A deep learning framework for four-dimensional ocean sound speed field prediction.

Area of Science:

• Oceanography
• Geophysics
• Artificial Intelligence

Background:

• Accurate prediction of ocean sound speed fields (SSFs) is vital for underwater communication, marine exploration, and environmental monitoring.
• Deep learning models show promise for SSF prediction but struggle with high-dimensional data, limiting them to 3D feature extraction and incomplete spatiotemporal information capture.

Purpose of the Study:

• To develop a novel deep learning model for four-dimensional (4D) SSF prediction, capturing complete spatiotemporal information.
• To enhance the accuracy and capabilities of existing SSF prediction methods.

Main Methods:

• Proposed the Swin Transformer-UNet (ST-UNet) model, integrating U-Net and Swin Transformer networks.
• Utilized Swin Transformer for spatiotemporal feature extraction via multi-head self-attention.
• Employed U-Net to refine spatial details through convolutional feature recovery.

Main Results:

• The ST-UNet model achieved a root mean square error of 0.783 m/s for 24-hour SSF prediction using 7-day historical data from the South China Sea.
• Demonstrated superior performance compared to baseline architectures, with improvements ranging from 33% to 72%.

Conclusions:

• The ST-UNet model effectively predicts 4D SSFs, outperforming existing methods.
• This advancement holds significant potential for improving underwater communication, marine resource exploration, and environmental monitoring.