YOLO Based Ship Detection in SAR Images: In-Depth Analysis on HRSID, RSDD and SSDD Datasets

Accurate ship detection is essential for effective monitoring, control, and management of maritime systems. In this study, we conduct a comprehensive evaluation of advanced deep learning models from the YOLO family-specifically YOLOv10 and YOLOv11-applied to Synthetic Aperture Radar (SAR) imagery. Using three established datasets (HRSID, RSDD, and SSDD), we analyze key performance metrics, including precision, recall, and mean average precision (mAP). Our experimental results demonstrate that YOLOv11 variants consistently outperform YOLOv10 models across all datasets. Notably, YOLOv11l achieved the highest mAP50-95 of 0.686 and a precision of 0.905 on the HRSID dataset, indicating superior detection capabilities in densely populated maritime scenes. On the RSDD dataset, YOLOv11n attained the highest recall of 0.899 and a mAP50 of 0.964, while YOLOv11l recorded the highest precision at 0.957, reflecting robust performance in handling diverse ship orientations and environmental conditions. Furthermore, on the SSDD dataset, YOLOv11l achieved a mAP50 of 0.971 and mAP50-95 of 0.706, effectively detecting small, isolated ships in cluttered backgrounds. These enhancements in feature extraction and noise suppression are particularly valuable for real-world maritime applications, where overlapping targets and environmental clutter frequently challenge detection accuracy. The robustness and high accuracy of the proposed models, incorporating automatic hyperparameter tuning, transfer learning, and hard negative mining, thereby enhance maritime safety and environmental monitoring.