Detection of Huanglongbing-Infected Citrus Leaves Using Statistical Models with a Fluorescence Sensor

Abstract

A handheld fluorescence sensor was tested as a sensing tool to identify Huanglongbing (HLB), a citrus disease, in both symptomatic and asymptomatic stages. Features such as yellow, red, and far-red fluorescence at UV, blue, green, and red excitations, and other fluorescence ratios were acquired from the healthy and HLB-infected leaves of different cultivars. The classification studies were performed with these features as well as selective fluorescence features. Results indicated that the bagged decision tree classifier yielded 97% classification accuracy in case of the healthy and symptomatic samples. Although the asymptomatic samples from the HLB-infected trees could not be classified based on polymerase chain reaction (PCR) analysis results, the Naïve-Bayes classifier grouped most of the asymptomatic samples as HLB. We found that a few fluorescence features such as yellow fluorescence (UV), far-red fluorescence (UV), yellow to far red fluorescence (UV), simple fluorescence ratio (green), and yellow fluorescence (green) could result in classification accuracies similar to those of the entire dataset.

Keywords

Leaf Fluorescence Classification Symptomatic And Asymptomatic Leaves Fluorescence Sensor

INTRODUCTION

Citrus greening or Huanglongbing (HLB), a bacterial disease caused by Candidatus Libertibacter asiaticus, is causing serious problems for citrus production in Florida. This is a vector-based disease that spreads from one tree to another through insect (Asian citrus psyllid) feeding. Currently, management strategies include disease detection and pesticide application to control the spread of HLB.^1,-2 The most accurate method used for HLB detection so far is polymerase chain reaction (PCR) analysis.^3,–6 Due to cost and the need for specific expertise, there are ongoing efforts to develop sensors for monitoring HLB in citrus orchards. Moreover, given the rate at which HLB is spreading, it is difficult to monitor an entire field with PCR analysis, which is labor-intensive and time-consuming.^7,–9 Also, PCR analysis requires minimal bacterial population for accurate detection.^{8, 10}

HLB infection begins when an insect carrying the bacterial inoculum starts feeding on the citrus leaves. Depending on the age, cultivar, season, and physiological status of the citrus tree, leaf symptoms begin to appear between six months and about two years after exposure. The typical symptoms include leaf chlorosis (yellowing), blotchy mottle, and asymmetric, smaller fruits. Infection eventually affects the fruit yield and internal quality.^{2, 9, 11,-12} Orchards are regularly monitored by workers inspecting for HLB symptoms (scouting). The scouting process is slow, expensive, and often inaccurate depending on worker expertise and the season. Therefore, there is a need to develop rapid sensors capable of identifying HLB in citrus orchards. Such sensors could aid scouts in simpler and faster detection in the field. Thereby, fewer samples would need to be verified using laboratory PCR analysis. Fluorescence sensing is one potential method that could be related to physiological stress in crops using chlorophyll and other pigments/structures in leaves as indicators.^13,–17

Chlorophyll fluorescence is one of the most used sensing methods for early in vivo detection of plant stress. It could be related to various physiological factors such as root growth potential, leaf injury, gas exchange, photosynthetic activity, and stem water potential.¹⁸ Red and far-red fluorescence values result from chlorophyll, whereas blue and green fluorescence are emitted from the plant phenolics in the epidermal cells, major leaf veins and cell walls of green mesophyll cells.¹⁹ The soluble plant phenolics (such as flavones, flavonols, cinnamic acids, coumarins, and stilbenes) in the epidermal cells are sensitive to stress conditions, which can cause blue-green fluorescence to increase, followed by a decrease in chlorophyll and carotenoid content in mesophyll cells. Blue to red and blue to far-red ratios are particularly sensitive to environmental changes and can aid in early detection of stress. Soluble plant phenolics in the epidermis increase blue-green emissions.²⁰ Simple fluorescence ratios (SFR) represent changes in chlorophyll content.²¹

