Computers and Electrical Engineering 103 (2022) 108357

Contents lists available at ScienceDirect

Computers and Electrical Engineering

journal homepage: www.elsevier.com/locate/compeleceng

Computer aided agriculture development for crop disease

detection by segmentation and classification using deep
learning architectures
Nirmal Raj a, Senthil Perumal b, Sanjay Singla c, Girish Kumar Sharma d,
Shamimul Qamar e, *, A. Prabhu Chakkaravarthy f
a
School of Electrical and Electronics, Sathyabama Institute of Science and Technology, Chennai, India
b
SRM Institute of Science & Technology, Ramapuram, Chennai, India
c
UIE CSE Chandigarh University, Punjab, India
d
Department of Computer Applications, Bhai Parmanand DSEU Shakarpur Campus-II (A Constituent Campus of Delhi Skill and Entrepreneurship
University) GNCT of Delhi, Delhi 110092, India
e
Computer Science & Engineering Department, College of Sciences & Arts, Dhahran Al Janoub Campus 64261 King Khalid University, Abha,
Kingdom of Saudi Arabia, (KSA)
f
Department of Computer Science and Engineering, Saveetha School of Engineering, Saveetha Institute of Medical and Technical Sciences, Chennai
602 105, India

A R T I C L E I N F O A B S T R A C T

Keywords: The growth of the Indian economy is mainly based on agricultural production. A method for crop
Disease detection disease detection based on pre-processing and segmentation processes using filtering and neural
Neural network network techniques is proposed. The dataset here has been collected based on the pre-historic
Pre-processing
cultivation data and disease-affected data of the crop. Live images from the field have been
Convoluted Gaussian filtering
collected and the dataset has been created. This data has been initially processed using a pre-
DCACNN
Loss functions processing technique based on convoluted Gaussian filtering. Then the processed image has
been segmented using a deep active contour convolutional neural network (DACCNN) to
formulate new loss functions which incorporate the region and information about size in the
disease detection while training. From the results of the experiment, the proposed method is a
vigorous method for crop disease detection and also segments main diseases of plant leaves like
Cercospora Leaf Spot, Bacterial Blight, Powdery Mildew, and Rust.

1. Introduction

Production and quality of agricultural products are reduced due to diseases. Therefore, earlier diagnosis of diseases in plants is
more important for curing and controlling them. The best seeds of plants are taken, and a perfect environment is delivered that is
suitable for the plant’s growth. Unfortunately, many diseases affect the plants [1]. Earlier detection of diseases in plants is very
important in agriculture for minimising the damage, decreasing the cost of production, and increasing income. The human eye does not
only identify diseases. Many years ago, diseases were only observed with the naked eye, with examples of pretentious plants or people
with disease prediction skill observing the farm, and farmers took conservative measurements based on their suggestions [2]. A skilled

* Corresponding author.
E-mail address: [email protected] (S. Qamar).

https://doi.org/10.1016/j.compeleceng.2022.108357
Received 14 April 2022; Received in revised form 25 August 2022; Accepted 26 August 2022
Available online 19 September 2022
0045-7906/© 2022 Published by Elsevier Ltd.
N. Raj et al. Computers and Electrical Engineering 103 (2022) 108357

person’s identification is challenging, and their suggestions will not often cure the diseases. A huge period is required, and costs are
very high due to the need for experts. Any country’s roots in agriculture means that disease identification in agricultural products is
more important [3]. Thus, few accurate, fast, automatic, and low-cost approaches are used for detecting diseases. The latest ad­
vancements in image processing and machine learning (ML) technology will provide farmers with economic knowledge of pesticides.
In the products of agriculture, diseases are caused mainly by two factors: non-living and living agents. Living agents are viruses, fungi,
bacteria, and insects. Non-living agents are temperature changes, increased moisture, light insufficiency, decreased nutrients, and air
pollution. Identifying leaves, detecting leaf diseases and fruits, etc., are the diseases for which some agricultural applications are
established. The digital camera captures digital images that are required by these applications. The information needed for examining
diseases is extracted by the images captured, which are applied with image processing and machine learning techniques [4]. The
utilisation of crops can be classified into four types: cash, food, plantation, and horticulture. Plants get affected by two major diseases:
biotic and abiotic. Fungi, bacteria, and viruses in plants are caused by biotic diseases [5], whereas abiotic can cause plants in terms of
weather conditions, chemicals, etc. Leaves of different plants bear different diseases that must be identified with the support of colour,
texture, and shape. The histogram technique was used on paddy leaves based on colour intensity to identify the infected regions. The
disease detection mechanism involves several phases. It consists of two major parts, such as training and testing. The training part
begins with the collection of images from the stem, root, leaves, etc. These images are pre-processed by evacuating the blur effect, the
noise effect, and even correcting the RGB/grey level [6]. The segmentation phase removes the background image from the ROI and
detects the affected part during training. The feature extraction phase is utilised for extracting the features and producing feature
vectors. These feature vectors are utilised for training the classifier. In the training part, the test image goes through all the phases and
is recognised as either infected or healthy by the trained classifier. The effectiveness and compatibility of the model are evaluated using
performance metrics. It is also called the recognition rate or success rate. These rates depend upon the comparison of a model, type of
classifier, techniques used, and accuracy of recognition from one to another [7].
The contribution of this paper is as follows:

