International Journal of Information and Electronics Engineering, Vol. 2, No.

3, May 2012

Classification of Rice Leaf Diseases Based on

Morphological Changes
S. Phadikar, J. Sil, and A. K. Das

blast and the brown spot diseases based on the morphological

Abstract—Rice is widely cultivated economical crop in the changes caused by the diseases. The accuracy of the
world. During cultivation the earliest and accurate diagnosis of proposed system is compared with the existing classifier,
the rice plant diseases able to reduce the damage, resulting which shows comparable results.
environment protection and better return. In the work, an
automated system has been developed to classify the leaf brown
The paper, is organized in four Section II describes the
spot and the leaf blast diseases of rice plant based on the process of image acquisition and feature extraction. Section
morphological changes of the plants caused by the diseases. III describes the classification process. Section IV represents
Radial distribution of the hue from the center to the boundary results and discussion and the paper is concluded in section
of the spot images has been used as features to classify the V.
diseases by Bayes’ and SVM Classifier. The system has been
validated using 1000 test spot images of infected rice leaves
collected from the field, gives 79.5% and 68.1% accuracies for
Bayes’ and SVM Classifier based system respectively. II. IMAGE ACQUISITION AND FEATURES EXTRACTION
The images are collected from different parts of East
Index Terms—Rice leaf disease, Bayes classifier, leaf blast, Midnapur, a major rice producing district of South Bengal,
brown spot, support vector machine, radial hue distribution. India and the features for classification of the rice leaf
diseases has been extracted through execution of the
following steps:-
I. INTRODUCTION
a) Acquire the images of the rice leaf of infected
Rice (Oryzae Sativa) is considered as the main crop in the diseased plants from the field.
east India and believed to be the second central crop after
b) Preprocess the images to remove noise from the
wheat, in the world. Rice blast caused by the fungus infected portion of the leaf image.
Pricularia Grisea [1] - [3] one of the major rice diseases
c) For a diseased leaf, execute steps (d), (e)
occur in most of the rice growing areas in the world. It causes
severe problem, especially in the areas where the resistant d) Extract the infected portion of the image by
segmentation.
varieties are not available and the environment condition is
favorable towards out brake of the disease. The damages e) Compute radial hue distribution vectors of the
caused by the Blast, depend on the degree of severity of the segmented regions, which are used as feature
disease. Another important rice disease Leaf brown spot vectors.
caused by the fungus Bypolaris Oryza [1], [2], [4] is visible A. Image Acquisition
throughout the growing season. Brown spots generally The images are grabbed using Nikon COOLPIX P4 digital
become severe when rice is grown on silicon-deficient soils. camera in macro mode. The samples of the leaves affected by
New trends of research in agriculture aim towards the Diseases (blast, brown spot) and the uninfected rice leaves
development of the disease resistant variant of seed using are shown in Fig 1.
gene technology, which increase productivity and food
quality at reduced expenditure [4]. With the advancement of
information technology, remote sensing techniques have
been used in the field of crop management described in [5] –
[7]. A relation among the ground disease index and remote
sensing data is established in [8] to classify the diseases.
Very recently, image processing techniques are applied to Fig. 1. Acquired rice (oryzae sativa) leaf images (a) Uninfected rice leaf
detect the rice diseases automatically [9] In the paper, an (b) Rice leaf infected by blast(pricularia grisea) (c) Rice leaf infected by
brown spot (bypolaris oryza)
automated system has been designed using image processing
and pattern recognition techniques to distinguish the leaf
B. Preprocessing
Manuscript received February 24, 2012; revised April 26, 2012. The automated system first classifies the acquired image
S. Phadikar is with the National Institute of Standards and Technology,
Boulder, CO 80305 USA (e-mail: author@ boulder.nist.gov). either as an uninfected or a diseased leaf. For an uninfected
J. Sil was with Rice University, Houston, TX 77005 USA. He is now with leaf the distribution of color is nearly uniform while for the
the Department of Physics, Colorado State University, Fort Collins, CO diseased leaf the distribution is not uniform because the pixel
80523 USA (e-mail: author@lamar. colostate.edu).
A. K. Das is with the Electrical Engineering Department, University of values of the diseased leaves varies widely and differs from
Colorado, Boulder, CO 80309 USA, on leave from the National Research the normal intensity value of pixels. In order to distinguish
Institute for Metals, Tsukuba, Japan (e-mail: author@nrim.go.jp). between the uninfected and diseased leaves, the aim of the

