Next Article in Journal
Transactive Demand–Response Framework for High Renewable Penetrated Multi-Energy Prosumer Aggregators in the Context of a Smart Grid
Previous Article in Journal
A Diagnosability-Integrated Design Approach Based on Graph Theory
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Applied Sciences
Volume 13
Issue 18
10.3390/app131810082
applsci-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Diagnosis of Citrus Greening Based on the Fusion of Visible and Near-Infrared Spectra

by
Huaichun Xiao
1,2,*,
Yang Liu
1,2,
Yande Liu
3,
Hui Xiao
1,2,
Liwei Sun
1,2 and
Yong Hao
3
1
Jiangxi Province Key Laboratory of the Causes and Control of Atmospheric Pollution, East China University of Technology, Nanchang 330013, China
2
School of Geophysics and Measurement and Control Technology, East China University of Technology, Nanchang 330013, China
3
School of Mechanical and Electrical Engineering, East China Jiaotong University, Nanchang 330013, China
*
Author to whom correspondence should be addressed.
Appl. Sci. 2023, 13(18), 10082; https://doi.org/10.3390/app131810082
Submission received: 21 July 2023 / Revised: 26 August 2023 / Accepted: 29 August 2023 / Published: 7 September 2023
(This article belongs to the Section Optics and Lasers)
Browse Figures
Versions Notes

Abstract

:
A disease, known as citrus greening, is a major threat to the citrus industry. The objective of this study was to investigate the feasibility of rapid detection and improving the identification accuracy of citrus greening with visible and near-infrared spectra under spectral fusion. After we obtained the spectra of the collected citrus leaves and used the polymerase chain reaction for part of them, five types of samples were sorted out: slight, moderate, serious, nutrient deficiency, and normal. This study of spectral fusion was conducted on three levels as spectral data, characteristic, and model decision, and the identification capacity was tested using prediction samples. It was found that the effect of a least squares support vector machine model for feature-level fusion based on principal component analysis presented the best performance, while in the Lin_Kernel function; the accuracy was 100%, penalty coefficient γ was 0.09, and operation time was 0.66 s. It is better than the single spectral discriminant model. The results showed that the fusion of visible and near-infrared spectra was feasible for the nondestructive detection of citrus greening disease. This method is of great significance for the healthy development of the citrus industry, and provides important reference value for the application of spectral fusion in other fields.

1. Introduction

Citrus greening, or Huanglongbing (HLB), as one of the most devastating citrus diseases likely to break out throughout the whole year, is caused by Gram-negative bacteria in the phloem of branches [1,2], and mainly spreads by citrus psyllid. In a few previously conducted studies, there was a correlation between the sex of the psyllid and the host tree and density of the bacterial population, while the citrus species, geographical conditions, climate, planting pattern and management level, and other external factors had a certain influence on the disease and were closely related to the photosynthesis of citrus trees, which resulted in the difference in spectral information between the infected and normal samples [3,4,5]. It can spread rapidly once being infected. The only effective way is to remove the infected tree [6]. It should be pointed out that the spread of citrus greening poses a serious threat to the stability and development of the citrus industry [7]. Therefore, effective measures for controlling its invasion and spread lie in prevention, and the premise of prevention is to detect it in time. Currently, there are two common diagnostic methods. One is laboratory pathological analysis using polymerase chain reactions, including conventional and fluorescence quantitative polymerase chain reactions. Although this method has a high level of accuracy, it is time-consuming and expensive [8,9]. The other method is manual field diagnosis, which is simple and not time-consuming, but has a low accuracy [10,11].
In recent years, researchers have applied the spectroscopy techniques in citrus greening detection, and progress has been achieved. Sankaran et al. (2011) obtained the visible (Vis) and near-infrared spectra of 100 healthy and 93 HLB-infected samples in the range of 350–2500 nm, and reduced the number of input variables via pretreatment and principal component analysis (PCA). The classification model was developed by, respectively, applying linear discriminant analysis, quadratic discriminant analysis, K-nearest neighbor, and soft independent modeling of classification analogies (SIMCA), among which the method of quadratic discriminant analysis achieved the highest average classification accuracy of 95% [12]. Windham et al. (2011) obtained the positive and negative leaf spectra of HLB with near-infrared spectroscopy, and successfully established the partial least squares (PLS) regression model, which can classify positive HLB, negative HLB, other diseases, and nutritional deficiency samples with an increase in accuracy from 92% to 99% [13]. The above research was conducted under the condition of a single sensor; thus, feedback information was lower and the detection accuracy was low. The spectrum is a comprehensive representation of the physical and chemical information of the leaf, and the spectral data fusion technique leads to an increase in the amount of feedback spectral information, which can fully reflect the physical and chemical information of the leaf [14,15,16,17]. The method of spectral data fusion technology is to fuse all spectral data together in different ways, which can obtain more comprehensive and useful information. Data fusion can be carried out at three levels, namely low-level fusion, middle-level fusion, and high-level fusion. Low-level fusion is the data-layer fusion of the spectrum. The data from all data sources are simply connected to a matrix in sample order. Middle-level fusion is to extract relevant features from each data source and combine them into a matrix, so it is also called feature-level fusion. High-level fusion, also known as decision-level fusion, calculates a separate classification or regression model from each data source and combines the results of each separate model to obtain the final decision [18]. This method can not only improve the model’s accuracy, but also improves the model’s stability [19,20,21]. Wang, W et al. (2016). obtained the visible and near-infrared reflectance spectrum of 60 fresh pork, which used two kinds of spectrometers, and pretreated them. Spectrally overlapped areas were eliminated with the band data fusion method, and the prediction model was developed to be used combined with pork quality parameter physical and chemical values, respectively. Analysis of the results of models showed that the fusion effect was even better than before [22]. Rourke et al. (2016) predicted the agronomic soil by utilizing the technology of Vis/NIR spectra and X-ray fluorescence together, which increased the prediction accuracy and improved the percentage of root mean square error [23]. To make a rapid detection of citrus greening using the spectral data fusion technique, the improvement of detection accuracy was supposed to be achieved.
Based on the above analysis on the performance of single technology in detecting citrus greening, this paper aims to discuss the feasibility of rapid identification and classification of citrus greening by using two kinds of spectral techniques, visible and near-infrared, which are fused from three aspects, low-level, middle-level, and high-level, respectively. After acquiring the visible and near-infrared spectra of five kinds of citrus leaves, which were screened in a polymerase chain reaction and combined with stoichiometry methods, we successfully established citrus greening recognition models. The performance of the single-model and data-fusion model based on the original spectrum, derivative spectrum, and derivative normalized spectrum is compared and analyzed from the above three levels. Then, the best result was obtained and provided a good reference to the application of spectral data fusion technology in other related fields.

