Study of feature discrimination effects due to time varying channel for passive sonar

The performance of an automatic marine vessel classification system is dependent on the uniqueness of the features in characterizing the classes and robustness to distortions introduced due to the propagation of the radiated sound in the channel. This paper studies the effect of a time varying channel distortion on several spectral as well as cepstral features extracted from the acoustic signatures of marine vessels. Different types of channels with varying range, depth and bathymetric profiles have been considered as representative of the diverse channel manifestations. The features for each class are modeled both at the source location and at a distant receiver using Gaussian mixture models (GMM). We have evaluated the class discriminating capability of each feature by computing the Fisher discriminant ratio (FDR) and the divergence measure both at the source location and at the receiver. The distortion effect of the channel has been reduced using certain averaging mechanism. An analytical evaluation of the advantages of the averaging mechanism is presented for the FDR features selection method