🌿 Plant Disease Prediction System

Architecture: EfficientNetB3 + Transfer Learning

Why EfficientNetB3?

Model	Params	Val Accuracy	Notes
VGG16	138M	~90%	Too heavy, overfits
ResNet50	25M	~95%	Good baseline
DenseNet121	8M	~95%	Memory intensive
EfficientNetB3	12M	~98%	Best balance

EfficientNet uses compound scaling — balancing depth, width, and resolution together — which is optimal for fine-grained leaf texture classification.

---

Quick Start

1. Install dependencies

pip install -r requirements.txt

2. Get the dataset

# Option A: Check dataset stats for the augmented dataset
python dataset_setup.py --stats

3. Train

You can train the model in two ways:

Option A: Interactive Jupyter Notebook (Recommended)

Launch the interactive Jupyter notebook plant_disease_training.ipynb to train the model, view custom data visualizations, generate detailed performance evaluation metrics, and explore Grad-CAM explainability maps step-by-step:

jupyter notebook plant_disease_training.ipynb

Option B: CLI Script

Run the script to train the model through the command-line interface:

python plant_disease_model.py

4. Predict on new images

from plant_disease_model import PlantDiseasePredictor

predictor = PlantDiseasePredictor("checkpoints/best_model.pth")
result    = predictor.predict("my_leaf.jpg")
print(result)
# {
#   "predicted_class": "Tomato___Late_blight",
#   "confidence": 97.42,
#   "top_k_predictions": [...]
# }

5. Grad-CAM explainability

from plant_disease_model import visualize_gradcam, PlantDiseaseClassifier
import torch, json

class_names = json.load(open("logs/class_names.json"))
model = PlantDiseaseClassifier(len(class_names), pretrained=False)
ckpt  = torch.load("checkpoints/best_model.pth")
model.load_state_dict(ckpt["model_state"])

visualize_gradcam(model, "my_leaf.jpg", class_names, "results/gradcam.png")

6. Flask web app integration

Run the Flask web app to add user registration, login, image upload, and prediction result pages.

pip install -r requirements.txt
python app.py

Then open http://127.0.0.1:5000 in your browser.

• Register a new account
• Login
• Upload a leaf image (jpg/jpeg/png)
• View prediction and confidence results

---

Dataset: New Plant Diseases Dataset (Augmented)

• 87,867 images across 38 disease classes (70,295 training, 17,572 validation)
• 14 plant species: Apple, Blueberry, Cherry, Corn, Grape, Orange, Peach, Pepper, Potato, Raspberry, Soybean, Squash, Strawberry, Tomato
• Covers both diseased and healthy leaves

---

Training Pipeline

Raw Leaf Images
      │
      ▼
Data Augmentation (RandomCrop, Flip, Rotation, ColorJitter, RandomErasing)
      │
      ▼
EfficientNetB3 Backbone (ImageNet pretrained, frozen)
      │
      ▼ [Epoch 5: unfreeze backbone for fine-tuning]
      │
Custom Head: Linear(1536→512) → BN → SiLU → Dropout(0.4)
           → Linear(512→256)  → BN → SiLU → Dropout(0.2)
           → Linear(256→38)
      │
      ▼
CrossEntropyLoss (label smoothing=0.1)
AdamW + CosineAnnealingLR
Early Stopping (patience=7)

---

Output Files

checkpoints/
  best_model.pth        ← best checkpoint (by val accuracy)
logs/
  class_names.json      ← class index mapping
results/
  training_history.png  ← loss & accuracy curves
  confusion_matrix.png  ← per-class confusion matrix
  gradcam.png           ← Grad-CAM visualization

---

Expected Performance (Augmented Dataset)

Metric	Value
Test Accuracy	~97–98%
Training Time (GPU)	~45 min
Training Time (CPU)	~8 hours
Model Size	~48 MB
Inference Speed	~5ms/image (GPU)