Ai2's OlmoEarth v1.3 Turns a Year of Satellite Imagery Into 128 Numbers
Ai2 released OlmoEarth v1.3 with precomputed global satellite embeddings at 10m resolution and a browser tool that trains classifiers without a GPU.
- Ai2 released OlmoEarth v1.3, a satellite foundation model with precomputed global embeddings at 10m resolution.
- 128-dim embeddings lead AlphaEarth Foundations and Tessera v2 on 20 of 27 evaluations across six continents.
- Browser-based Embeddings Explorer trains classifiers in seconds, exports GeoTIFFs, no GPU or signup.
- Training cost dropped to roughly $2,720 of H100 time, 18% of the original v1 compute.
- Precomputed data covers 2023-2025 globally and 2017-2022 for select countries, around 150 TB per year.
- Fully open: weights on Hugging Face, data on Source Cooperative, pretraining code on GitHub.
Ai2, the Allen Institute for AI, has released OlmoEarth v1.3, an Earth-observation foundation model that compresses a year of radar and multispectral imagery into one vector for each 10-meter output cell. Ai2 has already generated those embeddings across global land areas and adjacent coastal waters, allowing teams to train lightweight classifiers without running the foundation model over raw imagery first.
Its companion Embeddings Explorer runs in a browser without an account or local GPU. Users can select a location, search for visually similar areas, label examples, fit a classifier in seconds, and export predictions as a GeoTIFF, a standard georeferenced raster format.
A year of imagery in 128 numbers
An embedding is a learned numerical summary that preserves features useful for downstream tasks. A classifier can use those features to distinguish crops, land cover, fire damage, mining activity, or other patterns from a relatively small set of labels.
OlmoEarth v1.3 processes annual time series from Sentinel-1 radar and the multispectral sensors on Sentinel-2 and Landsat. Radar captures information about surface structure and can observe through clouds, while multispectral sensors measure reflected light across several wavelength bands. The model combines those signals into a 128-dimensional vector for each 10-meter cell.
The representation is nested, so its first 64 dimensions also function as a standalone embedding. The 64-dimensional option uses roughly half the vector storage, while the full representation provides the strongest benchmark results.
Three routes from pixels to maps
| Workflow | Best suited to | Requirements | Typical output |
|---|---|---|---|
| Embeddings Explorer | Interactive prototypes and small studies | Web browser and labeled points | Similarity results or GeoTIFF |
| Precomputed archive | Regional or batch analysis | Relevant embedding tiles and standard machine-learning tools | Classifier predictions and maps |
| OlmoEarth Studio | Custom dates, sensor combinations, or fine-tuning | Source imagery, storage, and GPU access | Custom embeddings or task models |
Earlier OlmoEarth workflows centered on fine-tuning a model for each task and then running inference across satellite imagery. Precomputed embeddings remove that repeated inference step, letting multiple classifiers reuse the same features for land-cover mapping, crop classification, and change detection.
OlmoEarth leads 20 of 27 evaluations
Ai2 reports that the 128-dimensional embeddings achieve the highest score on 20 of 27 evaluations against Google DeepMind’s AlphaEarth Foundations and Tessera v2, with the strongest aggregate rank. The 64-dimensional embeddings lead 18 evaluations despite their smaller storage footprint.
The release also introduces the PLANTEUR benchmark, which tests crop-type classification in tropical agricultural regions. OlmoEarth’s embeddings produced the strongest results when classifiers were evaluated in locations excluded from training, a demanding test for models used beyond their labeled regions.
Ai2 says v1.3 required 53% as much training compute as v1.2 and 18% as much as v1. Its associated carbon emissions were 50% of v1.2 and 22% of v1. The organization estimates that rented H100 GPU time for the final pretraining run would cost about $2,720.
From crop labels to mining signals
- Map smallholder crops in Kenya from a few dozen labeled points
- Locate brick kilns around Dhaka
- Assess fire damage in Indonesia
- Track illegal gold mining in the Peruvian Amazon
Ai2’s Madre de Dios example compares each cell’s embedding between 2017 and 2020, then between 2020 and 2025. The resulting change maps align with Peru’s Operation Mercurio crackdown and the subsequent return of mining activity. The analysis uses vector differences across years, while causal attribution still requires field observations or administrative evidence.
Each global year occupies about 150 TB
The Source Cooperative archive stores annual embeddings at 10-meter resolution across land and adjacent coastal waters. Each global year occupies roughly 150 TB. At launch, 2023, 2024, and 2025 have global coverage, while data from 2017 through 2022 cover selected countries. Ai2 plans to expand the historical coverage.
Every tile includes provenance tables identifying the Sentinel-1, Sentinel-2, and Landsat scenes used to generate it, with links to the corresponding public catalogs. Given the archive’s size, most projects will download the tiles covering their study area rather than mirror an entire year.
Annual summaries leave temporal gaps
Annual embeddings compress seasonal observations into one representation. Projects that require a specific date range, monthly changes, a different sensor combination, or finer temporal resolution need to run the model through OlmoEarth Studio with the relevant source imagery.
The 10-meter grid also defines the output scale and does not provide direct subpixel measurements. Analyses requiring auditable spectral values should retain the source scenes and provenance records alongside any embedding-based predictions.
Code, weights, data, and terms
Ai2 distributes the release under the OlmoEarth Artifact License used for previous versions. Teams should review its terms before incorporating the weights, code, or embeddings into commercial or redistributed products.
Spatial evaluation requires care because nearby pixels often share terrain, weather, and land-use patterns. Holding out entire regions for testing gives a more realistic estimate of how a classifier will perform in unfamiliar locations than randomly splitting neighboring pixels between training and evaluation.
Comments
