Underwater target classification based on the combination of dolphin click trains and convolutional neural networks.

Sonar remains a major way to detect and discriminate underwater targets by interpreting the echoes. In this study, we used broadband dolphin clicks to detect and classify targets. The peak and notch features of the echo spectra were coded, and echoes were obtained using five-click trains, with the number of clicks changing from 1 to 50. Codes containing the target interpretation were classified by convolutional neural networks (CNNs). Compared to a single click, the increasing number of clicks to 5, 10, 20, and 50 in a train would gradually improve the classification rate of targets by 3%, 6.1%, 8.2%, and 10.5% on average with a signal-to-noise ratio ranging from -10 to 15 dB. The 50-click train outperformed other click trains in target detection and classification. The CNNs achieved an average classification accuracy of 95.2% for a 50-click train, higher than that of the nearest neighbor method by 10.3% across signal-to-noise ratios. Therefore, the usage of dolphin clicks and CNN-based echo encoding technologies constitutes an effective method for enhancing target classification, offering valuable insights for future applications in detecting underwater targets.