Rice Plant Diseases Classification using a New

Convolutional Neural Network based Approach

2024 First International Conference on Electronics, Communication and Signal Processing (ICECSP) | 979-8-3503-6459-0/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICECSP61809.2024.10698072

Chandan Kumar Jha

Department of Electronics & Communication Engineering
Indian Institute of Information Technology Bhagalpur, Bihar-813210
Email: [email protected], [email protected]

Abstract—In India, agriculture is a significant source of income learning techniques have achieved a notable performance in
and essential for livelihood. The majority of India’s land is used image pattern recognition, they still encounter few limitations
to grow rice which is a fundamental component of diet. Plant including inadequate data and the requirement of segmentation
diseases have a significant impact on rice harvesting which results
in significant losses for the agricultural. Plant pathologists are and feature extraction [5]. In different agricultural images,
trying to find a precise and trustworthy method to diagnose rice segmentation of diseased regions with the same capacity is
plant diseases. In past few years, machine learning techniques difficult. It makes a challenge for a typical machine learning
have been proven to be useful in the automated categorization of approach to classify agricultural diseased images with satis-
crop diseases. Presently, deep learning is widely used for disease factory results [6]. In [4], deep convolutional neural network
detection in agriculture. In this paper, a new convolutional neural
network (CNN)-based technique is developed to identify four is used to extract features of rice leaf images. Further, these
different types of rice plant diseases: brown spot, bacterial blight, feature are utilized by SVM to distinguishes different classes
blast, and tungro. The proposed CNN architecture used in this of rice plant diseases. In agricultural pattern recognition, the
study distinguishes diseases of rice plants based on deep features development of deep learning techniques makes the classifi-
extracted from the five consecutive convolution layers with the cation approach more robust. Deep learning techniques may
activation function RELU. For experimentation of the proposed
technique, a rice plant dataset of 5932 images are used. For the be used to classify agriculturally contaminated images with-
training of the CNN model, 4152 rice plant images are used out requiring the use of prerequisite procedures like feature
while testing is performed using 890 images of four different extraction and segmentation [4].
classes. The classification technique used in this paper achieves Convolutional neural network (CNN) is the basic building
98.10% accuracy, 96.10% sensitivity, 98.72% specificity, 96.36% block of deep-learning based methods [7]. Recently, it is
precision and 96.21% F1-score. The proposed technique offers
better performance than many state of the art techniques. widely used in object recognition, image classification, and
Index Terms—Rice Plant Diseases, Deep-Learning, CNN, Clas- video classification. Plant diseases detection using CNN is the
sification, Accuracy. subject of numerous researchers over the past few years. A
hybrid model using autoencoder and CNN is proposed in [8]
I. I NTRODUCTION for plant diseases detection of any plant. In this technique,
For hundreds of millions of people worldwide, rice is a convolutional autoencoder is used to encode the input
regarded as a basic diet today. Over half of the world’s plant images data in the compressed domain representation.
population diet depends on rice that makes it a significant Further, the encoded data is utilized by the CNN architecture
commercial crop. Rice constitutes 36% of global staple food for classification. An attention mechanism based depthwise
consumption, and its demand is expected to rise in future separable neural-network (ADSNN) using MobileNet structure
[1]. In India, rice is grown on most of it’s land [2]. Failure is developed in [7]. In this technique, hyper-parameters of
of the rice crop might result in hunger. One of the most the ADSNN are tuned using Bayesian optimization (BO) that
dangerous plant diseases, rice blast, is a key factor in the improves the classification performance in categorization of
failure of rice crops [3]. Typically, laboratory testing or visual four different types of rice leaf diseases. An ensemble learning
examination are used to oversee the diagnosis of rice plant method using CNN and SVM is proposed in [4] that offers
diseases. The visual inspection takes a lot of time and is 97.96% classification accuracy to categorize four classes of
only done by an expert. Laboratory-based diagnosis of rice rice leaf diseases. Despite the fast development of CNN-based
plant diseases necessitates the use of chemical reagents and a techniques, a suitable paradigm is required for reliable and
complex method [4]. To effectively address and streamline this efficient rice plant diseases classification.
crucial issue of farmers, it is essential to develop a simple and This papers presents a new CNN architecture to classify
efficient technique. Automated rice plant diseases detection four different classes of rice leaf diseases. The proposed
techniques based on machine-learning and deep-learning can CNN architecture consists of five convolution layers with
resolve this issue. maxpooling and a classification layer to distinguish different
In past, numerous automated techniques based on image classes of images. The proposed architecture is experimented
processing and machine-learning tools have been developed using a dataset consists of 5932 images representing four
for the detection of rice plant diseases. Although machine unique rice leaf diseases: brown spot, bacterial blight, blast,

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on February 01,2025 at 07:47:29 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1: Images of rice plant diseases

TABLE I: Distribution of images in the training, validation

and testing sets

Total
Classes Training Validation Testing
Images
Brownspot 1600 1040 230 230
Bacterialblight 1584 1120 232 232
Blast 1440 990 225 225
Tungro 1308 902 203 203

and tungro. All the classes of images obtained from the

dataset are uniformly resized to 128x128x3 pixels. Further,
the resized images of the dataset are divided into three parts:
70% for training, 15% for validation, and 15% for testing of
the proposed CNN architecture. In testing, it is found that
the proposed technique offers better classification results in
comparison with the many existing techniques.

