Measurement: Sensors 24 (2022) 100581

Contents lists available at ScienceDirect

Measurement: Sensors
journal homepage: www.sciencedirect.com/journal/measurement-sensors

Performance analysis of deep learning architectures for plant leaves

disease detection
Sachin Dahiya a, *, Tarun Gulati a, Dushyant Gupta b
a
Department of Electronics & Communication Engineering, Maharishi Markandeshwer Engineering College, Maharishi Markandeshwer (Deemed to be University),
Mullana, Ambala, Haryana, India
b
Electronics Department, Institute of Integrated & Hons. Studies, Kurukshetra University, Kurukshetra, India

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

Keywords: Plant disease is a major concern in agriculture field. Different pathogens are causes of plant disease. Manual
CNN inspection sometimes can cause errors in the disease detection over large cultivation area. Large number of
Machine learning resources and high level of expertise is required for correct prediction of diseases. Sometime the experts are not
Deep learning
available in the nearby. Some diseases are so common and have same symptoms even the different experts may
give different opinion about the same disease.
In this work different deep architectures are studied and tested for plant leaves disease detection. For carrying
out the study plant village dataset of 20640 images are used which represents 15 class and 3 species namely
pepper, potato, and tomato. Total images in the dataset are divided in three sets i.e., 70% for training, 20% for
validation and 10% for testing purpose. MATLAB R2019b is utilized for experimentation. Appropriate value of
hyper parameter is determined initially by varying number of epochs, learning rate. Performance of Adam
optimizer is compared with Sgdm Optimizer. Finally, accuracy of eight deep learning architectures is calculated
at 16 & 32 minibatch size.

1. Introduction transformation that has transformed many managerial functions into

artificially intelligent systems to extract value from the growing amount
Indian economy is basically an agriculture-based economy. Majority of data coming from diverse sources. Deep learning, a branch of machine
of the Indian population earn their livelihood from the agriculture learning subset of artificial intelligence, has a great deal of promise to
sector. Agriculture sector contributes a significant proportion towards overcome many obstacles in the development of knowledge-based
the GDP of the country. Latest research and use of technology in the field farming systems that will boost crop production and so resolve all
of agriculture makes the country self-sufficient and a net exporter of problems.
agricultural and related goods, as opposed to being a food-begging bowl
following independence. India’s growing population, rising average 2. Application of machine intelligence in agriculture
income and the consequences of globalization will raise demand for a
greater variety, quantity, and quality of food. To produce more food in In the recent times the computer vision and machine learning algo­
greater quantity, variety, and quality put pressure on the limited amount rithms have made a tremendous growth in the plant leaves detection
of arable land. India is fortunate to have a lot of land that is available, 15 area. Machine learning methods such as artificial neural network, sup­
agro-climatic zones that, according to the ICAR, have practically every port vector machine and k-nearest neighbours are employed in the field
sort of weather and soil type, and the ability to cultivate a wide range of of plant leaves detection. Deep learning is an advancement of machine
crops. Despite these facts, India’s average crop productivity is still learning approach. The most attracting feature of deep learning over
extremely poor. The population of the nation is anticipated to grow to be machine learning is that it is free from feature engineering. Deep neural
the largest in the world in the upcoming years, making the provision of networks basically create mapping between input which is an image and
food for all is a top concern. Use of artificial intelligence, computer with the output in the form of crop disease pair. Deep learning archi­
vision in the field of agriculture has undergone a tremendous amount of tectures consist of convolution neural network. CNN consists of large

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

https://doi.org/10.1016/j.measen.2022.100581
Received 16 October 2022; Received in revised form 1 November 2022; Accepted 19 November 2022
Available online 22 November 2022
2665-9174/© 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
S. Dahiya et al. Measurement: Sensors 24 (2022) 100581

expensive, many industrialized nations now deploy robots to control

weeds.
• Disease Detection: Crop disease is a serious risk to the agriculture
industry. Manual disease identification in crops is a laborious and
lengthy process. With the advancement of deep learning and its ar­
chitectures, it is now a quicker process.
• Yield Prediction: Crop production estimation using machine
learning and deep learning techniques gives a fair judgment about
total yield of the crops. The decision of what to cultivate and when to
cultivate it is made easier for the farmers.
• Fruit Quality and Quantity Estimation: Computer vision has made
considerable strides recently and is primarily used to recognize ob­
jects. Use of latest machine intelligence-based techniques in the large
farms will be very helpful in producing good quality of fruits and in
larger quantity.

