Oceanic eddies are ubiquitous phenomena that play a crucial role in linking large- and small-scale ocean dynamics. We propose a deep learning approach to detect oceanic eddies (radius >20 km) directly from daily sea level anomaly (SLA) data at 0.125 deg grids from the Haiyang-2 (HY-2) satellite, eliminating the conventional need for geostrophic current field computation. Two deep learning models, i.e., the U-shaped network (U-Net) and You Only Look Once v8 (YOLOv8), are trained using extensive samples from 2022 to 2024, aligned with current field simulations from the Copernicus Marine Environment Monitoring Service reanalysis in the Northwest Pacific. Using the daily HY-2 SLA product in 2024, two key parameters of mesoscale eddies in the Northwest Pacific, i.e., the geographic location of the center and the radius, are obtained. For comparison, mesoscale eddies (radius >100 km) are also identified using a vector geometry algorithm applied to AVISO SLA data, derived from geostrophic dynamics. The correlation of bias distribution (AVISO minus deep learning methods) indicates YOLOv8 has better performance than U-Net due to high correlation across the <30 km bias range. YOLOv8-based eddy identification in 2024 reveals distinct seasonal patterns: anticyclonic eddies exhibit pronounced zonal distribution relative to Japan, whereas cyclonic eddies demonstrate clear meridional distributions that are most frequently observed near 21°N.
Leng et al. (Mon,) studied this question.