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

Abstract

Accurate daily three-dimensional (3D) ocean sound speed prediction is essential for underwater acoustic applications. Existing data-driven studies often focus on global field errors, leaving localized thermocline-related errors insufficiently addressed. Using ten years of daily GLORYS12 reanalysis data from an equatorial Pacific region, we developed a 3D ConvLSTM predictor for background spatiotemporal evolution. We then proposed the Full-Depth Dynamic Gradient-Guided Residual Correction Network (FDGRC-Net), which derives a continuous mask from the previous-day regional temperature gradient profile to guide residual learning and output gating. On the 2009 test set, FDGRC-Net reduced the overall RMSE from 0.4209 to 0.3916 m/s and the MAE from 0.2302 to 0.2180 m/s. It also demonstrated improved prediction performance in the northern South China Sea cross-region evaluation and achieved annual RMSE reductions of 5.12–7.90% in frozen 2010–2015 tests. Against 521 collocated Argo profiles, it slightly but consistently improved the observational agreement. In a controlled Bellhop diagnostic, the transmission loss MAE decreased from 2.825 to 2.489 dB and the mean arrival time error from 4.099 to 3.653 ms. These results demonstrate that FDGRC-Net provides an effective and physically interpretable approach to thermocline-aware daily 3D sound speed prediction.