Plant Leaf Disease Detection using

Transformer for Sustainable Agriculture

Report
For Evaluation 2

Bachelor of Technology
Computer Science and Engineering

Submitted By
Harsh Gupta (2021369516)
Khushi Rani (2021440769)
Rounaq Parvez Bhat (2020407187)

September 2024

Under the mentorship of

Dr.Pawan Kumar Verma
(Associate Professor)

Department of Computer Science

and Engineering
School of Engineering and
Technology
Sharda
University,
Greater Noida
Problem Statement:
Traditional methods of plant disease detection are labor-intensive and prone to error, while existing
automated systems based on convolutional neural networks struggle to capture complex spatial
dependencies in plant leaf images. To address these limitations, there's a need for a robust and
scalable solution leveraging Transformer models to accurately detect plant diseases, overcoming
challenges such as complex spatial patterns, limited labeled data, interpretability, and scalability to
diverse plant species and disease types.
Key Problem:
The question is: How can we design an AI-based system that automates the detection of plant
diseases and provides improved predictions using image data?

Team:

S.N Name Branch System

O ID
1. Harsh Gupta CSE 20213695
16
2. Rounaq Parvez Bhat CSE 20204071
87
3. Khushi Rani CSE 20234407
69

Guide:
Dr. Pawan Kumar Verma
(Associate Professor, Computer Science and Engineering)

Outcome:
Research Paper / Hackathon Project
 TABLE OF CONTENT

Sr Conte Pa
. nt ge
N no.
o.

1 Abstract 01

2 Motivation 02

3 Literature Review 03

4 Problem Formulation/Objectives 05

5 Methodology 06

6 Software/Hardware required 07

7 References 08
 ABSTRACT

In the context of sustainable agriculture, effective plant disease detection plays a crucial role in
maintaining crop health and optimizing yield. Leveraging advancements in deep learning, this
project proposes a approach using transformer-based models for automated detection of plant leaf
diseases. The methodology involves collecting a diverse dataset of labeled images encompassing
healthy leaves and leaves affected by various diseases. Pre-processing techniques standardize the
images and augment the dataset to enhance model robustness. A transformer- based architecture,
adapted from natural language processing tasks to image classification, is trained on the dataset using
supervised learning. Through rigorous validation and fine-tuning the model achieves high accuracy in
distinguishing between healthy and diseased leaves. Upon deployment, the model offers real-time inference
capabilities, facilitating timely intervention in agricultural practices. Continuous monitoring and feedback
mechanisms ensure the model's adaptability to evolving environmental conditions and emerging diseases
This transformative approach not only aids in sustainable crop management but also reduces reliance on
chemical treatments, promoting environmentally friendly agricultural practices. Our method involves pre-
training a Transformer model on a large dataset of plant leaf images, utilizing self-attention
mechanisms to learn discriminative features. We then fine-tune the model on a smaller dataset
specific to the target plant diseases, enabling it to recognize disease patterns with high accuracy. By
employing Transformers, we aim to enhance the robustness and generalization capability of disease
detection systems, thereby contributing to sustainable agriculture practices. Furthermore, we
explore the integration of Transfer Learning techniques to alleviate the need for extensive labeled
data, particularly beneficial in scenarios where annotated datasets are limited. Through experiments
conducted on real-world plant disease datasets, we demonstrate the effectiveness of our proposed
approach in accurately identifying various leaf diseases across different plant species.
The proposed method offers several advantages over traditional CNN-based approaches, including
improved feature representation, enhanced interpretability, and scalability to diverse plant species
and disease types. By leveraging advanced deep learning architectures like Transformers, we strive
to facilitate early and accurate detection of plant diseases, ultimately promoting sustainable
agricultural practices and ensuring global food security.
 MOTIVATION

The motivation behind developing a project for plant leaf disease detection using transformers in
sustainable agriculture is multifaceted. Firstly, it addresses the critical issue of crop health
management. Plant diseases can cause substantial yield losses and economic strain for farmers,
highlighting the urgent need for early detection and intervention. By automating this process, we
aim to provide a proactive solution that enables timely action to mitigate losses and ensure food
security. Moreover, traditional disease management practices often rely on chemical treatments,
which can have detrimental effects on the environment, soil, and human health. By shifting towards
non- invasive, automated detection methods, we can reduce reliance on such treatments and
promote environmentally friendly agricultural practices. Additionally, manual inspection of crops
for disease symptoms is labor-intensive and prone to errors. Leveraging advanced deep learning
algorithms offers a more efficient and precise alternative, allowing for early detection of diseases
even before visible symptoms appear. Furthermore, the emergence of transformer-based models,
originally developed for natural language processing, presents an opportunity to harness cutting-
edge technology for image classification tasks in agriculture. This convergence of technology and
agriculture aligns with the broader goal of enhancing global food security by improving crop health
management and productivity sustainably. In summary, the motivation for developing a project for
plant leaf disease detection using transformers lies in addressing challenges related to crop health,
environmental sustainability, efficiency, precision, technological advancements, and global food
security.

