This full-text paper was peer-reviewed and accepted to be presented at the IEEE WiSPNET 2017 conference.

Leaf Disease Detection on Cucumber Leaves Using

Multiclass Support Vector Machine
P. Krithika1 and S. Veni2
Department of Electronics and Communication Engineering, Amrita School of Engineering, Coimbatore
Amrita Vishwa Vidyapeetham, Amrita University, India-641 112
Email: 1 krithikaguru93@gmail.com 2 s_veni@cb.amrita.edu

Abstract—In India, smart organic farming is gaining impor- in the production of cucumbers are Karnataka, Tamil Nadu,
tance. There may be problems due to environment, temperature, and Andhra Pradesh. Karnataka grows 60 percent of India’s
humidity or nutrient deficiency in this farming. If we have a cucumbers, with the other two states accounting for 20 percent.
monitoring system for this farming it is possible to produce
healthy plant. The aim is to address this issue using computer The demand for cucumber in overseas is greater than the
aided image processing technique. Main solution is to create an demand here domestically. The health benefits of Cucumber
automation system which can detect the disease present in the leaf are as follows:
of the plant. In this paper, a first level attempt is made to detect • Vegetable with low calorie.
diseases present in the leaf of salad cucumber. The most common
• Good source of dietary fiber.
diseases which are present in salad cucumber are Alternaria
leaf blight, Bacterial wilt, Cucumber green mottle mosaic, Leaf • Very good source of potassium.
Miner, Leaf spot, Cucumber Mosaic Virus (CMV) disease and so • Contains unique anti-oxidants.
on. K-means clustering, an unsupervised algorithm along with • Have mild diuretic property.
Support Vector Machine(SVM) is used in this work to address • Have a high amount of vitamin K.
this problem.
Index Terms—k-means clustering, Multiclass SVM, Image The rest of the Section is organized as follows. The related
processing. works are discussed in Section II. In Section III, a brief
description of diseases present in cucumber leaf is given. The
I. I NTRODUCTION methodology is explained in Section IV. The discussion on
the results is carried out in Section V. The future scope of
Weather and other environmental conditions cannot be con- this work is discussed in the last Section.
trolled by farmers. Disease mitigation is the prime factor to
be considered in the case of farming practices. Immediate II. R ELATED W ORKS
attention has to be given to crops which gets affected by Researches have been carried out to use digital image
pest or disease. If proper monitoring of leaves are made at processing techniques in the agriculture field. This section will
the beginning stage itself, further spread of the disease in the discuss on the existing methods.
plants can be avoided. On observation by naked eyes it is Chaudhary, Piyush, et al. [3] segmented the region of
difficult to recognize the disease in leaf. This might result in disease from the RGB image by performing color transform.
the wrong application of the pesticides and ultimately it results Here comparisons are done on three color spaces. They are
in the crop failure. YcbCr, CIELAB and HSI color spaces. Median filter is used
Due to several factors in the environment there are several for image smoothening and this resulted in noise free algo-
diseases which are affecting the crops. This leads to reduction rithm. The experiment was conducted in dicot and monocot
in the quality and productivity of the plants. Hence an auto- leaves. Rothe, P. R., and R. V. Kshirsagar [4] extracted colour
matic system which can detect diseases is mandatory since layout descriptors and classified using Neural Networks. This
this system will be useful in monitoring the crops and hence experimentation was done in cotton leaves. Revathi, P., and
immediate actions can be taken. M. Hemalatha [5] used edge detection method for detection
It is found that image processing techniques will give of diseases. The features that were used for the analysis
successful results in disease detection [1]. The machine learn- purpose were color, boundary, shape and texture which were
ing methodologies for classification purpose, and k-means taken from the disease spots. D.S. Guru, P.B. Mallikarjuna
clustering can be used in agricultural researches like disease and S. Manjunath [6] used gray-level co-occurrence matrix
detection. Earlier this kind of disease detection has been (GLCM) for feature extraction purpose and diseases in to-
carried out in cotton [1] and paddy leaves [2]. bacco leaves were determined. Schikora, Marek, and Adam
In the total cucumber production, India occupies 30th po- Schikora [7] proved that the performance of SVM were
sition and it accounts for less than one percent of the world’s better than Neural Network for classification. Jian, Zhang,
supply. Entire year it is seen that cucumber is available on and Zhang Wei [8] considered the entire leaf instead of
a cheaper cost. The warm climate in India is suitable for diseased portion. The tests were conducted using Radial Basis
development of cucumbers. The three main states involved Function (RBF) and it gave better performance using Support

