Seabed classification with ResNet-18 using multi-channel ship noise spectrograms

Abstract

Transiting cargo ships in the ocean generate noise that can be used to infer the effective seabed type below. Recently this has been demonstrated with a residual convolutional neural network trained on synthetic data that displays frequency and range-dependent transmission loss generated by a range-independent normal model and empirical source spectrum for ship noise. The generalizability of the trained model was tested on ship noise measured from the middle sensor of a 16-channel vertical line array near shipping lanes on the New England continental shelf. In contrast, this current work utilizes multiple sensors and seeks to determine the optimal combination of sensor data. It was discovered that including sensors near the seabed improved model performance. Best results were obtained when the model’s average sensor depth was 75%–80% of the water depth. The distance between sensors had no discernible effect on performance. Although the almost isovelocity sound speed profile used in this study indicated little difference between one-channel and two-channel model performance, future analyses will test the hypothesis that the use of multiple sensors will improve model performance for greater sound speed variation. [Work supported by the Office of Naval Research and the National Science Foundation REU program.]