2
 LITERATURE REVIEW

Introduction:
Plant diseases pose significant challenges to global food security, leading to substantial yield losses
and economic repercussions for farmers. Timely and accurate detection of these diseases is crucial
for effective disease management and sustainable agricultural practices. In recent years, there has
been growing interest in leveraging deep learning techniques, particularly transformers, for plant
leaf disease detection. This literature review aims to explore the current state of research in this
field, highlighting methodologies, findings, and future directions.
Overview of Plant Leaf Disease Detection:
Traditionally, plant leaf disease detection relied on manual inspection by agronomists, which is
labor- intensive, time-consuming, and subjective. With advancements in computer vision and deep
learning, automated disease detection systems have emerged as a promising alternative. These
systems analyze images of plant leaves to identify disease symptoms, enabling early intervention
and precision agriculture practices.
Transformer-Based Approaches:
Transformers, originally developed for natural language processing tasks, have shown promise in
computer vision applications, including image classification and object detection. In the context of
plant leaf disease detection, transformers offer several advantages, such as the ability to capture
long- range dependencies in images.
Comparison with Other Approaches:
Several studies have compared transformer-based models with traditional approaches, such as
convolutional neural networks (CNNs) and handcrafted feature-based methods. While CNNs have
been widely used for image classification tasks, transformers have shown superior performance in
capturing global image context and handling variable-sized inputs. Additionally, transformers offer
interpretability through attention mechanisms, allowing for better understanding of model
predictions.
Challenges and Future Directions:
Despite their potential, transformer-based approaches for plant leaf disease detection face
challenges such as computational complexity, scalability, and the need for large annotated datasets.
Future research directions include exploring techniques to mitigate these challenges, incorporating
multi- modal data sources for enhanced disease detection, and developing user- friendly tools for
real-world deployment in agricultural settings.
Conclusion:
In conclusion, plant leaf disease detection using transformers represents a promising approach for
sustainable agriculture. By leveraging deep learning techniques, particularly transformers,
researchers can develop accurate, efficient, and interpretable disease detection systems that
empower farmers to make informed decisions and mitigate yield losses. Continued research in this
field is essential to address remaining challenges and realize the full potential of transformer-based
approaches in plant health monitoring.
Review of literature
1. Image classification using Boosted support vector machine-based Arithmetic optimization
algorithm (BSVM-AOA) M Prabhu et al. (2022) [1]

The implementation is done in Python. The approach combines a boosted Support Vector Machine
(SVM) with the Arithmetic Optimization Algorithm (AOA). Boosted SVM: Enhances the SVM's
performance by combining multiple weak classifiers to form a strong classifier. Arithmetic
Optimization Algorithm (AOA): An optimization technique used to fine-tune the parameters of the
SVM for better accuracy in disease detection.
Methodology: Boosted Support Vector Machine (SVM): The study employs a boosted SVM model,
which enhances the classification performance by combining multiple SVM classifiers. This approach
is beneficial for handling complex datasets typically found in plant disease detection. Arithmetic
Optimization Algorithm (AOA): The paper introduces an Arithmetic Optimization Algorithm to
optimize the SVM parameters. AOA is inspired by mathematical arithmetic operators and aims to find
the best solution by balancing exploration and exploitation in the search space. The primary dataset
used in this study is comprised of images of plant leaves affected by various diseases. These images are
sourced from publicly available plant disease datasets such as the PlantVillage dataset, which contains
a diverse collection of images for training and testing purposes
Result: The accuracy of the proposed BSVM-AOA technique obtains excellent performance with an
accuracy rate of about 98.6% while the InceptionResNetV2-CNN, Random Forest, NSGA-II, H2K
methods obtain 95.67%, 79%, 78.23%, and 96% respectively in which the values are lower than the
proposed technique. It revealed that the proposed BSVM-AOA technique achieves a higher percentage
of accuracy than other techniques. The integration of the Boosted Support Vector Machine (SVM) with
the Arithmetic Optimization Algorithm (AOA) significantly improved the accuracy of plant leaf
disease detection. This was demonstrated through comprehensive experiments and comparisons with
traditional SVM and other machine learning techniques

