Near-infrared spectroscopy analysis of compound fertilizer based on GAF and quaternion convolution neural network

Abstract

This paper combines near infrared spectroscopy with deep learning theory to propose a rapid identification method for compound fertilizer based on gramian angular field (GAF) image coding and quaternion convolutional neural network (QCNN). First, the near-infrared (NIR) spectra of 200 samples of two types of compound with and without polyglutamic acid (γ-PGA) were collected and pretreated by multivariate scattering correction(MSC) and first derivative methods. Then, the one-dimensional (1D) NIR spectra were transformed into two-dimensional (2D) images by GAF coding method to express the spectral information more intuitively. Finally, three images of Gramian angular difference field (GADF), Gramian angular summation field (GASF) and their average images were formed into a pure quaternion matrix for parallel representation and were analyzed by the designed QCNN. The peak features and minor features were mined based on the automatic learning function of multi-layer structure of QCNN to establish a high-performance, qualitative model for identifying compound fertilizer. The experimental results show that the classification accuracy, sensitivity, and specificity of the identification model were 95.84%, 95.96% and 94.25%, respectively. Compared with the classification results based on traditional partial least square discrimination analysis(PLS-DA), principal component analysis combing with support vector machine classification(PCA-SVM), 1D convolution neural network (1DCNN) and GAF-CNN methods, the classification accuracy and adaptability of the model based on the proposed GAF-QCNN method have been significantly improved. The combination of GAF image of NIR spectra and QCNN developed in this study provides a novel approach for the intelligent modeling algorithm of NIR spectroscopy technology.