Chained Nowcasting of High Spatiotemporal Resolution Urban Rainfall and Inundation Risk Using Deep Learning

A critical challenge in urban hydrometeorology is achieving accurate nowcasting of storms and inundation risks, hindered by the high heterogeneity of both urban surfaces and precipitation fields, which jointly induce nonlinear and abrupt spatiotemporal patterns of inundation risk. To address this, we develop a cascaded deep learning framework that performs end-to-end, high-resolution forecasting from radar extrapolation to inundation risk. First, a ConvLSTM-UNet model is trained on a decade-long (2015–2025), highspatiotemporal-resolution (1 km, 6 min) radar echo mosaic dataset over Wuhan, China, to generate skillful short-term quantitative precipitation nowcasts, thereby capturing fine-scale rainfall heterogeneity. Second, using a large set of historical waterlogging points collected from online platforms as labeled data, another deep learning model is trained to learn the complex coupling between nowcasted rainfall and high-resolution urban features with 12.5 m DEM, 30 m Local Climate Zone maps, and fine road networks, thereby quantifying how surface heterogeneity modulates runoff accumulation and flood susceptibility. By chaining these two stages, the framework produces high spatiotemporal resolution, probabilistic inundation risk nowcasts directly from radar observations. This data-driven approach offers an effective and novel tool for real-time early warning and refined risk management in complex urban environments.