“Diagnosis and classification of grape leaf diseases using neural networks”

Abstract

Plant diseases cause significant damage and economic losses in crops. Subsequently, reduction in plant diseases by early diagnosis results in substantial improvement in quality of the product. Erroneous diagnosis of disease and its severity leads to inappropriate use of pesticides. The goal of proposed work is to diagnose the disease using image processing and artificial intelligence techniques on images of grape plant leaf. In the proposed system, grape leaf image with complex background is taken as input. Thresholding is deployed to mask green pixels and image is processed to remove noise using anisotropic diffusion. Then grape leaf disease segmentation is done using K-means clustering. The diseased portion from segmented images is identified. Best results were observed when Feed forward Back Propagation Neural Network was trained for classification.

Full text

“Diagnosis and Classification of Grape Leaf
Diseases using Neural Networks”
Sanjeev S Sannakki1, Vijay S Rajpurohit2, V B Nargund3, Pallavi Kulkarni4*
sannakkisanjeev@yahoo.co.in vijaysr2k@yahoo.com nargund56@gmail.com pallavideshpande66@gmail.com
1,2,4 Dept. of Computer Science, Gogte Institute of Technology Belgaum, Karnataka, India
3Department of Plant Pathology, University of Agricultural Sciences, Dharwad, India
Abstract—Plant diseases cause significant damage and
economic losses in crops. Subsequently, reduction in plant
diseases by early diagnosis results in substantial improvement
in quality of the product. Erroneous diagnosis of disease and
its severity leads to inappropriate use of pesticides. The goal of
proposed work is to diagnose the disease using image
processing and artificial intelligence techniques on images of
grape plant leaf. In the proposed system, grape leaf image with
complex background is taken as input. Thresholding is
deployed to mask green pixels and image is processed to
remove noise using anisotropic diffusion. Then grape leaf
disease segmentation is done using K-means clustering. The
diseased portion from segmented images is identified. Best
results were observed when Feed forward Back Propagation
Neural Network was trained for classification.
Index Terms—plant disease identification, Feed forward
neural network, image processing, k-means, co-occurrence
matrix, feature extraction
I. INTRODUCTION
Grape (Vitis vinifera) cultivation-Viticulture is one of the
most remunerative farming enterprises in India [1]. Grapes
originated in Western Asia and Europe. Fruit is eaten fresh
or made into juice, fermented to wines and brandy and dried
into raisins [2]. Grapes also have medicinal properties to
cure many diseases.
Grapes generally require a hot and dry climate during its
growth and fruiting periods. It is successfully grown in areas
where the temperature range is from 150-400C. High
temperatures above 400C during the fruit growth and
development reduce fruit set and consequently the berry
size. Low temperatures below 15 C followed by forward
pruning impair the budbreak leading to crop failure. Grapes
can be cultivated in variety of soils including sandy loams,
red sandy soils, sandy clay loams, shallow to medium black
soils and red loams [16].
Grape suffers from huge crop losses on account of
downy mildew, powdery mildew and anthracnose [1]. In
case of downy mildew, the losses are very high when the
clusters are attacked before fruit set. Entire clusters decay,
dry and drop down [16]. Plant disease is one of the crucial
causes of reduction in quantity and degrades quality of the
product. The naked eye observation of experts is the main
approach adopted in practice for detection and identification
of plant diseases. This approach is prohibitively expensive
and time consuming in large farms [3]. Further, in some
developing countries, farmers may need to go long distances
to contact experts. Diseases are managed by adjusting the
pruning time and using various fungicides [2]. Observations
during research at NRCG, Pune show that precision farming
i.e. using information technology for decision making has
improved the yield and quality of crops.
II. LITERATURE SURVEY
Various researchers have proposed image-processing and
pattern recognition techniques in agricultural applications
for detection of weeds in a field, sorting of fruits and veget-
ables, detecting diseases etc. Automatic detection of plant
diseases is an essential research topic as it may prove bene-
fits in monitoring large fields of crops, and detect the symp-
toms of diseases as soon as they appear on plant leaves.
Therefore; looking for fast, less expensive and accurate
method to detect plant disease cases is of great significance.
Excessive use of pesticides for plant disease treatment in -
creases costs and raises the danger of toxic residue levels on
agricultural products. This requires that the disease must be
identified accurately and also the stage in which the disease
is. Hence an efficient disease identification and diagnosis
model is required.
Al-Bashish, D., M. Braik and S. Bani-Ahmad have
proposed the leaf disease Detection and classification for
early scorch, cottony mold, late scorch and tiny whiteness
with K-means-based segmentation, CCM to consider texture
feature classification and BPNN to classify diseases into one
of the six disease classes [4]. A. Camargo, J.S. Smith
discuss converting the RGB image of the diseased plant or
leaf, into the H, I3a and I3b color transformations. The
transformed image is then segmented by analyzing the
distribution of intensities in a histogram. The extracted
region was post-processed to remove pixel regions not
considered part of the target region. Then the neighborhood
of each pixel and its gradient is analyzed [5]. S. S.
Sannakki, V. S. Rajpurohit, V B Nargund, et.al proposes K-
means clustering for segmentation and Disease grading by
Fuzzy Logic where the intensity of disease is decided by
the area of diseased portion [6]. H. Al-Hiary, S. Bani-
Ahmad, M. Reyalat, et.al propose masking of green pixels
before extracting feature which results in more accurate
classification [7].
IEEE - 31661
4th ICCCNT 2013
July 4-6, 2013, Tiruchengode, India

