RESUMEN
Near-infrared (NIR) spectroscopy and chemometric methods were used to predict the chemical properties of decomposing eucalyptus harvest residues to better understand the decomposition process of these materials. Leaves, twigs, branches, and bark from a decomposition experimental set up in commercial plantations were sampled for one year. The contents of carbon (C), nitrogen (N), extractives (EX), acid-soluble lignin (SL), Klason insoluble lignin (KL) and holocellulose (HC) were determined by the reference method in the collected samples. Principal component analysis (PCA) was employed to distinguish the types of harvest residues throughout the decomposition period. Multi-residue regression models were built from the NIR spectra using partial least squares regression (PLS). Two feature selection methods, i.e., ordered predictors selection (OPS) and genetic algorithm (GA), were applied and compared. The OPS and GA did not differ statistically; however, compared with the GA, OPS was more computationally efficient and selected fewer variables. Using the PLS-OPS models, the root mean square errors of prediction (RMSEP) for C, N, EX, SL, KL and HC were 19.70, 0.08, 0.74, 0.39, 28.13 and 33.99, respectively, and the prediction correlations (Rp) for these properties were 0.94, 0.99, 0.99, 0.99, 0.96 and 0.98, respectively. PLS-discriminant analysis (PLS-DA) was used to classify the samples over the decomposition time and provided a good separation. Some mismatches obtained in the modeled classes were explained by the differences in the decomposition rate and changes in the chemical composition of the different harvest residue components that were evaluated. The results showed the feasibility of NIR spectroscopy and chemometric methods to evaluate the chemistry of decomposing eucalyptus harvest residues, indicating that these methods can be used as rapid and inexpensive alternatives to conventional methods to help understand the decomposition process.