摘要
Diseases are critical factors that affect the yield and quality of crops. Therefore, it is of great research value to develop rapid and quantitative methods for identification of common agricultural diseases. This exploratory study involved data analysis of common fungal pathogens using identification modeling based on terahertz spectrum technology. The selected pathogens were Physalospora piricola, Erysiphe cichoracearum, and Botrytis cinerea, which are common fungal pathogens that cause apple ring rot, cucumber powdery mildew, and grape gray mold blight, respectively. Taking polyethylene as the control, the terahertz time-domain spectra, and frequency-domain spectra of samples of the three pathogens were both measured. The absorption and refraction characteristics of these samples in the range of 0.1–2.0 THz were calculated and analyzed, and samples were then divided using the KS algorithm. Terahertz spectrum-image data blocks of the pathogen samples were preprocessed, and the dimensions of data were reduced using non-local mean filtering and the SPA algorithm, respectively. K-nearest neighbors (KNN), support vector machine (SVM), and BP neural network (BPNN), and other algorithms were used for analysis of terahertz images at characteristic frequencies, and for investigating the identification model. The model was quantitatively evaluated, and its imaging visualization was studied. The results suggest that there are significant differences among P. piricola, E. cichoracearum, and B. cinerea in absorption and refraction in the terahertz band. SVM modeling identification results of the three pathogens at the frequency of 1.376 THz were satisfactory, with an Rp of 0.9649, RMSEP of 0.0273, and a high (93.8212%) comprehensive evaluation index F1-score, and a clearly identifiable visualization effect. This study demonstrated the potential of terahertz spectroscopy to be used for identification of common crop pathogens and has provided technical references for the rapid diagnosis and early warning of agricultural diseases.