• To create an ensemble model using pre-processing and segmentation techniques.

• The pre-historic cultivation data, disease-affected data of the crop, and live images from the field have been collected, and the
dataset has been created.
• To develop a pre-processing and segmentation process using convoluted Gaussian filtering and a deep convoluted active contour
neural network (DCACNN) for the detection of crop diseases from the leaves.

This paper’s organisation is followed as Section 2 exhibits related works on crops. Section 3 describes the proposed crop disease
detection design. The experimental results are depicted in Section 4. The conclusion is presented in Section 5.

2. Related works

This session explains how to use an image segmentation technique to detect and classify different forms of leaf diseases in plants.
According to an article [8], the grape plant is home to 80% to 85% of illnesses. Grape productivity and output are gradually declining.
For picture segmentation, the study’s authors employed the K-means cluster technique and neural networks; they used neural networks
for illness classification. According to a study [9], there are four processes for identifying plant leaves and classifying them. This is how
the phases are. (1). The pre-processing step of the image (2) K-means clustering: image segmentation through the use of a clustering
technique. (3) Feature extraction (4) Disease Classification Statistical A Grey-level co-occurrence matrix is used to extract features from
photos. The categorization is done with the help of a Support Vector Machine (SVM). The design technique in the paper [10] is divided
into five parts. Some phases include image capture, pre-processing, classification, extraction of features, and segmentation. This
approach employs a colour-based and cluster-based segmentation procedure and Support Vector Machine (SVM) tools for classifi­
cation. Cotton is one of India’s most significant crops, according to the paper [11], and most illnesses affect cotton plant leaves only in
their early stages. The colour-based approach is used to segment pictures to identify the discovered region for identifying the many
forms of illness that affect cotton plants. The characteristics are extracted using a grey-level co-occurrences matrix. The SVM classifier
is used to categorise the illness. The author of [12] proposes that image processing techniques such as pre-processing pictures, image
segmentation, image feature extraction, and image classification might help identify chilli plant illness sooner. Early disease detection
and image processing techniques are economical and simple for farmers. The real-time edge detection of rubber plant diseases is shown
in the article [13]. They employed the Sobel edge detection technique to detect the edges of plant leaf diseases.
Production losses arise due to the critical problems of diseases occurring in plants in farming, as described by Sharath et al. (2019).
It also affects farming production. Plant health monitoring is very difficult to monitor and classify the various infections manually. For
this purpose, experts are required, and this process will consume a lot of time. Several stages for detecting the infection are incor­
porated into this method. Depending on the obtained output, plants that are affected by the disease are monitored using these changes.
Recognition of plant infection is implemented with techniques that use infectious plant images [14]. Through image processing,
various plant diseases are characterised by Sandhu et al. (2019). In recent years, several researchers have been attracted to this method
of plant disease detection. These algorithms classify normal plant leaves as infected plant leaves.
Moreover, various problems arise in this method. The problems are The images captured without light are present in the recog­
nition system with the conversion of these images into digital form and environmental conditions. The accuracy and capability of these
recognition approaches to detect disease are concluded. The limitations present in this method require more future research. Farming
is a major food source and is not a technique, as stated by Devaraj et al. in [15]. Agriculture is the main source of income for 70% of the

2
N. Raj et al. Computers and Electrical Engineering 103 (2022) 108357

population in Asian countries. But, various kinds of diseases decrease the quality of the crop. This efficient method for disease detection
will prevent farming losses. Classification and disease detection are performed by software developed in this work. Various phrases are
incorporated into the process of disease detection. So, the implementation of techniques in image processing makes detecting and
classifying disorders in leaves easier in the farming sector.

