Classificação de vinhos produzidos no Vale do São Francisco baseada em imagens digitais e quimiometria

Abstract

This work proposes a simple, fast and low-cost methodology based on digital images and pattern recognition techniques for the classification of wines produced in the São Francisco Valley (VSF), a wine region located in the Brazilian Northeast, contributing to the certification process of the wines geographical indication. Red wines samples, purchased from Natal/RN and João Pessoa/PB supermarkets or donated by winegrowers and research institutions located in the Region of the São Francisco Valley were used. Three classification approaches were considered, taking as reference the geographical origin (with two classes: VSF and Word), winegrowers (with three VSF winegrowers) and grape varieties (Cabernet Sauvignon, Syrah and Touriga Nacional). In the classification approach according to the geographical origin of the wines, a hundred samples were used, and the models presented Correct Classification Rate (CCR) of the test set of 80.7% and 93.6%, respectively, with the RGB and HSI data and modeling with SPA-LDA and 61.3% and 83.9%, respectively, with the RGB and HSI data and PLS-DA modeling. In the classification approach as a function of winegrowers, seventy samples were used and the CCR values of the test sets were equal, with RGB and HSI data, 100% and 95.5%, respectively, for the models constructed with SPA-LDA and PLS-DA. In the classification approach using the varieties of grapes present in the composition of varietal wines, forty-eight samples were used, and the models constructed with SPA-LDA and PLS-DA obtained CCR values equal to 100% in the test sets for the Cabernet Sauvignon versus Touriga Nacional and Syrah versus Touriga Nacional, both with the RGB and HSI data. In the comparison between Cabernet Sauvignon versus Syrah, the SPA-LDA models achieved equal results with the RGB and HSI data, with CCR equal to 72.7%, for the test sets, while the models obtained with PLS-DA achieved equal CCR of 72.7% and 81.8% in the test sets, respectively, with the RGB and HSI data. The results demonstrate a feasibility of using digital images associated to chemometrics for wine classification, in a simple, fast, low sample consumption, without chemical reagents or dilution of the samples and with low generation of residues, according to their origin geographical, winegrower or varietal composition.