3. Related works

Yanli.et al. [1] discussed that lightweight models are much faster in
the identification of disease and requires less memory space. HLNet deep
learning model is proposed based on lightweight CNN and for fast and
Fig. 1. Deep CNN(Google images). efficient identification of disease. Mohammad Suleman et al. [2] dis­
cussed the identification of cotton leaf disease. A model based on the
number of layers such as input layer, output layer and several hidden meta deep learning is proposed which is capable enough to identify the
layers as shown in Fig. 1. Several hidden layers provide greater level of several crop diseases. Dataset contains 2385 images of healthy and
accuracy as compared to shallow networks. Convolution neural net­ diseased leave. The proposed model has achieved 98.53 accuracy. Plant
works are a class of DL methods. CNN can learn the complex problem image recognition using deep learning techniques based on leaf vein
and perform the pattern recognition when provided with a large dataset. patterns was first developed [3]. Three species of leguminous
Various DL methods now a day’s find application in various fields plants—white bean, red bean, and soybean—were classified using 3–6
because it facilitates the development of automated application that layers of CNN technology. Deep learning models was trained to evaluate
requires negligible amount of human intervention. the model’s performance on the Plant Village dataset in Ref. [4]. The
The machine intelligence-based approach like machine learning, performance of the two well-known CNN architectures AlexNet and
deep learning, and use of computer vision along with application of IoT GoogLeNet were evaluated using the three scenarios (color, grey and
technology greatly helps the farmers and makes farming much easier. segmentation), and it was determined that GoogleNet performed better
Use of drones powered by artificial intelligence (AI) is particularly than AlexNet. On the test set, the trained model had a 99.35% accuracy
beneficial for farming since it makes it simpler to monitor, scan, and rate. To conduct symptom-wise recognition of four cucumber illnesses,
analyses the crops by supplying high quality photographs. These images Ma et al. [5] used a deep CNN (i.e., downy mildew, anthracnose, pow­
may be applied to machine intelligence-based algorithms for tracking dery mildew, and target leaf spots). The recognition precision was
the growth of crops at one hand and makes the farmer aware about the 93.4%. A system based on CNN was presented by Kawasaki et al. [6] to
threats of any spreadable disease at the earliest. The main application identify cucumber leaf disease, and it achieved an accuracy of 94.9%.
area of the machine intelligence-based approaches is: Muammer et al. [7] used DL methods for recognizing the plant leaf
disease and pasts. Nine Architecture of deep learning was used for
• Crop Management: Crop management is a crucial activity for feature extraction. Features were then classified by SVM, ELM and
raising crop quality. In today’s agriculture, crop management func­ k-nearest neighbor method. Highest accuracy of 97.86 was achieved
tions including crop monitoring, field scanning, and so forth, drones with ResNet 50 model and SVM, classifier. The main limitation of this
play a critical role. Deep learning algorithms are employed in smart theory was that the data set utilized for experimentation was very small.
agriculture to track the temperature and moisture content of the Only 1965 images of 8 different plant leaf diseases were used. A deep
crops. Farmers may also keep an eye on their fields from anywhere in CNN framework was created by Ghosal et al. [8] to recognize and
the world. categorize 8 different types of soybean stress. Additionally, it will offer
• Water Management: For a crop to grow properly, water is neces­ an explanation mechanism and make predictions using the top-K
sary. Regularly inspection of moisture level in the soil is extremely high-resolution feature maps that isolate the visual symptoms. With
important for the development of the crops. With the use of machine no comprehensive symptom annotation by experts, the unsupervised
intelligence approaches, we can effectively manage the water level recognition of visual symptoms allowed for identification (a type of
based on the crop and soil categories. foliar stress), classification (low, medium, or high stress), and creation of
• Soil Management: One can quickly identify the appropriate crops a quantitative measure of stress intensity, instead of considering the
and the optimum kind of fertilizer for the soil by using machine entire leaf, Barbedo [9] and Lee et al. [10] suggested the usage of lesions
intelligence-based techniques. For more effective soil management, and spots. The benefits of this approach include the ability to identify the
SVM (Support Vector Machines), Bayesian Networks, and ANN Ma­ presence of multiple diseases on the same leaf and the ability to enhance
chine learning techniques are most used techniques. Regular moni­ the data by segmenting the leaf picture into various sub-images. The
toring of pH level and soil nutrients parameters is very helpful in GoogLeNet model was utilized in this study by J. Wang [11] to identify
determining the condition of the soil because it determines the 79 diseases in 14 kinds of plants that were present in challenging field
quality and quantity of the crops. and experimental conditions. The total accuracy rate for a single lesion
• Weed Detection: Weed detection and elimination are the primary and area was 94%, which was higher than the 82% accuracy rate for the
issues in the agricultural sector. Weed removal used to be done full image. To identify the areas affected by the wheat disease, Qiu et al.
manually in older times. Although it is time-consuming and [12] employed the Mask-RCNN, whose feature extraction network was
either ResNet50 or ResNet101. The average accuracy on the test dataset

