Accuracy Assessment of a Land-Cover Map of the Kuparuk River Basin, Alaska: Considerations for Remote Regions

An accuracy assessment of a Landsat MSS-derived land-cover map of the Kuparuk River basin, Alaska was performed. We used a stratified systematic transect-based sampling design with a homogeneous 3by %pixel block sampling unit. The ramifications of the sampling strategy are discussed. Sample sites were located using a helicopter and a Y-Code GPS receiver. Estimates of overall classification accuracy (P), Tau (T,), producer's accuracy, and user's accuracy were calculated from an error matrix. Assessment methods based on fuzzy sets theory were used to supplement the error matrix. The accuracy estimates indicate a classification with high accuracy. However, they are likely to have a fair degree of optimistic bias and can only be applied reliably to homogeneous 3 by 3 blocks of pixels. The combined use of an error matrix and fuzzy sets allowed for a more precise analysis of errors. Based on this analysis, changes were made to the final map. Several methodological advantages contributed to the high classification accuracy. Purpose Quantifying and documenting the accuracy of maps and spatial data are important components of any mapping process. However, assessing the accuracy of a map can be a time-consuming and expensive process. This is especially true for maps of remote areas, such as the North Slope of Alaska, which can involve considerable financial, logistical, and technical constraints. This paper presents an accuracy assessment of a satellite-derived land-cover map of the Kuparuk River basin on the North Slope of Alaska and examines how the choice of sampling strategy and analysis methods affects estimates of classification accuracy and their usefulness. S.V. Muller, D.A. Walker, and N.A. Auerbach are with the Tundra Ecosystem Analysis and Mapping Laboratory, Institute of Arctic and Alpine Research, University of Colorado, CB 450, 1560 30th St., Boulder, CO 80309-0450 ([email protected]). F.E. Nelson is with the Department of Geography, University of Delaware, Newark, DE 19716. J.G. Bockheim is with the Soils Department, University of Wisconsin, 1525 Observatory Dr., Madison, WI 53706. S. Guyer and D. Sherba are with the Bureau of Land Management, Division of Cadastral Survey (AK-924), 222 West 7th Ave. #13, Anchorage, AK 99513-7599. The Kuparuk River Basin Land-Cover Map As part of the of the National Science Foundation's (NSF) Land-Atmosphere-Ice-Interactions (LAII) Flux Study (Weller et al., 1995), a land-cover map of the Kuparuk River basin was derived from Landsat Multi-Spectral Scanner (MSS) satellite image data. The Kuparuk River basin, located on the North Slope of Alaska, covers approximately 9,201 km2 and extends 216 km from north to south, and 78 km from east to west (Figure 1). Except for the Dalton Highway and oil drilling bases along the Arctic coast, this area of the North Slope remains undeveloped, remote, and difficult to access. General vegetation land-cover types were derived by classification of the Landsat MSS satellite data (Plate 1; N. Auerbach et al., unpublished data, 1996). A framework for the land-cover type designations was provided by BraunBlanquet vegetation analysis of a tussock tundra landscape in the Brooks Range Foothills, Alaska (Walker et al., 1994). Derived classes include (1) Barrens, (2) Moist nonacidic tundra (MNT), (3) Moist acidic tundra (MAT), (4) Shrublands, (5) Wet tundra, (6) Water, (7) Clouds and Ice, and (8) Shadows. To expedite image processing, the digital data for a rectangular region encompassing the Kuparuk River watershed were extracted from an existing mosaic of MSS frames covering the Central Arctic Management Area (CAMA) and Arctic National Wildlife Refuge (ANWR), northeast Alaska, produced by the U.S. Geological Survey, EROS Data Center, Sioux Falls, South Dakota. Images for the entire mosaic were acquired during the snow-free growing seasons of 14 August 1976 through 2 August 1985. Due to prevalent cloud cover over the North Slope during most growing seasons, single-time-period (e.g., one week) mosaics of imagery from sun-synchronous satellites are generally not feasible. The mosaic (80 m nominal spatial resolution) was resampled to 50-m pixels, and was geometrically corrected using cubic convolution interpolation by means of a second-order polynomial registration, with a resultant root-mean-square error (RMSE) of 57.4 m. An IsoData unsupervised classification approach was implemented based on input of the green, red, and infrared spectral bands of the MSS image. Forty cluster classes were initially generated and then aggregated into the eight land-cover classes. We used first-hand experience and familiarity with the area, as well as geobotanical maps and earlier Landsat-derived maps of the region, as supplementary information to interpret the spectral classes (Walker et al., 1982; Walker and Acevedo, 1987; Walker et al., 1989; Walker and Walker, 1991; Photogrammetric Engineering & Remote Sensing, Vol. 64, No. 6, June 1998, pp. 619-628. 0099-1112/98/6406-619$3.00/0 O 1998 American Society for Photogrammetry and Remote Sensing I PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING