A Comparison and Rating of Conditioned Bayesian Discriminant Classifiers

by Quantitative Term of Training Representability

Abstract

Representability of training areas is primarily focused on its qualitative spectrum features. But it was revealed by this study that the representability by a quantitative application of prior frequencies of each land-cover item also plays an important role for more accurate classification in remote sensing, increasing the rate by ten to fifteen percent for the whole scene in pixel-by-pixel evaluation.
Complete enumeration for accuracy assessment in test field was realized such that detailed digital land-use data of 10 m×10 m resolution, prepared by the Japanese Geographical Survey Institute, were aggregated to 50 m×50 m cell size in match for geocoded Landsat MSS pixels.
Changing the sampling sizes in the merged file, coupled with eight application types of Bayesian discriminant method, produced sixteen systemat-ically-conditioned classifiers to compare, and resulted that there is a great variation of accuracies by the classifier especially in terms of the applicable suitabilities in case of separating individual land-cover items.