UXO-YOLO: A Lightweight Attention-Enhanced YOLOv8 Framework for Real-Time UXO Detection in UAV Imagery

Unexploded ordnance (UXO) poses a persistent threat to safety in post-conflict areas, where manual clearance is slow, dangerous, and financially expensive. UAV-based real-time detection provides a more secure and scalable approach. However, UXO typically appear as small, low-contrast targets embedded in complex backgrounds, which significantly degrade the performance of conventional YOLO-based detectors. To address these challenges, this paper proposes UXO-YOLO, a lightweight, attention-enhanced YOLOv8s framework specifically designed for UXO detection in UAV imagery. A Spatial-Efficient Channel Attention (SECA) module is introduced in the neck to enhance fine-grained feature extraction and improve UXO-to-background discrimination while maintaining real-time efficiency. In addition, a C2f Coordinate Attention (C2f CA) module is integrated into the backbone to enable direction-aware feature encoding and improve localization robustness for small UXO targets. Furthermore, the large-scale P5 detection head is removed to eliminate unnecessary computation and focus the network on small-scale target detection. Owing to the sensitive and specific nature of UXO in military contexts, publicly available UXO-specific datasets are extremely limited. To address this limitation, a custom Unexploded Aerial Bomb (UAB) dataset was constructed using UAV-captured RGB images under controlled testing conditions. The proposed model is evaluated on the UAB dataset and further assessed on the public UAVDT benchmark to examine cross-domain robustness. Experimental results show that UXO-YOLO achieves 95.8% mAP50, 68.2% mAP50:95, and 91.9% recall on the UAB dataset, and 86.7% mAP50, 47.9% mAP50:95, and 81.8% recall on the UAVDT dataset, while preserving real-time inference efficiency. These results indicate the suitability of UXO-YOLO for UAV-based humanitarian UXO detection applications. The reported real-time performance is evaluated on a GPU-equipped laptop platform, and on-board validation on embedded UAV hardware remains part of future work.