Near infrared hyperspectral imaging for the evaluation of endosperm texture in whole yellow maize (Zea maize L.) kernels

Abstract

Near infrared hyperspectral images (HSI) were recorded for whole yellow maize kernels (commercial hybrids) defined as either hard, intermediate, or soft by experienced maize breeders. The images were acquired with a linescan (pushbroom) instrument using a HgCdTe detector. The final image size was 570 x 219 pixels in 239 wavelength bands from 1000 to 2498 nm in steps of approximately 6.5 nm. Multivariate image cleaning was used to remove background and optical errors, in which about two-thirds of all pixels were removed. The cleaned image was used to calculate a principal component analysis (PCA) model after multiplicative scatter correction (MSC) and mean-centering were applied. It was possible to find clusters representing vitreous and floury endosperm (different types of endosperm present in varying ratios in hard and soft kernels) as well as a third type of endosperm by interactively delineating polygon based clusters in the score plot of the second and fourth principal components and projecting the results on the image space. Chemical interpretation of the loading line plots shows the effect of starch density and the protein matrix. The vitreous and floury endosperm clusters were used to make a partial least-squares discriminant analysis (PLS-DA) model, using four components, with a coefficient of determination (R(2)) for the y data (kernel hardness category) for the training set of over 85%. This PLS-DA model could be used for prediction in a test set. We show how the prediction images can be interpreted, thus confirming the validity of the PCA classification. The technique presented here is very powerful for laboratory studies of small cereal samples in order to produce localized information.