Computers and Electronics in Agriculture 110 (2015) 221–232
Contents lists available at ScienceDirect
Computers and Electronics in Agriculture
journal homepage: www.elsevier.com/locate/compag
Original papers
An optimum method for real-time in-field detection of Huanglongbing
disease using a vision sensor
Alireza Pourreza a, Won Suk Lee a,⇑, Reza Ehsani b, John K. Schueller c, Eran Raveh d
a
Department of Agricultural and Biological Engineering, University of Florida, Gainesville, FL 32611, United States
Citrus Research and Education Center (CREC), University of Florida, Lake Alfred, FL 33850, United States
c
Department of Mechanical and Aerospace Engineering, University of Florida, Gainesville, FL 32611, United States
d
Department of Fruit Trees Sciences, ARO Gilat Research Center, Negev, Israel
b
a r t i c l e
i n f o
Article history:
Received 9 June 2014
Received in revised form 11 November 2014
Accepted 23 November 2014
Keywords:
Citrus greening
Clustering
Disease identification
HLB
Image analysis
Starch accumulation
a b s t r a c t
Huanglongbing (HLB) or citrus greening is a bacterial infection which is spread by a citrus psyllid. No
effective cure for this disease has been reported yet, and the HLB-infected tree will eventually die.
Therefore, the infected tree must be detected and removed immediately to stop the spread of the disease.
One of the symptoms of HLB is the accumulation of starch which creates blotchy mottles in an asymmetrical pattern on infected citrus leaves. These blotchy mottles symptoms may be confused with the deficiency of certain nutrients such as zinc or magnesium. We showed in a previous study that the unique
capability of starch to rotate the polarization planar of light can be employed to identify the HLB-infected
citrus leaves and differentiate them from zinc or magnesium deficiency. In this study, a vision sensor was
developed for the purpose of real-time HLB detection for use under field conditions. The sensor included a
highly sensitive monochrome camera, narrow band high power LEDs, and polarizing filters. The sensor
was first tested and calibrated in a simulated field condition in a laboratory. Then, it was tested in a citrus
grove. Two simple image descriptors; mean and standard deviation of gray values, were used for the purpose of classification. The results showed that the sensor clearly highlighted the starch accumulation in
the HLB-infected leaf and differentiated it from visually analogous symptoms of zinc deficiency. HLB
detection accuracies which ranged from 95.5% to 98.5% were achieved during the laboratory and field
experiments.
Ó 2014 Elsevier B.V. All rights reserved.
1. Introduction
Huanglongbing (HLB) or citrus greening is one of the most
severe infections affecting citrus production. It is caused by an
Asian citrus psyllid-spread protobacterium of the genus Candidatus
Liberibacter (Albrecht and Bowman, 2008). The insect can pick up
the bacteria from a citrus greening-infected tree and transfer the
disease to other trees when feeding on them. In the Western Hemisphere, it was first seen in 2004 in Brazil (Texeira et al., 2005) and
then in August 2005, it was reported in Florida (Halbert, 2005).
Since then, the HLB disease has been reported in all citrus producing counties in Florida and has been reported at some locations in
California and Texas as well. From 2004 to 2011, the Florida commercial citrus acreage and the number of trees decreased by 28%
⇑ Corresponding author at: Department of Agricultural and Biological Engineering, University of Florida, 1741 Museum Road, Gainesville, FL 32611, United States.
Tel.: +1 (352) 392 1864x227; fax: +1 (352) 392 4092.
E-mail addresses: apourreza@ufl.edu (A. Pourreza), wslee@ufl.edu (W.S. Lee),
ehsani@ufl.edu (R. Ehsani), schuejk@ufl.edu (J.K. Schueller).
http://dx.doi.org/10.1016/j.compag.2014.11.021
0168-1699/Ó 2014 Elsevier B.V. All rights reserved.
(Salois et al., 2012), and HLB was one of the major reasons for this
loss. Early fruit drop which caused an average citrus production
loss of 10 percent in 2012 was another consequence of HLB disease
in Florida (Choi et al., 2013). Leaf yellowing that appears as blotchy
mottle is one of the early HLB symptoms. Once a branch of a tree
becomes infected, the disease gradually spreads through the entire
tree, which will then die within 2–3 years. Severely affected trees
have smaller leaves with some nutrient deficiency symptoms as
well as yellow veining. Additionally, fruits from an HLB-infected
tree show abnormal colors and uneven shapes and have a bitter
taste. These symptoms may be used for disease diagnosis; however, they are very inaccurate, especially the older trees which suffer from multiple problems (Chung and Brlansky, 2009). Although
no effective cure for this disease has been reported yet, early detection and removal of infected trees or branches were highly recommended (Buitendag and Von Broembsen, 1993) to prevent further
spread of the disease.
The efficiency of an HLB diagnosis performed by several professional inspectors was evaluated in DeSoto County, Florida (Futch
222
A. Pourreza et al. / Computers and Electronics in Agriculture 110 (2015) 221–232
et al., 2009). The results showed that the accuracy of identifying an
HLB-infected tree by visual inspection is between 47% and 59%.
HLB-infected leaves contain a high level of starch accumulation
(Etxeberria et al., 2009). It was shown that a leaf starch content
evaluation could reveal its HLB status (Gonzalez et al., 2012). A
quantitative real-time polymerase chain reaction (qrt-PCR) test is
another HLB diagnosis method which can identify the HLB status
with the highest accuracy (Hansen et al., 2008). Both the starch
measurement and qrt-PCR tests are laboratory-based diagnostic
methods and require crop scouting and leaf sampling which are
expensive and time consuming. On the other hand, the HLB status
of the entire grove should be observed continuously to make
timely decisions and prevent huge losses. Therefore, an easy-touse, fast, accurate, and inexpensive HLB diagnostic approach is
greatly needed, especially for small growers to monitor their
groves and control the spread of the disease.
A laser-induced fluorescence (LIF) spectroscopy method was
employed to discriminate water-stressed and citrus cankerinfected citrus leaves from healthy ones (Marcassa et al., 2006).
However, their method was not able to identify citrus variegated
chlorosis (CVC) disease from citrus canker disease. Later, they
developed another method which was able to distinguish between
mechanical interference stress and citrus canker disease stress
(Belasque et al., 2008). Despite the limitations of employing LIF,
some feel that this technology has much potential for citrus disease
detection (Lins et al., 2009).
The potential use of green, red, and near-infrared spectral bands
were determined in an HLB-infected tree identification study
(Mishra et al., 2007). Later, they developed a four-band optical sensor which was able to measure the reflectance of citrus trees at
570 nm, 670 nm, 870 nm, and 970 nm. Using this sensor, they conducted multiple measurements from the same tree and obtained
an error of less than 5% in the identification of HLB-infected trees
(Mishra et al., 2011).
In another spectroscopy study, it was shown that the reflection
of dried ground leaves in the mid-infrared band can be used to
determine the HLB status of a sample with >95% accuracy
(Hawkins et al., 2010). Mid-infrared spectroscopy was also
employed in the classification of ground citrus leaves into three
classes: HLB-infected, nutrient deficient, and healthy which
resulted in an accuracy of >90% in HLB detection (Sankaran et al.,
2010). In another study, near-infrared reflectance of ground citrus
leaves was used to classify samples into four classes: HLB-negative,
HLB-positive, nutrient deficient, and other citrus disease. The true
classification rates, ranging from 92% to 99% for HLB-negative and positive samples, were acquired using a partial least squares
regression model (Windham et al., 2011). Although comparatively
high detection accuracies were achieved in these three studies,
their methods contained sample preparation and processing,
which were time-consuming and required laboratory equipment.