978-1-5090-4442-9/17/$31.00 2017
c IEEE 1276
 This full-text paper was peer-reviewed and accepted to be presented at the IEEE WiSPNET 2017 conference.

Fig. 1. Leaf spot disease. Fig. 2. Leaf miner.

Vector Machine for classification purpose. Priya, C. Arun, T.

Balasaravanan, and Antony Selvadoss Thanamani [9] extracted
twelve features and comparison was performed between k-NN
method and SVM classification. Omrani, Elham, et al. [10]
analysed the diseases in apple. Both SVM and Artificial Neural
Networks(ANN) are used for the classification purpose. It has
been found that classification using SVM gave better results
than ANN. Phadikar, Santanu, and Jaya Sil [11] used pattern
recognition for disease detection in rice plants. The infected
part is segmented using image segmentation from the image
captured using digital camera. The classification is done using
Fig. 3. CMV.
Neural Network in this case. D. S. Guru, P. B. Mallikarjuna
and S. Manjunath used Probabilistic Neural Network(PNN)
for classification purpose.
III. D ISEASES P RESENT IN C UCUMBER L EAF
Alternaria leaf blight, Bacterial wilt, Cucumber green mot-
tle mosaic, Leaf Miner, Leaf spot, CMV, Anthracnose and
Aster yellows are some of the diseases which are present
in cucumber leaf [12]. A brief description of some these
diseases is provided in this section. Real time data was
collected for diseases such as leaf spot, leaf miner and CMV
in cucumber leaf.
A. Leaf Spot Fig. 4. Healthy leaf.

The leaf spot disease in cucumber is because of bacteria

called as Pseudomonas syringae. This bacteria can affect any problems due to leaf miner can be avoided. Fig. 2 shows the
parts of the leaf such as stem, leaf or the fruit. In case of leaf miners which are present in the leaves of cucumber.
leaves the affected area is in angular shape and it appears
C. CMV
as watersoaked area. In the later stage this turns into brown
colour. Ultimately this results in leaf withering. During rain the CMV creates mosaic symptoms in cucumber leaves.
bacteria can spread from one part to other parts of the plant. Because of this virus the young leaves will get affected and
The temperature suitable for the disease is 24◦ –28◦ C [13] becomes narrow which further affects the entire plant. This
and humidity favors this disease. Fig. 1 shows the cucumber will result in pale and bumpy cucumbers. The cucumbers
leaf which is infected with the leaf spot disease. produced from the crops which are affected by CMV will
generally be very bitter in taste. The temperature which is
B. Leaf Miner suitable for this disease is 79◦ –84◦ F. The only way to prevent
Leaf miner is an insect that feeds on leaves. This leaf miner this is by detection and removal. Fig. 3 shows the leaf with
can affect the yield of the crop. Generally it is present in plants CMV in it. Healthy leaf of cucumber is shown in Fig. 4.
with large leaves. This insect feeds in cucumber leaves also.
Considering small amount, this is not a big deal of problem. D. Alternaria Leaf Blight
But when found in large scale, this will weaken the crop and it The first appearance of spots can be seen in oldest leaves.
results in poor harvest. If proper maintenance is ensured then The spots are generally small and will be in yellow-brown

1277
 This full-text paper was peer-reviewed and accepted to be presented at the IEEE WiSPNET 2017 conference.

Fig. 6. Bilinear interpolation [14].