Several sensing techniques have been investigated for their ability to detect HLB such as visible-near infrared spectroscopy, mid-infrared spectroscopy, etc.^22,–25 Each of these methods offer unique benefits, but has certain limitations. For example, mid-infrared spectroscopy is a destructive method requiring sample preparation. Fluorescence offers non-invasive sensing and is more specific to the physiological changes that occur in a plant. In our previous work, we evaluated a handheld fluorescence sensor for detecting citrus stress (nutrient deficiency) and HLB.²⁶ The research findings indicated a good potential of fluorescence sensing in detecting nutrient deficiency as well as HLB-infected citrus leaves. Moreover, it was found that features such as yellow fluorescence with UV excitation contribute greatly to the ability to discern plant infection. Our previous study was conducted with two citrus cultivars, ‘Hamlin’ and ‘Valencia’. In this study we evaluated the fluorescence sensor with a larger data set from multiple cultivars to verify whether the performance of the fluorescence sensor could be repeated. The present work answers the following questions: Can fluorescence sensing be used to monitor HLB infection independent of citrus cultivars? Can a set of selective fluorescence features yield high classification accuracies? In this study, a large set of fluorescence data from citrus leaves of multiple cultivars were collected using a commercially available fluorescence spectrometer. In addition, a set of fluorescence data was collected from asymptomatic leaves from HLB-infected trees. The classification studies were performed with the entire set of fluorescence features, and multiple classifier algorithms to evaluate the classification accuracies. Fluorescence sensing is a suitable method for non-invasive sensing of HLB in citrus. Moreover, these findings will significantly help in customizing the sensor module for the desired application and also reduce the processing time using suitable classification algorithms. The use of improved statistical methods will enable the rapid detection of citrus disease and allow automation.

EXPERIMENTAL DETAILS

Fluorescence Sensor. A commercial handheld fluorescence sensor (Multiplex3, Force-A, Orsay, France) was used to collect fluorescence data from citrus leaves. The sensor comprises four excitation wavelength ranges with peak maxima at UV (375 nm), blue (465 nm), green (520 nm), and red (635 nm) regions and three fluorescence emission wavelengths with peak maxima at yellow (590 nm), red (688 nm), and far-red or near infrared (750 nm) regions. The sensor has six UV and three RGB LEDs for providing excitations. The photodiodes were synchronized with the LEDs to capture emissions resulting from yellow, red, and far-red fluorescence. The yellow fluorescence at RGB excitations included a small amount of reflectance especially in the red region due to some overlap between the excitation and emission wavelengths.^† Although the sensor is capable of acquiring several fluorescence features, due to saturation of far-red fluorescence at blue and red excitations, only the following features were used for analysis: yellow fluorescence with UV, blue, green, red excitation (YF_UV, YF_B, YF_G, YF_R); red fluorescence with UV, blue, green, red excitation (RF_UV, RF_B, RF_G, RF_R); far-red fluorescence with UV, green excitation (FRF_UV, FRF_G); simple fluorescence ratio with green excitation (SFR_G); and yellow to far-red fluorescence ratio with UV excitation (YFR_UV). In addition, the red to far-red fluorescence ratio (RFR_UV) at UV excitation was calculated by dividing RF_ UV by FRF_UV.^17,27,-28 Further details on the sensor system can be found in the literature.²⁶

Data Collection. Fluorescence data were collected in the laboratory from healthy and HLB-infected (symptomatic) citrus leaves. Each sample comprised four to five leaves placed at a distance of about 8 cm away from the sensor. The leaves were placed on the table and the sensor, with a sampling mask, was placed above it such that the leaves covered the entire area of interest (47.8 cm²). Thus, the sampling area remained constant. The number of sample data replicates was 10. The leaf samples were from a number of citrus cultivars, which include Hamlin, Valencia, Navel oranges, Murcott (Honey tangerines), Sunburst, Pineapple orange, Rhode Red Valencia and grapefruit. The number of healthy and HLB samples were 497 and 468, respectively. In addition to the data from the healthy and HLB-infected symptomatic leaves, fluorescence data were collected from asymptomatic leaves from the HLB-infected trees. Data collection involved 56 sample sets. All the leaves (healthy, symptomatic HLB-infected, and asymptomatic HLB-infected) were analyzed using the PCR method.

