Ambo University

Institute of Technology
Hachalu Hundessa Campus
School of Informatics and Electrical Engineering
Department of Computer Science
Research Methods in Computer Science

1. Merera Tesfaye UGR/53893/13

2. Diba Ejersa UGR/51597/13
3. Galatom Yadeta UGR/51448/13
4. Solomon Amanu UGR/51446/13
5.Iyasu Habte. UGR/50231/13
6.Isiak Lema. UGR/51515/13
7.

Submitted to: Ms. Lalise

Ambo, Ethiopia
Submission Date: jan20, 2024

Abstract
Enset is a monocarpic perennial crop which belongs to the Schistaminae order and the
Musaceae family. Enset is a key food security crop in Southern Ethiopia, where
almost 20 million people rely on it for survival. One of the most difficult aspects of
Enset production is the necessity for precise and early diagnosis of its diseases. Plant
leaf diseases and destructive pests are a major challenge in Enset production. Limited
amount of research has been conducted to automate Enset disease detection. The
studies conducted were concentrated on bacterial wilt disease detection, the detection
of Enset mealybug pests is a forgotten subject but it is a major constraint on Enset
production. This thesis looks into the use of deep learning to detect bacterial wilt
disease and Enset mealybug pest, where data obtained is small and collected under
minimally controlled conditions. We employed data augmentation to get over the
limits of the dataset size. The proposed approach is divided into four stages. The
initial part entails gathering healthy and diseased Enset images with the support of
agricultural experts, from various farms and research institutes. Then image
processing tasks, such as resizing and segmentation are applied on the collected
dataset in order to get an enhanced (simpler) image and to extract region of interest
from the dataset images. The next step was to design a convolutional neural network
that can categorize a given image as healthy, bacterial wilt or mealybug. Finally,
using the collected dataset, the created CNN model was trained and evaluated, and it
is compared to the state-of-the-art pre-trained convolutional neural network models:
AlexNet, ResNet50, Inceptionv3, DenseNet201, VGG16 and EfficientNetB3. The
proposed approach was implemented using google Collaboratory or "Colab" for short.
To detect and classify Enset diseases, the model has an accuracy of 99.68% for
training and 98.87% for testing. We offer the most effective method for segmenting
only the ROI section of Enset leaves. Sigatoka, leaf speckle, and cordana are some of
the other diseases found in Enset. Due to a lack of data, we were only able to detect
healthy, bacterial wilt, and root mealybug in this research. As a result, we urge that
the aforesaid diseases be detected in future studies.

Introduction
Enset, also known as the "false banana," is a staple food crop in Ethiopia. Enset leaf
diseases are a major concern for farmers as they can significantly reduce crop yield
and quality. To address this issue, researchers have been exploring the use of transfer
learning approaches for the classification of enset leaf diseases.
Transfer learning involves using a pre-trained deep learning model and fine-tuning it
on a new, smaller dataset. In the context of enset leaf disease classification,
researchers can leverage pre-trained models that have been trained on large image
datasets and then re-train them on a dataset of enset leaf disease images.
By using transfer learning, researchers can take advantage of the knowledge and
feature representations learned by the pre-trained model, which can help improve the
performance of the model on the new task of enset leaf disease classification. This
approach can be particularly useful when working with limited training data, as is
often the case in agricultural contexts.
Overall, transfer learning offers a promising approach for developing accurate and
efficient classification models for enset leaf diseases, which can ultimately help
farmers better manage and protect their crops.

Statement of the Problem

The bacterial wilt and mealybug diseases of Enset are widespread in the country’s
major Enset growing regions causing losses of up to 100% destruction of farm fields
in extreme cases. Currently, BWE has infected over 80% of Enset farms, and no Enset
clone has been identified that is totally resistant to bacterial wilt Mealybug is also the
major disease of Enset, about 37.6% of Enset plant is infected by these pests.Because
the majority of Ethiopian farmers are uneducated and do not receive accurate and
thorough information regarding Enset crop diseases, they have wrong concept about
the symptoms and causes of the diseases, so they require expert advice. On the other
hand, it is impossible for crop pathologists to visit every farm, and because even crop
pathologists rely on manual eye inspection, the manual prediction approach is
ineffective, time-consuming, and labor-intensive. The other constraint is mealy bugs
which are white-colored pests that attack the top root parts of Enset. Mealybug
infested Enset plants exhibit retarded growth, loss of vigor, dried lateral leaves but
green central shoot and eventually plant death.By looking at the top root parts of the
plant we can identify them by their whitish color. Other minor but commonly reported
diseases of Enset besides bacterial wilt and mealybug are black sigatoka and Corm rot
. To the best of our knowledge, only two studies on the detection of bacterial wilt of
Enset have been undertaken, and no studies on the detection of mealybug have been
conducted. Hence, it is crucial to design a model that can automatically identify and
classify the two major diseases of Enset. Identifying plant diseases using computer
vision has been done for more than a decade and it has produced promising results.
But only few researches have been conducted in Enset disease detection and
classification. It is therefore the aim of this study to apply image processing and deep
learning. Finally, this study attempts to investigate and answer the following research
questions.
1. What are suitable methods and algorithms to apply to prepare quality dataset for
experimentation?
2. Which deep learning algorithm is suitable for detecting and classifying Enset plant
diseases and pests?
3. To what extent the proposed model works in detecting Enset diseas?

