Underwater Bottom Mine Shells Target Classification Based on Relevance Vector Machine

摘要

The problem of classifying underwater bottom mines from acoustic backscattered signals is addressed here. Standard short-time fourier transform (STFT) is applied to convert the echo signal into the time-frequency plane to precisely depict the echo spectrogram, then time-frequency feature extraction scheme based on modified STFT is introduced to deal with mine shell echo signals corrupted by impulse, non-Gaussian noise. The scheme provides a robust estimation of STFT in the reverberation and suppresses it. The target echo features are extracted to reflect different target strengths of two mine shell types influenced by reverberation. The overall system classification performance is benchmarked on two mine shell echo data sets with 25 kHz-50 kHz bandwidth. Echo features are sent to relevance vector machine (RVM)classifier which represents a Bayesian extension of support vector machine (SVM). Compared with SVM, the case study shows RVM yields a much sparse solution and improves classification accuracy. The lake experiment exploits the robustness of feature extraction scheme and effectiveness of classifier with the analysis of the echoes from the two shells underwater bottom.