Zhang et al. (2026) Interpretable machine learning framework for urban flood susceptibility assessment: a multi-model comparison with spatial heterogeneity analysis in Yancheng

Identification

• Journal: Scientific Reports
• Year: 2026
• Date: 2026-05-09
• Authors: Xuan Zhang, Dongdong Guo
• DOI: 10.1038/s41598-026-47925-5

Research Groups

• Yancheng Kindergarten Teachers College, Yancheng, China

Short Summary

This study develops an interpretable machine learning framework to assess urban flood susceptibility in Yancheng, China, demonstrating that XGBoost provides the highest predictive accuracy and that flood drivers vary significantly across different geomorphic zones.

Objective

• To develop and compare a multi-model machine learning framework for urban flood susceptibility assessment and analyze the spatial heterogeneity of the driving factors using SHAP interpretability.

Study Configuration

• Spatial Scale: Yancheng City, China (coastal plain city)
• Temporal Scale: Not specified (based on historical flood point data)

Methodology and Data

• Models used: XGBoost, Random Forest (RF), Support Vector Machine (SVM), and SHAP (SHapley Additive exPlanations) for model interpretability.
• Data sources: 486 historical flood points and 10 conditioning factors (including topographic wetness index, elevation, and impervious surface ratio).

Main Results

• Model Performance: XGBoost outperformed other models with an AUC of 0.938 and accuracy of 0.891, compared to RF (AUC = 0.912) and SVM (AUC = 0.876).
• Key Drivers: The topographic wetness index (TWI), elevation, and impervious surface ratio were the most influential factors, with a combined contribution of 47.6%.
• Spatial Heterogeneity: Flood mechanisms differed by geomorphic unit:
  • Urban built-up areas: Dominated by impervious surfaces (31.5% high-risk).
  • Lixiahe lowland: Dominated by TWI (28.6%).
  • Coastal tidal flats: Dominated by elevation (23.1%).
  • Yellow River paleo-channel: Exhibited the lowest risk (11.2%).
• Susceptibility Mapping: 19.2% of the total study area is categorized as high to very high susceptibility, with Tinghu District showing the highest risk at 37.0%.

Contributions

• The research provides an integrated framework that combines predictive modeling with interpretability and spatial heterogeneity analysis, enabling the development of differentiated, zone-specific flood control strategies for coastal plain cities.

Funding

• Not mentioned in the provided text.

Citation

@article{Zhang2026Interpretable,
  author = {Zhang, Xuan and Guo, Dongdong},
  title = {Interpretable machine learning framework for urban flood susceptibility assessment: a multi-model comparison with spatial heterogeneity analysis in Yancheng},
  journal = {Scientific Reports},
  year = {2026},
  doi = {10.1038/s41598-026-47925-5},
  url = {https://doi.org/10.1038/s41598-026-47925-5}
}