2
S. Dahiya et al. Measurement: Sensors 24 (2022) 100581

was 92.01%. Based on the neural structure search technique, Huang Table 1
et al. [13] approached for identifying plant leaf diseases can automati­ AlexNet architecture details.
cally learn the neural network’s structure to the necessary depth on the Layers No. of filters Filter Size
PlantVillage. According to the results of the research, a suitable network
1st convolutional layer 96 11 x 11
structure was found on the datasets of imbalanced and balanced, and the 2nd convolutional layer 256 5x5
model’s recognition accuracy was 98.96% and 99.01%, respectively. 3rd convolutional layer 384 3x3
However, accuracy dropped to 95.40% if the grey image balance was not 4th convolutional layer 384 3x3
corrected. In order to detect camellia leaf diseases, Long et al. [14] 5th convolutional layer 384 3x3

employed AlexNet for two types of training: training from scratch and
transfer learning from the ImageNet (4 kinds of diseases and healthy). 4.1.1. Filter size
The findings demonstrated that transfer learning may considerably Various filter size is available in the CNN as 1x1, 3x3 & 5x5 etc.
enhance the models’ classification performance and convergence speed, Depending upon the application and the datasets, various filter size
with classification accuracy reaching 96.53%. Saliency maps were first determines the features sets to be detected by the CNN.
used to show the signs of plant diseases by Brahimi et al. [15] using the
precision (P), recall (R), F1 score, and overall accuracy metrics, Cruz 4.1.2. Padding
et al. [16] employed segmentation and edge mapping to identify plant Padding is required in the CNN so that after convolution the output
diseases to discussed detect olive plant diseases using the upgraded of the layer is not reduced in size. Since CNN contains multiple convo­
LeNet model. A new DL model network was created by Brahimi et al. lution layers, hence without padding it may lose some area after several
[17] to identify the spots of plant illnesses, and when compared to the convolution steps.
current plant disease treatment approaches, the new method produced a
clearer visualization impact. In Ref. [18] J Arun Pandian proposed a 4.1.3. Polling
14-layer deep CNN model for detection of leaf disease using leaf images. Polling layers in the CNN is required to reduce the number of pa­
Three augmentation techniques are used. Dataset contains 147500 im­ rameters for layers presents inside the CNN. It is also helpful in reducing
ages of 58 healthy and diseased classes and on no leaf class. The pro­ the calculations during the learning phase.
posed DCNN model achieved 99.96555% accuracy. The proposed model
gives better performance than other existing learning approaches. 4.1.4. Stride
Waleed albattah in Ref. [19] has introduced custom CentreNet frame­ Stride determines the movement of filters across the pixel of image
work with DenseNet-77 as a base network. Proposed method adopted horizontally. Stride is denoted by S.
three stages strategy i.e use of annotation to extract region of interest, at
second stage proposed CentreNet in which DenseNet-77 is used for key 4.1.5. Convolution arithmetic
point abstraction. Finally, CentreNet is used for categorization of dis­ The dimension of the output is determined by the following equa­
ease. In Ref. [20] C. Jackulin et al. carried out a review study on various tions.
ML and DL techniques in plant disease detection using AI and ML. author
has also compared the performance of ML and DL techniques and their Lo/ =
(N − F)
(1)
usage in various study. Dahiya et al. in Ref. [21] discussed about the p S+1
deep learning architecture, various CNN related parameters. Various
factors that affect the performance of DL model in plant leave disease (N + 2P − F)
Lo/ = +1 (2)
detection is also discussed. p S

where, N: length on side of activation layer; S: Stride; F: Filter size; P:

