LeafiVision

Deep Learning Based Pest and Disease Detection in Crops

LeafiVision – Automated Groundnut Pest & Disease Detection Using Custom CNN

LeafiVision is a deep-learning–based image classification system designed to automatically detect major pests and diseases affecting groundnut (peanut) crops. Using a custom Convolutional Neural Network (CNN), the project achieves 97% accuracy across 8 classes, offering a lightweight model suitable for real-time or smartphone-based crop diagnosis.

---

📑 Table of Contents

• Introduction
• Features
• Dataset
• Methodology
• Model Architectures
• Results
• Future Scope

---

🌱 Introduction

Groundnut (Arachis hypogea) is a globally important oilseed crop grown across 120+ countries. Biotic stresses—particularly leaf diseases and damaging insect pests—can drastically reduce yield and quality. Early diagnosis is essential but difficult for non-experts.

This project provides a computer-vision–based automated diagnosis system using a Custom CNN trained on field-collected images from ICRISAT. The model distinguishes between multiple diseases, pests, healthy leaves, and even detects when an image is not groundnut.

---

⭐ Features

• Custom-designed 3-layer and 6-layer CNN architectures
• Achieves up to 97% accuracy
• 8-class classification:
  • Alternaria Leaf Spot
  • Late & Early Leaf Spot
  • Rust
  • Leaf Hopper & Jassids
  • Leaf Miner
  • Tobacco Caterpillar
  • Healthy
  • Not Groundnut
• Includes data augmentation (random rotation, flipping)
• Lightweight architecture suitable for mobile deployment
• Trained using TensorFlow + Keras on Google Colab

---

🗂 Dataset

• Total images: 4867
• Collected from ICRISAT fields using mobile devices.
• Additional “Not Groundnut” images sourced from Kaggle datasets.
• Split:
  • 80% Training
  • 10% Validation
  • 10% Testing

Class	Description	Images
0	Alternaria Leaf Spot	663
1	Healthy	590
2	Leaf Hopper & Jassids	771
3	Leaf Miner	849
4	Not Groundnut	510
5	Late & Early Leaf Spot	414
6	Rust	301
7	Tobacco Caterpillar	769

---

🔧 Methodology

1. Image Acquisition: Field collection using mobile cameras
2. Manual Preprocessing: Removing blurry images, mixed infections
3. Data Augmentation: Random flip & rotation
4. Model Training: Multiple custom CNN models
5. Evaluation: Selecting the best performing model
6. Prediction: Softmax confidence scoring on outputs

---

🧠 Model Architectures

Model 1 – Custom CNN (3 Layers)

• Convolutional blocks: 3
• Filters: 16 → 32 → 64
• Max pooling layers
• Dense layer (128 units)
• Output: 7 classes

Model 2 – Custom CNN (6 Layers)

• Convolutional blocks: 6
• Filters: 32, then five layers of 64
• Max pooling after each convolution
• Dense layer: 64 units
• Softmax output: 8 classes
• Achieved highest accuracy (97%)

---

📊 Results

Model	Layers	Data Augmentation	Epochs	Accuracy	Classes
Model 0	3	No	15	89%	6
Model 1	3	Yes	15	91%	7
Model 2	6	No	30	93%	8
Model 3 (Best)	6	Yes	30	97%	8

---

🔮 Future Scope

• Expand dataset with more diseases/pests
• Increase variability in Not Groundnut class using open datasets
• Build a hierarchical classification pipeline:
  • Plant vs Non-plant
  • Groundnut vs Other Plant
  • Healthy vs Diseased
  • Disease/Pest classifier
• Deploy as a mobile application with on-device lightweight model
• Use explainability methods (GradCAM) to visualize disease regions

---

✨ Acknowledgments

• ICRISAT for real-field image collection and expert labeling