NeonTreeEvaluation: Benchmark for Individual Tree Detection from Airborne Imagery

NeonTreeEvaluation Benchmark provides multi-modal airborne remote sensing data for developing and testing individual tree detection algorithms. The dataset includes RGB images, LiDAR tiles, 426-band hyperspectral files, and 1-meter canopy height models for multiple geographic sites identified by NEON codes. Annotations for tree locations are provided separately, and the evaluation set includes unannotated data for testing or semi-supervised learning.

Use Cases

• Training tree detection models using the provided RGB, LiDAR, and hyperspectral data.
• Benchmarking algorithm performance against the provided ground-truth annotations.
• Developing semi-supervised learning methods using the unannotated evaluation data.
• Studying forest ecology and structure across different NEON field sites.

Strengths

• Provides a multi-modal benchmark with four complementary data types (RGB, LiDAR, hyperspectral, CHM).
• Includes data from multiple geographic sites under the NEON program, suggesting ecological diversity.
• Contains a structured evaluation set with both annotated and unannotated data for flexible use.

Limitations

• Key metadata such as row count, file size, and specific update dates are missing from all sources.
• The description notes that not all evaluation files have corresponding annotations, which may limit supervised use for some files.
• The dataset's geographic and temporal coverage is not explicitly detailed beyond the use of NEON site codes.

Provenance

Source

University of Florida, authored by Ben Weinstein.

Collection Method

Airborne remote sensing data collection (RGB, LiDAR, hyperspectral) under the NEON program.

Geography

Multiple geographic sites within the NEON (National Ecological Observatory Network) domain, identified by four-letter codes (e.g., HARV for Harvard Forest).