Boundary-aware dual-attention U-Net for multi-class oil spill segmentation in SAR imagery

Oil spills pose serious environmental risks, hence accurate and rapid detection is critical for marine monitoring. Synthetic Aperture Radar (SAR) imagery is commonly utilized for oil spill detection due to its all-weather, day-and-night imaging capacity. However, speckle noise, low contrast, and uneven spill boundaries make it difficult to distinguish between oil spills and visually similar look-alike occurrences. In this paper, we present a deep learning-based system, Dual-Attention Bottleneck U-Net (DAB-UNet), for multi-class oil spill identification and segmentation utilizing Sentinel-1 SAR imagery from the publicly accessible M4D dataset. The proposed framework is based on a U-Net encoder-decoder architecture that includes a Dual Attention Block implemented with the Convolutional Block Attention Module (CBAM) to improve oil-related features while lowering background and look-alike interference. A boundary-aware learning technique is used along with a composite loss function that combines categorical cross-entropy, Jaccard loss, Tversky loss, and boundary loss to solve class imbalance and complex spill shapes. Experimental results reveal that the suggested method outperforms baseline and current approaches, with an overall accuracy of 96.19%, precision of 95.00%, F₁-score of 95.94%, along with improved recall and fewer erroneous detections. The results reveal that DAB-UNet is a reliable and effective method for identifying and segmenting multi-class oil spills in SAR data.