Robust neural network approach to source-range estimation in a simulated Arctic environment

Abstract

Underwater source-range estimation is conventionally performed via matched field processing (MFP), which relies on precise modelling of the propagation environment to produce accurate estimates. Here, we design a convolutional neural network (CNN) for long range (up to 50 km) source-range estimation in a complex Arctic propagation environment and compare its performance and robustness to MFP. The model parameters are optimized through cross-validation with regularization and other precautions implemented to deter overfitting. It is trained and tested with simulated data recorded on a vertical line array and its robustness to environmental mismatch is examined by introducing deviations to the original sound speed profile (SSP) when generating the test data. Results show that the CNN is more robust to environmental changes than MFP but at the expense of worse performance when the environmental parameters are accurately modelled. Insights into how the CNN performs range estimation is discussed as well. [Work supported by Office of Naval Research.]Underwater source-range estimation is conventionally performed via matched field processing (MFP), which relies on precise modelling of the propagation environment to produce accurate estimates. Here, we design a convolutional neural network (CNN) for long range (up to 50 km) source-range estimation in a complex Arctic propagation environment and compare its performance and robustness to MFP. The model parameters are optimized through cross-validation with regularization and other precautions implemented to deter overfitting. It is trained and tested with simulated data recorded on a vertical line array and its robustness to environmental mismatch is examined by introducing deviations to the original sound speed profile (SSP) when generating the test data. Results show that the CNN is more robust to environmental changes than MFP but at the expense of worse performance when the environmental parameters are accurately modelled. Insights into how the CNN performs range estimation is discussed as well. [Work...