This dataset is part of a Publisher Catalog, and not managed by Google Earth Engine. Contact biosphere-models@google.com for bugs or view more datasets from the Nature Trace Catalog. Learn more about Publisher datasets.
- Catalog Owner
- Nature Trace
- Dataset Availability
- 2020-01-01T00:00:00Z–2020-12-31T23:59:59Z
- Dataset Producer
- Tags
Description
The Forest Typology (ForTy) v1 dataset consists of a global per-class probability map at 10 m resolution covering all land areas between 65°S and 84°N latitude for the year 2020.
The six-class typology is aligned with FAO and EU Deforestation Regulation (EUDR) definitions:
- Class 1 – Primary Forest: Natural forest of native species with no visible indications of human-caused disturbance.
- Class 2 – Naturally Regenerating Forest: Forests predominantly established through natural regeneration, including secondary forests recovering from disturbance and mixed forests where natural regeneration dominates.
- Class 3 – Planted Forest: Forests established predominantly through planting or seeding, with an operational rotation period exceeding 40 years.
- Class 4 – Plantation Forest: Intensively managed forests characterized by one or two species, even age, regular spacing, and a rotation period of 40 years or less.
- Class 5 – Tree Crops and Agroforestry: Agricultural plantations producing commodities other than wood (e.g., oil palm, rubber, fruit orchards), including agroforestry systems. Classified as agricultural land use under EUDR.
- Class 6 – Other Land: All non-tree land uses, sparse woody vegetation (canopy cover < 10%), other wooded land, trees in urban environments, and small tree patches that do not meet forest thresholds.
The primary data record is a five-band raster containing unsigned 8-bit integer values (0–250) representing quantized probability values for each of five forest type classes. The probability for the sixth class (Other land) is obtained as the complement: p_other = 250 − sum(p_i). The map uses the Universal Transverse Mercator (UTM) coordinate system.
To get the probability score of the sixth, "Other land" class, you to subtract from 250 the sum of the first five probabilities. A categorical map can be derived by assigning each pixel to the class with the highest probability (argmax). The probability representation enables users to apply custom thresholds, assess prediction confidence, and compute uncertainty estimates for downstream applications.
The source data is provided as Cloud Optimized GeoTIFFs (COGs) stored
on Google Cloud Storage at
gs://nature-trace/export_rasters/forest_typology_2020_v1_0/. The
Earth Engine ImageCollection is backed by these COGs.
Limitations: While this map provides a valuable global baseline, users should be aware of several limitations:
- Temporal reference date: The map represents the state of land cover as of the year 2020, based on multi-temporal satellite composites. Rapid land-use changes occurring within the compositing period may not be fully captured.
- Definitional ambiguity: The FAO-based definitions underlying the EUDR classification rely on management intent and history, which are not directly observable from satellite imagery. The model infers these categories from proxy indicators (spectral signatures, canopy structure, spatial patterns), introducing irreducible uncertainty at class boundaries. For example, mature naturally regenerating (secondary) forests can be spectrally and structurally indistinguishable from undisturbed primary forests.
- Agroforestry ambiguity: Mixed agroforestry systems—where tree crops such as cocoa or coffee grow beneath a native canopy—are among the most challenging classes to resolve. Shaded cocoa and coffee agroforests can still be misclassified as natural forest in some regions.
- Planted forest confusion: Planted forests exhibit the lowest classification accuracy among all forest classes (F1 = 58.2%). This class can be spectrally and structurally similar to naturally regenerating forests at one end and to plantation forests at the other, resulting in frequent misclassification in both directions.
- Small patches and threshold canopy coverage: Isolated tree patches near the FAO forest area threshold (0.5 ha) and areas at the canopy cover boundary (10%) are inherently ambiguous.
- Probability calibration: The per-class probabilities represent ensemble-averaged softmax outputs and have not been post-hoc calibrated. Users requiring well-calibrated probabilities for quantitative risk assessment should consider regional recalibration using independent reference data.
- Wildfires: It is not always possible to determine the land use after a wildfire.
- Conservative primary forest mapping: The primary forest class boundary is conservative near human infrastructure. Due to a spatial buffering effect inherited from the training data, the model systematically excludes forest edges near roads, settlements, or agriculture from the primary forest class, often labeling them as naturally regenerating instead.