PCR analysis. For PCR analysis, primers and probes are used for amplification of the 16S rDNA target gene. The DNA extraction and detection procedure used for real-time PCR analysis is described in the literature.⁴ The presence or absence of disease-causing bacteria is assessed using the threshold cycle (Ct). A Ct value of <32 is labeled ‘PCR positive’ and value of >32 is labeled as ‘PCR negative’.⁴ The PCR results of the leaf samples are depicted in Fig. 1.

Fig. 1.

PCR analysis results of leaf samples. (a) healthy and HLB-infected symptomatic leaves, (b) asymptomatic leaves from HLB-infected trees.

Classification. The data analysis was performed with Matlab (ver. 7.14, R2012b, The MathWorks, Inc., Natick, MA) using the Statistics toolbox from Matlab and PLS toolbox (Eigenvector Research, Inc., Wenatchee, WA). Thirteen fluorescence features acquired from the sensor were used as the input data for classification. In our previous study, we found that some fluorescence features exhibited higher discriminatory capability than others. These fluorescence features were selected based on forward sequential feature selection and tested with Naïve-Bayes (NB classifier) and Bagged Decision Tree (BDT) classifiers.²⁶ In additional to the initial evaluation of all the fluorescence features, the selected fluorescence features were further evaluated using a larger dataset and multiple classifiers. The features selected were: YF_UV, FRF_UV, SFR_G, YFR_UV, and YF_G; and classifiers tested were NB, BDT, support vector machine (SVM) with kernel (rbf), linear discriminant analysis (LDA), and quadratic discriminant analysis (QDA). The classification results were assessed based on the overall and individual classification accuracies.

The randomized data were separated into training and test datasets. Preliminary evaluation was performed using a few training datasets by testing dataset ratios such as 75:25, 60: 40, 50: 50, 25 : 75, and 10: 90 using the data from healthy and symptomatic HLB-infected leaves. The corresponding number of data samples in the training/testing datasets for each of the above mentioned ratios were 724/241, 579/386, 483/482, 242/723, and 97/868 (Fig. 2). Based on the overall classification accuracies, the ratio resulting in maximum accuracy was selected. Because the 75% training dataset with 25% test dataset resulted in the highest classification accuracy with BDT, this ratio was used for the other classification studies.

Fig. 2.

Overall classification accuracy of different training datasets to test dataset ratios separated from data from healthy and HLB-infected symptomatic leaves. (a) complete set of fluorescence features, (b) selective set of fluorescence features.

For the classification of symptomatic samples, 75% and 25% of the healthy and symptomatic HLB-infected data were used for training and testing, respectively. In the case of asymptomatic samples, 75% of the healthy and symptomatic HLB-infected data were used for training and the entire asymptomatic data were used as the test dataset. The ‘PCR negative’ samples were labeled as ‘healthy’ and the ‘PCR positive’ samples were labeled as ‘HLB’. This was true for asymptomatic samples even if the samples were from the HLB-infected trees because in such a tree, part of the tree can be infected while other parts remain healthy. In addition to the classification of data collected in this study, the model developed (using the 75% as the training dataset) was independently evaluated with the fluorescence data collected as a part of our previous work.²⁶ The validation dataset (entire dataset of previous work) was used as the test dataset. Table I summarizes the datasets used for classifications. It should be noted that the 724 samples used for training the classifiers were all from 497 healthy and 468 symptomatic HLB-infected samples.

TABLE I.

Details on the data used in classification.

