OlmoEarth v1.3 delivers open satellite embeddings for the entire planet

Hugging Face has released OlmoEarth v1.3, a new state-of-the-art model that generates vector embeddings from satellite imagery. Alongside the model weights, the team published precomputed annual embeddings for the entire planet and a browser-based exploration tool. This launch aims to lower the barrier for developers and researchers working with geospatial data by removing the need for heavy computational resources.

What happened

The core of this release is the OlmoEarth v1.3 model, which processes satellite time series data into compact, reusable feature vectors. Unlike previous versions that required users to fine-tune models on raw imagery using significant GPU power, this version provides ready-to-use embeddings. Hugging Face computed these embeddings for every 10-meter pixel on Earth for multiple years and made them available via Source Cooperative. The release also includes the OlmoEarth Embeddings Explorer, an interactive web tool that allows users to visualize changes over time and train simple classifiers without writing code or managing infrastructure.

This approach shifts the workflow from training large vision models from scratch to performing lightweight modeling on top of existing features. Developers can now build applications such as crop mapping, land-cover classification, or change detection by training small classifiers on the provided vectors. The team evaluated the new embeddings against competitors like AlphaEarth Foundations and Tessera v2 across 27 different benchmarks. OlmoEarth v1.3 achieved the highest overall score, leading in 20 of the 27 evaluations with its 128-dimensional vectors and in 18 evaluations with its smaller 64-dimensional variant.

The release emphasizes openness and efficiency. All model weights, pretraining data, and code are publicly available under the OlmoEarth Artifact License. The team reported that training OlmoEarth v1.3 was significantly cheaper and more energy-efficient than prior iterations, consuming only 53% of the compute time of version 1.2 and 18% of version 1. This efficiency makes it feasible for other organizations to reproduce the results, with the final training run estimated to cost approximately $2,720 in rented H100 GPU time.

How it works

OlmoEarth v1.3 operates by reading annual time series data from three major satellite sources: radar data from Sentinel-1, and multispectral data from Sentinel-2 and Landsat. The model uses a vision transformer architecture to process this input. An Attention Pooler component aggregates the information from patches of land into a single representation, which is then reduced to a 128-dimensional vector by a linear layer. These vectors, or embeddings, capture the semantic meaning of the landscape at each 10-meter pixel. The first 64 dimensions of this vector can also function independently as a smaller, lighter embedding.

Because the embeddings compress complex satellite data into fixed-size numbers, they enable efficient comparison and classification. Users can compare vectors from different years to detect changes, such as deforestation or urban expansion, by measuring the distance between them. The system also maintains provenance records, linking each embedding back to the specific satellite scenes used to generate it. This transparency allows users to verify the source data for any given location.

Key details

• OlmoEarth v1.3 produces 128-dimensional embeddings for every 10-meter pixel on Earth, with a usable 64-dimensional subset.
• Precomputed global embeddings are available for 2023, 2024, and 2025, with select country coverage for 2017 to 2022.
• The model outperformed AlphaEarth Foundations and Tessera v2 in 20 out of 27 benchmark evaluations.
• Training costs were reduced by 47% compared to v1.2, with carbon emissions cut by half relative to the previous version.
• The OlmoEarth Embeddings Explorer allows users to train classifiers and export GeoTIFFs directly in the browser without GPU access.
• All assets, including weights, code, and data, are open-source under the OlmoEarth Artifact License.

Why it matters

For software engineers and data scientists, this release dramatically reduces the friction involved in building geospatial applications. Traditionally, creating accurate land-use maps required downloading terabytes of raw satellite imagery, setting up complex preprocessing pipelines, and training large models on expensive hardware. With precomputed embeddings, these steps are eliminated. Developers can focus on solving specific domain problems, such as identifying smallholder crops or tracking mining activity, by training simple, fast classifiers on top of the provided vectors. This democratization of access means that smaller teams and organizations with limited resources can now perform planetary-scale analysis.

The availability of high-quality, open embeddings also accelerates research and innovation in environmental monitoring. By providing a standardized, high-performance baseline, the community can build more reliable tools for tracking climate change impacts, illegal logging, and agricultural trends. The inclusion of provenance data ensures that results are auditable, which is critical for scientific and regulatory applications. Furthermore, the efficiency gains in training demonstrate that high-performance foundation models do not always require unsustainable levels of energy consumption, pointing toward a more sustainable future for AI development.

What you can do

• Visit the OlmoEarth Embeddings Explorer to visualize planetary changes and test similarity searches without installing any software.
• Download the precomputed 64- or 128-dimensional embeddings for your region of interest from Source Cooperative.
• Train a lightweight classifier on the embeddings to create custom maps for tasks like crop detection or land-cover analysis.
• Use the provenance tables to trace specific embeddings back to their original Sentinel-1, Sentinel-2, or Landsat source scenes.
• Experiment with OlmoEarth Studio if you need to generate embeddings for specific time windows or resolutions not covered in the precomputed set.
• Review the technical report and PLANTEUR benchmark details to understand the evaluation methodology and model limitations.