A Cross-City Transferable Convolutional Neural Network Framework for Street-Scale Urban Flood Risk Assessment

Urban flooding has intensified under rapid urbanization and climate change, posing a growing threat to infrastructure and public safety. Conventional flood risk assessment approaches that are based on fluid dynamics simulations and physics-driven models are computationally demanding and not well-suited real-time applications, particularly in highly dynamic urban settings. This study develops a convolutional neural network (CNN) framework that integrates historical flood inventories, hydrometeorological data, topographic information, and urban morphological characteristics for faster and more accurate prediction of flood extent and depth. The CNN model is trained on a dataset from Shenzhen, China and then applied to Hong Kong, China, demonstrating robust spatial transferability for cross-city flood risk assessments. The framework is used to simulate flood inundation for four design rainfall scenarios . The results indicate that short-duration, high-intensity rainfall events significantly increase the extent of flooding and the depth of inundation. Flood-prone areas in Hong Kong expand to 64.1 km² during a 100-year rainfall event with a 60-minute duration, accounting for 5.79% of the urban area, and the mean inundation depth reaches 0.15 m. In addition, a complementary road-level flood vulnerability analysis identifies 501 flood-prone roads, primarily located in districts with aging infrastructure and high population density. This study highlights the potential of CNNs for rapid flood prediction with strong cross-city transferability, and provide decision-makers with timely insights for targeted flood prevention and disaster mitigation strategies.