Beyond discrete indicators: Modeling intersectional flood vulnerability

Social vulnerability to flooding is shaped by intersectional social marginalization, yet most quantitative assessments employ indicators of single populations. This study applies spatial machine learning to examine how the intersectional social vulnerability indicators of poverty-race, poverty-housing tenure, and race-housing tenure compare with traditional discrete indicators of single populations in predicting flood exposure in California. Using geographically weighted random forests and partial dependence plots, we model spatial heterogeneity and non-linear relationships between social vulnerability and exposure. We quantified flood exposure using a population-adjusted measure derived from building footprints and modeled 500-year fluvial and pluvial flood hazard. The results reveal distinct explanatory power of discrete and intersectional indicators. Variable importance analysis shows that intersectional indicators, such as Poor Renters and Non-white Renters, have stronger predictive importance than their discrete counterparts, particularly in urban regions, with mean local IncMSE values of 15.6–16.9 % compared to 12.3–14.8 %. Partial dependence analysis revealed threshold effects of non-linear indicator influence, with predicted exposure increasing sharply once intersectional populations exceed ∼60 % of tract-level representation. Our findings highlight limitations of assuming uniform indicator effects, and the need for non-linear, spatially adaptive models that increase conceptual alignment between social vulnerability theory and indicator modeling by integrating intersectional dimensions.