How Meta is uniting with research org WRI to map the world’s forests

A new collaboration between a technology company and a global research nonprofit offers nongovernmental organizations and researchers free, high-resolution data that maps previously inaccessible forest areas. Announced amid a growing urgency for better monitoring capabilities, could it help reach the 2030 target of zero forest loss? The canopy height map — developed by Meta and the World Resources Institute’s Global Restoration Initiative and Land & Carbon Lab — aims to set a new baseline for tree monitoring, helping stakeholders to make better-informed decisions about land use. With only six years left for the leaders of 145 countries to meet the 2030 target of reaching zero forest loss, the 4% decrease achieved in 2023 will not do. Accurate and detailed forest mapping is key, but current satellite data is at 10- or 30-meter resolution — which fails to capture detailed images of over a third of global forest cover. This map has a 1-meter ground pixel resolution, meaning that each pixel in the satellite image represents an area of 1x1 meter on the Earth's surface. This level of resolution, if not obstructed, allows for the detection of single trees at a global scale, enabling detailed observation and analysis of tree coverage. “For the first time, we have the tools to closely monitor restoration projects at a granular level, enabling better-informed decisions and ultimately driving greater success in global restoration efforts,” John Brandt, senior data science associate at WRI’s Land & Carbon Lab, told Devex. How it works One of the main problems facing remote monitoring of canopy height, Brandt explained, is the limited availability of Light Detection and Ranging, or LiDAR, sensors. These use laser light to measure distances and create detailed, high-resolution maps of physical environments — but they are expensive. The idea for the map started with Meta's foundational artificial intelligence model DINOv2, which was originally designed to estimate depth in 2D images. Foundational AI models are advanced machine learning models that serve as a basis for various AI applications. Meta’s research team realized that mapping canopy height is quite similar to what DinoV2 was doing. A satellite image can be used to measure the height of trees by comparing the distance from the satellite to the ground and to the top of the canopy, Brandt said. To do so, Brandt explained that they trained DinoV2 using 18 million satellite images to recognize image patterns across the entire planet and refined it with 5,600 LiDAR measurements of canopy height. Once trained, the model analyzed over 100 trillion pixels of land area, providing estimates of canopy height for each 1-meter pixel, Brandt told Devex. Micro-level: Monitoring forest cover changes The map is valuable to remotely tracking restoration efforts, Tamara Coger, senior associate at WRI’s Global Restoration Initiative, explained. This includes monitoring the impact of TerraFund — a coalition co-managed by WRI that supports local African organizations in restoring degraded land through grants and loans. “The WRI/Meta tree canopy data is definitely a game-changer … particularly in the developing countries and in the tropics where cloud is a major challenge with optical data and lidar data is either too expensive or beyond reach,” said Justice Camillus Mensah, a project manager responsible for remote sensing/GIS and MEL at Hen Mpoano — a two-time TerraFund AFR100 grant recipient focusing on restoring mangroves in southwestern Ghana. Optical data refers to information encoded using light. Mensah said that the canopy height maps will provide baseline data for research on forest dynamics, species interactions, and habitat requirements. High-resolution data also allows for the precise identification of forest boundaries and small, fragmented, or dispersed forest patches that lower-resolution data might miss. This enables more targeted interventions to protect vulnerable areas, explained Basma Albanna, a lecturer at Ain Shams University and co-founder of Data Powered Positive Deviance. Albanna noted that this could unlock significantly more climate and biodiversity finance for critical but previously hard-to-monitor areas. Macro-level: Identifying positive outliers in forest restoration The map covers the entire planet and is particularly valuable for areas lacking access to LiDAR data. If combined with a data powered positive deviance, or DPPD, approach, it could help design more targeted and effective interventions for forest conservation and restoration, Albanna explained. DPPD uses traditional and nontraditional data to identify outliers — or positive deviants. These are individuals or communities achieving better outcomes despite facing similar challenges, Albanna explained, emphasizing how “by studying these positive outliers, you can uncover innovative solutions that can then be amplified and scaled.” For forest conservation, high-resolution data can help DPPD identify successful small-scale, community-led projects that are performing well, smallholders who are