Enhancing Surface Current Retrieval From Sequential Imagery With Multiscale Deep Learning

Accurate retrieval of ocean surface currents from sequential satellite imagery is crucial for advancing our understanding of upper-ocean dynamics and their role in climate variability. In this study, we propose CN-RAFT, a novel deep learning framework that integrates a multi-scale coarse-to-fine optical flow strategy based on the ConvNeXt architecture and Recurrent All-Pairs Field Transforms (RAFT), to enhance surface current estimation from sequential sea surface temperature (SST) observations. The performance of CN-RAFT is rigorously evaluated against two benchmark methods: the RAFT optical flow model and the correlation-based Window Deformation Iterative Multigrid (WIDIM) technique, using high-resolution SST and surface velocity fields from the MITgcm simulation. CN-RAFT consistently outperforms the benchmark methods, achieving lower root-mean-square error (0.9). Sensitivity experiments further demonstrate its robustness under varying spatial resolutions and noise conditions. When applied to real Himawari-9 SST data in the Kuroshio Extension, CN-RAFT successfully resolves fine-scale eddies and frontal meanders, exhibiting superior agreement with SST gradient structures. Our findings highlight CN-RAFT’s potential as an effective tool for high-resolution, real-time ocean surface current monitoring from satellite observations.