OlmoEarth V1.3: Mapping Earth With AI Embeddings

🔍 Read the full analysis: OlmoEarth V1.3: Mapping Earth With AI Embeddings on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

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, paired with annual global data at 10-meter resolution and a browser-based map explorer. The team says the model ranked highest on 20 of 27 evaluations, a result that could help researchers assess land cover and landscape change without processing large satellite archives themselves; those performance findings are the developers’ own evaluation and have not been independently verified in the supplied material.

OlmoEarth v1.3 turns yearly image time series from Sentinel-1 radar and multispectral imagery from Sentinel-2 and Landsat into a 128-number representation for each 10-meter pixel, an approach explored in AI-powered Earth data analysis. The first 64 dimensions can be used as a smaller representation. These compact numerical representations, called embeddings, can be reused as input for different mapping tasks rather than requiring users to start each project by processing the underlying imagery.

The team says the precomputed data is available through Source Cooperative. Its 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 and train a small classifier. The team presents the browser tool as usable without a sign-in or advanced technical expertise.

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

Lowering Barriers to Satellite Mapping

Precomputed embeddings could reduce the storage and computing work needed to develop land-cover maps or compare landscapes over time. Instead of repeatedly downloading and processing extensive image archives, a researcher could use the shared representations to train a smaller classifier or examine differences between years. That may be useful to organizations without dedicated GPU resources, although the results still need to be checked against suitable imagery, labels and local knowledge.

The developers also say the model weights, pretraining data and code are available. They estimate that reproducing the final pretraining run for OlmoEarth v1.3 Base would cost about $2,720 in rented H100 computing time. That figure is the team’s estimate, not a guaranteed price; actual costs depend on computing provider, availability and setup. Open materials may help others inspect the work, but the availability of code alone does not establish that the reported benchmark advantage will hold in every application.

The headline evaluation counts are relevant to users choosing a general-purpose representation, but performance for a specific task depends on its location, labels and conditions. Independent comparisons could clarify whether the reported advantage persists across new regions and operational uses. The release presents a tool and reusable data, not a substitute for validating maps before decisions are made.

From Task Models to Reusable Embeddings

Earlier OlmoEarth work centered on fine-tuning a model for a particular task in OlmoEarth Studio, then running it over satellite imagery to produce a map. The team says that workflow requires GPU access and downloading and storing imagery. The embedding approach changes the sequence: imagery is processed into reusable vectors, which can then support multiple analyses.

The developers say v1.3 used 53% of the computing time of v1.2 and 18% of the computing time of the original v1 to train. They also report emissions equal to 50% of v1.2 and 22% of v1, and say v1.3 matches v1.2 in fine-tuning performance across research and partner labels. These are comparisons from the team; the supplied release material does not give the underlying emissions-accounting details.

One example in the release uses embeddings to examine landscape changes near La Pampa in Peru’s Madre de Dios region, comparing maps for 2017–2020 and 2020–2025. The team describes the maps as showing changes around an area affected by mining. The release references Amazon Conservation’s MAAP monitoring and the 2019 Operation Mercurio crackdown as relevant history. A change in embeddings can flag a difference in the landscape, but cannot by itself establish that mining caused it.

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

— The OlmoEarth team

Questions About Validation and Coverage

The reported evaluation results come from the OlmoEarth developers. Although the team’s technical report contains its methodology, the supplied source does not establish that outside researchers have reproduced the 27 evaluations. How task selection, regional variation and benchmark conditions affect the result will matter when assessing the model for a particular use.

The release description also does not specify how often the annual data will be updated, what gaps or quality limitations apply in every location, or what access and computing requirements users may face for large-scale analysis. The source material says the data is available through Source Cooperative but does not give enough detail here to assess all practical limits of access.

In the Peru example, a difference between annual embeddings is a signal of landscape change, not proof of its cause. Attribution to mining, forest loss or recovery needs corroboration from underlying imagery and other evidence, such as monitoring data or field observations. The supplied material does not establish independent verification of the example’s interpretation.

Testing the Maps in Practice

Researchers and other users can explore the precomputed data through Source Cooperative and try the browser-based explorer described by the team. The next useful test is whether the representations perform well on tasks and locations outside the developers’ reported evaluations, including when users supply their own labels.

Further scrutiny will depend on access to the technical report and open materials, as well as independent comparisons with existing models. Users evaluating the maps should check apparent changes against source imagery and relevant local evidence. The supplied release material does not give a schedule for new annual datasets or an independent evaluation date.

Key Questions

What is OlmoEarth v1.3?

It is a satellite embedding model that converts yearly Sentinel-1, Sentinel-2 and Landsat imagery into numerical representations for 10-meter pixels. Those representations can be reused in mapping and classification tasks.

What did the OlmoEarth team report in its evaluations?

The team says its 128-dimensional embeddings ranked highest in 20 of 27 evaluations, while the 64-dimensional version led in 18. These are developer-reported results; the supplied material does not show independent verification.

Can I view the embeddings without installing a model?

The release includes a browser-based explorer for viewing false-color maps, switching years and comparing with Sentinel-2 imagery. The team says it can be used without signing in.

Do the maps identify why a landscape changed?

No. Embeddings can indicate differences between images or years, but they do not establish the cause of a change. Users need to check interpretations against imagery and other relevant evidence.

Are the reported results independently confirmed?

The supplied material describes the evaluations as the OlmoEarth team’s own results and does not provide an independent replication. The team’s technical report contains its methods and results.

Primary source: Hugging Face · via ThorstenMeyerAI.com