(IJACSA) International Journal of Advanced Computer Science and Applications,

Vol. 13, No. 4, 2022

Plant Disease Detection using AI based VGG-16

Model
Anwar Abdullah Alatawi1, Shahd Maadi Alomani2, Najd Ibrahim Alhawiti3, Muhammad Ayaz4
Industrial Innovation and Robotic Center (IIRC), University of Tabuk, Tabuk 47731, Saudi Arabia 1, 2 3
Faculty of Computers and Information Technology, University of Tabuk, Tabuk, Saudi Arabia1, 2, 3
Sensor Networks and Cellular Systems (SNCS), Research Center, University of Tabuk, Tabuk, 47431, Saudi Arabia4

Abstract—Agriculture and modern farming is one of the fields occur in plants affecting their overall growth and might lead to
where IoT and automation can have a great impact. Maintaining plant death in extreme cases. Plant diseases can occur
healthy plants and monitoring their environment in order to throughout the different stages of plant development including
identify or detect diseases is essential in order to maintain a seed development, seedling, and seedling growth [3]. When
maximum crop yield. The implementation of current high diseased, plants go through different mechanical,
rocketing technologies including artificial intelligence (AI), morphological, and biochemical changes [4, 5]. Truthfully,
machine learning, and deep learning has proved to be extremely there are two main types of plant stress classified as biotic
important in modern agriculture as a method of advanced image stress represented by living creatures that interact with plants in
analysis domain. Artificial intelligence adds time efficiency and
a way that negatively affects their growth [6] such as bacteria,
the possibility of identifying plant diseases, in addition to
monitoring and controlling the environmental conditions in
viruses, or fungi [7], or abiotic stress represented by the
farms. Several studies showed that machine learning and deep collection of non-living factors or the environmental factors.
learning technologies can detect plant diseases upon analyzing Fig. 1 illustrates the collection of factors that contribute to
plant leaves with great accuracy and sensitivity. In this study, plant diseases.
considering the worth of machine learning for disease detection, Typically, the commonly used approach for farmers,
we present a convolutional neural network VGG-16 model to scientists, and even breeders, to detect and identify plant
detect plant diseases, to allow farmers to make timely actions disease was the manual inspection of plants. Of course, this
with respect to treatment without further delay. To carry this
process requires expertise and knowledge for the proper
out, 19 different classes of plants diseases were chosen, where
15,915 plant leaf images (both diseased and healthy leaves) were
detection. With time, manual inspection became tiresome and
acquired from the Plant Village dataset for training and testing. time consuming and not as quite efficient especially when large
Based on the experimental results, the proposed model is able to amounts of plants needed to be inspected. Another factor that
achieve an accuracy of about 95.2% with the testing loss being proves the inefficiency of manual inspection is the similar
only 0.4418. The proposed model provides a clear direction conditions that might be caused by different pathogens that
toward a deep learning-based plant disease detection to apply on might look alike in their effect on the plant [6].
a large scale in future.
Fungi
Keywords—Machine learning; VGG-16; disease detection;
convolutional networks; Plant Village; modern farming Bacteria
I. INTRODUCTION
Biotic Viruses
Agriculture has always been a basic human need ever since
humans’ existence as plants were a primary source of food.
Even nowadays, agriculture is still considered an essential food Phytoplasmy
resource and is the center of several aspects in humans’ lives
[1]. As a matter of fact, agriculture serves as the pillar of
Nematodes
economy in many countries regardless of their developmental Plant Disease
stages. The various domains that show the importance of Causes
agriculture include the fact that agriculture is a main source of Nutrition
livelihood where approximately 70% of the population depends
on plants and their cultivation for livelihood. This great
Moisture
percentage reflects on agriculture being the most important
resource that can actually stand a chance in the face of the Abiotic
rapidly increasing population [2]. Temperature
One of the most critical challenges that face agriculture and
affects it trade is plant diseases and how to timely detect them Toxic Chemicals
and deal with them to improve the health of crops. By
definition, plant disease in a type of natural problems that Fig. 1. Plant Disease Causes Detailed as Biotic and Abiotic Factors.

718 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 4, 2022

Fig. 3. Plant Disease Detection Phases through Deep Learning.

