A Binary Tree Classifier for Ship Targets

This paper describes the design of a binary tree classifier for ship targets. The design methodology is general enough so that it can be utilized for other classification problems. A hierarchical clustering procedure is employed to i) discover the underlying structure of data, and ii) construct the binary tree skeleton. The best feature subset, at each nonterminal node of the tree skeleton, is selected through a multivariate stepwise procedure which attempts to maximize the class separability. Further, this stepwise approach continues, until the probability of error at each nonterminal node with respect to a quadratic discriminant function is minimized. The proposed tree classifier has been evaluated against 1300 samples and classification accuracy of 85% versus 62% for the single stage classifier is achieved.