Ai2 releases OlmoEarth v1.3 global satellite embeddings and browser mapping tool

OlmoEarth v1.3 compresses a year of satellite observations into a numerical fingerprint for each patch of ground, including a 10-by-10-meter patch size at its smallest resolution, with 64-square-meter and 128-meter options, Ai2 said. The fingerprints capture patterns that other software can use to distinguish crops, forests, mines and other types of land cover.

Previously, the main way to use OlmoEarth was to fine-tune it in OlmoEarth Studio, train a model for each task and run it over satellite data to make a map. That required access to large graphics processing unit resources and the ability to download and store imagery. The new OlmoEarth Embeddings Explorer runs in a browser, requires no sign-in and lets users track change over time or find similar places, according to Ai2.

An embedding is a representation of what a model learned from millions of satellite images. OlmoEarth reads a year of radar images and optical data from Sentinel-1, Sentinel-2 and Landsat, missions that provide complementary views of Earth: Sentinel-1 measures radar reflections, while Sentinel-2 and Landsat capture visible and infrared light.

Ai2 said v1.3 is its cheapest OlmoEarth model to train. It used 53% of the computing time of v1.2 and 18% of v1, with 50% and 22% of their carbon emissions, respectively. Ai2 said v1.3 also remains a strong fine-tuning model, making v1.2's performance valuable for research and partner labs.

In Peru's Madre de Dios region, illegal gold mining around La Pampa peaked in 2017 and 2018 and cleared more than 1,200 hectares of forest in the buffer zone of the Tambopata National Reserve, according to the Amazon Conservation Team's MAAP monitoring program cited by SiliconANGLE. Peru's Operation Mercurio shut most of the activity down in February 2019. SiliconANGLE reported that comparing pixels across years with embeddings, without model training, showed a clear eastward push in the southern mining belt and 34 square kilometers of significant change from 2016 to 2020; the same comparison from 2020 to 2024 showed that area fell to 9 square kilometers.

Ai2 said an agricultural analyst could use the same approach to map oil palm cultivation and identify areas for closer examination where plantation expansion may have replaced forest. For many teams, testing an idea against OlmoEarth previously meant a fine-tuning run that took hours; in the explorer, a first pass takes seconds, and analysis of a region from a handful of labels requires no GPU, according to the company. When a task calls for fine-tuning, the embeddings can help decide whether it deserves compute time and where the model falls short.