Near-Infrared Spectroscopy with Data Analysis Model for Rapid Identification of Huanglongbing in Gannan Navel Orange

Huanglongbing (HLB) is a devastating disease of citrus. Near-infrared spectroscopy (NIRS) combined with the data analysis model was proposed to nondestructively diagnose HLB in Gannan navel orange. The near-infrared spectra of four kinds of Gannan navel orange leaves including slight HLB, serious HLB, nutrient deficiency and health were collected in this work. Principal components analysis (PCA) combined with linear discriminant analysis (LDA), k-nearest neighbors (KNN), support vector machine (SVM), and back propagation neural network (BPNN) were used to build the discriminant analysis models. The results showed that the PCA combined with BPNN model was superior to others. The discriminant accuracies for four kinds of samples were 100.00%, 99.17%, and 95.83% in the training set, validation set, and test set, respectively, and the overall discriminant accuracy reached 98.75%.

Introduction

Navel oranges are popular among consumers because of their bright color, crisp flesh, sweet flavor, and rich nutrition. Nowadays, Ganzhou in the south of Jiangxi province of China has the world’s highest planting area and the global third greatest yield of navel oranges. It is worth mentioning, Gannan navel orange brings substantial economic benefits for local income, and it has been listed as one of the products of national geographical indication. Unfortunately, navel oranges have been plagued by Huanglongbing (HLB) disease that caused significant loss to the citrus industry globally. HLB is caused by the phloem-limited and psyllid-vectored ‘Candidatus Liberibacter’ spp. (1). Furthermore, the pathogen can be spread from infected citrus to healthy citrus by grafting, the Asian citrus psyllid (Diaphorina citri), and dodder (2). Typically, trees infected with HLB do not show symptoms immediately. There is an incubation period of several months and even many years. The symptoms of serious HLB-infected trees are yellow new leaves, misshapen fruits, and blotchy mottled leaves, which can lead the tree to death eventually (3). At present, there is no effective solution to cure diseased fruit trees. Once the tree is infected, it must be cut down, dug, and burned to prevent the further spread of HLB. For a long time, HLB has been regarded as the “cancer” of citrus. Under this circumstance, it is urgent to diagnose citrus HLB especially in the early period of this disease.

Some traditional methods including field diagnosis, starch coloration, and indicator plant were used commonly for detecting citrus HLB (4). These methods are simple and fast, but there may be missed detection, or easy to be confused with other diseases. In addition, microscopic techniques and serology techniques have been developed for detecting HLB disease (5, 6). However, there is the possibility of false negatives from sampling because HLB pathogen is distributed unevenly in plant tissues and can fluctuate with time. The most reliable method for HLB diagnosis is the nucleic acid molecular detection method (7, 8), such as polymerase chain reaction (PCR) for the bacterial deoxyribonucleic acid (DNA). Nevertheless, it is costly and destructive to samples. Although PCR can achieve high accuracy, it is restricted by factors such as expensive equipment, high cost, and professional operations. It must be pointed out that family-based management is decentralized and the added value of products is low under the current situation in China. In this case, it has great significance to research a rapid and environmentally friendly diagnosis method of citrus HLB especially in the early period of the disease.

Near-infrared spectroscopy (NIRS) provides a rapid, non-invasive sensing tool for the detection of disease, biotic and abiotic stresses in plants (9–12). In the field of citrus HLB, Sankaran et al. investigated the applicability of Vis-NIRS for detection of HLB in citrus canopy leaves with an average classification accuracy rate of 87% by using support vector machine (SVM) (13). Windham et al. utilized NIR reflectance spectroscopy to detect HLB of orange and grapefruit leaves and the accuracy rate was ranging from 92% to 99% by using partial least squares (PLS) regression models (14). Liu et al. reported visible light and near-infrared spectra combined with PLS-DA to detect citrus Huanglongbing, which accuracy rate reached 97% (15). Xu et al. studied the LIBS-NIR joint spectra of HLB-infected Gannan navel orange leaves, and the accuracy was 89.5% and 95.7% in training and test sets respectively by using principal components analysis (PCA) and multilayer perceptron (MLP) (16).