Dataset*	Number of samples	Fluorescence features	Training/Testing samples
CFF	497 healthy, 468 symptomatic HLB-infected	13	724/241
AS-CFF	35 PCR negative, 21 PCR positive	13	724/56
SFF	497 healthy, 468 symptomatic HLB-infected	5	724/241
AS-SFF	35 PCR negative, 21 PCR positive	5	724/56
VCFF^†	60 healthy, 58 symptomatic HLB-infected	13	724/118
VSFF^†	60 healthy, 58 symptomatic HLB-infected	5	724/118

CFF—Complete fluorescence features, AS—Asymptomatic data, SFF—Selective fluorescence features, VCFF—Validation using complete fluorescence features, VSFF—Validation using selective fluorescence features.

†

Details of validation dataset available in literature.²⁶

Principal component analysis (PCA) was performed on the entire group of samples as well as selective fluorescence features of symptomatic and asymptomatic samples. PCA is an unsupervised learning technique capable of reducing the multidimensionality within the data and projecting them as a set of uncorrelated variables called principal components (PCs). PCA helps in observing the distribution of the PC scores among classes. The selected fluorescence features were also analyzed to determine correlations and tested with rank features to identify the features among the ones tested with maximum discriminatory power.

RESULTS AND DISCUSSION

Classification of the Entire Fluorescence Data Set. During the classification of fluorescence features, multiple classifiers were assessed based on the overall and individual classification accuracy (Table II). The HLB classification accuracies are important as they are directly related to the number of false negatives. Although all the classifiers performed well, with an overall classification accuracy of 92% or higher, the BDT resulted in maximum classification accuracy among the classifiers. These findings were similar to our previous work where we tested NB and BDT with a small dataset collected from Hamlin and Valencia orange trees.²⁶

TABLE II.

Classification results from different datasets.

Classifiers, %	NB	BDT	SVM	LDA	QDA
All fluorescence features in healthy-symptomatic HLB-infected dataset (CFF)
Overall classification accuracy	93.4	97.1	91.7	91.7	92.5
Healthy class classification accuracy	93.5	97.6	90.3	93.5	91.9
HLB class classification accuracy	93.2	96.6	93.2	89.7	93.2
All fluorescence features in healthy-asymptomatic HLB-infected dataset (AS-CFF)
Overall classification accuracy	25.0	39.3	33.9	46.4	39.3
Healthy class classification accuracy	0.0	28.6	8.6	40.0	8.6
HLB class classification accuracy	66.7	57.1	76.2	57.1	90.5
Selective fluorescence features in healthy-symptomatic HLB-infected dataset (SFF)
Overall classification accuracy	95.9	97.5	97.1	91.7	95.0
Healthy class classification accuracy	95.2	96.8	97.6	98.4	96.8
HLB class classification accuracy	96.6	98.3	96.6	84.6	93.2
Selective fluorescence features in healthy-asymptomatic HLB-infected dataset (AS-SFF)
Overall classification accuracy	32.1	41.1	48.2	46.4	42.9
Healthy class classification accuracy	8.6	28.6	51.4	37.1	42.9
HLB class classification accuracy	71.4	61.9	42.9	61.9	42.9
Validation with all fluorescence features (VCFF)
Overall classification accuracy	61.9	81.4	54.2	79.7	68.6
Healthy class classification accuracy	25.0	63.3	10.0	60.0	38.3
HLB class classification accuracy	100.0	100.0	100.0	100.0	100.0
Validation with selective fluorescence features (VSFF)
Overall classification accuracy	83.9	86.4	92.4	97.5	92.5
Healthy class classification accuracy	68.3	73.3	85.0	95.0	85.0
HLB class classification accuracy	100.0	100.0	100.0	100.0	100.0

The overall classification accuracy and individual classification accuracies from BDT were about 97%. The LDA resulted in the lowest classification accuracy, suggesting weaker performance of linear models. The classification results indicated that even when the fluorescence data from multiple cultivars were combined, the sensor along with the suitable classification algorithm could be used in correctly identifying HLB. This shows the relevance of the fluorescence sensor independent of the cultivar.

