GAL-MMF: A GAN-LSTM-Based Multimodal Framework for Dynamic Urban Waterlogging Risk Prediction

Abstract

In view of the challenges in urban waterlogging risk prediction, such as difficulty in multi-source heterogeneous data fusion, insufficient capture of spatiotemporal dynamics, and low accuracy of extreme event prediction, this study proposes a GAN-LSTM multimodal dynamic prediction model (GAL-MMF). The generator adopts a ConvLSTM architecture and the discriminator a spatiotemporal CNN. LSTM effectively learns long-term dependencies of multimodal data including meteorology, hydrology, topography, pipe network operation, and real-time monitoring. The GAN framework enhances the model’s ability to generate complex spatiotemporal patterns, especially under rare rainstorm events, and improves robustness through adversarial training, reducing bias caused by sample scarcity. The UW-RiskBench dataset (7,677 samples: 5,584 training, 2,093 testing) is used for validation. Results show that GAL-MMF reduces RMSE of water depth prediction by 18.2%, increases F1-score for high-risk area identification by 15.7%, and improves recall of extreme events by more than 25%. Compared with SWMM, single LSTM, ConvLSTM, and MMF-STGCN, GAL-MMF achieves higher accuracy, better extreme event detection, and faster response (15-minute resolution), providing strong support for refined waterlogging prevention and emergency management.