The entire Sentinel-2 archive. Monthly aggregated. Embedded with Clay v1.5. Free.

Anyone can now access 15.2 billion pre-computed embeddings covering the full globe from January 2017 through April 2026 — no paywalls, no proprietary formats, CC BY 4.0 on Source Cooperative.

Huge thanks to Amazon Web Services (AWS) and Renaissance Philanthropy for supporting this effort.

Why This Matters

Geospatial AI has a compute problem. Training foundation models or generating embeddings at scale requires infrastructure most teams can't access. That bottleneck has kept serious Earth observation work inside a small circle of well-funded institutions.

We processed 2.8 million rasters (2PB of imagery) to analyze 2.49 quadrillion pixels and created over 15 billion embeddings with a peak throughput of 3.5 million embeddings per minute. The 15 billion embeddings were published with at-rest compression ratio of 232:1 (99.5% reduction in size!). 

We're open-sourcing the result so no one else has to.

What's in the Dataset

Every embedding was generated using Clay v1.5 — a 632M-parameter Vision Transformer trained on Earth observation data via self-supervised learning. Clay encodes not just what a chip looks like, but where and when it was captured. That spatiotemporal awareness is what makes these embeddings useful for real-world geospatial tasks.

The monthly-aggregated dataset gives you one embedding per Major TOM grid cell per month — cloud-noise reduced, consistent, and ready for time-series work. Data is stored as GeoParquet 1.1, hive-partitioned by model version, collection, chip size, dims, geohash, year, and month.

Each row includes the 256-dim embedding vector, chip geometry (WKB, EPSG:4326), acquisition datetime, and a stac_item_id you can join back to the source Sentinel-2 scenes via the accompanying rasters catalog.

What You Can Build

With embeddings in hand, the expensive part is already done. From here you can:

• Similarity search — find locations that look like anywhere on Earth, across any time period
• Change detection — diff embeddings over time to surface anomalies
• Fine-tune downstream models — use embeddings as input features instead of raw pixels; dramatically less compute, faster iteration
• Power AI agents — plug directly into RAG pipelines and multimodal workflows that need to reason about Earth's surface

Show us what you build!

For the Community

Clay is Apache-2.0. This dataset is CC BY 4.0. We generated these embeddings as a core part of building LGND — and rather than sit on them, we're giving them back to the researchers, developers, and scientists who can put them to work. 👉 source.coop/clay/lgnd-embeddings