II. M ETHODOLOGY
A. Dataset
The proposed method utilizes a dataset consisting of 5932
images including four rice leaf diseases such as brown spot,
blast, bacterial blight, and tungro [9]. These images were
captured from several rice fields of Odisha, India. For image
Fig. 2: Proposed CNN architecture
acquisition, a high resolution Nikon DSLR-D5600 camera
equipped with 18 to 55 mm lens is used. The dataset contains
1584 rice leaf images of bacterial blight, 1440 of blast, 1600
B. Proposed CNN Architecture
of brownspot and 1308 of tungro. Figure 1 shows images
of four types of rice leaf diseases taken from the dataset. The CNN architecture proposed in this work is shown in
Images obtained from the dataset are resized to 128x128x3. Figure 2 using a block diagram. Images obtained from the rice
Further, the four classes of images are divided into three sets: leaf disease dataset are fed to four consecutive convolution
training, validation and testing. The training set contains 70% layers with 64 kernels of 3x3 size. The convolution layers
images of each class. The validation sets contain 15% images in this model employ the rectifier linear unit (RELU) as the
of each class while the testing sets contains 15% images of activation function. The features obtained from each convolu-
each classes of the dataset. The splitting of images among the tion layer are reduced using a maxpooling layer of size 2x2.
training, validation and testing sets is presented in Table I. Reduced features obtained from the fourth convolution layer

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on February 01,2025 at 07:47:29 UTC from IEEE Xplore. Restrictions apply.
 is again sent to a convolution layer with 128 kernels of 3x3
size. This layer is also activated using the RELU function
and deep features are reduced using maxpooling layer of 2x2
size. These features are used by a classification layer that is
made of a flatten layer, a dense layer with 256 neurons and
activated by RELU function, a dropout layer, and a dense layer
with 4 neurons and activated using the softmax function. The
classification layer produces decisions of in terms of the four
classes of rice leaf diseases.
C. Performance metrics
Classification performance of the proposed method is eval-
uated using the following performance metrics: accuracy
(Aacc ), sensitivity (Ssen ), specificity (Ssp ), precision (Pprec )
and F1 -score. These metrics are defined using the following
mathematical expressions:
TP + TN
Aacc = × 100% (1)
TP + TN + FP + FN

