Interpretable machine learning framework for urban flood susceptibility assessment: a multi-model comparison with spatial heterogeneity analysis in Yancheng.

Under the combined effects of global climate change and rapid urbanization, low-lying coastal plain cities face increasingly severe flood threats. This study developed an urban flood susceptibility assessment framework integrating multi-model comparison, SHAP interpretability analysis, and spatial heterogeneity evaluation, using Yancheng City as a case study. Based on 486 historical flood points and 10 conditioning factors, the predictive performance of Random Forest (RF), XGBoost, and Support Vector Machine (SVM) was compared. Results showed XGBoost performed best (AUC = 0.938, accuracy = 0.891), significantly outperforming RF (AUC = 0.912) and SVM (AUC = 0.876). SHAP analysis identified topographic wetness index (TWI), elevation, and impervious surface ratio as key driving factors, with a cumulative contribution of 47.6%. Spatial heterogeneity analysis revealed distinct flood mechanisms across geomorphic units: urban built-up areas were dominated by impervious surfaces (31.5% high-risk), Lixiahe lowland by TWI (28.6%), coastal tidal flats by elevation (23.1%), while the Yellow River paleo-channel showed the lowest risk (11.2%). Susceptibility mapping indicated 19.2% of the study area at high to very high levels, with Tinghu District reaching 37.0%. The proposed integrated framework provides scientific support for differentiated flood control strategies in coastal plain cities.