All these researchers gather the leaf samples and take
images in controlled environment i.e. plane background,
good lightening conditions and click from constant distance.
A. Meunkaewjinda, P. Kumsawat et.al propose SOFM
and BPNN to recognize grape leaf color, MSOFM for
segmentation, GA and SVM for classification[8], process
the images with complex background but use many machine
learning algorithms which makes the system complex.
III. SYSTEM DESIGN
Figure 1: Decision Support System
The proposed system aims at processing the images with
complex background, varying lightening conditions and
clicked from various distances. This makes the system more
dynamic to work under various climatic conditions. Figure 1
shows the corresponding block diagram which is an image
processing system for grape leaf in which the following
steps are undertaken.
A. Image acquisition: The First step is to collect the sample
images which are needed to train system. In the working
system, this step indicates the input or query image. Leaf
images are captured using digital camera, Nikon Coolpix
P510, 16.1 Megapixel and are used for training and testing
the system. All the images are stored in standard jpg format.
In the present study, images are captured from different
regions like Pune, Bijapur, Sangali and expert advice is taken
for identification. Some images are downloaded from
internet to have diverse environment. Images gathered
include the leaves affected by two major diseases found in
India, downy mildew and powdery mildew.
B. Background removal: In this step, the input image is
resized to standard size 300x300. Then, mostly green colored
pixels are identified. If the green color component is less
than threshold i.e. 70 in present work, then all red, green and
blue value of that pixel is assigned zero and green channel of
that image is assigned to 255. This is called masking green
pixels. This fastens the processing in the next step and also
improves accuracy [7].
A mask is a black and white image of the same
dimensions as the original image (or the region of interest
you are working on). Therefore, each of the pixels in the
mask can have a value either 0 (black) or 1 (white) as shown
in Figure 2. When executing operations on the image the
mask is used to restrict the result to the pixels that are 1
(selected, active or white) in the mask and the operation
restricts to some parts of the image.
Figure 2: The query image, RGB channels, mask and resulting image
after masking.
C. Preprocessing: Then it is enhanced by five iterations of
Anisotropic Diffusion [10] to preserve the information of
affected portion. The diffused image is shown in Figure 3.
Anisotropic diffusion is a generalization of this diffusion
process; it produces a family of parameterized images, but
each resulting image is a combination between the original
image and a filter that depends on the local content of the
original image. As a consequence, anisotropic diffusion is
a non-linear and space-variant transformation of the original
image.
Non-linear, Space variant method:
(1)
(2)
(3)
Equation 1 gives gradient of the brightness function. We
have an edge/not edge estimation method “E” in equation 2.
Equation 3 is used which will not only preserve but also
sharpen the edges if g(.) is chosen properly. First equation
amongst the two equations proposed by Perona and malik is
used. H component from HSV color space is extracted to
reduce the illumination effect.
Figure 3: Filtered image after Anisotropic diffusion.
D. Segmentation: Clustering of data is a method by which
large sets of data are grouped into clusters of smaller sets or
IEEE - 31661
4th ICCCNT 2013
July 4-6, 2013, Tiruchengode, India

