抄録
Among automatic supervised classification methods, the best linear discrimination (BLD) is superior in the theoretical sense because it can utilize statistical features (covariance matrix) extracted from each category. However, as pairwise comparisons were needed, processing speed of BLD became very slow when the number of categories increased.
In this paper, we proposed following two efficient algorithms: (a) adaptive ordering of pairwise comparisons by utilizing classified result of neighbor pixels (BLD/adp), and (b) utilization of the BLD as the decision rule of the binary decision tree (BDT/bld). As a theoretical approximation of processing speed of these algorithms, we showed the number of multiplication required for classifying a pixel, which depended on the number of categories (N) and variables (M). It was shown that the number of multiplication ranges from O (logN×M) to O(N×M) in the case of BDT/bld, while the number was not less than O(N×M) in the case of BLD/adp.
In order to see the actual performance, we applied these algorithms to multispectral data of LANDSAT 5. The conventional methods (maximum likelyhood, Fisher's linear discrimination, and binary decision tree) were also applied to the same data and were compared with proposed methods. It was shown that the performance of BDT/bld was better than that of Fisher's linear discrimination when the number of categories was large.