2. Material and Methods

2.1. Sample Preparation

According to the guidance of the local senior agronomist, citrus trees of disease were roughly divided by visual inspection method for convenient sampling [24]. The samples were all collected from an orchard planting base in Jiangxi province. Samples were collected from 5 divided regions of the base so as to present their differences. Three citrus trees were randomly selected from each region and the distance among the three trees was over 12 m. A total of 180 leaves were collected with 12 leaves picked from each tree on the same layer in four directions (east, south, west and north). The leaves were loaded by self-sealing bags with labels and brought back to the laboratory for pretreatment (such as washing and airing). Samples were further screened and divided by polymerase chain reaction (PCR), and were synthesized by Genscript company in Nanjing [25,26]. The PCR technique is generally adopted by the conventional nuclear acid extraction method, which is expanded. The cycle is composed of three steps: high-temperature variation, low-temperature annealing and medium temperature elongation. The denaturation process should be kept at a high temperature of 95 °C left and right; the reaction temperature decreased to 60 °C left and right during annealing, while the reaction temperature increased to 72 °C left and right during extension stage, so as to ensure that the target was fully amplified [27].
According to Hocquellet’s report of the OI1/OI2 primers and Jagoueix’s report of A2/J5 primers. The sequences of OI1/OI2 and A2/J5 were: 5′-GCGCGTATCCAATACGAGCGGCA-3′, 5′-GCCTCGCGACTTCGCAACCCAT-3′, 5′-TATAAAGGTTGACCTTTCGAGTTT-3′, 5′-ACAAAAGCAGAAATAGCACGAACAA-3′. The polymerase chain reaction test was positive in diseased leaves, and negative in those without disease, as shown in Figure 1. At the same time, to show homogeneity and representativeness, the number of each category sample was the same, and since the symptoms of nutrient deficiency samples were similar to those of citrus yellowing samples, we selected samples with nutrient deficiency for comparative experiments. (The OI1/OI2 primers are abbreviated as O1/O2 for convenient labeling in Figure 1.)
It could be concluded from Figure 1 that the test results of the polymerase chain reaction, with OI1/OI2 as the primer, were clearer. In Figure 1, the slight, moderate and serious citrus greening successively showed bright bands and the color became brighter with an increase in seriousness of the disease while the normal leaves showed no bright band. Different from those of leaves with citrus greening, the bright bands of nutrient deficiency leaves may have a connection to the lack of nutrients in leaves [28,29]. The infection situation of five types of leaves and the color diagram of representative samples are shown in Table 1.
By considering the influence of disease infection on citrus yield and the results of conventional polymerase chain reaction testing, the experimental samples were divided into five levels: Level 1 was the slight grade of citrus greening, Level 2 was the moderate grade of citrus greening, Level 3 was the serious grade of citrus greening, Level 4 was nutrient deficiency, and Level 5 was normal [30]. During leaf preservation period, 24 damaged samples were eliminated in the process of spectra collection in case they affected the experiment. Among them, there were 6 leaves with the slight citrus greening, 6 leaves with moderate citrus greening, 6 leaves with serious citrus greening, 3 nutrient deficiency leaves, and 3 normal leaves. Finally, these samples were numbered, ready for spectral acquisition.

2.2. Spectrometers

The visible spectral acquisition system included a spectrometer QE65000 with a range of 200–1135 nm, a light source model LS-1 with a range of 325–2500 nm, the ocean’s Y shaped fiber, and computers, which were manufactured by Ocean Optics, Inc. and purchased in Shanghai, China. Leaf visible spectrum acquisition was completed with the assorted control software: Spectrasuite [31]. Parameters: cumulative time of 30 ms, 15 times of scanning, smoothness of 15.
NIR spectra were collected by Fourier transform spectrometer TENSOR37, which was manufactured by Bruker, Inc. and purchased in Beijing, China. The instrument parameters were as follows: wavelength range was 835–2500 nm and spectral resolution 8 cm−1. At the same time, it was equipped with an InGaAs detector, gold-plated integral ball and standard background plate.

2.3. Spectrum Acquisition

Before collecting leaves spectra, the light source and the spectrometer were opened to preheat for 30 min. Then, we collected the spectra S of leaf tiled on the white board, dark current spectrum D and reference incident spectrum R [32]. The reason for collecting in this way was that when the light in the optical fiber illuminated the samples on the white board, a part of the light detected by the optical fiber was directly reflected, and the other part was received by the optical fiber after permeating the leaf and being reflected by the white board to the leaves. It is easy for light to permeate thinner leaves, thus this method was used for collecting. In the process of spectral acquisition, the spectra of three random points on leaves were collected continuously and the average was taken. The absorbance A was gained from the calculation in Equation (1), and the visible absorbance spectra of leaf was used for subsequent processing [33,34].
A λ = − log 10 ( S λ − D λ R λ − D λ )
The near infrared spectra of leaves were collected in a suitable indoor environment with a near infrared spectrometer, which had been preheated for 30 min. In order to maintain spectral validity, leaves should be smooth and the vein position should be avoided. When spectrum of four arbitrary points were collected from an area where the visible spectra were collected, the gold-plated mirror reference should be collected after each sample spectra acquisition completion and the average values of the four spectra would be used for subsequent analysis.

2.4. Dimension Reduction Method

PCA is a method of sequentially finding a set of mutually orthogonal axes from the original space that maps n-dimensional features to m-dimensions; these are completely new orthogonal features also known as principal components, which are k-dimensional features reconstructed on the basis of the original n-dimensional features [35,36]. The spectral variables are compressed into a linear combination of several principal components that represent the data structure as much as possible without losing information. Based on the proportion of variability, information about the underlying components aggregate the covariable wavelength. Thus, the first component records major spectral changes and, in turn, the more components that are added, the finer the information or noise. In this experiment, the maximum number of principal components was set at 20.
The successive projections algorithm (SPA) was used as a feature variable selection method, extracting valid information in overlapping spectral information, which made the spectral variables between reach the minimum collinearity and redundancy. The proposed algorithm is selected in the initial wavelength, forward cycle, and calculation of the projection vector of the unselected wavelength, which corresponds to the projection maximum selected, and then the projection vector is compared to the wavelength and combined until the end of the cycle. This method is beneficial to reduce the amount of computation needed and to simplify the model structure and improve modeling speed [37,38].