The above methods provided valuable references in the analysis of citrus HLB. However, navel orange HLB detection is influenced by varieties, territory, matrix effect, and other factors. It is worth mentioning that the slight HLB without obvious symptoms is usually similar to healthy navel orange and the serious HLB with yellowing symptoms is similar to nutrient deficiency navel orange. Therefore, the model stability and generalization ability may be insufficient when faced with the influences. Based on this, Gannan navel orange leaves were selected to confirm the feasibility of NIRS in the detection of HLB. This study analyzed the near-infrared spectrum characteristics of four types of navel orange leaves including health, slight HLB, serious HLB, and nutrient deficiency (lack of K and Mg). Multiple classification discriminant models combined with PCA were constructed and analyzed comparatively. It is expected to provide a stable and reliable reference method for the rapid and nondestructive detection of navel orange HLB.

Experimental

Sample Preparation

Four kinds of leaves samples with health, slight HLB, serious HLB, and nutrient deficiency (lack of K and Mg elements) were collected from Newhall Gannan navel orange trees in the navel orange orchard (Xinfeng, Ganzhou, Jiangxi, China) in July 2019 under the guidance of experienced horticultural experts. In order to obtain the heterogeneous representative samples, the navel orange orchard was divided into four zones, three orange trees (each tree was about 10 meters apart) were randomly selected from each zone to pick the same layer tree branches in four directions (east, south, west, and north). All fresh leaves were cleaned the surface with deionized water and dried naturally at room temperature, then marked and stored to categorize test in a refrigerator at 4 °C. All waste leaves were processed by sterilization treatment to prevent the spread of disease.

Acquisition of Spectral Data

A Fourier transform near-infrared spectrometer (Antaris II, Thermo Fisher Scientific, USA) was used to collect the sample spectrum. The acquisition range was 4000~10000 cm^-1 with a resolution of 8 cm^-1. Each spectrum was acquired in integrating sphere (diffuse reflectance) mode as the average of 32 co-added scans. The reflectance values (R) were converted into the absorbance values (A) using the formula A=log (1/R). The spectral interval was 3.856 cm^-1, and each spectrum contained 1557 variables. The average of the three spectra collected from three different positions of each leaf by avoiding the veins of the leaves at an indoor temperature of 25 ± 1 °C and a relative humidity of 50 ± 1% was used for further analysis.

PCR Identification and Main Chemical Contents Determination

After spectral data collection, the leaves were investigated using lab-based methods. Primer A2/J5 was applied in the polymerase chain reaction (PCR) to identify HLB leaves.

The sequences of A2/J5 were 5’-TATAAAGGTTGACCTTTCGAGTTT-3’/5’-ACAAAAGCAGAAATAGCACGAACAA-3’ (17).

Main chemical contents were detected to analyze the differences of nutrient contents in leaves. Chlorophyll and soluble sugar were detected according to the references of Shinano et al. (18) and Wang et al. (19). Besides, total nitrogen (N), total phosphorus (P), and potassium (K) were determined according to Chinese Agricultural Industry Standard NY/T 2017-2011 (Determination of Nitrogen, Phosphorus and Potassium in Plants), and magnesium (Mg) was determined using a method adapted from this standard.

Sample Set Division

Combining the result of PCR with main chemical contents determination, 120 leaf samples were selected from each kind of leaves, and a total of 480 leaf samples were obtained. Before data processing, all leaf samples were divided into the training set, validation set, and test set with the ratio of 2:1:1 by Kennard-Stone algorithm (20). The number of leaf samples in the training set, validation set, and test set were 240, 120, and 120 respectively.

Data Analysis

Principal Components Analysis

Principal components analysis (PCA) can compress the original spectral variables into a linear combination of multiple principal components. It can reduce the data dimension while retaining most of the original data features, making the data set easier to use. The variables obtained after processing are called the principal components (PCs) and the information between the PCs is independent of each other. The PCA equation used is as follows (21).

where X is the spectral variable of the sample, T is the sample score variable, P is the loading matrix, and E is the residual matrix.

Linear Discriminant Analysis

Linear discriminant analysis (LDA) generalizes Fisher's linear discriminant method, which is a commonly used classification method. There are three steps to achieve LDA classification (22). The first step is to calculate the between-class variance. The second step is to calculate the within-class variance. The third step is to maximize the between-class variance and minimize the within-class variance. The transformation of data sets and the classification of test vectors are achieved by class-dependent transformation or class-independent transformation (23). Class-dependent transformation was applied in the research. However, LDA is influenced by the Small Sample Size (SSS) and linearity problems sometimes (24).

