Frequency difference multi-branch multi-task learning for underwater source localization in mismatched environments

Model mismatches including array tilt and sound speed profile (SSP) mismatch are common in practice and degrade the performance of model-based deep learning methods for underwater source localization. This paper proposes a frequency-difference-based multi-branch multi-task learning method for source range and depth estimation, which uses frequency-difference processing to reduce mismatch effects and multi-branch feature fusion to improve localization accuracy. Trained on simulated data from a non-tilted array and average SSP, the network performs well on mismatched test sets and generalizes to experimental data.