Journal of Real-Time Image Processing

https://doi.org/10.1007/s11554-020-00987-8

SPECIAL ISSUE PAPER

A novel PCA–whale optimization‑based deep neural network model

for classification of tomato plant diseases using GPU
Thippa Reddy Gadekallu1 · Dharmendra Singh Rajput1 · M. Praveen Kumar Reddy1 · Kuruva Lakshmanna1 ·
Sweta Bhattacharya1 · Saurabh Singh2 · Alireza Jolfaei3 · Mamoun Alazab4

Received: 24 January 2020 / Accepted: 16 May 2020

© Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
The human population is growing at a very rapid scale. With this progressive growth, it is extremely important to ensure
that healthy food is available for the survival of the inhabitants of this planet. Also, the economy of developing countries is
highly dependent on agricultural production. The overall economic balance gets affected if there is a variance in the demand
and supply of food or agricultural products. Diseases in plants are a great threat to the yield of the crops thereby causing
famines and economy slow down. Our present study focuses on applying machine learning model for classifying tomato
disease image dataset to proactively take necessary steps to combat such agricultural crisis. In this work, the dataset is col-
lected from publicly available plant–village dataset. The significant features are extracted from the dataset using the hybrid-
principal component analysis–Whale optimization algorithm. Further the extracted data are fed into a deep neural network
for classification of tomato diseases. The proposed model is then evaluated with the classical machine learning techniques
to establish the superiority in terms of accuracy and loss rate metrics.

Keywords Tomato leaf disease classification · Deep neural networks · Grid search hyperparameter tuning · Principal
component analysis · Whale optimization algorithm

1 Introduction
* Alireza Jolfaei
alireza.jolfaei@mq.edu.au
Tomato (Lycopersicon esculentum) is an extremely impor-
* Mamoun Alazab tant agricultural crop having its requirement second to
alazab.m@ieee.org
potato. Our world at this present day and age has the neces-
Thippa Reddy Gadekallu sity of around 100 million tons of fresh tomatoes [5]. The
thippareddy.g@vit.ac.in
growing period of this daylength neutral crop is quite fast of
Dharmendra Singh Rajput about 90–150 days with average temperature constraint of
dharmendrasingh@vit.ac.in
18–25 °C of daylight and 10–20 °C for the night time. The
M. Praveen Kumar Reddy crop is relatively sensitive to high humidity and reduced
praveenkumarreddy@vit.ac.in
sunshine having derogatory effect on the quality of produc-
Kuruva Lakshmanna tion. If the humidity is high, it automatically makes crops
lakshman.kuruva@vit.ac.in
prone to pests, diseases and rotting. Hence, dry climatic
Sweta Bhattacharya condition is considered prudent for the production of good
sweta.bhattacharya@vit.ac.in
quality tomatoes. It is particularly essential to give immense
Saurabh Singh attention toward possibilities of pest attacks and tomato wilt
saurabh89@dongguk.edu
in the form of fungi, bacteria and viruses. The disease can
1
Vellore Institute of Technology, Vellore, Tamil Nadu, India be prevented if necessary mitigation steps are taken at an
2
Department of Industrial and System Engineering, Dongguk early stage [21]. Some of the diseases that affect the foliage
University, Seoul, South Korea and fruit of this crop are discussed in the following section.
3
Macquarie University, Sydney, Australia Early blight is a disease that affects parts of the tomato plant
4 and fruit with symptom of dark spots with concentric rings
Charles Darwin University, Casuarina, Australia

13
Vol.:(0123456789)
 Journal of Real-Time Image Processing

visible on older leaves. The leaves affected face premature blossoming of fruits being subjected to over fertilization of
deaths and fruit getting damaged as well. This disease is gen- nitrogen-based chemicals along with under watering. The
erally soil borne and develops during wet climatic condition. nutrient content gets unbalanced and often bares minimum
Maintaining a clean garden and removal of affected plants is showing up as black spots at the rear end of fruits. Removal
the basic solution to prevent contamination of the disease. of the affected fruits and regular watering is the only solu-
Gray leaf soft is a disease which affects the aged leaves of tion for the protection of plants from this disease. Finally,
the tomato plant. Small and dark spots evolve on the top and it is necessary to mention the existence of gray wall disease
bottom portions of the plant which eventually enlarge and that happens during riping of the tomato plant with symp-
progresses to a grayed brown color. The central part of such toms of grayed and brown patches inside the fruit. Plants
spots eventually collapses, and the affected leaf finally falls need to be shaded appropriately and provided with necessary
out. This disease normally occurring in warm and humid growing conditions for the prevention of this disease [33].
weather conditions results in deterioration in the fruit pro- The above-mentioned various types of crop diseases pose
duction. Removal of the affected plants and debris is the only severe threats on cultivation of food for the sustenance and
possible solution to stop spreading of the disease to a larger survival of the major species of the world. The identifica-
extent. The late blight disease is one of the most dangerous tion and prevention of these diseases require infrastructure
types which have the historical record of being responsi- and computational implementation. Machine learning [1, 2,
ble for the Irish famine. It contaminates other plants at an 4, 14, 35] has been adopted as a popular approach for the
accelerated pace with symptoms of erratic-shaped grey spots early detection of plant diseases. Majority of the traditional
on leaves with greasy surfaces. The spots have sometimes classifiers are implemented on smaller datasets depending
a white periferi during the winter season leading to leaves on hand-crafted features of images for the classification of
getting paper textured with eventual fall out. The fruits also data. However, applications of convolutional neural network
develop similar large greasy spots, and the possible solution directly on images from larger dataset have helped eliminate
is to get plants copper sprayed regularly. The Septoria leaf problems relevant to hand-crafted features. The use of visu-
spot have similar symptoms as late blight wherein the older alization methods for understanding of symptoms to localize
leaves get affected at the outset and then slowly affecting the disease has proved to be even more efficient and accurate
other parts of the plant. Copper spray in this case also is for the prediction of plant diseases [6, 34].
the best possible remedial step to be taken. On the contrary, In this work a novel PCA–WOA-based DNN model is
in case of Southern Blight, a white fungal infection type proposed to classify tomato plant disease datasets. The steps
of mold appears at the bottom end part of the stem almost involved in the present work are
touching the soil line. The interior and exterior part of the
rare end of the stems gets discolored and hence affects the • Transform the tomato image dataset using one-hot encod-
nutritional flow within the plant structure leading to a fatal ing approach.
outcome. Use of calcium, ammonia-based fertilizers and • Apply PCA for reducing the dimensions of the trans-
rotation of crops have been proven to be a plausible preven- formed data.
tive approach for this type of disease. Verticillium wilt is • Select the optimal features using WOA algorithm.
a soil-borne fungal infection in plants that has derogatory • Train the dimensionally reduced dataset using deep neu-
effect on crop cultivation including tomato plantations. The ral networks.
fungus is very difficult to completely eradicate and persists • Evaluate the performance of the proposed model and
in the soil for a prolong period where rotational crop culti- compare its performance against traditional approaches.
vation could be tactics to combat spreading of the disease.
The leaves in case of this disease get wilted in the day time
due to hot weather and alternatively recover in the night. 2 Literature review
The symptoms initially affect the aged leaves and spread to
the newer ones eventually extracting all nutrients and killing Studies have projected the importance of disease incidence
the entire plant. The removal of affected plants is the only and severity to be having significant effect on the prediction
solution to this disease. Anthracnose and Bacterial Speck are of plant diseases. The uses of smart phone-based applica-
similar types of fungal disease that rot tomatoes. The disease tion have helped to classify severity levels of diseased plant
starts spreading with small round spots on fruits which in images [3]. Different features extracted through the imple-
due course grow bigger in size. The disease is soil borne mentation of multiple extraction methods have resulted in
and any overripe tomato in close contact with the soil fall various results generated from the classifiers. The farmers
a prey to this disease. Copper spraying is as usual the only upload images of diseased plant leaves in the smart phone
way to prevent this disease. Calcium deficiency often leads application which helps to identify and predict diseases in
to the Blossom End Rot and Buckeye Rot disease during plant garden [26]. The quality of food is a major concern in

