A Comparative Study on Application of Computer Vision and Fluorescence Imaging Spectroscopy for Detection of Huanglongbing Citrus Disease in the USA and Brazil

e overall objective of this work was to develop and evaluate computer vision and machine learning technique for classi�cation of Huanglongbing-(HLB)-infected and healthy leaves using �uorescence imaging spectroscopy. e �uorescence images were segmented using normalized graph cut, and texture features were extracted from the segmented images using cooccurrencematrix. e extracted features were used as an input into the classi�er, support vector machine (SVM). e classi�cation results were evaluated based on classi�cation accuracies and number of false positives and false negatives. e results indicated that the SVM could classify HLB-infected leaf �uorescence intensities with up to 90% classi�cation accuracy.ough the �uorescence intensities from leaves collected in Brazil and the USA were different, the method shows potential for detecting HLB.


Introduction
In recent years, there has been an increasing use of machine vision and learning approaches in the domain of agriculture applied towards detecting and classifying various diseases on trees and fruits [1][2][3].Such techniques need to be developed for sustainable agriculture and to prevent great economic losses.Major citrus diseases such as citrus canker (Xanthomonas axonopodis pv.citri) and Huanglongbing (HLB) or citrus greening (Candidatus Liberibacter asiaticus) are a serious threat to citrus production worldwide including regions in Brazil and Florida.e citrus industry in both these countries, Brazil and the USA is making continuous efforts to control and contain these citrus diseases.e disease management practices include insect vector (Asian citrus psyllid/Diaphorina citri Kuwayama) control through pesticide application and eradication of symptomatic trees.erefore, there is a need for a reliable diagnostic tool to identify the disease symptoms and consequently eradicate these trees (source of inoculum) to prevent further spread of the diseases.However, the process of identifying diseased tree through scouting, the process of frequent �eld inspection for recognizing symptoms, followed by laboratory diagnosis using polymerase chain reaction (PCR) analysis is timeconsuming and expensive.A machine vision technique offers an alternative for disease detection under �eld conditions and thus better control of such diseases.Machine vision can aid in the scouting process and complement PCR analysis.Moreover, fewer samples can be sent for PCR analysis, thereby reducing the disease detection costs.
Pydipati et al. [4] used machine vision to identify citrus diseases such as greasy spot, melanose, and scab in leaves under controlled laboratory conditions.e researchers applied the color cooccurrence matrix (CCM) features to detect citrus diseases using the features recommended by Haralick et al. [5].e authors were able to outline very important properties of disease detection such as color and texture.ough there have been applications of machine vision for citrus disease detection under laboratory conditions, there is a need for a sensing technique that can be applied under �eld conditions.e �uorescence imaging spectroscopy (FIS) has a potential to obtain spatial and spectral (�uorescence) information than regular imaging techniques.With rapid development and miniaturization of new light sources and detectors in the last decade, its application under �eld conditions is growing.
In this paper, we present the development and evaluation of machine vision technique for detection of HLB in citrus leaves using FIS under �eld conditions.e present research results in a unique comparison of the performance of the FIS system for HLB detection in leaves from Brazil and Florida.

Leaf Sample Collection.
Fluorescence data were collected from HLB-infected citrus leaves in Brazil and the USA.In Brazil, twenty citrus trees from six different orchards (four municipalities in São Paulo State) were evaluated.e HLB-infected leaf samples were collected from trees of Citrus sinensis (L.Osbeck).e images were collected from both symptomatic and nonsymptomatic leaves from each �eld plant.A total of 100 samples of HLB-symptomatic leaves and 55 nonsymptomatic leaves (con�rmed as healthy) were collected.It should be noted that in order to avoid any detachment time effect, the �uorescence measurements were performed within �ve minutes aer the leaves were collected from each plant [6].e data were collected between June 18 and July 2, 2009.Aer the onsite measurements, the leaves were transported in closed styrofoam boxes to the laboratory to con�rm the presence of HLB using the laboratory method (PCR analysis).In the USA, leaf samples were collected from eight citrus trees at Citrus Research and Education Center's citrus orchard in Lake Alfred, Florida from 14 to 17 July 2011.e leaf samples were collected from the same cultivar of citrus trees as from Brazil.A total of 68 symptomatic (HLBinfected) and 17 non-symptomatic (healthy) leaf samples were collected.Fluorescence data were collected from HLBinfected samples in Brazil and the USA, to evaluate the applicability of �uorescence imaging spectroscopy under different conditions.Once infected with HLB, citrus leaves develop symptoms such as botchy mottle, yellowing and thickening of veins or the entire leaf.e time period between the HLB-infection and the appearance of symptoms vary with age, time of infection, cultivar, and physiological status of the tree.
e �uorescence spectroscopy imaging system was a portable custom-designed unit, which comprised of a monochromatic charged couple device camera (CCD) (model mvBlueFOX120a, Matrix Vision, Germany) that uses a USB communication port, a �lter wheel (model CFW-1-8, Finger Lakes Instrumentation, USA) that holds up to eight optical �lters and uses a USB communication port, four bandpass optical �lters (models FB570-10, FB610-10, FB690-10 and FB740-10, orlabs, USA; 570, 610, 690, and 740 nm emission wavelengths), an objective lens, high power light emitting diodes (LEDs) at different wavelengths (365, 405, 470, and 530 nm) as excitation sources, and a standard laptop computer.Figure 1 shows a schematic diagram of the portable FIS system with major components that operates on car batteries.All the parts of the FIS system, except the laptop, were placed inside a closed box.e CCD and �lter wheel were computer controlled using a program.

