IEEE Conference ID: 62700 13th – 15th Nov.

2024
Proceedings of the 2024 4th International Conference on Technological Advancements in Computational Sciences

Deep Learning Image Classification Abiotic Plant

Disease Detection Classification Using MobileNet

Ritu Aggarwal1, Eshaan Aggarwal2, Anurag Jain3, Rajanikanth Aluvalu4,

Uma Maheswari5, Tanupriya Choudhury6
Department of Computer Science and Engineering,
1

Seth Jai Prakash Mukand Lal Institute of Engineering and Technology Radaur, India
²Department of Computer Science and Engineering Maharishi Markendeshwar Engineering College (MM(DU), India
³School of Computer Science and Engineering, IILM University, Gurugram, Haryana, India
⁴Symbiosis Institute of Technology, Hyderabad campus, Hyderabad, Symbiosis International University, Pune, India
5
Department of Computer Science and Engineering, Chaitanya Bharathi Institute of Technology, Hyderabad, India
6
School of Computer Sciences, University of Petroleum and Energy Studies UPES Dehradun Uttarakhand India–248007
E-mail: [email protected], [email protected], [email protected],
5
[email protected], [email protected]

Abstract—In the field of agriculture sector the key

o n y
providing vital food supplies, income, and employment
l
i O
role of feature food which has a major role in the emergent opportunities. With a staggering 53% of the population
population with their economy. In the food production
n
s
engaged in farming, this sector contributes substantially,
for plant disease, it may cause significant loss for the era

r g
constituting 18% of the nation’s income. Over the past three
of diversity in species and the crop production rate in the

e
plant disease detection and diagnosis the accurate detecting
years, agriculture’s contribution to the Gross Value Added
n
(GVA) has shown a steady rise, climbing from 17.6% to
i
V
techniques could increase the superiority in food productivity
and reduce the economic loss for the image categorization and

a d
20.2%, underscoring its pivotal role in driving economic

ft
object recognition. As per computed results and analysis of the growth [1]. Despite its significance, agriculture faces
various techniques, deep learning is as best approach to get
r e
formidable challenges, particularly from plant diseases

ra
the accuracy means and classify the images. In this proposed
work to tune the parameters required to train load balancing
o f and pest infestations, which threaten both crop quality and
yield. Traditional methods of disease detection, reliant on

o
overload forecasting load balancing overload forecasting the manual inspection by experts, are time-consuming and

D r Pr
models using CNN. The model performance is evaluated by
subjective, leading to potential misdiagnoses [2]. Moreover,
confusion matrix parameters such as accolade balancing
overload forecasting accuracy, precision, Recall, sensitivity, erroneous identification may result in the inappropriate use
of pesticides, exacerbating environmental concerns and

Fo
specificity, f-measure, etc. In the plants, the stress will destroy
the crop’s growth, and yields are affected negatively due to crop quality issues. Enterprises in computer vision offer
the water supply’s suboptimal quantity, soil erection, and a promising avenue for revolutionizing disease detection
abnormal temperature due to other physiological conditions in agriculture. By harnessing algorithms to analyze plant
and changes in biochemicals. In this paper, the current model images, computer vision technology can swiftly and
Mobile, which is used for image detection and classification in accurately identify disease symptoms and patterns, enabling
disease, achieves an accuracy of 99.8%.
proactive intervention and management [3]. Researchers
Keywords: Climate change, Crop disease production, Drought,
Heat stress, Plant responses, Stress management
have proposed a variety of techniques to automate disease
detection, ranging from conventional image processing
I. Introduction methods to sophisticated machine learning algorithms that
In plants, Abiotic disorders are very common. In plant extract features and segment images. These advancements
samples, the clinical parameters for the different samples hold the potential to transform agricultural practices, offering
are collected to detect abiotic disorders. The abiotic disorder efficient and precise solutions to combat plant diseases and
is mainly due to soil factors, toxicity in plants means there safeguard food security. In the field of agriculture leveraging
are no nutrients, high temperatures of plants, and weather innovative technologies like computer vision holds the key
issues. As in the environment, the abiotic stress impacts to mitigating the impact of plant diseases and pests, ensuring
the plant’s growth and development. Different challenges sustainable agricultural practices and economic prosperity.
occur in the plant’s disease to survive and reproduce the Abiotic stressors pose significant challenges to crop
plants by drought, salinity, and temperature. The field of production, factors, particularly under conditions of regular
agriculture serves as a cornerstone of India’s economy, irrigation. This review specifically focuses on the European

