MULTIVARIATE PREDICTION OF WOOD SURFACE-FEATURES USING AN IMAGING SPECTROGRAPH

This paper presents a multispectral system for evolution of linear alg orithms for prediction of wood surface features important for automati c inspection of lumber. The selection of training samples, the imaging spectrograph scanning method, raw data representation, evaluation of linear algorithms and testing of performance is discussed. A possible on line implementation for high speed wood scanning with a smart senso r is outpointed. An example, showing the evolution of linear algorithm s for prediction of compression wood in softwood species (Picea abies, Pinus sylvestris), is reported, showing verified 92-94% correct class ification. It is shown that compression wood classification could be r educed to an uncomplicated linear model using just a few spectral comp onents where the most important one is around the limit for visible li ght going to the Ultraviolet spectra. This almost univariate behaviour for the model is not the common behaviour for other wood surface feat ures (Brunner et al., 1996; Hagman, 1995; Hagman, 1996).