Regulatory agencies within the USDA have expressed an interest in sing
le kernel analysis of grain for determination of basic intrinsic prope
rties such as protein content. A study was undertaken to determine the
feasibility of measuring protein content of intact wheat kernels usin
g near-infrared transmittance spectrophotometry. Six wheat classes wer
e examined: hard red winter, hard red spring, hard white, soft red win
ter, soft white, and durum. Five samples per class were used, with eac
h sample represented by 96 randomly drawn kernels. Within each class,
single kernel spectra (850-1,050 nm) were calibrated to single kernel
protein content using four samples; the remaining sample was set aside
for model validation. The chemometric method of partial least squares
(PLS) analysis was used. For reference values, the protein content (N
X 5.7) of each kernel was measured by combustion. Typical model accur
acies of r(2) in the range of 0.85-0.93 and standard errors of predict
ion of 0.4-0.9% protein were determined on the validation sets. The nu
mber of PLS factors required for optimal model performance ranged from
6 to 14, dependent on wheat class and the type of mathematical transf
ormation on the spectra before modeling. Repeatability measurements on
kernels that were reloaded into the kernel clamp indicated that the v
ariance of model-determined protein content caused by kernel positioni
ng was largest for the smaller wheat kernels. Improvement in model per
formance could probably be attained by the averaging of transmittances
taken from several positions on each kernel.