460
 International Journal of Information and Electronics Engineering, Vol. 2, No. 3, May 2012

preprocessing step is to reduce the difference among the Fig. 5 (a) and (b) are the segmented images of Figure 1 (b)
intensity value of pixels of the uninfected image while and (c) respectively, obtained by applying Otsu’s
increase the difference among the pixels of the diseased segmentation algorithm in the Hue plane and taking the color
image. The processed gray image has been obtained using (1) values corresponding to the segmented region.
B. Feature Selection and Extraction
MG=2×(G-(0.75×R-0.25×B)) (1) The morphological changes caused by the diseases are
studied extensively and it has been observed that the brown
spots are oval shaped while the blast spots are elliptical
where, R, G and B are Red, Green and Blue plane shaped. The shapes are distorted with the severity of the
respectively in the RGB color image. The processed images disease. The spots have different color at the center and the
are shown in Fig. 2. boundary. Radial distribution of the hue from the center
The quality of the image is enhanced further applying towards the boundary of the spot has been considered as
the[mean filtering technique [10] and the resultant images are feature to distinguish the spots since the color of the spots
shown in Fig. 3. vary from center towards the boundary. Hue represents the
true color and the hue distribution values are stored in vectors,
denoted as H = {h1, h2,…….h8}, where hi is the vector
containing the hue value of the spots corresponding to the
line number shown in Fig. 5.

Fig. 2. Modified Gray level images corresponding to Fig. 1. obtained

by applying the equation 2*(G-(0.75*R-0.25*B))

Fig. 5. Eight direction of radial features

III. CLASSIFICATION
Following two different phases of classification is
Fig. 3. Enhanced images corresponding to Fig. 2. obtained by performed.
applying the mean filter of size 3X3 • Classification of uninfected and diseased leaf
• Classification between diseased leaf
A. Classification of Uninfected and Diseased Leaf
Histogram represents the probability of occurrences of
different gray levels in the image. The high probability of a
particular gray level corresponds to a peak in the histogram.
For an uninfected image, the probability of a particular gray
level is very high since all the pixels have similar gray level
Fig. 4.(a) and (b) are the segmented images of Fig. 1(b) and (c) values. On the other hand, a diseased image exhibits high
respectively, obtained by applying Otsu’s segmentation algorithm in the Hue
plane and taking the color values corresponding to the segmented region.
probability for at least two sufficiently different gray levels,
one for the diseased region of pixels and other for the
uninfected region of pixels. Therefore, the histogram of an
A. Image Segmentation image with single peak is treated as uninfected and more than
Image segmentation is the process of subdividing the one peak is treated as the diseased leaf.
image into its constituent parts [10] of interest. There are
several methods [10], [11] of image segmentation like
thresholding [12] - [14], region based [15]-[18] methods,
morphological operation [10] and soft computing techniques
[12], [19], [20]. The selection of the algorithm generally
depends on the problem being solved. Here, Otsu’s threshold
based segmentation algorithm [13] has been applied which Fig. 6. Histogram of the images corresponding to Fig. 3
maximizes the between class variance parameter to detect the
threshold value. First, the number of peak is determined by translating the
Hue plane of the image has been used for the threshold histogram upward with the factor (m×n/250), where m×n is
selection. Using that threshold value the image has been the size of the image. It has been observed that in order to
segmented into two parts. Fig. 4 (a) and (b) show the reduce the noise, and to obtain clarity of the image size of the
segmented images correspond to Fig. 1 (b) and (c) infected portion of the leaf must be at least greater than
respectively obtained by taking the color values (1/250)th of the original image size. The histograms of the
corresponding to the black pixel of the segmented images. images are shown in Fig. 6. The number of peak is counted

461
 International Journal of Information and Electronics Engineering, Vol. 2, No. 3, May 2012

by considering change of direction of the histogram.

