A study of effects of sonar bandwidth for underwater target classification

Abstract

The problem of classifying underwater targets is addressed in this paper. The proposed classification system consists of several subsystems including preprocessing, subband decomposition using wavelet packets, linear predictive coding, feature selection and neural network classifier. A multi-aspect fusion system is introduced to further improve the classification accuracy. The classification performance of the overall system is demonstrated and benchmarked on two different acoustic backscattered data sets with 40- and 80-kHz bandwidth. A comprehensive study is then carried out to compare the classification performance using these data sets in terms of the receiver operating curves, error locations, and generalization and robustness on a large set of noisy data. Additionally, the importance of different frequency bands for the wideband 80-kHz data is also investigated. For the wideband data, a subband fusion mechanism is introduced which offers very promising results.