SPATIAL CLASSIFICATION ERROR RATES RELATED TO PIXEL SIZE

Abstract

This chapter reviews spatial classification error rates related to pixel size. The goal is to classify the central point of the square pixel at t. If one does not exploit any spatial structure and pretend that pixels are class homogeneous then the optimal classifier for t based on X(t) is the Fisher linear discriminant function appropriately modified by the prior probabilities Q and 1 - Q. Larger pixels will have the property that the residual or noise component of the observation, e(t), will have smaller variance but, at the same time, larger pixels are less likely to be class-homogeneous. These properties work oppositely with respect to classification error probabilities even with computational questions aside. For modest increases in pixel size it will turn out, for the most part, that larger pixels are better than smaller pixels.