JRC Tropical Moist Forests (TMF)

A cloud-native republication of the JRC Tropical Moist Forest (TMF) v1.2025 products — the canonical pan-tropical (30 m) forest-state and forest-change record, mapping moist-forest disturbance and recovery from 1990 to 2025.

ℹ️ The pixel values are identical to the JRC source. The only modification is the file packaging: JRC ships tiled GeoTIFFs that lack internal overviews (so they aren't valid COGs) over a click-through download CGI. We re-encode each tile as a valid Cloud-Optimized GeoTIFF (overviews + correct IFD order, ZSTD compression) and host it for anonymous, range-read access. See Provenance & modifications.

Products

6 product families → 41 layers (5 single-image products + AnnualChange, which JRC publishes as 36 separate yearly datasets, 1990–2025). Across ~86 published tropical-land tiles that's ≈ 3,500 COGs.

Tile-first layout — every product for a 10°×10° tile sits in one folder:

TILE_ID is the JRC NW-corner label, e.g. N0_E20, S10_W60, N20_W110.

Class tables are also carried in each per-product STAC index (STAC classification extension) and as the COG's embedded colour table where JRC publishes one. The legends below are reproduced from the JRC TMF v1 Data Users Guide (values + RGB are authoritative).

Class legend & symbology

TransitionMap_MainClasses — uint8

AnnualChange_<YYYY> — uint8, identical legend for every year 1990–2025

UndisturbedDegradedForest — uint8

DeforestationYear / DegradationYear — uint16, discrete years 1982–2025 rendered as a colour ramp (40,146,199)→(80,8,0); 0/NoData = never deforested/degraded.

TransitionMap_Subtypes — uint8, 39 discrete classes (click to expand)

Value	Label	RGB
10	Undisturbed tropical moist forest	0,80,0
11	Bamboo-dominated forest	10,100,10
12	Undisturbed mangrove	10,90,60
21	Degraded forest, short-duration (started before 2016)	30,120,0
22	Degraded forest, short-duration (started 2016–2024)	80,150,0
23	Degraded forest, long-duration (started before 2016)	100,160,40
24	Degraded forest, long-duration (started 2016–2024)	120,170,40
25	Degraded forest, 2/3 periods (last started before 2016)	100,160,40
26	Degraded forest, 2/3 periods (last started 2016–2024)	120,170,40

Scope — what's included

These 6 products are the complete set JRC distributes for download per the v1 Data Users Guide. The TMF dataset also defines several layers that are only available in Google Earth Engine (not via the tile download endpoint), so they are not mirrored here: Deforestation-after-degradation year, Duration, Number of annual disruption observations, Intensity, Areas of deforestation after degradation/regrowth, and the metadata layers StartMonitoringPeriod and ValidObs.

Deferred: 10 m Hybrid Transition Map (beta)

JRC also publishes a Hybrid Transition Map at 10 m for 1990–2022, combining the updated 2022 Landsat (30 m) TMF map with Sentinel-2 disturbance detections for 2022 (sharper degraded-forest delineation and finer linear features such as logging roads), plus a companion map of the detection source per pixel. It is distributed via FTP/GEE.

This product is intentionally not mirrored here because:

• It is a beta. JRC plans an enhanced beta (back-processing the 2016–2021 Sentinel-2 archive) and then a final version delivered with an accuracy assessment — i.e. the values will change.
• For EUDR / compliance use, the basis should be the stable, canonical 30 m products above (notably AnnualChange and DeforestationYear), not a preview that will be revised.
• It is a different resolution (10 m) and temporal extent (1990–2022), so it does not belong in this 30 m v1_2025 layout; it would warrant its own clearly-labelled beta release.

We intend to revisit it once JRC ships the non-beta version.

Access

Public, anonymous read via the source.coop data proxy.

Discover tiles via a product's STAC geoparquet index, then read the COGs:

Or read a tile directly if you know its id:

https://data.source.coop/epoch/jrc-tmf/v1_2025/N0_E20/AnnualChange_2020.tif

GDAL-free access (xarray + icechunk)

For analysis you can skip GDAL entirely. We publish a virtual icechunk datacube beside the COGs at epoch/jrc-tmf/v1_2025/icechunk/ — byte-range references into the same COGs, no pixel data copied. Reads go through zarr + obstore over HTTP range requests (typically faster than the GDAL path, and installs without GDAL).

It opens as an xarray DataTree with one group per published tile. (Per-tile, not one global array: the 10° tiles are 512-blocked but their pixel dimensions aren't 512-multiples and differ tile-to-tile, so they can't be losslessly stitched into a single Zarr array — but every tile is a clean cube, and AnnualChange's 36 years stack into a real (time, lat, lon) axis.)

Discover tiles for an AOI in the STAC parquet (above), open those groups, and — if you want a single stitched array — xr.combine_by_coords them at read time (works fine on decoded data, still GDAL-free):

Provenance & modifications

• Source: JRC TMF v1.2025 (https://forobs.jrc.ec.europa.eu/TMF/data), downloaded per-tile from the JRC download.py distribution CGI.
• Modifications (packaging only): each tile is re-encoded from a non-COG tiled GeoTIFF into a valid COG — internal overviews added (nearest-neighbour, since every product is categorical/discrete), IFD order corrected, ZSTD compression. No resampling, reprojection, or value change. South-up tiles, if any, are flipped to north-up (georeferencing adjusted; values unchanged).
• License: Copernicus / JRC open data. Cite Vancutsem et al. 2021 (https://doi.org/10.1126/sciadv.abe1603); state the source and that the files were re-packaged by Epoch, with no implication of EU endorsement.