End-to-End Geoacoustic Inversion With Neural Networks in Shallow Water Using a Single Hydrophone

This article presents a deep learning (DL) method to perform joint source detection and environmental inversion of low-frequency dispersed impulse signals recorded on a single hydrophone, in a fully automated way, with the inversion part covering both source localization (range and depth) and geoacoustic inversion (with the seabed modeled as a single sediment layer over a basement). The benchmark used for testing the resulting DL models are signals that were generated by navy explosives [signal underwater sound (SUS) charges] deployed during the Seabed Characterization Experiment 2022 performed in the New England Mud-patch (NEMP) off the coast of Massachusetts. A DL model based on a 1-D convolutional neural network is trained using simulated data. The resulting model is used to automatically process 816 h of acoustic data containing 289 SUS events. All the SUS events are detected (with no false positives), localized with a mean error of 400 m, and used to invert for seafloor geoacoustic parameters. The predicted parameters are in agreement with results obtained using classical inversion schemes. Using a trained DL model requires little to no computation time and power, compared to classical methods, which employ high-cost computational schemes. This advantage enables efficient inversion of enough SUS events (289) to spatially cover the NEMP, and inversion results suggest spatial variability in the mud sound speed.