Deep learning detection of submesoscale eddies from MODIS SST in the South China Sea

Submesoscale eddies play crucial roles in oceanic energy cascades and material transport; however, their detection faces substantial challenges owing to inadequate spatial resolution in traditional satellite altimetric measurements. This research introduces Sub-Eddy-Net, an innovative deep learning approach utilizing an enhanced YOLOv8-SGAM architecture for identifying submesoscale eddy features within MODIS-derived sea surface temperature (SST) imagery across the South China Sea region. By exploiting SST data's advantageous temporal coverage and observational continuity relative to altimetric sensors, synthetic aperture radar, and ocean color products, the developed algorithm demonstrates detection precision of 0.968, successfully identifying 20,443 submesoscale eddies throughout 2015–2024. Results indicate cyclonic eddy prevalence exceeds anticyclonic occurrences, exhibiting average radii of 13.1 km and 13.7 km respectively. Eddy distributions primarily cluster within coastal waters and Luzon Strait vicinity, displaying pronounced seasonal patterns. Comparison with Surface Water and Ocean Topography (SWOT) observations shows 64.3% matched features, 27.1% positional offsets, and 8.6% detected only in SST. Within the same eddies, coherent chlorophyll-a (Chla) signatures were present in a smaller fraction, indicating that SST is the more robust tracer for systematic detection. This Sub-Eddy-Net methodology offers a scalable approach for examining submesoscale oceanic phenomena, providing a decadal inventory for further study of their roles in ocean circulation and vertical exchange.