blog/free-building-blocks-forest-carbon-monitoring-licences-hero
← Back to blog

Free Building Blocks for Forest Carbon Monitoring: What Is Open, Under Which Licence, and Where the Gaps Are

Technology•13 November 2026•TREEO Indonesia•3 min read

Almost every step in forest carbon monitoring now has a free starting point: canopy height, embeddings, biomass equations, change detection, validation. Free does not mean the same thing across them. Some maps are open but their model weights are gated. Some code is permissive, some is copyleft, and some is archived. This guide maps the main options by need, with the licences as verified on 5 November 2026. Licences change, so re-check before you build on one.

Canopy height

ProductWhat it isLicence as verified
Canopy height v2 (Meta and WRI)Global canopy height at about 1 m from very high-resolution imageryMaps CC-BY 4.0; code and weights under the DINOv3 licence, gated
Canopy height v1 (Tolan et al.)First 1 m global mapApache-2.0, code and weights
Global canopy height (Lang et al.)10 m canopy top height for 2020 from Sentinel-2, trained on GEDIMIT, code

Reported accuracy differs by product and test. For the Lang map, the authors report an RMSE of 5.3 m and a bias of 0.5 m against independent airborne lidar. A later comparison reports that it overestimates low vegetation. For version 2, the paper reports a mean absolute error of 1.38 m and R² of 0.94 on a United States test set (an ablation test in the paper). The imagery behind a 1 m map has a date that varies by location, so the map is a single-date layer, not a time series.

Biomass from trees and from space

- BIOMASS (R package, GPL-2) converts tree measurements into plot biomass with error propagation, including wood density and height-diameter models. It builds in the Chave 2014 pantropical equation. Coffee and other shrub equations are not part of it. - GEDI L4A and L4B are NASA open data products. L4A holds footprint-level biomass and L4B a 1 km grid. Footprints are sparse and date from 2019. - ESA CCI Biomass is a free download with per-pixel uncertainty at 100 m. For the 2010 map, the validation paper found that patterns were well captured except above 250 Mg/ha. The licence text was not confirmed for this article, so read the terms before reuse.

Our post on what the benchmarks show about biomass from space explains why these products should be treated as predictors, not as field truth.

Embeddings and foundation models

ProductLicence as verified
Satellite embeddings (AlphaEarth), 64 numbers per 10 m pixel per yearData CC-BY 4.0; model weights not released
TESSERA, Sentinel-1 and 2 time-series embeddingsMIT, with open weights
Presto and Galileo (NASA Harvest)MIT
ClayApache-2.0, code and weights
TerraTorch fine-tuning toolkitApache-2.0

An independent benchmark (PANGAEA) found that geospatial foundation models do not consistently outperform a plain supervised baseline when labels are limited. The practical lesson is to compare "embeddings plus a simple model" with "indices plus a random forest" on your own labelled points before committing.

Commodity maps and change detection

- Forest Data Partnership models give 10 m probabilities for coffee, cocoa, palm and rubber, built on AlphaEarth; the code is MIT-licensed. On the project's own test, release 2025b scored 0.92 for palm, 0.96 for rubber, 0.94 for cocoa and 0.70 for coffee. The notes say coffee is overestimated on volcanic slopes and in tea-growing areas. - WorldCereal (MIT) is a reproducible crop-mapping system that you can retrain for your own classes. - Dynamic World's code repository is archived; the data remain in Earth Engine. - pyxccd (Apache-2.0) detects breaks in Landsat, HLS or Sentinel-2 time series in Python and is actively maintained. gee-ccdc-tools (MIT) runs the same family of methods in Earth Engine, but its last push was in 2023. - JRC tropical moist forest is free without restriction under the Copernicus regulation, at 30 m from 1990. The algorithms behind it, Hansen's forest change and RADD alerts are not released as code, although the data are open.

Validation tools matter as much as models

A biomass model that explains half the variance under random cross-validation can show almost no skill under spatial cross-validation. That is the finding of Ploton and colleagues (2020), and it is why spatial validation tools belong in the stack:

- blockCV (GPL-3) builds spatial or environmental folds. - spatialsample (MIT) does the same for tidymodels. - CAST (GPL 2 or later) estimates the area of applicability, flagging pixels unlike any training plot. - MAPIE (BSD-3) and quantile-forest (Apache-2.0) give prediction intervals. - GPBoost (Apache-2.0) combines boosting with random effects, useful when plots cluster by farm or village.

Where nothing open exists yet

The sources reviewed found no open options for:

- separating coffee monoculture from coffee agroforestry in Indonesia or Uganda; - splitting biomass into coffee-shrub and shade-tree pools, or coffee-shrub equations inside BIOMASS; - plot-to-pixel biomass upscaling code built for small numbers of field plots; - released code for the JRC, Hansen and RADD methods; - a 10 m forest history before 2016.

Frequently Asked Questions

The version 2 maps are CC-BY 4.0, which requires attribution. The model weights are gated under a separate licence. Read both before building on them.

Move beyond estimates

Verifiers test sampling design, uncertainty and whether a number traces back to the field. TREEO dMRV captures that evidence in real time, in one auditable chain.

Related Articles