Deep Learning-Based Source Separating for Target Detection in Underwater

Abstract

Marine clutter interfere with sonar systems and other underwater acoustic equipment by affecting the echo or target signal, thus impacting target detection and tracking performance. In marine environments with low noise levels, sonar systems can detect underwater targets at greater distances. Conversely, in noisy and turbulent marine environments, even nearby targets can be challenging to detect. Clutter, such as sounds from snapping shrimp (SS), can severely disrupt sonar echoes and hinder underwater communication. This research propose deep learning based method using AutoEncoder and U-Net to reduce clutter in received underwater vessel signals. We validate this method through 2D classification using Convolutional Neural Networks (CNNs) and present a technique for separating signal sources. Experimental results show that the method using AutoEncoder and U-Net to remove clutter demonstrated superior performance compared to methods that did not use these techniques, achieving over 90% classification accuracy at signal-to-noise ratios greater than -12 dB. © 2024 IEEE.