Raman spectrometry and neural networks for the classification of wood types—1

Abstract

In this work the coupling of near infrared (NIR) Fourier-transform (FT) Raman spectroscopy and neural computing for spectral feature extraction and classification of woods is reported. A NIR FT-Raman spectrometer operating at 1064 nm was used for all measurements; particular attention was paid to the effects of sample fluorescence and heating. It was demonstrated that fluorescence rejection is accomplished only for the lighter colored woods and that fluorescence was found to be severe for 10 of the 71 woods studied in this work even using excitation at 1064 nm. It was further found that hardwoods were no more or less susceptible to sample heating than softwoods. Feed-forward neural networks were used to extract the principal features of wood spectra at resolutions of 4, 8 and 16 cm⁻¹ and to classify spectra as either temperate hardwoods or temperate softwoods. Neural networks were constructed using zero and two processing elements in the hidden layer. It was shown that neural networks with two hidden processing elements perform near optimally, since each hidden layer processing element may function as either a hardwood or softwood feature detector. This work represents the first time that FT-Raman spectroscopy and neural network technology have been coupled for spectral feature extraction and classification.