Clustering of grapevine (<i>Vitis vinifera</i> L.) genotypes with Kohonen neural networks

Abstract

Self-organizing maps generated by Kohonen neural networks provide a method to transform multidimensional problems common in ampelography into lower dimensional problems. In this study the clustering efficiency of Kohonen neural networks was evaluated to characterize and identify 10 Sangiovese-related and 10 "coloured" (fruit gives intense red colour to the wine) grapevine accessions, on the basis of the elliptic Fourier coefficients of the leaves. The non-supervised learning algorithm used allowed apriori classification of the accessions. The results enabled us to distinguish between 16 accessions and to denote two pairs of synonyms, To obtain quantitative information regarding relationships among these accessions, Kohonen neural networks were trained with different numbers of neurons in the Kohonen output layer permitting the graphical representation of the similarity by construction of a dendrogram. In agreement with previous studies based on molecular markers and neural network technology, a high similarity was found for the ecotypes (1) Prugnolo acerbo, Prugnolo dolce and Prugnolo medio and (2) Brunelletto and Prugnolo gentile. Among the Sangiovese-related accessions the so-called Casentino ecotype diverged from all the others, probably indicating a different origin. Producing easily comprehensible low-dimensional maps, the Kohonen neural networks approach proposed here allows to study complex ampelographic data elucidating relationships that can not be detected by traditional data analysis tools.