PLS-DA-Based Classification of Red Wine Vintages via Digital Imaging and Chemical Properties

Classification of wine vintages plays a key role in verifying authenticity and product quality. Although advanced tools like FTIR spectroscopy combined with chemometrics are widely used, they require high-cost equipment and technical expertise. Digital image analysis offers a rapid, low-cost, and non-destructive alternative.

A total of 13 red wine samples from two vintages were analyzed. Total monomeric anthocyanin (TMA) content was determined using UV–Vis spectrophotometry, and total soluble solid (TSS) and alcohol levels were measured using a digital refractometer. Controlled lighting conditions were achieved using a custom-designed imaging cabinet, and CIE Lab color features were extracted from images using Python OpenCV. Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA) models were constructed in R with a nested cross-validation design (LOO outer loop with inner 5-fold hyperparameter tuning) to mitigate overfitting. Model significance was confirmed by permutation testing, and key variables were analyzed with Variable Importance in Projection (VIP) scores across folds.

The final model achieved 92% classification accuracy and a Q² of 0.56. The classification error corresponded to a single misclassified sample. Given the small sample size, this single error had a disproportionate impact on overall accuracy. Expanding the dataset would stabilize accuracy metrics significantly. The significance of the model was confirmed through permutation testing (100 iterations), yielding p < 0.05, indicating that the observed class discrimination is unlikely due to random chance. Two clusters were observed on the S-plot: one being a strong negative correlation with L_avg (older) and other being a strong positive correlation with TMA (younger). Hyperparameter optimization selected zero orthogonal components across all CV loops, resulting in a pure PLS-DA structure. Key variables contributing to class separation included L_avg and TMA (VIP > 1).

These findings demonstrate that digital imaging combined with chemometric modeling can effectively discriminate vintages, which offers a practical, non-invasive tool for rapid vintage authentication and quality monitoring.