K-Nearest Neighbors

K-nearest neighbors (KNN) is a pattern recognition method based on vector space. It classifies a sample by calculating the distance between the test sample and the training data sample (25). The distance metric is most commonly the Euclidean distance (26). K is the key parameter that determines the classification result. In this paper, K was varied from 1 to 10 in steps of 1. The application of KNN is as follows (27).

where x′ is a test example, N_k is the set of the k nearest neighbors, and the indicator function I(y_i = y) indicates that if x_i belongs to class y, it returns 1 (and 0 otherwise).

Support Vector Machine

Support vector machine (SVM) can be used for both linear and nonlinear classification and regression. The main idea of SVM is to establish a classification hyperplane to maximize the distance between positive and negative examples (28). SVM models were built using an open-source software package (LIBSVM 3.24). The radial basis function (RBF) kernel was adopted and the parameters were optimized by cross-validation in this paper (29). Two parameters (penalty parameter c and kernel function parameter g) can affect the accuracy of classification in the model, and the accuracy of the model can be improved effectively by optimizing the parameters. The RBF kernel function was as follows (30).

where (x_i, y_i) is the training sample set and σ is the kernel width parameter (g = 1/2σ²).

Back Propagation Neural Network

Back propagation neural network (BPNN) is a two-layer feedforward network with hidden neurons and output neurons. It simulates the neural signal conduction of the human brain to model complex variation (31). Available training patterns, examples, or input-output relations can be learned by neural networks (32). Therefore, it has obvious advantages in solving nonlinear and large-scale computing problems (33). Neural Pattern Recognition (nprtool) in MATLAB R2016a was used for modeling. During the modeling process, the top PCs were extracted as the input of BPNN model, the number of input PCs and the number of hidden neurons were adjusted to obtain the better discrimination accuracy, the number of neurons in the output was set to 4 (Four kinds of leaf samples), the transfer function of the hidden layer was set as the hyperbolic tangent sigmoid (tansig) function, the learning rate factor and momentum factor were all set to 0.1, the initial weight was set to 0.3, and the maximum number of training epochs was set to 1000.

Software

NIR spectra of the leaves samples were acquired and stored by software (RESULT software, Thermo Fisher Scientific, USA). Data difference analysis was carried out in SPSS Statistics 26.0 software (IBM Corp, Armonk, NY, USA) and all other algorithms were implemented in MATLAB R2016a (MathWorks Inc., Natick, MA, USA).

Results and Discussion

The Result of PCR and Main Chemical Contents

According to the results of PCR products after agarose gel electrophoresis, there were no bands of the leaf samples with health and nutrient deficiency. The bands of serious HLB leaf samples were brighter and wider compared with slight HLB. Furthermore, the contents of K and Mg in nutrient deficiency leaves were significantly lower than in healthy leaves. The contents of chlorophyll and Mg in HLB-infected leaves were significantly lower than in healthy leaves. However, soluble sugar content in HLB-infected leaves was significantly higher than in healthy leaves. The degree of losing chlorophyll and Mg as well as carbohydrate accumulation in serious HLB leaves was heavier than slight HLB leaves. The differences of main chemical contents in four kinds of leaves provided the basis for nondestructive rapid detection by near-infrared spectroscopy.

Combining the results of PCR and main chemical contents determination. Four types of leaves sample data sets were constructed for nondestructive discrimination and classification detection by near-infrared spectroscopy. Health subset was PCR-negative without symptoms of HLB; slight HLB subset was PCR-positive and asymptomatic; serious HLB subset was PCR-positive and symptomatic; nutrient deficiency subset was PCR-negative with yellowing symptoms and lacking K and Mg elements. After the four subsets of samples were verified, the subsequent analysis of spectral data could be performed.

