Evaluating the spatial application of multivariable flood damage models

Abstract Flood damage arises from complex interactions between flooding processes and socio‐economic elements. Damage assessments for elements such as residential buildings rely on a modelled representation of local damage factors. Multivariable model approaches are well suited for damage prediction using detailed information on flood hazard and building characteristics. While broad explanatory variable ranges can improve model prediction performance, model transfer across geographical contexts often causes performance loss. This study aims to determine if increasing explanatory damage variables in a multivariable model improves residential building damage prediction and whether models based on local variables transfer between locations. We used empirical damage observations from six flood events to train and evaluate random forest regression model prediction performance. Spatial transfer is tested by splitting event datasets with trained models applied to original and external events. Variable analysis demonstrates model performance improvement with up to seven flood hazard and building characteristics, decreasing thereafter. Event models showed highest prediction precision for the original event, while models trained on all events transfer with comparable predictions for urban stormwater flooding. Prediction precision reduces when models transfer between locations affected by different flood types. This indicates flood damage models must replicate variability in local damage factors for reliable spatial transfer.