Explainable machine learning for anticipating county-level disaster recovery burden in the built environment

Abstract

This study aims to develop and evaluate an explainable machine-learning framework to anticipate county-level high recovery burden from natural disasters one year ahead using only public administrative and social-vulnerability data. Its substantive aim is to help built-environment decision makers shift from reactive post-event allocation to earlier, place-based prioritization of recovery and mitigation effort. A county-year panel was assembled for 3,042 US counties over 2014–2022 by combining FEMA Disaster Declarations Summaries with county-level CDC/ATSDR social vulnerability index data. The outcome was a high-burden county-year, defined as repeated county-level declarations involving public assistance and/or hazard mitigation programme activity. Class-weighted logistic regression, random forest and XGBoost were benchmarked using a temporally ordered train-validation-test split. SHAP values and model coefficients were used to interpret predictive drivers. The benchmark produced a moderate but decision-relevant predictive signal. Class-weighted logistic regression achieved the strongest out-of-sample performance on the hold-out period (area under the receiver operating characteristic curve 0.626; area under the precision–recall curve 0.140; F1-score 0.262) and outperformed the more flexible tree-based models. Across models, stronger prior programme burden, recent major declarations, population size, mobile-home prevalence and social vulnerability were associated with higher next-year burden. The dependent variable is an administrative proxy for recovery burden rather than a direct measure of physical loss, repair cost or household recovery outcomes. The empirical setting is the USA; transfer to other countries should therefore be treated as conceptual and methodological rather than statistical. The framework offers a transparent annual screening tool for local governments, regional planning bodies and emergency-management agencies seeking to identify counties likely to face elevated recovery pressure. Because it relies on public tabular data rather than expensive proprietary inputs, it is potentially transferable to resource-constrained settings that need interpretable anticipatory analytics. The paper contributes a reproducible public-data pipeline that links explainable artificial intelligence to disaster-resilience planning in the built environment. It also responds to the journal’s growing interest in technology-enabled resilience by showing that interpretable models, rather than only complex algorithms, can support forward-looking preparedness and recovery prioritization.