Back to all News
Photo credit: 
Screenshot of the global canopy height layer displayed in a purple-orange color ramp and centered over the southern Amazon region in South America.

A partnership of the computer vision team at ETH Zurich and the Yale BGC Center have developed a brand new, 10-meter global canopy height model. 

This model is the most globally comprehensive and finest resolution canopy height model available to date, and it was created through a novel fusion of NASA’s GEDI lidar data and Sentinel 2 imagery and an end-to-end deep learning model. With the layer and the source code made fully and freely available by the authors, this layer will open numerous doors to new and improved analyses in fields such as carbon biomass estimations, annual tree cover and canopy height trends, and biodiversity modeling. 

The collaboration included BGC Director Walter Jetz from Yale University and Konrad Schindler, Jan Dirk Wegner, and the study’s lead author, Nico Lang from ETH Zurich presenting. As the researchers explain in their new paper, canopy height is a key metric used in a variety of ecological and conservation applications, but it is an exceedingly tricky metric to obtain at a large scale: canopy height measurements are not yet feasible to directly obtain for the entire globe, and the relationships between canopy height and other vegetation traits that are obtained globally is notoriously ambiguous. The NASA GEDI mission, which collected full-waveform lidar data between 2019 and 2023 in the tropical and temperate latitudes, was a crucial achievement towards this goal, but its coverage across the globe was sparse. The challenge, then, was to devise a way to use the limited GEDI data to accurately interpolate canopy height for the entire globe at a resolution useful for ecological and conservation applications – and this was the challenge taken on by Lang et al. 

By using canopy top height data measured from GEDI as a reference data set, the researchers trained an ensemble of deep fully convolutional neural network models to extract patterns from Sentinel-2 imagery that were predictive of canopy height. They then applied this model to a set of dense, global coverage Sentinel-2 imagery from 2019 to 2020 to estimate canopy height along with the associated uncertainty for the entire world at the Sentinel resolution of 10 meters. 

Careful quality control was a crucial part of the making of the new layer. In addition to testing the model on hold-out GEDI data, the researchers also compared the model to two independent airborne lidar datasets in North and Central America and in Europe. They found good agreement among the data and lower random error and bias in their new layer compared to other global canopy height models. Even in high latitude regions not covered by GEDI data (and therefore not seen by the model in the training dataset), they found low error in the canopy height estimations compared to the independent lidar data. 

The new global canopy height model revealed that 41% of the Earth’s land surface is covered with vegetation greater than 5 meters tall, 30% is covered with vegetation greater than 10 meters tall, and just 5% with vegetation greater than 30 meters tall. Additionally, about 34% of canopies 30 meters and taller fall within protected areas. 

According to the authors, “The [new modeling] approach can serve ongoing efforts in forest conservation, and has the potential to foster advances in climate, carbon, and biodiversity modelling.” To help make these advances possible, the authors have made the new layer publicly available. Any interested user can explore the global canopy height layer at the dedicated Google Earth Engine app and view and download the full paper, layer, and code base at the dedicated project page

Check out other articles covering this exciting development in Sciena and NASA Earth Observatory.