Oil Spill Segmentation in SAR Data Using ViT-UNet: Performance and Practical Insights

IEEE Access. 2026. Vol. 14. P. 121339–121357.

Denis Zuenko, Khaidarova I., Trofimova E.

Oil spill segmentation in Synthetic Aperture Radar (SAR) images is limited by noisy annotations in publicly available datasets and by architectural choices that interact with label quality in opposing directions. First, we introduce a manually refined version of the Deep-SAR Oil Spill (SOS) dataset, in which 36.25% of masks are corrected for false positives, missed spill areas, and coarse boundaries under a three-expert protocol (Fleiss’ κ=0.81 ). Second, we propose ViT-UNet, a hybrid architecture combining a pretrained ViT-Small encoder with a convolutional U-Net decoder, with skip connections placed at shallow, intermediate, and deep transformer blocks (layers 2, 5, 8, and 11) to match the multi-scale structure of SAR oil-spill signatures. To isolate the effects of annotation quality and architectural design, we evaluate nine models on both the original and refined labels, training each across three random seeds for 100 epochs. Dataset refinement has an architecture-dependent effect: transformer-based models gain 0.5–0.9 mean Intersection over Union (mIoU) points on the refined labels, while CNN attention models decline by 0.5–1.2 points, indicating prior overfitting to annotation artifacts. On the Refined-SOS validation split, ViT-UNet achieves the highest mIoU of 68.4% ± 0.2 %, outperforming SegFormer by 2.86 pp and task-specific DAENet and LRA-UNet by 4.83 and 5.17 pp, respectively. On the original Deep-SAR test split (776 PALSAR + 839 Sentinel-1), ViT-UNet ranks first by mIoU ( 0.6701 ± 0.0040 ) and Recall ( 0.7304 ± 0.0200 ), with a Recall margin of +14.58 pp over the next-best baseline, indicating that the proposed Recall-oriented design transfers without overfitting to the refined annotation distribution

Keywords: computer vision, Oil spill segmentation