Analysis of NIRS Characteristics of Navel Orange Leaves

The original near-infrared spectra of the four types (health, slight HLB, serious HLB, and nutrient deficiency) navel orange leaves were shown in Figure 1. There were obvious absorption bands near 6900 cm^-1 and 5200 cm^-1. The peak at 6900 cm^-1 is attributed to the first overtone of O-H stretching, with contributions from C-H combination bands (34, 35). The O-H bond was derived from water and starch in leaves, while the C-H bond was derived from sugar possibly. It may be that the clogging of the sieve tissue in the HLB-infected leaves led to the accumulation of starch and other sugars (36), which made the absorption peak of HLB-infected leaves higher than healthy leaves. The band near 5200 cm^-1 is attributed mainly to the O-H stretching plus bending combination band of water, with contributions from N-H combination bands (37). The O-H bond was derived from water in leaves, while the N-H bond was derived from protein in the leaves possibly. The absorption peak of HLB-infected leaves was higher than healthy leaves, which was caused by the increase in the expression of related amino acids in the leaves after being infected with HLB possibly (38). In addition, the loss of nutrient elements in nutrient deficiency leaves caused its absorption peaks to be lower than that of healthy leaves.

Figure 1: The original near-infrared spectra of four kinds of Gannan navel orange leaves. The spectral range was 4000~10000 cm^-1 with a resolution of 8 cm^-1. Each spectrum was an average of 32 scans. The leaves had obvious absorption bands near 6900 cm^-1 and 5200 cm^-1.

PCA Qualitative Analysis

PCA was used to analyze the original near-infrared spectrum. The first three principal component scores were taken as a scatter plot. As shown in Figure 2, the spectral PCs of four types of leaves including health, slight HLB, serious HLB, and nutrient deficiency, were overlapped and could not be distinguished distinctly. The variance contribution rates of the first three principal components were 90.84%, 6.15%, and 0.69% respectively, and the cumulative variance contribution rate was 97.68%. Effective discrimination could not be achieved with PCA alone, so it was combined with other methods for follow-up analysis.

Figure 2: Scatter plot of the first three PC scores of near-infrared spectra on four kinds of Gannan navel orange leaves. The replication samples were 120 of each kind of leaves.

LDA Model

The top PCs were selected for LDA modeling analysis. As shown in Figure 3, the recognition rate of the training set and test set increased gradually as the PCs increased. When the PCs was 10, the model was superior to others, and the recognition rates for four kinds of Gannan navel orange leaves in the training set, validation set, and test set were 93.75%, 61.67%, and 76.67%, respectively. When the linear relationship of the spectral variables is weak, the effect of the LDA may be poor.

Figure 3: The recognition accuracy rate analyzed by LDA of near-infrared spectra on four kinds of Gannan navel orange leaves with slight HLB, serious HLB, nutrient deficiency and health in training set, validation set and test set.

KNN Analysis Model

The top PCs were selected to construct the KNN analysis model. The results in Figure 4 showed that when the number of PCs was 9, the k value of the training set, validation set, and test set were 1, 4, and 4, KNN model had the best discrimination result. The recognition rates for four kinds of Gannan navel orange leaves in the training set, validation set, and test set were 74.17% (Figure 4a), 50.00% (Figure 4b), and 69.17% (Figure 4c). The construction of the model is closely related to the degree of dispersion of the sample data. Therefore, the insufficient degree of dispersion of data could result in low recognition rates in different data sets.

Figure 4: The recognition accuracy rate analyzed by KNN of near-infrared spectra on four kinds of Gannan navel orange leaves with slight HLB, serious HLB, nutrient deficiency and health in training set (a), validation set (b) and test set (c).

SVM Analysis Model

The top PCs were selected for SVM classification modeling. By cross-validation, when c=171.896 and g=0.020, the SVM model was relatively good. Figure 5 showed that the total accuracy increased gradually with the increase of PCs before the PCs were 12. When the PCs continued to increase, the recognition rate showed a trend of decline. Therefore, when the PCs were 12, the recognition rates in the training set, validation set, and test set reached 91.25%, 59.17%, and 82.50%, respectively. Sankaran et al. (13) used Vis-NIR combined with SVM for HLB detection in citrus trees. They found that SVM model generated the best classification performance with an average overall accuracy of about 87%. It can be seen that SVM has a certain potential for distinguishing HLB samples from healthy samples. Nevertheless, when the sample types are various with low differentiation, the effect may not be ideal.

Figure 5: The recognition accuracy rate analyzed by SVM of near-infrared spectra on four kinds of Gannan navel orange leaves with slight HLB, serious HLB, nutrient deficiency and health in training set, validation set and test set.

