Results of Approbation of the Method for Predicting the Classification Accuracy of Multichannel Images

The results of verification of the method for predicting the classification accuracy for three-channel remote sensing images are presented. This study is carried out based on pixel-by-pixel classification according to the maximum likelihood criterion. As a criterion for classification accuracy, the weighted total probability of correct recognition of classes is taken, where the weight coefficients are a priori probabilities of the occurrence of objects from the corresponding classes. The results of comparison of statistical estimates of a priori probabilities of classes and their predictions by the Monte Carlo method using “hard” and fuzzy classification are presented. Empirical dependences of the relative errors of prediction of the probabilities of correct recognition on the sample size are obtained. It is shown that predictions of the total probability of correct recognition, provided that the random sample is representative and of sufficient volume, produce the specified reliability. At the same time, the volume of processed data is significantly (at least by 1-2 orders of magnitude) reduced. This, in turn, leads to the corresponding reduction in the complexity of calculations and the time spent on obtaining predictions for the desired criteria.