On the other hand, deep architectures like CNN

(Convolutional Neural Networks) have also been heavily used
in studies that are concerned with plant disease detection.
These deep architectures differ from the shallow ones by not
Fig. 2. Steps in the Implementation of Machine Learning Models in Plant requiring hand-designed features since deep learning
Disease Detection. algorithms are able to learn the features themselves. Thus, deep
learning approaches undergo three basic stages in detecting
For this reason, humans needed a better suited technique plant diseases (see Fig. 3): classification, detection, and
that can deliver effective plant detection results in less time. segmentation [14].
Technology and the rise of both deep and machine learning After SVM machine learning approach was the most
came as a benefit in various fields, most importantly medicine. commonly used one for so long, approximately after the year
Nonetheless, advanced technological approaches can also be 2015 CNN replaced SVM as the most popular ML technique
implemented for the purpose of detecting diseases in plants. for detection of diseases. CNN is considered state-of-the-art
Thus, machine learning and deep learning approaches can be model that has been used in plant disease detection nowadays,
considered non-destructive disease detection methods since especially since this task requires dealing with image data
they are based on image-processing techniques, as opposed to applications [13].
destructive serology or molecular methods [8]. However, for
these techniques to work, the disease must have already caused CNN can execute tasks such as classification of images,
a visible change in the plant, particularly in the leaf area or segmentation, object detection, and recognition. In their
stem [9]. structure, CNNs are made up of artificial neural networks
where tens and even hundreds of layers are used [15, 16].
Artificial intelligence, computer vision and machine CNNs is made up of an input layer, several convolutional
learning utilizations can greatly enhance the process of plant layers, along with pooling layers in between them, and finally
disease detection, and is already applied in multiple research full connection layer in addition to activation function layers,
papers [10]. Such technologies are capable of not only and output layer. There exist several forms of CNN
detecting the presence of a disease, but it is also possible to architectures like VGG-16, Inception-V3, ResNet50, and
determine its severity, and to classify exactly which kind of AlexNet. However, CNN architectures need large data
disease is present in a given plant sample [11]. numbers which is often considered as a challenge [17-19].
Based on their depth, the plant disease detection methods Since agriculture is essential there’s a need to provide
can be divided into shallow architectures and deep methods that enhance the agricultural methods in terms of
architectures. Basic machine learning methods like Random planting, monitoring crop environment, detecting plant disease,
Forest (RF), Support Vector Machine (SVM), Naïve Bayes and even harvesting. These important details led to significant
(NB), and K-Nearest Neighbor (KNN) rely on specific design research to be conducted and several papers to be published
intended for features such that good features and patterns must with the purpose of providing solutions to these agricultural
be recognized. These specific features include hue saturation challenges. This study proposes a model based on CNN,
value (HSV), Histogram of Oriented gradient (HOG), linear namely VGG-16 architecture in order to detect and classify a
binary pattern (LBP), and red-green-blue RGB color features total of 19 plant conditions (several crop types and diseases)
[12]. In machine learning, according to the complexity of the with the best accuracy possible.
classifier, the more data is required for its training in order to
achieve satisfactory results. The Fig. 2 illustrates the general Our contributions in this study can be summed up as
flow of the implementation of machine learning techniques in follows:
detecting plant diseases. Initially, the data is acquired and 1) Updating a large dataset based on Plant Village. The
labeled according to disease classes or healthy class. A specific dataset comprises 15 thousand images of plant leaves which
dataset is then created for the model, where the input images
are captured on the field, which means that they are
can be pre-processed before feature extraction can take place.
Machine learning algorithms are capable of recognizing the photographed within their surroundings, and thus it is efficient
changes in features upon comparison, and thus determining the in terms of not needing to isolate the plant for disease
output as diseased or healthy [13]. detection.

