TL;DR
The Allen Institute for AI has released OlmoEarth v1.3, a satellite embedding model, alongside annual 10-meter embeddings covering the planet and a browser-based explorer. The team reports that its 128-dimensional embeddings led on 20 of 27 evaluations; independent verification and details of the global data’s coverage and update schedule are not provided in the release.
The Allen Institute for AI has released OlmoEarth v1.3, a model that turns satellite time series into reusable features, and published annual embeddings at 10-meter resolution covering the whole planet. The release also includes a browser-based OlmoEarth Embeddings Explorer, intended to let people compare places and observe changes without downloading satellite imagery or training a model first.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The model combines a year of radar observations from Sentinel-1 with multispectral data from Sentinel-2 and Landsat. It produces a 128-number vector for every 10-meter pixel; the first 64 dimensions can be used as a smaller representation. The team describes these vectors as compressed representations of patterns learned from satellite imagery, which can be reused for different mapping tasks.
Annual global embeddings are available through Source Cooperative, according to the report. Users can build applications such as crop or land-cover maps and change analyses by training a smaller classifier on the embeddings, rather than processing the underlying imagery for every task. The accompanying explorer displays the data as a false-color map, allows comparisons between years and provides access to underlying Sentinel-2 imagery for visual checks.
The report says OlmoEarth v1.3 was evaluated against AlphaEarth Foundations and Tessera v2 across 27 evaluations spanning six continents and multiple tasks. Its 128-dimensional version ranked highest on 20 evaluations and highest overall; the 64-dimensional version led on 18. These are results reported by the model’s developers. The team also introduces PLANTEUR, a benchmark for crop-type mapping in tropical farming, and says the model generalized best to locations absent from classifier training.
Making Satellite Analysis Easier to Reuse
The release shifts part of satellite mapping from repeated, resource-intensive image processing toward precomputed features that can support multiple tasks. That could lower the technical and computing burden for researchers, public agencies and other users who want to map land cover, monitor change or identify particular activities. The explorer offers a way to inspect patterns before building a custom analysis.
The potential value is practical: the report describes applications including smallholder crop mapping in Kenya, identifying brick kilns near Dhaka, assessing fire damage in Indonesia and monitoring gold mining in Peru. These examples are presented by the developers; they do not establish that the tool independently verifies activity or replaces field checks. Reliable results still depend on the task, training labels and validation.
Cost and access are also part of the announcement. The team says v1.3 required 53% of v1.2’s computing time and 18% of v1’s, with reported carbon emissions at 50% and 22% of those earlier versions, respectively. It says model weights, pretraining data and code are open. These comparisons concern the team’s own earlier models and training runs, not the cost of every downstream use.
Products Worth Considering
Apple Watch Ultra 4 GPS + Cellular, 49 mm
Apple Watch Series 10 GPS 46mm with Silver Aluminum Case and Blue Cloud Sport Loop
LG UltraGear 45GX950A-B 45-Inch 5K2K Curved OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
From Fine-Tuning to Global Embeddings
Before this release, the team says the main way to use OlmoEarth was through OlmoEarth Studio: users fine-tuned a model for a particular task and ran it over satellite imagery to make a map. That approach can require GPU access and the ability to download and store imagery. Embeddings offer a different workflow: compute a representation once, then reuse it for multiple applications.
The new release extends earlier OlmoEarth versions, which the team says were also designed for satellite-image analysis. It reports that v1.3 remains a strong fine-tuning model and matches v1.2 across research and partner labels. The technical report contains the full evaluation results and methodology. The developers estimate that reproducing the final pretraining run for OlmoEarth v1.3 Base would cost about $2,720 in rented H100 computing time; that estimate is for the stated run, not a general price for using the released embeddings.
The report illustrates year-to-year comparisons with Madre de Dios, Peru. It describes embedding changes around La Pampa, an area affected by gold mining, and links its interpretation to monitoring by Amazon Conservation’s MAAP program. The example is an illustration of how the explorer can reveal changing patterns; it does not establish that the model alone identifies the cause of every change.
“Compute once, reuse everywhere.”
— OlmoEarth v1.3 release team
Independent Checks and Data Coverage
The supplied release describes the team’s benchmark results but does not include an independent evaluation. Readers would need the technical report and its methodology to assess how the 27 evaluations were selected, how comparisons were run and how performance varies by geography or task. A high ranking across those tests does not guarantee equal accuracy in every location.
The announcement says annual embeddings cover the whole planet, but it does not specify here the years available for every region, the publication schedule for future years, or how missing or low-quality satellite observations are handled. It also does not detail data licensing and access terms beyond naming Source Cooperative as the distribution point and describing model materials as open.
Embeddings register patterns of change; interpreting those patterns may require imagery, local knowledge or other evidence. The Peru example, for instance, discusses mining and regrowth, but the release does not claim that a changed vector by itself proves a specific cause. The accuracy and practical limits of individual applications remain dependent on validation.
More Detail in Reports and Explorer
Users can examine the technical report, model weights, data and code linked from the release and access the global embeddings through Source Cooperative. The browser-based explorer is available for comparing years and inspecting locations against Sentinel-2 imagery. The release does not give a schedule for additional model versions or future annual embedding updates.
The next useful test is how well independent teams can reproduce the reported results and adapt the embeddings to local tasks. For readers assessing a map or analysis, the key checks are whether the relevant location and year are covered, whether the result has been compared with source imagery or ground observations, and whether the classifier was tested on places outside its training data.
Key Questions
What is OlmoEarth v1.3?
It is a satellite embedding model released by the Allen Institute for AI. It processes yearly radar and multispectral satellite time series into 128-dimensional vectors for 10-meter pixels, with a 64-dimensional option.
What are satellite embeddings used for?
They are reusable numerical representations of satellite data. A user can train a smaller classifier on them for tasks such as land-cover or crop mapping, rather than processing imagery from scratch for each application.
How did OlmoEarth v1.3 perform in evaluations?
The release team says the 128-dimensional embeddings led on 20 of 27 evaluations, while the 64-dimensional version led on 18. Those are developer-reported results; the technical report provides the evaluation details.
Can the embeddings identify the cause of a landscape change?
Not by themselves in every case. The embeddings can help flag or compare patterns, but attributing a change to mining, fire, farming or another cause may require imagery, local information and independent validation.
Where can people access the release?
The report points to Source Cooperative for the precomputed global embeddings and links to the model materials and OlmoEarth Embeddings Explorer. The supplied announcement does not state a complete coverage schedule or update timetable.
Source: rss