2. Detection of Rust Disease in Wheat Crop Using Convolutional Neural Networks via Transfer
Learning Technique Mikhail Genaev et al (2021) [2]

Algorithm: The study likely used Python as the primary programming language, along with libraries
such as TensorFlow or PyTorch for developing the neural network models.Hardware: High-
performance computing resources, including GPUs, would be necessary to handle the computational
load required for training deep learning models on large datasets.The core algorithm used is a
convolutional neural network (CNN) based on the DenseNet architecture. DenseNet is known for its
efficiency in image recognition tasks due to its dense connectivity between layers. They used the
Densenet201 model from the Cadene repository (https://github.com/Cadene/pretrained-
models.pytorch) pre-trained on 1000 classes of the ImageNet dataset.
Methodology: The researchers used transfer learning, which involves fine-tuning a pre-trained model
on a new dataset to leverage existing knowledge and improve performance on the specific task of rust
disease detection.
Techniques: Image preprocessing techniques were likely employed to enhance the quality of the input
images, and data augmentation methods were used to increase the diversity of the training data, thereby
improving the model's robustness. The dataset includes 876 images: 143 images of healthy wheat; 359
images of leaf rust; 377 images of stem rust. The rest of the data comes from publicly available images
found in Google Images database.
Result: The CNN model demonstrated high classification accuracy, with an F1 score of 0.9 and an
AUC (Area Under the Curve) of 0.98, indicating its effectiveness in distinguishing between healthy
wheat plants and those affected by leaf rust or stem rust. The study highlighted the potential of neural
networks in early disease detection, which is crucial for preventing the spread of rust diseases and
mitigating crop losses.

3. Lightweight Inception Networks for the Recognition and Detection of Rice Plant Diseases
Junde Chen , Weirong Chen, Adan Zeb , Shuangyuan Yang, and Defu Zhang (2022) [3]

Methodologies Used: Lightweight Inception Network (MobInc-Net): The study introduces MobInc-
Net, a lightweight convolutional neural network (CNN) designed for efficient and accurate recognition
of rice plant diseases. The network integrates Inception modules to capture features at multiple scales,
enhancing detection capabilities while keeping the model size manageable.Transfer Learning: Pre-
trained models were fine-tuned on rice disease datasets to leverage existing knowledge and improve
detection accuracy.Data Augmentation: Techniques such as rotation, flipping, and scaling were applied
to enhance the diversity of the training data.
Key Findings: High Accuracy: The MobInc-Net achieved an average accuracy of 99.21% on public
datasets and 97.89% on local datasets, demonstrating its effectiveness in real-world scenarios.
Efficiency: The model's lightweight design allows it to be deployed on devices with limited
computational resources, such as smartphones, making it accessible for on-site disease detection in
agricultural fields.
Broad Applicability: The model's robustness was confirmed across various rice diseases, showcasing
its potential for wide-scale agricultural use.These findings underscore the potential of MobInc-Net as a
practical tool for farmers and agricultural professionals to monitor and manage rice crop health
efficiently.

4. Deep learning system for paddy plant disease detection and classification Amritha Haridasan,
Jeena Thomas and Ebin Deni Raj (2022) [4]

Methodology Used in the paper "Deep Learning System for Paddy Plant Disease Detection and
Classification" involves the following steps: Image Acquisition: Collecting images of paddy plants
affected by various diseases. Preprocessing: Enhancing the image quality through techniques like
resizing and normalization. CNN Model Development: Utilizing Convolutional Neural Networks
(CNNs) to extract features and classify diseases. The methodology for detecting and categorizing rice
diseases includes image acquisition, data augmentation and data labeling, image pre-processing
segmentation, feature extraction, and classification. Brown leaf spot, leaf blast, sheath rot, false smut,
and bacterial blight are the five primary diseases of paddy crops that this work focuses on.
Key Findings: Methods such as SVM and CNN, as well as a few key image processing, image
segmentation, and feature extraction approaches, were used to accurately predict diseases impacting
rice harvests. The results reveal that CNN outperforms SVM in terms of performance measures like
validation accuracy and test accuracy. The performance of the models could be improved in the future
by including the most recent and advanced deep learning techniques like transfer learning.