13
Journal of Real-Time Image Processing

the present world wherein food is getting exposed to numer- of the existing studies conducted which acts as the primary
ous sources of chemical and physical contamination which motivation for this work.
are severe effects on human health. Sorting of good qual- As already mentioned, one of the many challenges in han-
ity food from the bad ones is extremely essential to ensure dling huge quantity of image datasets is the number of fea-
food quality. Based on disease symptoms pertinent to marks, tures a machine learning algorithm has to deal with [36, 41].
scars and spots diseased and defective tomatoes can be easily In this work, a novel PCA–WOA algorithm is introduced to
identified and segregated. The common approach to sorting reduce the features of the tomato leaf disease image data-
has been classically manual in nature with its associated set [18]. Then, the dataset with reduced features is trained
challenges. Manual sorting requires intensive labour time, using a deep neural network. The main contributions of this
is biased and lacks accuracy and precision being prone to paper are:
human errors [12]. Computer vision-based systems have
gained immense momentum in various types of food and • Study in detail about impact of dimensionality reduction
agriculture industry for classification purposed [25]. In line techniques on image classification.
with this, similar studies have been conducted for grading • Reduce the burden of the DNN by using a novel PCA–
the quality of tomatoes using a vision system based on RGB WOA hybrid optimization technique.
values of the images. Calyx and Stalk scar detection mecha- • Improve the accuracy and reduce the loss of deep neural
nism has been used for the detection of healthy and diseased networks.
tomatoes justified by higher accuracy percentages of the
results. RBF-SVM classifiers using LAB color pixel val- Rest of the paper is organized as follows. Section 3 nar-
ues have also helped to classify tomato plant image datasets rates the information of the algorithms and methods used
yielding reasonably accurate results. It has been observed in the proposed work. The proposed classification approach
that use of only color-related features often fails to grade is discussed in Sect. 4. Experimental results are discussed
accuracy with the gradual increase in grading classes. It is in Sect. 5. Finally, Sect. 6 discusses conclusion and future
recommended to use combination of color and texture fea- work.
tures for experimental analysis and classification with bet-
ter accuracy of the results [25]. The use of inadequate and
unbalanced data for prediction of tomato diseases is also a 3 Background
major issue which requires attention. The use of refinement
filter bank framework has been used to recognize and predict As huge amount of data is being generated these days,
tomato diseases. The authors in the study have implemented machine learning algorithms [11, 19] are being extensively
a primary diagnosis unit consisting of location data of the used in several sectors. Some of the applications of machine
disease and a secondary diagnosis unit for verification of the learning include weather forecasting, stock market analysis,
generated classification results. The CNN filter bank is used prediction of natural calamities, disease prediction, sales
finally for training and classification of the datasets [7, 40]. prediction, recommendation systems, etc. [8, 9, 28, 29, 31].
Many researchers have worked on classification of tomato Deep learning (DL) is one of the methods used in machine
disease datasets using convolutional neural networks [42], learning algorithms. Deep learning has more flexible ways
deep neural networks [10, 34, 37, 38]. A summary of related to connect layers, also has more neurons count for complex
studies conducted on plant image datasets is presented in models than existing networks and has more computational
Table 1. It is evident from the existing work done on plant power to train and future extraction automatically [16, 17,
image dataset that various machine learning and deep learn- 20]. Today, the growth of DL has increased dramatically in
ing models have been implemented and performance evalu- many applications, such as speech recognition, image pro-
ations have been conducted considering multiple perspec- cessing and computer vision [22, 43]. Figure 1 demonstrates
tives. On the contrary, it is observed that very less emphasis artificial intelligence development and expansion
has been given on improvement of the quality of the dataset
and conducts efficient feature engineering to achieve more 3.1 Activation function
accurate results. It becomes pertinent to mention that reli-
ability of the predictions generated from machine learning The main purpose of the activation function is used in an
model depends a lot on features considered. Improvement of artificial neural network to introduce nonlinear proper-
the quality of dataset and extraction of significant features ties [13]. For a neural network, (Xi , Wi ) where Xi stands for
using appropriate dimensionality reduction technique play a inputs, and Wi stands for weights, and f (Xi ) is the input func-
major role in achieving results with optimum accuracy. The tion that is passed to the output of the network. Then, this
proposed model focuses on the above-mentioned limitations input function may act as input for any other layers or final