When the entire set of fluorescence features of asymptomatic samples were used in classification, the classification accuracies were very low, especially for the healthy (PCR negative) classification. Because all the asymptomatic samples were from HLB-infected trees, the total number of samples classified as HLB was observed, irrespective of the PCR results. It was found that accuracy was higher than 85% (Table III), especially using the NB, SVM, and QDA classifiers. The QDA resulted in the highest classification accuracy with 91%, unlike the symptomatic samples, where BDT was better than the other methods. This is a very interesting observation, in spite of uncertainty with respect to the spread of bacteria in HLB-infected trees. Additional studies are needed to determine whether fluorescence can be used as a sensing tool for HLB even before symptoms appear.

TABLE III.

Total percent of asymptomatic samples classified as HLB.

	% samples classified as HLB

Classifiers	NB	BDT	SVM	LDA	QDA
All fluorescence features in healthy-asymptomatic HLB-infected dataset (AS-CFF)	87.5	66.1	85.7	58.9	91.1
Selective fluorescence features in healthy-asymptomatic HLB-infected dataset (AS-SFF)	83.9	67.9	46.4	62.5	51.8

During classification of the validation dataset, the model developed in this study was tested with the fluorescence data from Hamlin and Valencia healthy and symptomatic samples. Although the classification accuracies were not as high of those reported in this work, a moderate accuracy of about 81 % was achieved using the BDT classifier. It was interesting to note that although high classification accuracy of 90% and above could not be achieved, the HLB-infected samples were accurately classified.

Classification of Selective Fluorescence Features. Classification of the selective features indicated that the five significant features, YF_UV, FRF_UV, SFR_G, YFR_UV, and YF_G better represent the fluorescence features, while discriminating between healthy and HLB-infected samples. The accuracies of all the classifiers were comparable or even better than the entire fluorescence dataset. Similar to the results achieved using the entire fluorescence dataset, the BDT yielded a high classification accuracy of 97.5% with HLB classification accuracy of about 98%. Likewise, the classification accuracies of asymptomatic samples were slightly better or comparable to the entire dataset. Approximately 84% of the asymptomatic samples were classified as HLB using the NB classifier. Thus, the NB classifier may be used for classifying asymptomatic samples, while the BDT was good for the symptomatic HLB-infected samples.

Classification of selected features from the validation dataset resulted in accuracies higher than those using the entire set of fluorescence features, which was similar to the findings described above. However, the LDA, QDA, and SVM were best suited, resulting in overall classification accuracies of more than 92%. Among the classifiers, LDA resulted in the highest overall and individual classification accuracies. Similar to the results above, 100% HLB classification accuracies were achieved using each of the classifiers (Table II).

While analyzing the correlations between the selected variables (Fig. 3), it was found that there was a weak relationship between YF_UV and YFR_UV (0.70), and YF_UV and YF_G (0.76). When one or both features (YFR_UV, YF_G) were removed from the selected fluorescence features used as classifier input, the classification accuracies were reduced in most cases. The effect of removing YF_G was most drastic. When the selected features were ranked based on their discriminatory power between classes, it was found that the three features with maximum discriminatory power were YF_UV, YF_G and SFR_G. This further highlights the importance of these features during classification.

Fig. 3.

Covariance matrix representing correlation among variables.

The fluorescence sensor used in this study has various applications including fruit maturity prediction, assessment of plant stress, and foliar development.^21,29,–32 Zhang et al. evaluated corn nitrogen status with a fluorescence sensor in addition to other sensors used to monitor chlorophyll and nitrogen. Among the fluorescence features, they found that the fluorescence excitation ratios of anthocyanin relative index and SFR_G were statistically different, demonstrating the capability to assess corn nitrogen at early stages using analysis of variance. Moreover, they also found that shadowing and time of day did not affect these values.³² Bahar et al. found that the average SFR was correlated to grape total soluble solids with a correlation coefficient of 0.97.²¹ Moreover, SFR_R decreased with increasing maturity of both green and colored wine grapes. Similar to other studies, this study also presented SFR as one of the important parameters for HLB detection. The Candidatus Libertibacter asiaticus resulting in HLB-infection is a phloemlimiting bacterium that causes starch build-up in the leaves due to phloem blockage.^33,-34 The starch accumulation leads to degradation of the chloroplast thylakoid system and results in yellowing of leaves. SFR that indicates chlorophyll changes could be used as a valuable fluorescence indicator.

