No abstract

No abstract

The current US corn grading system accounts for the portion of damaged kernels, measured by timeconsuming and inaccurate visual inspection. Near infrared spectroscopy (NIRS), a non-destructive and fast analytical method, was tested as a tool for discriminating corn kernels with heat and frost damage. Four classification algorithms were utilized: Partial least squares discriminant analysis (PLS-DA), soft independent modeling of class analogy (SIMCA), k-nearest neighbors (K-NN), and least-squares support vector machines (LS-SVM). The feasibility of NIRS for discriminating normal or viable-germinating corn kernels and soybean seeds from abnormal or dead seeds was also tested. This application could be highly valuable for seed breeders and germplasm-preservation managers because current viability tests are based on a destructive method where the seed is germinated. Heat-damaged corn kernels were best discriminated by PLS-DA, with 99% accuracy. The discrimination of frost-damaged corn kernels was not possible. Discrimination of non-viable seeds from viable also was not possible. Since previous results in the literature contradict the current damagediscrimination results, the threshold of seed damage necessary for NIRS detection should be analyzed in the future. NIRS may accurately classify seeds based on changes due to damage, without any correlation with germination. KeywordsAgronomy, Seed Science Center, single seed, near infrared spectroscopy, discrimination The current US corn grading system accounts for the portion of damaged kernels, measured by timeconsuming and inaccurate visual inspection. Near infrared spectroscopy (NIRS), a non-destructive and fast analytical method, was tested as a tool for discriminating corn kernels with heat and frost damage. Four classification algorithms were utilized: Partial least squares discriminant analysis (PLS-DA), soft independent modeling of class analogy (SIMCA), k-nearest neighbors (K-NN), and least-squares support vector machines (LS-SVM). The feasibility of NIRS for discriminating normal or viable-germinating corn kernels and soybean seeds from abnormal or dead seeds was also tested. This application could be highly valuable for seed breeders and germplasm-preservation managers because current viability tests are based on a destructive method where the seed is germinated. Heat-damaged corn kernels were best discriminated by PLS-DA, with 99% accuracy. The discrimination of frost-damaged corn kernels was not possible. Discrimination of non-viable seeds from viable also was not possible. Since previous results in the literature contradict the current damage-discrimination results, the threshold of seed damage necessary for NIRS detection should be analyzed in the future. NIRS may accurately classify seeds based on changes due to damage, without any correlation with germination.

Four near-infrared spectrophotometers, and their associated spectral collection methods, were tested and compared for measuring three soybean single-seed attributes: weight (g), protein (%), and oil (%). Using partial least-squares (PLS) and four preprocessing methods, the attribute that was significantly most easily predicted was seed weight (RPD > 3 on average) and protein the least. The performance of all instruments differed from each other. Performances for oil and protein predictions were correlated with the instrument sampling system, with the best predictions using spectra taken from more than one seed angle. This was facilitated by the seed spinning or tumbling during spectral collection as opposed to static sampling methods. From the preprocessing methods utilized, no single one gave the best overall performances but weight measurements were often more successful with raw spectra, whereas protein and oil predictions were often enhanced by SNV and SNV + detrending. ABSTRACT: Four near-infrared spectrophotometers, and their associated spectral collection methods, were tested and compared for measuring three soybean single-seed attributes: weight (g), protein (%), and oil (%). Using partial least-squares (PLS) and four preprocessing methods, the attribute that was significantly most easily predicted was seed weight (RPD > 3 on average) and protein the least. The performance of all instruments differed from each other. Performances for oil and protein predictions were correlated with the instrument sampling system, with the best predictions using spectra taken from more than one seed angle. This was facilitated by the seed spinning or tumbling during spectral collection as opposed to static sampling methods. From the preprocessing methods utilized, no single one gave the best overall performances but weight measurements were often more successful with raw spectra, whereas protein and oil predictions were often enhanced by SNV and SNV + detrending.

No abstract