ASSESSING CLASSIFICATION PROBABILITIES FOR THEMATIC MAPS

In this paper we discuss the methods for estimating the probability th at a pixel is actually in category i given that it was classified into category j, and methods for calculating the probability that a pixel is classified into category j, given that it is actually in category i . We show that the proper method for calculating these probabilities d epends upon how the sample pixels are selected. If they are selected f rom the entire map, without regard to classification, then both of the above probabilities can be computed using standard multinomial formul ae. However, if separate random samples are taken from the pixels clas sified into each category, then the former probability can be estimate d using multinomial formulae, while the latter should be estimated emp loying Bayes theorem. We show how to do this, present an estimate of t he variance of the estimated probability, and conclude with two exampl e applications; one with TM data from New Jersey, and one with simulat ed data.