UAV-Based Multispectral Imaging for Surface Oil and Water Leak Detection over Soil Using Deep Learning
This project presents a UAV-based multispectral imaging framework for detecting and distinguishing surface oil and water leaks over soil using deep learning. The work addresses the challenge of differentiating oil and water leaks, which can exhibit similar visual characteristics under conventional RGB imaging but require different mitigation responses.
Multispectral data were acquired using a MicaSense RedEdge-P camera mounted on a quadcopter. The system captured Blue, Green, Red, Red-Edge, and NIR bands along with a high-resolution Panchromatic band. The imagery was processed through radiometric correction, geometric band co-registration, and pan-sharpening before being used for spectral analysis and detection.
I performed band-wise spectral analysis to identify the most informative bands for distinguishing oil and water over soil. The Red-Edge band showed the strongest individual-band separability, with a Jeffries-Matusita distance of 0.6545, followed by the Red band. The combined Red and Red-Edge bands provided improved multivariate discrimination.
A dataset of 512 Red and Red-Edge images was constructed and annotated with two classes: oil and water. The images were divided using an 80:20 training-validation split. Multiple object detection architectures were evaluated, including YOLOv11, RT-DETR, Faster R-CNN, and a fusion-based YOLOv11-RGBT model.
YOLOv11 and RT-DETR achieved the strongest overall detection performance. YOLOv11 achieved an overall mAP@50 of 0.989 and mAP@50:95 of 0.806, while RT-DETR achieved 0.992 and 0.798, respectively. YOLOv11 also achieved an inference time of approximately 0.016 seconds per image, demonstrating its potential for near-real-time deployment.
The system successfully detected oil leak volumes as low as 5 ml and maintained detection performance under partial occlusion and vegetation-induced shadowing. These results demonstrate the potential of UAV multispectral sensing combined with deep learning for localized and scalable surface leakage monitoring.