Principal Component Analysis. Figure 4 represents distribution of PC scores in different datasets. It can be observed that the distribution of healthy and HLB-symptomatic samples can be clearly delineated, while PCR negative and positive sample distribution could not. This explains the low classification accuracies of the asymptomatic samples versus those of symptomatic ones. Another observation is the variance in the datasets. The first principal component in the entire dataset could only account for 50-60% variance; however, while using the selective fluorescence data, the first PC could account for 93% and higher variance. This also indicates that the selected fluorescence features are more than sufficient for identifying HLB-infected trees.

Fig. 4.

PCA score plots of different datasets showing the distribution of samples. (a) entire fluorescence features (CFF), (b) selective fluorescence features (SFF), (c) entire fluorescence features of asymptomatic data (AS-CFF), (d) selective fluorescence features of asymptomatic data (AS-SFF).

CONCLUSIONS

In this study, we evaluated a handheld fluorescence sensor for detecting HLB in citrus. The study was performed to determine if fluorescence sensing can be used as a tool to identify HLB irrespective of the cultivar or appearance of symptoms. A high classification accuracy of about 97% was achieved with the BDT classifier in datasets when the entire set of fluorescence features and a few selective fluorescence features were used. Analysis of fluorescence features from asymptomatic leaf samples showed that although the classifiers could not separate PCR positive and PCR negative samples, most of the data from the asymptomatic samples were classified as HLB, especially with the NB classifier. Thus, the fluorescence sensing should be further investigated for its potential to detect HLB in asymptomatic conditions.

Footnotes

†

(Personal communication with Mr. Laurent Florin, Force-A, France)

ACKNOWLEDGMENTS

This research was funded by Citrus Research and Development Foundation (CRDF) and US Department of Agriculture-National Institute of Food and Agriculture (USDA-NIFA). We would like thank Ashish Mishra and Ratnesh Kumar for their help during this study. We would also like to express our special thanks to Mike Irey and his team in Southern Gardens Diagnostic Lab, United States Sugar Corporation, Clewiston, FL.

References

Tatineni

Sagaram

U.S.

Gowda

Robertson

C.J.

Dawson

W.O.

Iwanami

Wang

. “In Planta Distribution of ‘Candidatus Liberibacter Asiaticus’ as Revealed by Polymerase Chain Reaction (PCR) and Real-Time PCR”. Phytopathol. 2008. 98(5): 592–599.

Gottwald

T.R.

. “Current Epidemiological Understanding of Citrus Huanglongbing”. Annu. Rev. Phytopathol. 2010. 48: 119–139.

Teixeira

D.C.

Danet

J.L.

Eveillard

Martins

E.C.

Junior

W.C.J.

Yamamoto

P.T.

Lopes

S.A.

Bassanezi

R.B.

Ayres

A.J.

Saillard

Bové

J.M.

. “Citrus Huanglongbing in São Paulo State, Brazil: PCR Detection of the' Candidatus' Liberibacter Species Associated with the Disease”. Mol. Cell. Probes. 2005. 19(3): 173–179.

Hartung

J.S.

Levy

. “Quantitative Real-Time PCR for Detection and Identification of Candidatus Liberibacter Species Associated with Citrus Huanglongbing”. J. Microbiol. Methods. 2006. 66(1): 104–115.

Abad

J.A.

French-Monar

R.D.

Rascoe

Wen

Gudmestad

N.C.

Secor

G.A.

Lee

I.M.

