Gouldsbrough, Lily and Eastoe, Emma and Hossaini, Ryan and Young, Paul (2023) Statistical and machine learning modelling of UK surface ozone. PhD thesis, Lancaster University.
![[thumbnail of 2023lily_gouldsbroughPhD]](https://eprints.lancs.ac.uk/style/images/fileicons/text.png)
lily_gouldsbrough_phd_thesis.pdf - Published Version
Download (11MB)
Abstract
In addition to atmospheric observations, numerical models are crucial to understand the impacts of human activities on the environment, from attributing poor air quality to assessing climate change impacts. While process-based models, such as chemistry transport models (CTMs), are widely used, recent data science advances enable greater use of statistical and machine learning methods as alternatives to describe and predict atmospheric composition. State-of-the-art data science methods can be faster to run than CTMs and used at high temporal and spatial resolutions due to codebase efficiencies. This thesis focuses on modelling UK surface ozone and its drivers (high levels of which are detrimental to human and plant health) through the development and novel application of sophisticated statistical and machine learning techniques. Motivated by possible adverse effect of climate change on ozone concentrations, a temperature-dependent Extreme Value Analysis is used to explore the probability, magnitude, and frequency of extreme ozone events over recent decades. For 2010–2019, it is found that the 1-year return level of daily maximum 8-h mean (MDA8) ozone exceeds the ‘moderate’ health threshold (100 µg/m3) at >90% of sites, but that the probability of extreme ozone events has markedly decreased since the 1980s. A machine learning methodology to downscale and bias correct a CTM (EMEP4UK) ozone surface was developed and evaluated. Compared to the unadjusted CTM, the downscaled surface exhibits a lower bias in reproducing MDA8 ozone allowing more robust assessments of important policy metrics. Analysis of the downscaled product (2014–2018) reveals on average 27% of the UK fails the government long-term objective for MDA8 ozone to not exceed 100 µg/m3 more than 10 times per year, compared to 99% in the unadjusted CTM. A classification-based machine learning analysis into high-level ozone drivers was also performed and shows a robust relationship between ozone and temperature. The method is demonstrated to offer remarkable promise as a tool with which to forecast the presence of high-level ozone. Despite a UK focus, the data-driven methods developed and applied here are applicable to modelling ozone in other regions of the world where measurements exist.