Spatial Interpolation of Extreme Values

TL;DR

This paper presents a novel spatial interpolation method for extreme environmental values, combining limited measurement data with climate model outputs, and employs a modified Monte Carlo EM algorithm to estimate the model parameters.

Contribution

It introduces a new approach for interpolating extreme values using Gaussian process-modulated GEV parameters and a modified EM algorithm for parameter estimation.

Findings

Model accurately estimates extremal features of rainfall.

Interpolated maps effectively visualize extreme rainfall patterns.

Method integrates measurement data with climate model outputs.

Abstract

This paper introduces a method for spatial interpolation of extreme values, and in particular targets the case in which conventional data, resulting from a measurement for example, are available at only a few locations. To overcome this the conventional data are supplemented with output from a computer simulator. For environmental applications, such as extreme rainfall as we study here, the simulator could be a regional climate model. Annual maxima are studied and assumed to follow the generalised extreme value (GEV) distribution and dependence is accommodated between maxima by its parameters following Gaussian processes. The GEV's parameters are now random and so a modification to the Monte Carlo EM algorithm is presented so that the model's parameters can be estimated. Then a variety of checks for the model's goodness of fit are given. For the extreme rainfall application we find that the model estimates a variety of extremal features of interest well, and then show how specific features can be interpolated based on the model so that maps can be produced.