Nondestructive Detection of Citrus Greening by Near Infrared Spectroscopy

Liu Yande,Xiao Huaichun,Deng Qing,Zhang Zhicheng,Sun Xudong,Xiao Yusong

DOI: https://doi.org/10.11975/j.issn.1002-6819.2016.14.027

2016-01-01

Abstract:The feasibility was explored for identifying health, nutrient deficiency and citrus greening leaves based on near infrared (NIR) spectroscopy combined with machine learning methods. 232 samples were divided into the calibration and prediction sets for calibrating the models and accessing their performance according to the proportion of 3:1. The calibration set included citrus greening samples of 54, nutrient deficiency samples of 64 and healthy samples of 54. The prediction set included citrus greening samples of 21, nutrient deficiency samples of 17 and healthy samples of 22. The spectra of health, nutrient deficiency and citrus greening leaves were recorded in the wavelength range of 4 000-9 000 cm-1. After compared the representative spectra of health, nutrient deficiency and citrus greening, it was found that two significant differences appeared in the wavenumber bands of 5 100 and 6 880 cm-1. The peak around 6 880 cm-1 was caused by the stretching vibration of O-H first overtone of water and sugar. The difference between the spectra of health and citrus greening leaves was significant around 6 880 cm-1. The spectral intensity of citrus greening leaf was larger than health leaf. The ability of water absorption for citrus greening leaf was interfered with citrus greening. The peak around 5100 cm-1 was associated with the asymmetric vibration of N-H bond. Therefore, the spectral intensity of citrus greening leaf was lower than health leaf in the wavenumber of 5 100 cm-1. This may be related to the loss of nutrient elements in leaves of citrus greening. The study used different preprocessing methods as first derivative, smoothing and multiple scattered correction for spectral calibration. The preprocssing method of first derivative had removed baseline drift and enlarged the role of feature information. And the amplification characteristics of information can also lead to high frequency noise. Therefore, the further pretreatment was conducted by the method of smoothing. Then the scattering effect caused by the uneven thickness of the leaves was eliminated used the multiple scattering correction. Compared with other methods, it was found that the combination of first derivative, smoothing and multiple scatter correction can effectively eliminated the baseline drift and scattering phenomena. The machine learning methods of partial least square discriminate analysis (PLS-DA) and least square support vector machine (LS-SVM) were used to develop the classification models for identifying health, nutrient deficiency and citrus greening leaves. The principal component analysis (PCA) method was applied to optimize the input vectors of PLS-DA and LS-SVM models compared with full spectra. The first 14 and 11 principal components (PCs) were used to the input vectors for PLS-DA and LS-SVM models, respectively. And the regularization factor and the type of kernel function were optimized by the two-step grid search method. Compared to PLS-DA model, LS-SVM model yielded the best results with accuracy rate of 100% for identifying the health, nutrient deficiency and citrus greening. The kernel function type and regularization factor (γ) of the best LS-SVM model were linear kernel function and 2.25. The experimental results showed that it was feasible to identify health, nutrient deficiency and citrus greening leaves by NIR spectroscopy coupled with machine learning method of LS-SVM.