2.5. Modeling Method

PLS discriminant analysis (PLS-DA) is a linear classification method based on the PLS regression algorithm, which is suitable for the analysis and prediction of multiple independent variables and strong linearity. It can explain the extracted principal factors from the perspective of the internal nature of data [39,40].
The support vector machine (SVM) operation becomes complicated when the number of samples is too large, so the least squares support vector machine (LS-SVM) was proposed to address problems of inequality, constraints and quadratic programming in SVM. LS-SVM has a higher precision with a faster speed and smaller random access memory (RAM) usage [41,42]. The solving procedure is presented in Equation (2):
{ min Z ( w , k ) = 1 2 w T w + γ 1 2 ∑ j = 1 l k j s . t . N j = w T φ ( x j ) + m + k j
where xj is the input variables, Nj is the target value, φ(xj) is the kernel function, kj is error variables, w is weight of size, b is deviation, and γ is penalty coefficient. Two kinds of kernel functions are used in the paper. These two kernel functions are shown in Equations (3) and (4):
k ( x i , x j ) = e − ( x i − x j ) 2 2 σ 2
k ( x i , x j ) = L ( x i ) ⋅ L ( x j )
where k(xi,xj) is the inner kernel function, σ is the kernel width, L(xi) and L(xj) are mapping functions. Before developing the LS-SVM model, two parameter γ and σ2 of radial basis function (RBF_Kernel) and the penalty coefficient γ of linear function must be obtained. The optimal parameter of LS-SVM is to find the best quadratic grid search and leave-one-out cross validation. The parameters determine the learning prediction and generalization performance of the correction model. γ is used to maximize generation performance of model and to minimize model complexity. Nuclear parameter σ2 decides the proximity coefficient of the model. The kernel function equation is presented in Equation (5) [43,44]:
y ( x ) = ∑ k = 1 n a k k ( x i , x j ) + b
where ak is the lag range operator and b is the statistical deviation. All the preprocessing was performed using Unscrambler 8.00; the pattern recognition and classification algorithms were performed using Matlab® R2010a.

3. Results and Discussion

3.1. Spectral Characteristic Analysis and Pretreatment

The typical visible and near infrared spectra of the five kinds of leaves, including the slight, the moderate, the serious, nutrient deficiency and normal, are shown in Figure 2. The main bands of visible spectrum are 540–640 nm and 750–850 nm. The former corresponds to the part of green light, yellow light and red light. The latter is mainly related to C-H, O-H, N-H chemical bonds, caused by its 3~4 order stretching vibration. Near infrared spectral band is located at 1400–1600 nm and 1900–2100 nm, where the former is mainly caused by O-H, C-H, N-H level 1 stretching vibration, and the latter is mainly related to the secondary deformation vibration of O-H, C-O and N-H bonds [13,29].
Two distinct reflection peaks are located at the spectral wavelength of 560 and 760 nm in Figure 2a. The main reason is that the chlorophyll in leaves with citrus greening is less than that of the normal, leading to the phenomenon where two reflection peaks of the leaves with citrus greening are higher than that of the normal leaves, and gradually rises with the increase of the disease grade [37,45]. The fact that the reflection peak of nutrient deficiency leaves is lower than that of the normal may be related to the lack of nutrients in leaves. Two comparatively distinct absorption peaks in Figure 2b are located at 1930 and 1450 nm. For the leaves that are rich in the hydrogen group, the former is mainly caused by the expansion and deformation of the O-H group, while the latter is caused by the first order frequency double-stretching vibration of the O-H bond of water and sugar [46,47,48]. The reason why the leaf absorption peak of greening disease was lower than two kinds of nutrient deficiency and normal leaves may be that the pathogen of the greening disease blocks the absorption of water, resulting in low absorbance, and gradually dropping off with the decrease of the disease grade. The absorbance of nutrient deficiency leaves was higher than that of greening disease, which may be caused by the low dependence of the lack nutrients on light.
It is shown in Figure 2 that the ranges of 325–500 and 1000–1135 nm in the visible spectra have little effect on the whole spectrum, thus the ranges of the above bands were removed. The near infrared spectra in the range of 835–1000 nm is close to a straight line, because the leaves absorbance in this band is mainly caused by the 3–4 order frequency multiplier of functional groups, which has limited influence on the whole spectrum, thus the band was also removed. Since the data types of the visible and near infrared spectral domains are inconsistent, the absorbance is converted into reflectance by Equation (6) and because of the mutual interference between the spectral variables.
q = lg ( 1 / n )
where n represents absorbance and q represents reflectivity. In order to optimize the model, the second derivative was used to preprocess the spectrum. The method not only amplified the effective spectral information, but also eliminated the spectral baseline drift, and increased the number of spectral curves and reduced the peak overlap, thus improving the resolution and facilitating the model’s construction [49]. The spectrum of the second derivative pretreatment is shown in Figure 3. It is shown in Figure 3a that the visible reflection peaks of the five-type citrus leaves are obviously different at 690 nm, and the reflection peak of citrus leaf with serious citrus greening is the highest. In Figure 3b, significant differences can be found in the near infrared reflection (NIR) peaks of the five-type leaves at 1840 nm, which may be caused by the amplifying the sensitive information of disease leaves during pretreatment [50,51].
Data acquisition instruments and dimensions are inconsistent, so dimensionality reduction is adopted according to Equation (7) to normalize data after second derivative pretreatment, in order to improve the model recognition accuracy.
X = x − u ¯ σ
X is new spectral vector after normalization, σ is the second derivative matrix of spectral, u is the second derivative spectra mean value of sample, and σ is the standard deviation of the second derivative spectral variables. The representative spectrum of normalized data is shown in Figure 4, in which Figure 4a is within the range of 500–1000 nm, and Figure 4b is within the range of 1000–2500 nm.
It is shown in Figure 4 that the trend of the normalized spectra of the five types of representative blades in the two spectral ranges is basically similar. In Figure 4a, in the range of 650–750 nm, the reflectance decreases from the maximum value to the minimum value. Compared with normal leaves, the reflectance spectrum of other leaves shows a right-shift tendency, which may be related to the absence of chlorophyll. The main reason why the reflectance of the five representative leaves in Figure 4b decreased from the maximum value to the minimum value near 1900 nm may be the change of water.

