Pathology Detection in Crop Plants

Overview

This project investigates automated classification of diseases in apple leaves using advanced machine learning and deep learning techniques. Leveraging a high-quality image dataset, we combine traditional machine learning models and convolutional neural networks (CNNs) to develop scalable methods for plant disease detection. The proposed approaches support technology-driven solutions for sustainable agriculture.

---

Team Members

• Domenico Azzarito
• Guillermo Bajo Laborda
• Laura Alejandra Moreno
• Arian Gharehmohammadzadehghashghaei
• Michele Pezza

---

Dataset

The dataset is sourced from the Kaggle competition: Plant Pathology 2020 - FGVC7. It contains 1,821 labeled images of apple leaves categorized into four classes:

• Healthy: No visible symptoms of disease.
• Rust: Fungal infections with rust-like symptoms.
• Scab: Fungal lesions causing scab-like patterns.
• Multiple Diseases: Symptoms of more than one disease present.

---

Methodology

Exploratory Data Analysis (EDA)

• Class Distribution: Visualized using pie charts and histograms, highlighting the imbalance in the "Multiple Diseases" class.
• Color Analysis: Computed RGB histograms and channel statistics to identify class-specific patterns in leaf images.

Preprocessing

• Edge Detection: Applied Canny filtering to isolate leaf regions and remove background noise.
• Feature Engineering: Extracted RGB histograms and Gray-Level Co-occurrence Matrix (GLCM) features for machine learning models.
• Data Augmentation: Addressed class imbalance through synthetic oversampling and geometric transformations.

Machine Learning Models

1. Softmax Regression: Baseline multiclass prediction model.
2. Random Forest: Ensemble learning method to improve accuracy.

Deep Learning Models

1. ResNet-50
   • Fine-tuned residual connections to capture complex hierarchical features.
   • Addressed gradient issues through skip connections.
2. VGG-16
   • Optimized sequential layers for extracting fine-grained patterns.
   • Balanced depth and simplicity for efficient classification.

---

Results

Performance Comparison

Model	Accuracy
Softmax Regression	77.0%
Random Forest	81.1%
ResNet-50	83.52%
VGG-16	86.81%

Per-Class Performance

Class	ResNet-50 Recall	VGG-16 Recall
Healthy	81.0%	87.9%
Multiple Diseases	10.0%	30.0%
Rust	92.2%	90.6%
Scab	90.0%	92.0%

---

Key Observations

• VGG-16: Achieved the highest accuracy of 86.81%, handling fine-grained features effectively.
• ResNet-50: Performed well (83.52%) but struggled with rare categories due to fewer samples and complexity.
• Random Forest: Outperformed Softmax Regression, emphasizing the need for non-linear decision boundaries.
• Class Imbalance: The "Multiple Diseases" class remained a challenge, highlighting the need for oversampling and augmentation techniques.

---

Future Work

• Model Improvements: Fine-tune CNN architectures with learning rate schedules and dropout enhancements.
• Advanced Networks: Evaluate modern architectures like EfficientNet and Vision Transformers for improved scalability and accuracy.
• Preprocessing Enhancements: Integrate leaf segmentation and de-noising algorithms to enhance image clarity.
• Imbalance Handling: Expand augmentation methods, including generative approaches, to enrich minority classes.
• Mobile Deployment: Develop lightweight models for mobile and IoT platforms, enabling real-time disease detection in agricultural fields.

---

Conclusion

This study demonstrated the potential of combining traditional machine learning techniques with modern CNN architectures for plant disease detection. While CNNs like VGG-16 and ResNet-50 achieved high accuracy, challenges remain in handling underrepresented classes.

Key Findings:

• CNNs outperformed traditional models, achieving accuracies of 86.81% (VGG-16) and 83.52% (ResNet-50).
• Random Forest achieved 81.1%, highlighting the effectiveness of ensemble methods.
• Softmax Regression served as a baseline with 77.0% accuracy.
• Class imbalance, especially in the "Multiple Diseases" category, limits model performance, emphasizing the need for targeted augmentation and oversampling.

---

Roles

• Arian Gharehmohammadzadehghashghaei: Exploratory Data Analysis and Model Evaluation.
• Laura Alejandra Moreno and Guillermo Bajo Laborda: Machine Learning Models (Softmax, Random Forest).
• Domenico Azzarito and Michele Pezza: Deep Learning Models (CNNs - ResNet-50, VGG-16).

---

References

• Kaggle Dataset: Plant Pathology 2020 - FGVC7
• Mohanty, Sharada P., David P. Hughes, and Marcel Salathé. "Using deep learning for image-based plant disease detection." Frontiers in plant science 7 (2016): 1419.