On October 22, Ai2 released annual 10-meter embeddings from OlmoEarth v1.3. Its 128-number vectors summarize Sentinel-1, Sentinel-2 and Landsat imagery; they are not land-cover labels.

The big change

  • What changed: Precomputed annual vectors let analysts inspect patterns before classifier training.
  • Why it matters: Explorer accepts labeled points. Maps still need representative labels and ground evidence.
  • What to watch: Vector differences require imagery checks before assigning a cause. Our ForTy coverage examines forest classification.

Evidence and limits

Benchmarks are author-run.

Global embeddings cover 2023–25; earlier coverage is selective. Storage is about 150 TB yearly. Explorer is public; platform access requires an application. The OlmoEarth Artifact License restricts military, surveillance and extractive uses. We did not run it.