Using machine learning and process-based crop modelling for regional scale prediction

The aims of this thesis are to assess the effectiveness of machine learning techniques in comparison to process-based crop growth models for the purpose of prediction of the impact of climate variability on crops at the regional scale. Comparisons are made between popular and most effective machine learning methods to predict crop yields and the process-based crop growth model GLAM. Firstly, it is asked how much data is required for machine learning to outperform process-based crop modelling, and under which conditions? Secondly, the prediction performance of both methods for prediction of crop failures is compared as well as the effect of potential errors in climate data. Thirdly, machine learning and crop modelling are compared to bench-mark crop model sensitivity to climatic drivers of crop yield, hence providing a data driven approach to learn how to further improve crop model simulations. Results show that machine learning and process-based crop modelling have contrasting strengths and weaknesses. However, machine learning can be leveraged to improve process-based crop modelling through increased sensitivity to climatic drivers of crop yields. Furthermore, the effects of potential errors in data upon machine learning simulations is determined. In doing so it is shown that sensitivity of machine learning to climatological errors varies depending on model, and region, with different time-scales of effects depending on if errors are in temperature or rainfall. Overall, this thesis shows that machine learning can provide great benefit to regional scale crop yield prediction. However, due to disadvantages of reduced model interpretability and difficulty in predicting effects of extreme events, machine learning is not a perfect solution for regional scale crop yield prediction. Therefore, it is argued that machine learning should be used in cooporation with process-based crop modelling to improve understanding rather than replace existing methods or knowledge.