3. System design

Crop disease detection in the leaves is the major motivation of this work. An ensemble model based on pre-processing and seg­
mentation processes using convoluted Gaussian filtering and a deep convoluted active contour neural network (DCACNN) for the
detection of crop diseases from the leaves is proposed. Images are crop disease benchmark datasets of tomato and brinjal leaves. They
are used to perform experiments. The live dataset has been gathered, and the prehistoric dataset has been collected from the cultivation
land. 1000 Healthy Leaves images are present in this dataset and crop disease leaves. The architecture for crop disease detection using
convoluted Gaussian filtering and deep convoluted active contour neural network (DCACNN) has been given in Fig. 1.

3.1. Pre-processing using convoluted Gaussian filtering

The primary stage eliminates image noise. Grayscale conversion of the input picture is performed, and the filtering method in the
pre-processing stage is used for removing noise. Here the proposed model uses convoluted Gaussian filtering.
Gaussian filters are a class of convolution filters having the weights selected based on the Gaussian function shape. The best filter to
remove noise from the Gaussian distribution is Gaussian smoothing. The 1D zero-mean Gaussian function is as follows:
x2
g(x) = e− 2σ2 (1)
In 3-D, an isotropic Gaussian (i.e., circularly symmetric) with the expression
1 − x2 +y2 +z2
G(x, y, z) = e 2σ2 (2)
2πσ 2
The Gaussian filtering utilises this convolution used to achieve this. A group of discrete pixels performs image storage, so generating
a discrete approximation to the Gaussian function before performing convolution is needed. Theoretically, a non-zero Gaussian dis­
tribution is adopted, which needs an infinitely small standard deviation and mean, and at this point, the kernel is truncated. An
appropriate integer-valued convolution kernel approximates Gaussian with a of 1.0. The Gaussian is approximated by selecting the
obvious mask values. In this mask, a pixel centre with a Gaussian value is used, and it is inaccurate due to the non-linear variation of the
Gaussian value across the pixel. Over the entire pixel (summation of Gaussian at 0.001 increments), the Gaussian value is integrated.
Integrals are not in the form of integers: A corner value of 1 is obtained from array rescaling. Finally, all the values summed in the mask
are 273.

Fig. 1. Architecture for crop disease detection using 2DCNN.

3
N. Raj et al. Computers and Electrical Engineering 103 (2022) 108357

Noise is added and filtered by utilising the add noise built-in function, Gaussian Blur, and built-in function from Image J. The
"Gaussian Blur" includes the kernel convolution, defined by a Gaussian function, with the image pixels. The discrete case of convolution
is expressed as
∑∞
f ∗ g[n]=def m=− ∞
f [m] ⋅ g[n − m] (3)

The 2D Gaussian function generates the kernel. In the following function, the amplitude is defined by A, the centre is (x0, y0), and
the standard deviations in the x and y directions are σx, σy:
( )
2
(x− x0 ) (x− y0 )2
− +
2σx2 2σy2
f (x, y) = A ⋅ e (4)
In image processing, the convolution kernel is used to approximate the Gaussian distribution. Therefore, the convolutional matrix is
constructed by values from the distribution and then put into the original image. So, the heaviest weight (which has the highest
Gaussian value) is received by the value of the original pixel, and smaller weights are received by neighbouring pixels when increasing
the original pixel distance.
Image borders are smoothened by a low pass filter known as Gaussian kernel convolution. The Gaussian distribution’s variation is
the filter defined by parameter, and filtering results are dramatically affected. This work is mainly focused on Gaussian blur effects.
Results are obtained based on the filtered image comparison with the original image by the quality factor and are expressed as

σ f ,g Tg σ 2g ⋅ σ2f
Q(f , g) = 2
⋅ 2 ⋅ ⋅ 2 ⋅ (5)
σ2g σf f 2 + g− 2 σ 2f + σ2g

Where the gold standard (original image with no noise) is represented as g and the filtered image is represented by f. σf,gis the
covariance among the two images, σ 2f is the image f ’s variance, σ2g is the image g’s variance, and image f mean is represented by f and
image g mean represented by_. Covariance among the two images is compared by this quality factor, the luminance distortion (the
mean values) and the contrast distortion (the variance values).
The indicator we used to evaluate the noise level in the images is the SNR. The SNR evaluates the image’s noise level and is
represented as:
( )
σsignal
SNR = 20.log (6)
σnoise
Basic SNR compares signal intensity and noise intensity. The quality of the image is improved by increasing SNR.