719 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 4, 2022

2) Implementing the proposed VGG-16 model which is an of several hidden nodes with weights that connect and the
effective convolutional neural network architecture, and it complete neural network has no iterations. Image pre-
achieves a great accuracy. Our proposed model is capable of processing is done on the images in the dataset including image
scanning through thousands of leave images in order to resizing, segmentation, color space conversion, and HSV. A
scatter plot of HSV is done on the resized image of tomato late
identify if a plant has a certain type of disease based on its leaf
blight, where a mask can be performed on the segmented
image. The proposed model achieves a great accuracy of image. 7,000 images of tomato leaves were selected for
detection among 19 different disease classes in a short period training (70%) whereas 3,000 were selected for testing (30%).
of time, and it doesn’t require a long time in training either. The features HSV histogram, Haralick textures, and Colour
The arrangement of the current paper is as follows: section Moments were selected and the labels were transformed into
two is a description of some of the published similar studies numerical values via the One Hot Encoding method. In the
about ML and DL in plant disease detection. Section three ELM model, using 1024 neurons and Sigmoid activation
describes the proposed methodology including the dataset and function yielded the best disease detection accuracy of 84.94%.
the proposed model. Results are provided and compared When compared to Decision Tree model, ELM achieves better
theoretically with some of existing techniques in section four, overall accuracies, 77.8% vs. 84.94%, respectively, in all of the
while section five concludes the article by sharing the future 10 classes. On the other hand, SVM is slightly better in terms
research intentions. of accuracy (91.43%) when compared with this ELM model.
In their paper discussing plant disease detection, Bedi et al.
II. RELATED WORK [22] proposed a detection model based on convolutional auto
A large number of studies have been performed and encoder “CAE” and CNN as two parts of a hybrid system. The
published about the detection of diseases in plants, especially system was implemented on leaf images of peach plants
since disease detection and classification can help in treating specifically, but it can also be used to spot diseases in other
the diseased plant. The more technology emerges, the more it plants. The dataset consists of two different classes healthy
will be implemented in the agricultural field, one of the reasons peach leaves (2160) and diseased leaves with bacterial spots
being the continuous enhancement of already existing machine (2297) which sums up 4457 images in total. 70% of the images
learning and deep learning methods that identify and classify in the dataset were used for training the system whereas the
plant diseases. This section describes five studies conducted in remaining 30% were kept for testing the detection
this regard. performance. After gathering the data into the dataset, the usual
pre-processing procedures are performed on the images. The
The study by Ashwin et al. [20] discussed the development
next step is the training of the CAE model and getting
of a machine learning approach for the detection of diseases in
compressed domain representations of leaf images from which,
Soybean plant in Mazandaran province in Iran. The study
a CNN model will be trained afterwards. Upon testing, new
aimed to illustrate the importance of incorporating
leaf images will be used as input and their respective
physiological features as well as the morphological features in
compressed representations will be compared for the CNN to
realizing the diseases presence in a plant. The dataset used by
determine if the leaf is healthy or not based on whether a defect
the authors contains two randomly chosen subsets one for the
is found. The CAE network was made up of 14 layers while the
healthy soybean plants (negative) and one for the diseased
hybrid system was made up of 17 total layers through
soybean plants (positive), which makes up to 2,500 images
concatenating layers of the CAE with the CNN layers. Adam
total. These images were collected from 10 different regions in
optimizer and binary cross entropy were used on the proposed
the Mazandaran province. Several features were selected based
system, where parameters such as NRMSE loss and accuracy
on one-way ANOVA F1-score which include stem length, root
measures were used for evaluation. As a result, the testing
length, number of pods per plant, empty pods per plants and
NRMSE loss was 0.0607, the testing accuracy reached 98.38%,
seed oil content. Six different machine learning techniques
and the F1-score was 98.36%.
with ten-fold cross validation and internal cross validation,
which are the binary Logistic Regression, Multilayer Jeyalakshmi et al. [23] developed an approach based on
Perceptron with Adam optimization, Random Forest, Gradient machine learning for classification of diseases of potato and
Tree Boosting (GBT), and Support Vector Machine. These ML grape crops depending on their leaf images. The dataset was
techniques are evaluated based on confusion matrix, precision, acquired from Pant Village dataset, consisting of 1000 healthy
F1-score, AUC, etc. Out of the six techniques, GBT achieved leaf images, as well as 2000 images of diseased potato leaves,
the best results of 96.79% accuracy, 96.68% F1-score, and and 3270 of diseased grape leaves. In the potato subset, there
97% sensitivity and specificity depending on 21 features. The exist three different classes, while there are four classes in the
same technique achieved similar results (96.13%) depending grape leaves subset. From each RGB leaf images, the
on 12 morphological features. background was removed through using “enhanced GrabCut
algorithm”. Several features were selected from the images
Xian et al. [21] relied on a supervised machine learning
such as the intensities of red, green and blue, weighted mean
termed Extreme Learning Machine “ELM” to determine the
and weighted standard deviation (belonging to histogram
presence or absence of tomato diseases. The developed system
features), and texture features including contrast, entropy, etc.
was based on analyzing tomato leaves; the whole dataset was
After that, the features were normalized and fed into several
acquired from Kaggle, specifically the Plant Village dataset.
classifiers, namely KNN, Naïve Bayes, and SVM. It was
The dataset comprises 10 classes, which makes a total of 1,000
observed that the number of features (ranging between 7 and
images. The model is a feed forward neural network made up