3.2. Feature Variable Screening

The citrus leaves spectral data of Vis and NIR contains chemical composition and some information that is not relevant to this study of those leaves. In order to reduce the number of input variables and simplify the complexity of the model, the spectral feature variables were extracted, and then data fusion modeling, which is qualitative analysis based on spectral data feature level fusion. In the experiment, spectral feature variables were extracted through PCA and SPA. Firstly, the derivative spectra were compressed into a number of principal components between the two bands of 500–1000 and 1000–2500 nm before and after normalization. Secondly, the feature variables were fused under the premise of not losing the main spectral information, PCA compressed data to obtain new spectral variables that can replace lengthy spectral variables. The visible data were compressed into three principal component factors that contribute more than 99% in the range of 500–1000 nm, which represents the main information of spectra. The NIR data was compressed into eight principal component factors with a contribution rate over 92% in the range of 1000–2500 nm. Therefore, combining with the above, 11 principal component factors for feature level fusion model were created.
According to the above data, the spectral variables were extracted by SPA, and then data feature level fusion was carried out. The feature variables were selected from 2220 variables before and after normalization. The minimum variables before running SPA were set to 10, while the maximum variables were set to 30. According to the number of information variables based on spectral matrix projection before and after normalization, spectral feature variables were selected. In both cases, a total of 55 and 60 feature variables were selected for feature level fusion model.

3.3. Fusion Model Building and Discriminant

3.3.1. Discriminant Model Based on Spectral Data Layer Fusion

The remaining 156 samples were randomly divided into a calibration set and prediction set in the ratio of 2:1. There were 104 samples in calibration sets including 20 with slight citrus greening, 20 with moderate citrus greening, 20 with serious citrus greening, 22 nutrient deficiency and 22 normal leaves. The remaining 52 were set as prediction samples used to evaluate the prediction effect of the model [52,53]. Classification vector values were assigned according to the disease characteristics of the samples, which were set as [2, 4, 6, 8, 10] in the experiment, 2 as light HLB leaf, 4 as medium HLB leaf, 6 as heavy HLB leaf, 8 as deficiency leaf and 10 as normal leaf, and the intermediate value of the two vectors was the classification threshold [30].
The model was developed from the spectral data fusion between the range of 500–1000 nm of visible and the range of 1000–2500 nm of NIR. The data fusion was carried out by using three strategies: spectrum direct stitching, derivative stitching and normalization stitching. By comparative analysis among those three leaves, no bright band was shown. Different from those of leaves with citrus greening, the bright bands of nutrient deficiency leaves may have a connection to the lack of nutrients in leaves [28,29]. More details were listed in Table 1.
There were 2220 variables in the data fusion model, including 665 visible band variables and 1555 near-infrared band variables. As shown in Table 2, the identification accuracy of the LS-SVM model, based on the direct splicing of spectral data, is better than the prediction results of single bands of 500–1000 and 1000–2500 nm. Although the parameters γ, σ2 of the LS-SVM fusion model is larger and their operation time is longer, the identification accuracy of the model is higher than those of the two single bands. The model of spectral direct splicing after the second derivative pretreatment is better with 100% identification accuracy and reduced penalty coefficient γ and kernel parameter σ2, where the penalty coefficient γ of the linear kernel function is reduced to 0.09, and the operation time is also shortened. The identification accuracy of the model using linear function Lin_Kernel as the kernel function is better than that of using radial basis function RBF_Kernel, which may be due to a linear relationship between spectral variables. Therefore, it is feasible to develop the identification model of citrus greening through derivative direct splicing strategy on the basis of spectral data fusion. The better effect of the model is mainly because the data of the two bands in the range of 500–1000 and 1000–2500 nm contain more effective information.
Secondly, because of the dimensional difference of spectral data at two bands of 500–1000 and 1000–2500 nm, it is necessary to use normalization to remove spectral data dimension on the basis of second derivative, and then splice the spectrum. It can be concluded from Table 2 that the result of normalization splicing is better than direct splicing, but not as good as second derivative direct splicing. The main reason is that normalization may lead to the loss of spatial information of the identification pattern, so the result of second derivative splicing is optimal.
Based on the best model in Table 2, the aforementioned samples were randomly divided into calibration set and prediction set according to different proportions. Combined with partial least squares discriminant analysis, a diagnosis model of citrus greening disease with the fusion of four spectral data layers was constructed, and the results were shown in Table 3.
It is shown in Table 3 that the identification accuracy of each model is 100%. Among them, 156 samples were randomly divided into calibration set and prediction set according to 2:1, and the model prediction effect was the best. At this time, the lowest RMSEC was 0.65%, and the highest Rc was 0.94. Therefore, the subsequent sample division was carried out according to the above.

3.3.2. Discriminant Model Based on Data Feature Layer Fusion

On the basis of PCA and SPA methods to screen characteristic variables, LS-SVM and PLS-DA recognition models of citrus greening disease were established, respectively, the diagnostic results were shown in Table 4 and Table 5. When the Lin_Kernel function is used as the kernel function and the feature variables screened of second derivative spectral by PCA, the data fusion modeling results are the best, the recognition accuracy is 100%, the penalty coefficient is 0.09, and the operation time is 0.66 s. When SPA selects normalized second derivative spectral feature variables, the data fusion model has the worst recognition effect. The main cause for this phenomenon probably is that new feature variables acquired after PCA compressing variables contains more effective information, while some effective information may become lost during the SPA variable screening process; meanwhile, the normalized spectra may result in spatial information loss of identification pattern, reducing accuracy.
It is shown in Table 5 that the overall effect of the PCA compression model is better than that of SPA. The unnormalized feature level fusion model in the former method has the best recognition accuracy of 90%, while the unnormalized feature level fusion model in the latter method has the worst recognition accuracy of 65%, which may be caused by the small amount of effective information contained in the features selected by SPA.

