Statistical-based feature extraction and classification of active sonar data

Abstract

Sonar target recognition is difficult due to the potential nonlinear overlap within an acoustic color response due to various backscatter and clutter within the ocean. This talk presents initial results from using a statistical model of feature vectors in conjunction with machine learning classifiers. Canonical correlation analysis (CCA) seeks to find two linear combinations of data by maximizing the correlation between the linear combinations while maintaining unit variance. In this application, CCA is used as a feature extraction method before target classification of active sonar data experimentally collected during the Shallow Water Active Classification (SWAC)-1 and SWAC-2 sea trials in the Malta Channel. The database consists of 20 targets; three were analyzed using this method. The data are generated by taking windows of consecutive pings from the ping-vs-time domain and performing CCA. The intuition behind using CCA is that there are persistent features within the data that morph over time due to changing target aspect angles and platform positions which can be represented by the maximally correlated linear combinations of data among consecutive pings. The resulting linear combinations are feature vectors used to train a single hidden-layer neural network classifier. Results are reported as overall classification accuracy and confusion matrices.