Fig. 5. Flowchart.

colour. As the disease advances the lesions turn into big

patches. The pattern is repetitive. Finally the leaves will curl
and dies off.
E. Anthracnose Fig. 7. Contrast stretching—form of transformation structure [15].
The lesions will be brown in color with yellow edges. This
A. Image Acquisition
can be seen in leaves, stems and fruit. In case of resistant
varieties the lesions will appear tan with green edges. The diseased leaf images are captured using Sony Cyber-
shot DSC-W710. It has resolution of about 16.10 Megapixels
F. Aster Yellows was used to capture the image. The sensor type of the camera
This disease will show a unique feature of upright growth is charge-coupled device (CCD). The data were acquired on
and the stems will be rigid. The leaves are generally small 25th and 26th November 2016. The background was constant
but it will appear to be thick in nature. Flowers get disfigured, black so that leaf gets isolated from complex background. As
and leafy bracts are possessed by these flowers. Eventually the mentioned earlier there are several diseases which can affect
fruits will turn small and pale. cucumber. But the major diseases which cause huge reduction
IV. M ETHODOLOGY in the yield are leaf spot disease, Cucumber Mosaic Virus
and leaf miner. Hence leaf samples of these three diseases
The very first step is to acquire the disease affected leaf. along with healthy leaves were collected from Tamil Nadu
Required preprocessing is to be performed on the image ac- Agricultural University, Coimbatore.
quired. After preprocessing, the further steps are segmentation,
feature extraction and then classification. The step by step B. Image Preprocessing
approach for this is explained below. 1) Image Resizing: In case of practical applications the
1) Image acquisition. acquired image can be of different size. It is recommended
2) Preprocessing of the image. to resize the acquired images. The size of the resized image
• Image Resizing becomes 256 × 256 × 3. The interpolation method used for
• Contrast Enhancement resizing is bilinear interpolation beacuse in this method four
3) Color space conversion. nearest pixel values are averaged to find the new pixel value.
4) K-means clustering. Q1 + Q2 + Q3 + Q4
5) Feature Extraction P = (1)
4
• Using GLCM
where P is the new pixel value. Q1, Q2, Q3 and Q4 are four
6) Classification. neighboring pixels. This is depicted in Fig. 6.
• Muticlass SVM—1AA approach. 2) Contrast Enhancement: In order to distinguish between
The idea can be implemented as a mobile application, where object and background of an image contrast enhancement is
real time pictures are captured and is processed on the spot to used. Two types of contrast enhancement techniques are linear
get the result. At the very instant of time the disease can be and non linear. In this work contrast stretching is used which is
identified. The flowchart for the process is shown as in Fig. 5. a piecewise linear transformation function. Contrast stretching
The detailed explanation of the steps involved in the flowchart is done to extend the range of intensity levels in an image to
is given below. full intensity range of the medium and is depicted in Fig. 7.

