Underwater target classification using canonical correlations

A feature extraction method for underwater target classification is developed that exploits the linear dependence (coherence) between two sonar returns. A canonical coordinate decomposition is applied to resolve two consecutive acoustic backscattered signals into their dominant canonical coordinates. The corresponding canonical correlations are selected as features for classifying mine-like from non-mine-like objects. Test results are based on a subset of a wideband data set that has been collected at the Applied Research Lab (ARL), University of Texas (UT)-Austin. This subset includes returns from different mine-like and non-mine-like objects at several aspect angles in two different bottom conditions. The test results demonstrate the potential of the canonical correlation-based feature extraction for underwater target classification in difficult bottom conditions.