13
13
Table 1  Summary of related studies on plant image dataset
References Dataset Methods used Evaluation metrics Limitations

[34] Plant–village dataset Deep learning model, CNN Error percentage, precision, recall, accu- DL model for various illumination condi-
racy and loss tions not considered
Performance metrics not evaluated with
traditional models
[3] Plant–village dataset—50 K images of 14 Deep convolutional neural network model Full color training and validation accuracy Optimization algorithm not implemented to
crops and 26 diseases Grayscale training and validation accuracy enhance performance of existing models
Full color loss, grayscale loss Proposed model not evaluated with tradi-
tional models
[26] Cassava plant dataset from in situ of 7 K Hue histogram, histogram-oriented gradi- Color and orient fast, rotated brief (ORB) Farmers requiring to upload images through
images ent (HOG), scale invariant feature trans- smartphones
forms, speed up robust features (SURFs)
used for feature extraction
Linear support vector classifier (SVC),
KNN, extra trees used for classification
[12] Real time dataset from tomato sorting Convolutional neural network (CNN), Quantity, cost and accuracy Results confined to only dataset collected in
machine in the lab environment artificial neural network (ANN), self- lab environment
organizing map (SOM), learning vector Use of larger industrial dataset excluded
quantization (LVQ) and support vector
machine (SVM)
[25] Golden delicious apple image from Support vector machine (SVM), MLP, Recognition rate and accuracy Tuning of weight value for features in the
ESPRIT 3 and real-time data captured K-nearest neighbor (KNN) image dataset not included
through smart phone
[7] Tomato diseases and pest dataset by Dr. Convolutional neural network (CNN) False positives, precision, recall, accuracy Model is not evaluated against traditional
Alvaro Fuentes, Chonbuk National state of the art algorithms
University
[40] ImageNet-1 K dataset RexNeXt model Error percentage Model is not evaluated against traditional
state of the art algorithms
[42] Fer-2013 facial expression database, Convolutional neural network (CNN) Training time, testing time, recognition Methods to reduce network complexity not
labelled faces in the wild (LFW) dataset rate included
Journal of Real-Time Image Processing
Journal of Real-Time Image Processing

Table 3  Different types of optimizers

S. no Optimizers Function

1 Momentum 𝜔t = 𝛿𝜔t−1 + 𝜇∇𝜏(𝜃, a, b)

𝜃 = 𝜃 − 𝜔t
2 Nesterov acceler- 𝜃 = 𝜃 − 𝜔t
ated gradient 𝜔t = 𝛾𝜔t−1 + 𝜂∇𝜏(𝜃 − 𝛾𝜔t−1 )
(NAG) 𝜃 − 𝛾𝜔t−1 is the gradient of looked ahead
3 Adagrad-adaptive 𝜃t+1 = 𝜃t − 𝜂
.g
Gt +𝜖 t
√
gradient algo-
rithm

Fig. 1  Artificial intelligence development and expansion

output. Different types of activation functions are illustrated

in Table 2

3.2 Optimizers

To minimize the loss function, optimizers change the weight

parameters [23]. The loss feature serves as ground guides
telling optimizer if it moves in the right direction to enter
the valley’s edge, the global minimum. Different types of
optimizers are given in Table 3
Fig. 2  Max pooling with filter size (2 × 2) and stride = 3
3.3 Max Pooling

Max pooling in CNN [39] is used to reduce the pixel dimen- 3.4 Whale optimization algorithm (WOA)
sion of the source image as shown in figure. Each layer
comes with specification filters (3 × 3 matrix) in CNN which WOA is a metaheuristic-inspired algorithm [15, 24, 27].
is represented in Fig. 2. This algorithm is implemented using humpback whale hunt-
Average pooling is the other alternative for Max pooling. ing patterns. A whale-to-prey attack pattern is called a bub-
Max pooling algorithm is the most commonly used pooling ble-net feed mechanism, which occurs when bubbles form in
algorithm. Similarly, image flattening is used to transform a circle around the prey [32]. It consists of two stages: encir-
a two-dimensional input image representation array into a cling prey and attacking prey. The whales ’ hunting activity
single-dimensional array. is peculiar, having two strategies, one of which is upward
spirals, where whales dive 12 m down the water and start

Table 2  Different types of S. no. Type of activation function Function equation Function range
activation functions
1 Linear function g(a) = a (−∞, ∞)
2 Sigmoid function g(a) = 𝜎(a) = 1+e1 −a (0, 1)
3 Step function g(a) = 0 if a ≤ 0 (0, 1)
g(a) = 1if a ≥ 0
4 Leaky ReLU g(a) = 0.01 if a ≤ 0 (−∞, ∞)
g(a) = a if a ≥ 0
5 Hyperbolic tangent function g(a) = tanh(a) = ea −e−a
a −a
(−1, 1)
e +e
6 ReLU function g(a) = 0 if a ≤ 0 (0, ∞)
g(a) = a if a ≥ 0
7 Swish function g(a) = 2a𝜎(𝛼a) = 𝛼 = 0 if f (a) = a (−∞, ∞)
g(a) = 2a𝜎(𝛼a) = 𝛼 → ∞ if f (a) = 2max(0, a)

