Multi-Vintage Physicochemical Differentiation and Machine-Learning Classification of North Macedonian Wine Varieties

Source abstract

Routine physicochemical measurements may support wine varietal discrimination and quality control. This study assessed 614 red and 410 white wine samples from North Macedonia collected between 2008 and 2023 and analysed using standard OIV methods. Varietal differences were evaluated using one-way analysis of variance and Tukey’s post hoc test. Principal component analysis (PCA) identified major sources of variation, and five supervised classifiers were compared using the original and PCA-transformed feature spaces. Significant varietal differences were observed in alcohol content, specific weight, total extract, sugar, sugar-free extract and total acidity. Three principal components explained approximately 71% of the variance in red wines, while four explained approximately 81% in white wines. After hyperparameter optimisation, Random Forest achieved the highest accuracy using the original variables: 73% for red wines and 83% for white wines. The best accuracies after PCA transformation were 67% and 59%, respectively. These findings indicate that routine physicochemical parameters contain useful varietal information, but PCA may discard lower-variance information relevant to classification. The original feature space was therefore preferable for predictive performance, while PCA remained useful for exploratory analysis and visualisation.