🦅 BirdNest Detection System

📋 Project Background

The South Florida/Caribbean Inventory and Monitoring Network (SFCN) has monitored colonial nesting birds in Biscayne National Park since 2010 using low-level aerial photography. This deep learning model automates nest detection to increase monitoring efficiency and standardize detection across observers.

📊 Dataset

• Time Range: 2010–2024 (15 years)
• Total Locations: 866 monitoring sites
• Annotated Images: ~19,786 circled images
• Training Labels: ~16,330 YOLO format annotations
• Classes: occupied, eggs, chicks, non-occupied

🎯 Model Performance

• Architecture: YOLOv5s6 (normalized dataset)
• Best mAP@0.5: ~0.44 (curated dataset)
• Occupied Nests: ~62% recall at lower confidence thresholds
• Training: AWS SageMaker (ml.g5.12xlarge, 50 epochs)

✅ Current Status

Model Deployed: May 2025 (normalized dataset, 04192025_normalized_m)
Deliverables: Dataset, trained models, evaluation reports, and web interface completed
Next Steps: Dataset publication via FIU Envistor → Pelican → Dataverse (FAIR metadata)