Abstract
Extreme rainfall and flooding severely impact urban systems by disrupting access to critical services, interrupting mobility, and posing challenges for emergency management. Accurate road network flood prediction remains challenging due to complex flow dynamics, coarse-resolution traditional models, and limited data. The main objective of this study is to enhance road-network flood prediction using ensemble machine learning models trained on crowd-sourced flood datasets. Our results for the Washington, D.C. area show that stacked super-ensemble learning improves road flood prediction compared to the voting algorithm and several other base learners, including random forest, support vector machine, bagging, and boosting. Stacking algorithm achieved an accuracy of 0.84, precision of 0.82, and F1-score of 0.82. Shapley additive explanations indicate that elevation strongly influences model prediction accuracy. Stacking ensemble classifies around 5% of road networks as having very high likelihood and 11% as having high likelihood of flooding. We find that over 40% of energy and emergency services are located within high hazard networks. The insights gained from this study can help improve urban flood prediction which is crucial for enhancing community resilience to extreme weather events.