The Pulse
Ai2 Releases 10-Meter Satellite Embeddings for Earth
Ai2 released OlmoEarth v1.3 with annual satellite embeddings at 10-meter resolution and a browser-based explorer. The archive offers global coverage for 2023–2025, with selected-country data for earlier years.
AI.info Team · Written with AI · fact-checked ·
About 150 terabytes of satellite-derived data for each year: that is the size of the archive behind Ai2’s new global Earth-observation release. On October 22, the Allen Institute for AI published OlmoEarth v1.3, a model and precomputed annual embeddings that represent the planet at 10-meter resolution. Each pixel gets a compact numerical description drawn from a year of satellite observations, rather than a ready-made label such as “cropland” or “burn scar.”
Ai2 also released a browser-based explorer for comparing years, finding similar places and training small classifiers from labeled points. The data and tools give researchers and mapping teams a way to start with reusable features instead of first running a large model across satellite imagery themselves.
One vector for each 10-meter pixel
OlmoEarth v1.3 processes annual time series from three satellite sources: radar imagery from Sentinel-1, multispectral imagery from Sentinel-2, and Landsat data. It produces a 128-dimensional vector for each 10-meter pixel; users can also take the first 64 dimensions as a smaller representation. Ai2 says the model’s vision transformer processes the imagery, an attention pooler combines information across time and sensors, and a linear layer compresses the result into the vector.
The technical report describes the release in a sentence: “We use OlmoEarth v1.3 to compute global, annualized 10m embeddings, which we release under the OlmoEarth artifact license.” The team presents the embeddings as reusable features: a crop map, land-cover map or change analysis can be built by training a smaller classifier on top.
We use OlmoEarth v1.3 to compute global, annualized 10m embeddings, which we release under the OlmoEarth artifact license.
Team OlmoEarth, Allen Institute for AI
Ai2 reports wins against two other embedding sets
Ai2 compared OlmoEarth v1.3 with Google’s AlphaEarth Foundations and Tessera v2 across 27 evaluations spanning six continents and multiple tasks. The institute says the 128-dimensional version scored highest on 20 evaluations, while the 64-dimensional version led on 18. Those are Ai2’s reported benchmark results; they do not mean every map made from the embeddings will outperform work based on the alternatives.
The release also includes PLANTEUR, a benchmark for crop types in tropical farming. Ai2 says OlmoEarth generalizes particularly well to places a classifier has not seen during training. The team’s examples include mapping smallholder crops in Kenya, detecting brick kilns near Dhaka, assessing fire damage in Indonesia and tracking gold mining in Peru.
The explorer turns vectors into a first map
Users can pan around a false-color view, switch between years and compare the embeddings with underlying Sentinel-2 imagery. Clicking a location can surface similar places; users can also add a small number of labeled points, train a classifier in the browser and export its output as a GeoTIFF. Ai2 says an initial analysis in the explorer takes seconds and does not require an account or a GPU.
The browser tool sits alongside a much larger downloadable archive on Source Cooperative. The archive covers land and adjacent coastal waters, with 2023, 2024 and 2025 available globally at launch. Data from 2017 through 2022 covers selected countries, not the whole planet, and each 20-kilometer tile has provenance records listing the satellite scenes used.
Annual coverage sets the limits
These embeddings are not continuous monitoring or a finished environmental map. They summarize satellite inputs into annual vectors, and a user still needs labels and a classifier to produce many specific maps. Comparing annual vectors can help identify places that changed, but the archive’s yearly cadence does not itself provide a map of events as they unfold.
Ai2 says v1.3 took 53% of the computing time required to train v1.2 and 18% of the time required for the original v1; its reported carbon emissions were 50% and 22% of those earlier versions, respectively. The institute estimates that reproducing the final Base-model training run would cost about $2,720 in rented H100 time. Model weights, pretraining data and code are released alongside the embeddings under the OlmoEarth Artifact License, while the global archive’s broadest available coverage remains the three years from 2023 to 2025.