Fig. 2. ACNN is a framework training of an ACM which is automatically differentiable.

4
N. Raj et al. Computers and Electrical Engineering 103 (2022) 108357

This filtering method removes noise background so that image data is improved and, unlike distortion, is also suppressed. Ex­
amination and processing features of images are enhanced. In the RGB format, images are stored, and the standard size is resized.

3.2. Segmentation using deep active contour convolution neural network (DACCNN)

A technique for standard image examination is the Active Contour Model (ACM), in which several variants are concerned with a
massive quantity of research across numerous fields. The Eulerian functional energy of ACM with single-pixel parameter maps detected
by ACM is initialised by it. Importantly, Tensor Flow is used to fully implement both the components of CNN and ACM, and the whole
architecture of ACNN is differentiable automatically and with no user intervention, as shown in Fig. 2.
Level Set Active Contours:
Let a closed time-varying contour(t)= {(x, y)∣ϕ(x, y, t) = 0} denoted in Ω ∈ R2 by the signed distance map’s zero level set φ(x, y, t).
φ(x, y, t) is evaluated based on
{ ( )
∂ϕ ∇ϕ
= |∇ϕ|div ϕ(x, y, 0) = ϕ0 (x, y) (7)
∂t |∇ϕ
The initial level set represents φ(x, y, 0).
Active contour models (ACMs), denoted as snakes, are initially used to evolve contours by finding a solution to the energy mini­
mization problem. Compared to the snake’s parameters, contours are indirectly defined by level set-based ACMs. In recent years,
different types of ACMs have been established to enhance the image segmentation performance, among which is the region-based
Chan-Vese model that is widely used. The Chan-Vase model’s energy function is expressed as
∫
F(c1 , c2 , C) = μ ⋅ Length(C) + v ⋅ Area( inside (C)) + λ1 |u0 (x, y) − c1 |2 dxdy

(8)
inside (C)
∫
2
+λ2 |u0 (x, y) − c2 | dxdy
outside (C)

The raw image u0(x, y), a closed curve is C, C’s length is represented by the first term Length (C), the area inside C is denoted by the
second term, and regulated scalar parameters are µ,ν,λ1,λ2. Furthermore, the image u0(x, y)’s mean values inside and outside the curve
C are represented by c1, respectively.
I(x, y), image interest assumed model, comprises 2 regions of different intensities. The smoothed Heaviside function represents C’s
interior
( )
1 1 ϕ
He (ϕ) = + arctan (9)
2 π ϵ
And the exterior is represented by 1 − He. The smoothed Dirac delta function’s derivative is,
∂He (ϕ) 1 ϵ
δe (ϕ) = = (10)
∂ϕ π ϵ2 + ϕ2
The functional energy related to C is given by
E(ϕ(x,
∫ y, t)) = ∫
μδe (ϕ(x, y, t))|∇ϕ(x, y, t)| + νHe (ϕ(x, y, t))dxdy + λ1 (x, y)(I(x, y) − m1 )2 Hε (ϕ(x, y, t))dxdy
∫
Ω Ω (11a)
2
λ2 (x, y)(I(x, y) − m2 ) (1 − He (ϕ(x, y, t)))dxdy
Ω

C’s length is penalized by µ, the enclosed area is penalized by ν (ν = 0and (µ = 0.2 are set), and the inside and outside mean in­
tensities of the image are C m1 and m2.
Superior control over C is afforded, and λ2 and λ1are the generalized constants utilized in [3] to λ2(x, y) and λ1(x, y), which are
parameter functions in (4). The expansion and shrinkage of definite contour location (x, y) if λ2(x, y) > λ1(x, y) or λ2(x, y) < λ1(x, y)
are the location where the contour shrinks or expands. These functions of parameters are trainable and learnable straight through the
CNN backbone in DCACNN.
Given that φ(x, y, 0) and λ2(x, y), and λ1(x, y) is the parameter maps, the Active Contour Model is progressed by numerical time-
integration inside a narrow band from place to place C for increasing the efficiency of computation.