720 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 4, 2022

13) affects the accuracy of detection such that the higher the authors concluded that in binary classification, DL with
number of features, the better the accuracy. For instance, the Softsign is the best, whereas Softmax is better suited for multi-
highest accuracies were achieved by KNN at 94% considering class classifications. Table I compares between theses similar
13 features, and 96.8% by SVM similarly in potato crops. On systems.
the other hand, in grape crops, SVM was superior achieving
96.02% accuracy with respect to 13 features. From the analysis of these previous studies, a group of
limitations can be seen, as most of the studies focus on
Lamba et al. [24] published a paper discussing the detecting the diseases in one type of crops. On the other hand,
implementation of several machine learning algorithms as well logically, most farms are interested in growing more than one
as deep learning techniques to properly detect diseases in crop, especially if the farmer is using these crops for the local
crops. The images comprising the dataset are from Kaggle and market, whereas in some cases a field might be planted with
rice dataset. In fact, four different datasets are created which one specific type of crop which is grown for a huge demand
are: the rice dataset containing 120 images, the pepper dataset and can be used for exportation and national trade. However,
containing 1,994 images, potato dataset containing 2,152 the aim is to create a model that can be used by the various
images, and tomato dataset containing 16,072 images. As farmers, whether they have a small farm growing multiple
usual, the images are pre-processed accordingly, and then auto- crops, or they had large farms with more than one type of crop.
color correlogram filter is applied as an additional pre- For this reason, we chose to take into consideration a group of
processing step. After that, classification took place through plants to be analyzed and to train our model to detect their
machine learning techniques, or through deep learning methods diseases respectively. Furthermore, the second aim is to create
with various activation functions to assess the overall a reliable model which achieves very accurate results with
performances. With respect to machine learning, fifteen minimal loss and minimal false detection (both false positive
different algorithms were tested including bayes net, random and false negative), for this reason the majority of the studies
forest, and iterative optimized classifier to name a few. On the focus on CNN models for the disease detection class.
other hand, in the deep learning method, different activation Similarly, our chosen CNN architecture in VGG-16 still stands
functions were used including SoftMax, Softsign, ReLu, etc. as one of the best computer vision models. VGG provides
The evaluation of performance was done relying on confusion better results while having smaller convolutional layers since it
matrix, specificity, F1-score, and accuracy. As a matter of fact, doesn’t rely on a large number of hyper-parameters, yet it has
each of the models (ML and DL) was implemented on each several 3x3 filters with stride 1, and max layers and padding
one of the four datasets. In Pepper dataset, random forest and with stride 2 that are distributed throughout the architecture. In
DL-Softsign achieved the best results as ML and DL addition, the VGG16 model not only perform image
techniques respectively. In rice dataset, random forest and DL- recognition and detection, but it is also capable of localization,
Softmax achieved the best results. SVM and DL-Softsign were meaning that in our case and objectives, the model can be
responsible for the better results in the tomato dataset, and improved to detect the exact location of the classified disease
SVM and DL-Hard-tanh outperformed the rest in the potato on the plant leaf.
dataset. In conclusion, from the large set of experiments, the

TABLE I. COMPARISON BETWEEN SIMILAR SYSTEMS

Year of
Study Crop Type Dataset Technique Performance Results
Publication
A Machine Learning Approach to Real samples from LR-L1, LR-L2, MLP, Accuracy: 96.7%
2021 Soybean
Prediction of Soybean Disease” [20] Mazandaran province RF, GBT and SVM F1-score: 96.5%
Plant Diseases Classification using Extreme Learning Accuracy: 84.94%
2021 Tomato Plant Village
Machine Learning [21] Machine (ELM) F1-score: -
Plant diseases detection using hybrid
model based on convolutional Accuracy: 98.38%
2021 Peach Plant Village CAE and CNN
autoencoder and convolutional neural F1-score: 98.36%
network [22]
Naïve bayes, K nearest
An effective approach to feature
neighbor and support Accuracy: 96%
extraction for classification of plant 2020 Potato and grape Plant Village
vector machine F1-score: 95%
diseases using machine learning [23]
classifier
“Classification of plant diseases using Rice, pepper, Auto-color correlogram Accuracy: 99.4%
2021 Plant Village
machine and deep learning [24] potato and tomato and deep learning F1-score: 95%

