Detection of Mildew Degree of Wheat Using Terahertz Spectroscopy and Machine Learning

Abstract

In order to judge the mildew degree of wheat seeds quickly and accurately, this study proposed a qualitative analysis method for moldy wheat using terahertz time-domain (THz-TDS) spectroscopy combined with support vector (SVM), random forest (RF) or extreme learning machine (ELM). According to the content of aflatoxin B1 (AFB1), wheat seeds were classified into four types: normal, slight mildew, moderate mildew and severe mildew. Spectral data in the band of 0.1–4.0 THz were obtained using a CCT-1800 THz spectrometer. The effects of different spectral pretreatment methods on the results discrimination were examined, and three dimensionality reduction methods, principal component analysis (PCA), linear discriminant analysis (LDA), and t-distributed stochastic neighbor embedding (t-SNE), were used to reduce the dimension of spectral data. LDA was found to be the best method. Finally, a model based on SVM, RF or ELM was constructed. The SVM model had the best classification effect. When polynomial kernel function was chosen and error penalty coefficient was 1, the accuracy of discrimination was 98.61% and the root mean square error of prediction was 0.142 9. This study confirms that THz spectroscopy can be applied for accurate detection of wheat mildew, which can provide a detection approach for food safety and grain storage and detection.

Keywords

• terahertz time-domain spectroscopy; moldy wheat; spectral preprocessing; spectral dimension reduction; spectral classification