An Improved YOLOv8-Based Detection Model for Multi-Scale Sea Ice in Satellite Imagery

Sea ice detection is of vital importance for maritime navigation. Satellite imagery is a crucial medium for conveying information about sea ice. Currently, most sea ice detection models mainly rely on texture information to identify sea ice in satellite imagery, while ignoring sea ice size information. This research presents an improved YOLOv8-Based detection algorithm for multi-scale sea ice. First, we propose a fusion module based on the attention mechanism and use it to replace the Concat module in the YOLOv8 network structure. Second, we conduct an applicability analysis of the bounding box regression loss function in YOLOv8 and ultimately select Shape-IoU, which is more suitable for sea ice, as the loss function for bounding box regression. Third, we analyze the distribution characteristics of sea ice with different sizes in the NWPU-RESISC45 dataset. Based on these distribution characteristics, the bounding box information predicted by YOLOv8 are converted into evidence vectors for uncertainty quantification. Subsequently, information fusion is achieved by fusing these vectors with the probability of sea ice categories. Compared to YOLOv8 and other detection algorithm, our improved YOLOv8 achieves better detection accuracy on both the NWPU-RESISC45 and the Landsat-8-derived Sea Ice datasets.