An Enhanced Oil Spill Detection System Using Explainable Ai (XAI) and Transfer Learning on Synthetic Aperture Radar (SAR) Imagery

Abstract

Marine petroleum spills caused serious ecological degradation. It is urgent for automation early oil spill detection and identification. Although SAR is an active microwave sensor providing continuous day-and-night operation in adverse conditions such as cloudy days or night, the difficulty for differentiating between oil slicks and natural ocean look-alike has increased since their SAR backscatter properties are quite similar. This paper presents an interpretable deep neural network model for pixel-level oil slick segmentation that leverages a U-Net with a ResNet50 (pre-trained) feature extractor and Spatial/Channel Squeeze-and-Excitation (SCSE) attention modules. Gradient-weighted Class Activation Mapping (Grad-CAM) was used in the meantime to improve the visual transparency of automatic decisions to find important input regions by showing where and why the decision was made (for example, a pixel-level region of an oil spill region was found because its internal convolutional layer recognized a particular pattern). Through utilizing an official Sentinel-1 SAR dataset for testing and validation, the model reached an overall accuracy of 98.65% and mean Intersection over Union (mIoU) of 0.894, showing that accurate and reliable remote sensing-based oil spill monitoring is improved by the combination of attention processes and transfer learning.