Enhancing Underwater DOA Estimation Accuracy With Limited Datasets Using Task-Restructured Deep Mutual Learning

This paper aims to address the issue of low accuracy in underwater Direction of Arrival (DOA) estimation using Deep Learning (DL) methods, which arises due to the scarcity of underwater data caused by the difficulties in conducting underwater experiments. For multi-snapshot sampled signals, we segment the snapshots and reconstruct the task into a problem of processing few-snapshot data within an expanded dataset. By utilizing the new task, we employ a deep mutual learning (DML) model to enhance the accuracy of the original task's DOA estimates. Experimental results demonstrate that under conditions of small and limited datasets, our approach effectively improves the accuracy of DL-based DOA estimation methods.