1278
 This full-text paper was peer-reviewed and accepted to be presented at the IEEE WiSPNET 2017 conference.

The points (r1 , s1 ) and (r2 , s2 ) determine the shape of trans- underlying characteristics can be determined using this texture
formation structure and produces various degree of spread in analysis. The statistical texture features are calculated using
the intensity levels of the output image [15]. Gray level cooccurrence matrix (GLCM). The four main
features which are taken from GLCM for image analysis
C. Color Space Conversion are Contrast, Correlation, Energy and Homogeneity. Apart
This conversion is necessary because RGB color space is from this the other statistical features such as mean, standard
highly device dependent. Also it depends on the light inten- deviation, Entropy, RMS, Variance, Smoothness, Kurtosis, and
sity [16]. Image processing and analysis operations especially Skewness are also taken from the image. Choosing of statisti-
those involving color differences are easier to perform in cal parameter depends on requirements in output image [19].
L ∗ a ∗ b color space than in RGB. Also L ∗ a ∗ b space Consider p(i, j) as the element of co-occurrence matrix. It
is a device independent space. Initially color transformation represents the probability of moving from a pixel with gray
structure is created and then the structure is applied on the level i to a pixel with gray level j. Then parameters from
image to perform the conversion. The advantage of L ∗ a ∗ b GLCM can be represented by the below equations.
color model is that it is designed to approximate the human 1) Contrast: Measures local variations of gray levels
perception of light. L ∗ a ∗ b also separates the chrominance present in the image.
information than other models. 
Contrast = |i − j|2 p(i, j). (3)
D. K Means Clustering i,j
The colors in a ∗ b color space is segmented using k-means 2) Correlation: Determines correaltion of a pixel to its
clustering which is an unsupervised learning algorithm [17]. neighbour.
The number of clusters is fixed in advance. In our case k = 3
is chosen and the three clusters will have: the background,  (i − μi )(j − μj )p(i, j)
Correlation = . (4)
normal green leaf and the diseased portion of leaf. For each i,j
σi σj
cluster centroid is fixed. Placing the centroid is a crucial task.
The ideal way is to keep them away from one another. Each 3) Energy: Provides sum of squared elements in GLCM
data point is associated with the appropriate centroid. All the 
Energy = p(i, j)2 . (5)
points in the data is computed and first stage grouping is done.
i,j
In further steps, new centroids are determined from previous
step. Now with this new centroids the data point matching is 4) Homogeneity: Measures the closeness of distribution of
performed. This gets repeated in loop, until the location of k elements in GLCM to GLCM diagonal.
centroids cannot be changed anymore. The algorithm aims in  p(i, j)
minimizing the squared error function. Homogeneity = . (6)
i,j
1 + |i − j|
k 
 n
(j)
J= ||xi − cj ||2 (2) 5) Entropy: The amount of information which must be
j=1 i=1 coded is described by a quantity called as Image entropy.
(j) 
where: J = error function xi = data point cj = cluster Entropy = − Pi log2 Pi . (7)
(j)
center ||xi − cj ||2 = squared distance between data point
and cluster. The algorithm can be explained in the following Pi is the probability that the difference between 2 adjacent
steps: pixels is equal to i.
6) Average: Consider a region R. a[m, n] represents the
1) Initial k centroids are placed in the space represented by
pixel brightness. So sample mean of this over ∧ pixels will
the objects that are being clustered
give the average brightness ma in that region
2) The closest centroid to each data point is chosen and then
data points are associated with corresponding centroid. 1 
ma = a[m, n] (8)
3) New k centroid points are recalculated. ∧
(m,n)∈R
4) The steps 2 and 3 has to be repeated until no new points
can be found. where m, n ∈ R.
After this three different clusters for the given input image 7) Standard Deviation (S.D): Consider the brightness of
is generated. The appropriate cluster in which disease segment pixels in region R of an image. The sample standard deviation
is present will be selected for feature extraction. (sa ), is the estimate of standard deviation of this brightness
and is given by:
E. Feature Extraction 
1 
The main aim of feature extraction is to minimize the sa = (a[m, n] − ma )2 . (9)
resources which are required to represent large set of data ∧−1
accurately [18]. When large variables are considered the The other statistical parameters as mentioned earlier were
classification algorithm performs poorly. Certain unique and also extracted from the diseased portion of the leaf.

1279
 This full-text paper was peer-reviewed and accepted to be presented at the IEEE WiSPNET 2017 conference.

Fig. 8. 1AA approach.