- Temporarily unstocked areas: Under FAO definitions, temporarily unstocked areas (such as recently harvested plantations awaiting replanting or forests in early post-disturbance recovery) are still considered forest. However, because this map relies on a single year of observations (2020), these areas may appear as non-forest and be misclassified as "Other Land."
- Planted vs. plantation proxy: The distinction between planted and plantation forests uses an operational 40-year rotation threshold as a proxy. True EUDR/FAO management intensity criteria (e.g., even-aged class, regular spacing) are not directly observable from satellite imagery.
- Other Wooded Land (OWL) omission: The EUDR recognizes "Other Wooded Land" (OWL) as a distinct category, but in this dataset, OWL is not resolved as a separate class and is entirely subsumed into the "Other Land" category.
- Sub-pixel heterogeneity: Mixed pixels, particularly in complex smallholder landscapes, can reduce classification precision. Additionally, while the output is at 10 m, the effective resolution of the model's input data might be coarser than 10 m.
Bands
Bands
Pixel size: 10 meters (all bands)
| Name | Min | Max | Scale | Pixel Size | Description |
|---|---|---|---|---|---|
PrimaryForest |
0 | 250 | 0.004 | 10 meters | Quantized probability (unsigned 8-bit integer, 0–250) that the pixel belongs to the primary forest class. |
NaturallyRegeneratingForest |
0 | 250 | 0.004 | 10 meters | Quantized probability (unsigned 8-bit integer, 0–250) that the pixel belongs to the naturally regenerating forest class. |
PlantedForest |
0 | 250 | 0.004 | 10 meters | Quantized probability (unsigned 8-bit integer, 0–250) that the pixel belongs to the planted forest class. |
PlantationForest |
0 | 250 | 0.004 | 10 meters | Quantized probability (unsigned 8-bit integer, 0–250) that the pixel belongs to the plantation forest class. |
TreeCropsAndAgroforestry |
0 | 250 | 0.004 | 10 meters | Quantized probability (unsigned 8-bit integer, 0–250) that the pixel belongs to the tree crops and agroforestry class. The probability for the sixth class (Other land) is obtained as the complement: p_other = 250 − sum of all five band values. |
Terms of Use
Terms of Use
This dataset is licensed under CC-BY 4.0 and requires the following attribution: [PLACEHOLDER ATTRIBUTION].
Citations
Maxim Neumann, Anton Raichuk, Peter Potapov, Myroslava Lesiv, Matthew Overlan, Melanie Rey, Ravindran Rajakumar, Michelangelo Conserva, Radost Stanimirova, Michelle Sims, Sarah Carter, Elizabeth Goldman, Yuchang Jiang, Linus Scheibenreif, Ivelina Georgieva, Maria Shchepashchenko, Steffen Fritz, Nicholas Clinton, Charlotte Stanton, Dan Morris, Drew Purves: "Global forest typology at 10-meter resolution for forest and land-use monitoring", May 2026 (preprint: doi.org/10.31223/X58R27).
Explore with Earth Engine
Code Editor (JavaScript)
var dataset = ee.ImageCollection( 'projects/nature-trace/assets/forest_typology/forest_typology_2020_v1_0_collection') .mosaic(); // Compute argmax class: 1=Primary, 2=Naturally Regenerating, // 3=Planted, 4=Plantation, 5=Tree Crops & Agroforestry, 6=Other land. var b5 = ee.Image(250).subtract(dataset.select([0, 1, 2, 3, 4]).reduce('sum')); var classified = dataset.addBands(b5).toArray().arrayArgmax().arrayGet([0]).add(1); // Palette: Primary=dark green, Naturally Regenerating=light green, // Planted=blue, Plantation=pink, TreeCrops=orange, Other=yellow. var colors = ['1B7837', '7FBF7B', '1D91C0', 'E65FA9', 'E6AB02', 'FDE278']; Map.addLayer( classified, {min: 1, max: 6, palette: colors}, 'Forest Typology 2020 (v1)'); Map.setOptions('HYBRID'); Map.setCenter(116.21, -33.31, 12);