B. Classification between Diseased Leaves Therefore the Bayes’ classifier assigns a particular pattern
Classifiers based on Bayes’ theorem and Support Vector x to class ωi if
Machines (SVMs) are applied for diseased image p (x)- p(x/ ωi) p(ωi ) < p (x)- p(x/ ωj) p(ωj) )
classification and the performance are compared or
1) Bayes’ Classifier p(x/ ωi) p(ωi ) > p(x/ ωj) p(ωj) ,
Define Bayes’ classifier determines the belongingness of a j=1,2,3.....M; j≠i (8)
pattern in a particular class depending on the likelihood of the
pattern with that class [21]. The probability that a pattern x 2) Support Vector Machine (SVM)
belongs to class ωi is denoted by p(ωi /x). Assume pattern x SVMs [22], [23] are supervised machine learning
actually belongs to class ωi but the classifier detects that x technique that can be analyzed theoretically using concepts
comes from class ωj, and so it incurs loss equal to Lij. Since from computational learning theory while being able to
the pattern x may belong to any of the M classes under achieve good performance when applied to real-world
consideration, the expected loss incurred to assigning x to problems. The most attractive feature of the SVM is the
class ωj is formulated in (2). maximum-margin hyper plane, the soft margin and the kernel
M function. For classifying any two linearly separable classes
rj(x) = ∑ Lij p(ωi /x) (2) there may exist many separating lines that correctly classify
the data. Among these lines the SVM select the line, which
i=1 maximizes the distance between the separating hyper-planes.
This is often referred to as the conditional average risk in To explain it clearly we label the training data {xi, yi}, i =
decision theory terminology. The classifier has M possible 1,…….. , l, yi ∈ {-1, 1}, xi ∈ Rd where l denotes the total no of
categories to classify each pattern. Equation (2) computes the training sample and d denotes the dimension of the feature
quantities r1(x), r2(x),…… rM(x), for each x, and assign each vector. Suppose we have some hyper-plane, which separates
pattern to the class with the smallest conditional loss, which the positive from the negative examples (a “separating
produces total expected loss with respect to all decision is hyper-plane”). The points x which lie on the hyper-plane
minimum. The classifier, which minimizes the total expected satisfy w · x + b = 0, where w is normal to the hyper-plane,
loss, called Bayes’ classifier. From the statistical point of |b|/║w║ is the perpendicular distance from the hyper-plane
view the Bayes’ classifier represents the optimum measure to to the origin, and ║w║ is the Euclidean norm of w. Let d+ (d-)
performance using Bayes’ formula, given in (3) be the shortest distance from the separating hyper-plane to
the closest positive (negative) instance. Define the “margin”
of a separating hyper-plane to be d+ + d-. For the linearly
p(ωi /x)= p(ωi )p(x/ωi) (3)
separable case, the support vector algorithm simply looks for
p(x)
the separating hyper-plane with largest margin. This can be
formulated as follows: suppose that all the training data
Thus the (2) is rewritten as (4)
satisfy the following constraints:
1 M
xi* w + b = +1 for yi = +1 (9)
rj(x) = — ∑ Lij p(ωi)p(x/ωi) (4)
xi* w + b = -1 for yi = -1 (10)
p(x) i=1
These can be combined into one set of inequalities:
where p(x/ωi ) is called the likelihood function of class ωi .
Since 1/p(x) is a common factor so dropped from equation (4) yi(xi * w + b) - 1 ≥ 0 ∀i (11)
and the expression of average loss then reduce to (5). Now consider the points for which the equality of (9) hold
M (requiring that there exists such a point is equivalent to
choosing a scale for w and b) good. These points lie on the
rj(x) = ∑ Lij p(x/ ωi) p(ωi) (5) hyper-plane H1 : xi · w + b = 1 with normal w and
i=1 perpendicular distance from the origin |1 - b|/║w║. Similarly,
Therefore, a pattern x is assigned to a class ωi if ri(x) < rj(x) the points for which the equality of (10) holds false, these
for j=1,2,3,…..,M, where j≠i. points lie on the hyper-plane H2: xi · w + b = -1, with normal
However, a special type of loss function has been used in w, and perpendicular distance from the origin | - 1 - b|/║w║.
the paper, which computes loss to zero for the correct Hence d+ = d- = 1/║w║ and the margin is simply 2/║w║.
decision and same nonzero value for all other erroneous Note that H1 and H2 are parallel (they have the same normal)
decisions. The loss function is expressed as (6). and so no training points reside between them. Thus we can
find the pair of hyper-planes, which gives the maximum
Lij=1-δij (6) margin by minimizing ║w║2, subject to constraint (11), as
where δij =1 when i=j and δij =0 when j≠i. If we substitute the depicted in Fig. 7. Other important characteristic of the SVM
loss function of equation (5) using (6), it becomes as follows: is the soft margin that gives the user flexibility to choose the
parameter to roughly control the number of examples.
M
Another feature is the kernel function that projects the
rj(x) = ∑ (1-δij )p(x/ ωi) p(ωi)
non-linearly separable data from low-dimensional space to a
i=1
space of higher dimension so that they may become separable
= p(x)- p(x/ ωj) p(ωj) (7)

