The overall performance of a multi-stage decision- tree classifier has been shown to be better than that of the conventional single-stage classifier with the same number of features because different feature subsets can be selected at different stages. But, the classification time increases due to the complexity of computation. The linear binary tree classifier designed by the method proposed in this study takes the advantages of the accuracy of a decision-tree classifier and uses linear discriminant functions at decision stages to reduce the classification time. An application of this method to the multispectral remotely sensed data is presented. All ten classes under consideration are assumed to be gaussian distributed. The result from a test on about 7000 samples shows that the linear binary tree classifier is more accurate and much faster than the maximum-likelihood classifier with the same number of features.

Date of this Version