Satellite embeddings stop being per-task: OlmoEarth v1.3 computes one vector per 10-meter pixel worldwide and tops 20 of 27 evaluations
Synopsis
OlmoEarth v1.3 turns satellite embeddings from per-task fine-tuning into compute-once, reuse-everywhere: it reads a year of Sentinel-1 radar and Sentinel-2 and Landsat multispectral time series and outputs a 128-dimensional vector for every 10-meter pixel on Earth, with the first 64 dimensions usable on their own. Until now the primary way to use OlmoEarth was fine-tuning in OlmoEarth Studio, training a model per task, which requires GPU access and the ability to download and store imagery. Against AlphaEarth Foundations and Tessera v2 across 27 evaluations on six continents, the 128-dimensional embeddings score highest on 20 of 27 and rank highest overall, while the 64-dimensional version leads on 18 of 27.
Interpretation
Satellite imagery shifts from training one model per task to computing embeddings once and training a small classifier on top. Until now the primary way to use OlmoEarth was fine-tuning in OlmoEarth Studio, training a model for each task and running it over satellite imagery to make a map, which requires GPU access and the ability to download and store the imagery. Embeddings compress satellite data into reusable features, so a crop map, a land-cover map or a change analysis becomes a lightweight modeling problem: train a small classifier on top and run it locally.
The model compresses a year of radar and multispectral time series into one vector per 10-meter pixel. Embedding models compress satellite data into ready-to-use features that can be reused cheaply for many tasks. A vision transformer reads the satellite data, an Attention Pooler pools it into one vector per patch of land, and a linear layer reduces it to 128 numbers; the first 64 dimensions also work on their own as a smaller embedding.
The 128-dimensional embeddings rank highest on most evaluations, and the half-size 64-dimensional version still leads. The comparison is against AlphaEarth Foundations (AEF) and Tessera v2, covering six continents and many different tasks. Across 27 evaluations, the 128-dimensional embeddings score highest on 20 and rank highest overall, while the 64-dimensional version leads on 18; the evaluation also introduces PLANTEUR, a new crop-type benchmark for tropical farming.
This generation is the cheapest OlmoEarth model to train while remaining a strong fine-tuning model. Training compute time and carbon emissions both fall relative to the previous version v1.2 and the first version v1. It took 53% of v1.2's computing time and 18% of v1's, with 50% and 22% of the carbon emissions, and matches v1.2's performance across research and partner labels.
Perspective
The demonstrated scope is global annual embeddings at 10-meter resolution: 2025, 2024 and 2023 are global, 2017 to 2022 cover select countries, and more coverage is on the way; the embeddings cover land and adjacent coastal waters and are available in 128 or 64 dimensions. The intended readers are researchers and engineering teams doing regional mapping and change analysis from a handful of labels, either in the no-sign-in Embeddings Explorer or by running the same model in OlmoEarth Studio over chosen dates, modalities and resolution.
The evaluation conclusion is limited to the 27 reported evaluations across six continents, with the 128- and 64-dimensional versions leading on 20 and 18 of them respectively, not all of them. The training-cost figure refers to the final pretraining run, estimated at roughly $2,720 of rented H100 time. In the Madre de Dios example, the 2020-to-2025 embedding change map is not dark, showing the old mining zone keeps changing, an interpretation that relies on MAAP's monitoring records.