Multi-dimensional Attention-Based MOSSM Model for Marine Oil Spill Monitoring in SAR image Remote Sensing

Abstract

Marine oil spills pose severe threats to marine ecosystems, where rapid and accurate oil spill region segmentation is crucial for emergency response to disasters. Synthetic Aperture Radar (SAR), with its all-weather and day-night observation capabilities, serves as a vital data source for oil spill monitoring. However, SAR images are susceptible to the interference of speckle noise and complex background, limiting the accuracy of traditional segmentation methods. Therefore, this paper based on DeepLabV3 + proposes an enhanced model—Marine Oil Spill Segmentation Model (MOSSM). Firstly, the SE (Squeeze-and-Excitation) channel attention mechanism is introduced into the Bottleneck structure of ResNet50 to enhance the ability to extract critical features in oil spill regions by dynamically adjusting feature channel weights. Secondly, a High-Low Feature Fusion Module (HLFusion) is designed with spatial, channel, and pixel attention mechanisms incorporated to optimize detail preservation during high-low feature fusion. It effectively mitigates the interference between SAR image noise and background. Experiments based on the SOS oil spill dataset (containing 8,070 SAR images of 256 × 256 pixels) demonstrate that the indexes of MOSSM, including Intersection over Union (IoU) (74.47%), pixel accuracy (91.74%), and recall (85.91%), significantly outperform those of U-Net, FCN, SegNet, and the original DeepLabV3 + . The model particularly excels in segmenting complex oil spill boundaries and scattered oil slicks. This research provides a higher-precision and more robust segmentation method for SAR image oil spill monitoring, offering significant application value for marine environmental disaster warning and emergency decision-making.