Also these methods are impractical for in-orchard use.
Color images acquired through a digital microscope system
were found to be useful in determining the HLB status of citrus
leaves. Textural features were extracted from each image, and they
were used to identify symptoms of HLB infection in leaf samples.
An overall accuracy of 87% was achieved for classification of samples into several classes: green islands (HLB), greening blotchy
mottle (HLB), normal mature leaves, young flush leaves, zinc deficiency, manganese deficiency, and iron deficiency (Kim et al.,
2009). In another study, a laboratory system for laser-induced florescence imaging was developed in which the leaf samples were
excited with a solid-state blue laser system at 473 nm. The leaf
reflectance was then captured with an eight megapixel digital
camera, and color descriptors were used to identify the HLB
infected samples. Their results indicated a confidence level of
95% in the early stage HLB leaf identification (Pereira et al., 2011).
Airborne imagery is another approach in disease detection
which has been widely used in recent years. Several HLB detection
methods were conducted using airborne hyperspectral (HS) and
multispectral (MS) imagery. The spectral angle mapping (SAM)
classification method resulted in accuracies of 62% and 55% for
MS and HS images, respectively (Li et al., 2011). It was shown later
that the infected canopy had lower reflectance in the visible range
and higher reflectance in the NIR range compared to a healthy canopy. Also, it was determined that the severely infected areas in the
density map were easily detectable using most of the methods (Li
et al., 2012b). However, the mixture tuned matched filtering
(MTMF) method was proved to have a better performance compared to the SAM method in HLB detection using hyperspectral
images (Kumar et al., 2012). In another study, an extended spectral
angle mapping (ESAM) method was proposed for HLB detection
using HS images. Accuracies of 82.6% and 86.3% were obtained
for training and validation sets, respectively (Li et al., 2012a). A
multi-band imaging sensor carried by a multi-rotor unmanned aerial vehicle (UAV) was employed in another study to acquire airborne images of a citrus grove in six spectral bands between
530 nm and 900 nm. The NIR-R index and 710 nm reflectance values were found to be capable of discriminating HLB-infected trees
from healthy ones with accuracies ranging from 67% to 85%
(Garcia-Ruiz et al., 2013).
In our previous study (Pourreza et al., 2014), a customized
image acquisition system was developed which was able to highlight an increased level of starch accumulation in HLB-symptomatic leaves. The unique starch property of rotating the
polarization planar of light was used in this image acquisition system. A set of textural descriptors and step-by-step supervised classification models were used to identify HLB-positive leaves from
HLB-negative and nutrient deficient samples. Although an overall
accuracy of 90% was achieved using this method, some factors limited its performance in the field. The fixed automatic gain control
(AGC) property of the camera made the classification process more
complicated. Also, it was not able to perform an on-the-go diagnosis. Therefore, this study aimed to improve the performance of the
prior method by implementing a completely new sensing device.
Specific objectives were to (1) develop a vision sensor for real-time
in-field HLB detection, (2) add on-the-go diagnosis capability to the
sensor, (3) improve HLB detection accuracy, and (4) develop a simpler and more robust identification algorithm.
2. Materials and methods
2.1. Vision sensor
It was shown in the previous study (Pourreza et al., 2013) that
the starch accumulation in HLB-positive leaves can rotate the
polarization planar of light by 90° at 591 nm. This property was
used to design the vision sensor enclosed in a wooden box
(13 19 15 cm) including a camera and an illumination system
(Fig. 1a). A highly sensitive monochrome camera (DMK 23G445,
TheImagingSource, Bremen, Germany) with an ICX445 Sony CCD
sensor was used to measure the leaf reflectance. The spectral sensitivity curve of this CCD sensor had the quantum efficiency of
>90% at 591 nm which made it an appropriate option for our purpose. The camera was equipped with a wide lens (6 mm focal
length) that created a diagonal field of view of 53.1° and a rotating
linear polarizer and mounted inside the vision sensor housing. The
very short focal length was selected to increase the depth-of-field
so that more objects with different depths were in focus. Fig. 1a
also shows the LED panel of the vision sensor. Ten high luminous
efficiency LEDs (LED Engin, San Jose, California) at 591 nm (LZ400A100, 10 W) were mounted on an aluminum plate in a circular
pattern. The LEDs were powered with two 12 V car batteries (24
A. Pourreza et al. / Computers and Electronics in Agriculture 110 (2015) 221–232
223
Fig. 1. Vision sensor and LED circuit: (a) a schematic of the vision sensor with dimensions in which the polarizing filters are highlighted to emphasize their relative
perpendicular directions; (b) the LED circuit including 10 high power LEDs and five drivers.
V in total) in series and five 70 W LED drivers (RCD-48, RECOM,
Brooklyn, New York) in parallel as shown in Fig. 1b. The LED panel
was fixed on a side of the vision sensor, and a polarizing film (visible linear polarizing laminated film, Edmund Optics, Barrington,
New Jersey) was mounted in front of it. A hole in the center of
the LED panel and another one in the center of the polarizing film
were cut so that there was enough room for the camera lens to
come out. The direction of the camera’s linear polarizer was set
to be perpendicular to the direction of the LEDs’ polarizing film
as illustrated in Fig. 1a. Therefore, the camera was only able to
receive the minimum reflection.
2.2. Data collection
A set of citrus leaf samples (‘Hamlin’ sweet orange) was collected from a grove at the Citrus Research and Education Center
(CREC), University of Florida (Lake Alfred, FL) in September of
2013 by experienced HLB researchers. An experiment was conducted by acquiring images of 60 citrus leaf samples from four
classes: HLB-negative (20 samples), HLB-positive (20 samples),
zinc deficient HLB-negative (10 samples), and zinc deficient HLBpositive (10 samples) in a laboratory.
An in-field experiment was conducted in the CREC grove in
November of 2013 in which 20 images of HLB-positive citrus trees
and 10 images of HLB-negative citrus trees (the control) were
acquired. Eight out of 20 samples in the HLB-positive class were
also zinc deficient. The citrus trees and target leaves were located
and marked by experienced researchers in the morning before the
image acquisition. In order to verify the HLB status of the samples,
a qrt-PCR test (Hansen et al., 2008) was performed on a total of 90
samples, including 60 of the in-lab experiment samples and one
leaf sample from each image in the in-field experiment. The qrtPCR test was conducted at the United States Sugar Corporation
(USSC), Technical Operations, Southern Gardens (Clewiston, FL).
2.3. In-lab experiment
The lab experiment was conducted to evaluate several simulations of field imaging conditions and to determine the best settings
for the sensor. Since the vision sensor had its own illumination system, the in-field experiment was conducted after sunset to prevent
any interference from sunlight. Therefore, the lab experiment was
conducted in a completely dark room to simulate real lighting conditions. An exposure time of 0.1 s was set for the camera because
this was the shortest exposure time for capturing visually informative images without adding any gain (which increases the noise
level). In order to determine the effect of the object depth on its
histogram features, the images of one leaf were acquired from different distances, ranging from 50 cm to 150 cm. Then the histograms of the images taken at different depths were plotted and
compared with each other. Two main histogram features including
mean and standard deviation (SD) were considered for this evaluation, and the relationships between these features and the object
depth was modeled.
