A cross-city transferable convolutional neural network framework for assessing street-scale flood risks in urban...

Urban flooding is a major risk to infrastructure and public safety, particularly in densely populated cities facing rapid urbanization and climate change. This study introduces an AI-driven convolutional neural network (CNN)-based framework for assessing street-scale flood risks by integrating hydrometeorological, topographic, and urban morphological data. The CNN model, trained on Shenzhen data and applied to Hong Kong, demonstrates its strong spatial transferability and the potential of deep learning in urban flood risk assessment. In the application, the model predicts flood inundation under four design rainfall scenarios. Results show that under the extreme 100-year recurrence interval, 60-min rainfall scenario, the flood-prone area in Hong Kong expanded to 64.1 km² (5.79 % of the urban area), with average inundation depths reaching 15.6 cm and a maximum depth of 68.6 cm. A road-level flood risk assessment identified 501 high-risk road segments (18.9 % of the total network), concentrated in southern Kowloon Peninsula and northern Hong Kong Island. This framework emphasizes the need for tailored flood mitigation strategies for different urban areas and highlights the model's cross-regional transferability as an innovation for flood risk assessment in cities with similar conditions.