4. Artificial neural network Depth of padding.

Artificial neural networks are based upon the pattern of neuron 4.1.6. Confusion matrix
connection of the brain. These networks try to learn the things like brain. A Confusion matrix determines the accuracy of model under
Neuron nodes in artificial neural networks are linked together like a consideration. The structure of confusion matrix is N x N. This matrix is a
web, just like in the brain. There are millions of neurons in the human basically determines the performance of the model by comparing the
brain. A cell body, which is the component of each neuron, oversees actual class with the predicted class. Confusion matrix gives a fair rep­
processing information by transporting it into and out of the brain. The resentation of model performance.
neural network tries to seek pattern from the data supplied to produce
desired output from the input information, which receive a variety of
information based on an internal weighting system. Backpropagation 4.2. Deep CNN architectures
which stands for backward propagation of error, is a set of learning rules
that ANNs use to refine their output results, like human being learn 4.2.1. AlexNet
various rules and guidelines to produce a desired outcome. Development of AlexNet [22] has revived the interest of the
researcher towards CNN. It is a pretrained CNN.t AlexNet consists of 8
weight layers out of which are 5 layers are responsible for convolution
4.1. Convolutional neural network and three layers having fully connected connection (Table 1 & Fig. 2).
Three max-pooling layers are also used after the 1st, 3rd and 5th layer
A Convolutional Neural Network (CNN) is a special class of ANN. and computed nearly sixty million parameters. The architecture also
These are deep networks which accepts image as an input and assign contains some activation layer.
some internal weights and biases to different parts or the objects 4096 neurons are presents in fully connected layer. The second fully
appearing in the image and try to differentiate the different objects. The connected layer is connected to a SoftMax classifier which gives the
special feature which makes it a popular choice among the researcher is output in 1000 classes.
that it requires little or almost no preprocessing as compared to the The performance of the AlexNet is degrades due to overfitting.
traditional approaches. The pre-processing required in CNN is much Augmentation technique gives satisfactory results while dealing with
lower than the other classification algorithms. The various parameters overfittings. Two different types of augmentation techniques may be
which need to be suitably defined are as follows: utilized for the above purpose, one is positional augmentation comprises

3
S. Dahiya et al. Measurement: Sensors 24 (2022) 100581

Fig. 2. AlexNet (src: Google images).

Fig. 3. Naïve inception model.

of the image’s translation and reflection and the other one includes color architecture is four million which is much less than sixty million pa­
augmentation (Brightness, hue, etc.) generates images by using differing rameters of AlexNet. GoogleNet architecture consists of 22 deep
intensities in the images. The other techniques to avoid the overfitting convolution layers along with 27 global average pooling layers and 9
are to use dropout. Dropout leads to the generalization of the network. inception modules. The global average pooling layer relates to the
Due to dropout technique some of the units may be ignored leading to inception modules. Such models have repeating modules.
the different sample calculations and enable the model to extract more An inception network is a DL network that is made up of some
unique features/patterns. echoing modules called inception modules. Convolution neural net­
works extract the features from an image at different levels and scales.
4.2.2. GoogleNet The main idea is the same as the human visual cortex functions by
GoogleNet architecture which is also known as Inception came into extracting different information from different patterns; this informa­
existence in 2014 after AlexNet. Inception network has three versions, tion is accumulated at the end and form larger perceptions of objects. An
named Version 1, 2 and 3. The first version of the inception module was inception module is made up of 1x1, 3x3 and 5x5 convolution layer,
named GoogleNet [23]. GoogleNet has less error rate than other archi­ average max-pooling layer, and concatenation layer. Initially designed
tectures. This architecture uses the features of 1x1 convolution layers inception module was known as Naive inception module as shown in
and global average max pooling. These features enable it to go deeper Fig. 3. It was the first designed inception module.
within the architecture layers. The number of parameters in GoogleNet It consists of varying filter sizes by which it can learn about spatial

Fig. 4. Full inception model.

