Automatic eddy detection in the MIZ based on YOLO algorithm and SAR images

The automatic detection and analysis of oceanic eddies within the marginal ice zone using synthetic aperture radar present significant challenges, yet are crucial for both scientific research and practical applications. Thus, we explored the feasibility of automating the eddy detection process by applying YOLOv8, a state-of-the-art computer vision model, to high-resolution synthetic aperture radar data, specifically targeting the dynamic region of the Fram Strait. We specifically aim to distinguish between two distinct classes of eddies, based on their rotational direction: cyclonic and anticyclonic. The accurate identification of these eddy types is particularly important for collecting extensive statistical datasets, which are vital for understanding long-term oceanographic patterns and their impact on the Arctic climate. By fine-tuning of YOLOv8 model on an accurately labeled dataset, we achieved robust detection results with minimal training data. The performance of the different architectures within the model was evaluated using various metrics, and the best-performing one was selected through visual inspection and quantitative analysis. Experiments demonstrated the model’s robustness and precision in reliably identifying and distinguishing between cyclonic and anticyclonic eddies with different scales and in different sensing conditions. This work represents a significant advancement in automated eddy detection within the marginal ice zone, offering new insights into the dynamics of polar oceanography.