Open International Journal of Informatics (OIJI) Vol. 10 No.

2 (2022)

A Comparative Study on Plant Disease Article history

Detection Using Machine Learning Algorithm Received:

15 Sept 2022
Muhammad Solihin Kadir*1,Syahid Anuar2
Received in revised
1,2Razak form:
Faculty of Technology and Informatics, Universiti
25 Oct 2022
Teknologi Malaysia, Kuala Lumpur
1
[email protected], Accepted:
2 19 Nov 2022
[email protected]
Published online:
15 Dec 2022

*Corresponding
author
*muhammadsolinin
@graduate.utm.my

Abstract

The crop diseases are major problem in agriculture industry that requires an accurate and fast crop
disease detection method to prevent and limiting major loss. Many researchers utilize machine
learning algorithm to achieve this solution. Majority of the solution either using traditional machine
learning algorithm or deep learning-based algorithm. For traditional machine learning algorithm,
the algorithm usually used feature extraction algorithm paired with machine learning algorithm
such as Support Vector Machine, Logistic Regression and K-Nearest Neighbor. Deep learning-
based algorithm utilize either fully connected neural network or use convolution neural network as
feature extractor and paired it with machine learning classifier. However, evaluating those
algorithms are quite difficult due to different settings in each experiment done in evaluating deep
learning-based algorithm and traditional machine learning based algorithm. The purpose of this
paper is to evaluate those algorithms with same dataset which is Plant Village dataset to give them
fair comparison in performance. The results show that both machine learning and deep learning
algorithm achieve great result with the highest accuracy achieve around 97% accuracy.
Keywords: Crop diseases, deep learning, convolutional neural network, machine learning,
Support Vector Machine, K-Nearest Neighbor

1. Introduction
Agriculture industry is facing one major problem, which is crop diseases.
Due to climate change, crop diseases become deadlier than ever which according to
Food and Agriculture Organization of the United Nations (FAO), on 2018 crop
diseases account for an estimated 10-16 percent loss of global agriculture harvest
[1]. Crop diseases, without some sort of early detection method, will greatly affect
the agriculture industry. Current detection method is simply by using naked eye,
which expensive and also unscalable method to deploy in large farm. Thus, many
researchers are using artificial intelligent and computer vision to detect agriculture
disease in agriculture goods. By using artificial intelligence, result can be produced
faster, precise, and efficient compare to the human beings that prone to the human
error.
. On the current information era, devices around the world has more
computing power than ever. Machine learning based solution, which is computably
expensive before become more viable nowadays. Thus, many researchers use this
opportunity to use machine learning algorithm to classify crop diseases. K-Nearest
Neighbour (KNN) and Support Vector Machine (SVM) is ones of the popular

_________________________________________________________
Corresponding author: [email protected] 117
Open International Journal of Informatics (OIJI) Vol. 10 No. 2 (2022)

machine learning algorithm used in crop diseases classification [2], [3]. The
machine learning algorithm usually paired with some feature extraction algorithm.
However, this algorithm come with huge drawback which is reliance on pre-
processing technique that computationally demanding processing time and required
high domain of knowledge [4], [5].
Other popular Artificial Intelligence (AI) algorithm is Convolution Neural
Network (CNN). CNN, as name suggest is inspired from biological neural network.
CNN consist of various layers, which primally convolution, pooling and fully-
connected layer. The first layer in CNN will be convolution layer, which can be
followed by another convolution layer or pooling layers. The last layer will be fully
connected layer for classification purposes. The main purpose of CNN is to create
automated and trained models which greatly reduced human intervention in
finetuning an Artificial Intelligence algorithm. These algorithms can take up
clusters of data and also employ more data points to increase the accuracy.
However, CNN requires a lot of data and a lot more of training time to have an
acceptable accuracy [6]. One of the CNN architecture, Resnet50 takes around 14
days to train on ImageNet dataset [6]. And ImageNet itself have thousands of
images in their dataset [7].
On the most literature published to overcome agriculture crop diseases, most
of the algorithm used are usually machine learning algorithm such as KNN and
SVM coupled with feature extraction algorithm such as Gray Level Co-occurrence
Matrix (GLCM) and Scale-invariant Feature Transform (SIFT) or by using deep
learning algorithm such as CNN to identify crop diseases. Both of this method has
its own advantages over the others. However, it is difficult to compare both of this
method since different paper tend to have different settings and dataset for their
experiment. Thus, they are need for study to evaluate these two different algorithms
with same settings and dataset

