Evaluating the Performance of Automatic Detection for Urban Flood Levels Using Different Deep Learning Approaches

Abstract Urban flooding states have been intelligently detected in numerous studies via deep learning algorithms to identify objects associated with floods. Most automatic detections of urban flood depth reported in the literature have been conducted by constructing relationships between the inundation status of dynamic objects and the urban flood level with intelligent algorithms that possess the ability to process images quickly and accurately. However, knowledge gaps remain regarding the performance of detection models across different novel and advanced deep learning models that require different computational powers. Therefore, this study aimed to evaluate the performance of several state-of-the-art deep learning models, in detecting urban flood levels based on a dataset of flooded vehicles. Comprehensive experiments were conducted to compare these models in terms of performance metrics such as precision, recall, mAP50, and inference time were analyzed to determine the effectiveness of each approach. The results demonstrated that the YOLOv10 series models outperformed traditional approaches, making them the optimal choice for real-time urban flood risk detection. Among these, YOLOv10n strikes a balance between accuracy and low computational demands, whereas YOLOv10x offers the highest performance, making it particularly suitable for fixed urban flood monitoring facilities. Other YOLOv10 variants can be selected as needed based on the trade-off between accuracy and computational resources. Meanwhile, Faster R-CNN achieved higher recall at the cost of increased false positives, allowing it to detect more submerged vehicles than YOLOv10. These findings offer valuable insights that can guide the selection of suitable models for various practical scenarios in urban flood management.