Duan

Levy

. “Multiplex Real-Time PCR for Detection, Identification and Quantification of ‘Candidatus Liberibacter Solanacearum’ in Potato Plants with Zebra Chip”. J. Microbiol. Methods. 2009. 78(1): 59–65.

Lacava

P.T.

W.B.

Araú jo

W.L.

Azevedo

J.L.

Hartung

J.S.

. “Rapid, Specific and Quantitative Assays for the Detection of the Endophytic Bacterium Methylobacterium Mesophilicum in Plants”. J. Microbiol. Methods. 2006. 65(3): 535–541.

Wang

Yin

Yuan

Peng

Xia

. “Development and Application of Molecular-Based Diagnosis for ‘Candidatus Liberibacter Asiaticus’, the Causal Pathogen of Citrus Huanglongbing”. Plant Pathol. 2006. 55(5): 630–638.

Trivedi

Sagaram

U.S.

Brlansky

R.H.

Rogers

Stelinski

L.L.

Oswalt

Kim

J.S.

Wang

. “Quantification of Viable Candidatus Liberibacter Asiaticus in Hosts Using Quantitative PCR with the Aid of Ethidium Monoazide (EMA”. Eur. J. Plant Pathol. 2009. 124(4): 553–563.

Sankaran

Ehsani

. “Visible-Near Infrared Spectroscopy Based Citrus Greening Detection: Evaluation of Spectral Feature Extraction Techniques”. Crop Prot. 2011. 30(11): 1508–513.

10.

Kim

Wang

. “Characterization of Copy Numbers of 16S rDNA and 16S rRNA of Candidatus Liberibacter Asiaticus and the Implication in Detection in Planta Using Quantitative PCR”. BMC Res. Notes. 2009. 2: 37–1. 37–4.

11.

Gonzalez

Reyes-De-Corcuera

Etxeberria

. “Characterization of Leaf Starch from HLB-Affected and Unaffected-Girdled Citrus Trees”. Physiol. Mol. Plant Pathol. 2012. 79: 71–78.

12.

Zhang

Powell

C.A.

Guo

Doud

M.S.

Duan

. “Graft-Based Chemotherapy Method for Screening Effective Molecules and Rescuing Huanglongbing-Affected Citrus Plants”. Bacteriol. 2012. 102(6): 567–574.

13.

Lenk

Chaerle

Pfündel

E.E.

Langsdorf

Hagenbook

Lichtenthaler

H.K.

Van Der Straeten

Buschhmann

. “Multispectral Fluorescence and Reflectance Imaging at the Leaf Level and its Possible Applications”. J. Exp. Bot. 2007. 58(4): 807–814.

14.

Naumann

J.C.

Young

D.R.

Anderson

J.E.

. “Linking Leaf Chlorophyll Fluorescence Properties to Physiological Responses for Detection of Salt and Drought Stress in Coastal Plant Species”. Physiol. Plant. 2007. 131(3): 422–433.

15.

Chaerle

Leinonen

Jones

H.G.

Van Der Straeten

. “Monitoring and Screening Plant Populations with Combined Thermal and Chlorophyll Fluorescence Imaging”. J. Exp. Bot. 2007. 58(4): 773–784.

16.

Chaerle

Lenk

Hagenbeek

Buschmann

Van Der Straeten

. “Multicolor Fluorescence Imaging for Early Detection of the Hypersensitive Reaction to Tobacco Mosaic Virus”. J. Plant Physiol. 2007. 164(3): 25–262.

17.

Belasque

Gasparoto

M.C.G.

Marcassa

L.G.

. “Detection of Mechanical and Disease Stresses in Citrus Plants by Fluorescence Spectroscopy”. Appl. Opt. 2008. 47(11): 1922–1926.

18.

Willits

D.H.

Peet

M.M.

. “Measurement of Chlorophyll Fluorescence as a Heat Stress Indicator in Tomato: Laboratory and Greenhouse Comparisons”. J. Amer. Soc. Hort. Sci. 2001. 126(2): 188–194.

19.

