“Multi-Vision Deep Learning Model To

Classify Various Wheat Plant Leaf

Diseases”
An Internship Report submitted in partial fulfillment of the requirements
for the award of the degree of
BACHELOR OF TECHNOLOGY
IN
COMPUTER SCIENCE AND ENGINEERING
by
VU21CSEN0300065-M LAKSHMI ARAVIND REDDY
GITAM (Deemed to be University)
From June 2024 to October 2024

DEPARTMENT OF
COMPUTER SCIENCE & ENGINEERING
GITAM (Deemed to be University)
VISAKHAPATNAM
2024
 DEPARTMENT OF COMPUTER SCIENCE AND
ENGINEERING
GITAM SCHOOL OF TECHNOLOGY GITAM (Deemed to be
University)

DECLARATION
I hereby declare that the internship report entitled “Multi-Vision Deep
Learning Model to Classify Various Wheat Plant Leaf Diseases” is
an original work done in the Department of Computer Science and
Engineering, GITAM School of Technology, GITAM (Deemed to be
University) submitted in partial fulfillment of the requirements for the
award of the degree of B.Tech. in Computer Science and Engineering.
The work has not been submitted to any other college or University for
the award of any degree or diploma.

Date:05-11-2024
Registration No Name signature

VU21CSEN0300065 M L Aravind Reddy

I
CERTIFICATE

II
 Acknowledgement

My sincere gratitude goes to Dr. Raja Reddy, who generously offered me the opportunity to pursue
a research internship under his guidance on developing a Multi-Vision Deep Learning Model for
the classification of various wheat plant leaf diseases. His mentorship has been invaluable in
broadening my understanding of both agricultural disease classification and advanced model
implementation. Dr. Reddy’s insights into the practical aspects of deep learning and its
applications in agriculture have deeply influenced my research approach and inspired me to delve
further into this field. His patience and expertise allowed me to overcome challenges with
confidence, and he constantly motivated me to refine my analytical skills and technical knowledge.

I would also like to express my heartfelt gratitude to my college and professors, whose continuous
support and encouragement have been a cornerstone of my academic journey. Their dedication to
nurturing my learning and skill development has been instrumental in preparing me for this stage,
where I can now engage in practical, real-world applications of my studies. It is through their
guidance and the skills I acquired in college that I have been able to reach this point, equipped
with the knowledge and perspectives necessary to contribute meaningfully to my field.

Thank you to everyone who has played a role in my journey so far. This experience has enriched
my understanding and has provided me with valuable skills that I look forward to applying in the
future.

III
 Table of Contents
1. Introduction………………………………………………………… 1

1.1 Major Project Functions of the Project……………………….……... 1

1.2 Why LLMs?...................................................................................... 2

1.3 Importance of Time Series Forecasting…………………………….... 2

1.4 The role of LLMs in time series forecasting ………………………….. 3

1.5 Future Scope……………………………………………………...... 4

2. Problem Statement………………………………………………….. 5

3. Objective…………………………………………………………….. 6

4. Methodology………………………………………………………… 7

4.1 Requirement Analysis and Research………………………………... 7

4.2 System Design……………………………………………………... 9

4.3 Technologies……………………………………………………… 10

4.4 Model Implementation…………………………………………….. 12

4.5 Testing and Evaluation…………………………………………….. 13

5. Code…………………………………………………………………. 15

6. Results………………………………………………………………. 20

7. Conclusion…………………………………………………………... 23

IV
 Table of Figures

Figure No Figure Page No.

5.1 EfficientNet b0 15

5.2 EfficientNet v2 b0 16

5.3 SqueezzNet 17

5.4 DenseNet121 18

5.5 ResNet18 19

6.1 Performance of EfficientNet b0 20

6.2 Performance of EfficientNet v2 b0 20

6.3 Performance of SqueezzNet 21

6.4 Performance of DenseNet121 21

6.5 Performance of ResNet18 22

6.6 Confusion Matrix of SimpleEfficientNet 22

V
 ABSTRACT

This report presents the results and methodology of my research on classifying various wheat plant
leaf diseases using a Multi-Vision Deep Learning Model, incorporating Quantization-Aware
Training (QAT) to enhance both accuracy and efficiency. The primary objective was to assess the
capability of QAT in preserving model accuracy while significantly reducing computational and
memory requirements, thereby enabling deployment on resource-constrained devices, such as
mobile or embedded systems used in agricultural settings. The study involved a comprehensive
analysis of several convolutional neural networks (CNNs), including EfficientNetLite and
ResNet18, which were specifically adapted and fine-tuned to identify common wheat diseases
such as leaf rust, powdery mildew, and stripe rust with high accuracy.

