Check the mapping question and the licence before using OlmoEarth

Ai2's 22 October release presents OlmoEarth 1.3 with precomputed annual satellite embeddings, including 10-metre features in 64- and 128-dimensional forms. These are reusable representations for downstream work, not validated answers to every mapping question or evidence of a generally available commercial inference API. (Ai2, Ai2)

The model card names the OlmoEarth Artifact License. Its text restricts military/defence and specified extractive uses, requires attribution and distribution conditions, and does not clear other third-party rights. It is not Apache 2 or unconditional commercial permission; check the complete terms for the proposed use. (Ai2, Ai2)

Define what one output label would mean

A hypothetical land-cover project might ask whether an area belongs to a locally defined vegetation class. An embedding is a vector of features, not that label. The project needs reference examples, a decision rule and a plan to check where the rule fails. A visually convincing coloured map is not a substitute for those steps.

Resolution does not establish accuracy

A 10-metre representation describes a spatial scale; it does not promise that every 10-metre cell is correctly classified for your task. Land cover can be mixed, labels can be inconsistent and a model can respond to a change that does not mean the event you care about.

For an illustrative change analysis, compare predictions with independent reference evidence and inspect false detections. Use spatially separated validation where appropriate rather than scattering near-identical neighbouring samples between training and testing. Record uncertainty and areas where the reference data cannot support a conclusion.

Treat the rights review as part of the design

Record the exact artefact, licence version, intended use and what will be redistributed. A model's weights, an embedding dataset, source imagery and a derived map can involve different rights. If a licence restriction intersects with the project's purpose, stop and obtain a suitable clarification or agreement before proceeding. This is a due-diligence checklist, not an individual legal clearance.

Do not infer a licence from a platform badge or from another model with a similar name. Preserve the applicable attribution and downstream conditions in the project record. An accessible download does not establish that a particular customer-facing service is permitted.

Estimate the work around the features

Compare using precomputed embeddings with running inference or fine-tuning for the chosen task. List area, years, feature dimensionality, storage, transfer, labelling effort and compute requirements. The repository provides implementation material to inspect, but its existence does not mean your team already has a supported production pipeline. (Ai2)

• Test a small, permitted area using the intended workflow before estimating a larger rollout.
• Validate outputs against relevant independent reference data.
• Record failure cases, uncertainty and the review needed before decisions rely on the map.
• Cost the complete pipeline and ownership of updates, not a vendor headline about training efficiency.

No satellite map, model comparison or cost saving has been produced by KIURO for this article. For professionals evaluating the release from Switzerland, the decision is whether the task, evidence and rights fit together—not whether a global benchmark can be quoted as local accuracy.