Chemical analysis in conjunction with multivariate data evaluation methods was used to study elemental profiles and geographical origin of wines from central Balkan countries (Serbia, Montenegro, and Macedonia). Nine elements (Na, K, Mg, Ca, Fe, Mn, Zn, Cu, and Pb) chosen as chemical descriptors were analyzed in 41 commercial wine samples. Unsupervised pattern recognition methods—principal component analysis (PCA) and factor analysis—identified the main factors controlling the data variability, while the application of hierarchical cluster analysis (HCA) highlighted a differentiation between sample groups belonging to different variable inputs. Three PCs were shown to be the most significant, together accounting for 70.8% of the total variance. Supervised pattern recognition methods—linear discriminant analysis (LDA), k-nearest neighbor (kNN), soft independent modeling of class analogy (SIMCA), and artificial neural network (ANN)—applied to the classification of wine samples demonstrated different recognition and prediction abilities. The recognition rate for LDA was 97.6%, and the percentage of classification obtained by kNN, SIMCA, and ANN was 100%. However, the LDA method produced the best prediction rate of 83.3%, whereas kNN, SIMCA, and ANN gave much lower percentages of correctly classified samples, at 72.2, 61.1, and 55.6%, respectively. Trace elements seem to be suitable descriptors for wine samples studied by classification methods, since their concentrations comprising both natural and other sources of influence are attributed to grapegrowing and winemaking sites. Comparison of pattern recognition methods reveals the difference in their classification power.
- Copyright © 2010 by the American Society for Enology and Viticulture