Abstract

In several applications, the distribution of the extremes of one or more environmental variables across multiple spatial sites is of greatest concern: extremely high river flow levels cause flooding; very high air pollution levels or temperatures lead to higher mortality. Extreme value analysis provides a range of methods for estimating the joint tail distribution in such cases. However, most existing approaches are limited to scenarios with fairly moderate dimensions, in part due to a limited amount of data.
 
In recent years, one emerging area of research concerns the application of machine learning and dimension reduction techniques, in combination with existing extreme value methods, to analyse extreme events across large scales and/or in high dimensions. This talk outlines and illustrates two such approaches and discusses ongoing research. The first approach employs Bayesian clustering methods to estimate the tail behaviour of an environmental variable across several sites. Sites within an estimated cluster tend to have a similar tail behaviour and to frequently record extreme events in the same week. The second approach considers the application of ideas from principal component analysis to investigate the joint tail behaviour. If time permits, I will outline how these ideas can be used to generate hazard event sets (synthetic sets of extreme events) which are an important planning tool for risk analysis.

Dr Christian Rohrbeck, Lecturer in Statistics, University of Bath.

Dr Daniela Castro-Camilo from the Environmental Section and Glasgow local group.