All image acquisitions for the in-lab experiment were designed
to be conducted with a fixed depth, assuming that the mean and
SD features of any leaf at different depths can be computed accurately with its known depth.
In order to determine the optimum distance, three distances
(60 cm, 80 cm, and 100 cm) were examined. Also, four leaf positioning conditions, including separated, adjacent, and overlapped
leaves as well as the leaves on an artificial citrus tree were defined
to evaluate how the leaf position in the image can influence the
detection accuracy.
A circular area on each leaf was randomly selected from the
symptomatic areas of HLB-positive and zinc-deficient samples, as
well as a random area of HLB-negative samples. In order to select
the same spot on the images of the same leaf taken from three different distances, the sizes of 177, 112, and 52 pixels were chosen
for the circular areas on the images taken from 60 cm, 80 cm,
and 100 cm, respectively. Then, the histograms of the two symptomatic areas and the HLB-negative regions were compared to
each other to illustrate the dissimilarity of the histograms of the
three different types of leaves. In order to determine whether the
positioning condition of the leaves affects the identification accuracy of the symptomatic areas, a probability-based color transfer
function was developed according to the histograms of symptomatic areas. In this function, three probabilities (corresponding to
the three classes) were defined for each pixel value based on the
histogram analysis (Eqs. (1) and (2)):
Pc ðiÞ ¼ P
Hc ðiÞ
n2C H n ðiÞ
8 C ¼ fHLB; HLNþ; ZnDef :g
PHLB ðiÞ þ PHLBþ ðiÞ þ PZnDef : ðiÞ ¼ 1
ð1Þ
ð2Þ
where i is a gray value between zero and 255, Pc(i) indicates the
probability that pixel value i belongs to class c, and Hc(i) is the histogram value of class c for pixel value i. Then, a color transfer function was developed based on these probabilities to convert the
grayscale image to a red (R), green (G), and blue (B) image in which
the amount of R, G, and B represents the probabilities of HLB infection, healthiness, and zinc deficiency, respectively.
224
A. Pourreza et al. / Computers and Electronics in Agriculture 110 (2015) 221–232
2.4. In-field experiment
3. Results
In order to test the sensor in real in-field conditions, the images
of citrus trees were acquired after sunset. Images were taken at an
average distance of 80 cm from the trees and from the exact distance of 80 cm from the target leaves (Fig. 2). The target leaf from
each image was marked and collected for a qrt-PCR test to validate
its HLB status. The normalized histogram of the target leaf area in
each image was obtained for further analysis.
3.1. Dataset validation
2.5. Data analysis and classification
Two simple statistical histogram features, the mean and SD of
the gray value (Eqs. (3) and (4)), were extracted from the normalized histograms (h(i)) (Pourreza et al., 2012) of individual leaves
and leaves on the artificial tree from the lab dataset and the target
leaves from field dataset.
Mean of the gray values :
l¼
X
ihðiÞ
ð3Þ
i
SD of the gray values :
r¼
qX
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
ði lÞ hðiÞ
i
ð4Þ
In order to evaluate the separability between the classes, a twodimensional plot of the samples based on their means and SDs was
used. Then, a maximum margin method (Bishop, 2006) was conducted to find the best divider threshold line between the classes.
In this method, an objective function tries to find the optimum
divider threshold which maximizes the margin between each pair
of classes. A step-by-step classification model was designed based
on the scatterplots of samples (Fig. 3). A support vector machine
(SVM), which is also a maximum margin classifier, was trained
with means and SDs features and employed for all steps of the classification model. A three-fold cross validation method was
employed in the classification process in which the dataset was
randomly divided into three folds, which two folds were used for
training while the other fold was used for validation. This algorithm was repeated fifty times, and the average accuracies were
calculated for each class.
All the data analyses and feature extractions were performed in
MATLAB (version R2011a, MathWorks, Natick, MA). Also, the plot
visualizations of the features were carried out in Excel (Microsoft
Office, Microsoft, Redmond, Washington).
Fig. 2. In-field image acquisition condition. The vision sensor was placed at a
distance of 80 cm from one target leaf (the image acquisition was conducted after
sunset).
The cycle threshold (CT value) in a qrt-PCR test indicates the
number of required cycles for the fluorescent intensity to reach
the threshold. The threshold of 33 was selected for CT values to
determine the HLB status of the samples (Li et al., 2006). In other
words, the samples with CT values below 33 were considered as
HLB-positive leaves. Table 1 illustrates the interpretation of the
CT values and HLB status for the samples used for the in-lab experiment. The CT values in all HLB-negative samples were above 33
which confirmed they were not infected. Also, all HLB-positive
samples had CT values below 33 which verified their HLB infection.
The CT values for half of the zinc-deficient samples were below 33
and those of the other half were above 33. Thus, there were 10
HLB-positive and 10 HLB-negative samples within the zinc deficient class.
Table 2 shows the CT values for the samples in the field experiment dataset. The CT values for ten samples were above 33, so
they were categorized as the HLB-negative samples. Sample numbers one to 20 had CT values below 33, and they were considered
as HLB-negative. Eight out of 20 HLB-positive samples were also
zinc-deficient, and they were categorized in another subclass of
HLB-positive zinc-deficient samples. Since all the zinc-deficient
samples had CT values below 33, there was no zinc-deficient
HLB-negative class in this dataset.
3.2. In-lab experimental results
3.2.1. The object depth effect
Fig. 4 shows the leaf sample which was used to evaluate the
effect of depth and gray images acquired at different depths. In
order to find the relationship between the object depth and histogram features (mean and SD), a power regression method
(Gennadios et al., 1996) was employed in Excel. Figs. 5 and 6 show
a line fit, regression equation, and a coefficient of determination
(R2) value for mean (l) and SD (r) based on the object depth (d).
These curves and the very close coefficients of determination to
the value of one confirmed the close relationship between the
object depth and its histogram features. Therefore, these equations
can be used for feature calibration as the pre-processing step for an
on-the-go HLB diagnosis system when the depth information is
available.
Fig. 7 shows the normalized histograms of HLB-negative, HLBpositive, and zinc-deficient symptomatic areas at three different
distances: 60 cm, 80 cm, and 100 cm. The histograms of three classes were distinctive at all distances with a few overlaps; however,
the range of the gray values at the distance of 100 cm was shorter
than the other two distances which increased the amount of overlap between neighboring curves. At the distance of 60 cm, 92% of
pixels in the zinc-deficient class were saturated (gray value P
255) which overlapped with 3% of pixels in the HLB-positive. At
the distance of 80 cm, 45% of zinc deficient pixels were also
saturated; however, they did not overlap with the HLB-positive
pixels. The maximum range of the gray values (fmaxðijhðiÞ–0Þ
minðijhðiÞ–0Þg) was obtained at the distance of 80 cm as well.
Therefore, the distance of 80 cm was chosen as the optimum distance for HLB identification in both field and lab experiments.
