Abstract: The segmentation accuracy of marine oil spills from unmanned aerial vehicle (UAV) images is aimed to be enhanced. A segmentation and calibration method based on multi-scale attention fusion is proposed to achieve this. First, high-resolution images are rapidly detected and cropped using YOLOv5. Then, a multi-scale attention U-shaped network (MSA-UNet) is constructed, which integrates global semantic information from spatial pyramid pooling and local details from convolutional attention to enhance feature representation. Furthermore, the segmentation results for oil, water, and other classes are refined through the proposed hue-saturation-value region prototype-guided calibration mechanism (HSV-PGRM), which reduces interference from illumination variations and cloud cover. Experimental results on the De Kerf port dataset show that the proposed method achieves a mean intersection over union (mIoU) of 68.89% and an F1-score of 74.08%, with the F1-score improving by 3% over the U-shaped network (U-Net) baseline model. Good feature recognition capability is also demonstrated on an independently collected lake surface test set.