1. For computational expediency, the decision algorithm employed in the automatic classifier is based on a linear discriminant function. Although in principle more accurate classification should be attained with high order discriminants, it is found in practice that, because of the vagaries of actual remotely sensed data, the performance of the classifier on samples of known test data is as accurate as the more widely used maximum likelihood classifier(13)while the run time is much shorter. The success of any feature classifier designed by a supervised learning scheme (one in which training samples are used) is dependent upon the extent to which the training patterns may be generalized to the other scene patterns. In practice, the operational data' s characteristics tend to drift from those of the training data, and the sequential linear classifier appears to be more tolerant of this drift than the maximum likelihood classifier. In the land use sequential linear classification map shown in Figure 6 it was found that, after the original three band photography had been Iuminance corrected, the maximum likelihood classifier failed to identify Figure 5. Processing Flow in Feature Selection
2. . u;r,atIson of norn-jalized C Y Q- c r y e a t far tkis gi.cture whose sizs was approximilte!y 1700 x 400 took ahwt 18 minutes induding the detc:r-firati.t>! of minimi.x-!,and maxir:unx v;iuesun an XBM 7ri 94 when the filter parameters were I<5, K 2 4 , L;