Fluorescence Imaging Spectroscopy System.
2.3.Segmentation.e data analysis procedures were written in MATLAB (ver.7.6, e MathWorks Inc., Natick, MA). Figure 2 presents a sample �uorescence image at 690 nm emission wavelength using excitation light at 530 nm.It can be observed from Figure 2 that the image texture of healthy and HLB-infected samples was different, with healthy samples having a more uniform texture than those of the HLB-infected leaves.e �rst step in data analysis is the segmentation to identify leaf areas within an input image.In this step, the pixels are grouped into clusters.
ere are several techniques for segmentation.Some of the methods tested in this study were -means clustering, histogram thresholding, Markov random �eld (MRF), graph cuts, and 2D histograms.Histogram thresholding and means clustering did not produce immaculate boundaries.MRF, graph cuts, and 2D histograms represent a higher order statistics [7].Following segmentation, the leaf pixels were separated as rectangular areas, as shown in Figure 3. ese rectangular regions were further processed for feature extraction and classi�cation.
2.4.Feature Selection.e features of texture (such as autocorrelation, contrast, and energy) were extracted from the segmented image based on co-occurrence matrix [4,5].A cooccurrence matrix (or a gray tone spatial dependence matrix) is a matrix distribution that is de�ned over an image where,   ,   ,   ,   are the means and standard deviations of marginal distributions associated with  /.  is the matrix of relative frequencies with which two neighboring resolution cells are separated by distance "" on the image, one with gray tone "" and other with gray tone "."  is the number of neighboring resolution cell pairs used in computing a particular gray tone spatial dependence matrix.e descriptors (uniformity, contrast, correlation, and homogeneity) were computed for each offset corresponding to eight-point neighborhood.Following that, the mean value was calculated for a single descriptor over each offset.us, for an eight-point neighborhood, the number of feature dimensions was four.
e extracted texture features were used as an input in the classi�er support vector machine (S�M).e classi�cation of the data was performed with 20 sample data in the training dataset (10 healthy and 10 diseased).e S�M classi�er works on the concept of decision planes that de�ne decision boundaries.A decision plane is a hyperplane that separates between a set of objects that belong to different classes.A support vector machine [8] creates a plane in -dimensions to separate -dimensional points lying in different classes.For a new point, the class is predicted based on which side of the plane the point lays.For the nonlinear separable data, a nonlinear mapping function is used to map the data into a higher dimension feature space.e decision function is based on dot product of the input features with support vectors, and the mapping of data into higher dimensional feature space is not needed upon use of kernel [9].Some commonly used kernels are Gaussian radial e classi�cation of the data was performed using the SVM classi�er.e classi�er was used on the �uorescence data collected from Brazil and USA separately.e performance of the classi�er was evaluated using classi�cation accuracies, precision recall (PR) and receiving operating characteristics (ROC) curves.e ROC curve is a plot of false positive rate (FPR) on the -axis and the true positive rate (TPR) on the -axis.e FPR measures the fraction of negative samples that are misclassi�ed as positive, while the TPR measures the fraction of positive samples that are labeled correctly.e PR curve is the plot of recall on -axis and precision on -axis.e goal in the ROC curve is to have low false positive rates [10].

