Machine learning methods for the analysis of precipitation patterns
: (Alternative Format Thesis)

  • Andrew Barnes

Student thesis: Doctoral ThesisPhD


Extreme rainfall is becoming more common and is resulting in an increasing number of extreme flood events. Exploring why extreme rainfall events occur and how forecasts of such events can be improved is becoming key to improving society's preparedness and response to such events. This thesis evaluates and expands the capability of machine learning methods to aid both our understanding and forecasting of heavy rainfall events.

First, atmospheric trajectories were generated and clustered to identify the key moisture pathways relating to extreme rainfall events. This showed the applicability of new clustering methodologies to the identification of extreme rainfall circulation mechanisms. Following this, cluster analysis was then used to carry out an indepth analysis of the synoptic-scale meteorological conditions relating to extreme rainfall events in the UK. This resulted in the identification of a causal relationship between mean sea-level pressure and 2m air temperature patterns and extreme rainfall distributions. This highlights the importance of mean sea-level pressure anomaly polarisation across the North Atlantic, including strong and significant relationships between the synoptic patterns associated with extreme rainfall events and large-scale climatic indices such as the North Atlantic Oscillation and Atlantic Multidecedal Oscillation.

A neural network based sensitivity analysis technique was employed to identify which synoptic regions across the North Atlantic are important for understanding the difference between extreme and regular rainfall events in Great Britain. Following this a new model was developed using the gained knowledge to forecast regional, monthly rainfall values using forecasted synoptic meteorological images. This new image based forecasting model was shown to outperform the current state-of-the-art forecasting models; for example, at a one-month lead-time the image based model outperforms the ECMWF's mathematical model by 6mm. However, when comparing the models on the most extreme rainfall events the new image-based model produces errors 28mm lower than the ECMWF's model. Further to this, sensitivity analysis of the image-based model reveals a strong relationship between a low mean sea-level pressure anomaly followed by a high mean sea-level pressure anomaly in the North Atlantic results in greatly increased rainfall forecasts. A similar analysis of the termperature inputs highlights a key polarisation between warm and cool air to the south west of Great Britain and France can also lead to hightened rainfall.

Finally, a series of new forecasting techniques were developed which use a sequences of synoptic images in the form of a meteorological video to forecast regional, monthly rainfall. This video-based model when combined with the MetOffice's state of the art forecasting model can improve rainfall forecasts across Great Britain.

Throughout this thesis state-of-the-art machine learning techniques were expanded and evaluated in the interpretation and forecasting of extreme rainfall events. The results highlight the capability of machine learning methods to not only match but also improve our current understanding of rainfall based forecasting and analysis. This thesis also leads the way in opening new avenues of potential work including the investigation of the cost-benefit analysis of using these new, data intensive models. Further to this, new and exciting questions remain as to the sensitivity of this work to alternative meteorological variables such as integrated water vapour and wind vectors
Date of Award23 Mar 2022
Original languageEnglish
Awarding Institution
  • University of Bath
SupervisorThomas Kjeldsen (Supervisor), Ilaria Prosdocimi (Supervisor) & Nick McCullen (Supervisor)

Cite this