Verifying Wine Origin: A Neural Network Approach

¹ Departamento de Química, Faculdade de Ciências e Tecnologia, Universidade Nova de Lisboa, 2825 Monte da Caparica, Portugal.

jas{at}mail.fct.unl.pt

Neural networks are powerful computational tools that "learn" with training examples and have the capability for extrapolating their "knowledge" to new situations. Artificial back-propagation neural networks (BPNNs) were applied to two different problems of wine classification. In one case, each of five 16 x 2 x 2 BPNNs was trained with a different set of 21 samples to distinguish between young red wines of two Spanish Certified Brands of Origin (Ribera de Duero and Toro) on the basis of 15 anthocyanin contents. The networks were tested with 26 examples not involved in the training process and gave average correct predictions of 88% for Toro wines and 91% for Ribera wines. In another case, five 23 x 8 x 8 BPNNs were applied to classify eight Portuguese varietal wines of Vitis vinifera according to grape variety (Roupeiro, Manteúdo, Tamarez, Rabo de Ovelha, Moreto, Trincadeira, Periquita, and Aragonez). Percent composition of 22 free amino acids were available for 42 samples (of 7 different vintages) and were used as input data. Each network was trained with a different set of 26 examples and tested with the other 16, obtaining an average success rate of 73%. Two-layer BPNNs could be used to deduce which amino acids are characteristic of each variety of wine. In the two problems studied, the neural networks gave better predictions than linear discriminating analysis (LDA).

Key words: amino acids, anthocyanins, Certified Brand of Origin, chemistry, classification, neural networks, origin, wines

Submitted on September 26, 1995