Machine Learning for Predicting Grape Quality Using Spectral Imaging Techniques

The production of high-quality wine is strongly impacted by the quality of grape clusters, especially their sugar level, which is a key factor in deciding the best time to harvest. Destructive sample techniques that are labor-intensive, timeconsuming, and not scalable are traditionally used to evaluate grape maturity. This thesis addresses these limitations by investigating the non-destructive prediction of grape sugar concentration utilizing machine learning algorithms, such as regressions or neural networks, in conjunction with modern imaging technologies, with a primary focus on hyperspectral imaging (HSI). The study starts with a thorough overview of the principles of viticulture, covering grape biology, development phases, and the importance of sugar buildup during maturation. After that, it discusses the fundamentals of imaging techniques, including RGB, multispectral, and hyperspectral imaging, with a focus on how these technologies record and communicate important information about the product. This fundamental knowledge lays the groundwork for talking about how several imaging modalities, each with a different spectral resolution, can be used to forecast crucial quality metrics like the Brix Index, which measures the amount of sugar in grapes. The main focus of the thesis is the use of hyperspectral imaging in conjunction with machine learning models, like Partial Least Squares Regression (PLSR), to forecast grape sugar concentration. Preparing imaging data for predictive modeling required a large portion of the study in order to improve the signal-to-noise ratio and lower dimensionality. The results analysis is enhanced by philosophical thoughts on the models’ underlying assumptions and their suitability for various viticultural situations. An additional exploration was planned, potentially using RGB datasets to test whether simpler, more accessible imaging methods could provide comparable results to the more sophisticated hyperspectral approach. The results of this thesis demonstrate the potential of imaging-based techniques for viticulture’s non-destructive quality evaluation. This work lays the groundwork for future precision agriculture research targeted at enhancing the effectiveness and precision of grape quality monitoring by critically analyzing the methods employed and considering the types of data collected by various imaging technologies.