UFD-YOLO: a lightweight multi-reference model for urban flood risk level detection from surveillance images

Flood disasters pose persistent challenges to urban safety and emergency management. Traditional remote sensing technologies remain limited in real-time capability and spatial coverage. In surveillance-based urban flood monitoring, direct physical measurement of flood depth is often unavailable; therefore, flood risk level detection based on visual reference objects provides a feasible alternative for rapid flood assessment. However, existing lightweight YOLO models still struggle to accurately identify submerged targets under partial submergence, large-scale variation, and complex urban conditions. To address these issues, this study proposes UFD-YOLO, a lightweight framework for urban flood risk level detection using pedestrians and vehicles as multi-scale visual references. First, a reference-based flood risk level criterion was established according to the submergence states of pedestrians and vehicles, and a dataset covering diverse urban scenarios was constructed. Second, the DySnakeConv, AIFI, and DetectAux were progressively integrated into YOLO11n to improve local structural perception, strengthen feature interaction, and enhance multi-scale supervisory learning. Compared with YOLO11n, UFD-YOLO improves Precision, Recall, mAP50, and mAP50:95 by 2.5%, 5.1%, 5.8%, and 2.6%, respectively, while maintaining low computational cost. Additional evaluations further demonstrate its accurate and stable performance in challenging scenarios. Overall, this study provides an effective technical approach for urban flood risk assessment.