Clustering analysis of damage-related hydro-meteorological extreme events in southwest Germany

Clusters of different hydro-meteorological extremes occur particularly in the summer and are only statistically significant with an event definition of a fixed length.

Extreme weather events can cause considerable damage, especially when they occur together or in succession (link). In a current CEDIM project, funded by the Stiftung Umwelt und Schadenvorsorge, the temporal accumulation of different types of damage-related hydro-meteorological extremes is investigated: Flood events triggered by stratiformly (fluvial) or convectively (pluvial) dominated precipitation or a hybrid of these two classes, hail, large-scale storms and convective wind gusts. The aim of this work is to identify and quantify a temporal clustering of damage-related extremes and then to relate them to large-scale atmospheric flow patterns and weather conditions (e.g. weather regimes or teleconnections). This enables to estimate how often and under what conditions damage-related extremes cluster in a given year.

The study is based on insured building damage in Baden-Württemberg from 1986 to 2023 on a daily basis, which is aggregated into events of flexible length via percentiles using the peak-over-threshold method. Most major loss events occur in the summer months (hail, thunderstorms, convectively-dominated flood events) or winter months (stratiform-dominated flood events, large-scale storms). Looking at longer periods, e.g. 28 days in the course of a year, corresponding event clusters (i.e. two events or more) are detected in these two seasons, with the hail and windstorm clusters being the most relevant in terms of damage (Fig. 1a). In the summer months, there is an increased clustering of the individual event types from around the middle of the study period, i.e. at the beginning of the 2000s. This is even more evident when clusters of different event types are formed (Fig. 1b); clusters of the three convective event types are predominant. In the winter months, fluvial floods in particular occur in clusters with windstorms.

Fig. 1: (a) Clusters (= more than one event within 28 days) of loss events of the same type (pluvial, fluvial and mixed flood events, synoptic-scale storm events, hail and convective gusts) over the course of all years (vertical) and over the course of one year (horizontal). The size of the circles refers to the relative amount of damage. (b) As a), but with regard to combinations of different damage events.

The degree of clustering is statistically quantified using Ripley's K. Results show that for most event types there is no statistically significant clustering on the sub-seasonal to seasonal scale. Exceptions to this are fluvial flooding events in the winter months, which show clustering from around 23 days up to the entire season. Combinations of different event types, on the other hand, do not cluster significantly. However, if events are defined with a fixed event length (hours-clause method of the insurance industry, 72 hours for storm and hail and 168 hours for flood events), a statistically significant clustering can be seen for the combination of convective gusts and mixed flood events in the summer months and synoptic-scale storms in the winter months. In contrast to the Peaks-over-Threshold method, combinations of different event types also show significant clustering (see Fig. 2). This proves that the method of event definition (i.e. the event length) considered has a major influence on the degree of clustering of damage events of the same resp. various types.

Fig. 2: Visualisation of the clustering for loss event days based on the Ripley's K method, relative to the clustering window considered; based on data of the 10 % of the most expensive events of the respective combination in Baden-Württemberg between 1986 and 2023 (black line in each case). Blue represents the significance by median (dashed line) and 95 % confidence interval (shading) of a Monte Carlo sample with 1000 simulations of homogeneous Poisson processes. Event days are defined using the Peaks-over-Threshold method (top) or the Hours-Clause method (bottom). If the solid line is outside the confidence interval, the clustering is statistically significant.

Associated institute at KIT: Institute of Meteorology and Climate Research (IMKTRO)
Author: Katharina Küpfer (Jun. 2024)