Urban Morphology in Urban Flood Risk Prediction: A Deep Learning Framework for Resilient Planning

Existing flood risk models have improved predictive accuracy, but they prioritize natural and hydrological factors while giving limited attention to fine-grained urban morphology. This study develops an interpretable deep learning framework to examine how high-resolution, three-dimensional urban form shapes two dimensions of flood risk: inundation risk, measured by grid-level inundated area, and infrastructure risk, measured by flood-related disruptions, including water supply interruption, power outage, road blockage, and collapse-related damage. Using Zhengzhou, China, as a case study, we combine multi-source spatial data, convolutional neural networks, ablation analysis, SHAP interpretation, and Gaussian Mixture Model classification to examine how fine-grained urban morphology affects these two risk dimensions. Incorporating urban morphology improved inundation risk prediction, reducing MSE from 0.0431 to 0.0371. The improvement was greater for infrastructure risk, with accuracy increasing from 0.7327 to 0.8218, and ROC-AUC from 0.83 to 0.95. SHAP results show that inundation risk is associated with vegetation, elevation, hydrological proximity, and localized spatial disorder, whereas infrastructure risk is amplified by vertical intensity, imperviousness, building concentration, porosity, and shape. Spatially, very high infrastructure-risk areas accounted for only 2.30% of the city but 12.88% of the central districts, while 74.62% of very high infrastructure-risk zones were concentrated in dense mid- to high-rise morphology. These findings suggest that flood-resilient planning should move beyond hydrology-sensitive flood management toward morphology-sensitive planning.