Expert-based versus data-driven flood damage models: A comparative evaluation for data-scarce regions

The knowledge about potential flood damage is a key issue for disaster risk reduction. However, the scarcity of empirical data has limited flood damage modeling in several regions. As a result, studies in data-scarce regions have mostly been restricted to either building exposure assessment or identification of vulnerability indicators without a further linkage to probable damage. As expert-based approaches do not require empirical damage data, they have a high potential for flood damage modeling in data-scarce regions. In this study, we carried out a comparative assessment between an expert-based and a data-driven approach. The expert-based approach systematically combines the vulnerability indicator method and synthetic what-if analysis based on the knowledge of regional experts. The data-driven approach integrates empirical flood damage data in the analysis applying a multivariate random forest model. Flood damage data, collected through interviews after two flood events in 2017 and 2019 at separate locations in Nigeria, were used to evaluate the performance of both methods based on developed damage grades. Results from both methods showed i) a predictive accuracy of 30% and 38% for the expert-based and data-driven approaches respectively, ii) that distance to channel, wall material, building condition, and building quality are significant regional damage drivers, and iii) comparable model performance can be achieved even with a reduced number of variables. Furthermore, the study demonstrated how experts are likely to underestimate damage at low water depths and how a difference in conformity to building standards can add to challenges in flood damage prediction.