Identifying final demand points for aid in the aftermath of sudden-onset climate-related disasters in Peru: A supervised learning approach

Sudden-onset climate-related disasters, such as floods and landslides, pose significant challenges for disaster risk management in developing countries like Peru, which was selected as a case due to its geographical and economic diversity and the exposure of its cities to these natural hazards. Post-disaster performance and the upcoming impacts of future disasters depend on the effectiveness of disaster risk reduction strategies. Thus, having a robust screening of households at risk and accurate estimates of relief demand matters to disaster preparedness. This study proposes a supervised learning approach to train binary classifiers that predict household vulnerability to floods and landslides based on their socio-economic, geographic, health, and social characteristics. The classifiers are trained using XGBoost, a gradient boosting algorithm, with a custom objective function that prioritizes minimizing false negatives (unmet demand). Additionally, partial dependence plots and SHAP values provide interpretability to the models, allowing decision-makers to understand the underlying factors contributing to household vulnerability. The classifiers achieve promising accuracy on hold-out, one-year-ahead, test data, demonstrating their potential for informing disaster preparedness and response efforts. The study contributes to the literature by providing a data-driven method for demand estimation and vulnerability assessment in the context of recurrent climate-related disasters. The paper provides readers with an outline of the practical implications of disaster risk reduction.