Eligibility questions are historical. Was this land forest at a cut-off date? Was it cleared inside the lookback window? Small plots make a 10 m pixel the natural unit, so people ask for a 10 m forest history. The honest answer is that one does not exist as a single open product. What exists is a short 10 m record, a long 30 m record, and a few ways to join them.
What exists at 10 m
91%
overall accuracy of the JRC 10 m forest map for 2020 (v2 assessment, 21,752 sample units)
2017
first year of the annual satellite embedding layers
1990
first year of the JRC tropical moist forest annual change record, at 30 m
The layers side by side
| Layer | Resolution | Time coverage | What it can answer |
|---|---|---|---|
| JRC global forest cover 2020 | 10 m | One period, 1 January to 31 December 2020 | Forest or not in 2020 |
| JRC tropical moist forest | 30 m | Annual change, 1990 to 2025 | When moist forest was degraded or cleared |
| GLAD-S2 alerts | 10 m | Sentinel-2 era, primary forest only | Events of primary forest loss |
| RADD alerts | 10 m, radar | Africa from January 2019, South America and Southeast Asia from January 2020 | Disturbance events, even under cloud |
| Annual satellite embeddings | About 10 m | 2017 onward | Features for your own classifier |
Note what is missing: a 10 m layer that says, for each year, whether a pixel was forest. The 10 m forest map covers a single year. The long record is 30 m and comes from Landsat.
Why open 10 m history is short
The embedding layers that make a 10 m time series practical begin in 2017. AlphaEarth is a 64-dimension annual embedding, with a hosted copy covering 2017 to 2025, and the TESSERA project notes embeddings for 2017 to 2025. A ten-year lookback from today reaches back before that. Earlier years can only come from Landsat-era products at 30 m, and the sources reviewed for this article document no method that turns the 30 m Hansen loss record into a 10 m one. No source in the set measures a gain from super-resolution, rule-based fusion or Sentinel-2 re-mapping for that purpose.
The alert products do not fill the gap either. GLAD-S2 and RADD are event detectors inside a primary-forest mask. Their own documentation calls them conservative and says they are not meant for area estimates. They tell you where primary forest was being lost after 2019 or 2020. They do not tell you what was forest on a given legal date.
What the 10 m forest map does and does not prove
The JRC 2020 map builds forest presence from global tree and mangrove cover, then applies rules that use tree height, deforestation, cropland and agroforestry layers. According to the JRC accuracy assessment of version 2, 21,752 sample units gave 91% overall accuracy, 18% commission error for forest and 8% omission error. The assessment notes clusters of misclassification in open dry forests. JRC states that the map supports EUDR risk assessment but is not mandatory, not exclusive and not legally binding. The technical report for version 4 is expected in December 2026, so the version 2 numbers are the latest published accuracy.
Agroforestry is inside the exclusion logic, which matters for smallholder land. A preprint on rubber mapping reports that, in the first version of the JRC 2020 map, more than half of the area the authors mapped as rubber was classed as natural forest. The same preprint found low detection power for 12 rubber products in Malaysia and Indonesia. It is a preprint, not peer reviewed, and its findings are preliminary, but it points at a known failure mode: planted tree crops can look like forest to a map.
What embeddings add, and what they do not
An embedding is a fixed-length vector per pixel per year that summarises what the satellites saw. AlphaEarth gives 64 numbers per roughly 10 m pixel. TESSERA gives 128 numbers per 10 m pixel and combines radar with optical data. You train a small classifier on a few labelled points and apply it year by year.
Published accuracy is encouraging but not tropical. In the AlphaEarth paper, direct supervised change detection reached 78.4% and 79.3% balanced accuracy on two land cover and land use change tasks, against 72.0% and 71.5% for the next-best baselines. The reference labels are human-interpreted 30 m plots across the contiguous United States. They say little about Indonesian smallholder mosaics.
Cautions from independent work:
- A secondary summary of a May 2026 study reports that AlphaEarth embeddings were noisy and did not improve restoration-strategy prediction over environmental and spectral features in Brazilian Atlantic Forest restoration. - A coral-reef preprint, from a different domain and not peer reviewed, reports that recovery detected from AlphaEarth trajectories was directionally biased in the gain direction. If that carries over, regrowth or planting gain may be harder to detect than loss. - One embedding summarises a whole year, so it cannot date a clearing within a year. - TESSERA's authors argue the radar channel is essential under near-permanent cloud. That is a design argument, not a measured accuracy.
A defensible way to assemble the evidence
The sources support a layered approach rather than a single layer:
1. Use the 30 m tropical moist forest annual record for multi-year history where it applies. It covers moist forest only, which the JRC portal describes as about 60% of tropical natural forest, citing FAO's 2015 forest assessment. 2. Use the 10 m 2020 map as a baseline cross-check for that year. 3. Use radar and Sentinel-2 alerts for events after 2019 or 2020. 4. If you train an embedding classifier, test gain and loss separately, and validate with a probability sample of reference points that are more accurate than the map. 5. State each layer's known error direction in the eligibility file.
The one thing the sources do not support is calling the result a "10 m Hansen". That would claim a precision no one has shown.
For the differences between the main global forest layers, see which forest dataset do you mean. For how to read validation reports, see how accurate global land cover and biomass maps are.
Frequently Asked Questions
Not in the sources reviewed. The JRC 10 m map covers 2020 only. Annual embeddings start in 2017. Earlier history is 30 m.
Check the land before you commit
Eligibility turns on land history and forest cover at a cut-off date. TREEO Eligibility runs satellite-based forest cover detection on your area of interest — before a single seed goes into the ground.



