CBENet: contextual and boundary-enhanced network for oil spill detection via microwave remote sensing

Abstract In microwave remote sensing images, oil spills are generally distributed at various scales with blurry boundaries. To accurately detect variable oil spills from microwave remote sensing images, especially from synthetic aperture radar (SAR) images, we developed a contextual and boundary-enhanced network (CBENet) for oil spill detection from SAR observation images. The CBENet employs an encoder-decoder architecture that includes an encoder, a contextual fusion module, and a decoder. The encoder-decoder architecture intrinsically captures both global and local features of oil spills through the downsampling and upsampling processes inherent to the framework. The contextual fusion module enhances the contextual feature fusion using parallel dilated convolution branches. Furthermore, the utilization of a boundary-enhanced loss function further improves detection by focusing on the precise identification of oil spill boundaries. These properties strengthen the CBENet to effectively detect oil spills with blurry boundaries. The effectiveness of the CBENet has been validated through comprehensive empirical experiments including qualitative and quantitative evaluations. Comparative analyses demonstrate that CBENet outperforms several state-of-the-art detection models, significantly enhancing oil spill detection accuracy from SAR observation images.