3.3. 2D CNN based disease classification

A two-dimensional CNN design consists of a series of convolutions that are utilised to extract features from pictures. This type of
design frequently finishes with completely linked layers to forecast a specific class label or a group of class probabilities. Convolution
layers add filters to all pixels of the input picture to produce a set of highly abstract features; pooling layers limit the number of features
to prevent overfitting; and completely restructure the outcome into vectors of the same size as the number of classes. A 2D-CNN uses

5
N. Raj et al. Computers and Electrical Engineering 103 (2022) 108357

two activation functions for classification: softmax for such output nodes and Rectified Linear Unit (ReLu) for the remainder of the
layers. Softmax seeks to scale the outcomes between zero and one, indicating the likelihood of coverage belonging to a certain class. If
the input is positive, ReLu is a linear function that will produce it immediately. Otherwise, it will return a value of zero.In addition, 2D-
CNN 1 is made up of three convolution layers that are preceded by a max-pooling layer. The filter size (fs) for this sequence was picked
from among fs = [2 2, 4 4, 8 8,] as the first no., providing the best performance. Fig. 3 shows a graphical representation of the 2D-CNN
architecture.
The 2D-CNN 2, on the other hand, is made up of 3 convolution layers, which are preceded by a 2 2 max-pooling. A one-to-one
convolution layer is used in this design to extract more characteristics from the pictures without sacrificing information. Then,
dropout layers are used to disable a fraction of the neurons to prevent overfitting.
There are seven levels in the CNN model. Certain information is handled in each layer. The following are the 7 layers: The data is
stored in the case of images in the input layer. The image’s height, breadth, depth, and colour information are among the charac­
teristics (RGB). The picture input size is set to 224 × 224 RGB (Figs. 10 and 11).
The Convolution Layer: The feature extraction layer is also known as the convolution layer. This layer extracts the main charac­
teristics from the provided collection of photographs.
The Pooling Layer: By lowering (or) altering the shape of the featured matrix created by utilising dot products, the pooling layer
aims to limit the processing resources required to analyse the data. layer that is fully connected: Loads, neurons, and biases are all part
of it. It links neurons in different convolution layers.
Multi-classification is carried out via the Softmax Layer and Logistic Layer. The logistic layer carries out the binary classification. It
determines the likelihood of a certain object being present in the picture. The probability is "1" if the object is visible in the image; else,
it is "0".
ReLU’s Activation Function: It activates the node by transforming the whole weighted input through the node and putting it into
the operation. The Rectified Linear Unit (ReLU) is a convolution algorithm in neural networks.

Algorithm of DACCNN_2DCNN:
Given an input image f, select the parameters λ and μ.
→0
Initialize the algorithm by choosing u0 = 0, d0 = 0, b = 0, c01 = 1, c02 = 0, and k = 0 while stopping criteria are not met do
2 2
1: Define rk = (ck1 − f) − (ck2 − f)
→k →k
2: uk+1 = GSGcs(rk , d , b )
( )
→k 1
3: dk+1 = shrink ∇uk+1 + b ,
λ
→k+1 →k →k+1
4: b = b + ∇uk+1 − d
5: Find Ωk = {x: uk(x) ≥ γ}
∫ ∫
6: Update ck+1
1 = Ωk fdx, and ck+1
2 = (Ωk )c fdx
7: k = k + 1end while
(continued on next page)

Fig. 3. A schematic view of the proposed 2D-CNN.

6
N. Raj et al. Computers and Electrical Engineering 103 (2022) 108357

(continued )
Data: I: Image; G: Corresponding ground truth label; g: ACM energy function with parameter maps λ1,λ2; ϕ: ACM implicit function; L: Number of ACM iterations; W:
CNN with weights w; P: CNN prediction; L: Total loss function; η: Learning rate
Result: Trained TDAC model
while not converged, do
⃒
⃒ ∂ϕ
λ1 , λ2 , ϕ0 = W(I) P = Sigmoid(ϕ0 ) for t = 1 to L do ⃒⃒ t− 1
∂t
∂ϕt− 1
= g(ϕt− 1 ; λ1 , λ2 , I) ϕt = ϕt− 1
+ Δt end
∂t
L = LΛCM (ϕL ) + LCNN (P)
∂L
Compute and backpropagate the error
∂w
∂L
Update the weights of W : w ← w − η
∂w
K=k+1
End while