Image Preprocessing.
During the preprocessing of the �uorescence images, the data were segmented using normalized graph cuts in preference to other segmentation methods (-means clustering, histogram thresholding, MRF, and 2D histograms).e normalized cuts algorithm results in smooth spatial consistency in comparison to other methods such as -means clustering in color space and color histogram thresholding, which do not have high order information about spatial layout.In normalized graph cuts, the image is treated as a graph partitioning problem.e "cut" criterion measures both the total dissimilarity between the different groups as well as total similarity within the groups.e texture features such as uniformity, contrast, correlation, and homogeneity were extracted from the eight-point neighboring pixel of the segmented image.ese features were used in SVM classi�er.Preliminary evaluation of �uorescence bands indicated higher discriminatory power at 690 nm emission wavelength with ��0 nm excitation.e �uorescence at 690 nm that was measured represents chlorophyll �uorescence.As the disease affects plant physiology including chlorophyll �uorescence, this wavelength will be a useful in HLB detection.erefore, the segmentation and feature selection results from the �uorescence images at 690 nm were used as an input in the classi�er SVM.
3.�.��assi�ca�ion.e classi�cation results are presented in Table 1.An increase in the training sample size of higher than 20 samples did not improve the classi�cation accuracies.erefore, training was performed with 20 samples.e overall accuracies of SVM were 90% and 61% during the classi�cation of data collected from Brazil and the USA, respectively.e PR and ROC curves on the Brazil data are shown in Figure 4. e classi�cation accuracy was high (90%) for Brazil leaf samples with very low false negatives (9%).For selection of a suitable classi�er, it is more important to have a low false negative rate than false positive rate, because the HLB-infected samples should not be classi�ed wrongly as healthy.e false positive rate was also low (Table 1).
�hen the USA leaf samples� �uorescence intensities were compared, it was found that the leaf samples emitted �uorescence at a very low intensity under same illumination and detection conditions, with about 20% of �uorescence intensity as those of leaf samples from Brazil.e cause for this difference is under investigation.e possible reason for this observation could be due to variation in plant physiology resulting from soil differences in these two areas, Florida and S�o Paulo.Due to very low �uorescence intensity of USA samples, the SVM classi�cation accuracy was much lower (61%) than that of Brazil samples.e false positive as well as negative rates were also higher.Other possible reasons for variable classi�cation accuracies could be effect of seasonal variation in leaf symptoms, effect of sample size, and variation in growing conditions among others [11,12].Variability of symptoms has also been observed by a number of researchers [11,13,14].e irregularity in the appearance of symptoms can also occur within a tree, due to irregular distribution of the bacterial inoculum [11].Chamberlain and Irey [13] stated that variation in symptoms can occur periodically during a year, also from grove to grove and tree to tree at a given period of time.is signi�es the importance of calibrating the sensing systems under different conditions prior to �eld application.ese �ndings also suggest the need for such sensing techniques to improve the scouting detection efficiencies.

Conclusions
e present study explores the potential of �uorescence sensing as a means for rapid identi�cation of HLB infection in citrus orchards.A custom-designed �uorescence imaging sensor was evaluated with leaf samples collected from citrus trees in Brazil and the USA.e data were segmented using normalized graph cuts and texture features were extracted using cooccurrence matrix during data preprocessing.e extracted features were used as an input in the classi�er, support vector machine.e classi�cation results indicated that though the classi�er accuracy was low (61%) for USA samples, the accuracy was good (90%) when it came to classifying leaf samples from Brazil.e probable reason for this observation could be the low �uorescence intensity of leaf samples from the USA.e false negative rate for classifying Brazil samples was low, indicating 91% accuracy while classifying HLB-infected leaf samples.Our future studies will involve assessing other �uorescence features for HLB detection in citrus orchards and comparison of the FIS technique on other stress conditions (nutrient de�ciencies and other diseases).

F 2 :
Fluorescence intensity at 690 nm of healthy (a) and HLB-infected leaves (b).

F 3 :
Segmented images from Brazil (a) and the USA (b).to be the distribution of cooccurring values at a given offset.From this matrix, four metrics are computed from the image segment as shown in following equations: Uniformity or Angular Second Moment =

F 4 :
Precision-recall and ROC curves for the support vector machine classi�er.basis functions and polynomial and hyperbolic tangent.e radial basis function and polynomial kernel were used in this study.
T 1: Classi�cation results from support vector machine classi�er.