High-precision segmentation of marine oil spills based on an improved Mask2Former framework

Accurate segmentation of oil spills in sea surface imagery is of critical practical importance for pollution detection, damage assessment, and emergency response. Nevertheless, marine scenes pose substantial challenges, including severe background clutter, pronounced class imbalance, multi-scale target variation, and indistinct object boundaries, which collectively limit the accuracy and robustness of existing segmentation models. To address these challenges, this study proposes an enhanced semantic segmentation framework based on Mask2Former, incorporating five dedicated modules. The Multi-Scale Feature Enhancement (MSFE) module strengthens multi-scale target representation through scale-adaptive weighting and residual scale augmentation. The Oceanic Context Aggregation (OCA) module enriches mask features by integrating directional strip pooling, a multi-scale context pyramid, and noise-aware gating mechanisms. The Class-Balanced Query Attention (CBQA) module alleviates class imbalance via class-aware query reweighting. The Boundary Refinement Module (BRM) and the Adaptive Boundary Upsampling (ABU) module jointly enhance mask predictions in boundary regions through residual refinement and boundary sharpening, respectively. Comprehensive comparative and ablation experiments conducted on the LADOS dataset demonstrate that the proposed method achieves superior performance over established baselines, including classical segmentation models such as U-Net, FCN, and DeepLabv3+, attaining an mIoU of 74.48%, fwIoU of 76.05%, mACC of 83.56%, and pACC of 86.38%. Particularly pronounced improvements are observed in underrepresented categories such as oil platforms and ships. These results substantiate both the individual effectiveness and the complementary nature of the proposed modules.