Rapid Determination of Total Dietary Fiber and Minerals in Coix Seed by Near-Infrared Spectroscopy Technology Based on Variable Selection Methods

The feasibility of near-infrared (NIR) spectroscopy for determining total dietary fiber (TDF) and mineral elements (mainly K, Mg, P, and S) in Coix seed was investigated. Partial least squares regression (PLSR) was applied to establish quantitative models. Norris derivative smoothing (NDS) was used as pretreatment method. A comparison of three variable selection methods, namely competitive adaptive reweighted sampling (CARS), genetic algorithms (GA), and random frog (RF), showed that CARS obtained the best performances of PLSR models with the effective wavelengths mainly concentrated on around 12,000-11,000 cm(-1) and 6500-3600 cm(-1). For the quantitative determination models of TDF, K, Mg, P, and S, the optimal root mean square error of prediction (RMSEP) values were 0.0923, 182.7224, 75.4987, 162.6993, and 36.6278; the r values were 0.95, 0.88, 0.80, 0.96, and 0.96; the residual predictive deviation (RPD) values were 2.68, 2.05, 1.70, 3.24, and 3.04, respectively. It is concluded that the NIR spectral technique has a potential to determine TDF, K, Mg, P, and S in Coix seed.