721 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 4, 2022

III. IMPLEMENTATION OF THE MODEL are collected in a csv file, where each image is labeled with a
In order to maintain constant care of plants in any farm, path that indicates the health conditions of the crop within the
especially a vertical farm, you need to constantly keep a close image.
eye on the crops and their leaves as well as their stems and all
the surrounding conditions. Of course, detecting plant diseases
from looking into plant leaves manually takes a lot of time and
effort from the farmer, which is why automation could be a
game changer in the field of plant disease detection.
Computational techniques can be applied such that the images
of the plants are being taken and analyzed or screened for
diseases in a time efficient manner, without facing human
effort or human error challenges.
In this project, we implement several deep learning
technologies in order to classify the diseases in crops. Among
machine learning techniques, deep learning is an interesting
area since it can easily recognize patterns and perform
Fig. 4. Crop Disease Classes in Plant Village Dataset-1.
complicated procedures, which makes it suitable for the disease
classification task. In deep learning, deep neural networks are
created, among which is the CNN, short for Convolutional
Neural Networks, that can easily analyze images upon training.
In our project, VGG-16 was selected to be the CNN classifier
of different plant diseases.
A. Dataset
The chosen dataset for our study is the Plant Village dataset
which contains numerous images of plant diseases totaling for
15,915 images. From these images, the dataset is divided into
nineteen different classes, each class resembles a disease, and
these classes do not overlap, meaning that a single image only
belongs to one class and cannot simultaneously belong to two
or more classes.
The nineteen classes include tomato, grape, apple, tomato,
and corn diseases as well as healthy classes. The Fig. 4 and
Fig. 6 demonstrate in detail the different classes and to which
crop they belong.
The number of images per class varies between the classes, Fig. 5. Distribution of Images in Plant Village Dataset into Nineteen
as shown in the Fig. 5. After the data had been selected, they different Classes.

Fig. 6. Crop Disease Classes in Plant Village Dataset-2.

722 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 4, 2022

Fig. 7. Leaves with Diseases.

The Fig. 7 illustrates some images of the tomato late blight

class which refers to a disease in the tomato crop.
In order for the data to be convenient for our chosen model,
all the text value had to be converted into numeric values that
are suitable for the model. Moreover, the majority of dataset’s
images (80%) were utilized in the training procedure, and the
rest 20% were kept for testing.
B. Proposed Model
VGG-16 is one of the most commonly used CNN
architecture, especially since it works well with the ImageNet,
which is large project utilized for visual object recognition
procedures and it is considered one of the best models to be
proposed so far due to its extreme usefulness in the image
classification’s field in the deep learning domain. Initially, this
model was created by Karen Simonyan and Andrew Zisserman
in 2014, where they developed in during their work in Oxford
University titled “Very Deep Convolutional Networks for
Large-Scale Image Recognition”. In fact, “V” means Visual,”
G” Geometry while “G” stands for research group who
contributed in the development of this Convolutional Neural
Network model, whereas the number 16 refers to the neural
network layer’s number. ImageNet is so large that it contains
more than fourteen million images distributed over thousand
classes. This architecture is one of the top 5 models in terms of
performance achievement in the ImageNet dataset, where its (a) Bacterial Spot (b) Early blight (c) Late blight (d) Healthy (e) Leaf Mold
accuracy reached 92.7%. As an approach for the AlexNet
enhancement, this architecture was submitted to ImageNet.
(h) Two spotted (i) Yellow leaf curl
(f) Mosaic virus (g) Septoria leaf spider mite virus (j) Target spot
Large Scale Visual Recognition Challenge (ILSVRC), spot
where this model has replaced the large kernel-sized filters of Fig. 8. Training Steps.
numbers 11 and 5 in both first and second convolutional layer,
respectively by a multiple three × three kernel-sized filters Classifying the data: The proposed model is made up of
consecutively. Moreover, the training of this model was trained thirteen convolution layers, two batch normalization layers,
by utilizing NVIDIA Titan Black GPUs for several weeks. along with five max-pooling layers and three full connection
layers.
C. VGG-16 Architecture and Training Procedure
The training Procedure is made up of three consecutive The processed image passes through several convolutional
steps as shown in Fig. 8: layers that contains filters that are characterized by a receptive
field of size 3 × 3 for capturing the notions of left and right, up
 Preparing the images. and down along with the center. Despite its small size of the
mentioned filter, this filter is accompanied by the same
 Classifying the photos.
efficiency as that of a receptive field of size 7 × 7 due to its
 Printing the decision. deep characteristics such as including more nonlinearities and
lesser parameters. In addition to that, a 1x1 convolution filter
Image Processing: The input of the convents is 224 × 224 was used as an input channel’s linear transformation in a
RGB image with a fixed size where the value of each pixel is certain configuration. On the other hand, both spatial padding
subtracted from the RGB mean value of the training image.

