Individual canopy tree species maps for the National Ecological Observatory Network.

Area of Science:

• Ecology
• Computer Vision
• Remote Sensing

Background:

• Forest ecosystem health relies on accurate tree composition data.
• Individual tree data at broad scales is crucial for ecological analysis and monitoring.
• Deep learning offers a novel approach to extract fine-grained tree information from sensor data.

Purpose of the Study:

• To develop landscape-level predictions of individual tree species, crown location, area, and height.
• To create open-source datasets for ecological research and forest management.
• To leverage computer vision for large-scale forest ecosystem analysis.

Main Methods:

• Utilized over 40,000 individual tree stems for training deep neural networks.
• Applied hierarchical multi-temporal models fine-tuned for 24 National Ecological Observatory Network (NEON) sites.
• Generated predictions for over 100 million individual trees.

Main Results:

• Achieved an average accuracy of 79% for site-specific tree species predictions.
• Mapped 81 canopy tree species across 24 NEON sites.
• Produced open-source data in 1 km2 shapefiles, including crown attributes and species identification.

Conclusions:

• The generated individual tree data provides valuable insights for forest ecology and restoration.
• Openly archived data on Google Earth Engine enhances accessibility for the scientific community.
• Future work can improve predictions through targeted data sampling strategies.