A GeoAI-Based Physics-Enhanced Framework for Robust Short-Term Urban Waterlogging Prediction

Accurate short-term prediction of urban waterlogging depth is essential for real-time flood risk management in rapidly urbanizing areas under climate variability. Departures from quasi-stationary operating conditions, caused by changes in drainage efficiency, inflow patterns, or measurement quality, weaken historical rainfall–water depth relationships, making purely data-driven models prone to error accumulation. In this study, a GeoAI-based, physics-enhanced machine learning framework is proposed, which translates the water balance principle into Physical Violation Scores (PVSs) and incorporates them as additional input features. PVSs remain zero under expected rainfall–water depth behavior and become positive only under departure scenarios, providing sparse and lightweight diagnostic signals without modifying model structures or loss functions. The framework is implemented on five algorithms (Support Vector Machine, Multilayer Perceptron, Random Forest, Extremely Randomized Trees, and XGBoost) to construct physics-enhanced models (PEMs). These are evaluated against original feature models (OFMs) across 1 h and 2 h forecasting horizons. Results show that most PEMs improve prediction performance compared with their corresponding OFMs, with more pronounced gains at the 2 h horizon. Bootstrap analysis and RMSE-based error amplification factor further indicate comparable or lower R2 variability and reduced recursive error amplification for most PEMs. Interpretability analyses show that rainfall forcing and water-depth persistence remain dominant predictors, whereas PVSs act as auxiliary diagnostic signals. Overall, the proposed framework provides a lightweight, reliable, interpretable, and scalable GeoAI approach for incorporating water balance knowledge into short-term urban waterlogging prediction, supporting climate resilience and smart urban water management.