Ai2 releases OlmoEarth v1.3 state-of-the-art AI-ready satellite embeddings for the whole planet
SiliconANGLE Kyt Dotson
Ai2 shipped a new satellite embedding model for the whole planet, plus a browser tool anyone can use. It turns a year of Earth data into quick maps, with no sign-in and no GPU slog.
Based on reporting by SiliconANGLE, Kyt Dotson — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Ai2, the Allen Institute for Artificial Intelligence, has released OlmoEarth v1.3, a new satellite map embedding model, along with precomputed annual embeddings for the entire planet. The package comes with OlmoEarth Embeddings Explorer, a browser-based map-making tool that lets people use the model without signing in.
The core trick is simple to describe and hard to do at scale: OlmoEarth compresses a year of satellite observations into a numerical fingerprint for each patch of ground. At the smallest resolution, that patch is 10 by 10 meters. Larger resolutions of 64 square meters and 128 meters are also available. Those fingerprints can help software tell crops from forests, mines from other land cover, and one kind of change from another.
That matters because the old workflow was heavier. Users had to fine-tune in OlmoEarth Studio, train a model for each task, and then run it over satellite data to build a map. That took serious graphics processing unit resources and usually meant downloading and storing imagery. The new explorer is meant as a first pass that takes seconds, and Ai2 says no GPU is needed when someone is working from a handful of labels.
OlmoEarth reads a year of radar and optical data, including Sentinel-1, Sentinel-2 and Landsat. Sentinel-1 measures radar reflections, while Sentinel-2 and Landsat capture visible and infrared light. Ai2 says v1.3 is its cheapest model to train so far, taking 53% of the computing time of v1.2 and 18% of the first version, with 50% and 22% of the carbon emissions.
The real-world use case is clearer than the marketing gloss. In Peru’s Madre de Dios region, illegal gold mining around La Pampa peaked in 2017 and 2018, clearing more than 1,200 hectares of forest in the buffer zone of the Tambopata National Reserve, according to the Amazon Conservation Team’s MAAP program. Operation Mercurio shut most of it down in February 2019. Using embeddings, Ai2 says the change shows up without any model training: a comparison from 2016 to 2020 shows 34 square kilometers of significant change in the southern mining belt, while 2020 to 2024 drops to 9 square kilometers. The shutdown is visible. And so is the larger point: if a model can make that kind of change cheap to inspect, a lot more people can use satellite data without first buying a small mountain of compute.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of AI release: less fireworks, more useful plumbing. The industry loves to brag about bigger models; the planet usually benefits more from cheaper ones that people can actually run. Open access here beats another locked lab toy, which is refreshing for about five minutes until everyone starts asking for the next trillion-parameter miracle.
Read more about this at: SiliconANGLE