2.Literature Survey
AI as a tool have become a significant part for solving real problem
regarding problems related to prediction and regression. One of the branches of AI
is machine learning. One of problem that machine learning algorithm have been
applied by researchers is crop diseases classification. In image recognition field, a
great data collection is a significant part of the research. One of data collection used
by using handheld devices to take images and integrating image processing
technique to their methodology [8]. From the Table 1, this technique still leave room
for improvement. The summary of literature review for traditional machine learning
algorithm in plant disease detection are summarise in Table 1.
A part of deep learning algorithm, CNN is also used for crop disease image
classification by the researchers. CNN have no need for image segmentation and
feature extraction because of their layers have already fulfil similar role. However,
CNN usually need a huge amount of dataset for training purposes. This problem can
somehow be mitigated by using transfer learning. One of the researchers that use
this method is [9] in order to create a CNN model that can classify crop diseases.
Some other researchers are using CNN model that’s are trained on different dataset
as feature extractor and use machine learning classifier such as KNN and SVM to
classify [10]. The summary of CNN based algorithm are summarise in Table 2.

118
Open International Journal of Informatics (OIJI) Vol. 10 No. 2 (2022)

Table 1 Related work in Crop Disease Detection using Machine Learning

Algorithm
Author Title Process Dataset Result
[3] Detection of 1. Images in dataset were Various leaves Accuracy
plant leaf applied with image images of with
diseases using processing techniques. banana, beans, SVM
image 2.The datasets were lemons and 95.71%
segmentation applied with Green pixel rose were taken
and soft masking. using camera.
computing 3.Segmentation were used
techniques using GA.
4.Color co-occurrence
feature were computed.
5.Classify using Minimum
Distance Criterion and
SVM.
[11] Rice Blast 1. Images in dataset were Images of 90%
Disease converted using color paddy leaves testing
Detection and conversion. accuracy
Classification 2. K-means clustering were
Using Machine used for segmentation.
Learning 3. Extract feature such as
Algorithm Mean Value, GLCM and
Standard Deviation
4. Classifier used is
(Artificial Neural Network)
ANN
[12] Detection of 1. Pixel masking were used Plant Village 95% total
potato diseases for automatic region Dataset accuracy
using image segmentation.
segmentation 2. Feature extracted is
and multiclass colour and texture
support vector statistical analysis.
machine 3. SVM were used as
classifier.
[13] Improved 1. Color negation and Plant Village 99.25%
Segmentation conversion used for image dataset accuracy
Approach for enhancement.
Plant Disease 2. Colour thresholding and
Detection morphological operation
for image segmentation.
3. Histogram and GLCM
used for feature extraction.
4. Used fully-connected
layer classifier.

119
Open International Journal of Informatics (OIJI) Vol. 10 No. 2 (2022)

Table 2. Related work in Crop Disease Detection by using Convolution

Neural Network
Author Title Process Result Dataset
[14] Deep Learning 1. Inception V3 SVM achieve Cassava
for Image-Based CNN model is highest accuracy. Image
Cassava Disease tested with Dataset
Detection varieties of
classifiers.
[15] Basic 1. Image The CNN Cucumber
investigation on augmentation were achieve average Leaf images
a robust and performed on the accuracy of
practical plant dataset. 82.30%.
diagnostic 2. Two datasets
system were used to
trained their
custom CNN
models, which one
have bad images
and other have
good images.
[16] Unsupervised 1. In order to detect The SVM Plant Village
Convolutional plant diseases, classifiers dataset
Autoencoder- convolutional outperform the
Based Feature autoencoders is fully-connected
Learning for used for networks.
Automatic unsupervised
Detection of feature learning
Plant Diseases 2. The output are
then feed to SVM
classifier.

[17] Plant disease and 1. Pretrained CNN Resnet50 achieve Images of

pest detection models are tested highest accuracy leaves and
using deep with different of 97.86% but pest from
learning-based classifier. with SVM Turkey.
features Classifier used are classifier.
SVM, Extreme
Learning Machine
(ELM) and KNN.
[18] Experimental 1. 60 experimental GoogLeNet get Plant Village
study on crop configurations highest accuracy Dataset
disease detection were used for CNN of 99.35%
based on deep testing.
learning 2. AlexNet and
GoogLeNet CNN
model were used
for testing.

3. Experimental works and results

In this study, multiple general machine learning and convolutional neural
network algorithm were evaluated to classify plant disease. All algorithm will be

120
Open International Journal of Informatics (OIJI) Vol. 10 No. 2 (2022)

trained and tested with same dataset which is part of public plant disease dataset
named Plant Village Dataset [19]. For traditional based machine learning
algorithm, the dataset will first be segmented with green pixel masking. Then, the
feature will be extracted from the images. The feature extracted is Hu moments,
Haralick and histogram. To classify the image several classifiers will be used which
are Support Vector Machine, Logistic Regression, Linear Discriminant Analysis,
K-Neighbours, Decision Tree, Random Forests and Gaussian Naïve Bayes.
Six CNN models are also evaluated in this experiment. The CNN models
that are tested were VGG19, VGG16, Resnet50, Mobile net, InceptionResnetV2
and InceptionV3. All these models are fully connected layers with SoftMax at
Dense layers. All models were train for 40 epochs with 16 batch size. The training
will run on GPU to short the training time of the CNN models.

