|Delwiche, Stephen - Steve|
Submitted to: National Fusarium Head Blight Forum Proceedings
Publication Type: Abstract Only
Publication Acceptance Date: 12/9/2003
Publication Date: 12/13/2003
Citation: Delwiche, S.R., Hareland, G.A. 2003. Detection of scab-damaged wheat kernels by near-infrared reflectance. Proceedings of the 2003 National Fusarium Head Blight Forum (S.M. Canty, J. Lewis, and R.W. Ward, eds.), Dec. 13-15, 2003, Bloomington, MN. East Lansing: Michigan State University, p. 186.
Technical Abstract: Both wheat breeder and wheat inspector must currently deal with the assessment of scab in harvested wheat by manual human inspection. We are currently developing and examining the accuracy of a semi-automated wheat scab inspection system that is based on near-infrared (NIR) reflectance (1000 to 1700 nm) of individual kernels. Our initial work revealed that, for scanning, the kernels could be oriented in just a semi-random basis, in which the rotational angle about a kernel's long axis was arbitrary. Classification analysis has involved the application of various statistical classification techniques, including linear discriminant analysis, soft independent modeling of class analogy (SIMCA), partial least squares regression, and non-parametric (k-nearest-neighbor) classification. For the most recent year evaluated (2002), average cross-validation accuracy ranged from 82.1% (a wavelength difference, without kernel mass, model) to 89.6% (a k-nearest-neighbor, with kernel mass, model). Although the lower value in this range was indeed lower than that for a model using mass alone (83.8%), the corresponding accuracies of these models on a separate (fully independent) test set indicated that the spectrally based models, with accuracies in excess of 92% were clearly better than the mass alone model. Based on test set accuracy, there were only slight differences between models that were based on principal component scores and those that were based on a simple wavelength difference. Typically, test set accuracies were between 94 and 97 percent. For the k-nearest-neighbor model, the number of neighbors needed to achieve stable optimal accuracies was approximately 20. An exhaustive search of the most suitable wavelength pairs for the absorbance [A = log(1/R)] difference, [A(wavelength 1) minus A(wavelength 2)], revealed that the low-wavelength side of a broad carbohydrate absorption band (centered around 1200 nm) was very effective at discriminating between healthy and scab-damaged kernels, with accuracies at about 95%. The best wavelength difference was [A(1248 nm) - A(1140 nm)]. Although the average cross-validation accuracy was lower for this model (82.1%) compared to the k-nearest-neighbor model (83.3%), the test set accuracies were nearly identical (94.9% and 95.0%). Many other wavelength differences, [A(x) - A(y)], produced cross-validation accuracies that were within 0.5 percentage units of the optimal difference, with values for x favoring the 1150-1300 nm region, and values for y favoring the 1000-1150 region. Combined, these regions define the broad absorption band centered near 1200 nm, which is attributed to the second overtone of a carbohydrate CH stretch. The achieved accuracy levels demonstrate the potential for the use of NIR in inspection operations for wheat scab. Therefore, development of an automated, high-speed device utilizing as few as two wavelengths for wheat scab detection appears to be feasible.