Classificação de genótipos de café arábica usando espectroscopia de infravermelho próximo

• Ano de defesa: 2014
• Autor(a) principal: Marquetti, Izabele
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Campo Mourao; Medianeira; Programa de Pós-Graduação em Tecnologia de Alimentos
• Palavras-chave em Português: Café - Cultivo; Interação genótipo-ambiente; Café; Coffee plantations; Genotype-environment interaction; Coffee
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/1133

The environmental conditions in coffee cultivation, such as climate, soil type and altitude, associated with agronomic practices, are responsible for influence the final chemical composition of the bean. They directly influence the essential features of the beverage, increasing its aggregate price. Proof of geographic and genotypic origin of the coffee genotypes must be done using reliable methods. Thus, the near infrared spectroscopy (NIRS), in the 1100 to 2498nm range, was used for analyze different coffee genotypes that were cultivated in different cities (Brazil - Paraná State). As first approach linear methods, principal components analysis (PCA) and partial least squares with discriminant analysis (PLS-DA), were used for data interpretation due to the high complexity and amount of information contained in the spectra. The obtained PLS-DA models had an average sensitivity of 93.75% and a specificity of 100% for the geographical classification. While for genopyte classification, the PLS-DA performance was 93.75% for sensitivity and 97.13% for specificity. In an attempt to improve the performance and reliability of the developed classifiers, both the PCA scores and the PLS-DA latent variables were fed into two artificial neural networks, the multilayer perceptron (MLP) and radial basis function network (RBF), that are nonlinear models. The architecture parameters of these networks were optimized using the sequential simplex method. The two-stage models, linear with PLS-DA and nonlinear with RBF, were able to classify geographically and genotypically with 100% of selectivity and specificity all the training and test samples. The latent variables of the PLS-DA are determined by taking into account the desired response, so it contains more information than the scores of the PCA. While the RBF network, by having fewer free parameters and a simpler architecture compared to the MLP, has a faster and covergente training. The spectra analysis in near-infrared region showed better results than mid-infrared spectra. These results indicate that NIRS spectra contain important information that, combined with appropriate methods of pattern recognition, allow the classification of green arabica coffee samples by genotype and growing region. Besides, the PLS-DA loadings analysis allows associating which NIRS bands are specific of each class. This information can be correlated with the samples chemical composition, providing preliminary data to evaluate the effect of growing region and genotype in the selected green coffee chemical composition.