3.2.2. The positioning effect of the leaf
Fig. 8 shows three samples, one from each class: zinc-deficient,
HLB-negative, and HLB-positive, in four leaf-positioning conditions: individual, adjacent, overlapped, and leaves on the artificial
tree. The corresponding RGB images at each leaf positioning
225
A. Pourreza et al. / Computers and Electronics in Agriculture 110 (2015) 221–232
HLB-negative
All datasets
HLB-positive
1. HLB-positive
2. Zn-deficient
Zn-deficient
Zn-deficient
and
HLB-positive
Zn-deficient
and
HLB-negative
Fig. 3. The step-by-step classification model. In each step, the input samples were divided into two parts and at the final step all the dataset were divided into four classes.
Table 1
The qrt-PCR test results for citrus leaf samples in the lab experiment.
Zinc-deficient samples
HLB-negative samples
HLB-positive samples
ID
CT value
HLB status
ID
CT value
ID
CT value
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
40.0
24.9
37.7
40.0
40.0
23.3
40.0
23.6
22.6
40.0
27.8
40.0
24.6
23.2
40.0
40.0
40.0
22.1
24.3
21.4
+
+
+
+
+
+
+
+
+
+
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
40.0
40.0
40.0
40.0
40.0
40.0
40.0
40.0
40.0
40.0
40.0
40.0
40.0
36.5
40.0
40.0
40.0
40.0
40.0
40.0
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
28.0
23.2
25.4
32.3
21.9
21.5
22.3
26.5
24.4
23.1
21.9
30.8
24.8
22.0
22.7
23.2
26.6
21.9
22.8
21.9
Table 2
The qrt-PCR test results for citrus leaf samples in the field experiment.
HLB-positive samples
HLB-negative samples
Non-zinc deficient
Zinc-deficient
ID
CT value
ID
CT value
ID
CT value
1
2
3
4
5
6
7
8
12
13
14
19
21.9
24.1
24.3
21.0
19.5
22.4
22.9
23.0
24.3
26.1
22.6
21.4
9
10
11
15
16
17
18
20
20.7
26.0
24.2
23.2
26.5
23.3
22.4
21.5
21
22
23
24
25
26
27
28
29
30
40.0
40.0
36.0
37.2
40.0
36.2
40.0
40.0
40.0
40.0
condition were created using the color transfer function in which
the green, red, and blue colors indicate the HLB-negative, HLBpositive, and zinc-deficient areas, respectively. The color transfer
function was able to detect the symptomatic areas in all leaf positioning conditions.
3.2.3. Histogram features and classification results
The features of the mean and SD of the gray values which were
extracted from the normalized histograms of the images of the citrus leaves in the dataset for the in-lab experiment are shown in
Table 3. Both features for healthy samples in both leaf positioning
conditions (individual leaves and leaves on the artificial tree) were
generally smaller than HLB-positive and zinc-deficient samples.
Also, these features were normally greater in zinc-deficient samples compared to HLB-positive samples. A comparison between
the features of the same leaves in the two different leaf positioning
conditions indicated that the sample images acquired on the artificial tree had smaller gray value means for 83% of samples (50
out of 60 samples) and also smaller gray value SDs for 78% of samples (47 out of 60 samples).
Fig. 9 illustrates a scatter plot of samples in four classes based
on the means and SDs of the gray values for the images of the individual leaves. This plot shows a clear distinction between the HLBnegative class compared to the HLB-positive and zinc-deficient
classes. Three linear thresholds were acquired from the maximum
margin method and used to divide all samples into four classes of
HLB-negative, HLB-positive, zinc-deficient HLB-negative, and zincdeficient HLB-positive. The sample number 23 had the maximum
SD value equal to 8.0 in the HLB-negative class (Table 3), while
the minimum SD value of the classes was 11.0, which belonged
to the sample number 55 in the HLB-positive class. A linear threshold (r1 ¼ 0:28l þ 22) separated the HLB negative samples from
the rest of the dataset. The zinc-deficient samples generally had
larger mean and SD values compared to the HLB-positive samples.
A second linear threshold (r2 ¼ 0:81l þ 106:43) separated the
HLB-positive samples from the zinc-deficient leaves with one misidentified HLB positive sample (#58). The last threshold was set
within the zinc-deficient class (r3 ¼ 0:37l þ 15:43) to identify
the HLB-positive samples in this class. Using the optimum threshold, only one zinc-deficient HLB-positive leaf sample (#11) was
misidentified in the zinc-deficient HLB-negative class.
Fig. 10 shows a scatter plot of the samples in four classes based
on the means and SDs of the gray value for the leaf images on the
artificial tree. The maximum SD value in the HLB negative class
226
A. Pourreza et al. / Computers and Electronics in Agriculture 110 (2015) 221–232
Fig. 4. The leaf sample that was used for the object depth effect evaluation. (a) Normalized histograms of the leaf sample images acquired from several distances between
50 cm and 150 cm; (b) color image of the sample and its gray images acquired from several distances between 50 cm and 150 cm.
10
Mean of the gray values (µ)
30
25
µ = 45217d -1.861
R² = 0.998
20
15
10
5
Mean
Power (Mean)
0
40
50
60
70
80
90 100 110 120 130 140 150 160
Object depth (d )
Fig. 5. The power regression line fit curve, equation, and coefficient of determination (R2) for gray value means (l) at different object depths (d).
(sample #21) was equal to 8.1, and it was smaller than the
minimum SD value in the other classes (sample #55) which was
equal to 8.2. However, sample #21 was misidentified as belonging
to the HLB-positive class using the first linear threshold
(r1 ¼ 0:3l þ 17:74) because it was considered to be a trade-off
by the maximum margin method to create a more general threshold. The second threshold (r2 ¼ 0:1l þ 35:89) was set between
the HLB-positive and zinc-deficient samples which resulted in
one misidentified HLB-positive leaf sample (#58). The HLB-positive
and -negative samples within the zinc-deficient class were closer
Standard deviation of gray values (σ )
35
8
σ = 10015d -1.782
6
R² = 0.991
4
2
Standard Deviation
Power (Standard Deviation)
0
40
60
80
100
120
140
160
Object depth (d )
Fig. 6. The power regression line fit curve, equation, and coefficient of determination (R2) for gray value SDs (r) in different object depths (d).
to each other in the scatter plot of the leaves on the artificial tree
compared to the individual leaves. The best possible threshold
(r3 ¼ 0:4l þ 11:78) was able to separate these two subclasses
with two misidentified zinc-deficient HLB positive samples (#11
and #13).
As Fig. 9 suggests, except for one HLB-positive sample and one
zinc-deficient HLB-positive sample, all other samples were classified correctly in this dataset. Table 4 shows the classification
results confusion matrix for the data sets of individual leaves. On
average, 0.33 HLB-positive samples and 0.33 zinc-deficient
HLB-positive samples were misclassified in a three-fold cross-
227
A. Pourreza et al. / Computers and Electronics in Agriculture 110 (2015) 221–232
0.1
Distance: 60 cm
0.92
0.1
0.1
0.08
0.06
0.04
0.02
0.06
0.04
0.02
0
50
100
150
200
250
0.06
0.04
0.02
0
0
Distance: 100 cm
0.08
Probability
Probability
Probability
0.08
0.45
Distance: 80 cm
0
0
50
Gray value
100
150
200
250
0
HLB -
HLB +
50
100
150
200
250
Gray Value
Gray value
Zn-deficient
Fig. 7. Comparison of the histogram curves of symptomatic areas in three classes (HLB-negative, HLB-positive, and zinc-deficient) and at three different distances (60 cm,
80 cm, and 100 cm).