5. GTCLC: leaf classification method using multiple descriptors Cem Kalyoncu1 , Önsen Toygar
(2016) [5]

Methodology Used: The authors propose Geometric, texture and color based leaf classification, a
novel leaf classification method using a combination of geometric, shape, texture and color features
that are extracted from the photographic image of leaves. This method combines features that
complement each other to define the leaf. A new local binary pattern (LBP) variant, namely sorted
uniform LBP is also proposed for leaf texture description.
The paper describes the use of multiple image descriptors (e.g., shape, texture, color) to capture the
diverse characteristics of leaves. Common descriptors might include Local Binary Patterns (LBP),
Histogram of Oriented Gradients (HOG), and color histograms. The process of extracting features from
leaf images using these descriptors would be detailed. The authors might have used standard datasets
or collected their own for this purpose.
Key Findings: This section contains experimental results to demonstrate the superiority of the
proposed GTCLC method. Accuracy experiments are performed on three different databases.
Processing speed and detailed descriptor accuracy tables are computed from a single database. All
reported results are the average of 100 simulations and the expected variance is no more than 0.1%.
According to the experiments demonstrated in, the proposed method has superior accuracy compared
with the other state-of-the-art methods. They have also demonstrated the performance of proposed
sorted

6. Tomato Leaf Disease Identification by Restructured Deep Residual Dense Network

CHANGJIAN ZHOU et al. (2021) [6]

Methodology used: The paper discusses a specialized architecture combining deep residual networks
(ResNet) and dense networks (DenseNet). In this paper, the researchers proposed RRDN to solve the
problem of crop leaf disease identification. As the original model was used in image super-resolution,
the input images have no dimension reduction operation, which may work well in a single block. But
in the image classification task, tens of thousands of images are input, which will require considerably
more computing resources, as well as low efficiency. So in this paper, the input image is convolved
first in Res-DenseBlock (RDB) and the tensor is batch normalized after the convolution in the RDB
block.
Key findings: Experiments show that the RRDN can achieve satisfactory performance on the tomato
dataset as high as 95%; the results show that the method can improve the identification accuracy on the
tomato leaf diseases dataset. This model is only suitable for only plant leaf dataset

7. LAD-Net: A Novel Light Weight Model for Early Apple Leaf Pests and Diseases Classification
Xianyu Zhu , Jinjiang Li, (2023) [7]

Methodology Used: The paper introduces LAD-Net, a lightweight model designed for the early
detection and classification of apple leaf pests and diseases. The model focuses on achieving high
accuracy with minimal computational resources, making it suitable for mobile devices. The approach
includes several advanced techniques in deep learning, such as convolutional neural networks (CNNs)
optimized for small size and high performance, which allows for real-time disease recognition.
Key Findings: The LAD-Net model achieved a recognition accuracy of 98.58% while maintaining a
model size of only 1.25 MB, making it significantly more efficient than existing models. This makes it
an effective tool for early diagnosis of apple leaf diseases, potentially improving yield and reducing the
spread of pests and diseases.

8. Performance analysis of AI-based solutions for crop disease identification, detection, and
classification Divyanshu Tirkey , Kshitiz Kumar Singh, Shrivishal Tripathi (2023) [8]

Methodology used: This paper establishes a crop disease and insects identification model in split
segments. A system is developed based on image recognition using AI. The proposed solution will
help farmers apply insecticides accurately and on time. The proposed approach will provide the farmer
with precise information and reduce operational costs, improving crop yield and quality. The proposed
system architecture is divided into parts: system background, gathering dataset, data pre-processing,
and training deep-learning models to identify crop types and insects. The dataset is segmented to
remove the noise. One significant use of machine learning and computer vision in agriculture is the
identification and classification of crop diseases. The objective is to create algorithms based on images
of leaves or other plant elements that can automatically identify and categorize diseases in crops like,
InceptionV3 model, CNN model, YoloV5 model
Key Findings: The proposed algorithms’ results in various performance metrics such as accuracy,
mAP, precision, confusion matrices, etc., and illustrated their practical importance. After examining
the findings and observations of the trained models, it can be concluded that models trained on
segmented pictures outperform color and gray-scale images. The images have very little noise,
contributing to the models’ high accuracy.

