Global near infrared spectroscopy models to predict wood chemical properties of Eucalyptus

Global near infrared spectroscopy models (multiple-species, multiple-sites) were developed to predict chemical properties of Eucalyptus wood. The sample data set included 186 samples from four data sets (five species) originating from six countries: Eucalyptus urophylla from Argentina, Colombia, Venezuela, and South Africa; Eucalyptus dunnii from Uruguay; Eucalyptus globulus and Eucalyptus nitens from Chile; and Eucalyptus grandis from Colombia. The 186 samples were all preselected from larger collections of 400 to nearly 1800 samples to represent the range of chemical and spectral variation in each data set. The chemical traits modeled were total lignin, insoluble lignin, soluble lignin, syringyl-guaiacyl ratio (S/G), glucose, xylose, galactose, arabinose, and mannose. Single-species models and global multiple-species models were developed for each chemical constituent. For the global model, the R-cv(2) for total lignin, insoluble lignin and syringyl-guaiacyl ratio were 0.95, 0.96, and 0.86, respectively. An alternate expression of the syringyl-guaiacyl relationship (S/(S+G)) resulted in better near infrared calibrations (e.g., for the global model, R-cv(2) = 0.95). The global models for sugar content were also very good, but were slightly inferior to those for the lignin related traits, with R-cv(2) = 0.74 for glucose, 0.89 for xylose, and from 0.72 to 0.91 for the minor sugars. To investigate the utility of the global models to predict chemical traits for species not included in the calibration, three-species calibrations were used to predict each trait in a fourth species data set. The prediction fit statistics ranged from excellent to poor depending on the species and trait, but in general the predictions would be at least moderately useful for most species-trait combinations. For some species-trait combinations with poor initial predictions from the global model, the inclusion of 10 samples from the "new'' species into the calibration global model improved the fit statistics substantially. The global calibrations will be useful in tree breeding programs to rank species, families, and clones for important wood chemical traits.