Satlas

Satlas is a project of the Satlas team at AI2.⁶ It takes Sentinel-2 images, the free 10 metre pictures the European Space Agency takes of most land every week, and runs trained models over them to find particular things on the ground.¹ The results are published as maps anyone can download.^1,2

There are three kinds of map. Offshore wind turbines and offshore platforms are points in dated snapshots, tagged by kind.² Onshore wind turbines are points and solar farms are outlines.² Tree cover is a 10 metre raster that grades each pixel in five steps of cover, the lowest meaning no trees.² A combined file for the turbines and platforms gives each one a start and an end month, so a reader can see when something was built or taken down.² The team publishes precision and recall for each model by continent in a validation report.¹

Files come as GeoJSON, KML or shapefile, and tree cover as GeoTIFF tiles, all from a public bucket with no account.² An index file lists every snapshot with its date, so a script can fetch only what has changed.^2,8 On the day this entry was written the offshore series ran to July 2026, while the onshore wind and solar series stopped at January 2024 and tree cover at January 2023.^2,8,9 A model of a coast or an energy system could use the offshore points to count turbines and platforms through time; a city or watershed model could use the tree cover classes as a coarse canopy layer.

The models were first trained on SatlasPretrain.¹ It holds 302 million labels in 137 categories, paired with satellite images from Sentinel-1, Sentinel-2 and Landsat and with US aerial photographs.⁵ It was shown at the ICCV computer vision conference in 2023.^5,7 Models pretrained on it can be downloaded and fine-tuned to look for other things, and the fine-tuning labels and weights for the four Satlas maps are published as well.^1,5

The maps, the new SatlasPretrain labels and the model weights are under ODC-BY.^2,5,7 That open data licence allows any use, commercial included, with attribution.³ The code is under Apache 2.0.⁴ The Hugging Face copy of SatlasPretrain carries an apache-2.0 tag in its metadata, but its own page puts the dataset and the weights under ODC-BY, as the GitHub pages do.^5,7 Imagery and labels taken from other sources, such as OpenStreetMap under ODbL, keep their original licences.⁵