- Tree Level
US urban tree canopy height data
Lead Scientists
- ,
- ,

High-resolution tree canopy cover and height data
CTrees generates canopy cover and height maps using advanced deep learning U-Net models trained on airborne imagery and LiDAR data. The models use U-Net architectures optimized for tree height prediction, tree cover segmentation, and individual tree detection. Outputs are derived from the USDA NAIP program images, covering the conterminous United States (CONUS).
Our approach represents the frontier of remote sensing and deep learning modelling. CTrees’ high-resolution, tree-level datasets have applications in emerging use cases such as tree cover monitoring, utility asset management, urban resilience, and insurance.
Based on the availability and quality of additional training data, we also offer bespoke modeling training and validation tailored to specific user requirements for geographies, temporal frequency, and accuracy.
Data at a Glance
- Units
Meters(m)
- Spatial Resolution
0.6 meters (derived from USDA NAIP imagery)
- Temporal range
For anywhere in CONUS, data is available every 2 years from 2020 to the most recent image, aligned with NAIP imagery refresh cycles
- Data types
Unsigned 8-bit integer
- Scale factor
Height (in meters) = pixel value / 2.5
- Offset
0
- No data value
0 or NA
- File format
Cloud-optimized GeoTIFF (COG), tiled and internally compressed
Modeling tree canopies at scale
Input Data
Remote sensing inputs:
- NAIP imagery
- Four-band aerial imagery (Red, Green, Blue, Near-Infrared) at 0.6-meter resolution
- Source: USDA, USGS
- Orthorectified, radiometrically corrected
- Spatial reference: UTM (per tile)
Calibration and validation data:
- Airborne LiDAR-derived Canopy Height Models. Field plots or crown delineation maps (optional).
Validation Approach and Results
Validation of the canopy height results was conducted using independent LiDAR datasets. 90% of independent LiDAR datasets are used for training, and 10% for validation.
- RMSE: ~3.4 m
- MAE: ~2.2 m
- R²: ~0.82
Modeling Approach
Canopy height
A U-Net convolutional neural network was trained to regress tree height values from NAIP imagery. Key parameters of this step are below:
- Input: >1 million 256×256 image patches from NAIP four-band imagery
- Loss function: Mean Squared Error (MSE)
- Data augmentation: Random flips, rotations, brightness variations
- Encoding: Pixel values are scaled by a factor of 2.5 to encode height in 8-bit integers
Canopy cover
Canopy cover can be further developed based on tree height, depending on the availability and quality of training data. In this process, we leverage the canopy height results for post-processing: commission errors are filtered based on a 3-meter tree canopy threshold.
Individual crown segmentation
Individual crown segmentation and identification can be achieved using the U-Net-id model. The approach employs three interconnected prediction paths to delineate crown borders, segments, and inner segments, assigning a unique ID to each inner segment. During post-processing, a buffer is reapplied to restore the original crown size and ensure accurate spatial representation.
Access US canopy height data
Team
Tree-level data for science, policy, and investment
Applications of this tree-level data product include but are not limited to:
- Aboveground biomass estimation
- Tree inventory optimization
- Detection and identification of trees in proximity to utilities or infrastructure
- Habitat modeling
- Change detection
- (Experimental) plantation growth tracking
- Support for individual tree-level and stand-level analysis
Page last updated Nov 16, 2026
All datasets
Stay Connected
Subscribe to our monthly newsletter and be the first to know about new data releases, research, and events.