13
 Journal of Real-Time Image Processing

spiraling bubbles around the prey. The second technique has Each strategy is used to attack prey humpback whales. The
three separate loop capture, lobtail and loop capture phases. chance of capturing prey using these approaches is around
50%.
3.4.1 Encircling prey Shrinking encircling approach The value of “x” begins
to decrease from 2 to 0, which is directly propositional to X
The humpback whale tries to identify prey and prey’s loca- given in Eq. 3. X is an [− x, x] random number. The random
tion. The whale believes the current best option is targeted values for X may be within [− 1, 1] shown in Fig. 4.
prey and advises other search whales to change their position Spiral-updating approach In this approach, the distance
toward the target prey. Encircling prey mechanism is defined of the whale toward the targeted prey is calculated using
in Eqs. 1 and 2. Eq. 5.
L = |Y.S∗ (k) − S(k)| (1) S(k + 1) = L.emn . cos (2𝜋h) + S∗ (k) (5)
where L = |S∗ (k) − S(k)| is the distance of ith whale and the
S(k + 1) = S∗ (k) − X.L (2) targeted prey, which is best position, m is a constant which
where X, Y are the vector coefficients, k is the current itera- defines logarithmic spiral shape, and n lies in range [− 1, 1].
tion, S∗ is the best vector position, and S is the vector posi- Once the humpback whale identifies prey, it swims
tion. Coefficients X, Y are shown in Eqs. 3 and 4 where x around prey either in spiral shape or shrinking circle. The
is gradually reduced from 2 to 0, and 𝛼 is within [0, 1]. probability for choosing either spiral shape or shrinking cir-
Figure 3 shows the best solution S∗. cle is 50%. Updating the position of whales toward the best
solution is shown in Eq. 6.
X = 2x ∝ −x (3) {
S(k) − X.L if P < 0.5
S(k + 1) =
L.emn . cos (2𝜋h) + S∗ (k) if P ≥ 0.5 (6)
Y = 2.𝛿 (4)

3.4.3 Searching of prey (exploration phase)

3.4.2 Bubble‑net attacking (exploitation phase)
In this phase, the whale (search agent) searches globally
Humpback whale net attack consists of two methods. The for the best solution (prey) and changes its position to other
first solution is a diminishing, spiral-updating approach. whales. To force the search agent to move far into other
reference whales globally, X should be greater than 1 or
less than − 1. The random search is expressed in Eqs. 7
and 8. Algorithm 1 illustrates whale optimization algorithm
(WOA).

Fig. 3  Position vector of best solution Fig. 4  Possible best solutions

13
Journal of Real-Time Image Processing

L = ||Y.Srand − S|| (7) 4 Proposed work

Figure 5 depicts the working model of PCA–WOA with

S(k + 1) = Srand − XL (8)
DNN on the tomato leaf disease dataset. The dataset is col-
lected from publicly available plant–village image dataset.
The first step in the working model is encoding a tomato leaf
disease dataset with a one-hot encoding scheme to transform
the image data. One-hot encoding is a technique used to rep-
resent categorical values in a dataset into binary digits. The
reason for this transformation lies in the inability of machine
learning algorithms to work on categorical data values. The
naturally expected debate would be on the justification to
use integer coding upfront. But it is important to remember
that use of integer coding has its own limitations when it
is used on relationships of ordinal data types. This limita-
tion is addressed in one-hot encoding approach wherein the
categorical values are assessed to integers and integers are
converted to binary values. Hence, post one-hot encoding
implementation, “1” represents existent and “0” represents
nonexistent data values. This conversion helps in yielding
enhanced prediction results in comparison with the use of
categorical values.
The principal component analysis [30] is applied to the
image dataset for reducing the dimensions. The justification
behind the use of PCA is its versatility and simplicity to
be used in real time implementations. To be more explicit,
with the use of PCA, testing of the results and comparing of
the performance with and without using PCA are extremely
easy. The PCA algorithm helps to remove correlated features
in a dataset efficiently resulting in the components being
independent of one another which further reduces training
time when subjected to machine learning algorithms. Also,

Fig. 5  Proposed framework

13
 Journal of Real-Time Image Processing

since the most significant features are selected extremely 7. Evaluate the model using accuracy, loss, time complex-
efficiently, PCA reduces the chances of overfitting. The high- ity metrics.
dimensional dataset is converted into low dimensions when
PCA is used, and hence, it becomes very easy to visualize its
nature in a 2D plot and derive inferences. In order to further 5 Results and discussion
refine and elect optimal features from the reduced data, a
renowned algorithm, namely whale optimization algorithm The experimentation was carried out using the tomato dis-
(WOA), is used post-implementation of PCA. To train these ease dataset collected from plant–village dataset repository.
selected features from the tomato leaf disease dataset, a deep The dataset consisted of images of tomato leaves both dis-
neural network is used. Further, to select optimal parameters eased and healthy. The technical implementations used in
for DNN, a grid search hyperparameter tuning approach is the paper helped to identify the diseased leaves segregating
used. Then, evaluate the performance of PCA–WOA with them from the healthy ones. The attributes used in the study
DNN using metrics, accuracy, loss rate. The detailed meth- were collected from the same dataset of the plant–village
odology for the proposed approach is given below. dataset repository, and the details of tomato images in the
dataset are given in Table 4. The sample images for every
1. Load tomato leaf disease dataset into Google Colab, a disease are shown in Fig. 6. To carry out the experimenta-
GPU-based python notebook framework provided by tion, Google Colab, a GPU framework provided by Google
Google. is used, with 50 GB hard disk and 25 GB GPU-based RAM.
2. In order to transform the collected dataset, a one-hot The metrics used to evaluate the proposed model are
encoding scheme is applied to the input dataset. accuracy and loss function.
3. To reduce the burden on the DNN, PCA is applied to the Performance analysis For evaluation of the proposed
transformed dataset to eliminate less significant pixels approach, a sequential model was used to build the deep
from the images. neural network. Cross-validation was performed on a ran-
4. To select optimal pixels from the previous step, WOA is dom sample of 6000 images from the entire image dataset.
applied. 5000 images from the complete dataset were used for train-
ing the model, and the remaining 1000 images were used for
(a) Encircling Prey: The humpback whale tries to validation and testing purposes
identify prey and prey’s location. The whale The choice of the hyperparameters in DNN was computed
believes the current best option is targeted prey using the grid search method.
and advises other search whales to change their Grid search method considered the following
position toward the target prey. Encircling prey hyperparameters:
mechanism is defined in Eqs. 1 and 2.
(b) Shrinking Encircling Approach: The value of “x” • Optimizer: Adam
begins to decrease from 2 to 0, which is directly • Number of epoch: 20
proportional to X given in Eq. 3. X is an [− x, x] • Batch size: 128
random number. The random values for X may be • Number of layers: 5
within [− 1, 1] shown in Fig. 4. • Activation function at intermediate layers: Leaky ReLU
(c) Spiral-Updating Approach: In this approach, the • Activation function at output layer: sigmoid
distance of the whale toward the targeted prey is • Loss function: categorical_crossentropy
calculated using Eq. 5. Updating the position of
whales toward the best solution is shown in Eq. 6. Figure 7 illustrates the performance of the DNN without
(d) Searching of Prey (Exploration Phase): In this dimensionality reduction considering the accuracy metric.
phase, the whale (search agent) searches globally As shown in Fig. 7, the best training and testing accuracy
for the best solution (prey) and changes its posi- were observed at epoch number 15. Also, it is important
tion to other whales. To force the search agent to highlight that the model attained a training accuracy of
to move far into other reference whales globally, 99% and a testing accuracy of 86%. After 15th epoch, the
X should be greater than 1 or less than − 1. The testing accuracy started decreasing indicating the model has
random search is expressed in Eqs. 7 and 8. become over-fitted. Hence, the conclusion could be drawn
that the best training and testing accuracy would be achieved
5. Use the grid search method to select optimal hyperpa- at epoch 15. Figure 8 illustrates the performance of the DNN
rameters of DNN. without dimensionality reduction using loss percentage.
6. Using the selected hyperparameters, train the data from Also, it is evident from this figure, and the least testing loss
step 4 using DNN. was attained at epoch number 15, with a value of 5%.