723 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 4, 2022

and the convolution stride are fixed to 1 pixel for 3 × 3 loss function. Therefore, one of the most important loss
convolutional layers, in which the spatial resolution’s functions is the Cross Entropy Loss Function where it is used
preservation becomes easy to occur. Also, spatial pooling is for Classification model’s optimization and the complete
easier in case of a five max-pooling layers’ addition after some understanding of this loss function depends on the Softmax
of the convolutional layers and the Max-pooling layer takes activation function understanding. Moreover, in our project,
place over a 2×2-pixel window, with stride 2. the Sparse Categorical Cross Entropy is used for training our
model where it has the same loss function as that of the cross
In addition to that, a total of three varying FC (Fully entropy as in (3):
Connected) layers in depths are fixed behind a group of
convolutional layers, where the first two FC layers is made up ∑ ̂ (3)
of 4096 channels per FC layer, and the third performs 1000-
way ILSVRC classification and is made up of 1000 channels However, the truth labelling procedure is what differs
for each class. Finally, the final layer is the soft-max layer, it’s between the two loss functions, where in the case of a one hot
important to say the Fully Connected Layer’s configuration encoded true labels ([1,0,0], [0,1,0] and [0,0,1] in 3
does not vary among different networks. classification problem) the categorical cross entropy is used,
while the cross entropy is used in the case of an integer truth
The architecture of the VGG-16 model is portrayed in labels coding ([1],[2],[3]).
details in the Fig. 9.
F. Accuracy Calculation
The accuracy of the model is calculated as in (4):

(4)

IV. RESULTS
The performance of our proposed model was assessed in
terms of Loss Function and Accuracy, which was measured
during the training/validation step as well as during the testing
step. The Fig. 10 shows the performance during the training
phase.

Fig. 9. VGG-16 Detailed Architecture showing the Various Layer and their
Positioning.

D. Activation Function Used

Two activation Functions were used for our model training
where the Softmax activation and the ReLU function.
The ReLU function was used at the fully connected layers,
where the ReLU or “Rectified Linear Unit” is one of the
popular activation functions used in Neural Networks and
specifically in Convolutional Neural Networks and is defined
as in (1):
(1)
Moreover, the Softmax activation function is used for the
output layers and this activation function is a type of logistic
regression that is able of normalizing the inputted vector to a
new vector where its probability distribution is equal to 1 and it
is defined as in (2):

⃗ (2)
∑

E. Loss Functions used

In the machine learning domain, the cost functions tend to
optimize the model in the training procedure and the aim of the
training procedure is to minimize the loss function and the Fig. 10. The Performance of Proposed Model in Terms of Training and
Validation Loss and Accuracy.
model obtained is better as much as we tend to minimize this

724 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 4, 2022

In the testing phase, the proposed VGG-16 model achieves

the 93.5% accuracy, taking into consideration that the model
performs classification on 19 different classes. The
performance of our model is excellent for detecting diseases
automatically from plant images. On the test set, which is
excellent accuracy for classifying 19 items of plant images
where the Test loss has reached 0.44 while the Test accuracy
has reached 0.95. Moreover, the test accuracy was also
computed for each of the 19 cases as well as the estimated
distribution for the prediction and probability is shown in the
Fig. 11 and Fig. 12.
A. Discussion
In the light of these results, and taking into consideration all
of the parameters such as the big dataset and the variety of used
crops, we can compare our results with those of similar studies
that also used VGG networks for plant disease detection. For
instance, the study by Iftikhar Ahmad implemented the VGG-
16 model among others for the detection of diseases that infect
tomato leaves, where the VGG16 network achieved a 76.2%
accuracy by using feature extraction on augmented field
dataset of 15 thousand images, and 84% accuracy for VGG16
when using parameter tuning [25]. Another study also used
VGG and achieved an accuracy of 95.32% when using the
Plant Village dataset of 4062 grape leaf images [26]. Narayani
Patil and his colleagues also used VGG-16 networks for
detection of plant diseases, and it achieved 97.53 % accuracy
[27]. Finally, a pretrained VGG-16 model achieved an
accuracy of 94.3% compared to 96.5% for the fine-tuned
VGG-16 while using the Plant Village dataset but only for
tomato leaves [28]. Taking into consideration that in our model
we used a large dataset and multiple crops, it can be concluded
that our VGG-16 model performed better in terms of accuracy
than these studies.
Increasing agricultural movement is becoming more and
more important by the day due to the rapid increase in
population which increases the food demand, and ultimately
means the need for more crops. Maintaining crop health is a
very important field, since growing the plants in their ideal
conditions leads to better yield which comes with greater
benefits. New technologies are invested in the agricultural
domain to take care of the wellbeing of plants.

