ANALYSIS OF CLASSIFICATION RESULTS OF REMOTELY-SENSED DATA AND EVALUATION OF CLASSIFICATION ALGORITHMS

Classification results of remotely sensed data are usually summarized as confusion matrices, and various classification algorithms are used to improve results. Confusion matrices should be normalized to assess classification accuracies of remotely sensed data, and multiple compar isons are required to evaluate the classification algorithms. The clas sical iterative proportional fitting procedure, including eliminating zero counts, was scrutinized to normalize confusion matrices. The Tuke y multiple comparison method was used for the comparison of results fr om three classification algorithms: minimum distance, maximum likeliho od, and an artificial neural network. Normalized confusion matrices pr ovided uniform margins and accuracies for each classification category . The Tukey comparisons of the three algorithms were made simultaneous ly; results provided the overall classification accuracy for each algo rithm and showed no differences among the algorithms at a risk level o f 5 percent. Normalized confusion matrices can be compared entry by en try because of their uniform margins. Results of this study indicate t hat classification algorithms can be evaluated with the Tukey method, and the multiple comparisons of the algorithms should be made based on normalized category accuracies obtained with the iterative proportion al fitting procedure. Normalized confusion matrices provide a unified measure of producer's and user's accuracies.