Marcel Hensel,* Jochen Vestner, and Dominik Durner
*Dienstleistungszentrum ländlicher Raum, Breitenweg 71, Neustadt a.d.W./ Rhineland Palatinate/ 67435, Germany (Marcel.Hensel@dlr.rlp.de)

Wine color plays a major role for wine quality and consumer acceptance. CIEL*a*b* data is used in wine research for descriptive purposes. A new method is proposed here on how to extract more information from CIEL*a*b* data. Recent development and accessibility of multivariable statistics and machine learning (ML) led to the exploitation of these methods in various disciplines. ML in wine research is mostly used to predict intangible parameters such as wine quality or price based on complex, time-consuming chemical analysis. In this study, wine color measurement, which can be obtained quickly and easily, is used as an example on how to make ML applicable for winemakers. Blanc de noir wines are white wines from red grapes and are produced worldwide. The light color of Blanc de noir wines is very important for consumer acceptance. In this study, CIEL*a*b* coordinates from over 150 Blanc de noir, white, and Rosé wines were calculated from the recorded spectrometric data. A support-vector-machine (SVM) was trained and used to classify the wines. The SVM was trained with 70% of the data; the model was validated against 30% of the data. An algorithm optimization was performed by full factorial grid search. The algorithm predicted whether a given wine was a Blanc de noir wine or a white wine with up to 96% accuracy in the cross-validated training data set. Validation of the optimized model resulted in a classification accuracy of 95%. A method to differentiate Blanc de noir from Rosé was developed, defining the lower color boundary of Blanc de noir wines. The validated model was exploited in a user-friendly dashboard, making this approach applicable to winemakers.

Funding Support: This work was financially supported by the AiF, Arbeitsgemeinschaft industrieller Forschungsvereinigungen “Otto von Guericke” e.V. (Project Number 20964)