Application of quantitative non-destructive determination of protein in wheat based on pretreatment combined with parallel convolutional neural network

Abstract

With the increasing demand for wheat, the detection of wheat quality has become imperative. Protein content is an important indicator for wheat quality. Near infrared spectroscopy (NIRS) quantitative non-destructive testing technology has gained widespread application in agricultural field with the development of science and chemometrics technology. In this study, NIRS system was employed to measure the spectra of wheat, and the original spectra were pretreated using Savitzky-Golay smoothing (SG) pretreatment method. Subsequently, the NIRS prediction model of protein in wheat that using SG combined with parallel convolutional neural network (PaBATunNet) was established. PaBATunNet was composed of a one-dimensional convolutional layer, a parallel convolution module (Module), a flattening layer, four fully connected layers and a parameter regulator (PR). Module was made up of five submodules and a Concatenate function. The multidimensional features of the spectra were extracted by five submodules and spliced by Concatenate function. SG pretreatment combined with PaBATunNet (SG-PaBATunNet) was compared with commonly modeling methods, such as SG-partial least squares (SG-PLS), SG-principal component regression (SG-PCR), SG-support vector machine (SG-SVM) and SG-back propagation neural network (SG-BP). The results demonstrated that the modeling accuracy and prediction accuracy of SG-PaBATunNet were improved by 26.7%, 23.9%, 45.6%, 44.2%, and 38.4%, 39.6%, 60.1%, 58.0%, when compared with SG-PLS, SG-PCR, SG-SVM and SG-BP. The problems of low prediction accuracy and poor generalization ability with commonly modeling methods were effectively addressed by SG-PaBATunNet. This study provides an essential theoretical foundation for developing a fast, nondestructive and high-precision NIRS quantitative analysis model of protein in wheat.