Accuracy Sampling design Bias on Coarse Spatial Resolution Land Cover Data in the Great Lakes Region (United States and Canada
John S. Iiames, Jr.
US Environmental Protection Agency Research Triangle Park, NC, USA
Abstract: A number of articles have investigated the impact of sampling design on remotely sensed landcover accuracy estimates. Gong and Howarth (1990) found significant differences for Kappa accuracy values when comparing pure- pixel sampling, stratified random sampling, and stratified systematic unaligned sampling. This study compares accuracy assessment results for landcover derived from 2007 Moderate Resolution Imaging Spectroradiometer (MODIS) 250 m normalized difference vegetation index (NDVI) time-series data for the Great Lakes Basin (GLB), USA. Here, two sampling schemes are compared: (1) Pure-pixel sampling (center pixel within a 3 x 3 window) and (2) Isolated independent pixel sampling (i.e. 'edge pixels'). MODIS spectral characteristics typically 'bleed' from adjacent pixels, causing pure 'edge pixels' to be suspect with respect to their homogeneity. This study will explain the possible bias by inclusion of these fringe pixels within the assessment process. Our study focuses on the Northern Lakes and Forests Level III Omernik ecological region (115,934 km2) south of Lake Superior and existent within three states: Michigan, Wisconsin, and Minnesota.
Keywords: MODIS, NDVI, accuracy