Skip to main content
ARS Home » Research » Publications at this Location » Publication #144316


item Delwiche, Stephen - Steve
item Graybosch, Robert - Bob

Submitted to: Journal of Applied Spectroscopy
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 7/14/2003
Publication Date: 12/1/2003
Citation: Delwiche, S.R., Graybosch, R.A. 2003. Protein content and quality of wheat by near infrared reflectance: an examination of spectral pretreatments for PLS regression. Journal of Applied Spectroscopy. 57(12):1517-1527.

Interpretive Summary: The analytical technique known as near-infrared (NIR) spectroscopy has been widely used by the wheat industry for more than 20 years to measure basic nutiritional properties such as protein content. Its popularity stems from its rapidness (e.g., hundreds of samples per laboratory per day), accuracy, and ease of analysis. Of interest to the industry as well as to plant breeders is the application of this technology to estimate the functionality of intermediate products in cereals processing and the quality of the end products. Collectively described as wheat quality, these biochemically-routed properties are traditionally measured by electrophoretic or chromatographic procedures, which are time-consuming and difficult to perform. Hence, the prospect of using NIR spectroscopy for their evaluation has long been the desire of industry and breeders alike. An obstacle to the development of NIR procedures has been the identification of mathematical and statistical models that are tailored to each property. This study evaluated an automated inhouse computer procedure that allows the analyst to evaluate the potential of hundreds of mathematical transformations that are singly applied to the NIR spectra before a statistical regression procedure known as partial least squares (PLS) regression is performed. Using a set of nearly 400 samples of hard red winter and hard white wheat grown over two crop seasons, the procedure evaluated PLS regression models for protein content, sodium dodecyl sulfate sedimentation volume (an common protein quality index), and accumulated times that the developing plants were above or below critical temperatures (indicators of heat or cold stress). Our findings indicate that spectral pretreatments have widely varying affects on model accuracy, especially for protein quality indicators, as opposed to the protein content. Spectroscopists and plant breeders are the immediate beneficiaries of this work.

Technical Abstract: Use of near-infrared (NIR) diffuse reflectance on ground wheat meal for prediction of protein content is a well-accepted practice. Although protein content has a strong bearing on the suitability of wheat (Triticum aestivum L.) for processed foods, wheat quality, as largely influenced by the configuration and conformation of the monomeric and polymeric endosperm storage proteins is also of great importance to the food industry; however, the measurement of quality by NIR has been much less successful. The present study examines the effects and trends of applying mathematical transformations (pretreatments) to NIR spectral data before partial least squares (PLS) regression. Running mean smoothes, Savitzky-Golay second derivatives, multiplicative scatter correction, standard normal variate transformation, with and without detrending are systematically applied to an extensive set of hard red winter wheat and hard white wheat grown over two seasons. The modeled properties are protein content, sodium dodecyl sulfate (SDS) sedimentation volume, number of hours during grain fill at temperature < 24 deg. C, and number of hours during grain fill at temperature > 32 deg. C. The size of the convolution window used to perform a smooth or second derivative is also examined. The results indicate that for easily modeled properties such as protein content, the importance of pretreatment is lessened, whereas for the more difficult to model properties, such as SDS sedimentation volume, wide-window (>20 points) smooth or derivative convolutions were important in maximizing model performance. By averaging the best 50 PLS cross validation trial statistics (standard error) for each property, we have been able to ascertain the inherent modeling ability of each wheat property.