Robustness Analysis of a Convolutional Neural Network Approach to Source-Range Estimation in a Simulated Arctic Environment

This study presents a convolutional neural network (CNN) approach to underwater source-range estimation in an Arctic propagation environment and compares its performance to conventional matched field processing (MFP). The covariance matrices of simulated sources upon a vertical line array are used as input data and estimates through both classification and regression approaches are examined. The network architecture is designed to be intuitive and lightweight; regularization is implemented to prevent over-fitting. The training data consist of acoustic outputs from a near-surface monopole source placed at discrete range increments between 3-50 km away from the receiving array; the test data are generated by placing the source at random ranges within the training interval. Robustness of the CNN models to sound speed profile (SSP) variability is tested. Results show that the CNN approaches are more tolerant of SSP mismatch compared to MFP at the expense of worse range resolution when the SSP is modelled accurately. By examining the CNN models filter activations and intermediate layer outputs, we present insights into how they generate predictions and achieve their robustness.