The variation of underwater sound velocity greatly influences the distribution of acoustic fields for underwater sound signals, since the irregular distribution of sound velocity dictates the route that signals take as they propagate. To achieve efficient underwater communication and accurate positioning, it is essential to accurately and real-time obtain the distribution of sound velocity. For the method of obtaining sound velocity distribution, sound speed profile (SSP) inversion methodologies provide more rapid response times in comparison to direct measurement techniques. Nevertheless, these approaches predominantly concentrate on the development of spatial sound velocity fields and are significantly dependent on sonar observation data. Consequently, they impose stringent requirements on the implementation of data acquisition systems. In order to effectively model the variation patterns of SSPs and facilitate SSP prediction independent of sonar observation data, we propose a long short-term memory (LSTM) neural network enhanced by an attention mechanism (Att-LSTM). The proposed methodology facilitates the estimation of SSPs without the necessity for on-site data collection, which consequently leads to a reduction in time expenditure. The Att-LSTM demonstrates a root mean square error (RMSE) of less than 0.6 m/s in forecasting the monthly average SSP, which exhibits an average accuracy improvement of 23.5% and 33.0% when compared to LSTM and gated recurrent unit (GRU) neural networks, respectively. These results substantiate the efficacy of the proposed approach in predicting sound velocity distribution.
20 October 2025 in IEEE Geoscience and Remote Sensing Letters
The distribution of underwater sound speed is crucial for determining the propagation mode of underwater acoustic signals. In traditional works for predicting sound speed profiles (SSPs), attention is seldom paid to the situation under the influence of internal waves, due to the scarcity of data. However, the sound velocity under the influence of internal waves exhibits distribution characteristics at different time scales, making it difficult to be predicted by existing models, such as the long short-term memory neural network (LSTM). To address these issues, we propose a double basis function Kolmogorov–Arnold network (DBF-KAN) model for predicting sound speed distribution with the presence of internal waves in this paper. Different basis functions are used to capture features of time series at different scales. Real data experiments show that the DBF-KAN network can capture the changing characteristics of complex sound velocity fluctuations and obtain good accuracy performance.