Summary: Meta and the World Resources Institute released Canopy Height Maps v2 on March 10, 2026, replacing the DINOv2 backbone with DINOv3 and raising R² from 0.53 to 0.86. The release pairs the open model with world-scale maps and adds geographically diverse lidar training data. Meta says sparse regions, viewing geometry, and temporal coverage still need more work.
Forest monitoring has a measurement problem before it has a policy problem. You can see a forest from orbit, but seeing height, gaps, and canopy edges at scale is harder — and those are the numbers that feed restoration tracking, degradation detection, and carbon estimates. On March 10, 2026, Meta and the World Resources Institute published Canopy Height Maps v2 (CHMv2), an open source model plus world-scale maps generated from it.
The backbone swap is the whole story
CHMv2 keeps the shape of the 2024 release and replaces the DINOv2 backbone with DINOv3, pre-trained on SAT-493M, a large satellite imagery dataset. According to Meta’s announcement, the model’s R² moved from 0.53 to 0.86, with sharper maps and reduced bias for tall trees.
That jump is the part worth pausing on. Self-supervised pre-training on unlabeled imagery lets the model pick up the visual cues that correlate with height — shadows, texture, crown shape — without millions of hand-labeled examples. If you have ever priced a labeling pipeline for a geospatial task, you know why that matters more than a leaderboard point.
What changed beyond the encoder
Meta says the training data was expanded with more geographically diverse, high-quality lidar examples. To line satellite imagery up with real lidar measurements, the team built automated matching tools and a specialized loss function for canopy height estimation.
That is the unglamorous half of the release. A stronger encoder gets you better features; matched ground truth and a loss that respects the shape of the problem get you predictions you can act on. The R² gain is the headline, but the data and loss work is what makes the number reproducible across landscapes.
Where it is already being used
CHMv1 is in operational use, and the announcement names several public-sector efforts. Forest Research, the research agency of the Forestry Commission, uses the maps for national-scale forest inventory work in Great Britain. The European Commission’s Joint Research Centre used CHMv1 in its Global Forest Cover map for 2020 and hopes to use CHMv2 in future versions and other tree monitoring, including the 3 Billion Tree Initiative. In the US, the maps feed city planning tools tied to Cities for Smart Surfaces, and WRI Ross Center for Sustainable Cities is using them in Cool Cities Lab.
If you are building public-interest products, that list is the useful signal: the consumers here are agencies with inventory obligations, not demo apps. The UN System Data Commons post covers a similar pattern — official data work is where these models either earn trust or do not.
The limits Meta states plainly
The announcement does not claim the problem is solved. Meta says work continues on regions where data is sparse, on viewing-geometry effects, and on extending temporal coverage for change detection over time. Those are the three places a downstream builder should expect to add their own validation before shipping anything operational.
What to take from this
Two practical reads. First, a better pre-trained vision backbone can move a domain metric a long way when the downstream head and data pipeline are already sound — the encoder was the bottleneck, not the task framing. Second, open weights plus published maps mean you can test CHMv2 against your own ground truth instead of trusting the R². Start with the areas you care about most and the sparse-data regions Meta flags; that is where the number will tell you the least.