Fig. 8. Symptomatic areas detection results for three leaf samples (one sample from each class) in four different leaf-positioning conditions. The green, red, and blue colors
indicate the HLB-negative, HLB-positive, and zinc-deficient areas, respectively. (For interpretation of the references to colour in this figure legend, the reader is referred to the
web version of this article.)
validation. An overall accuracy of 97% was achieved using the
proposed classification model and the mean and SD features.
However, the purpose in this research was to detect HLB-infection.
As long as the average of 0.33 HLB-positive samples was misclassified in another HLB-positive class (zinc-deficient), their HLB statuses were identified correctly. The dotted lines in Table 4 (and
also Table 5) separated the HLB-positive classes from HLB-negative
classes. Therefore, the overall accuracy of 98.5% was obtained
when only HLB-infection was considered in the classification
results evaluation.
Based on Fig. 10, one HLB-positive sample was misidentified in
the zinc-deficient HLB-positive cluster and two zinc-deficient
HLB-positive leaves were also misidentified as zinc-deficient
HLB-negative samples. Table 5 also illustrates the confusion matrix
of the classification results for the leaves of the artificial tree dataset. Analogous to the scatter plot in Fig. 10, the average of 0.33
HLB-positive samples and 0.67 zinc-deficient HLB-positive samples
were misclassified in a three-fold cross validation. The overall
accuracy of 95.5% was achieved in the four-class classification.
Also, an overall HLB detection accuracy of 97% was achieved for
leaves on the artificial tree dataset.
3.3. Field experiment results
Fig. 11 shows three samples images (one sample image per
class) acquired in the field experiment. The distance between the
vision sensor and the target leaf in each image (specified with a
red boundary) was exactly 80 cm.
Table 6 shows the mean and SD gray value features which were
extracted from the normalized histograms of the leaf images in the
field experiment. Similar to the lab experimental results, the average of the feature values in the HLB-negative class was smaller
than that of the HLB-positive class. Also, these features were
mostly smaller for non-zinc deficient samples within the HLB-positive class.
A scatter plot of samples based on the mean and SD of the gray
values’ for the images from the field experiment is shown in
Fig. 12. Sample #27 had the maximum SD (9.5) in the HLB-negative
class which was smaller than the minimum SD in the HLB-positive
class (10.5) which belonged to sample #7. However, the first linear
threshold (r1 ¼ 0:24l þ 23:32) which was obtained by the maximum margin method was set above the sample #7 (HLB-positive).
In other words, one HLB-positive sample was misclassified as HLB-
228
A. Pourreza et al. / Computers and Electronics in Agriculture 110 (2015) 221–232
Table 3
The mean and SD gray value features extracted from the normalized histogram of images of the leaf samples from the in-lab dataset.
Zinc-deficient samples
ID
HLB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
b
c
d
b
l
+
+
+
+
+
+
+
+
+
+
Average
a
HLB-negative samples
a
ID
r
Ind.c
FTd
Ind.
FT
84.7
100.4
110.3
89.4
107.9
114.8
121.4
121.0
213.8
127.5
102.9
119.4
140.9
134.5
98.0
128.6
89.3
98.7
162.1
162.3
55.4
71.5
79.3
114.5
104.7
119.6
112.2
86.9
212.9
122.0
92.0
111.1
93.4
111.3
77.3
108.6
65.9
85.6
159.1
149.7
65.8
42.6
65.7
58.6
58.6
37.1
85.4
44.1
59.0
77.5
62.4
82.2
65.7
64.3
55.3
69.2
63.7
44.5
64.5
49.7
35.2
31.1
52.7
66.2
51.9
49.1
78.8
38.2
50.3
71.8
57.5
71.6
53.7
53.3
41.1
61.4
40.3
36.8
65.5
45.1
121.4
106.6
60.8
52.6
HLB-positive samples
l
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
ID
r
Ind.
FT
Ind.
FT
38.3
28.4
49.5
26.5
20.9
26.8
28.0
26.3
23.1
23.7
35.8
34.3
30.2
29.3
33.7
26.2
31.9
36.6
41.8
34.3
38.1
16.8
32.0
26.1
21.5
29.2
24.1
19.4
24.6
28.2
31.1
23.9
17.5
25.9
24.9
22.7
21.9
30.8
36.7
28.6
5.7
4.2
8.0
3.7
4.7
4.9
3.2
4.8
3.8
3.8
4.0
5.1
4.8
6.0
4.7
5.6
5.3
5.1
4.5
6.6
8.1
3.8
4.9
5.5
4.2
4.5
3.0
3.2
4.4
5.5
4.4
3.5
2.6
4.5
3.9
3.5
4.0
4.8
6.0
4.3
31.3
26.2
4.9
4.4
l
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
r
Ind.
FT
Ind.
FT
48.5
51.5
78.7
59.7
63.8
49.3
46.2
53.4
52.2
51.8
81.2
47.8
56.1
76.0
57.5
64.1
69.0
103.5
47.8
52.9
37.9
33.6
48.3
51.3
57.3
56.8
41.2
36.9
54.2
54.4
75.9
36.9
41.3
62.7
45.7
60.3
45.4
89.6
52.9
43.2
14.8
14.9
38.9
18.8
26.9
13.6
12.7
13.5
23.7
15.6
23.0
21.1
18.7
17.8
11.0
17.3
15.5
39.9
12.1
18.4
13.6
10.0
20.5
17.1
24.6
14.2
9.7
9.6
22.8
16.0
23.6
14.4
14.1
13.9
8.2
15.0
10.2
34.4
13.4
12.0
60.6
51.3
19.4
15.9
l: The mean of the gray values.
r: The SD of the gray values.
Ind.: Images of individual leaves.
FT: Images of leaves on the artificial tree.
90
Standard deviation of gray values (σ )
Standard deviation of gray values (σ )
80
80
70
60
50
40
30
20
10
0
0
50
100
150
200
250
Mean of gray values (µ)
HLB -
HLB +
Zn Def. HLB -
Threshold
Zn Def. HLB +
Support vectors
70
60
50
40
30
20
10
0
0
50
100
150
200
250
Mean of gray values (µ)
HLB Zn Def. HLB Zn Def. HLB +
HLB +
Threshold
Support Vectors
Fig. 9. Scatter plot of samples in four classes based on the means and SDs of the
gray value features of the individual leaf images.
Fig. 10. Scatter plot of samples in four classes based on the mean and SD gray value
features of the normalized histograms of the leaves on the artificial tree images.
negative in order to have a more general threshold. There were two
subclasses of non-zinc-deficient and zinc-deficient samples within
the HLB-positive class. Sample #6 in the non-zinc-deficient
subclass had the maximum mean gray value of 96.1, while the
minimum mean gray values in the zinc-deficient subclass was
equal to 105.4 (sample #11). The second threshold (r2 ¼
2:19l þ 263:5) was set between the two subclasses to separate
the zinc-deficient samples from the non-zinc-deficient samples
within the HLB-positive superclass. Using these two simple linear
thresholds, all the samples were clustered in three classes of
HLB-negative, HLB-positive and HLB-positive zinc-deficient with
only one misidentified sample.
