Variable selection methods to determine protein content in paddy using near-infrared hyperspectral imaging

The crude protein content is critical to quality assessment when screening for nutrients, taste quality and commercial value. Thus, this study performed a non-destructive and rapid prediction of protein content in paddy based on line-scanning near-infrared hyperspectral imaging (1001–2300 nm) technology. Partial least squares regression (PLSR), principal component regression (PCR) and multiple linear regression (MLR) predictive models were established to evaluate protein content (5.5037–8.2543 g 100 g −1 ) in 100 intact paddy, and the models achieved high performance. Spectral pre-processing with De-trending to a certain extent could enhance the smoothness of the spectrum and reduce spectral noise effectively. Successive projection algorithm (SPA) was used to extract characteristic wavelengths to simplify the models. A set of 18 feature variables were selected from the original wavelength, and the SPA-PLSR model has the best performance to predict protein content in paddy. In addition, the simplified performed model with a higher value of coefficient of determination (R 2 ) of R 2 C and R 2 P was 0.9078 and 0.8836, and the lower root mean square error (RMSE) of RMSEC and RMSEP was 0.0912 and 0.1675, respectively. The distribution maps of each sample protein contents in each pixel were obtained using the prediction model. Therefore, experimental results indicated the feasibility and possibility of a rapid and non-destructive hyperspectral imaging technique to detect the chemical component in paddy.