Western State College of Colorado

wsc > academics > academic programs > geology > student research > digital land-cover map

Digital Land-Cover Map of the Gunnison Basin

(116 kb) Ralph Falsetto, Joanna Soceka, John Sowell and Allen Stork Western State College, Gunnison, CO 81231  

Introduction

The Upper Gunnison Basin is a high-elevation ecosystem in Colorado ranging in elevations from 2200 to 4300 m above sea level. Present vegetation of this 11,000 km² basin ranges from sagebrush steppe at lower elevations, to montane and subalpine forests at mid-elevations, and alpine tundra above 3600 m. The low population density and the high percentage of federally-owned land (85% of the Basin) have helped prevent large-scale destruction except along rivers at lower elevations where riparian communities have been impacted by ranching.

Mapping of the Basin's vegetation has been limited by the diverse and often incompatible needs of the federal agencies managing resources in the Basin and the lack of digital technology. Numerous large-scale maps, typically of 1:24,000 scale, have documented the spatial distribution of potential vegetation in selected areas. The Soil Conservation Service (now the Natural Resource Conservation Service) has produced a Basin-wide map of land cover at a scale of 1:126,720 (USDA 1976a, 1976b, 1976c). While these maps delineate broad forest types, they combine all non-forested areas of the Basin as range. Falsetto et al. (1997) mapped the Basin using Landsat Multispectral Scanner (MSS) data. The map has an overall accuracy of 74% with eight land cover categories: coniferous forest, rock/bare ground, snow, water, grasslands, aspen, subalpine meadows, and tundra.

The MSS data has been insufficient in spatial detail (60 m pixels) and discrete reflectances (four spectral bands) to provide land cover maps of higher discrimination (Falsetto et al. 1997). Landsat Thematic Mapper (TM) data provides more spectral bands and resolution which may allow better characterization of the Basin's land cover (USDA 1995, USGS 1996b). Landsat TM images contain seven bands in the electromagnetic spectrum: three are visible light, three are infrared, and one is thermal infrared (USGS 1996b). The spatial resolution for Landsat TM imagery is 30 m per pixel for all bands except the thermal infrared, which is 120 m per pixel (Arnoff 1993).

The objective of this research was to further refine the land-cover map of the Upper Gunnison Basin using Landsat TM images and our GIS database for the Gunnison Basin. As with our previous work, of particular interest was the mapping of vegetation and the development of a map that would be useful for land management agencies (Stoms 1994).

Materials and Methods

The area of study included all portions of the Gunnison Basin that drain into the Gunnison River above the Black Canyon of the Gunnison River (Blue Mesa Dam). This basin, here referred to as the Upper Gunnison Basin, is delineated by the Continental Divide to the east and south, the Elk Mountains to the north, and the West Elk Mountains, Soap Mesa, Alpine Plateau, and the San Juan Mountains to the west. Mapping included the area between 37^o49'N and 39^o03'N latitude and 106^o13'W and 107^o13'W longitude (Falsetto et al. 1997).

The georeferenced terrain-corrected Landsat TM images, along with a Digital Elevation Model with the same resolution (30 m) were obtained from the EROS Data Center. A vector layer that outlines the Upper Gunnison Basin (Falsetto et al. 1997) was used to limit the image to the mapping area.

Landsat TM bands 3,4 and 5 were processed using the Idrisi cluster module (unsupervised classifications) to produce a base map with groups of unique spectral signatures. The map was subjected to iterations of field checking and corrective processing until a land cover map with meaningful vegetation discrimination was obtained. Several map clusters were inadequately distinguished with unsupervised clustering alone thus specific methods were used to increase map accuracy and cluster distinction.

  Distinct shadows on west-facing slopes hindered meaningful clustering. Shadow areas were separated and reclassed to a Boolean image that was used to create separate shadow and non-shadow images. The resulting two images were clustered separately to minimize the effects of the shadowed areas on the unsupervised cluster algorithm (Schneider et al. 1995). The base map is a combination of the two separate clustered images.

To better distinguish aspen forests from willows, meadows, and riparian areas, an aspen layer was created by supervised classification and overlaid onto the base map. Several signature files, which consist of a series of training sites, were created. Different training site combinations were used with the maxlike classifier to determine the most representative signature file. The combination of two signature files, one in the northwest portion of the Basin and the other to the south, near Slumgullion Pass, provided the best results.

The unsupervised classification produced a conifer cluster that excluded some areas of low tree density. A supervised layer was created with training sites from the shadow and non-shadow areas of the base map. A conifer layer was created using maxlike classifer and overlaid on the base map.

Development of the land cover map also utilized the DEM map layer to allow the discrimination of spectrally indistinguishable vegetation types by elevation (USGS 1996a). Such use of elevation in vegetation mapping can be effective in mountainous regions (Schneider et al. 1995). This technique was used to limit sagebrush to below 3350 meters. Also, grassland were limited to below 3700 meters, and tundra to above 3700 meters.

Within riparian hayfields were areas mistakenly classified as coniferous forest. The DEM was used to create an image where all areas with a slope of greater than 2 degrees were eliminated. A Boolean image was created from the DEM and used to restrict the upper elevation limit of the slope image to 3175 meters. The resulting image was overlaid onto the base map and assigned to the riparian category.

