OlmoEarth v1.3 turns a year of satellite data into one vector per pixel, leading 20 of 27 evaluations

Using satellite imagery shifts from training a model per task to taking a precomputed feature and training a small classifier on top. The primary way to use OlmoEarth had been 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. Across 27 evaluations covering six continents and many tasks, compared with AlphaEarth Foundations and Tessera v2, the 128-dimensional embeddings scored highest on 20 of 27 and ranked highest overall, while the 64-dimensional version led on 18.

An embedding compresses what the model learned from millions of images into reusable features, so a crop map, land-cover map or change analysis becomes a lightweight modeling problem. Embedding models compress satellite data into features that can be reused cheaply for many tasks, while foundation models and global embeddings are expensive to produce. OlmoEarth v1.3 reads a year of Sentinel-1 radar and Sentinel-2 and Landsat multispectral time series and produces one 128-dimensional vector for every 10-meter pixel on Earth, with the first 64 dimensions also usable alone as a smaller embedding.

The same embeddings support year-to-year comparison directly, reading landscape change without training a model. Illegal gold mining around La Pampa in Peru's Madre de Dios peaked in 2017 and 2018, and Operation Mercurio shut most of it down in February 2019; the question since has been whether it stayed down. Comparing each pixel's vector across years, the 2017 to 2020 pair lights up where mining spread southeast, and the same comparison from 2020 to 2025 shows the signal shifting, with the old mining zone still changing.

The demonstrated scope is global annual embeddings at 10-meter resolution, roughly 150 TB per year, covering land and adjacent coastal waters; at launch 2025, 2024 and 2023 are global, while 2017 to 2022 cover select countries, with more coverage on the way. Embeddings are retrievable in the full 128-dimensional version or the lighter 64-dimensional version, and each 20-kilometer tile and year carries provenance tables listing the Sentinel-1, Sentinel-2 and Landsat scenes the model read, with links to public catalogs. When a specific time window, modality or resolution is needed, OlmoEarth Studio runs the same model over chosen dates, and a tutorial shows how to generate embeddings on your own compute.

The evaluation conclusions are limited to the reported 27 evaluations and six continents, including the newly introduced tropical crop-type benchmark PLANTEUR. The model is described as keeping embeddings lightweight without giving up accuracy across tasks, and the technical report gives the bounds of that efficiency trade-off. Precomputed global coverage at launch reaches only 2023 to 2025, with earlier years covering select countries.