NeonTreeEvaluation: Airborne Imagery for Individual Tree Detection

426-band hyperspectral files, along with RGB imagery, LiDAR tiles, and canopy height data, form the training data for the NeonTreeEvaluation benchmark. This dataset provides co-registered, multi-sensor data for individual tree detection across multiple National Ecological Observatory Network (NEON) field sites. Annotations for the image tiles are available in a separate version-controlled repository.

Use Cases

• Training computer vision models for individual tree detection based on RGB imagery.
• Developing multi-sensor fusion algorithms using co-registered RGB, LiDAR, and hyperspectral data.
• Benchmarking geospatial object detection models against a standardized evaluation framework.
• Studying forest ecology and structure using detailed canopy height and spectral reflectance data.

Strengths

• Provides up to four co-registered data modalities per site, including a detailed 426-band hyperspectral file.
• Includes human-annotated ground truth for model training and evaluation, managed via version control.
• Covers multiple distinct geographic sites, each identified by a standardized NEON four-letter code.

Limitations

• Key metadata such as row count, file size, and last update date are not provided by any source.
• Sources conflict on the exact number of files per site, with descriptions listing either three or four data types.
• The dataset's scale and temporal coverage are unspecified.

Provenance

Source

Ben Weinstein, University of Florida, associated with the weecology GitHub organization.

Collection Method

Data appears to be compiled from airborne remote sensing campaigns conducted by the National Ecological Observatory Network (NEON).

Geography

Multiple NEON field sites across the United States, such as Harvard Forest (HARV).