9. Transfer Learning for Multi-Crop Leaf Disease Image Classification using Convolutional
Neural Network VGG Ananda S. Paymode , Vandana B. Malode [2022] [9]

Methodologies used: The study employs transfer learning by fine-tuning the VGG16 Convolutional
Neural Network, originally trained on the ImageNet dataset, to classify leaf diseases across multiple
crops. The researchers used a dataset of diseased leaf images, which were resized to fit VGG16's input
size. Data augmentation techniques, such as rotation, zooming, and flipping, were applied to enhance
the dataset's variability. The model's final layers were retrained on the specific task of leaf disease
classification, aiming to leverage pre-existing knowledge while adapting to the new dataset. The
model’s performance was evaluated using metrics like accuracy, sensitivity, specificity, precision,
recall, F1-score.
Key Findings: The proposed model achieved an accuracy of 98.40% for grapes and 95.71% for
tomatoes. The research demonstrated significant improvements in the classification of leaf diseases,
which can directly support agricultural productivity.
Datasets: The datasets were compiled from various sources, including publicly available datasets
(Plant Village) and images collected under controlled conditions . The datasets include images of leaves
that are affected by different as well as healthy
Limitations: The study was limited to specific crops (grapes and tomatoes), which may not generalize
well to other crops without further training. The model’s performance might be affected by variations
in environmental conditions such as lighting and background, which were not extensively covered in
the study. Training deep learning models like VGG requires significant computational resources, which
might be a constraint for widespread adoption.

Image processing based system for the detection, identification and treatment of tomato leaf
diseases. Sami Ur Rahman [2022] [10]

Methodologies: Images of tomato leaves affected by various diseases were collected from local
tomato crop fields. Feature Extraction: Thirteen different statistical features were calculated from the
images using the Gray Level Co-occurrence Matrix (GLCM) algorithm. Classification: The extracted
features were classified into different diseases using a Support Vector Machine (SVM) classifier.
Disease Recognition: The processed leaf images were compared with stored features to recognize the
disease. Implementation: The proposed method was implemented as a cell phone application for
practical use.
Key Findings: The proposed method achieved an accuracy of 100% for healthy leaves, 95% for early
blight, 90% for septoria leaf spot, and 85% for late blight. Practical Application: The implementation
of the method as a cell phone application makes it accessible for real-time use by farmers and
agricultural professionals.
Dataset: The dataset used in this study consisted of images of tomato leaves affected by various
diseases, collected from local tomato crop fields.
Limitations: The dataset size may limit the model’s ability to generalize to different conditions.
Environmental Variability: Variations in lighting and background conditions could impact detection
accuracy. Computational Resources: The method requires significant computational resources for
feature extraction and classification. Specific Disease Focus: The model was trained specifically on a
 limited set of tomato leaf diseases, which may limit its applicability to other crops and diseases.

Table for Literature Review

Reference Year Methodology Dataset Outcomes/Findings Limitations
[1] 2022 Boosted Support Plant Village BSVM-AOA technique High
Vector Machine (multiple crop) achieves a higher Computational
and Arithmetic percentage of accuracy Cost &
Optimization (95.67%) than other Low Recognition
Algorithm techniques. Speed

[2] 2021 CNN based on Google Images Achieved an accuracy of Unfit for handle
DenseNet Databases 90.6% large dataset
architecture (wheat crop)

[3] 2022 MobInc-Net, a Plant Village Achieved an avg Less efficient &
lightweight CNN (rice crop) accuracy of 95.62% no real-time
applications
[4] 2022 Integration of Real-world Achieved an accuracy of Large model
SVM with CNN paddy fields in 91.45% Volume & High
Kerala computational
complexity
[5] 2016 Local Binary ICL database, ICL database - 83% Environmental
Patterns (LBP), Flavia, & Flavia - 95.1% variability &
Histogram of Swedish Leaf Swedish – 98.1% feature
Oriented Gradients database dependency
(multiple leaf)
(HOG),

[6] 2021 Combination of Dataset from RRDN achieved an Trained and

deep residual AI Challenger accuracy of 95% tested on specific
(ResNet) and dense 2018 (Tomato dataset which
network leaf) limits its
(DenseNet) generalizability

[7] 2023 LAD-Net utilizes Real world High accuracy 98.58% Performance on
AD Convolution Apple orchards The model is lightweight other platform
and LAD- with a size of only and in different
Inception to 1.25MB environmental
enhance the Real Time Performance conditions were
extraction of multi- on mobile device not tested
scale features
 [8] 2023 Convolutional Real world Test accuracy, Inception Low recognition
Neural Networks soybean field V3 model – 95% speed
(CNNs), CNN – 97%
InceptionV3, and YOLOv5 – 98.75
YOLOv5

