🥔 Potato Plant Disease Detection using CNN & Transfer Learning

Overview

This project leverages Deep Learning to accurately identify potato leaf diseases — a crucial step in ensuring sustainable crop yield and food security.
Using a Convolutional Neural Network (CNN) built from scratch and compared against transfer learning (EfficientNet), this work demonstrates how model architecture and data augmentation significantly affect performance in real-world agricultural imaging.

---

Dataset

• Source: Kaggle – PlantVillage Dataset
• Classes:
  • Potato Early Blight – 1000 images
  • Potato Late Blight – 1000 images
  • Potato Healthy – 152 images → augmented to 1000 images

After augmentation, all classes were balanced with 1000 images each (3000 total).

---

Data Preprocessing

• Image Size: 224 × 224 pixels
• Normalization: Pixel values scaled to [0, 1] using keras.layers.Rescaling
• Augmentation Techniques:
  • Random rotation, shear, and zoom
  • Width/height shifts
  • Brightness variation
  • Horizontal flipping
• Data Split:
  • 80% Training
  • 10% Validation
  • 10% Testing

---

Model Architectures

Custom CNN Model

A robust CNN architecture was designed with multiple convolutional blocks for hierarchical feature extraction.

model = Sequential([
    rescale,
    data_augment,

    # Block 1
    Conv2D(32, (3,3), activation='relu'),
    MaxPooling2D((2,2)),
    Dropout(0.2),
    Conv2D(32, (3,3), activation='relu'),
    MaxPooling2D((2,2)),
    Dropout(0.2),

    # Block 2
    Conv2D(64, (3,3), activation='relu'),
    MaxPooling2D((2,2)),
    Dropout(0.2),
    Conv2D(64, (3,3), activation='relu'),
    MaxPooling2D((2,2)),
    Dropout(0.2),

    # Block 3
    Conv2D(128, (3,3), activation='relu'),
    MaxPooling2D((2,2)),
    Dropout(0.2),
    Conv2D(128, (3,3), activation='relu'),
    MaxPooling2D((2,2)),
    Dropout(0.2),

    # Fully Connected Layers
    Flatten(),
    Dense(64, activation='relu'),
    Dropout(0.2),
    Dense(3, activation='softmax')
])

Training Configuration

• Optimizer: Adam
• Loss Function: Categorical Crossentropy
• Batch Size: 32
• Epochs: ~20

Adding Batch Normalization led to unstable convergence, so it was excluded from the final architecture.

---

Transfer Learning (EfficientNet)

A pre-trained EfficientNet model was fine-tuned for comparison.
However, due to domain differences and small dataset size, the model underperformed, achieving:

Metric	Value
Training Accuracy	0.33
Validation Accuracy	0.33

---

📈 Performance Metrics

Model	Train Acc	Val Acc	Test Acc	Test Precision	Test F1 Score
Custom CNN	0.95	0.94	0.96	0.96	0.96
EfficientNet (TL)	0.30	0.30	–	–	–

---

Key Insights

• Custom CNN achieved superior accuracy (96%), proving that well-tuned small models can outperform complex pre-trained networks for specific domains.
• Batch Normalization disrupted stability — likely due to limited batch size or aggressive learning rate.
• Data augmentation played a pivotal role in balancing the dataset and improving generalization.
• Transfer learning struggled due to domain mismatch (generic ImageNet features vs. leaf texture patterns).

---

Conclusion

• A carefully engineered CNN can outperform modern transfer learning models for focused, domain-specific tasks.
• Achieved 96% test accuracy with strong precision and F1 score, validating the model’s real-world reliability.
• Demonstrates the potential of AI to assist farmers and agronomists in early disease detection.

---

Future Work

• Experiment with ResNet, InceptionV3, and MobileNetV2
• Deploy the trained model as a web or mobile app for real-time disease diagnosis

---