Abstract
Currently, coffee fermentation is visually operated, which results in incomplete or excessive processes and coffees with undesirable characteristics. In front of it, pH and total soluble solids (TSS) have been shown to be good fermentation indicators, although this requires rapid, accurate, and chemical-free measurement techniques such as NIR spectroscopy. However, the complexity of the NIR spectra requires optimization steps in which variable selection techniques simplify profiles and subsequent models. This work tests a new covering array feature selection (CAFS) approach on NIR spectra to optimize prediction models in coffee samples during fermentation. Spectral profiles in the range 1100-2100 nm were extracted from coffee beans (Typica, Caturra, and Catimor varieties) raw and during fermentation (4, 8, 12, 16, 20, and 24 h). Partial least-squares regressions (PLSR) were performed using full spectra using a five-fold cross-validation strategy for training and validation. The relevant wavelengths were then selected using the β coefficients, the important projection of variables (VIP), and the CAFS method. Finally, optimized models were performed using the relevant wavelengths and compared among these using their statistical metrics. The models performed using the selected variables (22-47) of CAFS showed the best performance in predicting pH (R (2) = 0.825-0.903, RMSE = 0.096-0.158, RPD = 6.33-10.38) and TSS (R (2) = 0.865-0.922, RMSE = 0.688-1.059, RPD = 0.94-1.45) compared to the other methods. These findings suggest that simple and efficient models could be performed and implemented in routine analysis due to the maximum coverage and minimum cardinality of CAFS.