Estimating the impact of seasonal extremal dependence with a Max-Stable process for modeling extreme precipitation events over Berlin-Brandenburg

Recent developments in Extreme Value Theory have led to the adoption of multivariate methods for modeling extreme rainfall in a way that the spatial dependence between the different measuring stations is used to "borrow“ information. A commonly used method, Brown-Resnick Max-Stable processes, extends the geostatistical concept of the variogram to suit block maxima, allowing to explicitly model the spatial extremal dependence shown by the data. This extremal dependence usually stems from physical processes that generate rainfall in such a way that several stations are affected simultaneously by the same extreme event, such as convective storms or frontal events. Depending on the region, this dependence can change in time, as different meteorological processes dominate the rainfall generation process for different seasons.

In this study, we analyze in the Berlin-Brandenburg region the change in extremal dependence for annual block maxima. We consider two different seasons – winter and summer – to investigate the effects of two different rainfall generating processes: frontal rainfall is more likely to occur in winter, while convection is dominating in summer. Furthermore, we investigate how this extremal dependence affects the accuracy of the estimation of return levels by using a Brown-Resnick Max-Stable process and comparing the estimated return levels to the results of a covariates model assuming spatial independence. We obtain the uncertainty of our estimates within a Bayesian modeling framework. The bivariate extremal coefficient shows a notable difference in the extremal dependence for summer and winter. Moreover, we observe a difference in the skill of the model when comparing the two seasons, suggesting that the difference in the extremal dependence has an impact on the marginal estimates from the model.