PhD Project: Structured Multivariate and Spatial Extreme Value Models for Environmental Science
Supervisors: Prof. Jonathan A. Tawn (Lancaster University), Dr. Ingrid Hobæk Haff (University of Oslo) and Simon J. Brown (Met Office)
Extremes in environmental science in processes such as precipitation, wind temperatures have often devastating consequences, i.e. floods, drought or wildfires. Developing a spatio-temporal model of extremes of these processes would help us understand their potential future impact. The impact questions of most interest are the so-called severity-area-duration-frequency questions and the current models only answer these questions partially. For example, current temporal models do not tell us about the area affected by a potential future extreme event hence its severity is also limited to an observation at a given point of interest. On the other hand, the spatial models suggest which areas are most at risk but not the temporal extent of a single event. A combined spatio-temporal model is still novel in the extreme value research and current methods only work in low number of dimensions.
There are several possible research directions to tackle this problem of high dimension to simplify the modelling using graphical models, clustering or vine copulas. One example is to use the known underlying structure by using graphical models and identifying pairs of variables that are conditionally independent given all other variables. An example of this would be an assumption of precipitation in two sites being independent given all other sites.
As a Data Scientist at the Leeds Institute for Data Analytics (LIDA) (2021-2022), I worked in geographic information science (GIScience):
- Urban Transport Modelling for Sustainable Well-Being in Hanoi (UTM-Hanoi) with Prof Nick Malleson and Prof Lex Comber. This included creating an R Shiny dashboard as a communication tool of the results of the further study. The code of an example dashboard is available on GitHub.
- Enterprise Car Club hire records with Dr Ian Philips and Prof Greg Marsden. This included Heckman selection models to determine factors associated with car club usage. The project findings are helpful in the planning of future lot locations.