Abstract Urban pluvial floods have become increasingly frequent under the combined effects of climate change and urbanization, leading to increased disaster losses. This study developed a data-driven urban pluvial flood prediction model using convolutional neural networks (CNNs) to enhance computational efficiency while maintaining simulation accuracy. A high-resolution cellular-based flood model generated the training dataset through systematic patch-based sampling combined with fixed step size selection strategies. The established framework enabled flood simulations through integrated analysis of topographic features and rainfall processes. Shapley additive explanations (SHAP) and Group masking analysis (GMA) were implemented to interpret the decision-making mechanisms of CNN model. The model was validated in a relatively independent drainage area, demonstrating strong agreement with conventional cellular model outputs across six design storm scenarios and two historical rainfall events. Computational experiments showed that the CNN model reduced simulation time from minutes to seconds compared to process-based approaches, while maintaining low absolute errors in water depth predictions. Both SHAP and GMA interpretation revealed that topographic features, particularly building, digital elevation model (DEM), and aspect, exert dominant influence on model predictions. This data-driven framework established an efficient computational paradigm for urban flood modeling, with SHAP and GMA analysis guiding input variable selection while explaining model behavior. The methodology demonstrated potential for real-time monitoring integration, supporting rapid flood risk assessment and resilience enhancement.
Chen et al. (Wed,) studied this question.