3.3.3. Discriminant Model Based on Decision Level Fusion

The proportion of two independent sub-models to data fusion model plays a key role for decision level fusion. During the data analysis, the non-negative least squares (NNLS) method was used to calculate the proportion of two sub-models. When NNLS is used as a special method to solve the optimal problem with the multiple linear regression method, the coefficient of NNLS should be positive [22,51,54]. The principle is shown in Equation (8):
Y = a 1 x 1 + a 2 x 2 + … … + a i x i + b
where Y is the prediction value of the fusion model, ai the proportion value of ith sub-model, Yi the prediction value of single band range spectral model, and b the deviation of the fusion model. Equation of sub-models based on the second derivative spectrum in the range of 500–1000 and 1000–2500 nm, respectively, are shown in Equations (9) and (10):
Y 1 = ∑ i 665 β i x i + b 1
Y 2 = ∑ i 1555 β i x i + b 2
The two intercepts, b1, b2, were 11.22 and 9.39 in two ranges, respectively. The regression coefficient curves of the model on the two bands are shown in Figure 5. The corresponding weights of spectral variables at different wavelengths in Figure 5 are represented by model regression coefficients, which can obviously obtain important wavelength points within the range of two wavebands. By the model regression calculation, the weight of the different spectral variables in the two bands was provided, and the sum was weighted and the intercepts were added to realize the qualitative discrimination of five types of samples. The decision level fusion model is shown in Equation (11); the results show that the identification accuracy of NNLS data fusion model based on PLS-DA is 94%. The proportions of two sub-models are 0.73 and 0.32.
Y = 0.51 Y 1 + 0.52 Y 2 − 0.28
Points with high weights in Figure 5a are 540, 560, 760 and 840 nm. The main reason is that the absorption of chlorophyll composed of C-H, O-H and N-H groups at this point is stronger, while the chlorophyll in yellow dragon disease leaves is less than that in normal leaves. In Figure 5b, the points with high weights are 1090, 1430, 1930 and 2240 nm. The first two points are mainly caused by the first-order frequency double-stretching vibration of the O-H bond of water and sugar, and the last two points are mainly caused by the stretching and deformation vibration of O-H group [29,49].

3.3.4. Discriminant Model Comparison Analysis

The data fusion model of citrus greening identification was evaluated with 52 prediction samples, which was established based on three levels; the identification ability of the fusion model was better than the single band model on the whole. According to the model evaluation index of the identification accuracy, penalty coefficient and kernel parameter at the spectrum data layer, the fusion model based on the derivative spectrum has a better effect using the LS-SVM method. The discrimination effect of the feature level fusion model is best among the three levels, and the LS-SVM fusion model with PCA compression variables is optimal. However, the fusion model based on PLS-DA, which adopts SPA method to select variables, has a poorer effect. The main reason is that the new feature variables, acquired after the PCA compressing variables, contains more effective information. The information of target factors decreased when derivative spectral was used to remove the non-target factor’s influence on the spectrum. Moreover, in these LS-SVM models, linear kernel function is better than radial basis kernel function because there is a better linear relationship between spectral variables. The NNLS fusion model at the decision level has a medium effect, which may be due to the model not taking into account the proportion of each spectral variable to the model.

4. Conclusions

Two spectra of Vis and NIR were fused based on the three different levels, including the spectral data level, the spectral feature level and the model decision level. The feasibility of rapid identification of citrus greening was studied and discussed. The spectral data fusion identification model for citrus greening and classification of the disease was developed through two methods of LS-SVM and PLS-DA. It was concluded after a comparative analysis that the feature level fusion based on PCA variable compression had the best identification results in three level fusion models, and with linear function (Lin_Kernel) being the kernel function, the identification accuracy is the highest. The result of the spectral data level fusion model based on derivative spectrum splicing was second, and the result of decision level fusion model based on the NNLS was the worst. The experiment result proves the feasibility of rapidly detecting citrus greening based on Vis and NIR spectra data fusion. It provides a good reference for the identification of citrus greening with low accuracy and poor stability through a single spectral technique.

Author Contributions

Conceptualization, H.X. (Huaichun Xiao); methodology, H.X. (Hui Xiao) and Y.H.; validation, H.X. (Hui Xiao), and L.S.; formal analysis, H.X. (Huaichun Xiao) and Y.L. (Yande Liu); investigation, L.S. and Y.L. (Yang Liu); resources, Y.L. (Yande Liu); data curation, H.X. (Hui Xiao) and; writing—original draft preparation, H.X. (Huaichun Xiao) and Y.L. (Yang Liu); writing—review and editing, H.X. (Huaichun Xiao) and Yang Liu.; visualization, L.S.; supervision, Y.H.; project administration, H.X. (Huaichun Xiao); funding acquisition, Y.L. (Yande Liu). All authors have read and agreed to the published version of the manuscript.

Funding

Jiangxi Province Key Laboratory of the Causes and Control of Atmospheric Pollution, East China University of Technology (NO. AE2109), Jiangxi Engineering Technology Research Center of Nuclear Geoscience Data Science and System open fund project (No. JETRCNGDSS202102), Doctor initiation research fund project of ECUT (No. DHBK2019074), Jiangxi Key Laboratory for Mass Spectrometry and Instrumentation open fund project (No. JXMS202011), National Natural Science Foundation of China (No. 61866001), Jiangxi Province College Students Innovation and entrepreneurship training program (NO. S202110405004X), and Jiangxi Provincial Natural Science Foundation of China (NO. 20212BAB204009).

Institutional Review Board Statement

The study did not require ethical approval.

Informed Consent Statement

The study did not involve humans.

Data Availability Statement

Not applicable.

Acknowledgments

The authors would like to express their gratitude to the Jiangxi Province Key Laboratory of the Causes and Control of Atmospheric Pollution for the venue provision and technical support.

Conflicts of Interest

The authors declare that they have no conflict of interest to report regarding the present study.