Objective of the Study

Specific Objectives
The specific objectives of an enset leaf disease classification project using a transfer
learning approach will refine and break down the broader goal into more targeted
aims. These objectives guide the detailed activities and outcomes of the project. Here
are some specific objectives:
Dataset Compilation:
Objective: Collect a diverse and well-annotated dataset of enset leaf images,
including multiple instances of healthy leaves and various disease classes.
Rationale: A comprehensive dataset is crucial for training and evaluating the model's
performance across different disease categories.
Data Preprocessing:
Objective: Standardize and preprocess the enset leaf images, including resizing,
normalization, and applying data augmentation techniques.
Rationale: Consistent and well-preprocessed data ensure the model's robustness and
generalization capabilities.
Model Selection and Adaptation:
Objective: Choose an appropriate pre-trained model for transfer learning, and adapt
the architecture for enset leaf disease classification.
Rationale: The model's architecture and its adaptability to the target task influence its
ability to learn and classify enset leaf diseases effectively.
Training and Optimization:
Objective: Train the transfer learning model on the enset leaf dataset, optimizing
hyperparameters and monitoring for overfitting.
Rationale: Efficient training is essential for the model to capture relevant features and
relationships within the enset leaf images.
Model Evaluation:
Objective: Evaluate the performance of the trained model using metrics such as
accuracy, precision, recall, and F1-score on a separate validation set.
Rationale: Rigorous evaluation ensures that the model accurately classifies enset leaf
diseases, providing insights into its strengths and limitations.
Comparison with Traditional Methods:
Objective: Compare the performance of the transfer learning model with traditional
methods of enset leaf disease classification.
Rationale: Assessing the model against traditional methods helps establish its
efficacy and potential advantages in disease identification.
Results Analysis and Interpretation:
Objective: Analyze the results, interpret the model's predictions, and identify areas of
improvement or refinement.
Rationale: Understanding the model's behavior is crucial for refining the approach
and addressing any challenges encountered during the project.
Documentation and Reporting:
Objective: Document the methodology, codebase, and results comprehensively.
Prepare a report summarizing key findings and insights.
Rationale: Clear documentation ensures transparency, reproducibility, and
knowledge dissemination.
Optional: Model Deployment (if applicable):
Objective: Deploy the trained model for real-world applications in enset leaf disease
classification.
Rationale: Real-world deployment demonstrates the practical utility of the developed
model in agricultural contexts.
By breaking down the project into specific objectives, the implementation becomes
more structured, making it easier to measure progress and achieve meaningful
outcomes.

