Pan et al. (2026) A Study on Rapid Dynamic Flood Forecasting in Small Watersheds Using a GNN-Transformer Approach Integrated with Spatial Physical Information

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-03-01
• Authors: Xinxin Pan, Jingming Hou, D. Li, Yanhong Wang, Xiaodong Li, Jiantao Sun, Chenchen Fan, Yongping Yang
• DOI: 10.1007/s11269-025-04360-x

Research Groups

• State Key Laboratory of Eco-Hydraulics in Northwest Arid Region of China, Xi’an University of Technology, Xi’an, Shaanxi, China
• Nanchang Institute of Technology, Nanchang, Jiangxi, China
• Drainage Management Center of the Hengshui Municipal Urban Management and Law Enforcement Bureau, Hengshui, China
• Shaanxi Provincial Water and Drought Prevention Center, Xi’an, Shaanxi, China

Short Summary

This study develops a novel GNN-Transformer deep learning model for rapid flood forecasting in small watersheds, integrating static physical information and dynamic rainfall data. The model achieves high accuracy (NSE > 0.99, RMAE < 7%) and significantly improved computational efficiency (100-200 times faster) compared to traditional hydrodynamic models.

Objective

• To develop a rapid, accurate, and generalizable flood forecasting model for small watersheds by integrating Graph Neural Networks (GNN) and Transformer architectures, leveraging static physical attributes of surface grids and dynamic rainfall time series.

Study Configuration

• Spatial Scale: The Wangmaogou watershed in Suide County, Shaanxi Province, China, covering approximately 5.74 km². The main channel extends 3.75 km, with elevations ranging from 940 m to 1,200 m. The computational domain uses a 2 m × 2 m grid resolution (1,606 × 1,710 cells), with 1,435,271 valid grids for the model.
• Temporal Scale: Design rainfall events with 2-hour durations, followed by a 1-hour zero-rainfall period, resulting in a total simulation time of 10,800 seconds (3 hours). Flood data was generated for 190 time steps with an output interval of 600 seconds. Hydrodynamic model validation used a 5-hour rainfall event with a 10-hour simulation duration.

Methodology and Data

• Models used:
  • Proposed: GNN-Transformer model, which integrates Graph Neural Networks (GNN) to capture spatial relationships among grids and their physical attributes, and Transformer architectures to model dynamic rainfall time series and flood evolution.
  • Reference/Baseline: A two-dimensional shallow water equations (SWEs) based hydrodynamic model, accelerated using GPU (CUDA/C++), employing the Harten Lax van Leer Contact (HLLC) approximation Riemann solver, Monotonic Upstream-Centered Scheme for Conservation Laws (MUSCL) scheme, and Runge-Kutta method.
• Data sources:
  • Rainfall Data: Design rainfall events generated using the Chicago rainfall pattern generator based on the Yulin City storm intensity formula, covering return periods from 1 to 250 years.
  • DEM Data: High-resolution Digital Elevation Model (DEM) with a 2 m × 2 m spatial resolution, acquired via UAV-based LiDAR survey. Derived static physical attributes include elevation, slope, aspect, and curvature for each grid cell.
  • Land Use Data: Classified into eight categories (water bodies, orchards, forests, grassland, villages, roads, bare land, terraced fields) based on high-resolution satellite imagery, with assigned infiltration rates (e.g., 0 to 4.5 mm·h⁻¹) and Manning’s roughness coefficients (e.g., 0.01 to 0.23 s/m^(1/3)).
  • Hydrodynamic Simulation Outputs: Water depth (m) and flow velocity (m/s) for all grids and time steps, generated by the hydrodynamic model, served as the training and reference dataset for the GNN-Transformer model.
  • Data Augmentation: Geometric transformations (horizontal flipping, vertical flipping, 180° rotation) were applied to both input static physical features and output flood variables to enhance model generalization.

Main Results

• The proposed GNN-Transformer model demonstrated high accuracy in flood forecasting, achieving Nash-Sutcliffe Efficiency (NSE) values no lower than 0.990, Mean Absolute Error (MAE) values no greater than 1.220, and relative MAE (RMAE) values no greater than 7.063 when compared to hydrodynamic simulation outputs.
• The average relative error for water depth and flow velocity across all grids remained within 15%, effectively capturing the spatial and temporal dynamics of flash flood events.
• The model significantly improved computational efficiency, achieving speedups of 111.34 to 126.12 times compared to the GPU-accelerated hydrodynamic model for flood event predictions. While training took 1,120 minutes, prediction for a 3-hour event was completed in approximately 1.3-1.4 minutes.
• The hydrodynamic model, used as a baseline, was validated against an idealized V-shaped watershed (NSE = 0.98) and field observations from the Wangmaogou watershed (NSE = 0.78), confirming its reliability for generating training data.

Contributions

• Developed a novel deep learning framework (GNN-Transformer) that effectively integrates static spatial physical features of watersheds with dynamic rainfall time series for rapid and accurate flood forecasting.
• Achieved a substantial improvement in computational efficiency (100-200 times faster) compared to traditional hydrodynamic models, providing crucial lead time for emergency flood management and early warning systems.
• Demonstrated high forecasting accuracy (NSE > 0.99, RMAE < 7%) for key flood variables (water depth, flow velocity, discharge), meeting the stringent requirements for small-watershed flood prediction.
• Enhanced the generalization and extrapolation capabilities of flood forecasting models by explicitly learning both spatial adjacency relationships through GNNs and temporal dynamics through Transformers.
• Offers a robust and intelligent technical solution for advancing flood management systems, including forecasting, warning, rehearsal, and contingency planning.

Funding

• National Key R&D Program of China (2024YFC3012402; 2024YFC3012403)
• National Natural Science Foundation of China (52409104)
• China Postdoctoral Science Foundation (Grant Number 2024M762625)

Citation

@article{Pan2026Study,
  author = {Pan, Xinxin and Hou, Jingming and Li, D. and Wang, Yanhong and Li, Xiaodong and Sun, Jiantao and Fan, Chenchen and Yang, Yongping},
  title = {A Study on Rapid Dynamic Flood Forecasting in Small Watersheds Using a GNN-Transformer Approach Integrated with Spatial Physical Information},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-025-04360-x},
  url = {https://doi.org/10.1007/s11269-025-04360-x}
}