An end-to-end deep learning approach for joint detection, source localization and environmental characterization using a single hydrophone in shallow water

Abstract

Environmental characterization of underwater environments is crucial for modeling acoustic propagation in the ocean, which in turn enables underwater applications such as performing environmental impact studies on the effect of ocean noise on marine life, evaluating sonar performance and more. In this work, the aim is to automate and improve run-time performance of environmental characterization, compared to classical methods which employ high-cost computational schemes and sometimes require manual inputs from a skilled operator. To do so, we use 1D Convolutional Neural Networks (1D-CNN), a type of Deep Learning (DL) model, to acoustically characterize the underwater environment using unprocessed pressure time-series recorded on a single hydrophone. We notably propose an end-to-end approach to detect, classify and localize acoustic sources along with estimating environmental parameters using a single DL model. The experimental data used for testing the resulting 1D-CNN models are signals that were generated by navy explosives (SUS charges) deployed during the Seabed Characterization Experiment performed in the New England Mud-patch off the coast of Massachusetts.