Deep Learning for Mesoscale Eddy Detection With Feature Fusion of Multisatellite Observations

Accurate oceanic eddy detection is crucial for understanding their dynamic behavior. In this study, we apply attention dual-U-net, a specialized deep learning model, to simultaneously detect the location and contours of mesoscale eddies in the South China Sea (SCS). This model integrates various features from satellite-observed absolute dynamic topography and sea surface temperature anomaly (SSTA), and is established separately for anticyclonic eddies (AEs) and cyclonic eddies (CEs) detection. Eddy contours from the delayed-time altimetric mesoscale eddy trajectories atlas are used for model training and evaluation. Results indicate that the model excels in detecting the shape and location of mesoscale eddies in the SCS, achieving success detection rates (SDRs) of 95.2% for AEs and 94.7% for CEs. Incorporating SSTA as an additional input enhances the accuracy of eddy shape and aids in further distinguishing normal from abnormal eddies. Abnormal eddies, characterized by cold AEs and warm CEs, constitute 16.8% and 29.8% of total AEs and CEs, respectively, with SDRs of 95.3% and 94.7%, underscoring the model robustness to abnormal eddies. Moreover, the mean absolute errors of AEs (CEs) are notably smaller than those estimated by the pyramid scene parsing network and EddyNet, with reductions of 49.1% (45.1%) and 67.6% (70.8%), respectively. These reductions are particularly pronounced in coastal areas and deep waters exceeding 200 m in depth. The amalgamation of the accurate eddy detection model and high-resolution multisatellite observations presents an effective approach to capturing eddy occurrences, contributing to a comprehensive understanding of eddy dynamics in marginal seas and open oceans.