Hyperspectral imaging and machine learning for monitoring grapevine physiology

“Nebbiolo” is a red winegrape variety well known to produce monovarietal wines in Piemonte, Valle d’Aosta, and Lombardia regions, taking part to 7 DOCG (Denominazione di Origine Controllata e Garantita) and 22 DOC (Denominazione di Origine Controllata) protected designations of origin (PDO) [1,2].

The crunchy texture of table grapes is one of the key quality parameters during production. This varies from cultivar to cultivar, stage of harvest and vineyard performance. Cell wall properties are key drivers of berry quality (e.g., pericarp firmness and intactness) at harvest and beyond. Common practise amongst producers is to continuously monitor firmness by evaluating pericarp appearance of cross-sectioned berries prior to harvest. These qualitative methods can be quite arbitrary and imprecise in their execution, but more quantitative, yet simple and high-throughput methods to evaluate these cell wall polymers are not yet readily available.

Syrah is a grape with weak aromatic expression. This atypical grape variety as a fruit allows the production of wines of great reputation for which the aromatic particularity plays an important role. The varietal aroma consists of volatile substances directly perceptible by the olfactory mucosa and of aroma precursors, of which the glycosides constitute an important class.

The need to rationalize agricultural inputs has recently increased interest in assessing vineyard variability in order to implement variable rate input applications, so-called ‘precision viticulture’. In many viticultural areas globally, precision viticulture is already widely used such as for selective harvesting and variable rate application (VRA) of inputs such as irrigation and/or fertilizer. Robust VRA relies on having a geostatistically accurate map (of one or more vineyard attributes) requiring high sampling densities, which can be cost- and time-prohibitive to obtain. Previous work on spatial interpolation using kriging have upscaled ground-based measurements, but such upscaling strategies are applicable only when vineyard conditions are spatially continuous and satisfies the assumption of second-order stationary processes. Alternatively, mixed models that combine kriging and auxiliary information, such as the regression kriging (RK) method, are more instructive for spatial predictions. In order to improve prediction accuracies, it is therefore necessary to incorporate additional information to achieve accurate spatial patterns with low error.

Analytical methods for dating wines often rely on assessing anthropogenic and cosmogenic radionuclides, including 14C and 137Cs [1,2].