Alleviating negative transfer using unsupervised adversarial partial domain adaptation for underwater source range estimation

An underwater source range estimation method based on unsupervised partial domain adaptation (PDA) is proposed. Unlike existing unsupervised domain-adaptation-based methods (UDA), this approach assumes the measured data's range space as a subset of the simulated data's, which is more representative of practical scenarios. Under this configuration, UDA suffers from severe performance degradation which is known as negative transfer. To address this issue, we incorporate a PDA-based simulated data weighting mechanism into the UDA framework. The deep neural network processes measured data, outputting a probability distribution across the range space. The expected probability across the entire measured dataset is employed as the weight for the domain discrimination loss associated with the simulated data. Consequently, simulated data that share a nearly identical space with the measured data are given greater importance, while other simulated data receive less emphasis. As a result, the negative transfer is mitigated. Experimental results demonstrate that when the range space of the measured data significantly differs from that of the simulated data, the PDA-based range estimation method outperforms UDA and exceeds the performance of matched-field processing across various scenarios, including different adaptation data sampling methods and mismatches in environmental parameters, thereby highlighting the effectiveness of the proposed method.