Abstract: Traditional eddy detection methods based on SAR data require manually setting thresholds and feature parameters, which makes the process complex and difficult to automate. Existing deep learning models also suffer from high false negatives and false positives during detection, failing to meet the accuracy and efficiency requirements of eddy detection. To address these issues, this paper proposes an improved model based on YOLOv8, namely EddyDetNet, to overcome the above limitations and enhance both detection accuracy and efficiency. The model introduces an Adaptive Feature Compression Module (AFCM) in the Backbone and Neck to achieve lightweight and efficient feature extraction, and uses a Multi-scale Feature Spatial Pyramid Module (MFSP) to enhance the effect of multi-scale feature fusion. By optimizing the Neck structure and adding a small-object detection head in the Head part, the model improves the detection accuracy of eddies at different scales. Experimental results show that EddyDetNet outperforms YOLOv8 by 2.4%, 3.2%, and 5.5% in precision (P), recall (R), and mean Average Precision (mAP), respectively, while reducing the parameter size and computational complexity by 38.1% and 15.8%. Compared to YOLOv8, EddyDetNet reduces computational complexity and parameter size while maintaining high detection accuracy, making it suitable for eddy detection tasks in multi-target and complex background scenarios.

Key words: Oceanic eddies, Synthetic Aperture Radar, Deep learning, YOLO, Object Detection