Attention-driven acoustic properties learning for underwater target ranging

Underwater acoustic ranging (UAR) plays a crucial role in estimating object distances for ocean exploration. However, a reliable UAR method remains elusive, with current approaches either being reliant on inadequate hand-crafted features or neglecting the unique underwater acoustic properties. To address this, we propose Multi-attentional Underwater Acoustic Ranging (MUAR), a highly effective and robust UAR framework. MUAR incorporates multiple attention mechanisms tailored to the acoustic properties. Specifically, to better leverage the rich channel information in UAR data, we design a grouped channel attention module that can efficiently capture informative channels of the input data. Then, a feature-balancing strategy based on spatial-attention is introduced to mitigate information redundancy and conflicts, thereby enhancing the multi-level expressive capability of the model. We further theoretically analyze the connection between the self-attention mechanism and the acoustical signal correlations, such that achieving a better interpretation for the extracted features. Through extensive experiments and analysis on three authentic datasets, we show that MUAR outperforms previous approaches by obtaining state-of-the-art performance, i.e, achieving a MSE of 0.44 (vs. 2.72) and a MAPE of 0.97 (vs. 2.42). The source code of the proposed MUAR is released at https://github.com/TiernosChu/MUAR.