462
 International Journal of Information and Electronics Engineering, Vol. 2, No. 3, May 2012

in the higher dimensional space too. feature vector and N is the number of training samples. Since
number of samples generally much larger than the dimension
of the feature vector, therefore the proposed system is time
efficient compare to SVM.

REFERENCES
[1] S. H. OU. In: Rice Diseases. Kew Surrey, England, Commonwealth
Mycological Institute, 1972
[2] C. Eric. In: Report on Study Tour to Examine Symptoms of Rice
Diseases in Southern USA and California. Principal Research Scientist
NWS Agriculture, Agricultural Institute, Orange NSW, 2002
[3] R. K. Webster, In: Rice Blast Disease Identification Guide. Davis,
Fig. 7. Hyper-plane separated by the SVM Dept. of Plant Pathology, University of California, 2000
[4] H. Sato, I. Ando, H. Hirabayash, Y. Takeuchi, S. Arase, J. Kihara, H.
Kato,T. Imbe, and H. Nemoto, “QTL Analysis of Brown Spot
Resistance in Rice(Oryza Sativa,” Breeding Science, vol. 58, pp. 93-96,
IV. RESULT AND DISCUSSION 2008.
Five hundred samples of each data class are considered to [5] H. L. Khatibl, F. Hawels, H. Hamdi, and N. L. Mowelhi. “Spectral
Characteristics Curves of Rice Plants Infected by Blast,” IEEE
test the system. For the first level classification (i.e. Geoscience and Remote Sensing Symposium, vol. 2, pp. 526-528,
uninfected leaf or a diseased leaf), it has been found that 1993.
success rate is 92%, 96.4% and 84% respectively for [6] T. Kobayashi, E. Kanda, K. Kitada, K. Ishiguro, and Y. Torigoe,
“Detection of Rice Panicle Blast with Multispectral Radiometer and
uninfected leaf, leaf with brown spot and leaf with blast, the Potential of Using Airborne Multispectral Scanners,”
summarized in Table I. Next, the diseased leaves are Phytopathology, vol. 91, pp.316-323, 2001.
considered in tenfold way to testing using Bayes’ Classifier. [7] P. J. Pinter, Jr. J. L. Hatfield, J. S. Schepers, E. M. Barnes, M. S. Moran,
C. S. T. Daughtry, and D. R. Upchurch, “Remote Sensing for Crop
Taking first 450 of the data set as training set and remaining Management,” Photogrammetric Engineering and Remote Sensing,
50 as test set the likelihood for both classes has been calculate vol. 69, pp. 647-664, 2003.
and then assign the pattern to the class, which have the [8] Z. Qin, M. Zhang, T. Christensen, W. Li, and H. Tang, “Remote
Sensing Analysis of Rice Disease Stress for Farm Pest Management
maximum likelihood value. Again the same procedure is Using Wide-band Airborne Data,” IEEE Geoscience and Remote
repeated for the first 400 and last 50 of the data set as training Sensing Symposium, vol. 4, pp. 2215-2217, 2003.
set while remaining 50 of the data set is considered as the test [9] P. Sanyal, U. Bhattacharya, and S. K. Bandyopadhyay, “Color Texture
set. Thus, 10 different combinations of training and testing Analysis of Rice Leaves Diagnosing Deficiency in the Balance of
Mineral Levels towards Improvement of Crop Productivity. IEEE 10th
data set are applied resulting classification accuracy 79.5%. International Conference on Information Technology, pp. 85-90, 2007.
