Deep learning for submesoscale surface flow retrieval from geostationary satellite observations

A wide range of problems in oceanic mass and energy transport involve learning submesoscale surface flow fields from diurnal geostationary satellite observations. Yet, traditional methods, such as the Maximum Cross-Correlation (MCC) algorithm, suffer from limited spatiotemporal resolution and extensive post-processing. Here, we present the RAFT-Ocean architecture, a deep neural network-based approach for learning submesoscale flow fields in pixel-to-pixel manner, to retrieve submesoscale surface flow fields from geostationary satellite data. Compared to the MCC algorithm, the RAFT-Ocean architecture significantly improves these methods, reducing the end-point error (EPE) uncertainty by more than 65% and the absolute angular error (AAE) by more than 55%. The RAFT-Ocean architecture, when transferred to the geostationary ocean color satellite (GOCI/CMOS and GOCI-II/GK2B) sea surface chlorophyll-a products for diurnal hourly flow field retrieval, produced more realistic, continuous, and refined sea surface flow field data compared to geostrophic flow data from altimeter data. The refined diurnal hourly flow field matched well with the filamentous structure of surface phytoplankton, demonstrating an advantage in spatiotemporal resolution for kinetic energy transfer across scales. This approach enhances flow field retrieval quality and opens new avenues for real-time marine environment monitoring and modeling.