13
Journal of Real-Time Image Processing

Fig. 6  Tomato leaf images [18]

13
 Journal of Real-Time Image Processing

Table 4  Tomato leaf images in the dataset

S. no Name of the disease No. of images

1 Healthy images 1591

2 Bacterial spot disease 2127
3 Early blight diseases 1000
4 Late blight disease 1908
5 Leaf mold disease 952
6 Septoria leaf spot disease 1771
7 Mosaic virus disease 373
8 Yellow leaf curl virus 5357
9 Two spotted spider mite 1676
10 Target pot 1404

Fig. 9  Accuracy with dimensionality reduction using PCA

The choice of the hyperparameters in DNN with PCA was

achieved using the grid search method.
Grid search method chose the following hyperparameters:

• Optimizer: RMSprop
• Number of epoch: 20
• Batch size: 128
• Number of layers: 5
• Activation function at intermediate layers: Leaky ReLU
• Activation function at output layer: sigmoid
• Loss function: categorical_crossentropy

Fig. 7  Accuracy without dimensionality reduction Figure 9 illustrates the performance of the DNN with dimen-
sionality reduction using PCA for accuracy measure. As
shown in Fig. 9, the best training and testing accuracy was
observed at epoch number 17. The model attained a training
accuracy of 99% and a testing accuracy of 90%. After 17th
epoch, the testing accuracy kept decreasing indicating the
model has become over-fitted. Hence, the conclusion could
be drawn that the best training and testing accuracy would be
achieved at epoch 17. Figure 10 illustrates the performance
of the DNN with dimensionality reduction using PCA using
loss percentage. Also, it is evident from this figure, and the
least testing loss was attained at epoch number 17, with a
value of 2%.

The choice of the hyperparameters in DNN with hybrid

of PCA and WOA was achieved computing the grid search
method.
Grid search method chose the following hyperparameters:
Fig. 8  Loss without dimensionality reduction
• Optimizer:RMSprop
• Number of epoch: 20
• Batch size: 128
• Number of layers: 5

13
Journal of Real-Time Image Processing

Fig. 10  Loss with dimensionality reduction using PCA Fig. 12  Loss with dimensionality reduction using PCA and WOA

Fig. 13  Time comparison of the models for the cases

Fig. 11  Accuracy with dimensionality reduction using PCA and
WOA

from this figure, and the least testing loss was attained at
• Activation function at intermediate layers: Leaky ReLU epoch number 18, with a value of 1.8.
• Activation function at output layer: sigmoid The time comparison of the models for the cases: DNN
• Loss function: categorical_crossentropy without dimensionality reduction, DNN with dimensionality
reduction using PCA and DNN with dimensionality reduc-
Figure 11 illustrates the performance of the DNN with tion using hybrid of PCA and WOA, is shown in Fig. 13. As
dimensionality reduction using hybrid of PCA and WOA for presented in the table, DNN with dimensionality reduction
accuracy measure. As shown in Fig. 11, the best training and using hybrid of PCA and WOA performed the analysis in
testing accuracy was observed at epoch number 18. Also, it lesser time for training and testing of the data.
is important to highlight that the model attained a training From the result analysis, the following points can be
accuracy of 99% and a testing accuracy of 94%. After 18th observed:
epoch, the testing accuracy kept decreasing indicating the
model has become over-fitted. Hence, the conclusion could 1. The training and testing accuracy was better with supe-
be drawn that the best training and testing accuracy would be rior performance in case of dimensionality reduction
achieved at epoch 18. Figure 12 illustrates the performance using PCA.
of the DNN with dimensionality reduction using hybrid of 2. The loss percentage has reduced significantly due to the
PCA and WOA using loss percentage. Also, it is evident use of PCA.

