Interpretable machine learning framework for urban flood susceptibility assessment: a multi-model comparison with...

Area of Science:

• Environmental Science
• Urban Planning
• Geospatial Analysis

Background:

• Low-lying coastal cities are increasingly vulnerable to severe flooding.
• Global climate change and rapid urbanization exacerbate flood threats.

Purpose of the Study:

• To develop an integrated urban flood susceptibility assessment framework.
• To compare machine learning models for flood prediction.
• To analyze spatial heterogeneity of flood risk in coastal plains.

Main Methods:

• Utilized historical flood data and 10 conditioning factors for Yancheng City.
• Compared Random Forest (RF), XGBoost, and Support Vector Machine (SVM) models.
• Employed SHAP analysis for factor interpretability and spatial heterogeneity evaluation.

Main Results:

• XGBoost demonstrated superior predictive performance (AUC=0.938, accuracy=0.891).
• Topographic wetness index, elevation, and impervious surface ratio were key flood drivers (47.6% contribution).
• Distinct flood mechanisms were identified across geomorphic units, with urban areas and Lixiahe lowland showing higher susceptibility.

Conclusions:

• The integrated framework effectively assesses urban flood susceptibility in coastal plains.
• Findings support the development of differentiated flood control strategies.
• The study highlights the importance of considering spatial heterogeneity in flood risk management.