OlmoEarth v1.3: State-of-the-art embeddings for the whole planet

What it claims

Ai2's OlmoEarth v1.3 Base is an embedding model whose precomputed annual embeddings can be used for mapping tasks with only a small classifier on top. Fine-tuning a model per task is not required. Against AlphaEarth Foundations (AEF) and Tessera v2 on 27 evaluations across six continents, the 128-dimensional embeddings score highest on 20 and rank highest overall. The 64-dimensional truncation, at half the storage, still leads on 18 of 27. The release also introduces PLANTEUR, a new crop-type benchmark for tropical farming. On PLANTEUR, Ai2 reports that its embeddings generalize best to places the classifier has never seen.

How it works

A vision transformer reads one year of Sentinel-1 radar plus Sentinel-2 and Landsat multispectral time series. An Attention Pooler collapses the output into one vector per patch of land. A linear layer then reduces that vector to 128 numbers.

The model builds on the same foundation as earlier OlmoEarth versions, tuned to keep embeddings lightweight. Ai2 states it still matches v1.2 as a fine-tuning model on research and partner labels.

Change detection needs no training. Comparing a pixel's vector across years highlights change. The post demonstrates this on illegal gold mining at La Pampa, Madre de Dios, Peru. The 2017 to 2020 comparison shows mining spreading southeast. The 2020 to 2025 comparison shows that area going quiet after Operation Mercurio in February 2019, with continued change in the older zone to the west.

Evidence

• Embedding benchmarks: wins on 20 of 27 evaluations (128-d) and 18 of 27 (64-d) against AEF and Tessera v2. Per-task metrics are only in the technical report.
• Example applications named: smallholder crops in Kenya, brick kilns west of Dhaka, fire damage in Indonesia, and gold mining in the Peruvian Amazon.
• Training compute: 53% of v1.2's compute time and 18% of v1's. Carbon emissions are 50% and 22% of those versions respectively.
• Reproduction cost: pre-training the final v1.3 Base run costs about $2,720 of rented H100 time.

How to use it

Precomputed embeddings

• Hosted on Source Cooperative at 10 m resolution, in 64-d or 128-d form.
• Covers land and adjacent coastal waters, at roughly 150 TB per year.
• At launch, 2023, 2024 and 2025 are global. 2017 to 2022 cover only select countries.
• Each 20 km tile and year has an open provenance table. It lists the Sentinel-1, Sentinel-2 and Landsat scenes used, with links to the public catalogs that host them.

Embeddings Explorer

• Browser tool with no account and no GPU.
• Supports similarity search from a clicked point, year switching, and a Sentinel-2 imagery overlay.
• Can train a small classifier from a few labeled points and export the result as a GeoTIFF.

Custom time windows

• OlmoEarth Studio runs the same model over user-chosen dates, modalities and resolution.
• A tutorial covers generating embeddings on your own compute.

Model weights, pretraining data and pretraining code are all released under the OlmoEarth Artifact License, the same license as v1 and v1.2.

Limits

• Comparisons cover only AEF and Tessera v2. The post gives no per-task scores, and AEF or Tessera v2 lead on 7 of 27 evaluations, which the post does not name.
• Embeddings are annual composites, so sub-year timing needs Studio or self-generated embeddings.
• Global coverage exists only for 2023 to 2025. Earlier years are partial.
• At 150 TB per year, local use realistically means pulling regional subsets.
• The license is a custom Ai2 license, not a standard open-source one. Its terms should be checked before commercial use.