BPNN Analysis Model

The top PCs were selected to construct the BPNN model. The results showed that as the number of input PCs increased from 3 to 10, the total recognition accuracy generally increased (from 95.42% to 98.75%), with the optimal number of hidden-layer neurons ranging from 6 to 9. When the number of PCs of the input layer exceeded 10, the accuracy of model recognition decreased.

Table 1: The recognition accuracy rate analyzed by BPNN of near-infrared spectra on four kinds of Gannan navel orange leaves under different principal components and hidden layer neurons.

Note: Results were calculated 10 times to verify that every sample would be chosen.

As shown in Table 1, when the number of PCs of the input layer was 10 and the number of neurons in the hidden layer was 9, the BPNN model was superior to others. The recognition rates in the training set, validation set, and test set were 100.00%, 99.17%, and 95.83% respectively, with an overall recognition accuracy of 98.75%. Analogously, Xu et al. (16) found that an MLP-PCA model identified HLB-infected leaves more accurately than a DA-PCA model. It can be seen that neural networks have a preferable ability in classification. The discriminant results of the BPNN model for different subsets of samples were further analyzed. Practically, only a small number of samples were misjudged in the test set. The misjudged samples were mainly concentrated between the HLB and the nutrient deficiency navel orange leaves. Due to the lack of nutrients during the growth process of navel orange trees with HLB or nutrient deficiency, the physical and chemical index contents of the leaves were relatively close. Therefore, these samples often lead to some errors in the identification process.

In order to test the stability and generalization ability of the BPNN model, 60 leaf samples in each subset were collected from a different navel orange orchard (Ganxian, Ganzhou, Jiangxi, China) in August 2020 for external test in this work. The accuracy rate of the external test was lower than that of the internal test set; misjudgments were mainly caused by HLB and nutrient deficiency. There was inherently a high degree of similarity between these misclassified leaves. When the leaf samples were collected from different navel orange orchards with different varieties, soil conditions, and other factors, misclassification could easily occur. Therefore, more leaf samples from different sources will be utilized for model optimization of the NIRS detection of HLB in the subsequent study.

Compared with the LDA, KNN, and SVM models in Table 2, the BPNN model was superior to others. The growth of navel orange trees is affected by many environmental factors (such as soil properties, fertilizer management, time of light, temperature, growth situation, etc). In addition, HLB can clog the tissue fiber channels in navel orange trees leading to nutritional deficiency, leading the tree to death eventually. Hence, the rapid detection of the content of these components in navel orange trees is crucial to assess their physiological status. The different components content in navel orange trees can be related to certain absorption bands in near infrared spectroscopy (NIR). Thus, there exists a certain correlation between the content of components in navel orange trees and NIR data (see Figure 1). Besides, the change of the content of these components in navel orange leaves during the tree growth process is a dynamic process. Hence, the relationship between the content of these components in navel orange leaves and the characteristic variables from NIR data is very complicated and tends to be nonlinear. BPNN and SVM are nonlinear algorithms, whereas LDA is a linear method and KNN is a nonparametric distance-based classifier. As the change of multi-components in navel orange leaves in different physiological status is a complex dynamic process, and their spectral differences are also very complex, a linear algorithm alone cannot provide a solution to such a calculation problem. In general, a non-linear method is stronger than a linear method in terms of the level of self-learning and self-adjustment; besides, the topological network architecture of BPNN can be more suitable for the solution to this calculation problem compared with SVM in this work. Consequently, the BPNN model produced better performance than the other methods.

Table 2: The discrimination comparison results of LDA, KNN, SVM and BPNN.

Note: Results were calculated 10 times to verify that every sample would be chosen.

Conclusion

In this study, near-infrared spectroscopy combined with different pattern recognition models was explored to detect HLB of Gannan navel orange leaves. Compared with the LDA, KNN, and SVM models, the BPNN model had the highest accuracy with an overall recognition accuracy rate reached 98.75%. It showed that the PCA combined with BPNN classification model had high efficiency; however, its robustness to samples from other orchards still needs improvement. The application of near-infrared spectroscopy combined with the BPNN method is feasible for the detection of citrus Huanglongbing. It is a nondestructive and rapid method that has the potential for real-time, in situ measurement and could significantly improve the efficiency of disease diagnosis and control in citrus and other plants. It also provides the basis for the exploitation of portable near-infrared spectroscopy and offers a new approach for practical disease detection in plants.

