Multiscale dynamics-informed deep learning reconstruction of global ocean offshore near-surface current

Achieving exceptional accuracy and high-resolution reconstruction of sea surface currents is critical for investigating complex oceanic dynamic processes. Traditional approaches rely on physical differential equations and assimilation methods specifically designed for the global ocean, utilizing geostrophic and Ekman currents from satellite observations of sea level and wind stress fields. Nevertheless, dynamic processes such as mesoscale eddies and small-scale waves in marine environments remain significant challenges to the precise reconstruction of oceanic currents, and neglecting wave motion contributions renders any surface current product fundamentally incomplete. In this paper, we present an accurate method for reconstructing sea surface currents at a depth of 15 m. It is designed using a multiscale dynamics-informed neural network that captures the complex interactions among Geostrophic currents, Ekman currents, Stokes drift, and Tide currents, herein named GESTNet. It is designed based on the complex coupling relationships among diverse ocean surface physical processes and the actual conditions of the marine environment. Additionally, it integrates tree-based models and neural networks, enabling the model to learn intricate representations of latent features across different current components. GESTNet outperforms existing products and models significantly. It achieves a root-mean-square error (RMSE) of 7.029 cm/s, which represents an approximate reduction of 15 cm/s compared to current solutions. The improvement is even more substantial in the equatorial region, where GESTNet delivers a performance boost of around 37%. These results verify the feasibility of exploiting the interaction among multiscale ocean dynamical processes to develop deep learning models for reconstructing ocean surface currents.