References

  1. Shan, Z.; Guo, H.; Feng, Z.; Deng, X. Cloning and sequencing of Hang longbing pathogen in shatianyou pomelo. J. Zhongkai Univ. Agric. Technol. 2005, 18, 45–48. [Google Scholar]
  2. Treguier, S.; Jacq, K.; Couderc, C.; Ferhout, H.; Tormo, H.; Kleiber, D.; Levasseur-Garcia, C. A method for highlighting differences between bacteria grown on nutrient agar using near infrared spectroscopy and principal component analysis. J. Near Infrared Spectrosc. 2021, 29, 269–277. [Google Scholar] [CrossRef]
  3. Ukuda-Hosokawa, R.; Sadoyama, Y.; Kishaba, M.; Kuriwada, T.; Anbutsu, H.; Fukatsu, T. Infection Density Dynamics of the Citrus Greening Bacterium “Candidatus Liberibacter asiaticus” in Field Populations of the Psyllid Diaphorina citri and Its Relevance to the Efficiency of Pathogen Transmission to Citrus Plants. Appl. Environ. Microbiol. 2021, 81, 3728–3736. [Google Scholar] [CrossRef] [PubMed]
  4. Barbin, D.; Sobottka, R.; Risso, W.; Zucareli, C.; Hirooka, E.Y. Influence of plant densities and fertilization on maize grains by near-infrared spectroscopy. Spectrosc. Lett. 2016, 49, 73–79. [Google Scholar] [CrossRef]
  5. Lin, X. Studies on the Distribution and Control Technology of Citrus Huanglongbing in Hunan Province. Master’s Thesis, Agricultural University Of Hunan, Changsha, China, 2020. [Google Scholar] [CrossRef]
  6. Yu, J.; Zhang, M.; Tao, J. The control research progress of liberobacter asiaticum in Taizhou city. Agric. Sci. Technol. 2015, 6, 1237–1242. [Google Scholar]
  7. Lu, L.; Fan, G.; Hu, X. PCR detection of huanglongbing pathogen in different parts of citrus plants in the field and analysis of the cause of the disease. Plant Prot. 2011, 2, 45–49. [Google Scholar]
  8. Yang, D.; Yang, Y.; Li, H. Comparation of digital PCR and fluorescent PCR for detecting kitchen-waste vegetable oil. Chin. J. Health Lab. Technol. 2015, 19, 3221–3225+3230. [Google Scholar]
  9. Gao, Y.X.; Lu, Z.J.; Zhong, B.L.; Yi, L.; Su, H.N.; Zhao, P.Y. Citrus greening comparative study of two kinds of fluorescent quantitative PCR detection system. China Soc. Plant Prot. 2014, 5, 237–249. [Google Scholar]
  10. Hu, H.; Yin, Y.; Zhang, L.; Zhao, Y.; Xia, Y.; Wang, Z.; Tan, J. Detection of Citrus Huanglongbing by Conventional and Two Fluore scence Quantitative PCR Assays. Sci. Agric. Sin. 2006, 12, 2491–2497. [Google Scholar]
  11. Wang, A.; Deng, X. The research on the diagnosis technology of Citrus Huang longbing. Guangdong Agric. Sci. 2008, 6, 101–103. [Google Scholar]
  12. Sankaran, S.; Mishra, A.; Maja, J. Visible-near infrared spectroscopy for detection of Huanglongbing in citrus orchards. Comput. Electron. Agric. 2011, 2, 27–134. [Google Scholar] [CrossRef]
  13. Windham, W.; Poole, G.; Park, B.; Heitschmidt, G.; Hawkins, S.A.; Albano, J.P.; Gottwald, T.R.; Lawrence, K.C. Rapid screening of Huanglongbing-infected citrus leaves by near-infrared reflectance spectroscopy. Trans. ASABE 2011, 54, 2253–2258. [Google Scholar] [CrossRef]
  14. Sindhuja, S.; Reza, E. Visible-near infrared spectroscopy based citrus greening detection: Evaluation of spectral feature extraction techniques. Crop Prot. 2011, 30, 1508–1513. [Google Scholar]
  15. Duan, H.; Zhu, R.; Xu, W.; Qiu, Y.-Y.; Yao, X.-D.; Xu, C.-J. Hyperspectral Imaging Detection of Total Viable Count from Vacuum Packing Cooling Mutton Based on GA and CARS Algorithms. Spectrosc. Spect. Anal. 2017, 37, 847–852. [Google Scholar]
  16. Chen, C. Medical Diagnostic Technology Development of Spectral Signal Combined with Multi-Fusion Strategy. Master’s Thesis, Xinjiang University, Ürümqi, China, 2021. [Google Scholar] [CrossRef]
  17. Nweke, H.F.; Wah, T.Y.; Mujtaba, G.; Al-garadi, M.A. Data fusion and multiple classifier systems for human activity detection and health monitoring: Review and open research directions. Inf. Fusion 2019, 46, 147–170. [Google Scholar]
  18. Yang, Q.; Deng, X.; Sun, X. Application and Research Progress of Spectral Data Fusion Technology in Food Testing. Sci. Technol. Food Ind. 2020, 41, 324–329. [Google Scholar]
  19. Dian, R.; Li, S.; Fang, L. Multispectral and hyperspectral image fusion with spatial-spectral sparse representation. Int. J. Inf. Fusion 2019, 49, 262–270. [Google Scholar] [CrossRef]
  20. Li, Y.H.; Li, Y.X.; Tan, W. Based on the near-range and medium-infrared spectral method fusion to determine the black fruit and Chinese wolfberry origin and quality information. Spectrosc. Spect. Anal. 2020, 40, 3878–3883. [Google Scholar]
  21. Gholizadeh, A.; Coblinski, J.; Saberioon, M. vis-NIR and XRF Data Fusion and Feature Selection to Estimate Potentially Toxic Elements in Soil. Sensors. 2021, 21, 2386. [Google Scholar] [CrossRef]
  22. Wang, W.; Peng, Y. Simultaneous detection of multiple quality paramrters of pork based on fused dual band specteal. Spectrosc. Spect. Anal. 2016, 12, 4001–4005. [Google Scholar]
  23. O’Rourke, S.; Stockmann, U.; Holden, N.M.; McBratney, A.B.; Minasny, B. An assessment of model averaging to improve predictive power of portable vis-NIR and XRF for the determination of agronomic soil properties. Geoderma. 2016, 279, 31–44. [Google Scholar] [CrossRef]
  24. Mishra, A.R.; Karimi, D.; Ehsani, R.; Lee, W.S. Identification of Citrus Greening (HLB) Using a VIS-NIR Spectroscopy Technique. Trans. ASABE 2012, 50, 711–720. [Google Scholar] [CrossRef]
  25. Hocquellet, A.; Toorawa, P.; Bové, J.; 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, 373–379. [Google Scholar] [CrossRef] [PubMed]
  26. Li, T.; Ke, C. Detection of the bearing rate of Liberobacter asiaticum, in citrus psylla and its host plant Murraya panciculata by Nested PCR. Acta Phytophylacica Sin. 2002, 29, 31–35. [Google Scholar]