Fig. 12. Probability Estimated Distribution.

One of the domains where technology can have a large

impact is the detection of plant diseases. Machine learning and
deep learning are often implemented in systems for this
specific purpose especially since ML techniques have an
important image processing capability.
V. CONCLUSION
In this study, we proposed a system for the detection of
plant diseases through analyzing leaf images of plants to
determine not only if they are healthy or diseased, but rather to
Fig. 11. Crops Accuracy. classify which kind of disease exists in each crop type. Our

725 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 4, 2022

model is based on a VGG-16 architecture that classifies 19 [8] A. M. Abdu, M. M. Mokji, and U. U. Sheikh, "Machine learning for
classes of plant diseases, according to the data acquired from plant disease detection: an investigative comparison between support
vector machine and deep learning," IAES International Journal of
the Plant Village dataset. The model was able to achieve a Artificial Intelligence, vol. 9, pp. 670-683, 2020.
95.2% accuracy with a loss of 0.4418. [9] L. Mohamed, E. Ahmed, and A. Mohamed, "Deep learning in plant
Despite achieving a high accuracy with a low loss, our diseases detection for agricultural crops: A survey," International
Journal of Service Science, Management, Engineering, and Technology,
model faces some limitations since the input images must have vol. 11, pp. 41-58, 2020.
certain illumination conditions and a complex background [10] R. D. S. Chowdary, V. Yogeshwari, V. Varsha, M. R. B. Priya, and S.
behind them due to the fact that they are collected from actual Chinnappan, "Plant disease detection using image processing and
leaves from planted plants. These conditions pose as a machine learning algorithm," Journal of Xidian University, vol. 14, pp.
challenge for any model used for plant disease detection and 79-91, 2020.
they can be considered as areas for improvement when [11] A. Kamilaris and F. X. Prenafeta-Boldú, "Deep learning in agriculture:
designing or trying to enhance the existing model. A survey," Computers and Electronics in Agriculture, vol. 147, pp. 70-
90, 2018.
Furthermore, in the future studies, our efforts will be focused
[12] B. S. Kusumo, A. Heryana, O. Mahendra, and H. F. Pardede, "Machine
on achieving more precise disease detection, particularly learning-based for automatic detection of corn-plant diseases using
through training our machine learning model to identify the image processing," in 2018 International Conference on Computer,
exact location of the disease on each leaf, especially if more Control, Informatics and its Applications (IC3INA), Tangerang,
than one disease is detected in one plant leaf. In addition to Indonesia, 2018, pp. 93-97.
that, the plant disease dataset can be further increased to take [13] H. F. Pardede, E. Suryawati, D. Krisnandi, R. S. Yuwana, and V. Zilvan,
into consideration even more plant diseases and to incorporate "Machine learning based plant diseases detection: A review," in 2020
International Conference on Radar, Antenna, Microwave, Electronics,
additional crop types. Moreover, we can consider some and Telecommunications (ICRAMET), Tangerang, Indonesia, 2020, pp.
advanced methods to increase the accuracy of processing of 212-217.
leaf images by applying technologies like Faster region-based [14] M. Turkoglu, D. Hanbay, and A. Sengur, "Multi-model LSTM-based
convolutional neural network (Faster R-CNN), which is a convolutional neural networks for detection of apple diseases and pests,"
unified network designed for object recognition. Faster R-CNN Journal of Ambient Intelligence and Humanized Computing, 2019.
creates a network that proposes a region for the detection [15] M. Syarief and W. Setiawan, "Convolutional neural network for maize
which is then fed to the developed model for training, and after leaf disease image classification," Telkomnika (Telecommunication,
Computing, Electronics and Control), vol. 18, pp. 1376-1381, 2020.
that according to the features, the optimal detection region is
selected for classification purposes. Another method that can [16] Sharif, Z., Jung, L. T., Razzak, I., & Alazab, M. (2021). Adaptive and
Priority-based Resource Allocation for Efficient Resources Utilization in
be implemented is the You Only Look Once YOLO Mobile Edge Computing. IEEE Internet of Things Journal.
