Quality characteristics in rice by near-infrared reflectance analysis of whole-grain milled samples

Various physical and chemical tests exist to assess the cooking and processing characteristics of rice. As new lines of rice are developed in the United States and elsewhere, plant breeders routinely test for amylose content, alkali spreading value (an indicator of gelatinization temperature), protein content, viscosity properties of the flour-water paste, and the appearance of milled gains (whiteness, transparency, and degree of milling). A study was undertaken to determine the extent to which near-infrared reflectance (NIR) spectroscopy on whole-grain milled rice could be used to measure such characteristics. Samples of U.S. rices (n = 196) from advanced breeders' lines and commercial releases, representing conventional and specialty short-, medium-, and long-grain classes, were milled and scanned in the visible and near-IR regions (400-2,498 nm). Reference chemical and physical analyses were also performed on each sample. Results of partial least squares modeling indicated that reasonably accurate models were attained for apparent amylose content (standard error of prediction [SEP]= 1.3 percentage units; coefficient of determination on the validation set [r2] = 0.89), protein content (SEP = 0.13 percentage units. r2 = 0.97), whiteness (SEP = 0.60 percent reflectance, r2 = 0.97), transparency (SEP = 0.15 percent transmittance, r2 = 0.93), and milling degree (SEP = 2.7 dimensionless units on a 0-199 scale, r2 = 0.97). To a lesser extent, alkali spreading value could be modeled by NIR (SEP= 0.43 units on a 2-7 scale, r2 = 0.82), however, this accuracy is probably sufficient for initial screening in breeding programs. Conversely, models for the five flour paste viscosity properties recorded by a rapid visco analyzer (RVA) were not sufficiently accurate (r2 < 0.75) to warrant replacement of the RVA procedure with an NIR model. Reducing the sample size for NIR scanning from approximately 100 to approximately 8 g did not significantly affect the model performance of any constituent.