Deep unsupervised adversarial domain adaptation for underwater source range estimation

Abstract

In this study, an underwater source range estimation method based on unsupervised domain adaptation (UDA) is proposed. In contrast to traditional deep-learning frameworks using real-world data, UDA does not require labeling of the measured data, making it more practical. First, a classifier based on a deep neural network is trained with labeled simulated data generated using acoustic propagation models and, then, the adaptive procedure is applied, wherein unlabeled measured data are employed to adjust an adaptation module using the adversarial learning algorithm. Adversarial learning is employed to alleviate the marginal distribution divergence, which reflects the difference between the measured and theoretically computed sound field, in the latent space. This divergence, caused by environmental parameter mismatch or other unknown corruption, can be detrimental to accurate source localization. After the completion of the adaptive procedure, the measured and simulated data are projected to the same space, eliminating distribution discrepancy, which is beneficial for source localization tasks. Experimental results show that range estimation based on UDA outperforms the match-field-processing method under four scenarios of few snapshots, few array elements, low signal-to-noise ratio, and environmental parameter mismatch, verifying the robustness of the method.