The research largely focused on thorough data preprocessing, careful model architecture selection,
and extensive hyperparameter tuning to maximize classification performance under quantization
constraints. Evaluation metrics included precision, recall, F1-score, confusion matrix, and Mean
Average Precision (MAP), with additional considerations for model size and inference speed to
ensure deployment viability. Findings demonstrated that while the full-precision models achieved
robust baseline accuracy, QAT-enabled models achieved comparable performance with significant
reductions in memory usage and faster inference times.

This study underscores the practical advantages of incorporating QAT in plant disease
classification tasks, revealing that it effectively balances accuracy with efficiency, which is
essential for real-time, on-device processing in agriculture. The insights gained highlight the
transformative potential of QAT-enhanced models, making deep learning solutions more
accessible, scalable, and operationally feasible for advancing precision agriculture.

VI
 1. Introduction

Today, with the increasing demand for precision agriculture, plant disease classification has
become a vital component in managing crop health and enhancing agricultural productivity. As
agricultural data grows in complexity, particularly with high-resolution images of affected plant
areas, the intricate patterns of disease symptoms are often challenging to capture effectively.
This creates a clear need for advanced deep learning methods capable of high accuracy in
classifying multiple disease types across large datasets. The Multi-Vision Deep Learning Model
meets this demand by leveraging Quantization-Aware Training (QAT), which combines
powerful CNN architectures with model efficiency, specifically suited for resource-limited
devices. This Python-based approach integrates modern quantized deep learning models, making
them more accessible for real-time, on-field disease identification.

1.1 Major Project Functions of the Project:

 Multi-Vision Deep Learning Model: This project includes the development and
application of a deep learning model to classify multiple wheat plant leaf diseases,
such as leaf rust, powdery mildew, and stripe rust. The model leverages convolutional
neural networks (CNNs) like EfficientNetLite and ResNet18, providing a robust
foundation for accurate disease classification across diverse visual patterns in plant
leaves.

 Quantization-Aware Training Techniques: Users can apply Quantization-Aware

Training (QAT) methods, enabling the model to maintain high classification accuracy
while reducing model size and computational requirements. This functionality is
essential for deploying the model on resource-limited devices, making it practical for
in-field applications where real-time classification is needed.

1
  Evaluation and Performance Metrics: The model’s effectiveness is evaluated
through precision, recall, F1-score, and confusion matrix, along with additional
metrics like inference speed and memory usage. These evaluations help ensure that
the model performs optimally for accurate and efficient disease detection in
agricultural settings.

1.2 Why Quantization-Aware Training?

 Quantization-Aware Training (QAT) enables deep learning models to operate

efficiently on resource-constrained devices, such as those commonly used in
agricultural settings. By applying QAT, the model can maintain high classification
accuracy while significantly reducing memory and computational requirements,
making it feasible for real-time disease detection in the field.

 QAT allows models to be fine-tuned with reduced precision, preserving performance

while optimizing for speed and storage. This enables the Multi-Vision Deep Learning
Model to deliver accurate results without sacrificing efficiency, providing an
accessible solution for farmers and agricultural specialists to identify plant diseases
promptly and accurately.

1.3 Importance of Plant Disease Classification

Classifying plant diseases accurately is essential across agriculture, agritech, and

environmental sustainability, as it directly impacts crop health, yield, and food security.
The timely identification and management of plant diseases drive better agricultural
practices, protect resources, and enhance productivity. For example:

 In Crop Management: Early disease detection enables farmers to apply targeted

treatments, reducing the spread of disease and minimizing crop loss.

2
  In Agritech Solutions: Accurate disease classification allows technology providers to
develop smart farming tools that support automated disease monitoring, reducing
dependency on manual inspections.

 In Food Security: Preventing disease outbreaks in crops helps stabilize food supply
chains, ensuring sufficient yield to meet global food demands and improve overall
agricultural sustainability.

