Photometric wine color measurement

Abstract

The color of wine is an important quality parameter essential for the first impression of consumers. The Organisation inernationale de la Vigne et du Vin (OIV) recommends two methods to describe wine color: color calculation according to Glories and the determination of coordinates in the L* a* b* color space of the Commission Internationale de l'Eclairage (CIE). In this work, the influence of photometer settings on the calculation of the CIE L*a*b* color space were determined. It was shown that the photometer settings influence the reproducibility of the measurement. Furthermore, the color measurement according to Glories is compared to the CIE L*a*b* color space. The results show a weak correlation in the light red wine and white wine color area. Therefore, Glories' color measurement and the CIE L*a*b* color cannot be used interchangeably. To determine, which of the methods is more suited for further investigation, the color measurement according to Glories and the CIE L*a*b* color space were compared to the visual perception of 112 red and white wines. The results indicate that the CIE L*a*b color space is better suited to depicting the color perceived by humans. Since its development, the CIE color spaces have undergone various changes. The possibility of comparing colors has been no exception. The Euclidean color difference is the formula currently recommended by the OIV to compare wine colors. However, the CIE recommends the CIEDE2000 color distance formula, which has been proven to be more precise. The reason why the Euclidean color difference is still used in wine research is the absence of reference values calculated with the CIEDE2000 color distance formula for the just noticeable difference (JND), or the visual color threshold, the minimum difference in color hue that is visible by the human eye. Therefore, the JND was re-evaluated with the CIEDE2000 color distance formula via triangle testing. Compared to Glories' color measurement, CIE L*a*b* more closely match the human perception, elevating the use of CIE L*a*b* over the use of the Glories method. Visual color thresholds were better expressed with CIEDE2000 but still varied depending upon the color area in the CIE L*a*b* color space. The results of these studies indicate that the CIE L*a*b* color space is better suited for further investigation. Machine learning (ML) and statistical modeling have emerged as important innovations in science. In wine research, ML is often used to predict abstract parameters such as wine quality based on complex instrumental chemical analysis. The presented study used spectrophotometric data and CIE L*a*b* coordinates from 176 commercial wines to distinguish Blanc de noir from rosé wine and white wine. The transmission spectra were used to train extreme gradient-boosted trees (XGBoost) and a support vector machine (SVM). CIE L*a*b* coordinates were used to train SVM and logistic regression. After parameter hypertuning, the combination of SVM on CIE L*a*b* data provided the optimal classification with a cross-validated accuracy of 0.88 and a F1 score of 0.93. The final classification model is deployed in a browser-based, user-friendly dashboard for winemakers and other users, such as wine laboratories. SVM was also applied in the context of classification of barrel-aged red wine. The transmission spectra of 363 red wines were measured and transformed into absorption spectra and CIE L*a*b* coordinates. Transmission spectra, absorption spectra, and CIE L*a*b* coordinates were used to train an SVM. Furthermore, the absorption spectra were used to train a multilayer perceptron model. The spectra were preprocessed and transformed with principal component analysis (PCA) to reduce dimensionality. The performance of SVM on transmission spectra was outperformed by SVM on absorption spectra and CIE L*a*b* coordinates. The best performance was achieved by the neural network/MLP, with an F1 score of 0.75.