Multispecies, multisite, multi-age PLS regression models of chemical properties of eucalypts wood using Fourier Transformed near-Infrared (FT-NIR) spectroscopy

Early AccessInternational audienceNear Infrared Spectroscopy (NIR) is often used to perform high throughput phenotyping on thousands of genotypes using prediction models with high variability. A study was therefore undertaken to analyze the potential of multispecies, multisite and multi-age NIR calibration models of seven chemical properties of eucalyptus wood. The models are based on 358 samples selected among more than 5000 samples that belong to five eucalypt species including hybrids. The samples were collected from trees aged 2-35 originating from four different countries. Spectra were measured on non-extracted wood powders using an FT-NIR spectrometer. Models were established in the spectral range of 9090-4040 cm(-1) using the PLS regression method, tested by repeated cross-validation and validated on independent test sets. The results showed that the robust models for total extractives (R-P(2) = 0.91, RMSEP = 1.20%, RPD = 3.3) and KL (R-P(2) = 0.89, RMSEP = 1.21%, RPD = 3.0) provided good predictions. These two properties were the best predicted, followed by the S/G ratio (R-P(2) = 0.84, RMSEP = 0.19, RPD = 2.5) and ASL content (R-P(2) = 0.81, RMSEP of 0.54, RPD = 2.3). For holocellulose, alphacellulose, and hemicelluloses contents, the models provided approximate predictions. The prediction errors were always less than twice of the laboratory errors except for ASL and S/G ratio. For total extractives and ASL, beta-coefficients of models were of approximately the same magnitude throughout the 9000-4000 cm(-1) region while for the five other properties, they were higher in the 7500-4000 cm(-1) region. Models were also established in narrower NIR regions, and the quality of models obtained was about the same as that of the models based in the 9090-4000 cm(-1) wide range. These established robust models can be used to make predictions based on samples of high variability