Mapping the World’s Trees in Unprecedented Detail with AI
By John Brandt, Justine Spore, Camille Couprie, Jamie Tolan and Tobias Tiecke (Meta)
This blog is part one in a series — stay tuned for part two on the AI foundation model. You can also find out more in our upcoming webinar on July 18.
Forests are vital ecosystems for fighting climate change, supporting livelihoods and protecting biodiversity. Yet critical gaps remain in the scientific understanding of the structure and extent of forests around the globe. While satellite data has made it possible to visualize and analyze timely, globally consistent information about the world’s forests, the majority of existing data has resolutions of 10 or 30 meters, which is not granular enough to see the details of more dispersed forest systems such as agroforestry, drylands forests and alpine forests, which together constitute more than a third of the world's forests.
But now, WRI’s Global Restoration Initiative and researchers from Land & Carbon Lab have partnered with Meta to develop a groundbreaking AI foundation model that we’ve used to produce the world’s first global map of tree canopy height at a 1-meter resolution, allowing the detection of single trees at a global scale.
This new high-resolution data sets a baseline for remotely monitoring changes at the level of individual trees, making it a critical advancement for measuring land use emissions and tracking progress on conservation and restoration projects, which are essential for achieving the world’s goals for climate, nature and people. While this initial data set has limitations, it demonstrates the power of foundation models — a type of AI model that can serve as the “foundation” for a variety of tasks — to pave a new path toward AI-driven earth monitoring.
At Land & Carbon Lab, we believe in powering collaboration and driving solutions for nature by making open data and technology available early and often. WRI, Land & Carbon Lab and Meta have made the new canopy height data and AI model free and publicly available to ensure anyone working on solutions for climate, nature and people have access to the latest innovations.
How was the 1-meter tree canopy height data developed?
The accelerating pace of breakthroughs in AI and foundation models are changing the ways in which we all interact with the world around us. In recent years, mapping forests through remote sensing has made rapid improvements in terms of scale, resolution and refresh rate (how often an area is imaged).
This new 1-meter tree canopy height data set creates a global baseline of where trees are located, including individual trees and forests with open canopies.
To create the maps at this resolution, both a globally robust model and the computational resources to generate 100 trillion pixels of data were needed. To do this, we used a state-of-the-art AI model called DiNOv2 based on methods developed by AI at Meta Research. The model was trained on 18 million satellite images encompassing more than a trillion pixels from across the globe.
We used the powerful approach of Self Supervised Learning (SSL) to obtain a globally consistent high-resolution foundation model — in this case, to predict tree canopy height from satellite images of earth. This approach teaches the AI model to extract general image features from unlabeled satellite images without the need for manually labeling images, which is expensive and time consuming (for example, it would take human annotators about 70 years to manually label 18 million images). The SSL architecture provides a backbone of visual perception that can be used to identify the specific features we are looking for from the images.
To predict tree canopy height from the images, we adapted the DINO approach that was achieving state-of-the-art performance on a classical computer vision task called depth estimation. This task consists of taking a monoscopic image (such as from a smartphone) and estimating the depth to each object in the image. For instance, if you take a photo of the room you are in, a depth estimation model will predict how far away the door to that room is. For mapping tree canopy height, a satellite takes a monoscopic image of the planet’s surface, and the difference between the distance from the satellite to the ground and the satellite to the canopy is the height of the tree.
For this application, the canopy height predictor was trained on top of the SSL architecture using a sample of LIDAR ground truth data from the United States (the NEON data set). The data set analyzes the best available satellite imagery from the years 2009 through 2020. While cloud cover and seasonality impose limitations on the analyzed image dates, 80% of the data is produced with imagery from between 2018 and 2020.
The resulting data set maps trees above 1 meter tall and with a canopy diameter of more than 3 meters. We found that more than one-third of the land on earth (50 million km2) is covered by trees taller than 1 meter, while 35 million km2 is covered by trees taller than 5 meters.