F. Classification: Multiclass SVM

SVM is a supervised classification technique. In supervised
classification technique, training data is created with label Fig. 9. Contrast enhancement.
based on the sample image collected. In this work, each
disease is considered as a separate class and labels are given
accordingly. This training data refers to a set of examples
that are used to create a database for the learning process.
In supervised classification, each training set is analyzed and
a corresponding output is formulated. When the test data is
given, the class labels should be assigned to the test data.
SVMs are effective and accurate even if the training samples
are small. Earlier works have also indicated that SVM gave
better results than Neural Network for classification. Hence in
this work, SVM is used to perform supervised classification.
SVMs are binary classifiers but they can be made to handle
multiple classifications. The two approaches followed for
this multiple classification are: One-Against-One (1A1) and
One-Against-All (1AA) techniques.
1AA involves division of N classes into N two-class
machines. If the three classes are labeled as 1,2 and 3 then
classification would be carried out by comparing class 1 Fig. 10. Leaf image after clustering.
against non-class 1 (class 2,3) or class 2 against non-class
2 (class 1,3) or class 3 against non-class 3 (class 1,2) as
depicted in Fig. 8. The other approach 1A1 involves creation identified by classifying it using multiclass SVM. The four
of N (N − 1)/2 combinations for classification. Considering key features which are extracted from GLCM are tabulated in
the reduced complexity in 1AA, it is used for classifying the Table I.
diseases [20]. The three classes in this work are leaf spot SVM is a discriminative classifier which separates two
disease, leaf miner and CMV. Apart from this another class classes using hyperplane. When the hyperplane deviates from
for healthy leaves is also chosen. linear to complex shapes, SVM projects the data from input
space which is non linear to higher dimension. Kernel is a
V. R ESULT function which is used to project the data from input space
The very first step is to load the test data (disease attacked to feature space. RBF kernel is used in this work. The leaf
cucumber leaf). It is improvised by performing pre processing disease is detected and is displayed in Fig. 11.
operations. The image is resized, contrast enhanced and is
VI. C ONCLUSION
depicted in Fig. 9. Color space conversion converts the RGB
image into L∗a∗b color space. Kmeans clustering with k = 3 This paper proposed a technique which uses 1AA approach
gives three clusters: background, leaf portion and diseased of multiclass SVM to classify several diseases which are
segment of leaf and is depicted in Fig. 10. Disease affected present in cucumber leaves. Diseases such as leaf spot disease,
cluster is chosen and the features such as Entropy, Variance, leaf miner and CMV of cucumber leaves are considered for
Kurtosis, Mean, Standard deviation and several others are the analysis. After preprocessing these images, the cluster
considered to classify the disease present. The thirteen features containing the diseased part of the leaf is chosen using
are extracted from the test data. Using this the disease is k-means clustering algorithm. From that cluster, features are

