🌾 Early Wheat Disease Detection Using Deep Learning (PyTorch)

📌 Project Overview

This project implements an Early Wheat Disease Detection system using ResNet-18, a convolutional neural network pre-trained on ImageNet, with transfer learning. The system classifies wheat leaf images into three categories:

• Healthy
• Septoria
• Stripe Rust

The goal is to support early identification of wheat leaf diseases and assist timely agricultural intervention.

---

🎯 Objectives

• Detect wheat diseases from leaf images at an early stage
• Use a pre-trained CNN model to improve classification performance
• Apply image preprocessing and data augmentation techniques
• Evaluate the model using standard classification metrics
• Demonstrate a complete deep learning workflow using PyTorch

---

🧠 Model Architecture

• Base Model: ResNet-18
• Pre-training: ImageNet weights
• Framework: PyTorch
• Approach: Transfer Learning
• Input Image Size: 224 × 224 RGB images

Transfer Learning Strategy

• The pre-trained convolutional layers of ResNet-18 were frozen.
• The final fully connected layer was replaced according to the number of wheat disease classes.
• Only the final classification layer was trained initially.

---

🛠️ Technology Stack

• Programming Language: Python
• Deep Learning Framework: PyTorch, TorchVision
• Image Processing: PIL, OpenCV
• Data Handling: NumPy
• Visualization: Matplotlib, Seaborn
• Model Evaluation: Scikit-learn

---

📂 Dataset

The dataset contains wheat leaf images categorized into healthy and diseased classes.

Classes Used

Class	Description
Healthy	Wheat leaves without visible disease symptoms
Septoria	Wheat leaves affected by Septoria disease
Stripe Rust	Wheat leaves affected by stripe rust disease

Dataset Source

Dataset: Mendeley Dataset Repository
Link: https://data.mendeley.com/datasets/wgd66f8n6h/1

---

🧪 Methodology

1. Image Preprocessing

Each image was resized to 224 × 224 pixels to match the input size required by ResNet-18.

Validation and test images were normalized using ImageNet normalization values:

mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]

2. Data Augmentation

To improve generalization, the following augmentation techniques were applied to the training images:

• Random horizontal flipping
• Random rotation
• Color jittering
• Normalization

3. Model Training

The model was trained using:

• Loss Function: CrossEntropyLoss
• Optimizer: Adam
• Learning Rate: 0.0001
• Batch Size: 32
• Epochs: 10

4. Model Evaluation

The trained model was evaluated using:

• Accuracy
• Precision
• Recall
• F1-score
• Confusion matrix

---

📊 Results

The model achieved a best validation accuracy of 58.33%.

Classification Report

Class	Precision	Recall	F1-Score	Support
Healthy	0.75	0.19	0.30	16
Septoria	0.75	0.19	0.30	16
Stripe Rust	0.55	0.97	0.70	32
Accuracy			0.58	64
Macro Avg	0.68	0.45	0.43	64
Weighted Avg	0.65	0.58	0.50	64

---

📉 Confusion Matrix

The confusion matrix shows that the model performed well on the Stripe Rust class but struggled to correctly classify Healthy and Septoria images.

Confusion Matrix Values

Actual \ Predicted	Healthy	Septoria	Stripe Rust
Healthy	3	1	12
Septoria	0	3	13
Stripe Rust	1	0	31

---

🔍 Result Analysis

The model achieved moderate performance overall, with strong recall for the Stripe Rust class. However, the model showed poor recall for the Healthy and Septoria classes.

Key Observations

• The model correctly identified 31 out of 32 Stripe Rust samples.
• The model misclassified many Healthy and Septoria images as Stripe Rust.
• This indicates a class bias toward the Stripe Rust category.
• The overall accuracy was 58%, which shows that the model needs improvement before practical deployment.

---

⚠️ Limitations

• The dataset appears imbalanced, with more Stripe Rust samples than the other classes.
• The model is biased toward predicting Stripe Rust.
• Healthy and Septoria classes have low recall.
• Only the final layer of ResNet-18 was trained, which may limit disease-specific feature learning.
• More training data and stronger validation are needed for reliable real-world use.

---

🚀 Future Improvements

• Use a balanced dataset or apply class weighting.
• Fine-tune deeper layers of ResNet-18, especially layer4.
• Increase the number of training epochs with early stopping.
• Use stronger augmentation techniques.
• Try other architectures such as EfficientNet, DenseNet, or MobileNetV2.
• Apply Grad-CAM for explainability and visual interpretation.
• Improve dataset quality by removing blurry, duplicate, or noisy images.

---

🧪 Sample Prediction

The trained model can predict a disease class for a single wheat leaf image using the saved model weights.

prediction, confidence = predict_image("sample.jpg")
print(f"Prediction: {prediction}")
print(f"Confidence: {confidence:.2f}%")

---

💾 Model Saving

The trained model weights were saved using:

torch.save(model.state_dict(), "wheat_disease_detection_pytorch.pth")

---

📚 Key Learnings

• Practical implementation of CNN-based image classification using PyTorch
• Use of transfer learning with ResNet-18
• Importance of image preprocessing and augmentation
• Evaluation using precision, recall, F1-score, and confusion matrix
• Understanding the impact of class imbalance on model predictions
• Need for validation and unbiased testing in deep learning projects

---

🤝 Acknowledgements

• Mendeley Dataset Repository for providing the wheat disease dataset
• PyTorch and TorchVision communities for open-source deep learning tools
• Scikit-learn for model evaluation utilities

---

📌 Conclusion

This project demonstrates a PyTorch-based wheat disease classification system using ResNet-18 and transfer learning. The current model shows promising performance for detecting Stripe Rust but requires further improvement for reliable classification of Healthy and Septoria classes.

The project provides a strong foundation for future development in agricultural disease detection using deep learning.