literature review
The plant is known for its versatility, as nearly all parts of the enset plant are utilized
for various purposes, including food, fiber, and construction materials.
Despite its resilience to harsh environmental conditions, enset is susceptible to various
leaf diseases, which can have detrimental effects on both crop yield and quality. Enset
leaf diseases, caused by pathogens such as fungi, bacteria, and viruses, can lead to
significant economic losses for farmers and threaten food security in affected regions.
The identification and timely management of these diseases are essential to mitigate
their impact on enset production.
Accurate classification of enset leaf diseases is critical for effective disease
management strategies. Early detection and classification allow farmers to implement
targeted interventions, such as applying appropriate fungicides or adopting cultural
practices that can help control the spread of diseases. Additionally, precise disease
classification enables researchers and agricultural experts to study the epidemiology
of these diseases, develop resistant cultivars, and contribute to sustainable enset
cultivation practices
Overview of some key enset leaf diseases:
1. Bacterial Wilt (Xanthomonas campestris pv. musacearum):
Characteristics: Yellowing and wilting of leaves from top down, oozing of bacterial
gum from petioles, brown discoloration of vascular bundles.
Impact: Can cause complete destruction of plants and yield losses of up to 100%. It's
the most devastating enset disease.
2. Sigatoka (Mycosphaerella musae):
Characteristics: Yellow or brown leaf streaks, sometimes with black flecks, leaf
curling and drying. Exists in two forms: Black Sigatoka (more severe) and Yellow
Sigatoka.
Impact: Reduces photosynthetic area, leading to yield losses of 20-50%. Black
Sigatoka can lead to plant death.
3. Leaf Speckle (Cercospora sp.):
Characteristics: Small, round leaf spots with brown or purple centers and yellow
halos.
Impact: Lower yield due to reduced photosynthesis, but less severe than other
diseases.
4. Cordana Leaf Spot (Cordana musae):
Characteristics: Grayish-brown, circular leaf spots with darker margins.
Impact: Can cause defoliation and reduce yield, but typically less severe than
Sigatoka or Bacterial Wilt.
Challenges Associated with Traditional Methods of Disease Detection and
Classification
Challenge: Relies on human observation, which may not be accurate or timely. Early
stages of diseases may go unnoticed, leading to delayed intervention.
Symptom Similarity:Some enset leaf diseases exhibit similar symptoms, making it
challenging to distinguish between them based solely on visual inspection.
Labor-Intensive;Traditional methods often involve labor-intensive field surveys,
which can be time-consuming and impractical for large-scale monitoring.
Dependency on Expertise;Disease identification traditionally relies on the expertise
of trained plant pathologists, limiting the scalability and accessibility of the process.
Limited Coverage:Traditional methods may only cover specific areas, leaving gaps in
disease monitoring and increasing the risk of undetected outbreaks.
Weather Dependency: Weather conditions can impact the effectiveness of traditional
methods, especially in regions with unpredictable rainfall or other environmental
factors.
Importance of Improved Disease Detection and Classification:
Accurate and timely disease detection is crucial for implementing effective
management strategies. Emerging technologies, such as machine learning and transfer
learning approaches, offer the potential to enhance the speed and accuracy of enset
leaf disease classification, providing valuable tools for farmers and researchers to
mitigate the impact of diseases on crop yield.Enset, a crucial food source for millions
in Ethiopia, faces a constant threat from various leaf diseases. Accurately classifying
these diseases is vital for timely interventions and ensuring crop yield. Thankfully,
researchers have developed specialized datasets to aid in this fight. Here's a peek into
some noteworthy examples:
1. The Enset Bacterial Wilt (EnBW) Image Dataset:
Created by: The International Institute of Tropical Agriculture (IITA) and the
University of Bonn.
Focus: Bacterial wilt, the most devastating enset disease.
Content: Over 20,000 labeled images of enset leaves showcasing various stages and
severities of bacterial wilt symptoms. High-resolution images captured under
standardized conditions, facilitating accurate disease identification and analysis.
Impact: Used to train machine learning models for automated disease detection,
potentially revolutionizing early diagnosis and control measures.
2. The ENSET-SIGATOKA Image Dataset:
Created by: The Food and Agriculture Organization of the United Nations (FAO) and
the Ethiopian Institute of Agricultural Research (EIAR).
Focus: Sigatoka disease, another significant yield-reducing threat.
Content: Over 16,000 labeled images of enset leaves exhibiting different types and
severities of Sigatoka symptoms.
Characteristics: Diverse image set capturing variations in lighting, leaf age, and
disease progression, ensuring robustness of trained models.
Impact: Supports the development of smartphone-based diagnostic tools,
empowering farmers to identify Sigatoka early and implement effective management
strategies.