1280
 This full-text paper was peer-reviewed and accepted to be presented at the IEEE WiSPNET 2017 conference.

TABLE I If this model is developed into a mobile app, farmers can

F EATURES FROM GLCM. utilize this and will be benefited.
Name of the Contrast Correlation Energy Homogeneity R EFERENCES
disease
[1] Dheeb Al Bashish, Malik Braik, and Sulieman Bani-Ahmad, “A frame-
Leaf miner 0.754979 0.790659 0.841598 0.963275 work for detection and classification of plant leaf and stem diseases,”
1.082384 0.831427 0.607212 0.915662 in Signal and Image Processing (ICSIP), 2010 International Conference
0.477895 0.953105 0.523771 0.92715 on. IEEE, 2010.
0.644393 0.932684 0.568736 0.927365 [2] Zainon Radhiah, “Paddy Disease Detection System Using Image
0.913817 0.835928 0.768592 0.948597 Processing,” (2012).
Leaf spot 0.533977 0.954996 0.584004 0.941629 [3] Piyush Chaudhary, et al. “Color transform based approach for disease
0.466023 0.945998 0.650916 0.947643 spot detection on plant leaf,” International Journal of Computer Science
0.703983 0.920203 0.654252 0.937726 and Telecommunications, vol. 3, no. 6, pp. 65–70, 2012.
0.47114 0.945399 0.659171 0.949605 [4] P. R. Rothe and R. V. Kshirsagar, “Cotton leaf disease identification
0.906342 0.846709 0.779882 0.957075 using pattern recognition techniques,” in Pervasive Computing (ICPC),
2015 International Conference on. IEEE, 2015.
Mosaic disease 0.902711 0.521799 0.649452 0.947986 [5] P. Revathi and M. Hemalatha. “Advance computing enrichment evalua-
0.875888 0.552953 0.582059 0.94039 tion of cotton leaf spot disease detection using Image Edge detection,”
1.291161 0.530282 0.72889 0.947129 in Computing Communication & Networking Technologies (ICCCNT),
1.208609 0.494147 0.733681 0.953626 2012 Third International Conference on. IEEE, 2012.
0.905178 0.548338 0.655565 0.954097 [6] D. S. Guru, P. B. Mallikarjuna, and S. Manjunath, Segmentation and
Healthy leaf 1.473254 0.864399 0.433547 0.890061 classification of tobacco seedling disease, IEEE, COMPUTE11, (2010),
1.841452 0.835821 0.428414 0.884871 2011.
0.491789 0.951453 0.380233 0.906061 [7] Marek Schikora and Adam Schikora, “Image-based analysis to study
0.544056 0.949096 0.462481 0.917352 plant infection with human pathogens,” Computational and Structural
1.492096 0.745099 0.604216 0.912317 Biotechnology Journal, vol. 12, no. 20, pp. 1–6, 2014.
[8] Zhang Jian and Zhang Wei, “Support vector machine for recognition of
cucumber leaf diseases,” in Advanced Computer Control (ICACC), 2010
2nd International Conference on, 2010, vol. 5. IEEE.
[9] C. Arun Priya, T. Balasaravanan, and Antony Selvadoss Thanamani,
“An efficient leaf recognition algorithm for plant classification using sup-
port vector machine,” in Pattern Recognition, Informatics and Medical
Engineering (PRIME), 2012 International Conference on. IEEE, 2012.
[10] Elham Omrani, et al. “Potential of radial basis function-based support
vector regression for apple disease detection,” Measurement, vol. 55,
pp. 512–519, 2014.
[11] Santanu Phadikar and Jaya Sil, “Rice disease identification using pattern
recognition techniques,” in Computer and Information Technology, 2008.
ICCIT 2008. 11th International Conference on. IEEE, 2008.
[12] Ravi Shankar, Seema Harsha, and Raj Bhandary, “A practical guide to
identification and control of cucumber diseases,”
[13] Mary Ann Hansen, “Angular leaf spot of cucumber,” (2009).
[14] Viktor Knutsson, “Terramechanics based wheel-soil model in a computer
game enviroment,” (2016).
[15] Rafael C. Gonzalez and Richard E. Woods, “Image processing,” Digital
Image Processing, vol. 2, 2007.
[16] Shitala Prasad, Sateesh K. Peddoju, and Debashis Ghosh.
“Multi-resolution mobile vision system for plant leaf disease diagnosis,”
Signal, Image and Video Processing, vol. 10, no. 2, pp. 379–388, 2016.
[17] R. Anand, S. Veni, and J. Aravinth, “An application of image processing
Fig. 11. Disease detection.
techniques for detection of diseases on brinjal leaves using k-means clus-
tering method,” in Recent Trends in Information Technology (ICRTIT),
2016 International Conference on. IEEE, 2016.
extracted using GLCM and this is used for classification of [18] P. Mohanaiah, P. Sathyanarayana, and L. GuruKumar “Image texture
diseases. feature extraction using GLCM,” International Journal of Scientific and
Research Publications, vol. 3, no. 5, May 2013, ISSN 2250–3153.
As a future work, this can be extended to bigger farms. To [19] Vijay Kumar and Priyanka Gupta, “Importance of statistical measures in
monitor large farms, cameras can be fixed in intervals which digital image processing,” International Journal of Emerging Technology
can capture the leaf images in definite intervals. By inspecting and Advanced Engineering, ISSN 2250–2459, vol 2, no. 8, Aug. 2012.
[20] K. P. Soman, R. Loganathan, and V. Ajay. Machine learning with SVM
these images regularly the disease can be detected in early and other kernel methods. PHI Learning Pvt. Ltd., 2009.
stage itself which can prevent spread of disease in large scale.

1281