ESA WorldCover 2021

ESA WorldCover 10 m 2021 uses supervised machine learning to transform global Sentinel-1 radar and Sentinel-2 optical observations into a consistent land-cover dataset for Earth-system and environmental research. AI enters after satellite-data preprocessing and feature extraction. Three gradient-boosted decision-tree models (CatBoost) are trained: a combined model using Sentinel-1 and Sentinel-2 features, an optical-only model, and a radar-only model. Their class predictions and probabilities are subsequently combined with documented expert rules to assign one of 11 land-cover classes, including tree cover, cropland, built-up areas, wetlands and permanent water. The 2021 v200 workflow used 319,415 training locations and 115 input features. Its principal output is a global 10 m land-cover map for 2021, accompanied by annual Sentinel composites and documentation. The map serves as a reusable scientific data layer for subsequent environmental, land-use and remote-sensing analyses.

The practice was enabled by an international consortium combining Earth-observation processing, radar and optical remote sensing, land-cover science, machine learning and independent validation expertise. VITO Remote Sensing led the algorithm development and  production with supported by Brockmann Consult and Gamma Remote Sensing, while IIASA and Wageningen University & Research contributed to training data collection and validation, respectively; the work was commissioned by ESA. The workflow depends on global Copernicus Sentinel-1 and Sentinel-2 archives and large-scale computing: production processed approximately two petabytes of satellite data. The v200 algorithm uses 319,415 training locations, 115 derived features and three CatBoost classification scenarios, followed by expert rules that address known ambiguities and poor-quality observations. Independent validation is organisationally separated from production: Wageningen University assessed statistical accuracy and IIASA spatial accuracy. Reuse is supported through a Product User Manual, Product Validation Report, public map viewer, AWS and Google Earth Engine access, and machine-readable map tiles. The 11-class legend is aligned with the UN-FAO Land Cover Classification System, strengthening interoperability with established land-cover terminology.

The practice produced an openly available global land-cover map for 2021 at 10 m resolution, providing a common baseline for environmental monitoring, ecosystem and agricultural research, land-use analysis and downstream modelling. Independent validation reported a global overall accuracy of 76.7%, with continental accuracies above 72%; the validation documentation also identifies classes and regions with lower accuracy so that users can judge fitness for purpose. Compared with the 2020 product, the 2021 v200 workflow improved the training dataset, features and expert rules, but the team explicitly warns that differences between the two annual maps reflect both real land-cover change and algorithm changes and should therefore not be interpreted directly as a change map. Openness and transferability are strong: the product is released under CC BY 4.0 with a persistent Zenodo DOI, Cloud-Optimized GeoTIFF distribution, web services, AWS and Google Earth Engine access, supporting composites and detailed documentation. The workflow is computationally intensive, but its large-scale automated classification makes globally consistent 10 m mapping feasible and produces a reusable data resource for many downstream scientific applications.

1. Dataset: https://zenodo.org/records/7254221
2. WorldCover data access and documentation: https://esa-worldcover.org/en/data-access
3. Product User Manual v2.0: https://esa-worldcover.s3.eu-central-1.amazonaws.com/v200/2021/docs/WorldCover_PUM_V2.0.pdf
4. Product Validation Report v2.0: https://esa-worldcover.s3.eu-central-1.amazonaws.com/v200/2021/docs/WorldCover_PVR_V2.0.pdf