🔍 Read the full analysis: What Can Planet-Wide Embeddings Do? Meet OlmoEarth V1.3 on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
The OlmoEarth team has released v1.3 of its satellite embedding model, annual global embeddings at 10-meter resolution, and a browser-based explorer. The team says its 128-dimensional version ranked highest in 20 of 27 evaluations, but the results have not been independently verified in the source material.
The OlmoEarth team has released version 1.3 of its satellite embedding model, alongside annual global data at 10-meter resolution and a browser-based map explorer, as described in the original analysis. The release could make some land-cover and change-mapping work easier to start without downloading large satellite archives or training a large model, though the team’s performance comparisons are developer-reported, not independently verified in the available source material.
OlmoEarth v1.3 turns yearly satellite image time series into a 128-number vector for each 10-meter pixel, with custom exports available for downstream analysis. The model uses radar observations from Sentinel-1 and multispectral imagery from Sentinel-2 and Landsat. Users can also use the first 64 dimensions as a smaller representation. The team says precomputed global embeddings are available through Source Cooperative.
The accompanying OlmoEarth Embeddings Explorer lets users view the data as false-color maps, switch between years, and compare the results with Sentinel-2 imagery. According to the release description, a user 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 signing in or having advanced technical skills.
In comparisons with AlphaEarth Foundations and Tessera v2, the OlmoEarth team reports that the 128-dimensional version ranked highest in 20 of 27 evaluations, while its 64-dimensional version ranked highest in 18. The evaluations covered multiple tasks across six continents. The team also introduced PLANTEUR, a benchmark focused on tropical crop types, and says OlmoEarth generalized best to locations not seen by the classifier during training. Its technical report contains the full methods and results.
Reusable Data for Land Mapping
Satellite imagery can support mapping of crops, forests, urban growth and other land changes, but working with large image archives can demand substantial computing, storage and technical capacity. With precomputed embeddings, a researcher can reuse compact representations to train smaller task-specific classifiers or compare locations across years, rather than processing the full imagery collection for every project.
That approach may lower the entry cost for teams with limited access to computing resources. It does not remove the need to check a map against source imagery and other evidence: an embedding is a model-derived representation, not a direct explanation of why a landscape changed. The value of the release will depend on how well it performs in particular regions and tasks, as well as on the data’s coverage and quality.
satellite imagery analysis software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Earlier OlmoEarth use centered on fine-tuning through OlmoEarth Studio, where a user trained a model for a particular task and applied it to satellite imagery. The team says that workflow requires GPU access and the ability to download and store imagery. The v1.3 embedding approach instead processes imagery into reusable pixel-level features that can be applied to different mapping tasks.
The developers report that v1.3 used 53% of the computing time of v1.2 and 18% of the time used to train v1. 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 figures are team-reported; the source material does not detail the emissions accounting. The team says model weights, pretraining data and code are available, and estimates that reproducing the final pretraining run for OlmoEarth v1.3 Base would cost about $2,720 in rented H100 computing time. Actual costs can vary by provider and setup.
The release illustrates the embeddings with maps of mining-related landscape change in Peru’s Madre de Dios region, including areas around La Pampa. The team compares changes across 2017–2020 and 2020–2025. Its example refers to monitoring by Amazon Conservation’s MAAP program and the 2019 Operation Mercurio crackdown. The mapped differences indicate detected change; they do not, on their own, establish its cause.
“OlmoEarth’s 128-dimensional embeddings score highest on 20 of 27 evals and rank highest overall.”
— The OlmoEarth team
As an affiliate, we earn on qualifying purchases.
Independent Checks Still Needed
The reported benchmark rankings come from the OlmoEarth team’s own evaluation. The technical report is identified as the source for methods and full results, but the supplied material does not establish that independent researchers have reproduced the comparisons. How results vary by region, task selection and available labels will matter when judging whether the reported advantage applies to a specific project.
The release description also does not specify how often the annual embeddings will be updated, what gaps or quality limitations apply in every location, or what access and computing requirements arise for large-scale use. In the Peru example, attributing a mapped difference to mining or recovery requires additional evidence, such as underlying imagery, monitoring records or field observations.
As an affiliate, we earn on qualifying purchases.
Testing the Embeddings in Use
Users can explore the precomputed data through Source Cooperative and inspect locations with the browser-based explorer. The team’s technical report provides the next reference point for readers who want to review the evaluation design and results. The available source material does not give a schedule for future data releases or a date for independent evaluation.
Practical testing will show how the embeddings perform on users’ own regions, labels and mapping questions. Researchers and other users can compare the representations with source imagery and established monitoring data, while reporting limitations such as missing observations or regional differences. Until independent testing is available, the rankings and training-cost and emissions comparisons should be treated as claims reported by the developers.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is new in OlmoEarth v1.3?
The release includes the v1.3 model, annual global satellite embeddings at 10-meter resolution, and a browser-based explorer. The embeddings represent each pixel with 128 numbers, with a 64-dimensional option using the first half.
What are satellite embeddings used for?
They are compact, model-generated representations of satellite imagery that users can reuse for tasks such as land-cover classification and change mapping. They can reduce the need to process large image collections for each new task, but results still need validation.
Did OlmoEarth independently prove its benchmark lead?
The supplied material reports rankings from the OlmoEarth team’s evaluation: the 128-dimensional version led in 20 of 27 evaluations and the 64-dimensional version in 18. It does not document independent replication of those results.
Can the embeddings identify why a place changed?
No. A difference between yearly embeddings can flag possible landscape change, but it does not establish the cause. Attribution, such as linking a change to mining, needs corroboration from imagery and other monitoring or field evidence.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
