OlmoEarth v1.3 Satellite Embedding Model Released with Global Precomputed Embeddings

New Model and Global Embeddings

OlmoEarth v1.3, a satellite embedding model, has been released. Alongside the model, annual embeddings for the entire planet have been made available. These embeddings are provided at 10-meter resolution in both 64- and 128-dimensions.

Accessibility and Ease of Use

A new browser-based tool, the OlmoEarth Embeddings Explorer, allows users to explore and customize maps using these embeddings. This eliminates the need for GPU access or downloading and storing large amounts of imagery, making the technology accessible for tasks such as tracking change over time and finding similar locations.

Previously, using OlmoEarth involved fine-tuning models in OlmoEarth Studio, which required significant computational resources. The new embedding models compress satellite data into features that can be reused for various tasks with less overhead.

Technical Details and Data Sources

OlmoEarth v1.3 processes a year of radar (Sentinel-1) and multispectral (Sentinel-2 and Landsat) satellite time series. It produces a 128-dimensional vector for each 10-meter pixel on Earth, with a 64-dimensional version also available. The embeddings are generated using a vision transformer, an Attention Pooler, and a linear layer.

The precomputed global embeddings are shared on Source Cooperative to ensure broad access, allowing users to train small classifiers locally for applications like crop mapping or land-cover analysis.

Performance Benchmarks

OlmoEarth v1.3's embeddings were evaluated against AlphaEarth Foundations (AEF) and Tessera v2 across 27 evaluations spanning six continents and various tasks. The 128-dimensional embeddings scored highest on 20 of these evaluations and ranked highest overall. The 64-dimensional version led on 18 of the 27 evaluations. A new crop-type benchmark for tropical farming, PLANTEUR, was introduced as part of this evaluation.