1.4 Role of QAT in Plant Disease Classification

The integration of Quantization-Aware Training (QAT) in plant disease classification

marks a significant advancement in precision agriculture. By enabling efficient, high-
performance models to operate on low-resource devices, QAT facilitates the on-field
detection of plant diseases at scale. Through QAT, deep learning models can handle high-
resolution image data effectively, supporting accurate and efficient disease diagnosis
while conserving memory and computational power. This project integrates popular
Python libraries, including TensorFlow and PyTorch for deep learning, and Scikit-learn
for preprocessing and evaluation, to enhance model development and deployment.

This report discusses the design, features, and applications of QAT in disease
classification, showcasing how it enables models to deliver robust performance in real-
world agricultural environments. QAT’s role in balancing accuracy with efficiency
addresses the challenges of deploying advanced deep learning solutions in agriculture,
where quick, reliable insights are crucial for proactive disease management and improved
crop outcomes. This underscores the importance of QAT in supporting accessible and
effective AI-driven solutions, crucial for the future of sustainable agriculture.

3
1.5 Future Scope

The future potential for plant disease classification using deep learning, especially with
Quantization-Aware Training (QAT), offers vast opportunities for enhancement and
innovation. Possible directions for future work include:

 Hybrid Models : Future research could focus on hybrid models that integrate
CNNs with Vision Transformers or other architectures to capture complex visual
patterns and dependencies in plant diseases. Such combinations may lead to better
classification performance on intricate datasets.

 Meta-Learning Approaches: Meta-learning can enable models to quickly adapt to

new crops or emerging disease patterns. This approach would be particularly
useful for diverse agricultural environments where disease symptoms or data
distributions may change frequently.

 Explainability and Interpretability: As models grow more complex, ensuring

interpretability in disease classification is crucial for user trust. Future research
can explore explainable AI frameworks that help farmers and agronomists
understand model decisions, fostering transparency in disease diagnosis.

 On-Device Real-Time Classification: Enhancing the capability for real-time, on-

device disease detection would make these models more practical for in-field use.
QAT can facilitate this by balancing speed with accuracy, enabling rapid, accurate
disease detection directly on mobile devices.

 Scalability and Deployability: Further research on scalable architectures suitable

for distributed systems can improve the accessibility of these models.

4
 2. Problem Statement

In the field of agriculture, particularly within precision farming, the accurate identification and
classification of plant diseases are essential for safeguarding crop health, yield, and quality.
Traditional disease identification methods, such as visual inspections, are not only labor-
intensive and time-consuming but also prone to human error, making them less feasible for
large-scale or real-time deployment in the field. Early and accurate disease detection is vital to
curbing the spread of infections and optimizing crop management, yet the complexity of visual
symptoms and variations across plant species present challenges that require advanced solutions.

Recent advancements in deep learning have shown significant potential in automating disease
classification; however, these models often demand substantial computational power and
struggle with deployment on resource-constrained devices such as smartphones and IoT devices.
This project addresses these challenges by exploring a Multi-Vision Deep Learning Model
integrated with Quantization-Aware Training (QAT) for the classification of wheat plant leaf
diseases. The QAT approach enables efficient model performance on low-resource devices by
optimizing models to maintain accuracy despite lower computational demands, making them
highly applicable for real-time use directly in the field.

By combining a multi-vision model with QAT, the project aims to maximize both accuracy and
efficiency, allowing for on-device disease classification suitable for farmers in remote or large-
scale agricultural settings. This involves rigorous experimentation across model architectures and
hyperparameters to strike an optimal balance between predictive accuracy and computational
efficiency, creating a scalable solution that is accessible and effective.

This project’s broader objective is to enhance the utility of deep learning in agriculture by
enabling faster, more reliable, and more resource-efficient disease detection, potentially
transforming how farmers manage plant health. With the goal of creating a deployable, high-
performance model, this project seeks to address the scalability and accessibility gaps in current
AI-powered agricultural tools. The findings and methodologies developed will contribute to
building sustainable, data-driven farming practices, offering farmers the technology needed to
detect diseases proactively and ensure healthier crops in a resource-effective manner.

5
 3. Objective

The goal of this internship was to develop and implement a Multi-Vision Deep Learning Model
to classify wheat plant leaf diseases with Quantization-Aware Training (QAT), combining
accurate disease detection with efficient model deployment on resource-constrained devices.
This project involved achieving the following objectives:

Understanding the Disease Characteristics: Developing a comprehensive understanding of

various wheat leaf diseases, including identifying visual symptoms, variations, and similarities to
ensure precise classification across disease categories.

