🌱 CAPSTONE-LAZARUS: AI Plant Disease Detector

CAPSTONE-LAZARUS is a robust, production-ready computer vision system that detects plant diseases from leaf images using deep learning. Designed for farmers, agronomists, and agricultural support systems to enable quick diagnosis and corrective action.

🚀 Recent Improvements (2025)

✅ Python 3.11+ Compatibility with pinned dependencies
✅ Defensive Programming throughout data pipeline
✅ Deterministic Training with reproducible seeds
✅ Comprehensive Testing with CI/CD pipeline
✅ Clean EDA Notebook replacing corrupted version
✅ Robust Model Loading with proper error handling
✅ Environment Validation scripts for quick setup verification

🔧 Quick Validation

# Validate environment setup
python scripts/validate_environment.py

# Test data loader functionality  
python scripts/validate_loader.py

# Test inference functions
python scripts/validate_inference.py

# Run smoke tests
python tests/test_data_utils.py

---

🎯 Objectives

• Robust multi-class classification across 19 plant disease classes
• High recall on critical diseases (>90%) to minimize false negatives
• Interpretable predictions with confidence scores and Grad-CAM visualizations
• Lightweight deployment for mobile/edge devices
• Interactive dashboard with Plotly visualizations

---

📊 Dataset

26,134 RGB leaf images across 19 disease classes:

🌽 Corn (Maize)

• Healthy
• Cercospora leaf spot / Gray leaf spot
• Common rust
• Northern Leaf Blight (+ oversampled/undersampled variants)

🥔 Potato

• Healthy
• Early blight
• Late blight

🍅 Tomato

• Healthy
• Bacterial spot
• Early blight
• Late blight
• Leaf Mold
• Septoria leaf spot
• Spider mites Two-spotted spider mite
• Target Spot
• Tomato mosaic virus
• Tomato Yellow Leaf Curl Virus

---

🛠️ Architecture

📂 Capstone-Lazarus/
├── 📊 data/                     # 26K+ plant disease images
├── 📓 notebooks/                # Jupyter notebooks for EDA & training
│   ├── eda_plant_diseases.ipynb      # Comprehensive data exploration  
│   └── model_training.ipynb          # Model training & evaluation
├── 🧠 models/                   # Saved trained models
├── 📱 app/                      # Streamlit dashboard
│   └── streamlit_app.py
├── 🔧 src/                      # Core ML modules
│   ├── data_utils.py                 # Data loading & preprocessing
│   ├── model_factory.py              # Model architectures
│   ├── training.py                   # Training pipeline
│   └── inference.py                  # Prediction & explanation
└── 📋 requirements.txt          # Dependencies

---

🚀 Quick Start

# 1. Install dependencies
pip install -r requirements.txt

# 2. Validate environment
python scripts/validate_environment.py

# 3. Explore data
jupyter notebook notebooks/eda_plant_diseases.ipynb

# 4. Train models (see Local Training & Dashboard below)
jupyter notebook notebooks/master_model_trainer.ipynb

# 5. Launch dashboard
streamlit run app/lazarus_console/__init__.py

---

🎓 Local Training & Dashboard

Training Models

Option 1: Jupyter Notebook (Recommended)

# Full training run
jupyter notebook notebooks/master_model_trainer.ipynb

# Or convert and execute directly
jupyter nbconvert --execute notebooks/master_model_trainer.ipynb

Option 2: Python CLI

# Fast smoke test (1 epoch, small sample)
python -m src.master_trainer --fast-test

# Full training from config
python -m src.master_trainer --config config.yaml

For Low-Spec Machines (HP ZBook 15 G5, 16GB RAM):

Recommended settings in notebook or config.yaml:

• batch_size: 2-4 (instead of 8)
• image_size: 160 (instead of 224)
• fast_test: true for initial runs
• Train one model at a time

Launching the Dashboard

# Primary method
streamlit run app/lazarus_console/__init__.py

# Or use convenience wrapper
python run.py

Dashboard Features:

• 🏠 Mission Readiness: Latest metrics, checkpoints feed
• 🔬 Inference Lab: Batch upload, ensemble mode, confidence thresholds
• 🧬 Explainability Studio: Grad-CAM with opacity control
• 📊 Model Comparison: Side-by-side metrics, confusion matrices

Running Tests

# Smoke test suite (PowerShell on Windows)
.\run_smoke_tests.ps1

# Or Bash (Linux/macOS/Git Bash)
./run_smoke_tests.sh

# Individual test files
pytest tests/test_smoke_train.py -v
pytest tests/test_streamlit_integration.py -v
pytest tests/test_master_trainer.py -v

Profiling Inference

# CPU profiling
python scripts/profile_inference.py \
  --model models/{run_id}/best.pth \
  --device cpu \
  --runs 100

# GPU profiling (if CUDA available)
python scripts/profile_inference.py \
  --model models/{run_id}/best.pth \
  --device cuda \
  --batch-size 8

For detailed troubleshooting, see docs/runbook.md.

---

📈 Success Metrics

• Macro F1 Score: ≥ 0.85
• Critical Disease Recall: ≥ 90%
• Model Size: ≤ 50MB
• Inference Time: ≤ 500ms

---

🔍 Key Features

• Transfer Learning with EfficientNet, ResNet, MobileNet
• Advanced Data Augmentation for field conditions
• Class-weighted Loss for imbalanced datasets
• Grad-CAM Explanations for model interpretability
• Uncertainty Estimation with Monte Carlo Dropout
• Interactive Dashboard with Plotly visualizations

---

🌟 Impact

• Early disease detection → reduced crop losses
• Precision agriculture → optimized treatment
• Farmer empowerment → AI-assisted decisions
• Sustainable farming → reduced pesticide usage