Chu et al. (2026) An interpretable AutoML–SHAP approach for rapid urban pluvial flooding prediction

Identification

• Journal: Environmental Modelling & Software
• Year: 2026
• Date: 2026-04-01
• Authors: W. P. Chu, Chaosen Jin, Chunxiao Zhang, Suchuang Di, Tianbao Wang, Heng Li, Yuqian Hu
• DOI: 10.1016/j.envsoft.2026.106985

Research Groups

• School of Artificial Intelligence, China University of Geosciences in Beijing, Beijing, China
• Hebei Key Laboratory of Geogeographical Digital Twin and Collaborative Optimization, China University of Geosciences Beijing, Beijing, China
• Ningbo Hydrological Monitoring and Forecasting Center, Ningbo, China
• Observation and Research Station of Beijing Fangshan Comprehensive Exploration, Ministry of Natural Resources, Beijing, China
• Department of Big Data Analysis and Application, Beijing Water Science and Technology Institute, Beijing, China
• State Key Laboratory of Remote Sensing Science, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China
• Key Laboratory of Virtual Geographic Environment (Ministry of Education of PR China), Nanjing Normal University, Nanjing, Jiangsu, China

Short Summary

This study introduces an interpretable AutoML-SHAP framework for rapid urban pluvial flooding prediction, achieving a 2000-fold speedup over traditional hydrodynamic models while providing transparent insights into key flood drivers and their nonlinear responses.

Objective

• To develop a novel, accurate, efficient, and transparent framework for urban pluvial flood forecasting by integrating Automated Machine Learning (AutoML), SHapley Additive exPlanations (SHAP), and Generalized Additive Models (GAM) to overcome the computational burden and black-box nature of existing methods.

Study Configuration

• Spatial Scale: Shijingshan District, Beijing, China.
• Temporal Scale: Focuses on rapid prediction capabilities for real-time flood forecasting and early warning systems.

Methodology and Data

• Models used: Automated Machine Learning (AutoML) for optimizing XGBoost, SHapley Additive exPlanations (SHAP), Generalized Additive Models (GAM). The physically-based hydrodynamic model MIKE FLOOD was used as a benchmark and for generating training data.
• Data sources: Simulation data generated by the MIKE FLOOD model for various rainstorms, used to train the machine learning surrogate model. Input features include elevation, topographic wetness index, and manhole density.

Main Results

• The AutoML-optimized XGBoost model demonstrated a 2000-fold speedup compared to MIKE FLOOD, maintaining strong generalization performance on unseen rainstorms.
• Integrated SHAP–GAM analysis successfully revealed nonlinear responses of flood drivers, which were consistent with hydrological expectations, thereby reinforcing the credibility of the learned decision behavior.
• Depth-stratified analysis identified elevation, topographic wetness index, and manhole density as a core set of drivers, collectively contributing over 43% across different flood depth bins.
• The hierarchical importance of these core drivers was observed to reorganize as flood severity increased.

Contributions

• Proposes a novel, integrated framework combining AutoML, SHAP, and GAM for urban flood prediction, effectively addressing both computational efficiency and interpretability challenges.
• Achieves a significant speedup (2000-fold) over physically-based hydrodynamic models, enabling rapid, real-time flood forecasting for operational deployment.
• Provides transparent and explainable insights into the nonlinear relationships between urban flood drivers and flood severity through the SHAP-GAM analysis, enhancing the trustworthiness and utility of machine learning surrogates.
• Identifies and quantifies the stage-specific dynamics and hierarchical reorganization of key flood drivers (elevation, topographic wetness index, manhole density) under varying flood severities.

Funding

Not provided in the given paper text snippet.

Citation

@article{Chu2026interpretable,
  author = {Chu, W. P. and Jin, Chaosen and Zhang, Chunxiao and Di, Suchuang and Wang, Tianbao and Li, Heng and Hu, Yuqian},
  title = {An interpretable AutoML–SHAP approach for rapid urban pluvial flooding prediction},
  journal = {Environmental Modelling & Software},
  year = {2026},
  doi = {10.1016/j.envsoft.2026.106985},
  url = {https://doi.org/10.1016/j.envsoft.2026.106985}
}