Near infrared spectroscopy (NIRS) have been utilized in a wide selection of single seed applications because it provides fast and non-destructive measurements. Despite the limitation of small seed sizes, NIRS has led to successful results. In this dissertation we explored the feasibility of NIRS for several discriminative applications for corn and soybean seeds. The first application focused on discrimination of conventional and genetically modified Roundup Ready RTM soybeans. Classification accuracies ranged from 75 to 99% percent. The highest accuracies were obtained with a light tube instrument and with locally weighted principal component regression (LW-PCR) models with few samles represented. Artificial Neural Network (ANN) and Support Vector Machines models gave simmilar accuracies. The technologies performing worse were the low ressolution chemical imaging unit and the Fourier Transform transmittance instrument due to their sensitivity to seed positioning. Discrimination within a single variety was possible above 95% accuracies for most of the varieties. Moisture was proven to impact the classification due to interactions between water and carbohydrates (fiber). For this reason, this application would be feasible for breeders working in controlled seed moistures. Other applications such as discrimination of damaged corn kernels (heat and frost damage) and viability of corn and soybeans with NIRS were analyzed. Only discrimination of heat-damaged corn kernels was successful (accuracies above 95% using partial least squares discriminant analysis, PLS-DA); frost-damaged kernels and non-viable seeds could not be discriminated with any of the tested algorithms. This indicates that NIRS only detects changes in seeds due to damage and there is no relationship with its viability. The final remaining question is what the extent of damage that a seed may suffer to be detected by NIRS would be.
展开▼
