Robustness of Models Based on near Infrared Spectra to Predict the Basic Density in Eucalyptus Urophylla Wood

Abstract

Scientific contributions have shown good results by using near infrared (NIR) spectroscopy as a rapid and reliable tool for characterising lignocellulosic materials. Many reports have evaluated the predictive power and the robustness of the NIR models by means of methods known to validate them. However, in most of these investigations, the samples were divided systematically into two non-independent groups: one group was used to build and the other to validate the NIR models. This approach does not adequately simulate a real situation in which the properties of unknown samples should be predicted by established NIR models. Hence, the aim of this paper was to evaluate the robustness of models based on NIR spectroscopy to predict wood basic density in Eucalyptus urophylla using two totally independent sample sets. Wood density and NIR spectra were measured in diffuse reflectance mode on transversal, radial and tangential surfaces of wood samples in two data sets. We used one data set to build partial least squares regression (PLS-R) models and another to validate them and vice versa. The predictive models developed from the radial surface NIR spectra proved satisfactory with r2 p varying from 0.79 to 0.85 and RPD ranging from 2.3 to 2.7, while the spectra measured on tangential and transversal wood surfaces generated less robust regression models. Our results showed that it is possible to assess wood density in unknown samples by established PLS-R models from solid wood samples preferably using radial surfaces.