Lanjie et al. (2025) Efficient urban flood surface reconstruction: integrating deep learning with hydraulic principles for sparse observations

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-10-23
• Authors: Xu Lanjie, Jingming Hou, Tian Wang, Qingyuan Guo, Donglai Li, Pan Xinxin
• DOI: 10.1016/j.jhydrol.2025.134439

Research Groups

• State Key Laboratory of Water Engineering Ecology and Environment in Arid Area, Xi’an University of Technology, China
• Xi’an Meteorological Bureau of Shanxi Province, China

Short Summary

This study proposes a novel deep learning framework, Sparse-Point Learning and Interpolated Surface Reconstruction (SPIR), to efficiently and accurately simulate high-resolution urban flood inundation by integrating a lightweight neural network with hydrodynamic-informed interpolation. The framework significantly reduces computational time while maintaining high prediction accuracy compared to traditional hydrodynamic models.

Objective

• To develop an efficient and accurate deep learning-based framework for high-resolution urban flood surface reconstruction from sparse observations, addressing the computational limitations of traditional two-dimensional (2D) hydrodynamic models.

Study Configuration

• Spatial Scale: Urban areas, high-resolution flood inundation mapping.
• Temporal Scale: Real-time flood forecasting and prediction.

Methodology and Data

• Models used:
  • Sparse-Point Learning and Interpolated Surface Reconstruction (SPIR) framework.
  • FC-DecMaskNet (Fully Connected–Deconvolutional Neural Network) for flood extent masks and water depth prediction.
  • GAST (GPU Accelerated Surface Water Flow and Associated Transport) model (for comparison).
• Data sources:
  • Implied use of high-resolution observational data for model training and evaluation.
  • Flood extent masks generated through thresholding and morphological filtering.
  • Representative locations selected based on frequency statistics and spatial clustering.

Main Results

• The SPIR framework can generate a complete flood inundation map in approximately 6.3 seconds, achieving a nearly 2657-fold speed-up compared to the physics-based GAST model.
• It significantly reduces training complexity while improving prediction accuracy compared to conventional deep learning approaches.
• Evaluation results show a Probability of Detection (POD) over 99 % and a False Alarm Ratio (FAR) below 0.78 %.
• The mean Root Mean Square Error (RMSE) for water depth prediction is 0.0018 m.
• The 99th percentile of grid-wise errors (Q99) averages 0.0064 m.

Contributions

• Introduction of the novel SPIR framework, which integrates deep learning with hydrodynamic principles for efficient and accurate urban flood surface reconstruction from sparse observations.
• Development of a lightweight FC-DecMaskNet for predicting flood extent and water depths at representative locations.
• Significant improvement in computational efficiency (2657-fold speed-up) compared to physics-based models, making it suitable for real-time applications.
• Demonstrated high accuracy in flood inundation mapping (POD > 99%, FAR < 0.78%, RMSE = 0.0018 m).

Funding

• Not specified in the provided text.

Citation

@article{Lanjie2025Efficient,
  author = {Lanjie, Xu and Hou, Jingming and Wang, Tian and Guo, Qingyuan and Li, Donglai and Xinxin, Pan},
  title = {Efficient urban flood surface reconstruction: integrating deep learning with hydraulic principles for sparse observations},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134439},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134439}
}