Designing an inundation monitoring and real-time urban flood forecasting system: a synthetic study

• A unified framework integrates flood modeling, ML, and sparse sensors for urban flood forecasting. • Stochastic rainfall model generates diverse precipitation scenarios for design. • Novel method optimizes sensor placement for accurate inundation map reconstruction. • ML model combines rainfall & observations, outperforming traditional models by 20–50% Urban flooding due to extreme precipitation poses an increasing threat to cities, necessitating accurate, timely inundation predictions. This study presents a novel machine learning-based framework for urban inundation forecasting, addressing the challenges in computational efficiency and predictive accuracy. The approach integrates an ensemble of space–time varying rainfall scenarios, a high-fidelity flood model, and a machine learning surrogate model. Stochastically generated synthetic rainfall scenarios include extreme precipitation with various return periods. They are inputs to an urban flood model tRIBS-Urban that produces an ensemble of synthetic inundation fields that are synthetic “observations”. We combine Principal Component Analysis and Karhunen-Loève Expansion to optimize flood sensor placement and accurately reconstruct inundation maps from “observations” at a few locations. Application of a spatio-temporal vision transformer (ViT) as a surrogate model for tRIBS-Urban demonstrates excellent performance in capturing spatiotemporal patterns of inundation depth, with real-time forecasts produced in a few seconds for 1–12 h lead times, showing high accuracy (with Root Mean Square Error approximately 0.15 m and Kling-Gupta Efficiency greater than 0.75) and minimal cumulative error. By incorporating both rainfall and inundation observations, the inundation forecasts improved by 20–50% compared to those generated by the model without considering inundation data. This framework therefore shows significant potential to enhance flood prediction capabilities, offering a promising solution for improving urban flood resilience. Key points: ● A novel machine learning framework integrates flood physical modeling and sparse sensor data for forecasting urban flood inundation. ● A stochastic rainfall model offers diverse spatiotemporal precipitation scenarios, addressing limitations in infrastructure design standards. ● A novel method optimizes flood sensor placement to accurately reconstruct inundation maps from (synthetic) sparse observations. ● A machine learning model combines rainfall and observations to ensure high accuracy in flood prediction, outperforming traditional models by 20–50%.