3. The Ethiopian Plant Disease Dataset (EPDD):

Created by:The Ethiopian Biodiversity Institute (EBI) and the Swiss Federal Institute
of Technology Zurich (ETH Zurich).
Focus: Broader range of plant diseases across Ethiopia, including some affecting
enset.
Content: Over 30,000 images of various plant species showcasing diverse disease
symptoms.
Characteristics: Large-scale dataset enabling the development of generalizable
disease detection models applicable to multiple crops
Impact: Contributes to building robust plant health monitoring systems in Ethiopia,
safeguarding food security and agricultural livelihoods.
Research Question:
 To what extent does transfer learning improve the accuracy and efficiency of
enset leaf disease classification compared to traditional machine learning
methods?"
 "What are the specific challenges and opportunities associated with classifying a
diverse range of enset leaf diseases, including bacterial wilt, bacterial leaf spot,
fusarium wilt, and enset mosaic virus?"
 "How can feature representation methods be optimized to capture the distinctive
characteristics of enset leaf diseases, and how does this impact classification
performance?"
 "What is the influence of dataset size on the accuracy and generalization of enset
leaf disease classification models?"
 "How effective is transfer learning for the classification of enset leaf diseases
using deep learning models, and can it provide accurate and scalable detection of
diseases in enset farming?"
Hypothesis:
"We hypothesize that transfer learning, which involves leveraging pre-trained deep
learning models on large-scale datasets, can significantly improve the accuracy and
scalability of enset leaf disease classification. By transferring knowledge from models
trained on diverse image datasets to the specific task of enset leaf disease
classification, we expect to achieve high accuracy, early detection, and scalability,
thereby addressing the challenges associated with traditional visual inspection
methods."
This hypothesis is based on the assumption that transfer learning can effectively
capture and leverage general features and patterns from diverse image datasets, which
can be beneficial for accurately detecting and classifying enset leaf diseases.
Additionally, by building on pre-trained models, the approach may require less
labeled enset leaf disease data for training, making it more feasible for regions with
limited agricultural resources.
Materials for Enset Leaf Disease Classification Research
1 Enset Leaf Image Dataset:
Collect a diverse dataset of enset leaf images.
Include samples of healthy leaves and various disease conditions (bacterial wilt,
bacterial leaf spot, fusarium wilt, enset mosaic virus, etc.).
Annotate the dataset with accurate labels specifying the type of disease for each
image
2 Computational Resources:
Utilize a computer with sufficient processing power (CPU/GPU) to handle the
training of machine learning models.
Optionally, consider cloud computing resources (e.g., AWS, Google Cloud) for
scalability.
3 Software and Libraries:
Choose appropriate machine learning frameworks (e.g., TensorFlow, PyTorch) for
model development.
Utilize image processing libraries (e.g., OpenCV) for data preprocessing.
Consider using version control tools (e.g., Git) for managing code.
4 Data Preprocessing Tools:
Image editing tools for resizing, cropping, and adjusting the enset leaf images.
Annotation tools for labeling and marking specific regions of interest in the images.
5 Documentation Tools:
Use tools for documenting your research methodology, code, and results (e.g., Jupyter
Notebooks, Markdown, LaTeX).
Methodology for Enset Leaf Disease Classification:
1. Data Collection:
a. Enset Leaf Image Dataset: - Source of the dataset (e.g., field surveys, existing
databases). - Inclusion criteria for images (various diseases, healthy leaves). - Image
acquisition details (camera specifications, resolution, lighting conditions).
b. Data Annotation: - Annotation process for labeling each image with the
corresponding disease class. - Quality control measures for accurate annotations.
2. Data Preprocessing:
a. Image Resizing and Standardization: - Resize images to a consistent resolution
suitable for model training. - Standardize pixel values to a common scale.
b. Data Augmentation: - Apply data augmentation techniques (rotation, flipping,
zooming) to increase dataset variability and improve model generalization.
3. Transfer Learning Model Selection:
a. Pre-trained Model: - Choose a pre-trained convolutional neural network (CNN)
suitable for transfer learning (e.g., VGG16, ResNet, Inception).
b. Model Adaptation: - Modify the last few layers of the pre-trained model to match
the number of classes in enset leaf diseases.
4. Model Training:
a. Transfer Learning: - Train the modified model on the enset leaf dataset, keeping the
weights of the early layers frozen to retain generic features.
b. Hyperparameter Tuning: - Optimize hyperparameters (learning rate, batch size,
dropout rates) for improved performance.
5. Model Evaluation:
a. Dataset Split: - Divide the dataset into training and validation sets.
b. Performance Metrics: - Evaluate the model using metrics such as accuracy,
precision, recall, and F1-score on the validation set.
c. Confusion Matrix: - Analyze the confusion matrix to understand the model's
performance for each disease class.
6. Comparative Analysis:
a. Traditional Methods: - Implement and evaluate traditional classification methods (if
applicable).
b. Statistical Comparison: - Use statistical tests to compare the performance of the
transfer learning model with traditional methods.
7. Results Analysis:
a. Visualization: - Visualize model predictions, misclassifications, and feature maps.
b. Comparative Analysis: - Compare the results of the transfer learning model with
traditional methods.
8. Interpretation and Discussion:
a. Interpretability: - Interpret the learned features and discuss how the model identifies
different enset leaf diseases.
b. Limitations and Future Work: - Discuss any limitations of the model and propose
potential areas for improvement in future research.
9. Documentation and Reporting:
a. Codebase: - Document the entire codebase, including preprocessing, model
training, and evaluation.
b. Report: - Prepare a comprehensive research report detailing the methodology,
results, and conclusions.
This methodology provides a structured approach to implementing enset leaf disease
classification using a transfer learning approach. Adjustments may be necessary based
on the specifics of your dataset and research goals.
Work plan(timeline)
1. Project Planning and Setup (1 week):
Define project goals and objectives.
Set up the development environment, including necessary libraries and tools.
Define roles and responsibilities within the team.
2. Data Collection and Preprocessing (2 weeks):
Collect enset leaf image dataset.
Annotate the dataset with accurate labels.
Resize and standardize images.
Apply data augmentation techniques.
3. Transfer Learning Model Setup (1 week):
Choose a pre-trained model suitable for transfer learning.
Adapt the model architecture for enset leaf disease classification.
Configure hyperparameters for training.

