Underwater source localization via active deep complex residual network in a shallow-water waveguide

The prevailing deep learning methods for source localization in underwater environments primarily rely on real-valued representation, and the data is not lost by retaining only the real part but probably not fully exploited. Moreover, extensive labeled sample requirements and limited adaptability to diverse environments pose significant obstacles to the widespread application of deep learning in this domain. In light of these issues, we propose an end-to-end active learning framework for underwater source localization, denoted as ADCRN, which reconstructs a deep complex residual network (DCRN-CCL) as its backbone. The framework unifies an adaptively weighted uncertainty–diversity query strategy with model-based transfer learning to improve generalization performance. A complex-domain circle loss and mean-based probabilistic fusion are further embedded to enhance complex-valued feature discrimination. Simulation and experimental results demonstrate that, compared to the deep learning algorithms utilizing real-valued representation, the proposed framework achieves approximately a 12% improvement in accuracy and reduces the mean absolute deviation by about 1.2. Furthermore, it maintains nearly the same performance as the DCRN-CCL on the SWellEx-96 dataset, while requiring labels for only 37.89% of the total samples and 69.87% of the training time.