Acknowledgments

The authors gratefully acknowledge the planter Lixin Gong, who provided the Gannan navel orange leaf samples. The research work was funded by Major Projects of Youth Science Foundation of Jiangxi Province (no. 2014ACB21013) and the Natural Science Foundation of Jiangsu Province (no. BE2017302).

References

1. Li, W.; Yao, Y.-N.; Wu, L.; Hu, B. Detection and Seasonal Variations of Huanglongbing Disease in Navel Orange Trees Using Direct Ionization Mass Spectrometry. J. Agric. Food Chem. 2019, 67 (8), 2265–2271. DOI: 10.1021/acs.jafc.8b06427
2. Bové, J. M. Huanglongbing: A Destructive, Newly-Emerging, Century-Old Disease of Citrus. J. Plant Pathol. 2006, 88 (1), 7–37. URL: https://www.sipav.org/en/16/Archive/828
3. Gottwald, T. R.; da Graça, J. V.; Bassanezi, R. B. Citrus Huanglongbing: The Pathogen and Its Impact. Plant Health Prog. 2007, 8 (1), 31. DOI: 10.1094/PHP-2007-0906-01-RV
4. Takushi, T.; Toyozato, T.; Kawano, S.; et al. Scratch Method for Simple, Rapid Diagnosis of Citrus Huanglongbing Using Iodine to Detect High Accumulation of Starch in the Citrus Leaves. Jpn. J. Phytopathol. 2007, 73 (1), 3–8. DOI: 10.3186/jjphytopath.73.3
5. Achor, D.; Etxeberria, E.; Wang, N.; et al. Sequence of Anatomical Symptom Observations in Citrus Affected with Huanglongbing Disease. Plant Pathol. J. 2010, 9 (2), 56–64. DOI: 10.3923/ppj.2010.56.64
6. Garnier, M.; Jagoueix-Eveillard, S.; Cronje, P. R.; Le Roux, H. F.; Bové, J. M. Genomic Characterization of a Liberibacter Present in an Ornamental Rutaceous Tree, Calodendrum capense, in the Western Cape Province of South Africa. Proposal of 'Candidatus Liberibacter africanus subsp. capensis'. Int. J. Syst. Evol. Microbiol. 2000, 50 (6), 2119–2125. DOI: 10.1099/00207713-50-6-2119
7. Cellier, G.; Redondo, C.; Cubero, J.; et al. Comparison of the Performance of the Main Real-Time and Conventional PCR Detection Tests for 'Candidatus Liberibacter' spp., Plant Pathogenic Bacteria Causing the Huanglongbing Disease in Citrus spp. Eur. J. Plant Pathol. 2020, 157 (4), 919–941. DOI: 10.1007/s10658-020-02052-3
8. Park, J.-W.; Louzada, E. S.; Braswell, W. E.; et al. A New Diagnostic Real-Time PCR Method for Huanglongbing Detection in Citrus Root Tissue. J. Gen. Plant Pathol. 2018, 84, 359–367. DOI: 10.1007/s10327-018-0793-4
9. Sankaran, S.; Mishra, A.; Ehsani, R.; Davis, C. A Review of Advanced Techniques for Detecting Plant Diseases. Comput. Electron. Agric. 2010, 72 (1), 1–13. DOI: 10.1016/j.compag.2010.02.007
10. Sheng, X.; Li, Z.; Li, Z.; et al. Nondestructive Determination of Lignin Content in Korla Fragrant Pear Based on Near-Infrared Spectroscopy. Spectrosc. Lett. 2020, 53 (4), 306–314. DOI: 10.1080/00387010.2020.1740276
11. Kumar, S.; Lahlali, R.; Liu, X.; Karunakaran, C. Infrared Spectroscopy Combined with Imaging: A New Developing Analytical Tool in Health and Plant Science. Appl. Spectrosc. Rev. 2016, 51 (6), 466–483. DOI: 10.1080/05704928.2016.1157808
12. Liang, P.-S.; Haff, R. P.; Hua, S.-S. T.; et al. Nondestructive Detection of Zebra Chip Disease in Potatoes Using Near-Infrared Spectroscopy. Biosyst. Eng. 2018, 166, 161–169. DOI: 10.1016/j.biosystemseng.2017.11.019