4
S. Dahiya et al. Measurement: Sensors 24 (2022) 100581

Fig. 5. Standard Convolution layer in MobileNet.

patterns at different scales. In naïve inception module an extra block of

1x1 convolution layer is added with the 3x3, 5x5, and average max
pooling layers as shown in Fig. 4. It will be helpful in learning the pa­
rameters at depth and reduces dimension. The insertion of this 1x1 layer
reduces the computational costs of this module. So, we can compare the
computational costs of naïve and full inception modules and find that
full inception module’s computational costs are the times lesser. In the
inception module all convolution layers work parallel and at the end
output of these layers is concatenated to give final output.

4.2.3. MobileNet Fig. 6. Depth wise separable Convolution layer in MobileNet.

Depth wise separable convolutions technique is utilized in MobileNet
[24] architecture. Its architectures contain three convolutional layers. 1
avoids connection from some of the layers and connected to the outputs.
× 1 point wise convolution layer is used after the initial depth convo­
Skip connection techniques improve the performance of the network as
lutions. In its earlier version known as V1 the point wise convolution
it avoids layers which performs poor by regularization. Hence the
keeps channels the same or doubled them, whereas in V2 it tries to keep
problem of vanishing gradient can be removed. ResNet50 is a type of
the total channels smaller. The layer performing this function is known
variant of ResNet model having 48 convolution layers along with one
as projection layer. It converts the high dimensional data into lower
convolution layer and 1 average pool layer. This is a commonly used
dimensions. Depth wise layer works on with 144 channels, which the
model. ResNet 101 is a deeper network having 101 layers.
projection layer may reduce the number of channels to 24 channels as it
reduces the data flow across the network. The function of 1 × 1
4.2.5. ShuffleNet
convolution layer is to enhance the channels availability for the data
ShuffleNet [26] a deep CNN trained on the ImageNet set and can
before applying to the depth wise convolution. The types of convolution
classify the 1000 different class. The network accepts an image of size of
layer are shown Figs. 5 and 6.
224x224 as an input. This network implements the pointwise convolu­
tion. It also utilizes the channel stuffing for reducing the cost factor. In
4.2.4. ResNet
the beginning the 224x224 input images is fed to the convolution layer
ResNet [25] stands for residual network. After the successful
having kernel of size 3x3 with stride of 2. This is a highly efficient
implementation of AlexNet researchers put more stress in designing the
network and provide good accuracy as it reduces the number of pa­
deeper network having more layers. But increasing number of.
rameters that is why it also efficient in reducing the computational
Layers also increase the rate of training and test errors. Residual
complexity.
network architecture also reduces the problem of vanishing gradient. A
special techniques called skip connection is used in ResNet [25], which

5
S. Dahiya et al. Measurement: Sensors 24 (2022) 100581

Table 2 from 30 to 50. The results are graphically represented in Fig. 7.

Parameters for experiments.
S. No. Parameter Value 5.2. Experiment 2
1. Mini Batch Size 64
2. Learning Rate 0.001 5.2.1. Determination of learning rate
3. Epochs 30,40,50 In this study number of epochs is set to 30 and minibatch size is set to
4. Optimizer Adam
64. GoogleNet architecture is used with Adam optimizer. The learning
rate is varied from 0.001(10− 3) to 0.000001(10− 6). Results showed that
at 0.0001(10− 4) the highest accuracy 99.06 is achieved with learning
Table 3 time 1:47:21 h. At 10− 5 and 10− 6 learning rate the accuracy starts
GoogleNet Architecture with 30,40,50 epochs. decreasing due to overfitting of models. At (10− 4) learning rate the
S. No. Epochs Accuracy (%) Training Time(hrs) learning time is slightly more than 10− 3 but less than 10− 5 and 10− 6
1. 30 97.63 1:46:27 learning rate. Results are summarized in Table 4 and represented in
2. 40 97.65 3:09:48 Fig. 8.
3. 50 95.61 3:13:17