526 Copyright © IEEE–2024 ISBN: 979-8-3503-8749-0

 Deep Learning Image Classification Abiotic Plant Disease Detection Classification Using MobileNet

region, which encompasses diverse climatic zones, ranging the multilabel feature technique. It identifies and diagnoses
from arid regions the impact of abiotic stressors on crop the disease by various deep learning techniques applied to
production in Europe is crucial given its varied climatic that dataset. The author developed a model by used CNN,
conditions and its importance in global agricultural markets Xception and Dense Net. It achieves accuracy in terms of
[4]. This paper proposes and studies deep learning models disease detection prediction at a higher rate. Pradeep, A.R.
such as convolutional network models for federate learning et. al [9] proposed a work for plant disease detection and
for disease detection in plants. In this work, samples of developed a model using an efficient multilabel disease
different leaves are collected from the village and this detection method and achieve accuracy at for higher rate.
sample is taken from an online repository [5]. The work is
III. Methodology
organized into many sections.
In this Proposed work provided a comprehensive
II. Literature Review overview of methodologies, models, and experiments
The abiotic disease found in plants is due to the eating related to plant disease classification using deep learning
conditions and the soil erection. The existing research techniques. Here’s a summary based on the provided text.
in disease detection of plants computed with the help 1) Problem Statement
of following literature review that easily evaluated the
The focus is on classifying plant diseases using image
proposed work.
datasets. This involves distinguishing between vigorous
Li, L Zhang et. al [1] have developed a model for
and polluted plants, as well as identifying specific diseases
plant disease detection and recognition by using AI and
within plant classes.

n
deep learning models. It identifies and extracts the various
features of plants by mapping leaf disease using deep 2) Challenges Addressed
y
learning. The various features are mapped by imaging
selection techniques and obtained high-resolution images
•

i O o n l
Long training durations and the need for large
datasets in CNN models.
to detect the disease and insects in leaves. Dawod, R. et. al
[2] have proposed a model on leaf accuracy in the sense of s•

r g
Complex Feature extraction requires deep CNN
models.
classifying and identified the plant disease images by their
e • Transfer learning is employed to mitigate these
i n
V
classification tools and feature color that is primary and challenges by utilizing pre-trained networks.
secondary features for the leaf using deep learning models. d
3) Multi-Class Classification
a•
ft
Khan, M. A. et. al [3] have studied & implemented to compute
the segmentation models on the image enhancement. It used
r e Images of healthy and infected plant samples are

ra
f
mapped to specific classes.
the SHSB method and deep learning models such as VGG
and VGG-M19 for feature selection and extraction. It trains
o o • For example, the banana class may have subclasses
representing different diseases like Xanthomonas

D r Pr
and tests the model on the entropy and obtains an accuracy wilt, fusarium wilt, etc.
of 98.08% for approx. 10.52 seconds. Scientist D. et. al [4]
used the neural network and deep learning to develop a
4) Transfer learning facilitates the classification of
target images which extracts the different images as

Fo
model. With the help of CNN, the detection of potato disease
in plants achieves an accuracy of 99 % and 839,203 by their per their features.
trainable parameters using training time. Panigrahi, K. P. et. 5) Transfer Learning Approach
al [5] have developed a CNN model for detecting the corn • Utilizing pre-trained CNN models significantly
leaf images to rectify linear unit activation function. It used reduces training time and improves accuracy
the Adam optimizer by various parameters and activation compared to training from scratch.
functions. The author used the evaluation parameters for the • These models are trained by MobileNet and use
detection of the disease and achieved an accuracy of 98.78%. the plant dataset, which contains a vast array of
Aldhyani T. H. et. al [6] identified a technology for feature images across different categories.
extraction using artificial intelligence by implementing
classification techniques. The field of agriculture for the 6) Description of Models
crop disease detection develops a model using CNN for • ResNet-50: A CNN model with 50 deep layers,
the dataset. It trained and tested the model with the help utilizing residual connections to mitigate vanishing
of ML techniques and their classifiers. Panigrahi, K.P. et. gradient issues.
al [7] proposed a work by using various machine learning • VGG-16: A deep convolutional network with
techniques for plant disease detection by classification 16–19 weight layers, known for its simplicity and
techniques. It analyzes and compares the model with the effectiveness.
higher accuracy prediction classification techniques and • DenseNet-121: Utilizes dense connections between
obtained an accuracy of 79.23%. Mohsin Kabir et. al [8] layers, allowing each layer to receive input from
proposed a model for the recognition of plant disease using all preceding layers.