TP
Ssen = × 100% (2) Fig. 3: Model accuracies vs epoch plot
TP + FN

TN TABLE II: Confusion matrix

Ssp = × 100% (3)
TN + FP Actual Predicted Classes
Classes Brown Bacterial
TP Blast Tungro
Pprec = × 100% (4) Spot Blight
TP + FP Brown Spot 226 2 2 0
2TP Bacterial Blight 2 215 13 2
F1 − score = × 100% (5) Blast 1 2 222 0
2TP + FP + FN
Tungro 2 2 6 193
where TP and TN represent the occurrences of true positive,
true negative events respectively. FP , and FN indicates the TABLE III: Classification Performance
false positive and false negative events respectively.
Classification Brown Bacterial
III. R ESULTS AND DISCUSSION Blast Tungro Overall
Performance Spot Blight
The proposed classification technique is implemented using
python and Keras library with TensorFlow and Google Colab- Aacc (%) 98.99 97.45 97.3 98.65 98.10
oratory Jyupter notebook. For the training of the classification Ssen (%) 98.26 92.67 98.68 95.07 96.17
model, 100 epochs with batch size 32 have been taken and Ssp (%) 99.24 99.09 96.84 99.71 98.72
losses were optimized using adam optimizer. Figure 2 depicts
training and validation accuracies versus epoch plots. It shows Pprec (%) 97.83 97.28 91.36 98.97 96.36
that the proposed model offers training accuracy more than F1 -score (%) 98.05 94.92 94.87 96.98 96.21
90% after 20 epoch. Testing results of the proposed model
is shown in Table II using a confusion matrix that is created
using TP , TN , FP and FN events. The classification results in categories of rice plant diseases. A technique using color
terms of Aacc , Ssen , Ssp , Pprec and F1 -score is shown in Table features and SVM is developed in [11] to accurately categorize
III for the four classes of rice leaf diseases. The overall Aacc , four types of rice leaf diseases with a precision rate of 94.65%.
Ssen , Ssp , Pprec and F1 -score are 98.10%, 96.17%, 98.72%, A technique based on ADSNN with BO is developed in [7]
96.36% and 96.21% respectively. that classifies four classes of rice leaf diseases with 94.65%
Table IV shows comparison of the classification efficacy accuracy. An ensemble model using deep-learning based
of the proposed technique with previously developed recent models: Densenet121, EfficientnetB7 and Xception (DEX) is
techniques. Sethy et al. proposed a technique using CNN based developed in [12] that offers 98% classification accuracy for
deep features and SVM classifier for the four classes of rice nine classes of rice leaf diseases. In comparison of all these
leaf diseases. It offers 97.96% classification accuracy. Patidar techniques, it is found that the proposed technique offers
et al. developed a technique using residual neural network 98.10% classification accuracy which is either better than
that gives 95.83% classification accuracy for three distinct others.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on February 01,2025 at 07:47:29 UTC from IEEE Xplore. Restrictions apply.
 TABLE IV: Comparison of classification performance with [8] P. Bedi and P. Gole, “Plant disease detection using hybrid model
existing techniques based on convolutional autoencoder and convolutional neural network,”
Artificial Intelligence in Agriculture, vol. 5, pp. 90–101, 2021.
Method [9] P. K. Sethy, “Rice leaf disease image samples,” Mendeley Data, vol. 1,
Literature Aacc (%) 2020.
No. of [10] S. Patidar, A. Pandey, B. A. Shirish, and A. Sriram, “Rice plant disease
Features Classifier
classes detection and classification using deep residual learning,” in Machine
Learning, Image Processing, Network Security and Data Sciences:
Deep Second International Conference, MIND 2020, Silchar, India, July 30-
Sethy et al. [4] SVM 4 97.96 31, 2020, Proceedings, Part I 2. Springer, 2020, pp. 278–293.
features
[11] V. K. Shrivastava and M. K. Pradhan, “Rice plant disease classification
Residual neural using color features: a machine learning paradigm,” Journal of Plant
Patidar et al. [10] NA 3 95.83 Pathology, vol. 103, pp. 17–26, 2021.
network [12] M. T. Ahad, Y. Li, B. Song, and T. Bhuiyan, “Comparison of cnn-based
Color deep learning architectures for rice diseases classification,” Artificial
Shrivastava et al. [11] SVM 4 94.65 Intelligence in Agriculture, vol. 9, pp. 22–35, 2023.
features
Wang et al. [7] NA ADSNN with BO 4 94.65
Ensemble Model
Ahad et al. [12] NA 9 98
based on DEX
Proposed NA CNN 4 98.1

IV. C ONCLUSION
This paper presents a novel CNN-based classification tech-
nique for the detection of four distinct rice plant diseases:
blast, tungro, bacterial blight, and brown spot. The proposed
CNN architecture employed in this work is able to differ-
entiate between the four distinct rice leaf diseases based on
deep features obtained from the five consecutive convolution
layers with the activation function RELU. A collection of
5932 images of rice plants is used in experimentation of the
proposed approach. Out of 5932 images, 4152 images are
utilized to train the CNN model while 890 images of four
distinct classes are used for testing. The proposed technique
achieved a classification performance of 98.10% for accuracy
(Aacc ), 96.17% for sensitivity (Ssen ), 98.72% for specificity
(Ssp ), 96.36% for precision (Pprec ), and 96.21% for the F1-
score (F1 -score). In comparison to the recently developed
techniques, the classification methodology utilized in this
study achieves better performance than others.

R EFERENCES
[1] OECD/FAO, “Oecd-fao agricultural outlook 2020–2029,” Oecd, 2020.
[2] S. K. Upadhyay and A. Kumar, “A novel approach for rice plant diseases
classification with deep convolutional neural network,” International
Journal of Information Technology, pp. 1–15, 2022.
[3] W.-L. Chen, Y.-B. Lin, F.-L. Ng, C.-Y. Liu, and Y.-W. Lin, “Ricetalk:
Rice blast detection using internet of things and artificial intelligence
technologies,” IEEE Internet of Things Journal, vol. 7, no. 2, pp. 1001–
1010, 2019.
[4] P. K. Sethy, N. K. Barpanda, A. K. Rath, and S. K. Behera, “Deep feature
based rice leaf disease identification using support vector machine,”
Computers and Electronics in Agriculture, vol. 175, p. 105527, 2020.
[5] J. Chen, J. Chen, D. Zhang, Y. Sun, and Y. A. Nanehkaran, “Using
deep transfer learning for image-based plant disease identification,”
Computers and Electronics in Agriculture, vol. 173, p. 105393, 2020.
[6] Y. Lu, S. Yi, N. Zeng, Y. Liu, and Y. Zhang, “Identification of rice
diseases using deep convolutional neural networks,” Neurocomputing,
vol. 267, pp. 378–384, 2017.
[7] Y. Wang, H. Wang, and Z. Peng, “Rice diseases detection and classifi-
cation using attention based neural network and bayesian optimization,”
Expert Systems with Applications, vol. 178, p. 114770, 2021.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY JALANDAR. Downloaded on February 01,2025 at 07:47:29 UTC from IEEE Xplore. Restrictions apply.