Rapid techniques for screening wood properties in forest plantations

Abstract

The use of three nondestructive techniques (NDT) for the assessment of wood properties in forest plantations: i) acoustics; ii) near infrared (NIR) spectroscopy; and iii) SilviScan, was studied. In this dissertation, a major focus was given to aspects related to the development of NIR calibration models using SilviScan data. The utilization of acoustics in Pinus taeda plantations was useful for the classification of trees based on wood stiffness. NDT values compared well to those obtained from traditional laboratory tests after accounting for differences in moisture content and the type of acoustic waves employed. For NIR modeling, using pulp yield data from Eucalyptus nitens trees, it was found that models fitted with a reduced number of selected samples showed similar performance compared to a regression fitted using all samples. Selection of samples based on NIR spectra was as successful as selection based on pulp yields. The partial least squares (PLS) algorithm was modified allowing incorporation of grouped data. It was shown that repeated measurements of wood strips could induce autocorrelation in the PLS residuals. Models fitted using the modified approach showed that the serial correlation can be completely removed; however the strength of the autocorrelation will determine whether PLS or its modified version is preferable. Net analyte signal and figures of merit were introduced to help interpreting the effects of different pre-processing techniques on the NIR data. Sensitivity, selectivity, and signal-to-noise ratio proved to be more useful statistics than R2, RMSECV, and RMSEP for this objective. Two approaches based on randomization tests were proposed for determining the dimensionality of principal component (PCA) and PLS models. Compared to these tests, it was found that cross-validation can lead, in some situations, to model underfitting. Alternatively, the development of nonlinear NIR calibrations based on kernel methods showed that some problems detected with PLS were due to nonlinearities in the properties. Kernel regressions were able to model better these properties, improving their predictive ability compared to PLS. Finally, the use of NIR values as control points for spatial interpolation was demonstrated by generating within-tree maps showing the variation of density and microfibril angle in P. taeda trees.