Comparative Evaluation of Deep Learning Models for Urban Flood Forecasting

A rapid and accurate urban inundation prediction system is important for reducing the impacts of urban flooding. This study comparatively evaluates three deep learning models—a Transformer, CNN-LSTM, and GNN—as surrogate models that link rainfall time series with the spatial distribution of inundation. A dataset of 1000 rainfall–inundation pairs was constructed using the physics-based EPA-SWMM/FLO-2D framework under a range of synthetic and observed rainfall events with durations of 1 h, 2 h, and 3 h. The three deep learning models were trained to approximate the inundation responses generated by the physics-based simulations, and their predictive performance was evaluated using RMSE, NSE, IoU, and F1 score, together with computational time, threshold sensitivity, and robustness to different train–test partitions. For the 30 common evaluation events, the Transformer achieved an RMSE of 0.034 ± 0.011 m, NSE of 0.908 ± 0.018, IoU of 0.903 ± 0.020, and F1 score of 0.949 ± 0.016, showing the highest overall predictive performance among the three models under the conditions considered in this study. The training times of the three models were relatively similar, ranging from 3268.1 to 3407.5 s, while inference times ranged from 0.031 to 0.044 s, substantially shorter than the mean EPA-SWMM/FLO-2D simulation time of 293.1 s. Threshold-sensitivity analysis and repeated random train–test splits showed a similar relative performance pattern among the models, while 50-run Monte Carlo cross-validation of the Transformer indicated relatively consistent performance across repeated data partitions. These results demonstrate the potential of deep learning surrogate models for rapid urban inundation prediction and provide a comparative assessment of the suitability of different architectures for approximating physics-based inundation responses.