An edge-guided hybrid CNN–transformer framework for accurate marine oil spill segmentation in synthetic aperture radar images

Synthetic Aperture Radar (SAR) imagery plays a crucial role in marine safety and environmental monitoring, particularly for detecting and mapping marine oil spills. However, accurate oil spill segmentation from SAR images remains challenging due to speckle noise, low-contrast boundaries, and the presence of look-alike dark features such as low wind areas and biogenic films. To address these issues, this study proposes a hybrid CNN–Transformer architecture incorporating edge-guided and uncertainty-aware learning for improved marine oil spill segmentation. The proposed model combines a convolutional neural network encoder to capture local texture information with a transformer module to model long-range contextual dependencies. A cross-attention fusion mechanism is introduced to effectively integrate multi-level features. To enhance boundary delineation, an edge guidance module is employed to emphasize spill contours during training. In addition, an uncertainty refinement head identifies ambiguous regions and improves segmentation reliability. The model is evaluated on a SAR oil spill dataset containing 6,455 training images and 1,615 testing images. Experimental results demonstrate superior performance, achieving a Dice score of 0.942, IoU of 0.891, precision of 0.951, recall of 0.928, and accuracy of 0.968, outperforming U-Net, DeepLabV3+, Attention U-Net, and TransUNet. Qualitative and ablation analyses confirm improved boundary preservation and reduced false detections.