An Integrated SAR Image Segmentation and IoT-Based Monitoring Framework for Marine Oil Spill Detection and Severity Classification

Accurate segmentation of oil spill areas in Synthetic Aperture Radar (SAR) images is achieved by distinguishing between regions contaminated by oil and similar areas of the ocean surface. This research proposes an AI-IoT approach to segment and assess the severity of oil spillage from Sentinel-1 SAR images. The proposed methodology involves logarithmic transformation, normalization, and patch generation prior to training a U-Net convolutional neural network with ground truth segmentation masks. Performance of the network is measured on a separate test set using the Dice coefficient and Intersection over Union (IoU). An obtained Dice coefficient of 0.718 and IoU of 0.593 measure the spatial correspondence of predicted and reference areas of spills, while inference time is about two seconds per tile. Mask predictions are converted into area estimates and classified into three levels – minor, moderate, and severe levels. An IoT communication module transmits data related to spill location, severity, timestamp, and confidence factor to a cloud server for geospatial visualization and alerts. The system uses pixel-based SAR data analysis with an organized environmental monitoring and coastal response process.