. R., Wylie, B. K., & VanDriel, N. (2001). Completion of the 1990s National Land Cover Dataset forthe conterminous United States from Landsat Thematic Mapper data andancillary data sources. Photogrammetric Engineering and Remote Sensing,67, 650−651.

Vogelmann, J. E., Sohl, T., & Howard, S. M. (1998). Regional characterizationof land cover using multiple sources of data. Photogrammetric Engineeringand Remote Sensing, 64, 45−57.

Wang, J., Shang, J., Brisco, B., & Brown, R. J. (1998). Evaluation of multidateERS-1 and multispectral Landsat imagery for wetland detection in southernOntario. Canadian Journal of Remote Sensing, 24, 60−68.

Weiss, G. M. (2004). Mining with rarity: A unifying framework. SigkddExplorations, 6, 7−19.

Western Regional Climate Center (2004). Retrieved October 28, 2004 from http://www.wrcc.dri.edu/cgi-bin/cliMAIN.pl?wyyell

Wickham, J. D., Stehman, S. V., Smith, J. H., & Yang, L. (2004). Thematicaccuracy of the 1992 National Land-Cover Data for the western UnitedStates. Remote Sensing of Environment, 91, 452−468.

Wilen, B. O., & Bates, M. K. (1995). The U.S. Fish and Wildlife Service'sNational Wetlands Inventory project. Vegetatio, 119, 153−169.

Wilson, P. A. (1997). Rule-based classification of water in Landsat MSS imagesusing the variance filter. Photogrammetric Engineering and RemoteSensing, 63, 485−491.

Wilson, J. P., & Gallant, J. C. (2000). Secondary topographic attributes. In J. P.Wilson & J.C. Gallant (Eds.), Terrain analysis: Principles and applications(pp. 87−131). New York: Wiley.

Winter, T. C. (2000). The vulnerability of wetlands to climate change: Ahydrologic landscape perspective. Journal of the American Water ResourcesAssociation, 36, 305−311.

Work, E. A., & Gilmer, D. S. (1976). Utilization of satellite data for inventoryingprairie ponds and potholes. Photogrammetric Engineering and RemoteSensing, 5, 685−694.

Wright, C. K. (2004). Remote Sensing of Wetlands in Yellowstone NationalPark. Ph.D. dissertation. Bozeman, MT: Department of Ecology, MontanaState University.

Zhao, G., & Maclean, A. L. (2000). A comparison of canonical discriminantanalysis and principal component analysis for spectral transformation.Photogrammetric Engineering and Remote Sensing, 66, 841−847.

Page 25: Improved wetland remote sensing in Yellowstone …Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental

本文献由“学霸图书馆-文献云下载”收集自网络,仅供学习交流使用。

学霸图书馆(www.xuebalib.com)是一个“整合众多图书馆数据库资源,

提供一站式文献检索和下载服务”的24 小时在线不限IP

图书馆。

图书馆致力于便利、促进学习与科研,提供最强文献下载服务。

图书馆导航:

图书馆首页 文献云下载 图书馆入口 外文数据库大全 疑难文献辅助工具