Desenvolvimento de metodologias analíticas para a classificação e a determinação do índice de cor em amostras de açúcar, usando imagens digitais, espectroscopia NIR e quimiometria

Abstract

Analysing the colour of sugar is very important because it can serve as an indicator of its level of purity, quality and suitability for different industrial and food uses. This work proposes two new analytical methods, one based on digital images and the other on NIR spectroscopy, both associated with chemometric techniques, to determine one of the main quality parameters of sugar, its color. Furthermore, to classify sugar in terms of its suitability for different types of market based on its color. To this end, strategies for building classification models were employed on images decomposed into grey scale, RGB and HSI channels and their combinations. Two classification studies were evaluated. The first involved discriminating between refined sugar (RS) and unrefined sugar (US) samples. The second study involved classifying sugar samples into three different classes: refined for the external market (EMRS); refined for retail sugar (RRS) and US. The best results obtained for the two studies carried out were constructed by GA-LDA. With two classes (RS and US), it showed 97.4 per cent accuracy for the combination of Gray+RGB+HSI channels; when inserting the third class, the RGB+HSI combination led to 95.7 per cent accuracy. The models also showed good predictive capacity, achieving the best result for the model built from Gray+RGB+HSI/SPA-MLR, which resulted in r2pred values of 0.976, RMSEP of 4.882 UI and REP of 9.879%. The study based on NIR spectroscopy used two pieces of equipment, one bench-top and one portable, where the recorded spectra were submitted to a classification study with the aim of classifying the samples as standard refined sugar for the external market or unrefined sugar. The results were satisfactory on both pieces of equipment, recording up to 98.9% overall model accuracy with 100% correct classification in the test set for some models, such as those whose spectra were subjected to LBC and modelled by PLS-DA (LBC/PLS-DA) and pre-processing by smoothing and derivation for application of iSPA-PLS-DA (SG-1d-2o-21w/iSPA-PLS-DA) with data recorded by the bench equipment. The models based on iSPA-PLS-DA obtained 100% accuracy when the spectra were pre-processed using LBC and SNV (LBC/iSPA-PLS-DA and SNV/iSPA-PLS-DA) for data obtained by the portable equipment. The calibration models built with data obtained on the benchtop equipment showed lower results than those obtained on the portable equipment, with the best results showing an RMSEP of 5.087 UI, r2pred of 0.971 and REP of 9.375% for the OFF/iSPA-PLS model. It can therefore be concluded that, given the results obtained, the methods presented are an interesting alternative for obtaining the color index of sugar.