Orchards in Colima, Mexico
Agroforestry in Concepcion, Alajuala Province, Costa Rica
At Land & Carbon Lab, we intentionally release data early and often to help the monitoring community see advances as they happen, test out and build upon new applications, and provide feedback so we can collaboratively improve faster. Land & Carbon Lab and Meta have released both the data and model under licenses permitting commercial use, enabling anyone to build on top of the data to further facilitate accountability and transparency for restoration projects, as well as other conservation and carbon accounting applications. You can find the data on AWS and Google Earth Engine, the model on GitHub, and the publication here. Because the data and model are open, users can improve the output for tree canopy height such as through training the model on additional data sets like country-specific data.
How can the data be used to improve restoration monitoring?
Restoration projects present an increased challenge to monitor tree growth because monitoring young trees, sparse trees (such as in agroforestry) or small project areas (such as in community-led efforts) requires individual tree-scale sensitivity across large areas. The increased granularity as compared to existing tree cover data sets of the 1-meter canopy height map will allow for better monitoring of these complicated systems.
How does the 1-meter canopy height data compare to other tree cover data?
Characteristics like geographic and temporal coverage, resolution, and how they define tree cover all inform which tree cover data sets are best suited for different use cases. The table below compares this new 1-meter tree canopy height data with three other tree cover data sets that are available on Global Forest Watch (GFW). Read more here about these differences and how all three data sets provide valuable information on the extent of forests across the tropics and around the globe.
The 1-meter canopy height map can be used as a starting point to estimate aboveground biomass and to establish baselines for conservation and restoration projects. WRI is a managing partner of TerraFund for AFR100, a partnership that has invested in 192 local nonprofits and enterprises who are restoring Africa’s degraded land. In total, they are restoring more than 29 million trees across tens of thousands of sites. Land & Carbon Lab is using the data — and the foundation DINO AI model — to connect high-resolution satellite data with extensive field-based forest inventory data. This information will allow the initiative’s managers to ensure projects are meeting their targets, identify and provide capacity support to those that are falling behind, and transparently communicate impact.
Additionally, accurate emissions accounting for land use is essential to meeting nature-based climate goals, so it is critical to improve monitoring and verification of forest-based carbon emissions and removals worldwide, particularly by improving the spatial resolution of forest structure data. Improving remotely sensed data with AI can both help mitigate the gap between reported and measured land use emissions for applications like corporate and national inventories and carbon markets, and enable monitoring across international, national, local and corporate scales of conservation and restoration projects. The 1-meter canopy height data has already been integrated into online earth observation platforms such as Restor, Earthblox and the Earth Engine Community Catalog.
What’s next for the data?
This data set and the underlying model represent a significant step forward in the global, high-resolution mapping of tree cover. For the first time, researchers and practitioners can evaluate the utility of 1-meter data, and the model we used to develop this data, for their needs.
However, being the first global high-resolution map comes with its own limitations. The data set is a mosaic of predictions across approximately 580,000 different observations, and some inconsistencies persist at the edges of images. Additionally, some geographies with persistent cloud cover may have incorrect canopy height predictions as cloud-free observations were not always available.
Examples of data set limitations
Left: Areas with persistent cloud cover may have gaps in the predictions due to clouds.
Middle: Some areas have square artifacts that are 150 x 150 meters in size due to the model input image sizing.
Right: Cutlines between image acquisitions in the underlying images sometimes remain in the generated data due to differences in season or image characteristics.
WRI, Land & Carbon Lab and Meta are currently working on the next phase of research to enable counting individual planted trees on restoration sites and tracking the removal of individual trees so users can accurately monitor restoration progress over time. We are improving the ability of the model to distinguish individual tree crowns in closed canopy forest systems, and are developing methods to jointly predict canopy height and stem count to enable the assessment of aboveground woody biomass.
Additionally, the AI foundation model can be used for many diverse earth observation tasks, and we are looking forward to seeing creative applications of the model in the future.