Classification of Sound and Stained Wheat Grains Using Visible and near Infrared Hyperspectral Image Analysis

Abstract

Near infrared hyperspectral image analysis has been used to classify individual wheat grains representing 24 different Australian varieties as sound or as being discoloured by one of the commercially important blackpoint, field fungi or pink stains. The study used a training set of 188 grains and a test set of 665 grains. The spectra were smoothed and then standardised by dividing each spectrum by its mean, so that the analysis was based solely on spectral shape. Penalised discriminant analysis was first used for pixel classification and then a simple rule for grain classification was developed. Overall classification accuracies of 95% were achieved over the 420–2500 nm wavelength range, as well as reduced ranges of 420–1000 nm and 420–700 nm.