Model Development and Implementation: Designing and implementing a deep learning model
tailored for leaf disease classification, incorporating Quantization-Aware Training (QAT) to
achieve a balance between accuracy and computational efficiency suitable for deployment on
low-resource devices.

Performance Evaluation: Conducting rigorous testing to assess model performance, comparing

the QAT-optimized model against traditional models in terms of accuracy, precision, and
computational efficiency.

Data Preprocessing and Feature Engineering: Handling essential data preprocessing and feature
engineering tasks, including preparing and augmenting image datasets, to create optimal input
for the multi-vision model and improve classification accuracy.

Research and Documentation: Documenting findings, methodologies, and results to provide

insights into the effectiveness of QAT in agricultural applications and contribute to
advancements in real-time, resource-efficient plant disease detection models.

6
 4. Methodology

The development process of the Multi-Vision Deep Learning Model for Wheat Plant Leaf
Disease Classification with Quantization-Aware Training (QAT) was meticulously structured,
beginning with requirement analysis and progressing through design, model development, and
testing. This systematic approach ensured that the final model was robust, accurate, and capable
of efficiently identifying various disease patterns in wheat leaves while being optimized for
deployment on resource-constrained devices. Each stage contributed to a reliable classification
system suitable for real-world agricultural applications.

4.1 Requirement Analysis and Research

This phase focused on an in-depth analysis of project requirements to define the main
objectives and select appropriate techniques for disease classification. This process
involved an extensive study of image classification models and quantization-aware
training (QAT) techniques for an optimized deep learning model suitable for on-device
applications.

 Core Functionalities: The project's goal was to develop a model that could
accurately classify various wheat plant leaf diseases while being efficient for
real-world agricultural applications. Specific functionalities included:

 Accurate classification of wheat leaf diseases, capturing unique

patterns associated with each disease type.

 Robust performance under quantization-aware training to ensure

minimal loss in accuracy.

 Visualization of classification outcomes to provide users with

interpretable results.

7
 R&D and Model Selection: Various models were researched and
evaluated for their ability to classify leaf disease patterns. Selected models
include:

 EfficientNet-based architectures for image classification, given

their accuracy and efficiency.

 Lightweight models like MobileNet for potential deployment on

devices with limited computational resources.

 Application of quantization-aware training (QAT) to achieve

efficiency without compromising classification performance.

 Tool and Library Selection: Key tools and libraries were selected to meet
the project’s needs:

 Python as the primary programming language

 TensorFlow and Keras for implementing the EfficientNet and

MobileNet models and for utilizing QAT capabilities.

 TensorFlow Lite for model conversion and deployment,

particularly for mobile or edge devices.

 OpenCV and Pillow for preprocessing image data to ensure

compatibility with the deep learning models.

 Matplotlib and Seaborn for visualizing classification results and

performance across different diseases.

8
4.2 System Design

The system design stage ensured that the wheat leaf disease classification tool was
structured, capable of handling image data effectively, and optimized for accurate
disease detection. The architecture was divided into several modules, each responsible
for different aspects of the classification and quantization process.

 Data Processing Module: This module focused on preparing raw images for
model training and inference, including the following steps:

 Data Cleaning: Addressed image quality issues, such as removing

duplicates, correcting mislabeling, and handling images with poor
resolution.

 Image Preprocessing: Applied resizing, normalization, and

augmentation (e.g., rotation, flipping) to enhance model robustness
and generalization.

 Machine Learning Model Module: Designed and validated appropriate deep

learning architectures for image-based disease classification.

 EfficientNet: Selected due to its efficiency and accuracy, this model

was trained to recognize disease-specific patterns in wheat leaves.

 MobileNet: Implemented for its lightweight architecture, ideal for

resource-constrained devices. This model ensured that deployment on
mobile or edge devices remained feasible.

9
  Quantization-Aware Training (QAT) Module: Implemented quantization
techniques to ensure efficient model deployment without significant loss of
accuracy.

 QAT Implementation: Applied quantization techniques, reducing

model size and improving inference speed while preserving
performance.

 Model Conversion: Used TensorFlow Lite to convert the quantized

model, making it compatible with mobile and edge environments.

 Visualisation Module: This module visualized classification results, trends,

and performance metrics for easy interpretation by users.

 Matplotlib and Seaborn: Generated graphs and plots showing

classification accuracy, confusion matrix, and per-disease detection
rates.

