Remotely Sensed Data-Driven Deep Learning Assimilation for Wave in China Seas

Abstract

This letter proposes a U-shaped network (Unet)-based postprocessing data-assimilation framework to improve hourly significant wave height (SWH) hindcasts from the finite-volume community ocean model (FVCOM)–simulating waves nearshore (SWAN) in the China Seas. Unet is trained with 2024 SWH observations from Jason-3, Haiyang-2 (HY-2), and the surface waves investigation and monitoring (SWIM) instrument onboard the China–France Oceanography Satellite (CFOSAT), and applied to the 2014–2023 hindcasts. Validation against buoy observations during 2022–2023 shows that Unet reduces the root-mean-squared error (RMSE) from 0.65 to 0.53 m and increases the correlation coefficient (Cor) from 0.84 to 0.86. Unet also outperforms residual Unet (ResUNet), the Kalman filter (KF), and ensemble optimal interpolation (EnOI), whose RMSEs are 0.58, 0.63, and 0.64 m, respectively. Monthly mean SWH from the assimilated fields shows positive tendencies during 2014–2023, demonstrating the framework’s potential for regional decadal-scale variability analysis.