Transferência de calibração para identificação de madeiras tropicais por espectroscopia NIR independente da umidade

Abstract

The scientific and technological limitations for the rapid and reliable identification of wood from forest species based on their characteristics hinder the monitoring of the chain of custody of certified wood. The existence of solutions for species discrimination is of fundamental importance for the control and monitoring of the exploitation and transportation of native timber. Near-infrared spectroscopy (NIR) has shown promise for this purpose due to its ability to generate real-time information through reading wood samples combined with machine learning techniques. There are still gaps in the application of this technique to wood considering moisture variation, and it is necessary to understand its influence on the performance of models for the classification of these materials. Therefore, the objective is to evaluate the effect of wood moisture on the predictive capacity of the models for species discrimination based on the spectral signature in the NIR and apply calibration transfer. Eleven native species from the Amazon rainforest were used to produce 110 wood specimens with dimensions of 100 x 25 x 4 mm (length x width x thickness). NIR spectra were collected on the radial face of the wood at equilibrium moisture content (EMC%) using a benchtop FT-NIR spectrometer (Bruker, model MPA) with an integrating sphere and a portable MicroNir spectrometer (Viavi, model Onsite-W). After the saturation of the specimens, spectra were collected at maximum water content (MWC) and subsequently at every 10% of estimated water mass loss during drying. Principal component analysis, partial least squares discriminant analysis (PLS-DA), and calibration transfer will be applied to the spectral data using Chemoface® software to discriminate and group the wood species based on their spectral signatures and test the performance of the models. Preliminarily, the first principal component of the obtained spectral data captured 99.87% of the variation, being able to discriminate the species depending on the density gradient of the wood samples. The PLS-DA models, both for benchtop and portable equipment, showed prediction efficiency rates above 80%. The calibration transfer models between equipment presented prediction values between 31 and 85% of correct predictions. In view of the results, it is concluded that the spectral information provided by the two spectrometers enabled the generation of robust models for identifying the eleven native wood species, regardless of their moisture content. Furthermore, the calibration transfer between the spectrometers was successfully carried out and yielded satisfactory results.