5.3. Experiment 3
4.2.6. SqueezeNet
SqueezeNet [27] is an 18-layer deep network on ImageNet dataset. It 5.3.1. Comparison of optimizers
can classify images in to 1000 classes. SqueezeNet accepts 227*227 During this part of experimentation Adam optimizer is compared
images. with the sgdm optimizer with GoogleNet architecture with 30 number of
epochs, 10− 3 to 10− 6 learning rates. The results are summarized in
5. Experimental study Table 5 and represented in Fig. 9. Results in Fig. 9 indicates that Adam
optimizer is better than Sgdm optimizer since is it is not stuck over local
In this work, initially experiment is performed with the different set minima. The performance of Adam optimizer is best at 10− 4 Learning
of hyperparameters. For finding the best set of hyperparameters Goo­ rate.
gleNet architecture is used. For experimental study a version of Plant­
Village dataset containing 20640 images of 15 class and 3 species 5.4. Experiment 4
namely pepper, potato, and tomato are taken. Total images are divided
in three sets i.e 70% for training, 20% for validation and 10% for testing 5.4.1. Determination of appropriate minibatch size
purpose. Matlab R2019b is utilized for experimentation. Six experiments In this part, GoogleNet is studied with variation in minibatch size at
are performed: minibatch size values 16,32,64. Learning rate 10− 4, 30 epochs and adam
optimizer is used. Accuracy against minibatch size is plotted in Fig. 10
5.1. Experiment 1 and summarized in Table 6.
From the above results, it has been found that minibatch size
5.1.1. Setting UP OF number of epochs
In this, three experiments were carried out to find out the appro­ Table 4
priate number of epochs to calculate the accuracy of the models in Determination of learning rate with GoogleNet.
minimum time. The number of epochs taken in this study are 30, 40 & Sr. No. Initial Learning Rate Accuracy (%) Learning Time (hrs)
50. 64 minibatch size, learning rate 0.001 with Adam optimizer is used.
1. 10–3
97.63 1:46:27
Results of the study are summarized in Table 2 ad Table 3. Results 2. 10–4 99.06 1:47:21
showed that when the number of epochs increased from 30 to 50, there 3. 10–5 97.80 2:00:12
is significant impact on the accuracy of the models. But the training time 4. 10–6 90.07 2:20:45
is increased and almost double when we increased number of epochs

Fig. 7. Accuracy vs no. of epochs (GoogleNet).

6
S. Dahiya et al. Measurement: Sensors 24 (2022) 100581

Fig. 8. Determination of learning rate with GoogleNet.

and accuracy is calculated as shown in Table 7 and represented in

Table 5 Fig. 11.
Comparison of optimizers. Results in Fig. 11 shows that performance of Alexnet and SqueezNet
Sr. No. Initial Learning Rate Adam Accuracy (%) sgdm Accuracy (%) is poor than other networks. ResNet50 & ResNet101 achieves 100%
1. 10–3 97.63 98.74 accuracy whereas GoogleNet, MobileNet, ShuffleNet and ResNet18
2. 10–4 99.06 97.02 achieves more than 99% accuracy.
3. 10–5 97.80 86.14
4. 10–6 90.07 42.94
5.6. Experiment 6

variation has negligible effect on the accuracy of the architecture. 5.6.1. Comparison of different deep learning architecture at 16 minibatch
Variation in mini batch size increases only the learning time of the size
models. Eight deep learning architecture is simulated with minibatch size 16
and accuracy is calculated as shown in Fig. 12 and summarized in
Table 8. ResNet50 and ResNet101 achieves the highest accuracy
5.5. Experiment 5
whereas AlexNet and SquezeNet has achieved the lower accuracy.
5.5.1. Comparison of different deep learning architecture at 32 mini batch
size
Eight deep learning architectures is simulated with minibatch size 32

Fig. 9. Comparison of Adam and Sgdm.

7
S. Dahiya et al. Measurement: Sensors 24 (2022) 100581

Fig. 10. GoogleNet Architecture with variation in minibatch size.

