Real-Time Urban Flood Detection Using YOLOv5-Seg
Jaegeun Jang, Youngwoo Kwon · 2025
Focusing on urban flooding and the recent increase in associated risks, this paper proposes a real-time urban flood detection framework utilizing a lightweight instance segmentation model based on YOLOv5n-seg. The approach leverages existing CCTV infrastructure, eliminating the need for additional sensors. The YOLOv5n-seg model was trained using an open dataset provided by AI-Hub, a large-scale disaster image collection, enabling the simultaneous extraction of object locations and segmentation masks for accurate identification of inundated areas. Experimental results demonstrate that the model achieves a box$\text{mAP} {@} 0.5$of 0.72, a mask mAP of 0.64, and an average real-time processing speed of approximately 139 FPS. With only 2.0M parameters and 7.1 GFLOPs, the model's lightweight architecture is well-suited for resource-constrained environments such as edge devices. However, as the experiments were conducted solely on a high-performance GPU (NVIDIA RTX 4070 Ti), further evaluation is needed to verify real-time performance on low-end devices. Additional limitations include occasional false positives and negatives when detecting small objects or in complex backgrounds and the absence of quantitative comparisons with the latest YOLO models. Future research will focus on benchmarking against recent YOLO models and evaluating real-time performance on low-end devices to further enhance the proposed framework's robustness and practical applicability.