13
 Journal of Real-Time Image Processing

3. The training and testing accuracy performance was 5. Bosona, T., Gebresenbet, G.: Life cycle analysis of organic tomato
further enhanced when dimensionality reduction using production and supply in sweden. J. Clean. Prod. 196, 635–643
(2018)
hybrid PCA and WOA was performed. 6. Brahimi, M., Mahmoudi, S., Boukhalfa, K., Moussaoui, A.: Deep
4. The loss percentage reduced significantly due to the use interpretable architecture for plant diseases classification. arXiv
of hybrid of PCA and WOA method. preprint arXiv​:1905.13523​ (2019)
5. The time taken for training period initially got reduced 7. Fuentes, A.F., Yoon, S., Lee, J., Park, D.S.: High-performance
deep neural network-based tomato plant diseases and pests diag-
by implementation of PCA and observed further reduc- nosis system with refinement filter bank. Front. Plant Sci. 9, 1162
tion due to the use of hybrid of PCA–WOA method. (2018)
6. The final observation revealed that the performance 8. Gadekallu, T., Soni, A., Sarkar, D., Kuruva, L.: Application of
of the basic DNN has been enhanced by the proposed sentiment analysis in movie reviews. In: Rajput, D.S., Thakur,
R.S., Muzamil Basha, S. (eds.) Sentiment Analysis and Knowl-
hybrid PCA–WOA method. edge Discovery in Contemporary Business, pp. 77–90. IGI Global,
USA (2019)
9. Gadekallu, T.R., Khare, N.: Cuckoo search optimized reduction
6 Conclusion and future work and fuzzy logic classifier for heart disease and diabetes prediction.
Int. J. Fuzzy Syst. Appl. (IJFSA) 6(2), 25–42 (2017)
10. Gadekallu, T.R., Khare, N., Bhattacharya, S., Singh, S., Reddy
In this research, we have proposed hybrid PCA–whale opti- Maddikunta, P.K., Ra, I.H., Alazab, M.: Early detection of dia-
mization-based deep learning model for the classification of betic retinopathy using PCA-firefly based deep learning model.
tomato diseases. The dataset used in the study is collected Electronics 9(2), 274 (2020)
11. Garg, S., Kaur, K., Kumar, N., Rodrigues, J.J.: Hybrid deep-learn-
from the publicly available plant–village dataset. The model ing-based anomaly detection scheme for suspicious flow detection
commences with one-hot encoding technique being applied in SDN: a social multimedia perspective. IEEE Trans. Multimed.
to the dataset for transformation. This transformed datum 21(3), 566–578 (2019)
is exposed to PCA algorithm for achieving dimensionality 12. Haggag, M., Abdelhay, S., Mecheter, A., Gowid, S., Mushara-
vati, F., Ghani, S.: An intelligent hybrid experimental-based deep
reduction. The whale optimization algorithm is then applied learning algorithm for tomato-sorting controllers. IEEE Access 7,
to extract the most significant features in the dataset which is 106890–106898 (2019)
further fed into the deep neural network model. The results 13. Hayou, S., Doucet, A., Rousseau, J.: On the selection of initiali-
of this hybrid framework accurately classified tomato dis- zation and activation function for deep neural networks. arXiv
preprint arXiv​:1805.08266​ (2018)
eases which help to take proactive measures to combat risks 14. Iwendi, C., Khan, S., Anajemba, J.H., Bashir, A.K., Noor, F.:
associated with crop cultivations. The proposed model is Realizing an efficient IoMT-assisted patient diet recommenda-
then experimentally compared with the popular classical tion system through machine learning model. IEEE Access 8,
machine learning techniques, and the results highlight the 28462–28474 (2020)
15. Iwendi, C., Maddikunta, P.K.R., Gadekallu, T.R., Lakshmanna,
superiority of the model yielding higher accuracy and low K., Bashir, A.K., Piran, M.J.: A metaheuristic optimization
rate. As a future direction of work, the proposed model could approach for energy efficiency in the IoT networks. Softw. Pract.
be implemented on various other datasets for proactive pre- Exp. 1, 1–14 (2020)
dictions in various realms affecting human life. 16. Jindal, A., Aujla, G.S., Kumar, N., Chaudhary, R., Obaidat, M.S.,
You, I.: Sedative: SDN-enabled deep learning architecture for net-
work traffic control in vehicular cyber-physical systems. IEEE
Netw. 32(6), 66–73 (2018)
References 17. Jindal, A., Aujla, G.S., Kumar, N., Prodan, R., Obaidat, M.S.:
Drums: Demand response management in a smart city using deep
1. Alloghani, M., Al-Jumeily, D., Hussain, A., Mustafina, J., Baker, learning and SVR. In: 2018 IEEE Global Communications Con-
T., Aljaaf, A.J.: Implementation of machine learning and data ference (GLOBECOM), pp. 1–6. IEEE (2018)
mining to improve cybersecurity and limit vulnerabilities to cyber 18. Kaggle: Tomato Leaf Disease (2018). https​://www.kaggl​e.com/
attacks. In: Nature-Inspired Computation in Data Mining and emmar​ex/plant​disea​se. Accessed on 12 Dec 2019
Machine Learning, pp. 47–76. Springer (2020) 19. Kaljahi, M.A., Shivakumara, P., Hakak, S., Idris, M.Y.I., Anisi,
2. Alloghani, M., Baker, T., Al-Jumeily, D., Hussain, A., Mustafina, M.H., Rajan, D.: Saliency-based bit plane detection for network
J., Aljaaf, A.J.: Prospects of machine and deep learning in analy- applications. Multimedia Tools Appl. 1, 1–19 (2020)
sis of vital signs for the improvement of healthcare services. In: 20. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. Nature
Nature-Inspired Computation in Data Mining and Machine Learn- 521(7553), 436–444 (2015)
ing, pp. 113–136. Springer (2020) 21. Lu, J., Ehsani, R., Shi, Y., de Castro, A.I., Wang, S.: Detection of
3. Ashqar, B.A., Abu-Naser, S.S.: Image-based tomato leaves dis- multi-tomato leaf diseases (late blight, target and bacterial spots)
eases detection using deep learning. Int. J. Acad. Eng. Res. 2(12), in different stages by using a spectral-based sensor. Sci. Rep. 8(1),
10–16 (2018) 2793 (2018)
4. Bhattacharya, S., Kaluri, R., Singh, S., Alazab, M., Tariq, U., 22. Mahmud, M., Kaiser, M.S., Hussain, A., Vassanelli, S.: Applica-
et al.: A novel PCA-firefly based xgboost classification model for tions of deep learning and reinforcement learning to biological
intrusion detection in networks using GPU. Electronics 9(2), 219 data. IEEE Trans. Neural Netw. Learn. Syst. 29(6), 2063–2079
(2020) (2018)
23. Mendoza, H., Klein, A., Feurer, M., Springenberg, J.T., Urban,
M., Burkart, M., Dippel, M., Lindauer, M., Hutter, F.: Towards