6. Conclusion
Table 6
Comparison of minibatch size.
In this work, study is carried out for determining the appropriate
Sr. No. Minibatch Size GoogleNet Accuracy (%) hyper-parameters for comparing the different deep learning architec­
1. 64 99.06 tures for plant leaf disease detection using PlantVillage dataset. Goo­
2. 32 99.27 gleNet architecture is used for study, and it is found that 30 epochs,
3. 16 98.47 0.0001 learning rate is suitable for study of plant leaf disease detection
using PlantVillage dataset. Experiment results suggest that 30 numbers
of epochs are sufficient as it takes lesser time for training without any
Table 7 significant degradation in accuracy. 10− 4 learning rate gives better re­
Accuracy (%) at 32 minibatch size. sults without overfitting of the modals. Performance of Adam optimizer
is found better than Sgdm optimizer. Using these hyper parameters eight
Sr. No. Deep learning Architecture Accuracy(%)
different architectures GoogleNet, ResNet18, ResNet50, ResNet101,
1. GoogleNet 99.27
MobileNetv2, ShuffleNet, AlexNet and SqueezNet is analyzed, and it is
2. ResNet18 99.3
3. ResNet50 100 found that ResNet50 & ResNet101 performs better than other networks
4. ResNet101 100 whereas the performance of Alexnet and SqueezNet is poorer as
5. MobileNetv2 99.95 compared to other network at 32 & 16 Mini batch size.
6. ShuffleNet 99.25
7. SqueezNet 96.15
CRediT authorship contribution statement
8. AlexNet 96.97

Sachin Dahiya: Writing – original draft, Validation, Reviewing and

Fig. 11. Accuracy (%) at 32 minibatch size.

8
S. Dahiya et al. Measurement: Sensors 24 (2022) 100581

Fig. 12. Accuracy graph at 16 minibatch size.

