Modelling extremes over space can be quite cumbersome. This is because data are typically measured at a finite number of locations, but extrapolation is required at any other location of the study region. Moreover, since we want to focus on high (or low) quantiles, extrapolation is also required into the joint upper (or lower) tail of the distribution. Finally, a recurrent problem with large heterogeneous regions is the lack of spatial homogeneity.

The literature on spatial extremes is mostly divided into two mainstream approaches: the first approach defines extreme events as block (e.g., annual) maxima, which are typically modelled using max-stable processes. The second approach defines extreme events as high threshold exceedances, which are usually modelled using generalised Pareto processes. Both max-stable and generalised Pareto processes are asymptotic models, in the sense that they have a limiting characterisation for block maxima and threshold exceedances, respectively. However, their tail dependence structure is fairly rigid. Moreover, inference for these extreme-value models is known to be particularly hard.

In this project, we tackle these problems by using a divide-and-conquer strategy: we fit a set of locally stationary spatial models, which allows the tail dependence structure to flexibly change across space. Inference is possible in a reasonable time since the big-dimensionality problem is broken into a bunch of small-dimensionality problems.

To see more details on the modelling and inference as well as a cool application to precipitation data, visit this link.

First published: 25 October 2022

<< Completed projects