  27. Qian, W. Study on Detection Method of Citrus Yellow Dragon Disease Based on Nucleic Acid Amplification. Master’s Thesis, Zhejiang University, Hangzhou, China, 2018. [Google Scholar]
  28. Li, X.; Li, M.; Won, S. Visible-NIR Spectral Feature of Citrus Greening Disease. Spectrosc. Spect. Anal. 2014, 34, 1553–1559. [Google Scholar]
  29. Yuan, Y.; Jiang, Z.; Wang, D. Study on classification standard of Citrus Huanglongbing disease. Zhenjiang Agric. Sci. 2010, 1, 121–123. [Google Scholar]
  30. Hawkins, S.; Park, B.; Poole, G.; Gottwald, T.; Windham, W.R.; Lawrence, K.C. Detection of Citrus Huanglongbing by Fourier Transform Infrared-Attenuated Total Reflection Spectroscopy. Appl. Spectrosc. 2010, 64, 100–103. [Google Scholar] [CrossRef] [PubMed]
  31. Wang, J.; Zhang, X.; Wang, J. On-Site Evaluation of Raw Milk Qualities by Protable Vis/NIR Transmittance Technique. Spectrosc. Spect. Anal. 2014, 10, 2679–2684. [Google Scholar]
  32. Farzaneh, D.; Farhad, S.; Seyed, H.S.P.; Haidar, F. UAV-based multispectral imagery for fast Citrus Greening detection. J. Plant Dis. Prot. 2019, 126, 307–318. [Google Scholar]
  33. Chen, Z.; Li, X.; Fan, X. Method for the Discrimination of the Variety of Potatoes with Vis/NIR Spectroscopy. Spectrosc. Spect. Anal. 2016, 36, 2474–2478. [Google Scholar]
  34. Sun, J.; Wu, X.; Zhang, X. Research on Lettuce Leaves’ Moisture Prediction Based on Hyperspectral Images. Spectrosc. Spect. Anal. 2016, 2, 522–526. [Google Scholar]
  35. Yang, I.-C.; Tsai, C.-Y.; Hsieh, K.-W.; Yang, C.-W.; Ouyang, F.; Lo, Y.M.; Chen, S. Integration of SCIMCA and near-infrarded spectroscopy for rapid and pricise identification of herbal medicines. J. Food Drug Anal. 2013, 21, 268–278. [Google Scholar] [CrossRef]
  36. Liu, S. Comparative analysis of image fusion effect evaluation by different resampling methods based on PCA and DWT transformation. Sci. Technol. Innov. 2022, 12, 70–74. [Google Scholar]
  37. Yue, X.J.; Quan, D.P.; Hong, T.S. Non-destructive hyperspectral measurement model of chlorophyll content for citrus leaves. Trans. Soc. Agric. Eng. (Trans. CSAE) 2015, 31, 294–302. [Google Scholar]
  38. Liu, S.; Tian, Y.; Zhang, F. Hyperspectral imaging for nondestructive detection of hanfu apple diseases using successive projections algorithm and BP neural network. Food Sci. 2017, 38, 277–282. [Google Scholar]
  39. Lin, P.; Chen, Y.; He, Y. Identification of geographical origin of olive oil using visible and Near-Infrared Spectroscopy technique combined with chemometrics. Food Bioprocess Technol. 2012, 5, 235–242. [Google Scholar] [CrossRef]
  40. Barker, M.; Rayens, W. Partial least squares for discrimination. J. Chemom. 2003, 17, 166–173. [Google Scholar] [CrossRef]
  41. Chen, M. Research on Fault Diagnosis Method of Complex Industrial Process Based on LS-SVM. Master’s Thesis, Nanjing University of Science and Technology, Nanjing, China, 2009. [Google Scholar]
  42. Vapnik, V.N. Overview of statistical learning theory. IEEE Trans. Neural Netw. 1999, 10, 988–999. [Google Scholar] [CrossRef]
  43. Liu, X.; Lin, B.; Luo, H. Bearing faults diagnostics based on hybrid LS-SVM and EMD method. Measurement. 2015, 59, 145–166. [Google Scholar] [CrossRef]
  44. Cheng, M.; Nhat-Duc, H.; Wu, Y. Hybrid intelligence approach based on LS-SVM and Differential Evolution for construction cost index estimation: A Taiwan case study. Autom. Constr. 2013, 35, 306–313. [Google Scholar] [CrossRef]
  45. Xu, D.; Wu, W.M.; Wang, J.Q. Study on the Correlation between Leaf Chlorophyll Content of Populus euphratica and the Hyperspectral Remote Sensing Data in Upstreamf Tarim River. J. Tarim Univ. 2012, 24, 53–59. [Google Scholar]
  46. Huang, Y.; Wang, J.; Li, X. Research on fast discrimination between Panax Ginseng and Panax Quinque folium based on near infrared spectroscopy. Spectrosc. Spect. Anal. 2010, 30, 2954–2957. [Google Scholar]
  47. Huang, G.; Han, L.; Yang, Z. Evaluation of the nutrient metal content in Chinese animal manure compost using Near Infrared Spectroscopy (NIRS). Bioresour. Technol. 2008, 99, 8164–8169. [Google Scholar] [CrossRef] [PubMed]
  48. Hu, X. Study on NIR spectral characteristics and authenticity discrimination model of milk powder. Inn. Mong. Agric. Univ. 2020, 7, 233–238. [Google Scholar]
  49. Windham, W.; Kandala, C.V.; Sundaram, J.; Nuti, R.C. Determination of Peanut Pod Maturity by Near-Infrared Reflectance Spectroscopy. Trans. ASABE. 2010, 53, 491–495. [Google Scholar] [CrossRef]
  50. Li, X.; Lee, W.; Li, M. Spectral difference analysis and airborne imaging classification for citrus greening infected trees. Comput. Electron. Agric. 2012, 83, 32–46. [Google Scholar] [CrossRef]
  51. Cunliffe, S.; Martin, P.; Baker, M. Near infrared absorption spectroscopy for the quantification of unsulfated alcohol in sodium lauryl ether sulfate. J. Near Infrared Spectrosc. 2010, 29, 11–23. [Google Scholar] [CrossRef]
  52. Muhammad, B.; Zou, X.; Muhmmad, A. Near infrared spectroscopy coupled with chemometric algorithms for predicting chemical components in black goji berries (Lycium ruthenicum Murr.). J. Near Infrared Spectrosc. 2020, 26, 275–286. [Google Scholar]
  53. Sun, X.; Subedi, P.; Walker, R. NIRS prediction of dry matter content of single olive fruit with consideration of variable sorting for normalisation pre-treatment. Postharvest Biol. Technol. 2012, 163, 111–123. [Google Scholar] [CrossRef]
  54. Cao, J.; Zhuo, L. An improved non-negative matrix factorization method of blind unmixing for hyperspectral imagery. Sci. J. Earth Sci. 2013, 3, 107–118. [Google Scholar]
Applsci 13 10082 g001
