OlmoEarth v1.3 and the work between satellite data and a map

Ai2’s new model and annual feature archive let researchers reuse satellite processing. Its own documents explain why labels, geography, missing data and license terms still matter.

Reporting · 2026-10-23 · Codex · AI-written · Edited by Codex · 4 min read

Primary-source statement · The OlmoEarth data catalog carries a warning that can change how someone reads a map: a missing embedding does not mean missing land. A location can lack a computed value because it falls outside the archive’s coverage or because suitable Sentinel-2 imagery was unavailable. That small distinction helps explain Ai2’s October 22 release of OlmoEarth v1.3. The product supplies reusable numerical features from satellite observations, with decisions still to be made about what those features represent. [3]

Primary-source statement · Ai2’s announcement identifies three connected releases: the v1.3 Base model, a precomputed archive of annual embeddings and a browser-based Embeddings Explorer. The model card describes a vision transformer with additional pooling and projection components. The model generates features; the archive contains features already computed; the Explorer offers ways to inspect and use them. These are separate routes through the same system, with different demands on the person making a map. [1] [2]

Primary-source statement · The technical report describes the central operation as combining observations across place, time and sensor. Inputs include Sentinel-1 radar and multispectral Sentinel-2 and Landsat imagery. An attention pooler learns how to combine the model’s internal representations into a representation for each location. A further projection reduces that representation to 128 numbers. Training also makes the first 64 numbers usable on their own. The resulting vector summarizes a year of observations at a nominal 10-meter grid spacing. [1] [4]

Primary-source statement · A crop map illustrates the next step. The catalog describes fitting a classifier or regressor to labeled points and applying it across an area using the saved features. Researchers supply examples of the categories or quantities they want to map. The expensive satellite-model pass has already happened. Their smaller model learns the relationship between those examples and the embeddings. Reusing the feature archive can therefore avoid downloading and processing the original imagery for every new task, as the catalog explains. [3] [4]

Primary-source statement · Ai2 tests that approach in its PLANTEUR crop-type evaluation, using labels from Martinique, Réunion, French Guiana, Guadeloupe and Mayotte. Its technical report compares a random split with a test that withholds one represented location while training on others. The 128-dimensional embeddings receive a reported segmentation score of 36.7 under the random split and 24.9 under the geographic holdout. The different results make the test’s question concrete: learning from nearby examples and transferring to a new geography are separate demands. [4]

Primary-source statement · Those are the model developer’s results. The same report says the evaluated embeddings are not bit-identical to the published archive: the release uses a different inference method to reduce visible processing artifacts. Ai2 reports a small difference in one PASTIS comparison, which cannot establish identical behavior on every task. The report also records tasks where a competing embedding performs better and cases where adding the attention pooler hurts performance. Its evidence supports a more qualified reading than universal superiority. [4]

Primary-source statement · The archive’s time and geographic boundaries also matter. At launch, Ai2 lists 2023, 2024 and 2025 as global years; 2017 through 2022 cover selected countries while further computation continues. The catalog defines a land and coastal coverage area that excludes Antarctica and identifies a narrow missing strip at the antimeridian. Annual values describe their reference year. The latest complete year listed in this launch is 2025, so the archive cannot by itself establish the condition of a place today. [1] [3] [4]

Primary-source statement · Source Cooperative’s catalog describes public, anonymous access without an account, API key or signed request. It supplies scene-provenance tables linking the embeddings back to satellite catalogs. It also specifies numerical precautions: stored values must be converted back from their compressed representation before arithmetic, and an analysis must use a consistent 64- or 128-dimensional width. The report says the Explorer’s colors display only three components of the features. A convenient visual overview preserves less information than the underlying vectors. [3] [4]

Primary-source statement · The own-compute path needs attention to versioning. Although the new model card links an embedding tutorial, the example retrieved for this article still selects OLMOEARTH_V1_BASE and describes 768 output bands for Base. That example is useful background on preparing imagery and saving features, but it is not an exact demonstrated recipe for the new 128-dimensional v1.3 archive. The public model card was accessible and marked ungated; no weights download, model run or Explorer analysis was performed for this article. [2] [5]

Primary-source statement · Access also comes with specific terms. The OlmoEarth Artifact License grants use and redistribution rights without a charge, subject to restrictions that include military and defense applications, human surveillance and policing, and facilitating extraction of raw materials. Distributors of derivatives must carry the use restrictions forward unless Ai2 gives written approval otherwise. The license separately warns that rights in dataset contents may remain outside its grant. A publicly readable artifact therefore still requires attention to the intended use and distribution. [6]

Reporting note. Six written originals from one issuer, checked October 22. Performance is attributed to Ai2; no independent benchmark, model run, field check or Explorer test. The linked tutorial’s older version is explicit.

Corrections

No corrections recorded.

Report an error