Gu et al. (2025) FloodTransformer: Efficient real-time high-resolution flood forecasting

Identification

• Journal: Environmental Modelling & Software
• Year: 2025
• Date: 2025-12-11
• Authors: Zhanzhong Gu, Jeonghyun Kang, Wei-Guo Jin, Feifei Tong, Y. Jay Guo, Wenjing Jia
• DOI: 10.1016/j.envsoft.2025.106832

Research Groups

• School of Electrical and Data Engineering, University of Technology Sydney, NSW, 2007, Australia
• School of Architecture and Civil Engineering, The University of Adelaide, South Australia, 5005, Australia

Short Summary

This paper introduces FloodTransformer, a hybrid AI-hydrodynamic framework for accurate, real-time, and high-resolution flood forecasting, demonstrating superior performance in water depth and inundation classification compared to existing models.

Objective

• To develop an AI-hydrodynamic hybrid framework, FloodTransformer, that integrates fine-scale hydrodynamic simulations with advanced deep learning techniques to achieve rapid, high-resolution, and high-accuracy real-time flood predictions.

Study Configuration

• Spatial Scale: Wagga Wagga catchment, New South Wales, Australia, discretized into 346,000 variable-size computational cells with resolutions ranging from 5 meters × 5 meters (river channel) to 160 meters × 160 meters (areas unlikely to inundate).
• Temporal Scale: Real-time, non-autoregressive sequential predictions over a 24-hour forecast horizon (48 time steps at 30-minute intervals) in a single run.

Methodology and Data

• Models used:
  • Hydrodynamic model: 3Di hydrodynamic flood model.
  • AI model: FloodTransformer (Vision Transformer-styled architecture with variable-size cell embedding, tokenized time-sequence encoding, and physics-informed multi-task optimization).
  • Comparative models: CNN-RNN, SwinTransformer, CNN-base.
• Data sources:
  • Training data: 150 hydrodynamic simulations generated by the calibrated 3Di model under diverse initial water levels, boundary conditions, rainfall durations (30 minutes to 4 days), and Annual Exceedance Probabilities (AEPs)/Average Recurrence Intervals (ARIs).
  • Static geographical features: Geoscience Australia 5 Metre Digital Elevation Model (DEM) (LiDAR, 2001-2015), Australian Bureau of Agricultural and Resource Economics and Sciences (ABARES) land use data (50 meters × 50 meters), NSW Great Soil Group (GSG) Soil Type map, NSW Open Data Portal road data, and NSW SES levee study data.
  • Time-series data: Initial water depth, initial inundation state, and rainfall forecasts derived from 3Di simulations (NetCDF format).
  • Calibration and validation data: Observed rating curves from the Australian Bureau of Meteorology (BoM) and historical flood inundation maps (e.g., 2018 flood event, WMAwater, 2018).
  • Historical flood events: Real historical rainfall and flood water level records for Wagga Wagga (2022-10-24, 2022-10-08, 2022-11-01).

Main Results

• FloodTransformer achieved superior accuracy and efficiency compared to baseline models.
• Water Depth Prediction:
  • Nash–Sutcliffe Efficiency (NSE): 0.9440
  • Kling–Gupta Efficiency (KGE): 0.9759
  • Root Mean Square Error (RMSE): 0.2296 meters
  • Percent Bias (PBIAS): 0.0215 ± 0.0146
• Inundation State Prediction:
  • Intersection over Union (IoU): 0.8180
  • Precision: 0.87
  • Recall: 0.93
  • F1 score: 0.8997
• Historical Flood Events: Demonstrated strong predictive performance with IoU values above 0.8 and F1 scores around 0.9 (e.g., 2022-11-01: Precision 0.9384, Recall 0.9238, F1 0.9306).
• Computational Efficiency: Achieved real-time inference with a total inferencing time of 2.9766 seconds for a one-day lead time, using 91.8803 million parameters and 92.9715 TFLOPs.

Contributions

• Proposed an AI-hydrodynamic hybrid approach that integrates advanced deep learning models with physics-based hydrodynamic simulations for real-time, reliable flood forecasting while preserving physical consistency.
• Designed an efficient FloodTransformer model featuring novel variable-sized cell embeddings for fast and effective processing of high-resolution, multi-modality inputs across large-scale catchments and urban regions.
• Introduced an innovative tokenized time-sequence encoding and decoding architecture that captures long-term temporal dependencies with minimized computational costs, enabling sequential predictions in a single run.

Funding

• New South Wales (NSW) State Emergency Service (SES) via the NSW Digital Restart Fund.

Citation

@article{Gu2025FloodTransformer,
  author = {Gu, Zhanzhong and Kang, Jeonghyun and Jin, Wei-Guo and Tong, Feifei and Guo, Y. Jay and Jia, Wenjing},
  title = {FloodTransformer: Efficient real-time high-resolution flood forecasting},
  journal = {Environmental Modelling & Software},
  year = {2025},
  doi = {10.1016/j.envsoft.2025.106832},
  url = {https://doi.org/10.1016/j.envsoft.2025.106832}
}