Validação independente de modelos baseados em espectros no NIR para estimativa de densidade básica da madeira

Abstract

Forest-based industries are interested in developing technologies capable of estimating the density of their raw materials quickly, reliably and automatically. The objective of this study was to analyze the robustness of models in estimating the basic density of wood by means of independent validations from near-infrared (NIR) spectra. For this, NIR spectra of wood samples from independent lots were applied to pre-established predictive models. For sample lot A, wood from discs taken from the base and 1.30 m from the ground of commercial clones of Eucalyptus spp. 5-year-old animals grown in 3x1, 3x2, 3x3, and 3x3 m spacing were used to determine the basic density and acquire NIR spectra. Samples with dimensions of 20 x 20 x 25 mm were taken from the wooden discs in the pith, intermediate region and sapwood. The spectral acquisition was performed with the aid of a fiber optic probe on the transverse and radial faces of the wood, and the basic density was determined in the laboratory by the ratio between absolutely dry mass and saturated volume. The database of sample lot B is represented by clones of Eucalyptus spp. 6.5 years old planted in a single spacing of 9 m². Specimens with dimensions of 25 x 25 x 50 mm were used. The acquisition of spectra in the NIR was performed using a fiber optic probe and an integration sphere on the transverse face of the wood. The determination of the basic density was also performed by the ratio between dry mass and saturated volume. Partial least squares regression was established considering the NIR spectra as the independent variable (x) and the basic density as the dependent variable (y). The models were validated in independent wood lots, considering the differences in spectral acquisition pathways, clones, ages, spacings and sites. The model developed with spectra measured using an integration sphere in wood from batch B (model 3) presented estimates with better correlation with the value determined in the laboratory (Rcv² = 0.81). However, when applied to the other lot, the relationship between observed and predicted values was null (Rp² = 0). Model 3 was able to identify 7 of 15 samples of lighter woods and 8 of 15 samples of denser woods. On the other hand, the model developed with spectra measured by optical fiber in the woods of Lot A (model 1) presented estimates with median correlation with the value determined in the laboratory (R²cv = 0.53). However, this model generated estimates with R²p of 0.50 with observed values. Model 1 was able to identify 7 out of 15 samples of lighter woods and 11 of 15 samples of denser woods.