Copyright © IEEE–2024 ISBN: 979-8-3503-8749-0 527

IEEE Conference ID: 62700 13th – 15th Nov. 2024
Proceedings of the 2024 4th International Conference on Technological Advancements in Computational Sciences

• Inception V4: Incorporates inception modules with its simplicity and effectiveness, consists of 16–19 weight
varying kernel sizes to handle local and global layers. DenseNet-121 utilizes dense connections between
features efficiently [6]. layers, while Inception V4 incorporates inception modules
7) Experiments with diverse kernel sizes to efficiently handle local and
global features [8]. Experimental evaluations are conducted
• Conducted on a GPU NVIDIA GeForce GTX
on a GPU NVIDIA GeForce GTX workstation using
workstation with specific software and libraries.
specific software and libraries. Each model’s training
• Training and testing accuracy, as well as loss, were
and testing accuracy, as well as loss, are evaluated using
evaluated for each model.
the PlantVillage dataset, which comprises various plant
• PlantVillage dataset comprising 38 classes and
disease categories. The dataset is divided into training,
54,305 images was used, split into training,
validation, and testing sets to facilitate model evaluation
validation, and testing sets.
[9]. In summary, the study delves into methodologies,
• Pre-trained models were fine-tuned using 80%
model descriptions, experimental setups, and dataset details
of the dataset for training and evaluated on the
pertinent to plant disease classification using deep learning
remaining 20%
techniques, with a particular emphasis on transfer learning
8) Dataset Description approaches [10].
• The PlantVillage dataset contains various plant B. Preprocessing and Tuning of Parameters by
disease categories with a significant number of
Training
images.
Transfer learning offered accelerated learning compared

n
• Split into training, validation, and testing sets
to building models from scratch. Its standardized parameters
to facilitate model evaluation. Overall, the text
y
provides a detailed overview of methodologies,
i O o n l
for the different models and used the initial learning rate
0.001 for SGD optimization. Each model underwent 30

s
models, experiments, and dataset descriptions
epochs, with a fixed dropout rate of 0.5. Convergence was

r g
involved in plant disease classification using
deep learning techniques, with a focus on transfer observed after a few iterations, mitigating overfitting and
learning approaches [7].
e n
degradation concerns.
i
A. Proposed Model and Techniques Used
V C. Network_Architecture Method

a d network
t fr
The application of deep learning techniques for plant The model used for pretraining based

f e
disease classification poses several challenges despite requirements and the suitability of the classification disease

a o
the suitability of Convolutional Neural Network (CNN) model. The relevant features are selected by filtering the

r
models. One significant challenge is the extensive time special features by mapping the value. It utilized actual pre-

ro
trained network models with their respective convolution

DrP
required for model training, exacerbated by the necessity
for large datasets of samples 17860. Addressing these layer configurations and filter sizes [12] as shown in Fig.
issues necessitates the utilization of deep CNN models to 1. and Fig. 2.
extract intricate low-level features from images. However,

Fo
this exacerbates the complexity of model training. To
mitigate these challenges, transfer learning emerges as a
viable solution. Transfer learning leverages pre-trained
networks, enabling the transfer of learned model parameters
from one dataset to another. In this context, we explore
the methodologies employed in this study. Multi-class
classification involves categorizing images of healthy and
diseased plants into distinct classes. For example, within the
banana class, images of various diseases like Xanthomonas
wilt, fusarium wilt, etc., are categorized. During testing,
the model utilizes features extracted from source images
to accurately classify target images. Transfer learning
significantly reduces training duration and enhances
accuracy compared to training CNN models from scratch.
Various pre-trained models such as ResNet-50, VGG-16,
DenseNet-121, and Inception V4 are utilized, all trained
on the extensive ImageNet dataset, ResNet-50 a CNN
model with 50 layers, employs residual connections to
tackle vanishing gradient problems. VGG-16, known for Fig. 1: Predicted AMD_Category

528 Copyright © IEEE–2024 ISBN: 979-8-3503-8749-0

 Deep Learning Image Classification Abiotic Plant Disease Detection Classification Using MobileNet

7: End for Rt+1>=0

8: End G=0
To classify the images for different samples, transfer
learning process is computed for the different samples of
leaves the first three layers are trained as per the model. For
the training process freezing the last and first layers by tuning
the different parameters. In this model feature extraction
and classification are employed. Different CNN deep
learning models are used such as VGG-16, DenseNe-121,
ResNet-50, and Mobile Net is employed. In this proposed
work the basic approach that is transfer learning process
is implemented to are collected for different plants. As per
the different datasets of plants train and test the plant to
compute the validation results as per the plant selection as
shown in table 1.
Table 1: PLANT Dataset for Different Classes
Type_of_Plant_Diseases_ Training Test Validation
Fig. 2: Proposed Methodology Classes Results