Table 7 includes the classification accuracies and one misclassification error for the field dataset. The overall three-class classification accuracy of 97% was achieved using the mean and SD features
and the SVM classifier.
229
A. Pourreza et al. / Computers and Electronics in Agriculture 110 (2015) 221–232
Table 4
Average number of samples in the individual leaves dataset which were classified into the four classes and their corresponding classification accuracies and misclassification
errors (%). The last row and column illustrate the sum of samples in the corresponding rows or columns.
Actual class
Sum
HLB+
Prediction
Zn Def. HLB+
HLB+
Zn Def. HLB+
Zn Def. HLB
HLB
6.67 (95.29%)
0.33 (4.71%)
Sum
7
Zn Def. HLB
3.67 (91.75%)
0.33 (8.25%)
4 (100%)
4
4
HLB
7 (100%)
6.67
4
4.33
7
7
22
Table 5
Average number of samples in the leaves on the artificial tree dataset which were classified into the four classes and their corresponding classification accuracies and
misclassification errors (%). The last row and column illustrate the sum of samples in the corresponding rows or columns.
Actual class
Sum
HLB+
Prediction
Zn Def. HLB+
HLB+
Zn Def. HLB+
Zn Def. HLB
HLB
6.67 (92.29%)
0.33 (4.71%)
Sum
7
Zn Def. HLB
3.33 (83.25%)
0.67 (16.75%)
4 (100%)
4
4
HLB
7 (100%)
6.67
3.66
4.67
7
7
22
Fig. 11. One sample image from each class of the in-field dataset. The target leaf in each image is indicated with a red boundary. (For interpretation of the references to colour
in this figure legend, the reader is referred to the web version of this article.)
Table 6
The mean and SD gray value features extracted from the normalized histogram of leaf
sample images from the field dataset.
HLB-negative
HLB-positive
Non-zinc-deficient
Zinc-deficient
ID
l
r
ID
l
r
ID
l
r
21
22
23
24
25
26
27
28
29
30
37.6
29.7
37.0
33.4
32.8
29.1
44.2
29.3
29.8
26.4
5.0
5.5
7.0
6.9
5.9
6.4
9.5
6.3
6.8
3.9
1
2
3
4
5
6
7
8
12
13
14
19
73.7
39.0
89.3
45.1
57.4
96.1
40.5
86.5
94.2
69.1
73.5
78.0
22.3
22.6
38.9
22.1
16.5
24.5
10.5
37.6
28.1
27.5
18.9
50.3
9
10
11
15
16
17
18
20
119.8
126.7
105.4
159.2
122.1
211.6
126.6
127.1
38.9
27.4
38.5
69.3
37.4
58.1
72.0
40.2
Average
32.9
6.3
Average
70.2
26.7
Average
137.3
47.7
4. Discussion
The purpose of this study was to optimize the HLB detection
performance of a previously introduced method (Pourreza et al.,
2014) by developing a new vision sensor which could increase
identification accuracy and decrease algorithm complexity and
analysis time.
Our determination showed that the object depth had an
extreme effect on the image histogram. However, there was a close
relationship between the histogram features (mean and SD) and
the object depth. Since these two histogram features were used
for the classification purpose in this study, they could be easily calibrated and computed using the proposed regression equations
when the object depth information is available. Depth cameras,
such as an RGB-D (red, green, blue, and depth) camera, can measure the depth of each individual object in the image.
Khoshelham and Elberink (2012) evaluated the resolution and
accuracy of the depth information of a Kinect camera (Microsoft,
Redmond, Washington). They determined that the Kinect depth
resolution varied from 2 mm (at a distance of 1 m) to 25 mm (at
a distance of 3 m). Since the maximum imaging distance in our
application never goes beyond 3 m, a Kinect sensor can be used
in our system to acquire the depth information with an acceptable
resolution. Therefore, the mean and SD of the gray values of any
leaf in a citrus tree image can be computed and used for HLB detection. This method will be used in the future in an on-the-go HLB
diagnostic system. A comparison of three different distances
between the sensor and leaf sample showed that the symptomatic
areas in the three classes of HLB-negative, HLB-positive, and
zinc-deficient samples were clearly distinguishable in all three
230
A. Pourreza et al. / Computers and Electronics in Agriculture 110 (2015) 221–232
Standard deviation of gray value (σ )
70
60
50
40
30
20
10
0
0
50
100
150
200
250
Mean of gray value (µ)
HLB -
HLB +
Zn Def. HLB +
Threshold
Support Vectors
Fig. 12. Scatter plot of samples in the three classes based on their gray value mean
and SD features of the normalized histograms of the target leaves from the field
dataset.
Table 7
Average number of samples in the field dataset which were classified into each of the
three classes and their corresponding classification accuracies and misclassification
errors (%). The last row and column illustrate the sum of samples in the corresponding
rows or columns.
Actual class
HLB+
Prediction
HLB+
Zn Def. HLB+
HLB
Sum
Zn Def. HLB+
HLB
Sum
4 (100%)
4
3.67
3
4.33
11
3.67 (92%)
3 (100%)
0.33 (8%)
4
3
distances; however, the maximum separation was achieved at a
distance of 80 cm. The maximum range of the gray values (237)
was also obtained at the distance of 80 cm, while it was found to
be 219 and 229 at the distances of 60 cm and 100 cm, respectively.
Additionally, the number of pixels in each pair of neighboring classes (HLB-negative and HLB-positive or HLB-positive and zincdeficient), which had the same gray value decreased at a distance
of 80 cm.
The results of identifying symptomatic areas using the color
transform function confirmed that the positioning condition of
the leaves did not have a significant effect on the identification
accuracy. The results were slightly different for the leaves on the
artificial tree condition (Fig. 8) because the distance between the
leaves and the sensor was not exactly equal to 80 cm, and also
the surfaces of the leaves were not precisely perpendicular to the
line-of-sight of the camera. As a result, different locations on each
leaf had different depths; still the symptomatic areas were distinctive for the leaves on the artificial tree condition as well. Because of
the same reason, the mean and SD gray values of the leaves mostly
decreased when their images were acquired on the artificial tree
compared to the individual leaf positioning condition in which
the distance between the leaves and the sensor was exactly equal
to 80 cm (Table 3).
The mean and SD gray values of a histogram represent the overall intensity and root mean square (RMS) contrast in an image (Peli,
1990). Based on the results of this study, the mean gray values
were smaller for the HLB-negative citrus leaves compared to the
HLB-positive samples (Tables 3 and 6). Since the HLB-positive citrus leaves had some starch accumulation and starch has the capability to rotate the polarization planar of polarized light, the
designed sensor was able to highlight the HLB symptomatic areas
on the HLB-infected leaves. These highlighted HLB-positive areas
had brighter pixels which caused larger mean gray values for the
HLB-positive leaves. However, the portion of HLB symptomatic
area varied in each infected leaf and this variation influenced the
mean gray values. Therefore, although the mean gray value was a
satisfying feature for identifying HLB infection, the SD of the gray
value was also used for this purpose to increase the accuracy. Still,
the mean gray value alone was able to differentiate the zincdeficient leaves from non-zinc-deficient samples within the HLBpositive class for the field experiment (Fig. 12). The SD of the gray
value or RMS contrast of an image indicates the dispersion of gray
values from the mean. The HLB-negative (non-zinc-deficient) samples did not have any high intensity areas, so their pixel values
were mostly closer to the mean and consequently, they had
smaller SD. On the contrary, the HLB-negative and zinc-deficient
samples had both symptomatic (zinc or/and HLB) and non-symptomatic areas, and as a result, they had wider histogram curves
and larger SD values. Accordingly, as the scatter plots in Figs. 9,
10 and 12 suggest, simple thresholds only in the SD could effectively separate HLB-negative samples from HLB-positive and
zinc-deficient samples with zero error in all datasets.