in: Advances in Visual Computing, International Symposium on Visual Computing,

Table 8 ISVC, 2015, pp. 638–645.
Accuracy at 16 minibatch size. [7] Muammer Turkoglu, Davut Hanbay, Plant disease and pest detection using deep
learning based features, Turk. J. Electr. Eng. Comput. Sci. (2019), https://doi.org/
Sr. No. Deep learning Architecture Accuracy(%) 10.3906/elk-1809-181.
[8] Sambuddha Ghosal, David Blystone, K. Singh Asheesh, Baskar Ganapathy,
1. GoogleNet 98.47
Subramanian Ga, Arti Singh, Soumik Sarkar, An explainable deep machine vision
2. ResNet18 98.59
framework for plant stress phenotyping 115 (18) (2018) 4613–4618 (PNAS).
3. ResNet50 99.98
[9] Jayme G.A. Barbedo, Plant disease identification from individual lesions and spots
4. ResNet101 99.88 using deep learning, Biosyst. Eng. 180 (2018) 96–107. April 2019.
5. MobileNetv2 99.25 [10] Sue Han Lee, Hervé Goëau, Pierre Bonnet, Alexis Joly, New perspectives on plant
6. ShuffleNet 99.98 disease characterization based on deep learning, Comput. Electron. Agric. 170
7. SqueezNet 97.29 (March 2020), 105220.
8. AlexNet 96.07 [11] Chongke Bi, Jiamin Wang, Yulin Duan, Baofeng Fu, Jia-Rong Kang, Yun Shi,
MobileNet based apple leaf diseases identification, Mobile Network. Appl. 27
(2022) 172–180.
draft and Editing, Dr. Tarun Gulati: Supervision, Dr. Dushyant Gupta: [12] Ruicheng Qiu, Ce Yang, Moghimi Ali, Man Zhang, Brian J. Steffenson, Cory
D. Hirsch, Detection of Fusarium head blight in wheat using a deep neural network
Supervision. and color imaging, Rem. Sens. 11 (2019) 2658, https://doi.org/10.3390/
rs11222658.
[13] J. Huang, J. Chen, K. Li, J. Li, H. Liu, Identification of multiple plant leaf diseases
using neural architecture search, Trans. Chin. Soc. Agric. Eng. 36 (16) (2020)
Declaration of competing interest
166–173.
[14] M. Long, C. Ouyang, H. Liu, Q. Fu, Image recognition of Camellia oleifera diseases
The authors declare that they have no known competing financial based on convolutional neural network & transfer learning, Trans. Chin. Soc. Agric.
interests or personal relationships that could have appeared to influence Eng. 34 (18) (2018) 194–201.
[15] Mohammed Brahimi, Kamel Boukhalfa, Abdelouahab Moussaoui, Deep Learning
the work reported in this paper. for Tomato Diseases: Classification and Symptoms Visualization, vol. 16, May
2017, pp. 299–315 (Applied Artificial Intelligence).
Data availability [16] Albert C Cruz, Andrea Luvisi, Luigi De Bellis, Yiannis Ampatzidis, “Vision Based
Plant Disease Detection System Using Transfer and Deep Learning” , American
Society of Agriculture and Biomedical Engineers.
Data will be made available on request. [17] M. Brahimi, S. Mahmoudi, K. Boukhalfa, A. Moussaoui, Deep interpretable
architecture for plant diseases classification, in: 2019 Signal Processing:
Algorithms, Architectures, Arrangements, and Applications, SPA, 2019,
References pp. 111–116, https://doi.org/10.23919/SPA.2019.8936759.
[18] J.A. Pandian, V.D. Kumar, O. Geman, M. Hnatiuc, M. Arif, K. Kanchanadevi, Plant
[1] Y. Xu, S. Kong, Z. Gao, Q. Chen, Y. Jiao, C. Li, HLNet model and application in crop disease detection using deep convolutional neural network, Appl. Sci. 12 (2022)
leaf diseases identification, Sustainability 14 (2022) 8915, https://doi.org/ 6982, https://doi.org/10.3390/app12146982.
10.3390/su14148915. [19] Waleed Albattah, Marriam Nawaz, Javed Ali, Momina Masood, K. Saleh Albahli1,
[2] M.S. Memon, P. Kumar, R. Iqbal, Meta deep learn leaf disease identification model Disease detection using deep convolutional neural network, Appl. Sci. 12 (2022)
for cotton crop, Computers 11 (2022) 102, https://doi.org/10.3390/ 6982, https://doi.org/10.3390/app12146982.
computers11070102. [20] C. Jackulin, S. Murugavalli, A Comprehensive Review on Detection of Plant Disease
[3] L. Guillermo, Grinblat, Lucas C. Uzal, Mónica G. Larese, Pablo M. Granitto, Deep Using Machine Learning and Deep Learning Approaches” Measurements: Sensors,
learning for plant identification using vein morphological patterns, Comput. September 2022, https://doi.org/10.1016/j.measen.2022.100441.
Electron. Agric. 127 (September 2016) 418–424. [21] Sachin Dahiya, Tarun Gulati, Machine intelligence in plant leaf disease detection
[4] S.P. Mohanty, D.P. Hughes, M. Salathe, Using deep learning for image-based plant and classification: a review, J. Comput. Theor. Nanosci. 17 (6) (June 2020)
disease detection, Front. Plant Sci. 7 (2016) 1419. 2645–2652.
[5] Juncheng Maa, Keming Du, Feixiang Zheng, Lingxian Zhang, Zhihong Gong, [22] A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep
Zhongfu Sun, A recognition method for cucumber diseases using leaf symptom convolutional neural networks, in: F. Pereira, C.J.C. Burges, L. Bottou, K.
images based on deep convolutional neural network",, Comput. Electron. Agric. Q. Weinberger (Eds.), Advances in Neural Information Processing Systems, Curran
154 (November 2018) 18–24. Associates, Inc.), 2012, pp. 1097–1105.
[6] Yusuke Kawasaki, Hiroyuki Uga, Satoshi Kagiwada, Hitoshi Iyatomi, Basic study of [23] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed,
automated diagnosis of viral plant diseases using convolutional neural networks, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich,

9
S. Dahiya et al. Measurement: Sensors 24 (2022) 100581

Going deeper with convolutions, in: Proceedings of the IEEE Conference on [26] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, Jian Sun, ShuffleNet: an Extremely
Computer Vision and Pattern Recognition, 2015. Efficient Convolutional Neural Network for Mobile Devices, 2017, 01083v2 arXiv
[24] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, L.C. Chen, MobileNetV2: inverted preprint arXiv:1707.
residuals and linear bottlenecks, in: IEEE/CVF Conference on Computer Vision and [27] Forrest N. Iandola, Song Han, Matthew W. Moskewicz, Khalid Ashraf, William
Pattern Recognition, IEEE, 2018, pp. 4510–4520. J. Dally, Kurt Keutzer, SqueezeNet: AlexNet-Level Accuracy with 50x Fewer
[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun, Deep residual learning for Parameters and <0.5 MB Model Size, 2016, 07360 arXiv preprint arXiv:1602.
image recognition, in: Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, 2016, pp. 770–778.

10