Learning to estimate losses of compound inland flooding with Bayesian multilevel models

Research on natural hazards has increasingly become concerned with compound events, i.e. multiple hazards that may coincide in space and time or happen sequentially. Such events may lead to unexpected or unwanted amplifications of the impacts compared to those of individual hazards. To what extent the co-occurrence of hazards exacerbates impacts and losses is largely undocumented.

Fluvial, pluvial, and coastal floods are commonly understood as distinct hazards. However, floods can be further differentiated, for example, into river floods, urban floods or flash floods. Most flood-loss models follow such a distinction of flood pathways, assuming that the damaging processes are also different and disconnected from each other. Recent studies have shown that vulnerability varies between distinct flood pathways. But loss modelling under the co-occurrence of distinct flood pathways has not yet been further examined.

Germany has faced severe floods since 2002, including preconditioned events (e.g. the rain-on-snow floods of 2006 and 2011; the excessive rainfall on already saturated soil of 2013), co-​occurrence of multiple/consecutive hazards in the same geographical region, and spatially compound floods (such as in 2002, 2010 and 2016). Survey data collected after floods in Germany between 2002 and 2016 show that around 60% of 1150 surveyed households reported having been affected by more than one flood pathway indicating the process complexity at flooded properties.

With these survey data, we learned a model for estimating residential flood losses. We used Bayesian multilevel models that probabilistically incorporate uncertainty and allow for partial pooling of the data. Such models are capable of differentiating parameters for different flood pathways, but learn the parameters from all data simultaneously. One missing piece of information, however, is the contribution of each individual flood pathway to the overall financial impact. Although we cannot separate the magnitude of each flood pathway in our data, they are understood as distinct processes.

Bayesian inference is data driven and explicitly includes prior knowledge or beliefs. Our model thus assigns a prior belief of the extent to which co-occurrent pathways contribute to an increased loss. Therefore, five weight sets spanning a reasonable range, from averaged weighed to a total sum of effects, are implemented here in order to find eventual differences in the vulnerability of residential buildings to the different pathways and determine how they combine together into a single (potentially synergetic) impact.

This contribution introduces five model variants, their components, and shows the first differences across the model parameters. With this we also highlight the need to engage with the procedure of defining the weights sets, which still remains a challenge for the study of compound event' impacts.