Ai2 Releases OlmoEarth v1.3 Satellite Embedding Model and Global Annual Embedding Data
Ai2 has released the satellite map embedding model OlmoEarth v1.3 and precomputed annual embedding data for the globe at multiple resolutions, while also launching the browser-based mapping tool OlmoEarth Embeddings Explorer, which can be used without logging in. The model compresses a year of satellite observations into a numerical fingerprint for each 10×10-meter patch of ground, and at the finest resolution it also offers two additional tiers of 64 square meters and 128 meters, which can be used to distinguish land cover such as farmland, forest, and mining areas. Ai2 says v1.3's training compute time is only 53% of v1.2's and 18% of v1's, with carbon emissions at 50% and 22% of the two, respectively.
Why read this
AI score 60
The original text provides comparative data on the new version's training cost and carbon emissions, and explains how the browser tool lowers the barrier to use.
Topics
Body
Ai2, the Allen Institute for Artificial Intelligence, today released a satellite map embedding model called OlmoEarth v1.3 and precomputed annual embeddings for the whole planet at various resolutions with 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 the smallest resolution, with 64-square-meter and 128-meter also available. That fingerprint captures patterns that other software can use to distinguish crops, forests, mines and other types of land cover.
Until now, the primary way to use OlmoEarth was to fine-tune in OlmoEarth Studio, train a model for each task and run it over satellite data to make a map. That requires access to vast graphics processing unit resources and the ability to download and store imagery.
With today’s launch, users can use OlmoEarth Embeddings Explorer, a browser-based map-making tool. That means anyone can turn OlmoEarth into real work, from tracking change over time to finding similar places – no sign-in required.
An embedding represents what a model learned from millions of satellite images. OlmoEarth reads a year of radar images from data, including Sentinel-1, Sentinel-2 and Landsat. Sentinel and Landsat are satellite missions that provide complementary views of Earth: Sentinel-1 measures radar reflections, while Sentinel-2 and Landsat capture light in visible and infrared wavelengths.
The company said that, compared with previous generations of the OlmoEarth model, 1.3 is its cheapest model to train. It took 53% of the computing time of the latest version (v1.2) and 18% of the first (v1), with 50% and 22% of the carbon emissions. It also remains a strong fine-tuning model, making v1.2’s performance valuable for research and partner labs.
Where these models fit into the real world
These satellite models help compress large swaths of data across time and space, which is useful for understanding how the Earth changes.
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. The Amazon Conservation Team’s MAAP monitoring program reports this activity. In February 2019, Peru’s Operation Mercurio shut most of it down. How do we know that happened?
Embeddings give this a clear answer. Comparing pixels within the data across years, no model training required, looking across 2016 to 2020 and there’s a clear eastward push within the southern mining belt: 34 square kilometers of significant change. The same comparison from 2020 to 2024 shows that area falls to 9 square kilometers. The shutdown is visible.
Equivalently, an agricultural analyst could map where oil palm is being grown and identify areas for closer examination where plantation expansion may have replaced forest. These questions could determine where the environmental impact is affecting the world at large.
Ai2 said for many teams testing an idea against OlmoEarth has meant a fine-tuning run, which is an expensive, arduous venture on the order of hours. In the explorer, a first pass takes seconds.
Analysis of a region from a handful of labels requires no GPU. When a task calls for fine-tuning that experiment, the embeddings can help decide whether it deserves the compute time and where the model falls short.
Image: TheDigitalArtist/Pixabay