segments of similar data. In present work, k-means
clustering [10] is used to for segmenting an image into six
groups which is found to be optimum as shown in figure 4.
One or more (In case more diseases are present at a time)
clusters contain the diseased portion of leaf. a and b
component from L*a*b space are extracted before clustering.
K-means clustering algorithm:
x1,…, xN are observed data points or vectors.
Each observation (vector xi) will be assigned to
only one cluster.
C(i) is cluster number for the ith observation
For a given cluster assignment C of the data points,
compute the cluster means mk where
(4)
For a current set of cluster means, each observation
is assigned as:
(5)
Iterate last two steps until convergence.
Figure 4: Six clusters formed by K-means clustering.
E. Extract lesion: Once the image is divided into six
clusters, the mean of each cluster is calculated and means are
sorted in ascending order. It is observed that the downy
affected lesion is extracted at second (Figure 5) and powdery
affected lesion is extracted at sixth in the sorted clusters. It is
observed that this is true for the leaves having lesions of both
the diseases at same time.
(a) (b)
Figure 5: (a) Downy affected region (b) Powdery affected region
F. Feature extraction: The next step is to extract texture
features of the extracted diseased portions. This is done by
calculating Gray Level Co-occurrence Matrix (GLCM) [12].
The colour co-occurrence texture analysis method was
developed through the use of spatial gray-level dependence
matrices (SGDM’s). Co-occurrence matrices measure the
probability that a pixel at one particular gray-level will occur
at a distinct distance and orientation from any pixel given
that pixel has a second particular gray-level.
The SGDM’s are represented by the function
P(i,j,d,θ) where i represents the gray-level of location (x,y)
in the image, and j represents the gray-level of the pixel at a
distance d from location (x,y) at an orientation angle of θ.,
where i is the row indicator and j is the column indicator in
the SGDM matrix P(i,j,d,θ). The nearest neighbour mask,
where the reference pixel (x,y) is shown as an asterisk. All
eight neighbors shown are one pixel distance from the
reference pixel ‘*’ and are numbered in a clockwise direction
from one to eight as shown in figure 6. The neighbors at
positions 1 and 5 are both considered to be at an orientation
angle equal to 0◦, while positions eight and four are
considered to be at an angle of 45◦.
6 7 8
5 * 1
4 3 2
Figure 6: Directions considered in Co-occurrence matrix
Co-occurrence matrix is then normalized. The equation
for normalizing co-occurrence matrix is given in Equation
6.
(6)
In the above equation, P(i, j, 1, 0) is the intensity co-oc-
currence matrix and Ng represents the total number of in-
tensity levels.
Texture features can be used as useful discriminator when
target images do not follow well defined color or shape
[16]. Nine texture features listed in table I, are extracted for
each image which will be used for training Neural Network
in the next step for classification.
Classification: The feed forward Back Propagation Neural
Network classifier [13] [15] consisting of three layers
namely input layer, a hidden layer, and an output layer is
used.
Training Back propagation Neural Network
We need to train the network with available data. Initially the
network predicts an output for one input vector for which the
true output is known. This combination of known output and
input vector is called training sample. The predicted output is
compared to the known value. The weights on the arcs are
adjusted depending on how the prediction of the actual
.,,1,
)(:
Kk
N
x
m
k
kiCi
i
k




