A vision-enhanced framework for urban flood modelling and monitoring under data scarcity

Urban flood risk is increasingly exacerbated by climate change, while traditional models face accuracy issues from low-quality DEM and land use data and inefficient monitoring. This study proposes a computer vision-based collaborative framework for urban flood simulation and monitoring, which is demonstrated in a typical study area in Zhongshan City, China. In the one-dimensional drainage network modelling stage, the use of road buffer zoning, building-footprint elevation refinement, and combined slope-width-roughness calibration resulted in Nash-Sutcliffe efficiency of 0.802-0.917 for three observed heavy-rainfall events, with peak timing error ≤ 30 min and peak water-level error ≤ 0.12 m. By constructing a U-Net network for single-class segmentation of building footprints from satellite imagery and updating land use, the NSE values increased to 0.884-0.922, with peak-timing errors and peak value errors reduced to 15 min and 0.1 m, respectively, thereby improving model performance. A coupled SWMM-ITF-FLOOD model was developed to simulate inundation extent and depth. The improved CBAM-YOLOv8 was then used to automatically identify the inundation extent, with an area error below 8%. A novel visual water-level gauge was also designed, with inundation depth recognition error of 2 cm using U-Net binarization for reading. The results demonstrate that computer vision can supplement incomplete surface and monitoring data and improve the representation of model inputs and enable high-precision real-time monitoring, providing viable technical support for smart water management and urban flood control.