Background: The chemical compounds in coffee are important indicators of quality. Its composition varies according to several factors related to the planting and processing of coffee. Thus, this study proposed to use near-infrared spectroscopy (NIR) associated with partial least squares (PLS) regression to estimate quickly some chemical properties (moisture content, soluble solids, and total and reducing sugars) in intact green coffee samples. For this, 250 samples produced in Brazil were analyzed in the laboratory by the standard method and also had their spectra recorded.
Results: The calibration models were developed using PLS regression with cross-validation and tested in a validation set. The models were elaborated using original spectra and preprocessed by five different mathematical methods. These models were compared in relation to the coefficient of determination, root mean square error of cross-validation (RMSECV), root mean square error of test set validation (RMSEP), and ratio of performance to deviation (RPD) and demonstrated different predictive capabilities for the chemical properties of coffee. The best model was obtained to predict grain moisture and the worst performance was observed for the soluble solids model. The highest determination coefficients obtained for the samples in the validation set were equal to 0.810, 0.516, 0.694 and 0.781 for moisture, soluble solids, total sugar, and reducing sugars, respectively.
Conclusion: The statistics associated with these models indicate that NIR technology has the potential to be applied routinely to predict the chemical properties of green coffee, and in particular, for moisture analysis. However, the soluble solid and total sugar content did not show high correlations with the spectroscopic data and need to be improved. © 2020 Society of Chemical Industry.
Keywords: green coffee; multivariate regression; near infrared; predictive models; preprocessing.
© 2020 Society of Chemical Industry.