NimxiC
Kk
ki
,,1,minarg)(
1
2



00
450
1350
900
IEEE - 31661
4th ICCCNT 2013
July 4-6, 2013, Tiruchengode, India

result. Sigmoid transfer function is used for generating
output at each stage. The input layer has 9 nodes, which are
related to two 9 texture features—contrast, uniformity,
maximum probability, homogeneity, inverse difference,
difference variance, diagonal variance, entropy of H bands of
lesion area. Output layer contains two neurons. This module
assigns an appropriate disease class i.e. Downy or powdery.
Table I. Mathematical formulations of texture features [14]
No. Features Formula
1. Contrast ∑i ∑j|i − j|2 p(i, j, d, θ)
2. Uniformity(Energy) ∑i ∑j p(i, j, d, θ)2
3. Maximum probability max i,j p(i, j, d, θ)
4. Homogeneity ∑i ∑j p(i,j,d,θ)/(1+|i−j|)
5. Inverse difference
moment of order 2
∑i∑j1/(1+(i−j)2) p(i, j, d, θ)
6. Difference variation Variance of
∑i ∑j |i − j|p(i, j, d, θ)
7. Diagonal variance Variance of p(i, j, d, θ)
8. Entropy ∑i ∑j p(i, j, d, θ) log(p(i, j, d, θ))
9. Correlation ∑(i-µi)(j-µj)p(i,j)
i,j σiσj
IV. EXPERIMENTAL RESULTS
The data consisted of 16 images of powdery mildew (class
1) and 17 images of downy mildew (class 2). MATLAB 7.1
Neural network pattern recognition tool was used for train-
ing. In 33, 29 images are used for training and 2 each for
testing and validation. Data was given in two files namely
input and target. Input file with 9 rows representing texture
features and 33 columns representing sample images. Target
file consisted two rows with binary values. [0 1] for downy
and [1 0] for powdery. It gives good validation results as
shown in figure 7. Confusion matrix in figure 8(a) shows
images of class 1 and class 2 are classified properly. Figure
8(b) shows sensitivity (True positive) and specificity (False
positive) rate of the system.
Figure 7: Validation results after training
(a) (b)
Figure 8: Sensitivity and specificity of system and confusion matrix
The green line meeting at upper left corner in Figure 8(a)
shows the perfect result of binary classification test. Dur-
ing training it gave 100% correct results which shows the
system will work almost accurately. The green box indicates
correct classification and red box indicates wrong classifica-
tion. The model that used hue features gives accurate results
reaching a perfect 100%.
Figure 9: GUI developed using guide toolbox. Downy mildew is recog-
nized by the system
V. CONCLUSION AND FUTURE WORK
Study involved collecting leaf samples from different re-
gions. Work was carried out to investigate the use of com -
puter vision for classifying grape leaf diseases. Two classes
of grape leaves, i.e., downy mildew and powdery mildew
were considered in the experiments. Algorithms based on
image-processing techniques, feature extraction and classi-
fication, were deployed. The feature extraction process used
IEEE - 31661
4th ICCCNT 2013
July 4-6, 2013, Tiruchengode, India

