Impact of Data Enhancement Methods on Ship Detection Using YOLO11 in SAR Imagery

Vessel detection from SAR imagery with applying deep learning approached applications has gained a significant attention as an alternative approach for both military and civilian applications in recent years. The effectiveness of these methods is directly proportional to the quality of the datasets, thereby improving ship detection performance. In this study, a new dataset named SARMSSD, built by combining the open-source SSDD, HRSID and SRSDD datasets, was analyzed using the YOLO11 deep learning model to evaluate the effects of various image denoising and data augmentation techniques. Three denoising methods - Median filtering, BM3D, and SAR-CAM - were applied. Ship detection experiments showed that the SARMSSD-SAR-CAM dataset achieved the highest [email protected]:0.95 score of 0.770, outperforming other denoised datasets. Based on this result, data augmentation experiments were conducted, where mosaic augmentation proved to be the most effective, yielding the highest mAP score of 0.788. These findings emphasize the importance of selecting optimal denoising methods to balance noise reduction, as well as the effectiveness of data augmentation, particularly mosaic augmentation, in enhancing deep learning-based ship detection performance in SAR imagery. The study on code realm will be accessed at https://github.com/buyukkanber/SARMSSD