A multi-task neural network for source localization in shallow-water environment with depth classification

Limited high-quality training data can cause overfitting and weak generalization in deep-learning-based source localization, leading to degraded performance in realistic environments. In this study, a multi-task neural network with a depth-classification branch is developed, motivated by the depth resolution of a vertical line array and the depth-dependent structure of the localization ambiguity surface. The depth-classification branch regularizes depth estimation and reduces model complexity, thereby mitigating overfitting in the range-regression branch. Transfer learning from simulations to experiments further improves performance with limited measurements. The developed network yields improved localization performance, as demonstrated using both simulated and experimental data.