D. Tuning Details for Specific Models

n
Crop_Type Samples Samples Samples Samples

ly
Detailed specifications were provided for VGG-16,

o
Potato_early_grow 1500 710 200 80

i O
Inception V4, ResNet-50, and DenseNet-121[13]. For
instance, VGG-16 featured an input image dimension of
healthy_Potato 950
n 610 90 70

224x224x3, with specific layer configurations and filter

sizes outlined. Similar detailed descriptions were provided s
blight_Potato_atlate

r g
132 112 12 10

e
healthy_Raspberry 454 186 26 25
for Inception V4, ResNet-50, and DenseNet-121[14].
i n
V
Soybean_healthy 3226 3222 405 345
IV. Results & Discussion
d
powdery_mildew_

t fr ea
1438 1328 154 134
In this proposed study various deep learning models are Squash

f
employed. In this paper transfer learning approach is used

a o
to detect and diagnose plant disease. The 17860 samples
Strawberry_healthy 1112 787 101 97

r
of dataset are pre-trained by the deep learning models. leaf_scorch_
305 345 36 21

ro
Strawberry

DrP
Different function values are to be set to get the optimization
rate such as the learning rate of 0.01 and dropout function bacterial_spot_
2016 1523 202 171
Tomato
rate of 0.5 [15]. For computing better results split the dataset
as in testing and training datasets Actual learning rate and

Fo
early_blight_
1000 810 100 90
other function values computed by performance metrics. Tomato

A. Pseudocode for Client_Updates healthy_Tomato 1591 1546 191 172

1: Pro_ for Client_update (X, W) late_blight_Tomato 1509 620 96 76
2: DC← (Split Tk into batch_of_size DC) Tomato _leaf_
3: Forevery local_epochs I from 1 to ED do 856 1112 158 140
mold_Tomato
4: Forevery_batch_b in DCdo
Septoria _leaf_
5: W←W—PRql(W, h) spot_Tomato
1771 1357 168 151
6: Endfor
spider_mites_two-
7: Endfor 1224 1235 146 123
Tomato
8: ReturnWtoServer
9: Endprocedure Tom_target ratio 1203 435 325 395

B. Pseudocodefor Server_Updates virus_infected in

332 312 32 25
tomato
1: Pro_Server_Executes
2: Initialize T0 yellow_virus__
3109 1154 110 122
curl_virus_Tomato
3: Foreachiteration u=1,2,...,Ddo
4: Rt←G
In the below table 2 the various parameters and the
5: ForeachclientkinStinparalleldo
epochs values are to be tuned and computed with the
6: Rt+1←Client_updates(R,Tt)
attributes for obtained better results.

Copyright © IEEE–2024 ISBN: 979-8-3503-8749-0 529

IEEE Conference ID: 62700 13th – 15th Nov. 2024
Proceedings of the 2024 4th International Conference on Technological Advancements in Computational Sciences

Table 2: Parameters_ Hypertuned Results to increase accuracy and effectiveness of plant disease
Hyper Tuned _values Ch_Ephocs prediction some other model of deep learning and dataset of
Dropout_ rate 0.5 various plants are collected and optimizing it.
Epochs _rate 35 References
Active value ReLu [1] Contribution of Agriculture Sector Towards GDP Agriculture has
been the Bright Spot in the Economy despiteCOVID-19. Available.
Regularization_Method Normalization methods as in batch form
online: https://www.pib.gov.in/indexd.aspx(2022).
Optimizer-result (SGD) [2] Li, L.; Zhang, S.; Wang, B. Plant Disease Detection and Classification
LR Rate 0.001 by Deep Learning— A Review. IEEE Access2021,9,56683–56698.
[3] Dawod, R.G.; Dobre, C. Upper, and Lower Leaf Side Detection with
Output_classes 34
Machine Learning Methods. Sensors 2022, 22, 2696.
As shown in table 3 the various deep learning models [4] Khan, M.A.; Akram, T.; Sharif, M.; Javed, K.; Raza, M.; Saba, T.
with their different parameters are employed and computed An automated system for cucumber leaf diseased spot detection and
classification using improved saliency method and deep features
results as per following table 4 shows.
selection. Multimedia. Tools Appl.2020, 79,18627–18656.
Table 3: Deep Learning Models [5] Scientist, D.; Bengaluru, T.M.; Nadu, T. Rice Plant Disease
Model_used_ Identification Using Artificial Intelligence Inc. Electr. Eng.
VGG Incp4 ResNet-50 DenseNet-121 Mobile_Net
deep learning Technol.2020,11, 392–402.
Tot_layer 18 22 50 121 14 [6] Panigrahi, K.P.; Sahoo, A. K.; Das, H. A. CNN Approach for Corn
Leave Disease Detection to support Digital Agricultura l System.
Max_ pool layers 7 7 2 6 5 In Proceedings of the 4th International Conference on Trends in

n
DenseNet_layer 3 3 4 5 6 Electronics and Information, Tirunelveli, India, 15–17 June 2020;

y
pp.678–683.

