Multivariate regression methods for estimating basic density in Eucalyptus wood from near infrared spectroscopic data

Abstract

Near infrared (NIR) spectroscopy is a fast and efficient technique to predict a range of wood traits; however, methods for extracting useful information from the NIR spectra could be improved. Thus, the aim of this study was to evaluate the statistic performance of two regression methods for estimating the basic density in Eucalyptus urophylla x grandis wood from near infrared spectroscopic data. The predictive models calibrated by principal component regression (PCR) or partial least square regression (PLSR) method provided fine correlations. The coefficients of determination (R²cv) of the PCR models ranged from 0.78 to 0.85 with standard error of cross-validation (SECV) and the ratio of performance to deviation (RPD) varying from 32.8 to 41.2 kg/m³ and from 1.6 to 1.9, respectively. The PLSR models presented R²cv with relatively lower magnitude (from 0.65 to 0.78); but also lower SECV (from 29.8 to 38.9 kg/m³) and higher RPD values (from 1.6 to 2.1). In short, PCR method provides higher R² between Lab-measured and NIR-predicted values while PLSR produces lower standard errors of cross-validations. For both regression methods, the pre-treatments on NIR spectra, and the wavelength selection improved the calibration statistics, reducing the SECV and increasing the R²cv and the RPD values. Thus, PCR and PLS regression can be applied successfully for predicting basic density in Eucalyptus urophylla x grandis wood from the near infrared spectroscopic data.