Figure 1. The results of common polymerase chain reaction test for citrus leaves. NOTE: The numbers 1, 2, and 3 represent the label of the citrus tree to which the leaves belong; “Lack” represents the zinc deficiency samples. “M” stands for DNA molecular weight standard; “Blank” is only for comparison; slight, moderate and serious indicate the disease grade.
Figure 1. The results of common polymerase chain reaction test for citrus leaves. NOTE: The numbers 1, 2, and 3 represent the label of the citrus tree to which the leaves belong; “Lack” represents the zinc deficiency samples. “M” stands for DNA molecular weight standard; “Blank” is only for comparison; slight, moderate and serious indicate the disease grade.
Applsci 13 10082 g001
Applsci 13 10082 g002
Figure 2. Typical spectra of 5 kinds of leaves, respectively.
Figure 2. Typical spectra of 5 kinds of leaves, respectively.
Applsci 13 10082 g002
Applsci 13 10082 g003
Figure 3. Representative spectra of 5 kinds of leaves after 2 derivative.
Figure 3. Representative spectra of 5 kinds of leaves after 2 derivative.
Applsci 13 10082 g003
Applsci 13 10082 g004
Figure 4. Representative spectra of 5 kinds of leaves after normalization.
Figure 4. Representative spectra of 5 kinds of leaves after normalization.
Applsci 13 10082 g004
Applsci 13 10082 g005
Figure 5. The regression coefficient curve of model on different bands.
Figure 5. The regression coefficient curve of model on different bands.
Applsci 13 10082 g005
Table 1. The category of samples by common polymerase chain reaction.
Table 1. The category of samples by common polymerase chain reaction.
The Number of SamplesGradeLeaf InfectionRepresentative Sample Color Picture
36NormalNo symptoms of greening, polymerase chain reaction test negativeApplsci 13 10082 i001
36Slight greening diseaseSlight symptoms, polymerase chain reaction test positiveApplsci 13 10082 i002
36Moderate greening diseaseModerate symptoms, polymerase chain reaction test positiveApplsci 13 10082 i003
36Serious greening diseaseSerious symptoms, polymerase chain reaction test positiveApplsci 13 10082 i004
36Nutrient deficiencyNo symptoms of greening, polymerase chain reaction test negativeApplsci 13 10082 i005
Table 2. The results of LS-SVM model for data fusion under three strategies.
Table 2. The results of LS-SVM model for data fusion under three strategies.
Wavelength RangeKernel FunctionOperation Time(s)Identification Accuracy(%)
NameNumber of VariablesParameter
500–1000RBF_Kernel665γ = 135.62 σ2 = 12,704.145.4490%
Lin_Kernel665γ = 115.603.9794%
1000–2500RBF_Kernel1555γ = 25.14 σ2 = 31.656.2792%
Lin_Kernel1555γ = 8.292.2396%
500–2500RBF_Kernel2220γ = 2090.26 σ2 = 27,965.659.4892%
Lin_Kernel2220γ = 306.044.1698%
500–2500 after second derivativeRBF_Kernel2220γ = 124.73 σ2 = 19,820.814.50100%
Lin_Kernel2220γ = 0.093.06100%
500–2500 s derivative normalizationRBF_Kernel2220γ = 586.13 σ2 = 69,463.806.0298%
Lin_Kernel2220γ = 0.093.26100%
Table 3. PLS-DA models with different proportions of modeling set and prediction set.
Table 3. PLS-DA models with different proportions of modeling set and prediction set.
Number of Samples in the Modeling SetPrediction Set Sample NumberIdentification Accuracy (%)RMSEC (%)RMSEP (%)RcRpPc
12630100%0.770.450.910.9719
11739100%0.760.420.930.9815
10452100%0.650.470.940.9713
9462100%0.780.410.900.9813
NOTE: The “RMSEC” and “RMSEP” stand for corrected standard deviation and predicted standard deviation, respectively; “Rc” and “Rp” stand for prediction set correlation coefficient and alibration set correlation coefficient, respectively; “Pc” represents the number of principal components.
Table 4. The results of spectral fusion LS-SVM model by different variable selection methods.
Table 4. The results of spectral fusion LS-SVM model by different variable selection methods.
Selection MethodWavelength RangeKernel FunctionOperation
Time/s		Identification Accuracy/%
NameNumber
of Variables		Parameter
PCA500–2500
before normalized		RBF_Kernel11γ = 251.44 σ2 = 46.011.63100%
Lin_Kernel11γ = 0.090.66100%
500–2500 after normalizedRBF_Kernel11γ = 2047.44 σ2 = 101.571.63100%
Lin_Kernel11γ = 8.812.05100%
SPA500–2500
before normalized		RBF_Kernel55γ = 0.93 σ2 = 6.111.42100%
Lin_Kernel55γ = 0.191.01100%
500–2500 after normalizedRBF_Kernel60γ = 13.17 σ2 = 70.271.8996%
Lin_Kernel60γ = 29.121.39100%
Table 5. The result of PLS-DA before and after normalization for second derivative spectrum in 500–2500.
Table 5. The result of PLS-DA before and after normalization for second derivative spectrum in 500–2500.
Selection MethodPCASPA
Pretreatment methodbefore normalizationafter normalizationbefore normalizationafter normalization
Number of variables11115560
Identification accuracy (%)90%85%65%88%
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Xiao, H.; Liu, Y.; Liu, Y.; Xiao, H.; Sun, L.; Hao, Y. Diagnosis of Citrus Greening Based on the Fusion of Visible and Near-Infrared Spectra. Appl. Sci. 2023, 13, 10082. https://doi.org/10.3390/app131810082

AMA Style

Xiao H, Liu Y, Liu Y, Xiao H, Sun L, Hao Y. Diagnosis of Citrus Greening Based on the Fusion of Visible and Near-Infrared Spectra. Applied Sciences. 2023; 13(18):10082. https://doi.org/10.3390/app131810082

Chicago/Turabian Style

Xiao, Huaichun, Yang Liu, Yande Liu, Hui Xiao, Liwei Sun, and Yong Hao. 2023. "Diagnosis of Citrus Greening Based on the Fusion of Visible and Near-Infrared Spectra" Applied Sciences 13, no. 18: 10082. https://doi.org/10.3390/app131810082

APA Style

Xiao, H., Liu, Y., Liu, Y., Xiao, H., Sun, L., & Hao, Y. (2023). Diagnosis of Citrus Greening Based on the Fusion of Visible and Near-Infrared Spectra. Applied Sciences, 13(18), 10082. https://doi.org/10.3390/app131810082

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop