Ai2 released OlmoEarth v1.3 on October 22 with precomputed satellite embeddings and a browser Explorer. An accessible numerical representation is not automatic confirmation of events on the ground.

What the representation contains

The model card describes a 128-million-parameter ViT-Base for Sentinel-1, Sentinel-2 and Landsat imagery, including time series. Additional components reduce the representation to 128 dimensions. The weights follow the OlmoEarth Artifact License and Ai2's responsible-use guidelines, rather than an assumed universal permission for every possible use.

In the technical report, the authors explain three challenges: aggregating observations, working efficiently at high resolution, and compacting the final vector. They provide comparative evaluations, which remain their research results. We do not translate a ranking into an accuracy guarantee for Bulgarian farmland or an unfamiliar local problem.

For a general reader, an embedding can be understood as a feature description on which a smaller classifier can learn. The vector itself is not a finished label such as “forest” or “field.” Users still need a defined question and verified examples against which to evaluate the answer. That distinction matters when a convincing map can easily be mistaken for an established finding.

Coverage has specific boundaries

The public dataset describes annual vectors at 10-meter resolution, with the first 64 dimensions usable separately. At initial release, 2023–2025 have global coverage, while 2017–2022 cover selected countries. Antarctica is excluded. Provenance tables identify the contributing satellite scenes. Quantized values require the documented conversion before arithmetic; a missing embedding does not mean that no land exists.

The browser Explorer is announced as accessible without sign-in, supporting similar-place searches, labeled points and map export. It offers a useful exploratory starting point. An annual summary is nevertheless not a live observation and should not be presented as evidence that a change occurred today.

How we would evaluate a local task

For a pilot, we would choose a small region and a clearly bounded question, such as separating two familiar land-use types. First we would check the required year and location are covered. We would then reserve some reliable labels for evaluation, rather than judging the classifier only on the points used to build it. This is a proposed procedure, not an experiment we performed.

A second check would be spatial: success near the training points should not be the sole basis for judging a new area. We would record difficult boundaries, including mixed pixels, incomplete coverage and classes defined differently by different observers. That makes errors inspectable instead of hiding them behind one overall percentage.

For comparisons between years, we would seek the source imagery and another available check. A difference between vectors gives a reason to inspect a place; it does not establish the cause by itself. In an editorial or municipal analysis, we would retain the model version, year, vector width and labels so the result can be reproduced.

Responsibility for defining the task and checking the map remains with the team using it. For an owned project, we would first require a small, testable hypothesis and an acceptable result on held-out examples. Only then would we expand the analysis. Local knowledge belongs in the evaluation, particularly when two similar-looking areas serve different purposes.

Cover: conceptual AI illustration of satellite data; not an actual map or evidence of a change on the ground.

Checked against the listed sources: 2026-10-23T06:29:40.415508+00:00.