13
Journal of Real-Time Image Processing

automatically-tuned deep neural networks. In: Automated 42. Zhang, H., Jolfaei, A., Alazab, M.: A face emotion recognition
Machine Learning, pp. 135–149. Springer (2019) method using convolutional neural network and image edge com-
24. Mirjalili, S., Lewis, A.: The whale optimization algorithm. Adv. puting. IEEE Access 7, 159081–159089 (2019)
Eng. Softw. 95, 51–67 (2016) 43. Zhou, X., Gong, W., Fu, W., Du, F.: Application of deep learn-
25. Moallem, P., Serajoddin, A., Pourghassem, H.: Computer vision- ing in object detection. In: 2017 IEEE/ACIS 16th International
based apple grading for golden delicious apples based on surface Conference on Computer and Information Science (ICIS), pp.
features. Inf. Process. Agric. 4(1), 33–40 (2017) 631–634. IEEE (2017)
26. Mwebaze, E., Owomugisha, G.: Machine learning for plant dis-
ease incidence and severity measurements from leaf images. In: Publisher’s Note Springer Nature remains neutral with regard to
2016 15th IEEE International Conference on Machine Learning jurisdictional claims in published maps and institutional affiliations.
and Applications (ICMLA), pp. 158–163. IEEE (2016)
27. Pham, Q.V., Mirjalili, S., Kumar, N., Alazab, M., Hwang, W.J.:
Whale optimization algorithm with applications to resource
allocation in wireless networks. IEEE Trans. Veh. Technol. 69, Dr. Thippa Reddy Gadekallu is
4285–4297 (2020) currently working as Assistant
28. Reddy, G.T., Khare, N.: Hybrid firefly-bat optimized fuzzy artifi- Professor (Senior) in School of
cial neural network based classifier for diabetes diagnosis. Int. J. Information Technology and
Intell. Eng. Syst. 10(4), 18–27 (2017) Engineering, VIT, Vellore,
29. Reddy, G.T., Khare, N.: Heart disease classification system using Tamil Nadu, India. He obtained
optimised fuzzy rule based algorithm. Int. J. Biomed. Eng. Tech- his Bachelor of Technology
nol. 27(3), 183–202 (2018) degree in Computer Science and
30. Reddy, G.T., Reddy, M.P.K., Lakshmanna, K., Kaluri, R., Rajput, Engineering from Nagarjuna
D.S., Srivastava, G., Baker, T.: Analysis of dimensionality reduc- University, Andhra Pradesh,
tion techniques on big data. IEEE Access 8, 54776–54788 (2020) India, Master of Engineering in
31. Reddy, G.T., Reddy, M.P.K., Lakshmanna, K., Rajput, D.S., Computer Science and Engineer-
Kaluri, R., Srivastava, G.: Hybrid genetic algorithm and a fuzzy ing from Anna University, Chen-
logic classifier for heart disease diagnosis. Evolut. Intell. 1, 1–12 nai, Tamil Nadu, India, and com-
(2019) pleted his PhD in Vellore
32. Reddy, M.P.K., Babu, M.R.: Implementing self adaptiveness in Institute of Technology, Vellore,
whale optimization for cluster head section in internet of things. Tamil Nadu, India. He has 14 years of experience in teaching. He pro-
Cluster Comput. 22(1), 1361–1372 (2019) duced more than 25 international/national publications. Currently, his
33. Sabrol, H., Satish, K.: Tomato plant disease classification in digi- research interests include machine learning, deep learning, computer
tal images using classification tree. In: 2016 International Con- vision, big data analytics, blockchain.
ference on Communication and Signal Processing (ICCSP), pp.
1242–1246. IEEE (2016) Dr. Dharmendra Singh Rajput has
34. Saleem, M.H., Potgieter, J., Mahmood Arif, K.: Plant disease received his PhD in the year
detection and classification by deep learning. Plants 8(11), 468 2013 from NIT Bhopal, India.
(2019) He is currently working as an
35. Uddin, Z., Altaf, M., Bilal, M., Nkenyereye, L., Bashir, A.K.: Associate Professor in VIT,
Amateur drones detection: a machine learning approach utilizing India. His research areas are data
the acoustic signals in the presence of strong interference. Com- mining, artificial intelligence,
put. Commun. 154, 236–245 (2020) soft computing, automata, natu-
36. Venkatraman, S., Alazab, M., Vinayakumar, R.: A hybrid deep ral language processing.
learning image-based analysis for effective malware detection. J.
Inf. Secur. Appl. 47, 377–389 (2019)
37. Verma, S., Chug, A., Singh, A.P.: Recent advancements in image-
based prediction models for diagnosis of plant diseases. In: Pro-
ceedings of 3rd International Conference on Computer Vision and
Image Processing, pp. 365–377. Springer (2020)
38. Vinayakumar, R., Alazab, M., Srinivasan, S., Pham, Q.V., Padan-
nayil, S.K., Simran, K.: A visualized botnet detection system
based deep learning for the internet of things networks of smart
cities. IEEE Trans. Ind. Appl. 1, 1–8 (2020) Dr. M. Praveen Kumar Reddy is
39. Wang, S.H., Phillips, P., Sui, Y., Liu, B., Yang, M., Cheng, H.: currently working as Assistant
Classification of alzheimer’s disease based on eight-layer convo- Professor in School of Informa-
lutional neural network with leaky rectified linear unit and max tion Technology and Engineer-
pooling. J. Med. Syst. 42(5), 85 (2018) ing, VIT, Vellore, Tamil Nadu,
40. Xie, S., Girshick, R., Dollár, P., Tu, Z., He, K.: Aggregated resid- India. He was a visiting profes-
ual transformations for deep neural networks. In: Proceedings of sor in Guangdong University of
the IEEE Conference on Computer Vision and Pattern Recogni- Technology, China, in 2019. He
tion, pp. 1492–1500 (2017) had worked as Software Devel-
41. Zerdoumi, S., Sabri, A.Q.M., Kamsin, A., Hashem, I.A.T., Gani, oper in IBM in 2011. He worked
A., Hakak, S., Al-Garadi, M.A., Chang, V.: Image pattern recogni- in Alcatel-Lucent in 2013. He
tion in big data: taxonomy and open challenges: survey. Multime- obtained his B.Tech. in CSE
dia Tools Appl. 77(8), 10091–10121 (2018) from JNT University, A.P, and

