// ACADEMIC CONTENT PORTAL · Computer Science

Thesis Topic Ideas: Machine Learning Based Crop Disease Detection in Nepal

Level: Bachelor, MasterDifficulty: Intermediate★ Popular Choice

Applicable Universities:Kathmandu University (KU)Tribhuvan University (TU)Pokhara University (PU)University of CambridgeUniversity of Oxford

🎓 Academic Journey Progression Map

Grade 8-10 (SEE)
School Projects
+2 (Grades 11-12)
Stream Projects
Bachelor's Degree
Undergraduate Thesis
Master's Degree
Postgraduate Thesis
PhD Doctorate
Dissertation Guidance

1. Introduction & Problem Statement

Overview: Build a CNN-based image classification system to detect crop diseases in Nepali agricultural context — rice, wheat, potato. Train on local disease datasets.

Background Context (Nepal): Nepal's agriculture sector employs a majority of the workforce but faces crop losses due to diseases like late blight in potatoes and rice blast. Automated detection through mobile apps helps farmers diagnose early. Verify current NARC disease prevalence figures prior to thesis submission.

2. Research Objectives

›Collect and annotate crop disease image dataset from Nepal agricultural research stations
›Train CNN model (ResNet/MobileNet) for disease classification
›Achieve 85%+ validation accuracy on local field dataset
›Build mobile-friendly inference interface using TensorFlow Lite
›Compare model accuracy against expert manual diagnosis

3. Proposed Methodology

Literature review on plant disease detection via Kaggle and NepJOL publications
Dataset collection from NARC research stations and field surveys in Kavre and Jhapa
Image preprocessing including resizing, normalization, and data augmentation
Transfer learning using PyTorch/TensorFlow with MobileNetV3 architecture
Evaluation using precision, recall, confusion matrix, and mobile inference latency benchmark

$ Worked Example / Sample Scenario

Sample Scenario: A farmer in Kavre notices yellow spots on potato leaves. They photograph the leaf with a low-cost Android phone. The offline TFLite model detects Late Blight with 91% confidence and displays organic treatment advice in Nepali.

4. Thesis Chapter-by-Chapter Outline

Chapter 1: IntroductionTU/KU/NEB standard

Background, problem statement, research questions, objectives, scope, limitations, and significance of the study

Chapter 2: Literature ReviewTU/KU/NEB standard

Theoretical framework, conceptual models, previous empirical studies in Nepal and developing nations, review of CNN architectures and vision transformers, and gap analysis

Chapter 3: Research MethodologyTU/KU/NEB standard

Research design, population/sampling framework, data collection instruments, analytical tools, and ethical considerations

Chapter 4: Data Analysis & ResultsTU/KU/NEB standard

Empirical findings, statistical testing, model estimations, confusion matrix and class-wise accuracy metrics, and detailed discussion

Chapter 5: Conclusion & RecommendationsTU/KU/NEB standard

Summary of key findings, theoretical contributions, policy recommendations, and future research directions

5. Recommended Tools & Technologies

To implement the practical, technical, or analytical portions of this thesis topic, the following software tools, libraries, or APIs are recommended:

PythonTensorFlow/PyTorchOpenCVTensorFlow LiteFlutter / React NativeGoogle Colab

6. Core References & Academic Sources

[1]NepJOL (Nepal Journals Online) — Agricultural Sciences section papers on plant pathology
[2]Google Scholar — Deep Learning for Plant Disease Classification research papers
[3]NARC (Nepal Agricultural Research Council) Annual Technical Reports

7. Frequently Asked Questions (FAQs)

Q: Where can I collect local agricultural image datasets in Nepal?

You can request access to NARC (Nepal Agricultural Research Council) research stations or augment public datasets like PlantVillage with local field photos taken in Kavre or Terai districts.

Q: Is PyTorch or TensorFlow better for this thesis?

PyTorch is popular for academic research and quick experimentation, while TensorFlow Lite provides easier mobile export pipelines.

Q: How many images do I need per class?

Using transfer learning (e.g. MobileNet or ResNet), 300–500 quality images per disease class with data augmentation is typically sufficient for a Bachelor thesis.

Q: Can this be implemented as a web application instead of mobile?

Yes, using TensorFlow.js or ONNX Runtime Web, you can run model inference directly in a mobile web browser without requiring installation.