Zinc deficiency develops extensive chlorosis between the veins
which causes whitish yellow color in a symmetric pattern on the
zinc-deficient citrus leaf. This symptom was originally brighter
than the starch accumulation symptom in the HLB-positive leaves
(Figs. 7 and 8) which usually caused larger mean gray values for
zinc-deficient samples. Additionally, the SD values for non-zincdeficient HLB-positive samples were usually smaller than zincdeficient samples. A single threshold in SD for individual leaves
or leaves on the artificial tree datasets (Figs. 9 and 10) could separate the HLB-positive leaves from the zinc-deficient samples with
one misidentification; however, both of features were used with
the maximum margin method to achieve a more general threshold.
The HLB-positive samples within the zinc-deficient class (lab
experiment) had the averages of mean gray values equal to 135.1
and 118.2 for individual leaves and leaves on the artificial tree,
respectively, while these averages were equal to 107.7 and 95.1
for the zinc-deficient HLB-negative samples (Table 3). However,
the best possible threshold in the mean gray value (e.g.
l = 82.95) would result in average clustering error rate of 35% for
individual leaves and it would be worse for leaves on the artificial
tree. The averages of the SD of the gray values for HLB-negative
samples within the zinc-deficient class were also equal to 68.2
and 57.1 for individual leaves and leaves on the artificial tree, correspondingly, while these averages were equal to 53.4 and 48.1 for
the HLB-positive samples. However the maximum clustering accuracy rate using the optimum threshold in the SD would not be over
70% for individual leaves. The coefficient of variation (rl) is a measure of relative variability which shows the dispersion of the gray
values in relation to the mean (Kannan, 1981). The averages of the
coefficients of variation values for the zinc-deficient HLB-positive
samples were equal to 0.41 and 0.43 in individual leaves and leaves
on the artificial tree datasets correspondingly, while they were
equal to 0.64 and 0.60 for the zinc deficient HLB negative samples.
The zinc-deficient HLB-positive samples included both HLB and
zinc symptomatic areas, so they generally had smaller coefficients
of variation which illustrated less gray value dispersion. The pixel
values of the HLB symptomatic areas in an HLB-positive zincdeficient sample filled the gap between the pixel values of the
zinc-deficient symptomatic areas and healthy areas in the
A. Pourreza et al. / Computers and Electronics in Agriculture 110 (2015) 221–232
histogram of the leaf, and this was the reason for smaller coefficients of variation in this subclass. Therefore, it is necessary to
use both the mean and SD features in HLB-positive samples identification within the zinc-deficient class. The slopes of the threshold lines between the HLB-positive and HLB-negative samples
within the zinc-deficient class were analogous (Figs. 9 and 10) in
both leaves positioning conditions datasets. Sample #11 was misidentified in both lab datasets, and sample #13 was also misidentified in the leaves on the artificial tree dataset. The coefficients of
variation for both of these two samples were more similar to the
HLB-negative zinc-deficient samples, so their misidentifications
were unavoidable.
The zinc-deficient HLB-positive samples in all datasets had
higher mean gray values in all three datasets. Since there was no
zinc-deficient HLB-negative sample in the field dataset, a simple
threshold within a comparatively wide margin (Dl = 9.3) could
also cluster these two classes with zero error.
Between all HLB detection methods, airborne image analysis
can perform the fastest diagnosis in a large area. Li et al. (2012a)
obtained the best accuracy (86.3%) using airborne hyperspectral
imagery; however, their method was less accurate, more expensive, and more complicated comparing to the method presented
in this study. Sankaran and Ehsani (2012) also reported the best
overall accuracy of higher than 94% for field HLB detection; while
the vision sensor in this study was able to identify HLB infection
with less than 3% error. Additionally, compared with our previous
study (Pourreza et al., 2014), the HLB identification accuracy
within the zinc-deficient samples increased significantly. Using
only two simple statistical image descriptors in a step-by-step
classification model required a computationally inexpensive analysis algorithm which is an advantage in design and development
of the commercial diagnosis product.
5. Conclusions
In this study, an HLB detection method was introduced which
showed improved performance in different aspects compare to
the previous studies. No sample preparation such as leaf collection
or grinding was required in this method and the vision sensor
could detect the infection without being in contact with leaves.
Only two simple features were extracted from the leaf images
and used for classification purposes. This simplification decreased
the analysis expense and time, and facilitated the detection process. HLB-negative samples were classified with zero error in all
three datasets. Not only was the zinc deficiency accurately
detected, but also the HLB infection within the zinc-deficient
leaves was identified with increased accuracies. The two major
components of the vision sensor were 10 high power LEDs and
an inexpensive camera, and the whole sensor was assembled with
less than one-thousand dollars which made it an affordable diagnosis device even for small citrus growers. Two close relationships
were found between the leaf depth in the image and the gray values’ mean and SD features which were employed in leaf HLB status
determination in this study. Using these two equations and a depth
measurement sensor will enable this system to be used for on-thego HLB diagnosis in future studies. Compared to our previous
study, the HLB detection accuracy increased significantly in both
zinc-deficient and non-zinc-deficient classes.
Acknowledgements
The authors would like to thank the Citrus Research and
Development Foundation (CRDF) for supporting this research. We
would also like to express our appreciation to Dr. Hamidreza
Pourreza (Ferdowsi University of Mashhad, Mashhad, Iran), and
231
Mr. Michael Irey (United States Sugar Corporation, Clewiston, FL)
for their assistance in this study.
References
Albrecht, U., Bowman, K.D., 2008. Gene expression in citrus sinensis (L.) osbeck
following infection with the bacterial pathogen Candidatus Liberibacter asiaticus
causing Huanglongbing in Florida. Plant Sci. 175, 291–306.
Belasque, J.J., Gasparoto, M., Marcassa, L., 2008. Detection of mechanical and disease
stresses in citrus plants by fluorescence spectroscopy. Appl. Opt. 47, 1922–
1926.
Bishop, C.M., 2006. Pattern Recognition and Machine Learning, first ed. Springer
Science, New York.
Buitendag, C., Von Broembsen, L., 1993. Living with citrus greening in South Africa.
In: Proc. 12th Conference of the International Organization of Citrus Virologists.
University of California, Riverside, pp. 269–273.
Choi, D., Lee, W.S., Ehsani, R., 2013. Detecting and counting citrus fruit on the
ground using machine vision. In: ASABE Annual International Meeting. ASABE,
Kansas City, Missouri.
Chung, K.-R., Brlansky, R., 2009. Citrus diseases exotic to Florida: Huanglongbing
(citrus greening). University of Florida IFAS, Florida Cooperative Extension
Service, Gainesville, Florida.
