A Deep Learning-Based Framework for Buried Hazard Detection Using UAV-Mounted Ground Penetrating Radar

Ground penetrating radar (GPR) is a non-destructive electromagnetic sensor for imaging the subsurface. Its integration with unmanned aerial vehicles (UAVs) offers a safe and efficient approach for detecting buried hazards, such as unexploded ordnance, landmines, and vulnerable underground infrastructure. However, interpreting complex radargrams remains a major bottleneck due to environmental clutter, signal attenuation, and motion-induced noise. In this context, this thesis proposes an automated deep learning framework for detecting hyperbolic hazard signatures in UAV-mounted GPR data, addressing the intrinsic challenges of domain shift and ground-truth ambiguity. To overcome the lack of large and specialized datasets, a custom multi-modal labeling tool was developed to annotate more than 150,000 images sourced from three separate field experiments, using a confidence-based labeling system to explicitly quantify uncertainty. Additionally, to ensure a realistic deployment scenario, a cross-domain evaluation protocol was employed to assess three different object detection architectures: YOLO11n, RT-DETR, and Faster R-CNN. The analysis evaluates both their theoretical robustness and practical deployment capabilities through a comprehensive set of performance metrics, including a detailed site-specific comparison. The experimental pipeline includes B-scan filtering for input optimization, GPR-specific pre-training and offline test-time adaptation (TTA) via pseudo-labeling to improve network generalization capabilities. Results highlight the crucial role of radargram pre-processing, identifying the removal of the first three singular values as the optimal background subtraction strategy among the evaluated techniques. Other pipeline configurations are highly model- and site-dependent. Faster R-CNN, initialized with specialized pre-training, successfully detected 83% of the total number of buried threats while maintaining a relatively low false positives per image of approximately 0.2, proving to be the most robust model setup across all test sites. Furthermore, TTA was demonstrated to be an effective adaptation strategy, improving baseline performance across all tested architectures. Despite challenges such as limited data, label uncertainty, and environmental noise, this thesis highlights the complexities of airborne radar data interpretation and proposes a promising deep learning approach for automated detection of underground hazards.