Hierarchical Bayesian approach for flood loss modelling for Europe

14:00 Tuesday 28 May

OC057

Room S11

 

Nivedita Sairam (Germany) 1; Heidi Kreibich (Germany) 1; Kai Schröter (Germany) 1

1 - German Research Institute for Geosciences - GFZ, Potsdam

Reliable and accurate predictions of flood losses are imperative in making efficient risk based adaption strategies. Hence, flood loss estimation is a crucial step in Flood Risk Management. Loss models can either be based on empirical data from flood events or synthetic, i.e. based on engineering expert knowledge. Empirical flood loss models are generally developed using data pertaining to flood and exposure characteristics and reported loss. These models are temporally and spatially localized. However, in practice the models developed for one region are transferred to another, without testing their suitability for the region and purpose. Several validation studies show that these models are associated with high uncertainties when applied to regions outside the area or country for which the models were developed.

A hierarchical Bayesian approach is implemented for developing a flood loss model by combining empirical flood loss data from multiple regions and events in Europe. This approach tries to estimate losses while accounting for spatio-temporal variations in the damage processes. The hierarchical model is a multi-level regression model that estimates individual sets of coefficients using the predictors to model the outcomes in each group. Additionally, there is also a second probability distribution over these group-level parameters that govern the variability between the groups. In our hierarchical model structure, flood loss is modeled as a function of water depth (depth-damage curve) using two levels of groupings – region and event over which the second probability distribution for each region lies.

The hierarchical model is a middle-ground between generalized models (one set of loss estimation parameters for entire Europe) and localized models (independent set of loss estimation parameters for each event). Hence, the parameters of the hierarchical model become more independent in the presence of abundant data for a specific event. Similarly, when few data is available, the parameters shrink closer to the over-arching distribution parameters. We validate the predictions of the hierarchical model against official damage data and popular reference models such as FLEMOps for Germany and Multi-Colored-Manual for UK. Bayesian parameter estimation combined with hierarchical modelling allows quantification of uncertainty at each level in the hierarchy (region, region-event and household). As such, a consistent flood loss modelling approach for Europe can be achieved, enabling coherent uncertainty analyses.