Shang et al. (2025) A rainfall similarity-based dataset construction framework for enhanced urban inundation prediction using machine learning

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-10-10
• Authors: Yizi Shang, Li Hu, Yuxuan Gao, Wenming Zhang, Dongxin Liang
• DOI: 10.1016/j.jhydrol.2025.134395

Research Groups

• State Key Laboratory of Simulation and Regulation of Water Cycles in River Basins, China Institute of Water Resources and Hydropower Research, Beijing, China
• Department of Engineering, University of Cambridge, Cambridge, United Kingdom
• Department of Civil and Environmental Engineering, University of Alberta, Edmonton, Canada

Short Summary

This study proposes a rainfall similarity-based framework to construct high-quality datasets for machine learning models, significantly enhancing urban inundation prediction accuracy by incorporating process-oriented rainfall features. The framework improves predictive performance, with Random Forest models showing particular synergy and achieving inundation-extent accuracy exceeding 85 %.

Objective

• To develop and evaluate a rainfall similarity-based dataset construction framework that enhances data-driven urban inundation forecasting by incorporating process-oriented rainfall features.

Study Configuration

• Spatial Scale: A flood-prone metropolitan area (exemplified by Beijing, China).
• Temporal Scale: Individual rainfall events and their associated inundation, with dataset construction based on historical rainfall event characteristics (e.g., total rainfall depth, duration, intensity).

Methodology and Data

• Models used: Random Forest (RF), Deep Neural Network (DNN), Multi-distance fusion method (for rainfall similarity quantification).
• Data sources: Integrated datasets combining observational records and hydrodynamic simulations.

Main Results

• The rainfall similarity-based dataset construction framework significantly improves the predictive performance of machine learning models for urban inundation.
• Inundation-extent accuracy exceeded 85 % on average and approached 95 % in certain scenarios when using similarity-guided training.
• Both Random Forest (RF) and Deep Neural Network (DNN) models benefited from the framework, but RF consistently achieved higher accuracy than DNN, demonstrating strong synergy with the similarity-based dataset construction.
• The framework effectively bridges rainfall process analysis and machine learning modeling for reliable urban inundation prediction.

Contributions

• Introduction of a novel rainfall similarity-based framework for constructing high-quality, process-oriented datasets specifically designed to enhance data-driven urban inundation prediction.
• Development of a multi-distance fusion method to quantify rainfall event similarity based on key hydrological features (total rainfall depth, duration, maximum intensity, center location, spatial distribution pattern).
• Demonstration of significant improvements in urban inundation prediction accuracy (over 85 % average accuracy) by integrating this framework with machine learning models.
• Identification of Random Forest as a particularly effective machine learning model when combined with similarity-guided dataset construction for urban inundation forecasting.

Funding

• Not specified in the provided text.

Citation

@article{Shang2025rainfall,
  author = {Shang, Yizi and Hu, Li and Gao, Yuxuan and Zhang, Wenming and Liang, Dongxin},
  title = {A rainfall similarity-based dataset construction framework for enhanced urban inundation prediction using machine learning},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134395},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134395}
}