deforesting less, and protected areas with improved forest cover. Understanding what drives these successes can then guide the scaling of effective practices, tailoring strategies to local contexts, and allocating resources more efficiently, Albanna added. Potential shortcomings Iain Woodhouse, head of knowledge and outreach lead at Earth Blox, has identified some potential issues with the map’s processing, which he laid out in part in the analysis for Earth Blox. For one, Woodhouse told Devex that cloud cover presents a challenge for all-optical satellite imagery, including the WRI-Meta one. Clouds obstruct the satellite's view of the ground, making it difficult to get clear images. More vegetation often means more clouds due to rain, making it difficult to collect images in wetter regions such as tropical rainforests, he said. Satellites are bound by their orbits — they can't choose to collect images on sunny days, Woodhouse explained. Therefore, the data is more accurate in less cloudy regions. In dense forests, such as the Amazon, even advanced data products struggle to provide accurate measurements due to the forest density and cloud cover, he explained. This is why analyzing tree coverage in different regions such as Borneo and Malawi comes with different levels of success. Borneo's data, collected over many dates due to frequent clouds, is patchier, he told Devex. While Malawi, with fewer clouds, provides more consistent data. This is not just applicable to the WRI-Meta data, but is an issue that affects all optical satellites, he explained. Users also need to know the exact dates of the data to assess forest changes accurately, Woodhouse explained. While the WRI-Meta map now includes a date layer, some newer data products also include time series, which allows to track changes over time — this is important for monitoring and historical analysis, he said. A successful case of this is Planet’s NICFI program — which provides a monthly mosaic of tropics areas to help deforestation, he said. “Hopefully, the integration of other datasets will be made possible. It would also be great if canopy data from other periods could be added to estimate changes in the tree canopy,” Hen Mpoano’s Mensah added. Another issue comes from the map being divided into smaller sections, or satellite image tiles. Each tile is processed separately and satellite images of different tiles could be taken at different times, Woodhouse added. He provided an exaggerated example to explain how one tile might use data from January, March, and May, while an adjacent tile uses data from February, April, and June. This may create visible differences between where the tiles join. Such tiling issues are also consistent across different data sets, Woodhouse said. However, it can affect projects located at tile boundaries, he wrote in the Earth Blox analysis. For specific projects, Woodhouse emphasized that it’s crucial to calibrate global data sets with local data. This can be done using field measurements or higher resolution data to improve accuracy for the specific region, he said. For example, “combining the canopy data with drone images and field-based data will contribute to climate change models and carbon accounting,” Mensah told Devex, explaining how the integration will help to measure forest biomass and the amount of carbon it can capture. Lowering barriers with an open-access approach Just a few years ago, large organizations such as WRI could develop AI-derived products on their own. However, as fundamental AI models continue to scale, they now demand significantly more computing power than before, Brandt explained. “The bottleneck to this process was often validation, stakeholder engagement, and map generation,” he said, while now he explains, the bottleneck has shifted toward computational scale and AI expertise. This is why the data and underlying foundational model have been made free and publicly available. However, the first order might be better quality control. Woodhouse believes that data needs to be flagged as unreliable rather than simply assigning it a potentially misleading low value. The product offers 1-meter resolution, which means it contains a lot of pixels, but many of these may be unreliable, he explained. While there are limitations when it comes to tree coverage, Meta’s DiNOV2 foundational AI model aims to be adaptable for diverse geospatial applications. Brandt added that the model is already being used for tree counting, boundary identification, and disaster response mapping. Unfortunately, scaling isn’t always as straightforward as it sounds. While the open-source approach has its strengths, effectively coordinating widespread adoption and application of these tools may require additional capacity building and stakeholder engagement, Albanna highlighted. The bottom line is that the data is valuable to some reforestation projects because it's free, but its reliability varies – and is more suitable for projects in regions with distinct forest areas, such as the dry tropics, but is not that useful in densely forested regions with a lot of cloud coverage, Woodhouse said.