technology which presents a very fast detection in real time [17] A. Abdalla, H. Cen, L. Wan, R. Rashid, H. Weng, W. Zhou, et al.,
with approximately 45 frames for second. Another technique "Fine-tuning convolutional neural network with transfer learning for
similar to YOLO is the Single Shot Detector (SSD) which semantic segmentation of ground-level oilseed rape images in a field
provides a fast detection of objects from a single frame. SSD with high weed pressure," Computers and Electronics in Agriculture,
achieves its high accuracy by producing detection at different vol. 167, p. 105091, 2019.
scales and separates between the predictions by aspect ratio. [18] Uddin, M., et al. "Cloud-connected flying edge computing for smart
agriculture." Peer-to-Peer Networking and Applications 14.6 (2021):
REFERENCES 3405-3415.
[1] Ramsanthosh, A. S. Laxmi, S. A. Chepuri, V. Santosh, V. Kothareddy, [19] Sharif, Z., Jung, L. T., & Ayaz, M. (2022, January). Priority-based
and S. Chopra, "Review and further prospects of plant disease detection Resource Allocation Scheme for Mobile Edge Computing. In 2022 2nd
using machine learning," International Journal of Scientific Research in International Conference on Computing and Information Technology
Computer Science, Engineering and Information Technology, vol. 7, pp. (ICCIT) (pp. 138-143). IEEE.
105-115, 2021. [20] N. Ashwin, U. K. Adusumilli, N. Kemparaju, and L. Kurra, "A machine
[2] B. V. Gokulnath and G. Usha Devi, "A survey on plant disease learning approach to prediction of soybean disease," International
prediction using machine learning and deep learning techniques," Journal of Scientific Research in Science, Engineering and Technology,
Inteligencia Artificial, vol. 23, pp. 136-154, 2020. vol. 9, pp. 78-88, 2021.
[3] J. Liu and X. Wang, "Plant diseases and pests detection based on deep [21] T. S. Xian and R. Ngadiran, "Plant diseases classification using machine
learning: A review," Plant Methods, vol. 17, p. 22, 2021. learning," Journal of Physics: Conference Series, vol. 1962, p. 012024,
[4] T. Su, W. Li, P. Wang, and C. Ma, "Dynamics of peroxisome 2021.
homeostasis and its role in stress response and signaling in plants," [22] P. Bedi and P. Gole, "Plant disease detection using hybrid model based
Frontiers in Plant Science, vol. 10, p. 705, 2019. on convolutional autoencoder and convolutional neural network,"
[5] Ayaz, Muhammad, et al. "Internet-of-Things (IoT)-based smart Artificial Intelligence in Agriculture, vol. 5, pp. 90-101, 2021.
agriculture: Toward making the fields talk." IEEE access 7 (2019): [23] S. Jeyalakshmi and R. Radha, "An effective approach to feature
129551-129583. extraction for classification of plant diseases using machine learning,"
[6] S. Roy, R. Ray, S. R. Dash, and M. K. Giri, "Plant disease detection Indian Journal of Science and Technology, vol. 13, pp. 3295-3314,
using machine learning tools with an overview on dimensionality 2020.
reduction," in Data Analytics in Bioinformatics, R. Satpathy, T. [24] M. Lamba, Y. Gigras, and A. Dhull, "Classification of plant diseases
Choudhury, S. Satpathy, S. N. Mohanty, and X. Zhang, Eds. Beverly, using machine and deep learning," Open Computer Science, vol. 11, pp.
MA: Scrivener Publishing LLC, 2021, pp. 109-144. 491-508, 2021.
[7] D. B. Collinge, H. J. L. Jørgensen, M. A. C. Latz, A. Manzotti, F. Ntana, [25] I. Ahmad, M. Hamid and S. Yousaf, "Optimizing Pretrained
E. C. Rojas, et al., "Searching for novel fungal biological control agents Convolutional Neural Networks for Tomato Leaf Disease Detection,"
for plant disease control among endophytes," in Endophytes for a Complexity, 2020.
Growing World, B. R. Murphy, F. M. Doohan, M. J. Saunders, and T. R. [26] K. Akshai and J. Anitha, "Plant disease classification using deep
Hodkinson, Eds. Cambridge: Cambridge University Press, 2019, pp. 25- learning," in International Conference on Signal Processing and
51. Communication (ICPSC), 2021.

726 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 13, No. 4, 2022

[27] N. Patil, S. Kelkar and M. Ranawat, "Krushi Sahyog: Plant disease [28] H. Agh Atabay, "Deep residual learning for tomato plant leaf disease
identification andCrop recommendation using Artificial Intelligence," in identification," Journal of Theoretical and Applied Information
2nd International Conference for Emerging Technology (INCET), 2021. Technology , 2017.

727 | P a g e
www.ijacsa.thesai.org