3.1. Dataset
Plant Village dataset is used in this experiment, which is an online image
database which consist of variety images of crop diseases. In this experiment, the
dataset used is tomato leaves images which the leaves infected by diseases. The 5
categories of images are Bacterial Spot, Early Blight, Late Blight, Leaf Mold and
also healthy tomato leaves. 160 images were uses for training and 40 for testing for
each class. The training images are split to 80:20 of train and test ratio.

Table 3. Categories in plant diseases dataset

Category Original Dataset

Training Set Testing Set
Healthy 160 40
Early Blight 160 40
Late
160 40
Blight
Bacterial
160 40
Spot
Leaf Mold 160 40
3.2. Results
This section provides the result of experiment done to compare the
performance of CNN based algorithm and traditional machine learning algorithm.
The precision, recall, accuracy and f1-score of models and classifier are tabulated
in Table 4 and Table 5. Feature used for the classifier in Table 5 are Histogram, Hu
moments and Haralick feature. Based on the result below, for CNN based algorithm
InceptionResnetV2 gain the highest accuracy of 0.95 while for the traditional
machine learning algorithm, Decision Tree classifier also got accuracy of 0.95.
Mobile net accuracy is the lowest amount CNN based algorithm with 0.92 accuracy.
The Gaussian Naïve-Bayes classifier achieve lowest accuracy of 0.73 accuracy.

121
Open International Journal of Informatics (OIJI) Vol. 10 No. 2 (2022)

Table 4 Results for CNN algorithm experiment

Model Precision Recall Accuracy F1- Score
VGG19 0.89 0.89 0.89 0.89
VGG16 0.85 0.85 0.85 0.85
Resnet50 0.92 0.92 0.92 0.91
Mobile net 0.84 0.83 0.83 0.83
InceptionResnetV2 0.96 0.96 0.95 0.95
InceptionV3 0.94 0.93 0.94 0.94

Table 5 Result for traditional machine learning method

Classifier Precision Recall Accuracy F1- Score
Logistic 0.86 0.86 0.85 0.85
Regression
Linear 0.85 0.85 0.85 0.85
Discriminant
Analysis
K-Neighbours 0.82 0.83 0.83 0.83
Classifier
Decision Tree 0.81 0.81 0.81 0.81
Classifier
Random Forest 0.96 0.96 0.95 0.95
Classifier
Gaussian Naïve- 0.74 0.73 0.73 0.70
Bayes
Support Vector 0.85 0.86 0.85 0.85
Machine

5. Conclusion
From the experiment result in Table 4 and Table 5, both traditional machine learning
algorithms and CNN algorithms get very satisfying result. Random Forest Classifier
and InceptionResnetv2 both gain 95% accuracy. Both algorithms have its own
advantages. From the literature review, some research even gains higher accuracy
of 99.35% by using CNN algorithm which also tested on Plant Village Dataset [18].
In the end, it all depend on research ability to fine tune the model or pick relevant
features to extract. From this experiment, there were some advantages found on
CNN based algorithms and traditional machine learning algorithm. CNN algorithm
takes very long time to train compared to traditional machine learning algorithm
even though the CNN algorithm were run on GPU to further speed up the training
time. Traditional machine learning algorithm however depend heavily on green
pixel masking, which is not very robust algorithm since green pixel value were put
manually from the researcher. This make green pixel masking not effective on
dataset which taken outside of the lab which have more vibrant colour of green.
For the future work, some further experiment can be done on CNN based algorithm.
Right now, the models are fully connected model. The models can use which
machine learning classifier for more variant result. For the traditional machine
learning algorithm, the feature can be tested singularly instead of concatenated

122
Open International Journal of Informatics (OIJI) Vol. 10 No. 2 (2022)

together to feed the classifier. There also other feature extraction method that are
not tested yet such as GLCM and SIFTS.

Acknowledgments
While doing this paper, I meet many people from Universiti Teknologi
Malaysia (UTM) in search of guidance to finishing this paper. I wish to express my
great gratitude to Dr.Syahid Anuar for his guidance and kind words to push me to
finish this paper. I also in debt to my family members in giving their support and
motivation to me. Lastly, I would like to show my appreciation to UTM staff who
provide me with various help to complete this paper.

