ORIGINAL RESEARCH article
Front. Mar. Sci.
Sec. Physical Oceanography
Volume 12 - 2025 | doi: 10.3389/fmars.2025.1551823
A Deep Learning-based Reconstruction Model for 3D Sound Speed Field Combining Underwater Vertical Information
Provisionally accepted
- 1National University of Defense Technology, Changsha, Hunan Province, China
- 2Jiangsu Ocean Universiity, Lianyungang, China
- 3Shanghai Ocean University, Shanghai, Shanghai Municipality, China
Select one of your emails
You have multiple emails registered with Frontiers:
Notify me on publication
Please enter your email address:
If you already have an account, please login
You don't have a Frontiers account ? You can register here
The sound speed in the ocean significantly influences the propagation characteristics of underwater acoustic signals. Rapid acquisition of underwater three-dimensional (3D) sound speed fields is essential for target detection, acoustic communication, and underwater navigation. The usual used single empirical orthogonal function regression (sEOF-r) method, which reconstructs sound speed profiles (SSP) by establishing regression relationships between empirical orthogonal coefficients of SSP and sea surface environmental factors, may have several limitations: (1) The principal modes extracted by the EOF method may lose some sound speed information, resulting in low reconstruction accuracy; (2) The grid-by-grid inversion of SSP is computationally inefficient for acquiring large-scale 3D sound speed fields; (3) Oceanic dynamic activities cause disturbances in the sound speed field, and relying solely on sea surface environmental information can limit the accuracy of full-ocean-depth sound speed inversion. To address these issues, with the use of multi-source satellite remote sensing data and CMEMS temperature and salinity reanalysis data, this paper simultaneously incorporates sea surface environmental factors (SST, SLA, and EKE) and underwater information (average density) as inputs, and proposes a region-oriented reconstruction model (3dCNN-DEN) for 3D sound speed fields using the Convolutional Neural Network (CNN). Experimental results show that the 3dCNN-DEN achieves an average root mean square error (RMSE) of 0.7572 m/s and an average mean absolute error (MAE) of 0.5759 m/s, significantly outperforming conventional EOF-based methods. Incorporating underwater average density improves reconstruction accuracy, showing a 77.1% and 60.3% improvement over the sEOF-r and sEOF-CNN methods, respectively.
Keywords: Sound speed reconstruction, Convolutional Neural Network, average density, Empirical orthogonal function method, sea surface environmental information
Received: 26 Dec 2024; Accepted: 19 May 2025.
Copyright: © 2025 LI, Liu, Li, Wang, Zhu, Mao and Chen. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
* Correspondence: Ming Li, National University of Defense Technology, Changsha, Hunan Province, China
Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.