13
 Journal of Real-Time Image Processing

M.Tech. in C.S.E. from VIT, Vellore, Tamil Nadu, India, and PhD from Master’s degree in Information Security from Thapar University. He
VIT, Vellore, Tamil Nadu, India. He produced more than 15 interna- holds a strong academic record. He has published the paper in top
tional/national publications. Currently, Dr. Praveen Kumar Reddy is journals such as IEEE Access, IEEE Magazine, IEEE IoT, JNCA, JoS
working in the area of energy aware applications for Internet of things and Sustainability. His research interests include cloud security, IoT,
(IoT), machine learning and high-performance computing. SNS, smart home, blockchain, wireless sensor network and cryptogra-
phy. He is also a reviewer of IEEE IoT, Computer and Electrical Engi-
Dr. Kuruva Lakshmanna has neering, IJCS, FGCS, Journal of Human-Centric Computing and Infor-
received his B.Tech. in Com- mation Science (HCIS) and many other journals.
puter Science and Engineering
from Sri Venkateswara Univer- Alireza Jolfaei received the PhD
sity College of Engineering, degree in applied cryptography
Tirupathi, India, in the year from Griffith University, Gold
2006, M.Tech. in Computer Sci- Coast, Australia. He is currently
ence and Engineering (Informa- an Assistant Professor in cyber-
tion Security) from National security with Macquarie Univer-
Institute of Technology Calicut, sity, Sydney, Australia. He has
Kerala, India, in the Year 2009, authored more than 50 peer-
and PhD from VIT, India, in the reviewed articles on topics
year of 2017. He is working as related to cybersecurity. His cur-
an Assistant professor senior in rent research interests include
VIT, India. He was a visiting cybersecurity, the IoT security,
professor in Guangdong Univer- human-in-the-loop CPS security,
sity of Technology, China, in 2018. His research interests are data cryptography, AI and machine
mining in DNA sequences, algorithms, knowledge mining, etc. learning for cybersecurity. He
received the prestigious IEEE
Dr. Sweta Bhattacharya is cur- Australian Council Award for his research article published in the IEEE
rently associated with Vellore Transactions on Information Forensics and Security. He has served
Institute of Technology (Univer- more than ten conferences in leadership capacities, including the Pro-
sity), as an Assistant Professor in gram Co-Chair, Track Chair, Session Chair and Technical Program
the School of Information Tech- Committee Member, including IEEE TrustCom. He has served as a
nology and Engineering. She has Guest Associate Editor for IEEE journals and transactions, including
received her PhD degree from the IEEE Internet of Things Journal, the IEEE Transactions on Indus-
Vellore Institute of Technology trial Applications and the IEEE Transactions on Intelligent Transporta-
and Master’s degree in Industrial tion Systems.
and Systems Engineering from
State University of New York, Dr. Mamoun Alazab is Associate
Binghamton, USA. She has Professor at the College of Engi-
guided various UG and PG pro- neering, IT and Environment at
jects and published peer Charles Darwin University, Aus-
reviewed research articles. She is tralia. He received his PhD
also a member of the Computer degree in Computer Science
Society of India and Indian Science Congress. Her research experience from the Federation University
includes working on Pill Dispensing Robotic Projects, as a fully funded of Australia, School of Science,
Watson Research Scholar at Innovation Associates, Binghamton at Information Technology and
SUNY. She has completed six sigma green belt certification from Dart- Engineering. He is a cybersecu-
mouth College, Hanover. Her research interests include applications of rity researcher and practitioner
machine learning algorithm, data mining, simulation and modeling, with industry and academic
applied statistics, quality assurance and project management. experience. Alazab’s research is
multidisciplinary that focuses on
Dr. Saurabh Singh is a Post doc- cybersecurity and digital foren-
torate research fellow from sics of computer systems with a
School of Computer, Informa- focus on cybercrime detection and prevention including cyberterrorism
tion and Communication Engi- and cyberwarfare. He has more than 100 research papers. He delivered
neering, Kunsan National Uni- many invited and keynote speeches, 22 events in 2018 alone. He con-
versity South Korea. He received vened and chaired more than 50 conferences and workshops. He works
PhD degree in Division of Com- closely with government and industry on many projects, including
puter Science and Engineering Northern Territory (NT) Department of Information and Corporate
from Chonbuk National Univer- Services, IBM, Trend Micro, the Australian Federal Police (AFP), the
sity and done his research in the Australian Communications and Media Authority (ACMA), Westpac,
field of ubiquitous computing United Nations Office on Drugs and Crime (UNODC) and the Attorney
security. He received his Bache- General’s Department. He is a Senior Member of the IEEE. He is the
lor’s degree from Utter Pradesh Founder and Chair of the IEEE Northern Territory (NT) Subsection.
Technical University and holds a

13