Masking Modelling for Underwater Acoustic Target Recognition with Self-supervised Learning

Underwater acoustic target recognition (UATR) involves classifying targets in a marine environment using sonar signals. This task has been tackled using supervised learning techniques, including CNNs, RNNs, and Transformer models. However, these models often face challenges with generalization and performance due to limited labelled underwater signals and annotation difficulties. Simulated datasets for deep learning-based UATR algorithms are not practical. To overcome these issues, we propose a frame and patch masking Transformer (FPMT) combined with self-supervised learning (SSL) to exploit the high-level knowledge present in the underwater signals themselves. FPMT employs a hierarchical masking procedure on both frame and patch domains with acoustic Mel-spectrograms, allowing the model to learn supervision from joint time-frequency representations. Experimental results based on real-world underwater dataset demonstrate that our method outperforms state-of-the-art SSL methods in recognizing underwater acoustic targets.