Seminar: Compound climate events: statistical modeling and influence of multivariate bias correction on their evolution under climate change
By Grégoire Jacquemin

Grégoire Jacquemin will give a seminar on Tuesday 27/10 at 2pm in room 5A126.

Compound climate events: statistical modeling and influence of multivariate bias correction on their evolution under climate change

Compound events are commonly defined as the combination of multiple drivers and/or hazards that contributes to societal or environmental risk. They often result in amplified impacts compared to individual events. Accurate statistical modeling of multiple variables and their dependence in the extremes is required not to underestimate the risks associated to compound events. Moreover, the evolution in terms of frequency and intensity of compound events under climate change is yet unclear. First, a novel nonparametric approach relying on bivariate generalized Pareto distributions is proposed to estimate the return period of bivariate compound events. This new approach is motivated and its interest is assessed by comparison to a commonly used copula-based approach. Special attention is given to temporally dependent variables and to nonconcurrent compound events, defined as excess of variables over a threshold at a relatively close time. Two compound flooding events are then analyzed with the novel approach: one consists in simultaneous floods across two large watersheds in France and the other one consists in a devastating flood triggered by extreme precipitation over a saturated soil. Second, the outputs of climate simulations are used to project the evolution of these two compound flooding events in a climate change context. Climate simulations are often statistically biased, hence 2 multivariate bias correction methods (dOTC and R²D²) and one univariate bias correction method (CDF-t) are applied to raw simulations from 10 global climate models. Hence, bias correction methods are first evaluated and compared both in terms of return period of compound events and in terms of extremal dependence structure over the period 1992-2021. It is shown that the representations of the two compound events in raw climate simulations are biased and that bias correction methods improve their representations. Then, these methods are calibrated on reference data over the period 1992-2021 and applied over future periods in order to assess whether they can preserve the climate change signal of the two compound events. Finally, some extensions of the presented framework of analysis are discussed.