Reference Year Methodology Dataset Outcomes/Findings Limitations

[9] 2022 Transfer learning Plant village Accuracy of 98.40% for High
with VGG16 and images grapes and 95.71% for computational
Convolutional collected in tomatoes resources
Neural Network controlled Performance
conditions. affected by
variations in
environmental
conditions

[10] 2022 Gray Level Co- Local Tomato Accuracy of 100% for Specific Disease
occurrence Matrix field healthy leaves 85% to Focus &
(GLCM) and SVM 95% for diseased leaf Environmental
Variations

[11] 2023 DLMC-Net model Citrus Accuracy of 93.56%on

Cucumber Citrus , 92.34%on
State-of-the-Art Grapes Cucumber and 96.56%
(SOTA) Tomato on Tomato

[12] 2022 SVM (Support Plant Village An accuracy of rate of

Vector Machine) about 98.6% was
obtained from BSVM-
AOA (Arithmetic AOA technique
Optimization
Algorithm)

[13] 2023 Green ViTs plant The study The accuracy of 98% by The proposed
disease detection utilized two proposed Green Vision Green ViT
method popular Transformer demonstrate
standard superior overall
dataset PV , classification
DRLI accuracy
compared with
To test SOTA model ,
resilience a there are
new dataset misclassification
call PC dataset within both
was utilized categories
[14] 2023 pyramid network IDADP rice achieving an average
(DHLC- FPN) disease dataset accuracy of 97.44% on
the IDADP rice disease
Detection dataset
transformer
(DETR) algorithm

Detection
transformer
(DHLC- DETR)
methodology
[15] 2022 Data annotation Rocole An accuracy of
and processing 96.06%has been
obtained using Octave-
U-Net UNet model

Psp Net

Octave-UNet
 PROBLEM FORMULATION/OBJECTIVES

Problem Statement:
Challenges in Plant Disease Detection
 As per the Scientific literature, estimates suggest that on an average crop losses vary from 10-
35% annually due to diseases.
 Traditional detection methods are manual, slow, and prone to errors.
 Delayed detection leads to further disease spread and crop loss.
 Farmers in remote areas lack access to trained experts.
 Late detection causes overuse of pesticides, harming the environment.

Key Problem:
The question is: How can we design an AI-based system that automates the detection of plant diseases
and provides improved predictions using image data

Objective:
 Collection of dataset from authentic source and preprocess images for training and testing the
model.
 To develop fine tune-tuned ViT architecture for disease detection
 Compare the performance of the ViT model with state of the art approaches for leaf disease
detection methods.
 Methodology and Planning of Work

A Vision Transformer (ViT) is transformer designed for computer vision. A ViT breaks
down an input image into a series of patches (rather than breaking up text into tokens),
serializes each patch into a vector, and maps it to a smaller dimension with a single matrix
multiplication. These vector embeddings are then processed by a transformer encoder as if
they were token embeddings.
The advantages of using transformers, particularly in plant leaf disease detection for
sustainable agriculture, are multifaceted. Firstly, transformers excel at capturing long-range
dependencies in data, enabling them to discern complex patterns within images, crucial for
identifying subtle signs of plant diseases. Unlike traditional convolutional neural networks
(CNNs), transformers possess the flexibility to process inputs of varying sizes and
orientations, eliminating the need for image resizing and ensuring seamless handling of
diverse image datasets. Moreover, transformers demonstrate a remarkable ability to grasp the
global context of an image, allowing for a holistic understanding of the entire leaf surface
rather than focusing solely on local features. Their incorporation of attention mechanisms
further enhances interpretability by highlighting the most influential regions of the image,
aiding in understanding the model's decision-making process. Additionally, transformers offer
the potential for multi-modal learning, enabling the integration of diverse data sources such as
textual descriptions or environmental factors alongside images for more comprehensive
disease detection. These advantages collectively position transformers as a promising
approach for developing accurate, adaptable, and interpretable disease detection systems in
sustainable agriculture.
ViT has found applications in image recognition, image segmentation, and autonomous
driving.
Here are the steps we will follow while developing our project:
Data Collection: Gather a diverse dataset of labeled images containing healthy leaves and
leaves affected by various diseases across different crops to train the model effectively.
Data Preprocessing: Standardize the images, adjust brightness and contrast, and remove
background noise. Apply data augmentation techniques like rotation, flipping, and zooming
to increase dataset diversity.
Model Selection: Choose a transformer-based architecture suitable for image classification
tasks, such as the Vision Transformer (ViT), and adapt it to the plant disease detection
context.
Model Training: Train the selected model on the pre-processed dataset using supervised
learning techniques. Fine-tune hyper-parameters and optimize the model for accuracy and
generalization.
Validation and Evaluation: Validate the trained model on a separate validation dataset to
assess its performance. Evaluate metrics such as accuracy, precision, recall, and F1 score to
measure the model's effectiveness.
 Facilities required for proposed work:

