Hamer, Charlie Rhys (2025) A Machine Learning Approach to Rapeseed Yield Prediction and Shapley Value Analysis. Masters thesis, York St John University.
Preview |
Text (Masters by Research thesis)
A Machine Learning Approach to Rapeseed Yield Prediction and Shapley Value Analysis.pdf - Published Version Available under License Creative Commons Attribution. | Preview |
Abstract
Crop harvest prediction depends on many factors, including soil, irrigation, weather, crop genotype, and how all of it is maintained. This study aims to analyse the performance of a pre-existing formula for yield prediction, and build a new machine learning based model to compare it to. This study will use a dataset of rapeseed harvests over three years around the areas of South Wales, Derbyshire, and Lincolnshire to train the model. This dataset will have a dimensionality reduction performed on it, and the effects of that will be evaluated. This study will test the dataset on an independent Nottinghamshire-based dataset for generalisation purposes.
This study proposes a 1-dimensional Convolutional Neural Network, and it will evaluate its performance using the RMSLE prediction matrix. Finally, this study will use Shapley values to analyse the predictions of the neural network further, and use them to suggest improvements to farmers and data collection researchers. These Shapley values will show which features in the dataset provide the most influence over the final output prediction by the AI model. This research will also discover the importance of the vegetation matrix in such calculations.
The proposed CNN outperformed the original formula, broadcasting results of 96% prediction accuracy under ideal dataset conditions. However, this research observed failures to generalise to new areas with the testing of the Nottinghamshire dataset, and how removing the vegetation index variable can reduce the accuracy measures by almost 10%. Using Shapley values, the vegetation index, water stress and nitrate density variables were determined to generally be the most important, but some nuance exists based on number of features in the tested dataset and the omission of the vegetation index.
| Item Type: | Thesis (Masters) |
|---|---|
| Status: | Published |
| Subjects: | T Technology > T Technology (General) |
| School/Department: | York Business School |
| URI: | https://ray.yorksj.ac.uk/id/eprint/15411 |
University Staff: Request a correction | RaY Editors: Update this record
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)