4.3 Technologies

The project's success relied on selecting the most suitable tools and technologies to
address the challenges of wheat plant leaf disease classification efficiently. Each tool
and library was chosen for its specific strengths and its ability to integrate seamlessly
into the overall development process:

 Python was selected as the primary programming language due to its

simplicity, readability, and the vast ecosystem of libraries tailored for machine
learning and deep learning tasks. Python’s versatility made it ideal for
handling data preprocessing, model development, and deployment seamlessly.

10
 TensorFlow and Keras were chosen for building and training the deep
learning models, including EfficientNet and MobileNet architectures.
TensorFlow’s robust framework and Keras’s user-friendly API facilitated
rapid model prototyping and experimentation, enabling the implementation of
advanced neural network architectures with ease.

 TensorFlow Lite was utilized for converting the trained models into a format
suitable for deployment on mobile and edge devices. This was essential for
ensuring that the quantized models could run efficiently in real-time
agricultural environments without significant loss of accuracy.

 OpenCV and Pillow were employed for image preprocessing tasks. OpenCV
provided powerful tools for image manipulation, such as resizing,
normalization, and augmentation, while Pillow was used for handling image
formats and performing additional preprocessing steps to prepare the data for
model training.

 Quantization-Aware Training (QAT) Libraries within TensorFlow were

used to implement QAT techniques. These libraries enabled the optimization
of model weights and activations during training, ensuring that the final
models were both efficient and accurate when deployed on resource-
constrained devices.

 Scikit-learn was selected for its comprehensive suite of tools for data
preprocessing, including feature scaling, augmentation, and splitting datasets
into training and testing sets. Additionally, Scikit-learn was used for
implementing classical machine learning baselines, such as Random Forest
and Logistic Regression, to benchmark the performance of the deep learning
models.

11
  Matplotlib and Seaborn were used for creating detailed visualizations of the
model’s performance metrics, such as accuracy, precision, recall, and
confusion matrices. These visualizations were crucial for interpreting the
results and understanding the strengths and weaknesses of the classification
models.

 Jupyter Notebook provided an interactive development environment for

experimenting with different model architectures, preprocessing techniques,
and training strategies. Its ability to combine code, visualizations, and
narrative text made it an invaluable tool for documenting the development
process and sharing insights.

4.4 Model Implementation

Each model in this project was carefully crafted and individually tested to ensure it
could effectively classify wheat plant leaf diseases. The final system integrates these
models to handle various disease types, making it versatile and adaptable for
agricultural use.

 Deep Learning Models

 Convolutional Neural Networks (CNNs): These were primarily utilized

for their strength in image classification, as they excel in extracting
features from image data. CNNs, with architectures like EfficientNet
and ResNet, were employed to recognize disease patterns effectively
in wheat leaf images.

 Support Vector Machines (SVM): EfficientNet and MobileNet

Architectures: EfficientNet was chosen for its balanced performance in
terms of accuracy and efficiency, making it suitable for mobile
deployment in real-time scenarios. MobileNet, with its lightweight

12
 structure, was optimized for speed without sacrificing much accuracy,
crucial for field-level, on-device analysis.

 Quantization and Lightweight Models: Using quantization-aware

training, the project optimized CNN models to be efficient for edge
deployment, ensuring they retain accuracy while consuming less
memory and processing power. This approach allows these models to
perform on mobile devices in low-power environments, such as fields.

4.5 Testing and Evaluation

The testing and evaluation procedures were designed to ensure the reliability and
accuracy of the wheat plant leaf disease classification models across diverse datasets.
Each model's performance was assessed individually, and comparisons were made
across models using ensemble techniques to enhance predictive power.

 Accuracy Evaluation: The accuracy of each model was assessed using key
image classification metrics, including overall accuracy, precision, recall, and
F1-score. These metrics offered a comprehensive view of how effectively each
model distinguished between healthy and diseased wheat plant leaves. High
accuracy and precision scores indicated strong performance in correctly
identifying the disease categories, while recall and F1-score highlighted the
models' consistency in capturing all disease cases. This rigorous evaluation
ensured that the models met the critical requirements for accurate and reliable
disease classification, essential for effective disease management in agricultural
settings.

 Hyperparameter Tuning: A structured approach to hyperparameter tuning was

implemented to optimize the performance of each image classification model.
Techniques such as Grid Search and Random Search were employed
specifically to explore the optimal combinations of hyperparameters for image

13
 classification tasks. For each model, parameters like learning rates, batch sizes,
dropout rates, and network depth were fine-tuned to maximize classification
accuracy. This tuning process enhanced the efficiency of the models,
minimized classification errors, and improved the reliability of disease
identification across various wheat plant leaf categories.

