Single Wheat Kernel Size Effects on Near-Infrared Reflectance Spectra and Color Classification

Cereal Chem. 76(1):34-37 An optical radiation measurement system was used to measure reflectance spectra of single wheat kernels from 400 to 2,000 nm. Five classes of wheat were used for this study. Three kernel sizes (large, medium, and small) were used to determine how wheat kernel size affects visible and near-infrared (NIR) reflectance spectra and single wheat kernel color classification. Mean kernel weights ranged from 35.6 to 57.2 mg for large kernels, from 26.9 to 45.0 mg for medium kernels, and from 17.6 to 31.4 mg for small kernels. The results showed that wheat kernel size significantly affects visible and NIR reflectance spectra. Determination of the color class of red kernels increased in accuracy and that of white kernels decreased in accuracy as kernel size decreased. Data pretreatments such as multiplicative scatter correction (MSC), first or second derivatives, and first or second derivatives with MSC reduced the effect of kernel size on reflectance spectra and color classification. Many factors affect the accuracy of near-infrared (NIR) reflectance spectroscopy for the analysis of agricultural products such as grains. Although NIR spectroscopy is relatively simple when compared with other types of analyses (there are no reagents to prepare and weighing is not required for some analyses), nearly 40 sources of error have been identified (Hruschka 1990). Of these, sample factors and operational factors are the two major sources of error affecting measurement accuracy. Williams (1975) and Williams and Thompson (1978) indicated that the most important sample factors affecting the accuracy of NIR reflectance spectroscopy for the analysis of ground wheat samples were the mean particle size and particle size distribution. Sample factors are influenced by growing environment, which causes large variations in kernel size, kernel size distribution, kernel density, and kernel color. These variations affect spectral measurement by influencing the absorption of radiation by the wheat sample (Watson et al 1977). Norris and Williams (1984) studied the effect of particle size of ground hard red spring wheat samples on spectral measurements. They found that sample-induced errors in NIR reflectance spectral analysis could be caused by the variation of the sample surface. They reported that the log (1/R) values were significantly affected by particle size; coarse samples had higher absorption values. The effect of particle size was greater at longer wavelengths. They found that the second derivative divided by the second derivative at optimum wavelengths gave the best performance in predicting protein content of hard red spring wheat samples that varied widely in particle size. NIR spectroscopy research has shifted from ground and bulk sample analysis to single-kernel analysis and from NIR transmittance to NIR reflectance (Delwiche and Massie 1996). Part of the reason for the shift from transmittance to reflectance is that reflectance is easier to automate. To meet these changes and obtain high levels of accuracy, the effect of grain’s physical properties, such as kernel size, on single kernel spectra must be understood. Kernel size can influence the measurement of single-kernel NIR spectra by affecting the amount of light (F) reaching the sensor, which is determined by the equation F = f/D, in which f is the focal length and D is the diameter of the illuminated object, if it is smaller than the field of view. The radiant flux (φc) increases as the inverse of the square of F: φc = 1/F . The smaller the value of F, the more radiant flux is collected by the lens. To obtain high levels of accuracy during the measurement, the focal length f, or distance from the kernel to the fiber, must be kept constant. However, the variation in kernel size can cause the focal length to change during measurement. Also, differences in kernel sizes will result in different illumination and reflectance areas and direction of reflected light. The objectives of this study were to: 1) identify the effect of wheat kernel size on single wheat kernel reflectance spectra and color classification, and 2) find methods to reduce the effect of kernel size on the reflectance spectra. MATERIALS AND METHODS Five U.S. market classes of wheat, hard red spring, hard red winter, soft red winter, hard white wheat, and soft white wheat, were supplied by the Grain Inspection, Packers, and Stockyards Administration (GIPSA) Technical Center (Kansas City, MO) (Table I). Single wheat kernel reflectance spectra from 400 to 2,000 nm at 2-nm intervals were collected with an optical radiation measurement system (Oriel, Stratford, CT), which has been described by Wang et al (in press). A single wheat kernel was suspended horizontally from the germ end by a vacuum tube, and the crease side of the kernel was viewed. Two experiments were conducted. The first experiment studied the effect of kernel size on NIR spectra. For this experiment, 375 kernels (25 small, medium, and large kernels from each class) were selected. The second experiment studied the effect of kernel size on color classification. For this experiment, 375 kernels (25 from three cultivars of each class) were selected randomly. These kernels were used as a calibration set to develop calibration equations that were used to predict the color of kernels used in the first experiment. Wang (1997) gave a complete discussion of all samples used in these experiments. To keep the distance constant between the reflectance surface of a wheat kernel and the end of the optical fiber, a three-dimensional multi-axis precision translator (Newport, Irvine, CA) was used to adjust kernel position. To maintain proper orientation, a waveplate/polarizer holder (Newport) was used to adjust kernel orientation from 0 to 360°. The spectral data were first transferred to log (1/R) form by using the software package Grams/32 (Galactic Industries, Salem, NH). The spectral data were smoothed by the method of Savitsky and Golay (1964). A fifth-degree polynomial with 25 points was used for smoothing the spectral region from 400 to 1,450 nm, and a fifth-degree polynomial with 45 points was used for smoothing the spectral region from 1,450 to 2,000 nm. Derivatives were used to correct for overlapping peaks and large baseline variations. A 1 Grain Marketing and Production Research Center, USDA-ARS, Manhattan, KS 66502. Mention of a trademark or proprietary product does not constitute a guarantee or warranty of the product by the U.S. Department of Agriculture and does not imply its approval to the exclusion of other products that also may be suitable. 2 Corresponding author. Phone: 785/776-2753. Fax: 785/776-2792. E-mail: fdowell @usgmrl.ksu.edu 3 Agricultural Engineering Department, Texas A&M University, College Station