Considerations regarding the use of hyperspectral imaging data in classifications of food products, exemplified by analysis of maize kernels.

Development of robust analytical procedures is critical when using hyperspectral imaging technology in food technology and agriculture. This study used near-isogenic inbred corn lines to address two basic questions: (1) To what extent is classification accuracy increased by grinding maize kernels? (2) Can the classification accuracy of two near-isogenic inbred lines be increased by using a spectral filter to classify only certain hyperspectral profiles from each image cube? Whole kernels and ground kernels in two particle intervals, 0.250-0.354 mm (size 1) and 0.354-0.841 mm (size 2), were examined. Spectral profiles acquired from ground kernels had higher spectral repeatability than data collected from whole kernels. The classification error of discriminant functions from whole kernels was >3 times lower than that of size 1 ground particles. Applying a spectral filter to input data had negligible effect on classifications of hyperspectral profiles from whole kernels and size 2 ground particles, but for size 1 ground particles a considerable increase in accuracy was observed. Independent validation confirmed that distinction between wild type and mutant inbred maize lines could be conducted with >80% accuracy after the proposed spectral filter had been applied to hyperspectral profiles of size 1 ground particles. A combination of discriminant analysis and regression analysis could be used to accurately predict mixture ratios of the two inbred lines. The use of spectral filtering to increase the level of spectral repeatability and the use of hyperspectral imaging technology in large-scale commercial operations are discussed.