Flood Damage Evaluation for Buildings in a Small Island Developing State

Abstract

Flooding poses significant social and economic challenges for Small Island Developing States (SIDS). Despite frequent and damaging flood events, SIDS are underrepresented among global flood damage models. This study evaluated tree‐based learning algorithm performance for building damage prediction, using a new data set from 2012 Tropical Cyclone Evan in Apia, Samoa. Empirical building damage data was used to identify relationships with explanatory hazard and exposure variables, and test uni‐ and multivariable regression model performance in response to varying hyperparameter and explanatory variable combinations. Multivariable ensemble models showed higher precision and reliability than tree‐based deterministic and univariable ensembles. A high‐performing Extreme Gradient Boosting multivariable model showed prediction precision improvements for up to five variable additions, with reduced performance from variable additions thereafter. Water depth above floor level and building area caused the highest precision improvement. Building area importance for damage is a promising finding, warranting further investigation of geometric variable effects on building flood damage and damage model capacity for transfer between geographical locations. Such investigations should align with local knowledge of building damage processes to ensure appropriate explanatory variables are collected and applied in flood damage models.