Rezvani et al. (2026) An XGBoost-SHAP framework for interpretable and probabilistic flood susceptibility mapping

Identification

• Journal: Natural Hazards
• Year: 2026
• Date: 2026-02-25
• Authors: Hossein Rezvani, Atefe Arfa, Hossein Shafizadeh‐Moghadam, Masoud Minaei
• DOI: 10.1007/s11069-025-07908-7

Research Groups

• University of Tehran, Tehran, Iran
• Department of Water Engineering and Management, Tarbiat Modares University, Tehran, Iran
• Department of Geography, Faculty of Letters and Humanities, Ferdowsi University of Mashhad, Mashhad, Iran

Short Summary

This study develops an interpretable framework for flood susceptibility mapping by integrating Extreme Gradient Boosting (XGBoost) with SHapley Additive exPlanations (SHAP). Applied to the Karkheh Basin, Iran, the framework achieved high predictive performance (AUC of 0.89) and provided transparent insights into the influence and interactions of key environmental factors on flood susceptibility.

Objective

• To implement and validate a flood susceptibility map for the Karkheh Basin using the XGBoost algorithm.
• To interpret the XGBoost model using SHAP to elucidate global feature importance and local variable interactions.
• To quantitatively evaluate the relative influence of key environmental and conditioning factors on flood occurrence.

Study Configuration

• Spatial Scale: Karkheh Basin, Iran (approximately 52,000 km²), located between latitudes 30° 58′–34° 56′ N and longitudes 46° 06′–49° 10′ E.
• Temporal Scale: Historical flood event of 2018 for flood inventory mapping; susceptibility mapping represents a static assessment based on historical conditions.

Methodology and Data

• Models used: Extreme Gradient Boosting (XGBoost) for predictive modeling, SHapley Additive exPlanations (SHAP) for model interpretability.
• Data sources:
  • Flood extent: Derived from Sentinel-1 Synthetic Aperture Radar (SAR) imagery (Level-1 Ground Range Detected (GRD) products, Interferometric Wide (IW) swath mode) using Otsu’s method for the 2018 flood event.
  • Conditioning factors (9 geospatial and topographic variables): elevation, slope, plan curvature, profile curvature, flow accumulation, Stream Power Index (SPI), distance to rivers, stream density, and Normalized Difference Vegetation Index (NDVI).
  • Dataset: 2000 samples, split into 70% training and 30% testing sets.

Main Results

• The XGBoost model demonstrated high predictive performance for flood susceptibility mapping, achieving an Area Under the Curve (AUC) of 0.89, an overall accuracy of 81.83%, precision of 80.71%, recall (sensitivity) of 83.67%, and an F1-score of 82.16%.
• SHAP analysis identified slope, altitude, and Normalized Difference Vegetation Index (NDVI) as the most influential factors determining flood susceptibility.
• SHAP dependence plots quantitatively confirmed that flood susceptibility increases at lower altitudes and decreases on steeper slopes.
• The Stream Power Index (SPI) exhibited a highly conditional, non-linear relationship, with its effect on flood risk depending on interactions with other topographic factors.
• The SHAP interaction plot revealed that the protective effect of steep slopes against flooding diminishes significantly at higher altitudes, highlighting complex interactions between factors.
• The flood susceptibility map indicated that central and southwestern regions of the Karkheh Basin, characterized by lower elevations and proximity to water bodies, exhibit the highest susceptibility.

Contributions

• Presents a novel, transparent, and interpretable flood susceptibility mapping framework by integrating the high predictive power of XGBoost with the explainability of SHAP.
• Provides precise, post-hoc insights into how individual features influence flood probabilities, moving beyond traditional overall variable importance metrics.
• Quantitatively elucidates complex, non-linear relationships and interactions between hydro-environmental factors (e.g., slope and altitude interaction) that drive flood susceptibility.
• Enhances stakeholder confidence and promotes broader adoption of machine learning models in disaster management by bridging the gap between complex 'black-box' models and practical, actionable planning.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Citation

@article{Rezvani2026XGBoostSHAP,
  author = {Rezvani, Hossein and Arfa, Atefe and Shafizadeh‐Moghadam, Hossein and Minaei, Masoud},
  title = {An XGBoost-SHAP framework for interpretable and probabilistic flood susceptibility mapping},
  journal = {Natural Hazards},
  year = {2026},
  doi = {10.1007/s11069-025-07908-7},
  url = {https://doi.org/10.1007/s11069-025-07908-7}
}