Using climatic and imaging data to predict apple phenology

This thesis investigates the biological and environmental factors influencing apple maturity, with the aim of developing a non-destructive framework for accurate harvest prediction. It begins by reviewing phenology models commonly used to estimate flowering and harvest timing and introduces imaging as a promising alternative to directly assess apple maturity. Chapter 3 focuses on the parameterisation of multiple cultivars to determine the level of specificity during model fitting. This has not previously been done before, mainly due to the lack of training data. In this chapter, the PhenoFlex model is applied to a unique 85-year dataset of 26 apple cultivars grown at East Malling. The study demonstrates that generic (species-level) or grouped (based on flowering time similarity) parameterisation approaches can effectively predict flowering time, reducing the need for cultivar-specific calibration, especially with limited data. Chapter 4 explores the impact of flowering time variation on fruit maturity. The temperature experienced by individual flower clusters was tracked until their corresponding fruit was harvested. The Growing Degree Hours model was applied to each flowering to harvest period to determine the influence of temperature on growth. The variation in maturity was assessed considering flowering time, season, tree and canopy region, and it was found that flowering time accounts for up to 20% of maturity variation, depending on the cultivar. Smaller effects were observed from seasonal and tree-level effects. These findings highlight the limitations of using average flowering dates in harvest models and support the need for more precise, fruit-level assessments starting from the earliest possible harvest date. Chapter 5 evaluates hyperspectral imaging as a non-destructive method for assessing apple maturity. A large and diverse dataset of 5,756 apples was collected from different cultivars, seasons and countries. This dataset is approximately 1800% larger than the datasets used in previous studies studying apple traits. The Vision Transformer architecture achieved the highest accuracy in predicting Brix and firmness. Key spectral and spatial features were identified, and cultivar-specific information improved model performance. Imaging a single side of the fruit was insufficient; the model that was trained on images from all four sides yielded better results. The thesis concludes by discussing the results found in each chapter and proposes a way to integrate phenology and imaging into a framework for harvest window predictions.