Regionalization methods for compound events based on Wasserstein distance

Understanding compound events involves analyzing the interactions between different climate variables, assessing their probability of co-occurrence, and evaluating their cumulative impacts. 

This field of study has gained attention in recent years due to the increasing frequency and severity of extreme weather events, which are often linked to climate change.

Regionalization, in the study of compound events, refers to the process of tailoring analyses and models to specific geographic regions. This approach is vital because the characteristics and impacts of compound events can vary significantly across different areas due to variations in climate, geography, socio-economic conditions, and infrastructure resilience. Regionalization methods seek to identify sub-regions that display similar patterns in the variables of interest.

The objective of this talk is to offer a regionalization of intricate spatial climatological datasets, particularly when considering compound extremes.

To this end, a clustering algorithm is introduced to group time series of maxima for paired random variables observed at different stations. The approach requires different types of dissimilarity measures. In particular, it relies on the copula approach and on the use of the related Kendall distributions that are compared with the Wasserstein distance.

As an illustration, using data on daily maximum temperature (in Celsius) and daily maximum evapotranspiration (mm/day) from the ERA5 dataset, collected across various municipalities throughout Italy, enhanced estimation of climate-related metrics at specific locations are obtained by leveraging regions with statistically similar characteristics.