Full-Depth Dynamic Gradient-Guided Residual Correction for Daily Three-Dimensional Ocean Sound Speed Prediction

Area of Science:

• Oceanography
• Acoustics
• Data Science

Background:

• Accurate 3D ocean sound speed prediction is vital for underwater acoustic applications.
• Existing data-driven methods often overlook localized thermocline errors.

Purpose of the Study:

• To develop an effective method for daily 3D ocean sound speed prediction that addresses thermocline-related errors.
• To improve the accuracy and interpretability of sound speed predictions in oceanic regions.

Main Methods:

• Utilized ten years of GLORYS12 reanalysis data from the equatorial Pacific.
• Developed a 3D ConvLSTM for background spatiotemporal evolution.
• Proposed the Full-Depth Dynamic Gradient-Guided Residual Correction Network (FDGRC-Net) incorporating a gradient-guided mask for residual learning.

Main Results:

• FDGRC-Net reduced overall RMSE from 0.4209 to 0.3916 m/s and MAE from 0.2302 to 0.2180 m/s on the 2009 test set.
• Demonstrated improved cross-region prediction in the South China Sea and reduced RMSE by 5.12-7.90% in 2010-2015 tests.
• Improved observational agreement with Argo profiles and reduced transmission loss MAE and mean arrival time error in Bellhop diagnostics.

Conclusions:

• FDGRC-Net offers an effective and physically interpretable approach for thermocline-aware daily 3D sound speed prediction.
• The gradient-guided residual correction significantly enhances prediction accuracy, particularly in areas with strong thermal gradients.