Comparing the Performance of Convolutional Neural Networks Trained to Localize Underwater Sound Sources

Localizing underwater sound sources in the highly-variable ocean environment is challenging. The performance of conventional methods for localizing sound sources, such as matched-field processing, depends on accurate approximations of the environment while localization methods using neural networks typically generalize better, even when information about the environment is sparse. This paper directly compares the performance of convolutional neural networks (CNNs) trained to localize sound sources in simulation from real-valued (magnitude only) versus complex-valued pressure data. We also compared the performance of these two types of networks on data measured from single or multiple receivers in different spatial arrangements in the simulated water column. We also measured the CNNs’ performance loss as the number of sound speed profiles (SSPs) used to generate the simulated field replicas was increased. We found that when trained on data from only one SSP, the complexvalued network predicted the range and depth of the source more accurately than the magnitude-only network. However, as the number of receivers increased, this accuracy margin between the two network types decreased. All CNNs lost accuracy as we increased the number of SSPs, as predicted, but the magnitude-only CNNs still maintained relatively high accuracy for depth predictions even after training on data from a1115 different SSPs, exceeding the accuracy of the complex-valued networks. For range inferences, the complex-valued networks performed better than the real-valued networks, but when trained on more SSPs, the performance of the complex-valued networks decreased.