color co-occurrence methodology, which uses the texture of
an image to arrive at unique features, which represent that
image. nprtool of MATLAB 7.1 was used to train the neural
network for pattern recognition. This training achieved
training accuracies of 100% when using hue features alone.
Future work of this study can explore the utility of these
techniques to include samples of healthy and other diseases
of grapes like anthracnose. Instead of K-means, other seg-
mentation techniques can be used to extract the lesion more
accurately. With hue, models can be constructed with
the combination of other color components like Saturation
and intensity and the results can be compared.
ACKNOWLEDGMENT
Dr. S. D. Sawant, National Research Centre for Grapes,
Manjri Farm, Solapur Road, Pune.
Dr. A. M. Sheikh, Regional Agriculture Research
Station, Hitnalli Farm, Bijapur,
REFERENCES
[1] A report of the expert consultation on viticulture in Asia and the
Pacific. May 2000, Bankok, Thailand. RAP
publication:2000/13.
[2] K. Soytong, W. Srinon et.al. Application of antagonistic fungi to
control anthracnose disease of grape Journal of Agricultural
Technology May 2005.
[3] Weizheng, S., Yachun, W., Zhanliang, C., and Hongda, W.
(2008). Grading Method of Leaf Spot Disease Based on Image
Processing. In Proceedings of the 2008 international Conference
on Computer Science and Software Engineering - Volume 06
(December 12 - 14, 2008). CSSE. IEEE Computer Society,
Washington, DC, 491-494. DOI=
http://dx.doi.org/10.1109/CSSE.2008.1649.
[4] Al-Bashish, D., M. Braik and S. Bani-Ahmad, (2011). Detection
and classification of leaf diseases using K-means-based
segmentation and neural-networks-based classification. Inform.
Technol. J., 10: 267-275. DOI: 10.3923/itj.2011.267.275
[5] Camargo, A. and Smith, J. S., (2009). An image processing based
algorithm to automatically identify plant disease visual
symptoms, Biosystems Engineering, Volume 102, Issue 1,
January 2009, Pages 9-21, ISSN 1537-5110, DOI:
10.1016/j.biosystemseng.2008.09.030.
[6] Sanjeev S Sannakki, Vijay S Rajpurohit, V B Nargund, et.al
(2011) Leaf Disease Grading by Machine Vision and Fuzzy
Logic, Int. J. Comp. Tech. Appl., Vol 2 (5), 1709-1716
[7] H. Al-Hiary, S. Bani-Ahmad, M. Reyalat, M. Braik and Z.
ALRahamneh. Fast and Accurate Detection and Classification
of Plant Diseases. International Journal of Computer
Applications (0975 – 8887) Volume 17– No.1, March 2011
[8] A.Meunkaewjinda, P.Kumsawat, K.Attakitmongcol et.al. Grape
leaf disease detection from color imagery using hybrid
intelligent system. Proceedings of ECTI-CON 2008.
[9] Peitro Perona ,Jitendra malik, scale space and edge
detection using Anisotropic Diffusion, IEEE transactions on
pattern analyses and machine intelligence, vol.12, no.7, July
1990.
[10] Macqueen, J.B.,1967. Some methods for classification and
analysis of multivariate observations. Proc.Berkely Symp.
Math. Statist. Prob., 1:281-297.
[11] R. Pydipati, T.F. Burks, W.S.Lee. Identification of citrus
disease using color texture features and discriminant analysis.
Computers and Electronics in Agriculture 52 (2006) 49–59.
[12] I.A. Basheer, M. Hajmeer , Artificial neural networks
fundamentals, computing, design, and application, Journal of
Microbiological Methods 43 (2000) 3–31.
[13] Huang K.Y.Application of artificial neural network for detecting
Phalaenopsis seedling diseases using color and texture feature.
Computer and Electronics in agriculture. Pp. 3-11, 2007.
[14] Prasad Babu, M. S. and Srinivasa Rao, B. (2010) Leaves
recognition using back-propagation neural network - advice for
pest and disease control on crops. Technical report, Department
of Computer Science & Systems Engineering, Andhra
University, India, www.indiakisan.net on May 2010.
[15] A. Camago,J.S.Smith, Image pattern classification for the
identification of disease causing agents in plants, Computers
and Elecronics in Agriculture 66(2009) 121-125.
[16] http://nhb.gov.in/fruits/grape/gra012.pdf, gra002.pdf
[17]http://www.nrcgrapes.nic.in/zipfiles/GRAPE%20PROFILE%20-
%20NRC%20Grapes.pdf
IEEE - 31661
4th ICCCNT 2013
July 4-6, 2013, Tiruchengode, India