A Convolutional Neural Network-Weighted Cellular Automaton Model for the Fast Prediction of Urban Pluvial Flooding Processes

Abstract

Deep learning models demonstrate impressive performance in rapidly predicting urban floods, but there are still limitations in enhancing physical connectivity and interpretability. This study proposed an innovative modeling approach that integrates convolutional neural networks with weighted cellular automaton (CNN-WCA) to achieve the precise and rapid prediction of urban pluvial flooding processes and enhance the physical connectivity and reliability of modeling results. The study began by generating a rainfall-inundation dataset using WCA and LISFLOOD-FP, and the CNN-WCA model was trained using outputs from LISFLOOD-FP and WCA. Subsequently, the pre-trained model was applied to simulate the flood caused by the 20 July 2021 rainstorm in Zhengzhou City. The predicted inundation spatial distribution and depth by CNN-WCA closely aligned with those of LISFLOOD-FP, with the mean absolute error concentrated within 5 mm, and the prediction time of CNN-WCA was only 0.8% that of LISFLOOD-FP. The CNN-WCA model displays a strong capacity for accurately predicting changes in inundation depths within the study area and at susceptible points for urban flooding, with the Nash-Sutcliffe efficiency values of most flood-prone points exceeding 0.97. Furthermore, the physical connectivity of the inundation distribution predicted by CNN-WCA is better than that of the distribution obtained with a CNN. The CNN-WCA model with additional physical constraints exhibits a reduction of around 34% in instances of physical discontinuity compared to CNN. Our results prove that the CNN model with multiple physical constraints has significant potential to rapidly and accurately simulate urban flooding processes and improve the reliability of prediction.