How OlmoEarth V1.3 Turns Planetary Data Into Embeddings

🔍 Read the full analysis: How OlmoEarth V1.3 Turns Planetary Data Into Embeddings on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

The OlmoEarth team has released version 1.3 of its satellite embedding model, as detailed in the original analysis, accompanied by annual global data at 10-meter resolution and a browser-based explorer. The developers say the model ranked highest in 20 of 27 evaluations, but those performance results are from the team’s own testing and have not been independently verified in the supplied material.

OlmoEarth v1.3 processes yearly time series from Sentinel-1 radar and multispectral imagery from Sentinel-2 and Landsat. It produces a 128-number vector, or embedding, for each 10-meter pixel, which can support downstream analysis. The first 64 dimensions can be used as a smaller representation. The team says precomputed global embeddings are available through Source Cooperative.

The accompanying OlmoEarth Embeddings Explorer displays embeddings as false-color maps, lets users switch between years and compare results with Sentinel-2 imagery. According to the release description, users can select a location to find visually similar places or add labeled points to train a small classifier. The team presents the browser tool as requiring no sign-in or advanced technical knowledge.

In its comparison with AlphaEarth Foundations and Tessera v2, the team reports that the 128-dimensional embeddings scored highest in 20 of 27 evaluations, while the 64-dimensional version led in 18. The evaluations covered six continents and multiple tasks. The developers also introduced PLANTEUR, a crop-type benchmark focused on tropical farming, and say OlmoEarth generalized best to locations not seen by the classifier during training. The full results and methods are in the team’s technical report.

Making Satellite Mapping More Reusable

Precomputed embeddings may reduce the work involved in using satellite imagery for land-cover classification or change monitoring. Rather than downloading and processing large image archives for each research question, users can start with reusable numerical representations and train a smaller model for a particular task. That may help teams without dedicated GPU resources, although it does not remove the need to check the outputs against imagery and other evidence.

The release also makes the team’s model weights, pretraining data and code available. The developers estimate that reproducing the final pretraining run for OlmoEarth v1.3 Base would cost about $2,720 in rented H100 computing time. This is the team’s estimate, and actual costs can vary by provider, availability and setup. Researchers will still need to assess how the model performs in the regions and applications they care about.

The team’s evaluation counts are promising as a report of its testing, not a guarantee of results in operational use. Independent replication and task-specific checks would help establish whether the model’s performance holds for different regions, labels and data conditions.

From Task Models to Shared Embeddings

Earlier OlmoEarth workflows centered on fine-tuning through OlmoEarth Studio: users trained a model for a specific task and ran it over satellite imagery. The team says that approach requires GPU access and the ability to download and store imagery. Embeddings offer another workflow: process imagery into reusable vectors, then use those vectors for multiple mapping tasks.

The developers report that v1.3 used 53% of the computing time of v1.2 and 18% of the time used for v1. They also report emissions equal to 50% of v1.2 and 22% of v1, and say v1.3 matched v1.2 in fine-tuning performance across research and partner labels. The supplied material does not detail the emissions accounting, so those comparisons remain the team’s reported figures.

The release illustrates the approach with maps of mining-related landscape change in Peru’s Madre de Dios region, comparing 2017–2020 with 2020–2025 around La Pampa. It references Amazon Conservation’s MAAP monitoring program and the 2019 Operation Mercurio crackdown. The example shows how changed embeddings can flag differences over time; the maps alone do not establish the cause of those changes.

“OlmoEarth’s 128-dimensional embeddings score highest on 20 of 27 evals and rank highest overall.”

— The OlmoEarth team

Independent Tests Still Needed

The reported evaluation results come from the developers. Although the team’s technical report contains methods and results, independent replication is not described in the supplied material. How the evaluation tasks were selected and how performance varies by location will matter when judging the model for a particular use.

The release description also does not specify how often the annual global embeddings will be updated, what data gaps or quality limits apply across all locations, or what access and computing requirements users may face for large-scale projects. In the Peru example, a difference between yearly embeddings indicates a change in the mapped landscape; attributing that change to mining, recovery or another cause requires corroboration from imagery, monitoring records or field evidence.

Testing the Data in Practice

Researchers and other users can access the precomputed embeddings through Source Cooperative and explore them with the browser-based tool. The team’s technical report provides the basis for examining the evaluation methods and comparing the reported results with tests on local data.

The next practical test is whether the embeddings support reliable results across specific regions, years and mapping tasks, including areas with different imagery conditions or limited training labels. Independent evaluations, clearer information about data updates and coverage, and comparisons with local reference data would help establish where v1.3 is useful. The supplied material does not state a schedule for those tests or for future dataset updates.

Key Questions

What is OlmoEarth v1.3?

It is the OlmoEarth team’s satellite embedding model. It turns yearly imagery from Sentinel-1, Sentinel-2 and Landsat into numerical representations for each 10-meter pixel.

What are the embeddings used for?

The vectors can be reused for tasks such as land-cover mapping and monitoring changes over time. Users may train smaller classifiers on the representations rather than starting each project by processing extensive satellite image archives.

Has the model’s reported performance been independently confirmed?

Not in the supplied source material. The OlmoEarth team reports that its model led in 20 of 27 evaluations; the figures should be treated as the developers’ results pending independent testing.

Where can users view the embeddings?

The team says the precomputed data is available through Source Cooperative. Its browser-based Embeddings Explorer can display false-color maps, switch between years and compare results with Sentinel-2 imagery.

Can an embedding map identify why a landscape changed?

No, not by itself. A difference between years can indicate mapped change, but identifying a cause such as mining requires corroboration from other evidence, including satellite imagery, monitoring records or field data.

Primary source: Hugging Face · via ThorstenMeyerAI.com