4. Performance examination

Plant disorder detection is focused on in this work, which is detected using the ensemble classification method. This approach’s
performances are evaluated using convoluted Gaussian filtering and a deep convoluted active contour neural network (DCACNN). The
performance metrics are precision, recall, and accuracy.
Accuracy: The number of patterns properly segmented to the entire number of samples is defined as accuracy, and this accuracy is
calculated as follows:
TP + TN
A= (11)
TP + TN + FP + FN
Precision: The ratio of the positive number of samples to the total number of samples is known as precision.
TP
P= (12)
TP + FP
Recall: The ratio of true positive patterns to the total positive declared number of patterns is known as recall.
TP
R= (13)
TP + FN
The testing accuracy of the model will be generated after every epoch while the model is deployed. The confusion matrix for crop
disease predicted class has been shown in Fig. 4. The below Fig. 5 shows healthy and unhealthy leaves from the input dataset.
The below Fig. 6 shows the proportion of healthy and unhealthy leaves detected using the proposed classification technique.
The above Figs. 7–9 show the parametric examination of accuracy and recall between existing and proposed techniques. Table 1
shows the comparative examination in crop disease detection.
From the above table, a comparative analysis has been carried out for tomato and brinjal crop disease detection in terms of ac­
curacy, precision, recall and F-1 score between proposed and existing techniques. Here, the existing techniques compared are K-means,

Fig. 4. Confusion matrix of crop disease predicted class.

7
N. Raj et al. Computers and Electrical Engineering 103 (2022) 108357

SVM, and GLCM.

The above figures show the comparative analysis for tomato and brinjal crop disease detection using deep learning-based classi­
fication and segmentation. In tomato disease detection, the proposed technique achieved an accuracy of 98%, a precision of 93%, a
recall of 78%, and an F-1 score of 77%; in brinjal disease detection, the proposed technique achieved an accuracy of 98%, a precision of
90%, a recall of 75%, and an F-1 score of 76%. From this analysis, the proposed technique obtained optimal results in various crop
disease detections using the proposed technique.

5. Conclusion

Plant disease detection is the main challenge for deep learning and image processing. So this paper proposed crop disease detection
based on preprocessing and segmentation using filtering and neural network techniques. We presented the processing and filtering
procedure by preprocessing using convoluted Gaussian filtering, and then, based on the neural network, the image was segmented
using a deep convoluted active contour neural network (DCACNN) and classified using 2DCNN. Plant Disease Image’s benchmark
dataset is used for performing experiments. Healthy Leaves1000 images and crop diseases are present in this dataset. Automatic
detection of crop diseases is the major motivation of this system. 98% accuracy is obtained in the classification of experimental results
by the proposed system. The database for identifying diseases in plants is expanded in future research and also used in classification.
Training increments increase the system’s accuracy, so the training data needs to be improved in future research.

Statements and declarations

Ethical approval

This article does not contain any studies with animals performed by any of the authors.

Funding

The authors extend their appreciation to the Deanship of Scientific Research at King Khalid University, Abha, Kingdom of Saudi
Arabia for funding this work through Large Groups RGP.2/119/43.

Conflict of Interest

The authors declare that they have no conflict of interest.

Fig. 5. Detection of healthy and unhealthy leaves.

8
N. Raj et al. Computers and Electrical Engineering 103 (2022) 108357

Fig. 6. Proportion of healthy and unhealthy leaves detected using the proposed classification technique.

Fig. 7. Examination of accuracy for the existing and proposed technique.

Fig. 8. Precision examination for the existing and proposed technique.

9
N. Raj et al. Computers and Electrical Engineering 103 (2022) 108357

Fig. 9. Recall examination for the existing and proposed technique.

Fig. 10. Comparative analysis for Tomato crop disease detection in terms of (a) accuracy, (b) precision, (c) recall (d) F-1 score.

10
N. Raj et al. Computers and Electrical Engineering 103 (2022) 108357

Fig. 11. Comparative analysis for Brinjal crop disease detection in terms of (a) accuracy, (b) precision, (c) recall, and (d) F-1 score.

