In this work a fast method for the determination of the total sugar levels in samples of raw coffee was developed using the near infrared spectroscopy technique and multivariate regression. The sugar levels were initially obtained using gravimety as the reference method. Later on, the regression models were built from the near infrared spectra of the coffee samples. The original spectra were pre-treated according to the Kubelka-Munk transformation and multiplicative signal correction. The proposed analytical method made possible the direct determination of the total sugar levels in the samples with an error lower by 8% with respect to the conventional methodology.

A combinação da espectroscopia no infravermelho próximo (NIR) e calibração multivariada (método dos mínimos quadrados parciais - PLS) para a determinação do teor de proteína total em amostras de café cru, foi investigada. Os teores de proteína total foram inicialmente determinados usando-se como método de referência o de Kjeldhal, e, posteriormente foram construídos modelos de regressão a partir dos espectros na região do infravermelho próximo das amostras de café cru. Foram coletados 159 espectros das amostras de café cru utilizando um acessório de reflectância difusa, na faixa espectral de 4500 a 10000cm-1. Os espectros originais no NIR sofreram diferentes transformações e pré-tratamento matemático, como a transformação Kubelka-Munk; correção multiplicativa de sinal (MSC); alisamento (SPLINE); derivada primeira; média móvel e o pré-tratamento dos dados escalados pela variância. O método analítico proposto possibilitou a determinação direta, sem destruição da amostra, com obtenção de resultados rápidos e sem o consumo de reagentes químicos de forma a preservar o meio ambiente. O método proposto forneceu resultados com boa capacidade de previsão do teor de proteína total, sendo que os erros médios foram inferiores a 6,7%.

The combination of near infrared spectroscopy (NIR) and multivariate calibration using the partial least square - PLS method for the determination of the total protein level in raw coffee samples was investigated. The total protein levels were initially determined using the Kjeldhal method as the reference method. Regression models were built from the spectra in the NIR region of the raw coffee samples. Spectra of 159 samples were recorded, using an accessory of diffuse reflectance, in the range of 4500 and 10000cm-1 with 4cm-1 resolution. To the raw spectral data, different transformations and mathematical pretreatment such as Kubelka-Munk transformation; multiplicative sign correction (MSC); spline smoothing; first derivative and boxcar averaging were applied. The data was also preprocessed by scaling each column by its variance. The analytical method proposed is non invasive and the cost and time of analysis are very much reduced, making possible the fast and direct determination of the total protein content. The prediction error of protein levels given by the regression model were in average lower than 6.7%, indicating that this method is a good alternative for protein determination in raw coffee samples.