Aplicação de modelos de dois estágios em problemas de classificação de alta complexidade: segmentação geográfica e genotípica de café arábica

Abstract

Currently, Brazil is the largest producer of coffee, accounting for 33.6% of world production. The coffee belongs to the Coffea genus, from Rubiaceae family. The arabica and canephora (robust) species have great global economic importance, being the arabica responsible for 90% of production. In addition to the species, the coffee genotype also influences the quality of the beverage. The objective of this project was to develop a methodology to discriminate the different genotypes of arabica coffee, and also identify the cultivation region. Seventy-four samples of green beans of 20 genotypes of arabica coffee, grown in the cities of Mandaguari, Londrina, Paranavaí and Cornélio Procópio were provided by IAPAR (Londrina-PR). Spectra of samples were obtained by infrared spectroscopy with Fourier transform (FTIR). So, two-stage models were created using a first linear stage and a second nonlinear one. For the linear stage it was used the principal component analysis (PCA) and partial least squares method with discriminant analysis (PLS-DA). With PLS-DA, it was also possible to perform the classification of samples, providing a further comparison between the linear model and the two-stage model. For the second stage of the model it was used a regularized radial basis functions artificial neural network (RBF-R). In neural networks construction several parameters should be optimized and, in this work the sequential simplex method was used for this purpose. For geographical classification, the best model was the PLS-DA using the raw spectra in the range of 750 and 3750 cm-1. The obtained model classify correctly 100% of the samples and, had better performance confirmed by the thresholds established by Bayes' theorem. In genotypic classification, the best model found was the two-stage one using the first derivative of spectra in the range between 800 and 1900 cm-1 and PLS-DA as first stage. This model was able to correctly classify 89.04% of test specimens, and obtained better performance based on Bayes' theorem. Even performing a 100% correct geographical classification of samples, Bayes' inference showed that the models should still be modified in an attempt to find better results for sensitivity and specificity, and decrease the number of samples in the rejection region.