Table 1
Comparative examination in crop disease detection.
Crop Parameter k-means SVM GLCM Pro_2DCNN

Tomato Accuracy 95 96 97 98
Precision 89 90 92 93
Recall 74 75 76 78
f1-score 73 75 76 77
Brinjal Accuracy 94 95 96 98
Precision 86 87 89 90
Recall 71 72 74 75
f1-score 71 73 75 76

Informed consent

Informed consent was obtained from all individual participants included in the study.

Availability of data and material

All date available in Manuscript.

Code availability

All date available in Manuscript – Custom Mode.

11
N. Raj et al. Computers and Electrical Engineering 103 (2022) 108357

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Data availability

Data will be made available on request.

Acknowledgements

The authors extend their appreciation to the Deanship of Scientific Research at King Khalid University, Abha, Kingdom of Saudi
Arabia, for funding this work through Large Groups RGP.2/119/43.

References

[1] Supriya S, Patki GSS. A review: cotton leaf disease detection. IOSR J VLSI Signal Process (IOSR-JVSP) 2016;6(3):78–81.
[2] Sowmya BJ, et al. Utility system for premature plant disease detection using machine learning. Hybrid computational intelligence. Academic Press; 2020.
p. 149–72.
[3] Oo YM, Htun NC. Plant leaf disease detection and classification using image processing. Int J Res Eng 2018;5(9):516–23.
[4] Iqbal Z, et al. An automated detection and classification of citrus plant diseases using image processing techniques: a review. Comput Electron Agric 2018;153:
12–32.
[5] Sharif M, et al. Detection and classification of citrus diseases in agriculture based on optimized weighted segmentation and feature selection. Comput Electron
Agric 2018;150:220–34.
[6] Sherly Puspha Annabel L, Annapoorani T, Deepalakshmi P. Machine learning for plant leaf disease detection and classification - a review. In: Proceedings of the
2019 IEEE international conference on communication and signal processing, ICCSP. 2019; 2019. p. 538–42.
[7] Basavaiah J, Arlene Anthony A. Tomato leaf disease classification using multiple feature extraction techniques. Wirel Personal Commun 2020;115(1):633–51.
[8] Mkonyi L, Rubanga D, Richard M, Zekeya N, Sawahiko S, Maiseli B, Machuve D. Early identification of Tutaabsoluta in tomato plants using deep learning. Sci
Afr 2020;10:e00590.
[9] Hlaing CS, Zaw SMM. Model-based statistical features for mobile phone image of tomato plant disease classification”. In: Proceedings of the parallel and
distributed computing, applications and technologies PDCAT; 2018. p. 223–9. 2017-December.
[10] Nandhini SA, et al. Web-enabled plant disease detection system for agricultural applications using WMSN. Wirel Personal Commun 2018;102(2):725–40.
[11] Mukhopadhyay S, et al. Tea leaf disease detection using multi-objective image segmentation. Multimed Tools Appl 2021;80(1):753–71.
[12] Saeed F, et al. Deep neural network features fusion and selection based on PLS regression with an application for crops diseases classification. Appl Soft Comput
2021;103:107164.
[13] Ennouni A, Sabri MA, Aarab A. Plant diseases detection and classification based on image processing and machine learning. Intelligent systems in big data,
semantic web and machine learning. Cham: Springer; 2021. p. 283–90.
[14] Sharath DM, Akhilesh SAK, Rohan MG, Prathap C. Image-based plant disease detection in pomegranate plant for bacterial blight. In: Proceedings of the
international conference on communication and signal processing (ICCSP). IEEE; 2019. Year:| Conference Paper | Publisher.
[15] Devaraj A, Rathan K, Jaahnavi S, Indira K. Identification of plant disease using image processing technique. In: Proceedings of the international conference on
communication and signal processing (ICCSP). IEEE; 2019. F Year:| Conference Paper | Publisher:.

Further reading

[1] Cevallos C., Ponce H., Moya-Albor E., Brieva J. VisionBasedExamination on leaves of tomato crops for classifying nutrient deficiency using convolutional neural
networks. Proceedings of the international joint conference on neural networks; 2020. p. 0–6.
[2] Sandhu G.K., Kaur R. Plant disease detection techniques: a review. Proceedings of the international conference on automation, computational and technology
management (ICACTM). IEEE; 2019. Year:| Conference Paper | Publisher.

12