The impact of water sound speed on deep learning seabed classification from shipping noise

Previous studies have shown that ship noise spectrograms can be used for deep-learning seabed classification when the water sound speed profile (SSP) is known. This work addresses the challenge of seabed classification with unknown SSPs. Four synthetic training datasets were generated, each featuring the same catalog of 34 seabeds but with different SSPs. The four sets of SSPs are (1) slightly downward refracting SSPs measured during the Seabed Characterization Experiment (SBCEX) of 2017, (2) SSPs measured during SBCEX 2022, which exhibit greater variation with depth, (3) a combination of the measured SSPs from SBCEX 2017 and 2022, and (4) a collection of SSPs from the World Ocean Atlas database, sourced from locations where the water depth is similar to the experiment location. Each training dataset was used to train ResNet-18 models for seabed classification, employing different combinations of hydrophone depths from vertical line arrays (VLAs). Measured spectrograms of ship noise from both SBCEX 2017 and 2022 are used as testing data for the trained models. Comparison of the seabed predictions from the different cases demonstrated the importance of including sound speed variability in training deep learning models for ocean acoustics applications. [Work supported by ONR.]