Underwater acoustic localization using a modular differentiable model for acoustic wave propagation

Model-based algorithms for underwater acoustic localization are highly dependent upon environmental knowledge that is rarely available. Moreover, model-based localization usually involves optimization in high-dimensional spaces to find the best candidate for source localization. On the other hand, data-driven methods deliver poor generalization performance in data starved scenarios. Therefore, in order to improve generalization, we propose a modular, deep learning-based architecture that inherently learns the multipath structure, while learning the environmental properties from the training data. Furthermore, since ReLU-based fully connected networks cannot capture the high-frequency contents of the signals of interest properly, we use sinusoidal activations. Although the model needs an estimate of the carrier frequency as a hyper-parameter, it is observed that it can tolerate deviations about the true frequency. This model maps the source and receiver locations to received signals and is then used in a gradient descent manner exploiting the automatic differentiation toolbox in PyTorch, to find the source location. To evaluate the performance of the algorithm, we use Bellhop to generate data for a given set of environmental parameters. We show that our method outperforms matched-field processing and a deep fully connected network without a modular structure.