Accuracy assessments were performed on the processed land cover map using 173 points randomly selected throughout the basin (Eastman 1995). The points were checked using aerial photography, USDA aerial photos from the U.S. Forest Service, and field checking. To assist in referencing the points on the photos, roads, vegetation patterns, and the global positioning system were most often used. Road vector layers were obtained from Digital Line Graphs produced by the U.S. Department of the Interior (USDI 1993).

Results and Discussion

The Landsat TM data with its greater resolution and additional bands produced a very similar unsupervised cluster to the MSS data used in the previous Basin map (Falsetto 1997). The TM unsupervised cluster resulted in many of the same problems as the MSS data including the inability to distinguish between aspen, willow and riparian areas. While previous supervised clustering of MSS spectral bands was ineffective, such clustering of TM images was more successful.

The land-cover map units were broadly defined and included several communities. The sagebrush category included herb-dominated openings along with the different sagebrush-dominated communities. The conifer category included forests dominated by spruce, pine, and fir. The shrub mix category contained some sagebrush and other woody plants such as serviceberry and Gamble's oak with their associated understories. The rock category included talus, scree, and relatively barren areas as well as some man-made structures such as the airport runway. The grass category included high-productive areas dominated by herbaceous species. The riparian map unit contains stream-side vegetation and irrigated agricultural areas.

A total of 173 points in nine categories were verified, 141 of which were found to be accurately mapped, resulting in an overall accuracy of 81.5% ( Table 1). An additional snow category is present on the map but was not verified since the extent of snowcover changes so frequently. The individual categories varied greatly in accuracy. The accuracy of classifying sagebrush and conifer were above 85% which meets typical map accuracy standards (Eastman et al. 1993). All other categories are below 85% , therefore their usefulness will depend on the accuracy required for a given application.

Table 1. Percent accuracy of map categories, excluding snow.
Category	# Correct	Total	% Correct
Sagebrush	30	34	88.24
Conifer	32	34	94.12
Shrub mix	5	10	50
Aspen	32	38	84.21
Rock	5	9	55.56
Tundra	16	21	76.19
Grass	15	20	75
Water	1	1	100
Riparian	5	6	83.33
Total	141	173	81.5

Much of the error was due to misclassification of a few map units (Table 2). The mutual misclassification of rock and tundra may be because the boundary between rock and tundra is gradational and delineation is arbitrary. The communities included in the shrub mix category were often small vegetation units, often much smaller than the 30 meter pixel size used to create the map. This may have hindered the ability to adequately characterize these pixels which contained more than one land-cover type.

Table 2. Error matrix for map categories, excluding snow.
Map Category	Actual
	Sagebrush	Conifer	Shrub mix	Aspen	Rock	Tundra	Grass	Water	Riparian
Sagebrush	30		1		1	1	1
Conifer		32		2
Shrub mix	1	1	5	2			1
Aspen	1	1	1	32			2		1
Rock					5	4
Tundra			1		3	16	1
Grass	2		1	2			15
Water								1
Riparian			1						5
Total	34	34	10	38	9	21	20	1	6

Overall, the Landsat TM data was better than the Landsat MSS data in distinguishing land cover and separation of the categories. The map produced by MSS data had only eight categories and the TM data produced a map with ten categories. This indicates more accurate recognition of differences in reflectance of the wavelengths used.

Literature Cited

Arnoff, S. 1993. Geographic information systems: a management perspective. WDL Publications, Ottawa, Ontario, Canada.

Eastman, J.R. 1995. Idrisi for Windows: user's guide. Clark University, Worcester, Massachusetts, USA.

Eastman, J.R., P.A.K. Kyem, J. Toledano, and W. Jin. 1993. GIS and decision making. Explorations in geographic information systems technology, volume 4. UNITAR Palais des Nations, CH-1211 Geneva 10, Switzerland.

Falsetto, R.N., P.F. Kline, J.B. Sowell, and A.L. Stork. 1997. Upper Gunnison Basin land cover. Http://geology.western.edu/research/landcover/Welcome.html.

Schneider, K. , Robbins, P. eds. 1995. GIS and mountain environments. Explorations in geographic information systems technology, volume 5 UNITAR Palais des Nations, CH-1211 Geneva 10, Switzerland.

Stoms, D., F. Davis, C. Cogan, K. Cassidy. 1994. Assessing land cover map accuracy for gap analysis. National Gap Analysis Data Center. http://www.gap.uidaho.edu/gap/New/Publications/Handbook/LCA.htm.

USDA. 1976a. Land use and natural communities: Gunnison County, Colorado. United States Department of Agriculture, Soil Conservation Service, Portland, Oregon.

USDA. 1976b. Land use and natural communities: Hinsdale County, Colorado. United States Department of Agriculture, Soil Conservation Service, Portland, Oregon.

USDA. 1976c. Land use and natural communities: Saguache County, Colorado. United States Department of Agriculture, Soil Conservation Service, Portland, Oregon.

USDA. 1995. Guidelines for the use of digital imagery for vegetation mapping. United States Department of Agriculture, Forest Service Engineering Staff. Washington D.C..

USDI. 1993. 1:100,000 scale DLG Data hydrography/transportation. United States Department of the Interior. Reston, VA.

USGS. 1996a. 1-Degree USGS Digital elevation models. United States Geological Survey.
Http://edcwww.cr.usgs.gov/glis/hyper/guide/1_dgr_dem.

USGS. 1996b. Thematic Mapper Landsat Data. United States Geological Survey. Http://edcwww.cr.usgs.gov/glis/hyper/guide/landsat_tm.

This project has been funded to date by a Thornton Biology Research Grant.