Real-time estimation of underwater sound speed profiles with a data fusion convolutional neural network model

Abstract

Sound speed profiles (SSPs) characterize the spatial distribution of sound velocity, significantly impacting the trajectory of acoustic signals. Consequently, they play a crucial role in determining the efficiency and precision of communication, localization, and other practical applications. Recently, SSP inversion methods significantly improve real-time performance compared to the direct measurement method, but it is highly dependent on sonar observation data so that the application scope is limited. To obtain real-time sound speed distribution in larger scale areas, we proposes a data-fusion-driven multi-input multi-output convolutional regression neural network (DF-MIMO-RCNN) , which combines satellite-based real-time remote sensing sea surface temperature (SST), historical SSP feature vector, historical mean temperature below the sea surface with corresponding spatial coordinate information. The model gets rid of the dependence on sonar observation data and can be applied to a wider spatial region. Then two sets of experiments were conducted based on the Argo database and marine experimental data respectively. Experimental results based on Argo data show that the proposed method is not only suitable for deep-sea environments at the same depth, but even for shallow water of different depths. At the same time, compared with other algorithms, the root mean square error (RMSE) and RMSE distributions are smaller and more robust. Finally, we conducted deep-sea experiments in the South China Sea. The results show that even in non-strictly gridded sea areas, our proposed model still has lower RMSE and higher robustness, which demonstrates the superiority of our proposed algorithm.