Addressing Ship Detection Challenges: From Semantic Segmentation Limitations to YOLOv8 Success

Accurate ship detection in satellite imagery is essential for maritime applications. This paper addresses the challenges of applying UNet-based semantic segmentation, which exhibited limitations including a learning rate plateau (maximum Dice coefficient: 0.75805). To overcome these, the study transitioned to YOLOv8-based object detection. YOLOv8 demonstrated improved performance, achieving mAP50-95 values of 0.285, 0.328, and 0.387 for 5, 10, and 20 epochs, respectively. The findings highlight YOLOv8's efficacy as a robust alternative to semantic segmentation for ship detection.