Abstract
Marine oil spills pose a severe and persistent threat to ecosystems and coastal economies. Traditional manual or satellite detection is slow, laborious, and error-prone due to sensor limitations, noise, weather interference, small target sizes, and imbalanced datasets. To address these challenges, this paper proposes a novel, integrated framework for the rapid detection and monitoring of oil spills by using the Internet of Things, unmanned vehicles, and transfer learning. The proposed system uses a multi-layered architecture: a physical layer of visual, infrared, and acoustic sensors deployed on a network of Saildrone unmanned surface vehicles for real-time data acquisition; an edge layer for initial processing and low-latency response; and a cloud layer that uses deep transfer learning for accurate spill identification and classification. We fine-tuned pre-trained ResNet models using a Synthetic Aperture Radar oil spill dataset, achieving a peak accuracy of 97.89% with a three-layer transfer learning configuration, outperforming other tested configurations. The efficiency of the system in real-time data handling and leak localization was validated through a controlled experimental prototype. The results demonstrated a robust solution for minimizing response time and environmental impact. Our framework has been proven to gain about 98.3% model accuracy on drone images for oil spill detection.