Estudo dos genótipos de café arábica utilizando FTIR e redes neurais artificiais

Abstract

The climatic conditions of the coffee crop give special attributes to the beverage and could improve its value. However, it is essential to prove the geographical and genotype origin of the cultivar using reliable methods. Several mathematical methods have been developed in an attempt to reproduce the human capability of pattern recognition. The artificial neural networks (ANNs) are a set of techniques based on statistical principles, which is currently growing in the food science to perform tasks of regression and pattern recognition. ANNs are techniques that can represent complex and nonlinear input-output relationships. In this study three types of artificial neural network were used, namely the self-organizing maps (unsupervised learning), the multilayer perceptron (supervised learning) and the radial-basis function network (hybrid learning process), for the recognition and geographic and genotypic classification of arabic coffee. For this purpose, spectra obtained in the Fourier transform infrared (FTIR) were analyzed using different ANNs. Other techniques were also applied during neural network development: ensemble averaging to improve the network generalization and sequential simplex optimization to select parameters. It was concluded that the SOM were able to recognize the coffee genotypes and geographical origin using the chemical profile data. After the results it was concluded that the optimized multilayer perceptrons were able to classify the samples of arabica coffee geographically. However, for genotypic classification the performance was not satisfactory. The optimized radial-basis function networks showed more satisfactory results since these networks were able to classify the samples of arabica coffe geographically and genotypically. The radial-basis function networks presents the best results to genotypic classification and in addition a smaller number of free parameters compared with multilayer perceptrons, which has a high number of synaptic weights, thus a large degrees of freedom database is necessary to produce a network with generalization capability.