Global maps are easy to download and hard to judge. A layer can be sharp, recent and free, and still be wrong for your plot. The honest way to judge it is to read the validation report, if one exists, and see what was actually measured.
This post collects the published figures from a handful of widely used products. The pattern is consistent: accuracy depends on the class, the place and the question.
Published accuracy figures
76.7%
global overall accuracy of ESA WorldCover 2021 (v200)
50%
average standard error of CCI Biomass in tropical rainforest at 100 m
1 km
resolution of the GEDI L4B gridded biomass product
ESA WorldCover: good overall, uneven by class
ESA WorldCover 2021 (version 200) is a 10 m global land cover map with 11 classes, built from Sentinel-1 and Sentinel-2 imagery. Its validation report gives an overall accuracy of 76.7 ± 0.5%, based on 2.16 million sample units at 21,624 locations. The 2020 version (v100) scored 74.4%.
Some details in the report matter more than the headline:
- Continental accuracy is above 72% everywhere, with the highest value for Asia at 82.1%. - Tree cover, cropland, permanent water, snow and ice, and bare or sparse vegetation have high class accuracies. - Shrubs, herbaceous wetland, and moss and lichen are noticeably lower. - At global scale, the map slightly overestimates trees. - The 2020 and 2021 maps use different algorithm versions, so differences between them mix real change with algorithm change.
That last point is the one that catches people. Subtracting one WorldCover year from another does not give you a clean change map.
CCI Biomass: a target, and an error that depends on the scale
The ESA Climate Change Initiative (CCI) Biomass project maps above-ground biomass (AGB) at 100 m. Its stated target is a relative error below 20% where AGB is above 50 Mg per hectare. It publishes annual maps for 2005 to 2012 and 2015 to 2024, with a per-pixel standard deviation layer.
The project's own uncertainty budget is more cautious. For tropical rainforest, the average standard error is about 50% at 100 m resolution and about 35% at 1,000 m. Averaging over larger areas reduces the error, because random errors partly cancel. The maps are not masked to forest, and users are warned to validate change and not to rely on a single pixel.
AGB here is defined as oven-dry mass of all plant parts above ground, excluding stumps and roots, so it is not a carbon stock until you apply a carbon fraction and add the other pools.
GEDI L4B: sampled, not wall-to-wall
NASA's GEDI lidar samples the surface in footprints of about 25 m, spaced roughly 60 m along the track and about 600 m across it. The L4B product (version 2) turns those samples into 1 km mean biomass density for 19 April 2019 to 4 August 2021, with standard errors that reflect both the sampling and the model behind each footprint estimate.
A 1 km cell is useful for regional context. It is not a plot-level answer.
Smaller targets: oil palm and cocoa
Crop and plantation mapping shows how much the answer depends on the target and its age.
- A 30 m map of oil palm in Indonesia, Malaysia and Thailand for 2017 uses Sentinel-1 radar and a Landsat time series. It records the year of detection, and its authors note that young plantations only become detectable after a few years. - In Riau, a Sentinel-1 only map of mature oil palm without separating types reached 93.5% overall accuracy. - In Ghana, a random forest combining Sentinel-1 and Sentinel-2 beat either satellite alone, with 90.3% overall accuracy for industrial oil palm. - In a Thai study, an unmanned aerial vehicle (UAV) classifier beat Sentinel-2 for healthy oil palm, at 76.77% overall accuracy, and for mixed areas, at 67.22%. - For cocoa in West Africa, an evaluation of 14 datasets found that the official maps of Ghana and Côte d'Ivoire were the most accurate.
Each figure comes from a different place, class and method. They are not comparable with each other, and none is a general promise for your site.
What is still missing
The research behind this post looked for per-biome validation of several popular layers, including Hansen Global Forest Change, JRC Tropical Moist Forests and Dynamic World, and found none in the sources it could read. Absence of a figure in a quick search does not mean no study exists. It does mean you should not quote an accuracy number for those layers without opening the paper behind it.
A short way to read any map
1. Find the validation report, not the product page. 2. Check the class you need, not the overall figure. Overall accuracy can hide weak classes. 3. Check the scale. An error that is acceptable over a district can be unacceptable over a plot. 4. Check the year and the algorithm version before computing change. 5. Keep a field sample to test the map where you work.
Our post on validating satellite biomass with field plots describes the field side of this, and open satellite data sources compared covers the inputs.
Frequently Asked Questions
The 2021 version (v200) has a global overall accuracy of 76.7 ± 0.5%, according to its validation report. The 2020 version had 74.4%. Accuracy differs by class and continent.
Choosing a map for a compliance check?
A headline accuracy figure is only the start. If you are deciding which layers to trust for a plot-level check, we are happy to talk through the trade-offs.


