CLASSIFIER PERFORMANCE AND MAP ACCURACY

The distinction between the performance of a classifier and the accura cy of the resulting thematic map is discussed. Sampling the map itself is recommended as the preferred method for testing its accuracy. It i s shown, nonetheless, that map accuracy can also be determined from th e performance of a classifier, as assessed from testing data, provided the prior probabilities of class membership are well estimated. Unfor tunately, that is rarefy done in practice so that errors are often mad e in reported map accuracy figures if they are inferred from classifie r behavior. Examples are presented to illustrate how an apparently wel l-behaved classifier can give poor results for certain classes, becaus e of the proportions of those classes in the region being imaged.