Etxeberria, E., Gonzalez, P., Achor, D., Albrigo, G., 2009. Anatomical distribution of
abnormally high levels of starch in HLB-affected Valencia orange trees. Physiol.
Mol. Plant Pathol. 74, 76–83.
Futch, S., Weingarten, S., Irey, M., 2009. Determining HLB infection levels using
multiple survey methods in Florida citrus. Proc. Fla. State Hort. Soc., 152–158.
Garcia-Ruiz, F., Sankaran, S., Maja, J.M., Lee, W.S., Rasmussen, J., Ehsani, R., 2013.
Comparison of two aerial imaging platforms for identification of
Huanglongbing-infected citrus trees. Comput. Electron. Agric. 91, 106–115.
Gennadios, A., Ghorpade, V., Weller, C.L., Hanna, M., 1996. Heat curing of soy protein
films. Biological Systems Engineering. Papers and Publications, p. 94.
Gonzalez, P., Reyes-De-Corcuera, J., Etxeberria, E., 2012. Characterization of leaf
starch from HLB-affected and unaffected-girdled citrus trees. Physiol. Mol. Plant
Pathol. 79, 71–78.
Halbert, S.E., 2005. The discovery of Huanglongbing in Florida. 2nd International
Citrus Canker and Huanglongbing Research Workshop, Orlando, Florida, p. H-3.
Hansen, A., Trumble, J., Stouthamer, R., Paine, T., 2008. A new Huanglongbing
species, ‘‘Candidatus Liberibacter psyllaurous’’, found to infect tomato and
potato, is vectored by the psyllid Bactericera cockerelli (Sulc). Appl. Environ.
Microbiol. 74, 5862–5865.
Hawkins, S.A., Park, B., Poole, G.H., Gottwald, T., Windham, W.R., Lawrence, K.C.,
2010. Detection of citrus Huanglongbing by Fourier transform infrared–
attenuated total reflection spectroscopy. Appl. Spectrosc. 64, 100–103.
Kannan, K., 1981. Percentage Coefficient of Variation. CMFRI Special Publication, 149.
Khoshelham, K., Elberink, S.O., 2012. Accuracy and resolution of kinect depth data
for indoor mapping applications. Sensors 12, 1437–1454.
Kim, D.G., Burks, T.F., Schumann, A.W., Zekri, M., Zhao, X., Jianwei, Q., 2009.
Detection of citrus greening using microscopic imaging. Agricultural
Engineering International. The CIGR Ejournal.
Kumar, A., Lee, W.S., Ehsani, R.J., Albrigo, L.G., Yang, C.H., Mangan, R.L., 2012. Citrus
greening disease detection using aerial hyperspectral and multispectral
imaging techniques. J. Appl. Remote Sens., 6.
Li, H., Lee, W.S., Wang, R., Ehsani, R., Yang, C., 2012a. Spectral angle mapper (SAM)
based citrus greening disease detection using airborne hyperspectral imaging.
11th International Conference on Precision Agriculture, Indianapolis, Indiana.
Li, W., Hartung, J.S., Levy, L., 2006. Quantitative real-time PCR for detection and
identification of Candidatus Liberibacter species associated with citrus
Huanglongbing. J. Microbiol. Methods 66, 104–115.
Li, X., Lee, W.S., Li, M., Ehsani, R., Mishra, A.R., Yang, C., Mangan, R.L., 2011.
Comparison of different detection methods for citrus greening disease based on
airborne multispectral and hyperspectral imagery. In: ASABE Annual
International Meeting ASABE, Louisville, Kentucky.
Li, X., Lee, W.S., Li, M., Ehsani, R., Mishra, A.R., Yang, C., Mangan, R.L., 2012. Spectral
difference analysis and airborne imaging classification for citrus greening
infected trees. Comput. Electron. Agric. 83, 32–46.
Lins, E.C., Belasque Jr., J., Marcassa, L.G., 2009. Detection of citrus canker in citrus
plants using laser induced fluorescence spectroscopy. Precision Agric. 10, 319–
330.
Marcassa, L., Gasparoto, M., Belasque Jr., J., Lins, E., Nunes, F.D., Bagnato, V., 2006.
Fluorescence spectroscopy applied to orange trees. Laser Phys. 16, 884–888.
Mishra, A., Ehsani, R., Albrigo, G., Lee, W.S., 2007. Spectral characteristics of citrus
greening (Huanglongbing). ASABE Annual International Meeting. ASABE,
Minneapolis, Minnesota.
Mishra, A., Karimi, D., Ehsani, R., Albrigo, L.G., 2011. Evaluation of an active optical
sensor for detection of Huanglongbing (HLB) disease. Biosyst. Eng. 110, 302–
309.
Peli, E., 1990. Contrast in complex images. JOSA A 7, 2032–2040.
Pereira, F.M.V., Milori, D.M.B.P., Pereira-Filho, E.R., Veníncio, A.L., Russo, M.d.S.T.,
Cardinali, M.C.d.B., Martins, P.K., Freitas-Astúa, J., 2011. Laser-induced
fluorescence imaging method to monitor citrus greening disease. Comput.
Electron. Agric. 79, 90–93.
Pourreza, A., Lee, W.S., Raveh, E., Ehsani, R., Etxeberria, E., 2014. Citrus
Huanglongbing detection using narrow-band imaging and polarized
illumination. Trans. ASABE 57, 259–272.
232
A. Pourreza et al. / Computers and Electronics in Agriculture 110 (2015) 221–232
Pourreza, A., Lee, W.S., Raveh, E., Hong, Y., Kim, H.-J., 2013. Identification of citrus
greening disease using a visible band image analysis. ASABE Annual
International Meeting. ASABE, Kansas City, Missouri.
Pourreza, A., Pourreza, H., Abbaspour-Fard, M.-H., Sadrnia, H., 2012. Identification of
nine Iranian wheat seed varieties by textural analysis with image processing.
Comput. Electron. Agric. 83, 102–108.
Salois, M.J., Jauregui, C., Ferrell, L., Norberg, R.P., Barnhardt, V., Griffith, T., 2012.
Citrus reference book. Florida Department of Citrus, Gainesville, Florida.
Sankaran, S., Ehsani, R., 2012. Detection of Huanglongbing disease in citrus using
fluorescence spectroscopy. Trans. ASABE 55, 313–320.
Sankaran, S., Ehsani, R., Etxeberria, E., 2010. Mid-infrared spectroscopy for detection
of Huanglongbing (greening) in citrus leaves. Talanta 83, 574–581.
Texeira, D.d.C., Ayres, J., Kitajima, E., Danet, L., Jagoueix-Eveillard, S., Saillard, C.,
Bové, J., 2005. First report of a Huanglongbing-like disease of citrus in Sao Paulo
State, Brazil and association of a new Liberibacter species, ‘‘Candidatus
Liberibacter Americanus’’, with the disease. Plant Dis. 89, 107.
Windham, W.R., Poole, G.H., Park, B., Heitschmidt, G., Hawkins, S.A., Albano, J.P.,
Gottwald, T.R., Lawrence, K.C., 2011. Rapid screening of Huanglongbing-infected
citrus leaves by near-infrared reflectance spectroscopy. Trans. ASABE 54, 2253–
2258.