Sound source localization with sparse Bayesian-based feature matching via deep transfer learning in shallow sea

Intelligent and accurate underwater sound source localization in shallow sea remains a prominent research issue in the field of underwater engineering. Traditional localization methods are limited in computational efficiency. And deep learning methods are constrained by model interpretability. To address these dual challenges, a model combining the sparse Bayesian-based feature matching with deep transfer learning (DTL) via convolutional neural network (BFM-CNN), is designed for accurate and effective source localization in shallow sea. The proposed method is verified on four datasets of Swellex-96. The mean absolute percentage error of range localization is 3.3%, 1.8%, 2.4% and 2.2%, respectively, with the best accuracy of 99.4%. This method provides a promising approach for the practical applications of underwater source localization.