13. Sankaran, S.; Maja, J. M.; Buchanon, S.; Ehsani, R. Huanglongbing (Citrus Greening) Detection Using Visible, Near Infrared and Thermal Imaging Techniques. Sensors 2013, 13 (2), 2117–2130. DOI: 10.3390/s130202117
14. Windham, W. R.; Poole, G. H.; Park, B.; et al. Rapid Screening of Huanglongbing-Infected Citrus Leaves by Near-Infrared Reflectance Spectroscopy. Trans. ASABE 2011, 54 (6), 2253–2258. DOI: 10.13031/2013.40638
15. Liu, Y.; Xiao, H.; Sun, X.; et al. Study on the Quick Non-Destructive Detection of Citrus Huanglongbing Based on the Spectrometry of VIS and NIR. Spectrosc. Spectral Anal. 2018, 38 (2), 528–534.
16. Xu, F.; Hao, Z.; Huang, L.; et al. Comparative Identification of Citrus Huanglongbing by Analyzing Leaves Using Laser-Induced Breakdown Spectroscopy and Near-Infrared Spectroscopy. Appl. Phys. B 2020, 126, 43. DOI: 10.1007/s00340-020-7392-8
17. Hocquellet, A.; Toorawa, P.; Bové, J. M.; Garnier, M. Detection and Identification of the Two Candidatus Liberobacter Species Associated with Citrus Huanglongbing by PCR Amplification of Ribosomal Protein Genes of the β Operon. Mol. Cell. Probes 1999, 13 (5), 373–379. DOI: 10.1006/mcpr.1999.0263
18. Shinano, T.; Lei, T. T.; Kawamukai, T.; et al. Dimethylsulfoxide Method for the Extraction of Chlorophylls a and b from the Leaves of Wheat, Field Bean, Dwarf Bamboo, and Oak. Photosynthetica 1996, 32 (3), 409–415. URL: https://www.researchgate.net/publication/285785079
19. Wang, S.; Zhu, Y.; Jiang, H.; Cao, W. Positional Differences in Nitrogen and Sugar Concentrations of Upper Leaves Relate to Plant N Status in Rice under Different N Rates. Field Crops Res. 2006, 96, 224–234. DOI: 10.1016/j.fcr.2005.07.008
20. Morais, C. L. M.; Santos, M. C. D.; Lima, K. M. G.; Martin, F. L. Improving Data Splitting for Classification Applications in Spectrochemical Analyses Employing a Random-Mutation Kennard-Stone Algorithm Approach. Bioinformatics 2019, 35 (24), 5257–5263. DOI: 10.1093/bioinformatics/btz421
21. Dong, Y.; Sørensen, K. M.; He, S.; Engelsen, S. B. Gum Arabic Authentication and Mixture Quantification by Near Infrared Spectroscopy. Food Control 2017, 78, 144–149. DOI: 10.1016/j.foodcont.2017.02.002
22. Tharwat, A.; Gaber, T.; Ibrahim, A.; Hassanien, A. E. Linear Discriminant Analysis: A Detailed Tutorial. AI Commun. 2017, 30 (2), 169–190. DOI: 10.3233/AIC-170729
23. Balakrishnama, S.; Ganapathiraju, A. Linear Discriminant Analysis—A Brief Tutorial; Institute for Signal and Information Processing, Mississippi State University, 1998. URL: https://isip.piconepress.com/publications/reports/1998/isip/lda/lda_theory_v1.0.pdf
24. Baek, I.; Kim, M. S.; Cho, B.-K.; et al. Selection of Optimal Hyperspectral Wavebands for Detection of Discolored, Diseased Rice Seeds. Appl. Sci. 2019, 9 (5), 1027. DOI: 10.3390/app9051027
25. Yang, R.; Kan, J. Classification of Tree Species at the Leaf Level Based on Hyperspectral Imaging Technology. J. Appl. Spectrosc. 2020, 87 (1), 184–193. DOI: 10.1007/s10812-020-00981-9
26. Zhang, Y.; Wang, J. K-Nearest Neighbors and a Kernel Density Estimator for GEFCom2014 Probabilistic Wind Power Forecasting. Int. J. Forecast. 2016, 32 (3), 1074–1080. DOI: 10.1016/j.ijforecast.2015.11.006
