Azlina Kamaruddin, Norma Alias, Yahya Nur Ifriza
Oil spill detection faces critical challenges in environmental monitoring and maritime safety, where timely and accurate identification of spills is essential for ecological and economic impacts. This study presents a machine learning framework for oil spill detection and classification using Sentinel-1 Synthetic Aperture Radar (SAR) imagery. A dataset of 1,842 annotated images was pre-processed and used to train an enhanced Convolutional Neural Network (E-CNN) with data augmentation and dropout. E-CNN achieved 96.2% accuracy, 94.8 % F1-score, and 0.90 mean Intersection over Union, outperforming multiple baseline models, including standard CNN, ResNet-50, DAM-UNet, DAENet, YOLOv8, and CNN-Transformer hybrids. Receiver Operating Characteristic analysis confirmed robust discrimination capability with AUC values exceeding 0.95. The results demonstrated that attentionbased feature enhancement can deliver high accuracy with reduced computational complexity. The E-CNN provides an efficient and scalable framework suitable for near-real-time maritime oil spill monitoring and environmental surveillance. © 2025 IEEE.
Universiti Teknologi Petronas, Department of Computing, Seri Iskandar, Malaysia; Ibnu Sina Institute for Scientific and Industrial Research, Universiti Teknologi Malaysia, Skudai, Malaysia; Information Systems, Universitas Negeri Semarang, Semarang, Indonesia
