Huang et al. (2025) Urban flood prediction using a hybrid XGBoost-enhanced U-Net model

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-12-21
• Authors: Xiao Huang, Yanfen Geng, Peng Liu, Xinyu Hu, Zhili Wang
• DOI: 10.1016/j.jhydrol.2025.134822

Research Groups

• Department of Port & Waterway Engineering, School of Transportation, Southeast University, Nanjing, China
• Nanjing Hydraulic Research Institute, Nanjing, Jiangsu, China

Short Summary

This paper proposes an XGBoost-Enhanced U-Net (XGB-U-Net) model for timely and accurate urban flood prediction, integrating physical mechanisms with deep learning. The hybrid model demonstrates superior accuracy and efficiency compared to U-Net and XGBoost alone, particularly in complex urban environments under spatially heterogeneous rainfall.

Objective

• To develop and evaluate a hybrid XGBoost-Enhanced U-Net (XGB-U-Net) model for timely and accurate urban flood prediction, addressing the limitations of traditional time-consuming hydrological/hydrodynamic models and single deep learning approaches.

Study Configuration

• Spatial Scale: Regions with varying urbanization levels (non-built, low-density, and high-density built areas), with predictions made on a gridded representation of inundation.
• Temporal Scale: Designed for real-time urban flood forecasting, implying short-term prediction horizons.

Methodology and Data

• Models used: XGBoost-Enhanced U-Net (XGB-U-Net), U-Net (for comparison), XGBoost (for comparison), a hydrodynamic model (to generate training data), Spatially Variable Rainfall model (SVR-SS) (to generate spatially heterogeneous rainfall scenarios).
• Data sources: 11 rainfall- and terrain-related features (inputs); inundation results from a hydrodynamic model (outputs for training).

Main Results

• The XGB-U-Net model achieved superior accuracy compared to U-Net across all urbanization levels.
• For non-built areas, XGB-U-Net showed MSE of 0.0144 m² and AUC of 0.96, compared to U-Net's MSE of 0.0167 m² and AUC of 0.93.
• For low-density built areas, XGB-U-Net showed MSE of 0.0180 m² and AUC of 0.85, compared to U-Net's MSE of 0.0223 m² and AUC of 0.82.
• For high-density built areas, XGB-U-Net showed MSE of 0.0189 m² and AUC of 0.77, compared to U-Net's MSE of 0.0224 m² and AUC of 0.72.
• Building density significantly impacts model performance, with accuracy declining as density increases across all evaluated models.

Contributions

• Proposes a novel hybrid XGBoost-Enhanced U-Net (XGB-U-Net) model that effectively integrates physical mechanisms with deep learning for urban flood prediction.
• Combines the image-processing capability of Convolutional Neural Networks (U-Net) with the high-dimensional feature handling of Extreme Gradient Boosting (XGBoost), enhancing predictive precision and efficiency.
• Demonstrates the superior accuracy and robustness of the XGB-U-Net model across varying urbanization levels and under spatially heterogeneous rainfall scenarios, offering a robust solution for real-time urban flood forecasting in complex urban environments.

Funding

• No funding information was provided in the paper text.

Citation

@article{Huang2025Urban,
  author = {Huang, Xiao and Geng, Yanfen and Liu, Peng and Hu, Xinyu and Wang, Zhili},
  title = {Urban flood prediction using a hybrid XGBoost-enhanced U-Net model},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134822},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134822}
}