Samenvatting

By addressing the sub-research questions, the main research question and data mining goal were answered. Drought tolerance can be quantified using yield loss between control and drought treatments and robustly assessed by ranking varieties within each year. Varieties that consistently rank high or low across years can be identified as tolerant or sensitive. The most informative field trial data consisted of a subset of 12 drone-derived vegetation indices and spectral signatures measured throughout the growing season. Using these features, a Support Vector Machine achieved a mean absolute error (MAE) of 1.31, a Pearson’s r of 0.89, and an R² of 0.80 when predicting total yield, thereby exceeding the project’s success criteria. However, Meijer’s broader business objective—to predict abiotic stress early, or using DNA data alone, in order to reduce the time and resources required for field trials and the breeding process—was not achieved. The best-performing model relied on drone data from the full growing season, meaning predictions would only be available shortly before harvest, after most resources had already been invested. To meaningfully support Meijer’s objectives, predictions would need to be made using genetic data alone. This was explored by characterizing varietal yield variation using SHAP values; however, attempts to link these signals to specific genomic positions using a LASSO regression model were unsuccessful. Despite this limitation, the project provides a strong foundation for future work. It clarifies how drought tolerance can be quantified, identifies which data sources are most informative, and highlights the challenges of separating genetic and environmental effects. These findings form a clear starting point for further research, as discussed in the following chapter.