References
[1] “FAO - News Article: Global body adopts new measures to stop the spread of plant pests.”
https://www.fao.org/news/story/en/item/1118322/icode/ (accessed Nov. 15, 2022).
[2] S. B. Jadhav, V. R. Udupi, and S. B. Patil, “Soybean leaf disease detection and severity measurement using
multiclass SVM and KNN classifier,” International Journal of Electrical and Computer Engineering, vol. 9, no. 5,
pp. 4077–4091, Oct. 2019.
[3] V. Singh and A. K. Misra, “Detection of plant leaf diseases using image segmentation and soft computing
techniques,” Information Processing in Agriculture, vol. 4, no. 1, pp. 41–49, Mar. 2017.
[4] J. Francis, Anto Sahaya Dhas D, and Anoop B K, “Identification of leaf diseases in pepper plants using soft
computing techniques,” Oct. 2016, pp. 168–173.
[5] S. Zhang and Z. Wang, “Cucumber disease recognition based on Global-Local Singular value decomposition,”
Neurocomputing, vol. 205, pp. 341–348, Sep. 2016.
[6] Y. You, Z. Zhang, C.-J. Hsieh, J. Demmel, and K. Keutzer, “ImageNet Training in Minutes,” Sep. 2017.
[7] J. Deng, W. Dong, R. Socher, L.-J. Li, Kai Li, and Li Fei-Fei, “ImageNet: A large-scale hierarchical image
database,” Mar. 2010, pp. 248–255.
[8] C. G. Dhaware and K. H. Wanjale, “A modern approach for plant leaf disease classification which depends on leaf
image processing,” in 2017 International Conference on Computer Communication and Informatics, ICCCI 2017,
Nov. 2017.
[9] V. K. Shrivastava, M. K. Pradhan, and M. P. Thakur, “Neural Networks for Rice Plant Disease Classification,” pp.
1023–1030, 2021.
[10] J. Lu, R. Ehsani, Y. Shi, J. Abdulridha, A. I. de Castro, and Y. Xu, “Field detection of anthracnose crown rot in
strawberry using spectroscopy technology,” Comput Electron Agric, vol. 135, pp. 289–299, Apr. 2017.
[11] S. Ramesh and D. Vydeki, “Rice blast disease detection and classification using machine learning algorithm,” in
Proceedings - 2nd International Conference on Micro-Electronics and Telecommunication Engineering, ICMETE
2018, Sep. 2018, pp. 255–259.
[12] M. Islam, A. Dinh, K. Wahid, and P. Bhowmik, “Detection of potato diseases using image segmentation and
multiclass support vector machine,” in Canadian Conference on Electrical and Computer Engineering, Jun. 2017.
[13] M. A. Rahman, M. M. Islam, G. M. S. Mahdee, and M. W. Ul Kabir, “Improved Segmentation Approach for Plant
Disease Detection,” in 1st International Conference on Advances in Science, Engineering and Robotics Technology
2019, ICASERT 2019, May 2019.
[14] A. Ramcharan, K. Baranowski, P. McCloskey, B. Ahmed, J. Legg, and D. P. Hughes, “Deep learning for image-
based cassava disease detection,” Front Plant Sci, vol. 8, Oct. 2017.
[15] E. Fujita, Y. Kawasaki, H. Uga, S. Kagiwada, and H. Iyatomi, “Basic investigation on a robust and practical plant
diagnostic system,” in Proceedings - 2016 15th IEEE International Conference on Machine Learning and
Applications, ICMLA 2016, Jan. 2017, pp. 989–992.
[16] H. F. Pardede, E. Suryawati, R. Sustika, and V. Zilvan, “Unsupervised Convolutional Autoencoder-Based Feature
Learning for Automatic Detection of Plant Diseases,” 2018 International Conference on Computer, Control,
Informatics and its Applications: Recent Challenges in Machine Learning for Computing Applications, IC3INA
2018 - Proceeding, pp. 158–162, Jan. 2019
[17] M. Türkoğlu and D. Hanbay, “Plant disease and pest detection using deep learning-based features,” Turkish Journal
of Electrical Engineering and Computer Sciences, vol. 27, no. 3, pp. 1636–1651, 2019,
[18] X. Xia, Y. Wu, Q. Lu, and C. Fan, “Experimental study on crop disease detection based on deep learning,” in IOP
Conference Series: Materials Science and Engineering, Aug. 2019, vol. 569, no. 5.
[19] D. P. Hughes and M. Salathé, “An open access repository of images on plant health to enable the development of
mobile disease diagnostics,” 2015.

123