Coffee Disease Detection

Project Overview

Coffee Disease Detection is an advanced computer vision system designed to identify diseases in coffee plants from leaf images. Originally presented as a winning idea for Le Wagon’s Data Science & AI bootcamp, this project was selected as one of two final projects to be developed by a team of four students.

As project leader, I coordinated the development team and guided the creation of a comprehensive solution that can detect five different classes: healthy leaves and four common diseases (Cercospora, Leaf Rust, Miner, and Phoma) that significantly impact coffee crop yields.

Key Features

🧠 Multiple Model Architectures

• VGG16 Transfer Learning - Stable performance with anti-overfitting measures
• EfficientNetB0 - Optimized for accuracy and efficiency
• Custom CNN - Lightweight model for smaller datasets
• Automatic architecture detection and model loading

🎯 Disease-Focused Optimization

• Custom class weighting prioritizing disease detection
• Disease-specific recall metrics to minimize false negatives
• False negative analysis tools for model improvement
• Balanced training with adaptive learning rates

🚀 Production-Ready Deployment

• FastAPI backend with comprehensive API documentation
• Streamlit web application for easy image upload and prediction
• Google Cloud deployment with Dockerized API for scalable inference
• MLflow integration for experiment tracking and model versioning
• Modular Python package structure for maintainable codebase

Technical Architecture

Model Training Pipeline

• Multi-phase training with initial feature extraction and fine-tuning
• Custom callbacks for disease recall monitoring
• Adaptive learning rates based on dataset size
• Comprehensive evaluation with confusion matrices and detailed metrics

Data Processing

• Smart data augmentation preserving disease characteristics
• Letterboxing preprocessing to 224x224 resolution
• Automatic train/validation/test splitting
• Class distribution analysis and balancing

API and Deployment

• RESTful API with image upload and prediction endpoints
• Base64 and file upload support for flexible integration
• Comprehensive error handling and validation
• Production caching system for model loading optimization
• Google Cloud Platform deployment for scalable cloud-based inference
• Separate web frontend repository available at https://github.com/fermx3/coffeedd-web for the Streamlit application

Project Leadership & Team Coordination

Team Management

• Led a four-person development team through the complete project lifecycle
• Coordinated tasks and timelines ensuring timely delivery of milestones
• Applied software engineering best practices including version control and documentation
• Maintained comprehensive GitHub repository with clear structure and technical documentation

Le Wagon Bootcamp Achievement

• Winning project selection from multiple proposed ideas in Le Wagon’s Data Science & AI bootcamp
• One of two final projects chosen for development by the class
• End-to-end project delivery from initial concept to production deployment

Disease Classification

The system detects these coffee plant conditions:

1. Healthy - Normal, disease-free coffee leaves
2. Cercospora - Fungal disease causing brown leaf spots
3. Leaf Rust - Orange/yellow pustules on leaf undersides
4. Miner - Insect damage creating tunnels in leaves
5. Phoma - Fungal disease causing leaf blight

Technical Highlights

Smart Class Weighting

Custom weighting system that penalizes healthy misclassification while boosting rare disease detection:

• Healthy leaves: Reduced weight to prevent false negatives
• Rare diseases: Increased weight for better detection
• Adaptive weighting: Based on dataset size and distribution

Advanced Metrics

• Disease Recall Metric: Custom TensorFlow metric focusing on disease detection
• Binary classification analysis: Healthy vs. any disease performance
• Detailed confusion matrices with class-specific insights
• False negative analysis for continuous improvement

Model Management

• MLflow tracking for experiment management
• Automatic model registry with versioning
• Production model deployment with A/B testing capability
• Model architecture detection for seamless loading

Impact and Applications

• ✅ Early disease detection enabling timely treatment
• ✅ Reduced crop losses through accurate identification
• ✅ Farmer accessibility via simple web interface
• ✅ Scalable deployment for agricultural organizations
• ✅ Research foundation for agricultural AI development

The system represents a significant advancement in precision agriculture, providing coffee farmers with AI-powered tools to protect their crops and optimize yields through early disease detection.