Meta AI has unveiled Canopy Height Maps v2 (CHMv2), a significant upgrade in high-resolution global forest monitoring developed alongside the World Resources Institute. By swapping the previous backbone for DINOv3—a self-supervised vision model—the team has achieved a dramatic leap in detail. This version allows researchers to see the Earth’s green lungs with such clarity that they can distinguish individual tree crowns, gaps, and edges across the entire planet.

The technical progress is marked by an R² value—a statistical measure of how well the model's predictions align with real-world data—increasing from 0.53 to 0.86. This was made possible by training DINOv3 on SAT-493M, a massive dataset of 493 million satellite images. The model learns visual features like shadows and textures without needing millions of human-labeled examples, a process known as self-supervised learning. This makes the system far more scalable and trustworthy for tracking carbon storage and forest health.

Practical applications are already widespread, from cooling urban centers in the U.S. to supporting the EU’s initiative to plant 3 billion trees by 2030. In the United Kingdom, Forest Research is utilizing these maps to manage national forest inventories more effectively. By making the models and maps open source, Meta aims to provide governments and conservationists with the granular data needed to make informed land-management decisions and combat climate change.

Meta AI has unveiled Canopy Height Maps v2 (CHMv2), a significant upgrade in high-resolution global forest monitoring developed alongside the World Resources Institute. By swapping the previous backbone for DINOv3—a self-supervised vision model—the team has achieved a dramatic leap in detail. This version allows researchers to see the Earth’s green lungs with such clarity that they can distinguish individual tree crowns, gaps, and edges across the entire planet.

The technical progress is marked by an R² value—a statistical measure of how well the model's predictions align with real-world data—increasing from 0.53 to 0.86. This was made possible by training DINOv3 on SAT-493M, a massive dataset of 493 million satellite images. The model learns visual features like shadows and textures without needing millions of human-labeled examples, a process known as self-supervised learning. This makes the system far more scalable and trustworthy for tracking carbon storage and forest health.

Practical applications are already widespread, from cooling urban centers in the U.S. to supporting the EU’s initiative to plant 3 billion trees by 2030. In the United Kingdom, Forest Research is utilizing these maps to manage national forest inventories more effectively. By making the models and maps open source, Meta aims to provide governments and conservationists with the granular data needed to make informed land-management decisions and combat climate change.