Premium
Environmental Effects on Developing Wheat as Sensed by Near‐Infrared Reflectance of Mature Grains
Author(s) -
Delwiche Stephen R.,
Graybosch Robert A.,
Nelson Lenis A.,
Hruschka William R.
Publication year - 2002
Publication title -
cereal chemistry
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.558
H-Index - 100
eISSN - 1943-3638
pISSN - 0009-0352
DOI - 10.1094/cchem.2002.79.6.885
Subject(s) - partial least squares regression , chemistry , standard error , linear regression , near infrared reflectance spectroscopy , standard deviation , relative humidity , near infrared spectroscopy , analytical chemistry (journal) , zoology , mathematics , environmental chemistry , statistics , meteorology , physics , quantum mechanics , biology
For 30 years, near‐infrared (NIR) spectroscopy has routinely been applied to the cereal grains for the purpose of rapidly measuring concentrations of constituents such as protein and moisture. The research described herein examined the ability of NIR reflectance spectroscopy on harvested wheat to determine weather‐related, quality‐determining properties that occurred during plant development. Twenty commercial cultivars or advanced breeding lines of hard red winter and hard white wheat ( Triticum aestivum L.) were grown in 10 geographical locations under prevailing natural conditions of the U.S. Great Plains. Diffuse reflectance spectra (1,100–2,498 nm) of ground wheat from these samples were modeled by partial least squares one (PLS1) and multiple linear regression algorithms for the following properties: SDS sedimentation volume, amount of time during grain fill in which the temperature or relative humidity exceeded or was less than a threshold level (i.e., >30, >32, >35, <24°C; >80%, <40% rh). Rainfall values associated with four pre‐ and post‐planting stages also were examined heuristically by PLS2 analysis. Partial correlation analysis was used to statistically remove the contribution of protein content from the quantitative NIR models. PLS1 models of 9–11 factors on scatter‐corrected and (second order) derivatized spectra produced models whose dimensionless error (RPD, ratio of standard deviation of the property in a test set to the model standard error for that property) ranged from 2.0 to 3.3. Multiple linear regression models, involving the sum of four second‐derivative terms with coefficients, produced models of slightly higher error compared with PLS models. For both modeling approaches, partial correlation analysis demonstrated that model success extends beyond an intercorrelation between property and protein content, a constituent that is well‐modeled by NIR spectroscopy. With refinement, these types of NIR models may have the potential to provide grain handlers, millers, and bakers a tool for identifying the cultural environment under which the purchased grain was produced.