AI based assessment of flash flood damages to company

Flash floods like the flood in 2021 in the west of Germany result in particularly large numbers of fatalities and heavy asset damages. Among the several flood-exposed sectors, companies are severely affected by floods and constitute a significant component of overall flood damages. However, understanding and modeling the underlying processes influencing flash flood losses for companies is specifically challenging due to (1) heterogeneity in terms of sectors, building size and type, number of employees, and equipment, and (2) scarcity of company-specific flood loss data. In comparison to fluvial floods, the influence of flood characteristics and hydro-dynamic processes on damage is different in the case of flash floods.  To tackle this challenge, multi-variate probabilistic flash flood loss models are developed based on feature selection using empirical data from detailed surveys conducted with companies after the flash floods of 2002, 2016, and 2021 in Germany. The machine learning ensemble-based approach of feature selection revealed the significance of the following hazard variables (water depth, flow velocity, contamination), exposure variables (sector, number of employees, size of premise), and vulnerability variables (implementation of precautionary measures, early warning time, flood experience) in determining flood losses. The Bayesian Networks-based flood loss models developed in this study provide probability distributions of estimated losses and as such inherently quantify uncertainties.