Spectral Data Compression and Analyses Techniques to Discriminate Wheat Classes

Abstract

Near-infrared spectroscopy (NIRS) was evaluated to differentiate six different classes of wheat grown in western Canada. The original spectral data consisted of 2594 wavelength variables, and three data compression techniques, namely, principal component analysis (PCA), discrete Fourier transform (DFT), and discrete wavelet transform (DWT), were compared to reduce the dimensionality of the original dataset. Five different classifiers, namely, linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), k-nearest neighbor (k-NN) classifier, radial basis function probabilistic neural network (PNN), and least-squares support vector machines (LS-SVM), were compared based on correct classification rates. Classification was performed on principal component scores, Fourier coefficients, and wavelet coefficients of the original spectral data. For dimensionality reduction, PCA was most efficient technique. It was also corroborated that linear end point baseline correction was necessary to achieve efficient data compression using DFT. Classification accuracies achieved using LDA or QDA combined with PCA as a pre-processing method were consistently better than the other three classifiers. Classification results based on Fourier or wavelet coefficients were less favorable than those directly obtained from principal component scores of the original spectra. LS-SVM did not perform well on test samples. Fisher's criterion (FC) was used to select eight wavelength features, and it was demonstrated that LDA, k-NN, and PNN classifiers could effectively discriminate wheat classes based on reflectance spectra.