Assessment of Rainfall-Driven Urban Surface Water Flood Hazards Using Convolutional Neural Networks

Abstract

Rainfall-driven urban surface water flooding is one of the most common natural disasters that lead to traffic disruption, economic loss, and even casualties. Assessing its hazards is critical not only for flood management but also for urban and territorial planning. Physics-based models can simulate hydrological and hydraulic processes to predict floods; however, they are computationally expensive for large-scale and high-resolution simulations. This study presents a U-Net-based deep learning method for assessing the hazard levels of urban surface water flooding. The approach adopts three methods to improve the baseline U-net model: (1) Squeeze-and-Excitation Blocks that enhance feature representation; (2) Focal Loss, a loss function that mitigates the influence of data imbalance; and (3) Random Cutout, a data augmentation method that prevents overfitting. Catchment data are used as input to train the deep learning model against flood hazard targets under three different levels of annual exceedance events. The results showed that the models are capable of identifying mid and high hazards. The proposed three methods mutually constrained each other and can reduce the influence of data imbalance. The proposed model demonstrates potential for practical flood management through rapid and accurate identification of high-risk areas.