BP-ANN application to the model establishment of determination wheat protein using near infrared spectroscopy

Abstract

Near infrared spectroscopy calibration models of determination to the wheat protein concentrations were developed based on the back-propagation artificial neural network (BPANN). The spectra of 160 wheat samples were pretreated with detrend, normalization, then subjected to principal component analysis (PCA) to identify the principal components (PCs), the scores of the PCs were used as ANN input variables. Calibration models were established with the training set (80 samples) using various input variables and hidden nodes. The root mean square errors of prediction (RMSEPs) of the models to the prediction set (80 samples) were used to optimize the models. The RMSEP became stabilized when the input variables were up to 5, but changed little with varying hidden nodes. The optimal model with 9 input variables, 1 hidden node lead to the lowest RMSEP of 0.2869% and the highest correlation coefficient (R) value of 0.980 for the prediction set. Comparison of the calibration models developed with training sets of various sizes found that the simulation degree of model decreased slightly but prediction capacity improved with the increase of the training set data size and the optimal training set size was 80 samples. © 2006 IOP Publishing Ltd.