A canonical correlation-based feature extraction method for underwater target classification

A new feature extraction method for underwater target detection and classification is developed. This is accomplished by applying a canonical coordinate decomposition to the backscattered signals to maximize the mutual information between the outputs of the channels associated with consecutive aspects/pings. As a consequence of this information maximization the most coherent target features are extracted while reverberation effects are removed. The classification of targets/nontargets can then be made based on the extracted canonical coordinate features. Test results presented in this paper are based on a wideband data set that has been collected at Applied Research Lab (ARL), University of Texas (UT)-Austin. This data set consists of the backscattered signals of four different objects: two mine-like objects and two non-mine-like objects for several aspect angles and both smooth and rough bottom conditions. The extracted canonical coordinate features from every other aspect angle of backscattered signals in smooth bottom condition are used to train a back-propagation neural network (BPNN) classifier. The generalization ability of the trained network is then demonstrated by computing the classification rate statistics on the rest of the smooth data set. The performance of the classifier is investigated against environmental changes by testing the trained network on the rough bottom condition data. The results demonstrate the potential of the proposed method for feature extraction in difficult bottom/buried conditions.