4. Model Training (4 weeks):

Split the dataset into training and validation sets.
Train the transfer learning model on the enset leaf dataset.
Monitor training progress and adjust hyperparameters if necessary.
Save the trained model weights for future use.
5. Model Evaluation (2 weeks):
Evaluate the model using performance metrics (accuracy, precision, recall, F1-
score).Analyze the confusion matrix and model predictions.
Fine-tune the model if needed based on evaluation results.
6. Results Analysis and Visualization (1 week):
Compare the performance of the transfer learning model with traditional
methods.Visualize model predictions and misclassifications.
Interpret the learned features for each disease class.
7. Documentation and Reporting (2 weeks):
Document the entire methodology, including code and parameters.
Prepare a comprehensive project report.Create a presentation summarizing key
findings.
8. Review and Iteration (1 week):
Review the project with stakeholders or project advisors.Address any feedback and
make necessary adjustments.Iterate on the model or methodology if required.
9. Optional: Model Deployment (1-2 weeks):
If applicable, deploy the trained model for real-world applications.Integrate the model
into a user-friendly interface.
10. Final Review and Submission (1 week):
Conduct a final review of the entire project.
Ensure all documentation and reports are complete.
Submit the project deliverables.
BUNGET AND FUND
1. Data Acquisition and Preparation:
Image collection: Cost varies depending on the method (field acquisition, existing
datasets, crowd sourcing). Expect expenses for equipment, personnel, and potential
licensing fees.
Data annotation: Labeling images for disease classes requires skilled labor. Costs
depend on the dataset size, complexity, and desired annotation quality.Data
preprocessing: Cleaning and formatting images for model training involves software
tools and technical expertise.
2. Computational Resources:
Hardware: Training deep learning models requires powerful GPUs or cloud
computing resources. Rental costs vary based on model complexity and desired
training time.Software: Deep learning frameworks and specialized software tools
might require licensing fees.
3. Personnel and Expertise:
Researchers with expertise in deep learning and transfer learning: Salaries and
research funding contribute to personnel costs.
Domain experts in enset diseases: Consultation with plant pathologists or agricultural
specialists might be necessary.
4. Additional Costs:
Travel and fieldwork expenses: If data collection or validation requires field
visits, travel costs and logistics must be factored in.
Publication and dissemination fees: Sharing research findings through conferences or
publications might incur additional costs.
Funding Sources:
Several funding sources could support enset leaf disease classification research using
transfer learning:
Government grants: National agricultural research agencies or international bodies
like the Food and Agriculture Organization (FAO) might offer funding opportunities.
Non-profit organizations: Foundations focused on food security or agricultural
development could be potential sponsors.
Private sector partnerships: Agritech companies or private donors interested in
sustainable agriculture solutions might be interested in collaborating.
Crowd funding: Public fundraising platforms can be an alternative
approach, especially for early-stage research projects.
 Reference
Chatgpt

Abbas, A. (2016) ‘The effect of agriculture in Ethiopia’, (January 2011).

Aki, O., Gullu, A. and Ucar, E. (2016) ‘Classification of Rice Grains

Using Image Processing And Machine Learning Techniques’,

Proceeding(October), pp. 352–354