Evaluation of residential building damage for the July 2021 flood in Westport, New Zealand

Abstract

Reliable flood damage models are informed by detailed damage assessments. Damage models are critical in flood risk assessments, representing an elements vulnerability to damage. This study evaluated residential building damage for the July 2021 flood in Westport, New Zealand. We report on flood hazard, exposure and damage features observed for 247 residential buildings. Damage samples were applied to evaluate univariable and multivariable model performance using different variable sample sizes and regression-based supervised learning algorithms. Feature analysis for damage prediction showed high importance of water depth variables and low importance for commonly observed building variables such as structural frame and storeys. Overfitting occurred for most models evaluated when more than 150 samples were used. This resulted from limited damage heterogeneity observed, and variables of low importance affecting model learning. The Random Forest algorithm, which considered multiple important variables (water depth above floor level, area and floor height) improved predictive precision by 17% relative to other models when over 150 damage samples were considered. Our findings suggest the evaluated model performance could be improved by incorporating heterogeneous damage samples from similar flood contexts, in turn increasing capacity for reliable spatial transfer.