 Ensemble Techniques:The results from individual models were combined

using ensemble techniques such as bagging (Bootstrap Aggregating) or
boosting (e.g., XGBoost). These techniques aimed to enhance predictive
performance by leveraging the strengths of multiple models.

 Continuos Monitoring and Feedback: A feedback loop was established to

continuously monitor model performance in real-world applications. As new
data became available, models were retrained and evaluated to adapt to
changing disease patterns and environmental factors. This iterative process
ensured that the models remained relevant and accurate over time, ultimately
contributing to more effective disease management strategies in agriculture.

14
 5. Code

Figure 5.1 efficientenet b0

15
Figure 5.2 efficientnet v2b0

16
Figure 5.3 squezznet

17
figure 5.4 DenseNet121

18
figure 5.5 Resnet18

19
 5. Results

Figure 6.1 Performance of EfficientNet b0

Figure 6.2 Performance of EfficientNet v2b0

20
 Figure 6.3 Performance of Squezznet

Figure 6.4 Performance of DenseNet121

21
 Figure 6.5 Performance of ResNet18

Figure 6.6 Confusion Matrix of SimpleEfficientNet

22
 6. CONCLUSION

The project on “Wheat Plant Leaf Disease Classification Using Quantization-Aware Training
Models” demonstrates the significant potential of combining traditional classification techniques
with modern quantization-aware approaches to achieve highly accurate and efficient disease
identification in agricultural settings. By leveraging both classical statistical techniques and state-
of-the-art machine learning models, this project addressed the challenges of effective disease
classification on wheat leaf datasets with a particular emphasis on maintaining model efficiency
and interpretability.

Throughout this project, we applied a variety of models, from traditional image processing and
basic machine learning algorithms to advanced deep learning models like ResNet, MnasNet, and
EfficientNetLite. The classical models provided a foundational baseline, achieving reasonable
accuracy while remaining computationally inexpensive. However, due to the complexities and
non-linear patterns associated with visual disease symptoms, classical models alone were
insufficient for accurately distinguishing between disease types in high-resolution imagery. This
limitation led to the exploration and integration of deeper neural architectures, which enabled the
detection of subtle and complex patterns within the image data.

The introduction of quantization-aware training significantly improved the resource efficiency of

the models without compromising accuracy. By applying post-training quantization and
quantization-aware training to architectures like EfficientNetLite, the project succeeded in
optimizing the memory and computational requirements, making the models suitable for
deployment on edge devices, such as mobile or IoT-based agricultural monitoring systems. This
aspect of the project emphasized the value of efficient model designs for real-time applications in
rural and remote settings, where access to high-performance computing infrastructure may be
limited.

23
In conclusion, this project highlighted the importance of selecting and combining models based
on the specific requirements of agricultural disease classification, where efficiency, accuracy,
and interpretability are key factors. The project demonstrated that advanced deep learning
models, when optimized through quantization-aware training, can bridge the gap between
accuracy and deployability, offering robust solutions for large-scale disease monitoring and
management in the agriculture domain.

Future work may involve fine-tuning these models for specific types of wheat disease across
different geographic regions, considering the variations in climate and environmental factors that
may affect disease manifestations. Additionally, exploring ensemble approaches that combine
the strengths of classical and modern models could further enhance predictive power, especially
when working with multimodal datasets. Integrating multi-spectral image data or even
leveraging satellite imagery for early detection could extend the applicability of the model to
broader agricultural monitoring systems.

Interpretability will continue to play a crucial role in the refinement of these models. Developing
frameworks that emphasize the interpretability of model outputs will ensure that end-users,
particularly farmers and agricultural decision-makers, can trust and act upon the classifications.
Methods like SHAP and LIME could be further adapted for agricultural applications, helping
users understand which visual features are most indicative of specific diseases, potentially
supporting faster and more accurate diagnoses on the field.

Ultimately, this project establishes a foundation for highly efficient and accurate wheat plant leaf
disease classification, supporting sustainable agricultural practices and improved crop
management through advanced machine learning methodologies. As the agricultural industry
increasingly adopts digital tools, quantization-aware models such as these offer a promising
direction for scalable, precise, and user-friendly disease management solutions that can help
drive agricultural innovation forward.

24
25