Estimation and Mapping of Misclassification Probabilities for Thematic Land Cover Maps

Abstract

There is abundant evidence that thematic map accuracy may vary across a landscape in a manner that is only partially related to land cover type. Spatial variation in accuracy may be attributable to factors such as terrain, landscape complexity, and land use patterns. There may be serious consequences in the construction and application of thematic maps when the pattern and degree of spatial variation of map accuracy is not understood. Classification error matrices, the usual summary of thematic map accuracy, are not a natural tool for analyzing spatial variation in accuracy. This article presents a method of mapping spatially based estimates of map accuracy. We formulate a concept of misclassification probability and present a method of estimating misclassification probabilities at training observation locations. Misclassification probability estimates are then interpolated from the training locations to a lattice of points via Kriging. The last step uses the lattice to construct a contour accuracy map. We illustrate the method by constructing an accuracy map for a full Landsat Thematic Mapper scene covering a portion of central Idaho and western Montana, and a finer resolution accuracy map for a small area within that scene. The method is found to provide valuable information on the spatial distribution and variation of map accuracy for both decision-makers and GIS analysts.