Multivariate authentication of the geographical origin of wines: a kernel SVM approach

During the WineDB European project, wine samples from four countries and three different vintages have been collected and their chemical content for 63 parameters was analyzed. The possibility to determine the country of origin of wines based on their chemical content was investigated during the project and the results from two multivariate classification techniques, namely Partial Least Squares Discriminant Analysis and kernel Support Vector Machines, are described and compared. Attention has been paid to the development of efficient models in terms of cost of analysis and the problem of variable selection is considered. In particular, the kernel SVM approach leads to models which can reduce the annual updating effort of the classification models since a unique set of parameters can be used to discriminate authentic wines from different countries and different years of production, which is not the case when only PLS-DA is applied.