o
Drop_out_layer 1 2 2 3 3

l
i O
[7] Plant Village. Available online: https://plantvillage.psu.edu/

n
Flatten_ layer 1 1 1 2 1
(accessed on 29 September 2022).

s
1 × 1, [8] Aldhyani, T. H.; Alkahtani, H.; Eunice, R.J.; Hemanth, D.J.

r g
Filter_size, 3 × 3, 3 × 3,
2×2 3 × 3, 3×3 Leaf Pathology Detection in Potato and Pepper Bell Plant using
Stride_rate 1×1 2×2
5×5

e
Convolutional Neural Networks. In Proceedings of the 2022 7th
Train_parameters 46.2 92.6 26.9 8.69 98.9
i n
International Conference on Communication and Electronics Systems

V
(ICCES), Coimbatore, India,22–24 June 2022; pp.1289–1294.
As the below table 4 deep learning models such as [9]

a d
Panigrahi, K.P.; Das, H.; Sahoo, A.K.; Moharana, S.C. Maize

ft
Resnet, dense net etc. are computed for obtaining better leaf disease detection and classification using machine learning
results for the disease detection in plants leave precisely.
r e algorithms. In Progress in Computing, Analytics and Networking;

ra of
Springer: Singapore,2020.
The current model mobile net is implemented to achieve [10] Mohsin Kabir, M.; Quwsar Ohi, A.; Mridha, M. F. A Multi plant
accuracy as 99.88 after set for training and testing. disease diagnosis method using convolutional neural network.

DrP ro
Table 4: Training and Testing Data Accuracy and Loss arXiv2020, arXiv:2011.05151.
[11] Pradeep, A.R.; Hoque, A.M.; Kabir, M.M.; Rahman, M.S.; Mridha,
Deep learning_ Loss_ Percentage for Loss M.F. Plant Disease Identification from Leaf Images using Deep
Training_set
model sets and Accuracy CNN’s Efficient Net. In Proceedings of the 2022 International

Fo
Loss Accuracy Conference on Decision Aid Sciences and Applications (DASA),
Model Used Accuracy (%)
(%) (%) loss (%) Chiangrai, Thailand,23–25 March 2022; pp. 523–527.
[12] Onesimu, J. A.; Karthikeyan, J. An Efficient Privacy-preserving
InCP_4 89.79 0.12 93.29 0.0529
Deep Learning Scheme for Medical Image Analysis. J. Inf. Technol.
VGG-16 74.29 0.62 85.78 0.74 Manag.2020,12,50–67.
[13] Aggarwal, R., Aggarwal, E., Jain, A., Choudhury, T., & Kotecha,
ResNet-50 96.82 4.72 88.75 0.017 K. (2023, November). An Automation Perception for Cotton Crop
Disease Detection Using Machine Learning. In 2023 7th International
DenseNet-121 83.89 0.018 89.83 0.0167
Symposium on Innovative Approaches in Smart Technologies
Mobile_Net 99.88 0.019 99.9 0.0165 (ISAS) (pp. 1-6). IEEE.
[14] Mhathesh, T.S.R.; Andrew, J.; Sagayam, K.M.; Henesey, L. A3 D
V. Conclusion Convolutional Neural Network for Bacterial Image Classification. In
In this proposal, various samples of leaves are collected Intelligence in Big Data Technologies—Beyond the Hype; Springer:
Singapore, 2021; pp. 419–431.
to detect the disease in plants. In the agriculture field, deep [15] Aggarwal, R., & Kumar, S.; Missing Value Imputation and
learning is an advanced technology by which the automated Estimation Methods for Arrhythmia Feature Selection Classification
perception and detection of disease is possible to find and Using Machine Learning Algorithms, Machine Learning Methods
compute as in technology. The samples of leaf are analyzed for Engineering Application Development (2022) 1: 145https://doi.
by training and testing on the different values. In future, org/10.2174/9879815079180122010013.

530 Copyright © IEEE–2024 ISBN: 979-8-3503-8749-0