NeuroSpill: Attention U-Net++ Based Oil Spill Segmentation in SAR Imagery

Abstract: Marine oil spills pose a significant threat to ocean ecosystems, marine biodiversity, and coastal environments. Rapid detection and monitoring of oil spills are essential for minimising environmental damage and enabling effective response strategies. Synthetic Aperture Radar (SAR) imagery has emerged as a reliable remote sensing tool for oil spill detection due to its ability to capture ocean surface information under all weather and illumination conditions. However, accurately distinguishing oil spills from look-alike phenomena in SAR images remains challenging due to speckle noise, complex ocean dynamics, and limited annotated datasets. This paper presents NeuroSpill, a deep learning based framework for automated oil-spill detection using SAR imagery. The proposed method employs an Attention-UNet++ architecture to perform precise semantic segmentation of oil spill regions. To enhance feature representation, three SAR- derived channels—VV polarisation, VH polarisation, and the VV/VH ratio—are utilised as input features. A patch-wise prediction strategy is implemented to process large SAR scenes while improving segmentation accuracy efficiently. In addition, a verification mechanism and morphological post-processing are applied to reduce noise and refine spill boundaries. The framework also incorporates an oil spill analysis module that estimates spill coverage and categorises severity levels. An interactive Streamlit-based interface is developed to enable real-time visualisation and monitoring of SAR images. Experimental results demonstrate that the proposed system provides reliable oil spill detection and offers a practical solution for automated marine environmental monitoring.

Keywords: Attention U-Net++, Deep Learning, Environmental Monitoring, Oil Spill Detection, Remote Sensing, Synthetic Aperture Radar, SAR Semantic Segmentation