AUTHOR=Gao Jianhao , Zhou Feng , Tian Di , Zhou Muping , Guo Hailong 

TITLE=Identification of mesoscale eddies based on improved YOLOv8 model: a case study in the South China Sea

JOURNAL=Frontiers in Marine Science

VOLUME=Volume 12 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/marine-science/articles/10.3389/fmars.2025.1569781

DOI=10.3389/fmars.2025.1569781

ISSN=2296-7745

ABSTRACT=Mesoscale eddies play a crucial role in energy transfer and material transport in the ocean. Accurate identification of mesoscale eddies is crucial for a deeper understanding of ocean internal dynamics, the development of marine resources, and the prediction of changes in the marine environment. This study utilizes Absolute Dynamic Topography (ADT) data provided by AVISO and the YOLOv8 algorithm model to investigate the identification of mesoscale eddies in the South China Sea (SCS). Due to its feature analysis and generalization capability, the YOLOv8 can successfully captures some mesoscale eddies undetected by the PET, thus track more mesoscale eddy trajectories. By enhancing the model’s input features and loss function, the YOLOv8 algorithm model has achieved high-precision identification of mesoscale eddies in the SCS with 93.9% Recall and 96.4% AP0.5, radius and amplitude average errors kept under 5 km and 0.50 cm. The incorporation of sea surface current field has improved the characteristics of mesoscale eddies, resulting in a smaller bias. However, due to some obscured ADT information, there was a slight increase in the identification errors for eddies’ amplitude and radius. Under typhoon events, the model accurately captures the evolution of mesoscale eddy characteristics, demonstrating high reliability. The model’s high accuracy (90.5% Recall, 93.6% AP0.5) for the transfer application in the Arabian Sea. Moreover, its accuracy in the transfer application to high-resolution products is also commendable. After only a few additional training rounds, the model achieves a high level of accuracy (90.0% Recall, 94.9% AP0.5), highlighting its robust generalization capabilities and transfer potential. This study suggests that the improved YOLOv8 algorithm enables threshold-free identification of mesoscale eddies with strong prospects for generalization and transfer applications which are expected to provide richer and more accurate mesoscale eddy track data.