Satlas

Satlas is an open geospatial data platform from Ai2 (the Allen Institute for AI) that applies deep-learning models to imagery from the European Space Agency's Sentinel-2 satellites to map infrastructure and tree cover worldwide. Sentinel-2 images most of Earth's land every week, but at 10 m per pixel its images are hard to interpret even for people; Satlas uses AI to turn them into ready-made global datasets.

The models are pre-trained on Ai2's SatlasPretrain dataset (over 30 TB of imagery with 302 million labels across 137 categories), fine-tuned on hand-labelled examples, and read the three most recent Sentinel-2 images of each location. Satlas publishes four layers: marine infrastructure (offshore wind turbines and oil, gas, storage and mooring platforms), renewable energy infrastructure (onshore wind turbines as points and solar farms as polygons), tree cover, and Super-Resolution imagery that raises 2023 Sentinel-2 imagery from 10 m to 2.5 m per pixel. Users explore all four layers on the Satlas web map and can download the marine, renewable energy and tree cover layers free under the ODC-BY licence as GeoJSON, KML, shapefile or GeoTIFF (the super-resolved imagery is not available for bulk download). Ai2's own Skylight platform already uses the marine infrastructure data to detect vessels involved in illegal fishing. Training labels, model weights and Apache-2.0 code are also released, and a published validation report gives per-continent precision and recall (for example, 99% precision and 93.2% recall for onshore wind turbines as of October 2022).

Ai2 describes Satlas as an early prototype. Downloads currently run to April 2026 for marine infrastructure, January 2024 for renewable energy infrastructure, and 2016 to 2023 for tree cover.