SVM classification with 10 fold cross validation is carried [10] R. C. Gonzalez and R. E. Woods, In: Digital Image Processing. New
out for same data set and 68.1% classification accuracy is Delhi, India, Pearson Education., 2007.
[11] H. D. Chen, X. H. Jiang, Y. Sun, and W. J. Li, “Color Image
achieved, which has been shown in Table II. Segmentation: Advances & Prospects,” Pattern Recognition, vol. 34,
pp. 2259-2281, 2001
TABLE I: RESULT OF CLASSIFICATION OF UNINFECTED AND DISEASED [12] L. K. Huang and M. J. J. Wang “Image Thresholding by Minimizing
LEAVES the Measure of fuzziness,” Pattern Recognition , vol. 28, pp. 41-51.
Image Type Success Rate(%) 1995.
[13] J. N. Kapur, P. K. Sahoo, and Wong AKC. “A New Method for Gray
Normal Leaf Image 92 Level Picture Threshold Using the Entropy of the Histogram,”
Graphical Models and Image Processing, pp. 273-285. 1985.
Brown Spot Image 96.4 [14] N. Otsu, “A Threshold Selection Method from Gray Level
Histograms” IEEE Transaction on Systems, Man and Cybernetics, vol.
Blast Image 84
9, pp. 62-66., 1979.
[15] N. Ikonomakis, K. N. Plataniotis, and A. N. Venetsanopoulos, “A
Region Based Color Image Segmentation Scheme,” in Proc. of
TABLE II: RESULT OF CLASSIFICATION OF DISEASED LEAVES Electrical Imaging, vol. 3653, pp. 1202-1209, 1999.
Number of Success Rate [16] N. Ikonomakis, K. N. Plataniotis, and A. N. Venetsanopoulos,
Classifier
Samples (%) “Unsupervised Seed Determination for A Region-Based Color Image
Segmentation Scheme,” IEEE International Conference on Image
Bayes’ Classifier 1000 79.5 Processing, pp. 537-539, 2000.
[17] F. Y. Shih and S. Cheng, “Automatic Seeded Region Growing For
Support Vector Color Image Segmentation,” Image and Vision Computing vol. 23, pp.
1000 68.1
Machine 877-886, 2005.
[18] Y. Qixiang, G. Wen, and Z. Wei. “Color Image Segmentation Using
Density Based Clustering,” IEEE, ICASSP, vol. 3, pp. 345-348, 2003.
[19] Y. Chen and J. Z. Wang, “A Region Based Fuzzy Feature Matching to
Content Based Image Retrieval,” IEEE Transaction on Pattern
V. CONCLUSION
Analysis and Machine Intelligence, vol. 24, pp. 1252-1267, 2002.
In the work, an automated system has been developed for [20] D. P. Tuna, “Image Segmentation Using Probabilistic Fuzzy C-Means
Clustering,” International Conference on Image processing,
detecting two different types of rice diseases. In the first stage, pp.722-725, 2001.
uninfected and the diseased leaves are classified based on the [21] T. Touanf and R. C. Gonzalez, “In: Pattern Recognition Principles.
number of peaks in the histogram. Miss classification may London,” UK, Addition Wesley Publishing Company. pp. 110-1543,
occur due to shadow effect and color distortion of aging 1974.
[22] C. J. C. Burges, “A Tutorial on Support Vector Machines for Pattern
leaves. In the second level Bayes’ classifier and SVM are Recognition,” Data Mining and Knowledge Discovery vol. 2, pp. 1-47,
applied to classify the leaf diseases. Time complexity of the 1998.
[23] C. Nello and J. S. Taylor, In: An Introduction to Support Vector
Bayes’ classifier is O(N×D2) where as for the support vector
Machines and Other Kernel-based Learning, Cambridge University
machine it is O(D×N2) where D is the dimension of the Press. 2000.

463