Interpretable features for underwater acoustic target recognition

The major challenge of underwater acoustic target recognition is that the features clearly characterizing the underwater acoustic targets remain indistinct, where the sound signals are often submerged by intense noise. In this paper, we aim to discover an efficient interpretable feature set that can reveal the inherent mechanism, and result into eighty-eight efficient features. The performance of these features is evaluated with experiments by BP Neural Network. CNN is used as the baseline method. The accuracy of detection based on these features can reach 87.22%, 99.31%, and 91.56% in three sea areas, while CNN shows the accuracy of 64.17%, 99.17%, and 59.17%, respectively. The lowest MRE of ranging with these features collaborating with BP Neural Network is only 7.09%. The experimental results show that these features indeed exist validity for underwater acoustic target recognition with explicit physical interpretability, and lead to very low computational complexity in recognition.