Latest in AI
OlmoEarth v1.3: Global Satellite Embeddings, Coverage and How to Use Them
Ai2's new release lets teams build maps from reusable satellite features, but the available years and geographic coverage need a closer look.
Ai2 announced OlmoEarth v1.3 on October 22, releasing satellite embeddings, model weights and a browser-based explorer. The useful change is that teams can start mapping with precomputed features instead of first downloading satellite imagery and running a large model over it.
An embedding is a compact numerical representation. Here, it describes a location from satellite observations, so a smaller model can learn a task such as classifying land cover from those numbers.
What is available in OlmoEarth v1.3
The release announcement introduces the Base model, annual embeddings and the OlmoEarth Embeddings Explorer. In the explorer, users can compare years, look for similar places, label points to train a small classifier and export a GeoTIFF. Ai2 says the explorer requires no account.
This gives a geospatial team a useful first pass before committing to a larger fine-tuning project. A crop-mapping idea can begin with a few labels and a regional inspection rather than an infrastructure exercise.
But the archive is not a complete nine-year global record at launch.
| Property | Release detail |
|---|---|
| Spatial resolution | 10 meters |
| Embedding size | 128 dimensions, with a usable 64-dimensional prefix |
| Global years at launch | 2023, 2024 and 2025 |
| Earlier years | 2017 through 2022, selected countries |
| Main data format | Zarr v3 |
The dataset documentation makes that coverage distinction explicit. It excludes Antarctica, and some pixels have no computed embedding. A missing value must not be interpreted as evidence that an area contains no land or no change.
How the satellite model works
The model card describes a 128-million-parameter vision-transformer backbone, a 14.3-million-parameter Perceiver and a small linear projection. It accepts Sentinel-1, Sentinel-2 and Landsat imagery or time series, then compresses the representation to 128 dimensions.
The model uses the OlmoEarth Artifact License, rather than an Apache or MIT label. The weights are available, but teams should review the actual license and responsible-use terms for their application.
The technical report makes embeddings a first-class deployment target. Its architecture combines information across time and sensor types before compressing it. That separates the expensive representation-building step from the smaller downstream classifier.
For a team testing several kinds of maps over the same region, this separation is the main benefit. The next task can reuse the representation rather than repeat the full image-processing pipeline.
Benchmarks need the evaluation setup attached
In the report's eleven-task comparison using a kNN-20 classifier, the 128-dimensional embeddings lead on ten tasks against AlphaEarth Foundations and TESSERA v2. The 64-dimensional version leads on seven. These are Ai2's evaluations, not a universal ranking of every satellite-mapping workflow.
The classifier, labels and task all affect the comparison. A strong average result does not establish that the representation captures the small objects, seasonal differences or local classes a particular project needs.
Treat the explorer as a way to test a hypothesis, not as ground truth. Inspect the underlying imagery and validate against labels that were not used to build the classifier.
Accessing data without downloading the planet
Source Cooperative provides public, anonymous access with no API key. Its documentation recommends ranged reads of the Zarr data. The first 64 dimensions are stored in separate chunks, so choosing that representation cuts the embedding data read in half.
Provenance tables identify the satellite scenes behind each tile and year. That is valuable when a surprising change needs investigation: a map should have a path back to the observations supporting it.
For a production workflow, the next question is therefore quite specific. Can a small classifier built on these reusable features meet the project's accuracy needs on held-out locations? If it can, the expensive model run becomes shared infrastructure rather than a cost repeated for every new map.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help technical teams explain model releases and engineering work in clear stories that developers and founders can understand and share.