Aims: The aims of this study were to i) characterize the impact of geographical origin and viticulture treatments on Chilean Cabernet-Sauvignon, and ii) develop machine learning models to predict its quality.
Methods and Results: 100 vineyard plots representing the typical percentage distribution of geographical and viticulture impact factors on Chilean Cabernet-Sauvignon were monitored across two seasons, 2018 and 2019. Chemical analysis of grapes and wines included the quantification of phenolic compounds by liquid chromatography and UV-vis spectral measurements, aroma compounds by gas chromatography mass spectrometry (GC/MS), and maturity parameters. Spearman correlation and Principal component analysis (PCA) identified correlations of several non-volatile and volatile compounds with quality, mainly by means of their anthocyanins, flavonols, flavan‑3‑ols, total tannins and hydroxycinnamic acids. Furthermore by trans-2-hexenol, trans-3-hexenol, hexanal, 2-isobutyl-3-methoxypyrazine (IBMP), yeast assimilable nitrogen (YAN), total soluble solids and acidity. Experimental winemaking of 600 kg per plot followed a standardized procedure, and the wines were analyzed by an expert quality rating. A sensory quality profiling for the wines was performed through a Napping Ultra Flash Profile (UFP). It revealed the distinction of three different quality levels by mainly mouthfeel attributes, and fruity and green aromas. However, neither the observed correlations of chemical analysis and sensory quality ratings, nor origin or viticulture treatment could fully explain quality. Different clustering methods, namely k-means, k-medioids and spectral clustering were evaluated in order to find categories given by the chemical analysis data itself as unsupervised machine learning. Spectral clustering led to optimum results, and independently of sample origin and viticulture traits, quality ratings were characterized to be significantly different across the clusters allowing their interpretation as quality categories.
Conclusions: Chilean Cabernet-Sauvignon quality is associated with chemical quality markers known for this variety in Australia and California, including phenolic compounds, C6 alcohols and aldehydes, IBMP, maturity parameters and YAN. However, evaluation of sensory quality is fairly subjective and viticulture treatments in practical application contain interdependency, therefore it is challenging to establish supervised models involving this data. The application of unsupervised spectral clustering is proposed as an objective quality classification approach, which can be trained using supervised models for predictive purposes.
Significance and Impact of the Study: There is a high industrial need for objective quality classification. For the first time chemical quality markers for Chilean Cabernet-Sauvignon were determined, and an unsupervised machine learning approach based on these markers could be proposed for objective quality classification.
Authors: Doreen Schober1*, Martin Legues1,2, Hugo Guidez3, Jose Carlos Caris Maldonado1, Sebastian Vargas1, Alvaro Gonzalez Rojas1
1Center for Research and Innovation (CRI), Viña Concha y Toro, Ruta k-650 km 10, Pencahue, Región de Maule, Chile
2Pontificia Universidad Católica de Chile, Vicuña Mackenna 4860, Macul, Región Metropolitana, Santiago, Chile
3Institut National Supérieur des Sciences Agronomiques, Agroalimentaires, Horticoles et du Paysage, Agrocampus Ouest Campus d´Angers, France
Email: doreen.schober@conchaytoro.cl
Keywords: Cabernet-Sauvignon, spectral clustering, quality, terroir, vineyard management