From the source
Ai2 released OlmoEarth v1.3, a satellite map embedding model, along with precomputed annual embeddings for the whole planet and a browser-based map-making tool.
- The model compresses a year of satellite observations into a numerical fingerprint for each 10-by-10-meter patch of ground at its smallest resolution, with 64-square-meter and 128-meter resolutions also available, capturing patterns that other software can use to tell crops, forests and mines apart.
- Until now, using OlmoEarth meant fine-tuning in OlmoEarth Studio, training a model per task and running it over satellite data, which required vast GPU resources and the ability to download and store imagery.
- The new OlmoEarth Embeddings Explorer lets anyone make maps in a browser, from tracking change over time to finding similar places, with no sign-in required.
- Ai2 says v1.3 is its cheapest OlmoEarth model to train, taking 53% of v1.2's computing time and 18% of v1's, with 50% and 22% of the carbon emissions, while remaining a strong fine-tuning model.
- The embeddings read a year of radar imagery from Sentinel-1, Sentinel-2 and Landsat; comparing pixels across 2016 to 2020 showed 34 square kilometers of significant change in Peru's southern mining belt, falling to 9 square kilometers from 2020 to 2024.
- Ai2 says a first pass in the explorer takes seconds, and analysis of a region from a handful of labels requires no GPU.