Software Requirements:
1) Python 3.12: The primary language for developing machine learning and deep learning
models.
2) Deep Learning Frameworks:
PyTorch: A popular deep learning framework that supports Vision Transformers.
Tensor Flow: Another deep learning framework that can be used with Vit models.
3) Web Technologies:
HTML5: Markup language for structuring the web pages.
CSS3: For styling the web pages.
JavaScript: For adding interactivity to the web pages.
4) Version Control
Git: For source code management and collaboration.
GitHub or GitLab: For hosting the code repository.
Jupyter Notebooks for development and experimentation

Hardware Requirements:
Development Machines:
We will train our model using on both Graphics Processing Unit (GPU) and Central
Processing Unit (CPU) and a hard disk is required.
 References
1. Prabu, M., Chelliah, B.J. An intelligent approach using boosted support vector machine
based arithmetic optimization algorithm for accurate detection of plant leaf disease.
https://doi.org/10.1007/s10044-022-01086-z

2. Multispectral Plant Disease Detection with Vision Transformer–Convolutional

Neural Network Hybrid Approaches by Malithi De Silva and Dane Brown
https://doi.org/10.3390/s23208531

3. Kusuma, S., Jothi, K.R. Early betel leaf disease detection using vision transformer
and deep learning algorithms. Int. j. inf. tecnol. 16, 169–180 (2024).
https://doi.org/10.1007/s41870-023-01647-3

4. Vision Transformer Based Models for Plant Disease Detection and Diagnosis
November 2022 DOI:10.1109/ISIA55826.2022.9993508 Conference: 2022 5th
International Symposium on Informatics and its Applications
(ISIA)https://www.researchgate.net/publication/366665931_Vision_Transformer_B
ased_Models_for_Plant_Disease_Detection_and_Diagnosis

5. ORIGINAL RESEARCH article Front. Plant Sci., 26 May 2022 Sec. Technical
Advances in Plant Science Volume 13 -2022 |
https://doi.org/10.3389/fpls.2022.875693

6. Tomato Leaf Disease Identification by Restructured Deep Residual Dense Network

CHANGJIAN ZHOU et al. (2021)DOI: 10.1109/ACCESS.2021.3058947

7. https://www.researchgate.net/publication/356238093_Artificial_Cognition_for_Early

8. M. Genaev, S. Ekaterina and D. Afonnikov, "Application of neural networks to image

recognition of wheat rust diseases," 2020 Cognitive Sciences, Genomics and Bioinformatics
(CSGB), Novosibirsk, Russia, 2020, pp. 40-42, doi: 10.1109/CSGB51356.2020.9214703.

9. https://www.sciencedirect.com/science/article/abs/pii/S1574954123000547 DLMC-Net:
Deeper lightweight multi-class classification model for plant leaf disease detection

10. https://www.researchgate.net/publication/
362759136_An_intelligent_approach_using_boosted_support_vector_machine_based_arithm
etic_optimization_algorithm_for_accurate_detection_of_plant_leaf_disease An intelligent
approach using boosted support vector machine based arithmetic optimization algorithm for
accurate detection of plant leaf disease

11. https://www.mdpi.com/1424-8220/23/15/6949 Visual Intelligence in Precision

Agriculture: Exploring Plant Disease Detection via Efficient Vision Transformers
12. https://www.mdpi.com/2077-0472/13/7/1361 Disease Detection and Identification of
Rice Leaf Based on Improved Detection Transformer
13. https://www.mdpi.com/2073-4395/12/12/2933 Semantic Segmentation of Wheat Stripe
Rust Images Using Deep Learning
14. Wikipedia - https://en.wikipedia.org/wiki/Vision_transformer