Least-squares Curve Fitting of Near Infrared Spectra Predicts Protein and Moisture Content of Ground Wheat

TLDR: In this paper, a linear combination of spectra of known constituents of ground wheat was used to approximate the near infrared (log 1/reflectance) spectra and the appropriate coefficients of these linear combinations were then linearly correlated with the protein and moisture content of the samples.

Abstract: Near infrared (log 1/reflectance) spectra of samples of ground wheat were approximated by a linear combination of spectra of known constituents, the approximation satisfying a least-squares criterion. The appropriate coefficients of these linear combinations were then linearly correlated with the protein and moisture content of the samples. Two extensions of previously reported curve-fitting techniques were made. First, multilinearly correlating several of the curve fit coefficients with the chemical data improved the standard errors. Second, using sample spectra as components, rather than pure constituent spectra, improved the standard errors to a point where they became comparable to those obtained by currently used derivative methods. The samples covered a protein range of 10 to 19%. Correlation coefficients reached 0.998 for protein, corresponding to a standard error of prediction of 0.15%. Parameters examined included spectral region, smoothing, and wavelength shifting. Results with reflectance spectra of sample sets with large particle size variation and high noise are also reported.