Arabica coffee classification using near infrared spectroscopy and two-stage models

Abstract

Coffee quality depends on the environment al conditions of the growing area. Factors such as climate, soil type and altitude, associated with agricultural practices, directly influence the chemical composition of the coffee beans. This study developed two - stage models to determine the geographic and genotypic origin of the grain. For the first stage, the partial least squares with discriminant analysis (PLS - DA) and principal component analysis (PCA) models were tested. Then, two artificial neural network (ANN) non - linear models, i.e. multilayer perceptron (MLP) and the radial - basis function (RBF), were evaluated as the second stage. Samples from four genotypes, cultivated in four different cities within Parana State in Brazil, were analyzed using near infrared spectroscopy (NIRS) in the 1100 to 2498 nm range. Three preprocessing techniques were tested on the spectra, i.e. multiplicative scatter correction (MSC); the Savitzky - Golay second - derivative and both combined. The best models were obtained with the spectra treated using MSC plus the second - derivative, with PLS - DA as first stage followed by the RBF network. For geographic and genotypic classification the sensitivity and specificity values of 100% were obtained for the training and test sets. The NIRS spectra presented better class separation when compared with the FTIR spectra used in a previous work. These results demonstrate that NIRS spectra, allied with the right pattern recognition techniques, can be used as a quick and efficient technique to distinguish green coffee samples both geographically and genotypically.