Automated sorting and single kernel analysis by near-infrared hyperspectral imaging

An automated single kernel analysis (SKA) and sorting system was designed and constructed to singulate maize kernels from a large population, place them in the field of view (FOV) of a near-infrared hyperspectral imaging spectrometer for analysis, and sort according to predicted constituent values. The imaging and prediction processes were controlled using a personal computer, while a microcontroller was used to automate kernel singulation, placement, and sorting operations. The system was calibrated to predict crude protein using 360 kernels, selected randomly from a set of forty bulk maize standardization sets. Each kernel was analyzed with the imaging spectrometer, spectra were extracted, and a protein calibration was derived. Instead of regressing against reference chemistry assay values, the single-kernel spectra were regressed against the mean protein concentration of the bulk standardization sets from which each kernel was drawn. A test was performed to validate the accuracy of the protein calibration. The bulk protein concentration (known a priori) of a standardization set was statistically compared to the mean of the predicted protein values for that corresponding set; this was repeated for numerous standardization sets. The SKA was also tested for repeatability, kernel geometry, mechanical reliability, sorting capability, and kernel positioning effects. Results from tests showed the mechanism was able to make protein predictions and sort kernels according to those values, good repeatability, and a limited increase in standard error due to changes for both kernel placement and kernel geometry. The mechanism and software functioned well and allowed the concept to be demonstrated as a very practical method of automated single kernel analysis