Lichtenthaler

H.K.

Lang

Sowinska

Heisel

Miehé

J.A.

. “Detection of Vegetation Stress Via a New High Resolution Fluorescence Imaging System”. J. Plant Physiol. 1996. 148(5): 599–612.

20.

Lichtenthaler

H.K.

Miehé

J.A.

. “Fluorescence Imaging as a Diagnostic Tool for Plant Stress”. Trends Plant Sci. 1997. 2: 316–320.

21.

Bahar

Kaplunov

Zutahy

Daus

Lurie

Lichter

. “Auto-Fluorescence for Analysis of Ripening in Thompson Seedless and Colour in Crimson Seedless Table Grapes”. Aust. J. Grape Wine Res. 2012. 18(3): 353–359. doi: 10.1111/j.1755-0238.2012.00199.x.

22.

Sankaran

Ehsani

Etxeberria

. “Mid-Infrared Spectroscopy for Detection of Huanglongbing (Greening) in Citrus Leaves”. Talanta. 2010. 83(2): 574–581.

23.

Sankaran

Mishra

Maja

J.M.

Ehsani

. “Visible-Near Infrared Spectroscopy for Detection of Huanglongbing in Citrus Orchards”. Comput. Electron. Agric. 2011. 77(2): 127–134.

24.

Sankaran

Ehsani

. “Visible-Near Infrared Spectroscopy based Citrus Greening Detection: Evaluation of Spectral Feature Extraction Techniques”. Crop Prot. 2011. 30(11): 1508–1513.

25.

Hawkins

S.A.

Park

Poole

G.H.

Gottwald

T.R.

Windham

W.R.

Albano

Lawrence

K.C.

. “Comparison of FTIR Spectra between Huanglongbing (Citrus Greening) and Other Citrus Maladies. J. Agr”. Food Chem. 2010. 58(10): 6007–6010.

26.

Sankaran

Ehsani

. “Detection of Huanglongbing Disease in Citrus Using Fluorescence Spectroscopy”. Trans. ASABE. 2012. 55(1): 313–320.

27.

Lichtenthaler

H.K.

Wenzel

Buschmann

Gitelson

. “Plant Stress Detection by Reflectance and Fluorescence”. Ann. NY Acad. Sci. 1998. 851: 271–285.

28.

Cerovic

Z.G.

Samson

Morales

Tremblay

Moya

. “Ultraviolet-Induced Fluorescence for Plant Monitoring: Present State and Prospects”. Agron. 1999. 19: 542–578.

29.

Betemps

D.L.

Fachinello

J.C.

Galarça

S.P.

Portela

N.M.

Remorini

Massai

Agati

. “Non-Destructive Evaluation of Ripening and Quality Traits in Apples using a Multiparametric Fluorescence Sensor”. J. Sci. Food Agric. 2012. 92(9): 1855–1864.

30.

Baluja

Diago

M.P.

Goovaerts

Tardaguila

. “Assessment of the Spatial Variability of Grape Anthocyanins using a Fluorescence Sensor”. Precis. Agric. 2012. 13(4): 457–472.

31.

Whu

Zhu

Wang

. “Diagnosis of Freezing Stress in Wheat Seedlings using Hyperspectral Imaging”. Biosystems Eng. 2012. 112(4): 253–260.

32.

Zhang

Tremblay

Zhu

. “A First Comparison of Multiplex® for the Assessment of Corn Nitrogen Status”. J. Food. Agric. Environ. 2012. 10(1): 1008–1016.

33.

Schneider

. “Anatomy of Greening Diseased Sweet Orange Shoots”. Phytopathol. 1968. 58: 1155–1160.

34.

Schaffer

A.A.

Liu

K.-.C.

Goldschmidt

E.E.

Boyer

C.D.

Goren

. “Citrus Leaf Chlorosis Induced by Sink Removal: Starch, Nitrogen, and Chloroplast Ultrastructure”. J. Plant Physiol. 1986. 124: 111–121.