27. Kramer, O. K-Nearest Neighbors. In Dimensionality Reduction with Unsupervised Nearest Neighbors; Intelligent Systems Reference Library, Vol. 51; Springer: Berlin, Heidelberg, 2013; pp 13–23. DOI: 10.1007/978-3-642-38652-7_2
28. Yu, R.; Zhu, X.; Cao, S.; et al. Estimation of Chlorophyll Content in Apple Leaves Based on Imaging Spectroscopy. J. Appl. Spectrosc. 2019, 86 (3), 457–464. DOI: 10.1007/s10812-019-00841-1
29. Feng, G. H. Comparison of SVM Classification Kernel Functions and Parameter Selection. Comput. Eng. Appl. 2011, 47 (3), 123–124, 128 (in Chinese). URL: http://www.cqvip.com/qk/91690x/201103/36751280.html
30. Gholami, R.; Fakhari, N. Support Vector Machine: Principles, Parameters, and Applications. In Handbook of Neural Computation; Samui, P.; Roy, S. S.; Balas, V. E., Eds.; Academic Press, 2017; pp 515–535. DOI: 10.1016/B978-0-12-811318-9.00027-2
31. Gabriëls, S. H. E. J.; Mishra, P.; Mensink, M. G. J.; et al. Non-Destructive Measurement of Internal Browning in Mangoes Using Visible and Near-Infrared Spectroscopy Supported by Artificial Neural Network Analysis. Postharvest Biol. Technol. 2020, 166, 111206. DOI: 10.1016/j.postharvbio.2020.111206
32. Islam, M.; Chen, G.; Jin, S. An Overview of Neural Network. Am. J. Neural Netw. Appl. 2019, 5 (1), 7–11. URL: https://www.researchgate.net/publication/337137421_An_Overview_of_Neural_Network
33. Ma, H.; Liu, S. The Potential Evaluation of Multisource Remote Sensing Data for Extracting Soil Moisture Based on the Method of BP Neural Network. Can. J. Remote Sens. 2016, 42 (2), 117–124. DOI: 10.1080/07038992.2016.1160773
34. Workman, J., Jr.; Weyer, L. Practical Guide and Spectral Atlas for Interpretive Near-Infrared Spectroscopy, 2nd ed.; CRC Press: Boca Raton, FL, 2012; pp 77–85. DOI: 10.1201/b11894
35. Ciurczak, E. W.; Igne, B.; Workman, J., Jr.; Burns, D. A., Eds. Handbook of Near-Infrared Analysis, 4th ed.; CRC Press: Boca Raton, FL, 2021; pp 57–64. URL: https://www.taylorfrancis.com/books/9781351269889
36. Weng, H.; Lv, J.; Cen, H.; et al. Hyperspectral Reflectance Imaging Combined with Carbohydrate Metabolism Analysis for Diagnosis of Citrus Huanglongbing in Different Seasons and Cultivars. Sens. Actuators, B 2018, 275, 50–60. DOI: 10.1016/j.snb.2018.08.020
37. Zou, X.; Zhao, J.; Povey, M. J. W.; Holmes, M.; Mao, H. Variables Selection Methods in Near-Infrared Spectroscopy. Anal. Chim. Acta 2010, 667 (1–2), 14–32. DOI: 10.1016/j.aca.2010.03.048
38. Dala-Paula, B. M.; Plotto, A.; Bai, J.; et al. Effect of Huanglongbing or Greening Disease on Orange Juice Quality, a Review. Front. Plant Sci. 2019, 9, 1976. DOI: 10.3389/fpls.2018.01976

Sihui Chen, Wenkang Zhai, Manmin Zhang, and Lin Huang are with the College of Biological Science and Engineering / Jiangxi Engineering Laboratory for the Development and Utilization of Agricultural Microbial Resources / Institute of Applied Microbiology, Jiangxi Agricultural University, Nanchang, China. Jiang Xu, Mingyin Yao, and Muhua Liu are with the College of Engineering, Jiangxi Agricultural University / Jiangxi Key Laboratory of Modern Agricultural Equipment, Nanchang, China. Direct correspondence to Lin Huang at: huanglin213@126.com