146 / 2023-08-20 09:42:06

Deep Learning Based Oceanic Mesoscale Eddy Detection in the South China Sea

oceanic mesoscale eddy,deep learning,feature fusion,satellite altimetry,sea surface temperature anomaly,South China Sea

Accurate automatic eddy detection is crucial for monitoring the dynamics of mesoscale eddies. In this study, we proposed a deep learning model, which combines attention mechanism with a U-shaped network, namely Attention U-net network, to capture the daily occurrence of mesoscale eddies in the South China Sea (SCS) from multi-satellite observations of surface variables including absolute dynamic topography and sea surface temperature anomaly. Eddies from Archiving, Validation and Interpretation of Satellite Oceanographic data (AVISO) were used for training the network and evaluating the accuracy of estimated results. The Attention U-net model shows an excellent performance for the mesoscale eddy detection in the SCS. The accuracy between the estimated eddy signals and AVISO dataset is as high as 98%, and the average F1 score is 0.8874. Besides, the additional input of sea surface temperature anomaly helps to improve the detection accuracy of mesoscale eddies in the SCS, and the miss rates of cyclonic and anticyclonic eddies are reduced by 12.2% and 7.0%, respectively. The deep learning model has the strong ability to detect more smaller-scale eddies, which can provide the important complement to the widely used mesoscale eddies dataset.