A Rapid Inundation Simulation Method for Urban Waterlogging in Plain Polder Areas Based on Machine Learning

Frequent short-duration and high-intensity rainstorms during the monsoon season have caused recurrent pluvial flooding and severe localized ponding in the central urban district of Changshu City, Jiangsu Province. For emergency response, purely physics-based hydrodynamic models are often too slow to provide actionable forecasts. This study aims to develop a rapid simulation approach that balances accuracy and computational efficiency by integrating physics-based modeling with deep learning, thereby enabling timely warnings, situational awareness, and decision support for urban flood risk management in a typical plain polder city. A 1D pipe–2D overland coupled hydrodynamic model was built in InfoWorks ICM for the representative polderized core of Changshu, explicitly representing stormwater pipes, inlets, pumps, gates, and surface flow pathways over high-resolution terrain. Twelve design rainstorms with different intensities and temporal patterns were simulated, and the resulting high-fidelity inundation depths were extracted as training and validation samples. Two complementary surrogates were designed: (i) a long short-term memory (LSTM) network to forecast depth time series at critical nodes for operations, and (ii) a convolutional neural network (CNN) to regress full-domain flood-depth rasters for spatial situational awareness. Preprocessing ensured mesh–raster alignment and standardized rainfall and terrain inputs. Model skill was evaluated with metrics appropriate for temporal and spatial tasks, including coefficient of determination (<italic>R</italic>²), Nash–Sutcliffe efficiency (NSE), mean absolute error (MAE), root-mean-square error (RMSE), peak-depth error, and map-based indicators such as intersection-over-union (IoU), precision, recall, and F1 score. Cross-scenario tests assessed generalization across rainfall intensities and hyetographs. The framework achieved high predictive accuracy while drastically reducing computation time. The LSTM reproduced hydrograph evolution at most monitored locations with <italic>R</italic>² &gt; 0.90, capturing both rising and receding limbs. The CNN accurately reconstructed spatial inundation patterns governed by micro-topography and road embankments; peak-depth errors were controlled within ±5 cm, and the IoU for inundation extent reached 0.87. Compared with the full InfoWorks ICM solver, the surrogate-based workflow delivered ≥100× speed-up, enabling second- to minute-level inference on standard hardware. Error analysis shows residuals concentrate near sharp elevation transitions and curbed road segments where mesh–raster misalignment is most influential. Including near-dry states and first-flush periods in training reduces false positives and improves depiction of wetting–drying fronts. A demonstration case indicates that the method provides actionable depth maps and node-level forecasts early in the event window, supporting emergency routing, temporary storage scheduling, and pump dispatch. Coupling deep learning surrogates with a calibrated 1D/2D hydrodynamic model substantially enhances the timeliness of urban flood prediction without sacrificing accuracy, meeting operational demands for rapid warning and decision support in polderized plain cities. The workflow is reproducible and transferable to similar low-relief urban regions, provided a baseline hydrodynamic model is available for scenario generation and that training covers multiple storm intensities and early wetting phases. Future work will focus on probabilistic outputs via ensemble surrogates, tighter data assimilation with real-time sensors, and streamlined deployment within municipal emergency platforms.