A Systematic Review of Urban Flooding Prediction Models based on deep learning: Current Trends and Future Directions

In recent years, deep learning methods have rapidly gained prominence in urban pluvial flooding prediction due to their ability to perform automatic feature extraction and end-to-end mapping. This systematic review, drawing on 30 representative studies, classifies the literature into five application domains: sequential forecasting models, gridded spatial prediction models, spatiotemporal integration models, coupling with physical simulators, and image-based detection and semantic segmentation. Each category exhibits distinct data requirements, network architectures, and deployment scenarios, and collectively they offer advantages in high-calculation efficiency inference, physical consistency, and automated feature learning. Nonetheless, challenges such as data scarcity, uncertainty propagation, insufficient physical constraints, and difficulties in real-time implementation continue to hinder widespread adoption. Building on this critical analysis, we identify future research directions including deeper physics–data integration, enhanced model interpretability and generalization, incorporation of uncertainty quantification and network lightweighting, and the development of cross-regional multi-source transfer learning frameworks, all aimed at advancing urban flooding prediction technologies and facilitating their practical deployment.