HomeAIOlmoEarth Studio Adds Custom Embedding
AI

OlmoEarth Studio Adds Custom Embedding Exports

OlmoEarth Studio users can now compute and export Cloud-Optimized GeoTIFF embedding vectors generated by open-source Earth observation models.

WHAT YOU NEED TO KNOW
  • OlmoEarth Studio now exports custom Earth observation embedding vectors as Cloud-Optimized GeoTIFFs.
  • Users can select Nano (1.4M params), Tiny (6.2M params), or Base (89M params) encoder variants.
  • Configurations support spatial resolutions from 10 to 80 meters per pixel across 1 to 12 monthly periods.
  • A linear classifier trained on 60 labeled pixels achieved an F1 score of 0.84 in a Vietnam land-cover test.

Allen Institute for AI updated its OlmoEarth Studio platform to allow users to compute and export custom embedding vectors from its open-source OlmoEarth foundation models, according to reporting published by Hugging Face on August 12, 2026. The platform generates compact numerical representations of Earth-observation data for downstream geospatial analysis. Users can configure parameters in the Studio interface or via an API and download the results as Cloud-Optimized GeoTIFFs (COGs).

The exported COG files store vectors as signed 8-bit integers ranging from -127 to +127, with -128 reserved for nodata values. Users can recover floating-point vectors using the dequantize_embeddings function in the olmoearth_pretrain library. The source code, research paper, and model weights are publicly available alongside documentation.

Export options

Configuration settings in Studio allow users to select an area of interest by drawing or uploading a polygon. Time spans range from one to 12 monthly periods, allowing monthly tracking of seasonal changes rather than relying on annual snapshots. Users select spatial resolutions of 10, 20, 40, or 80 meters per pixel, with input imagery drawn from Sentinel-2 L2A, Sentinel-1 RTC, or both sources combined.

Studio offers three encoder variants for embedding generation: Nano, Tiny, and Base. The Nano encoder uses 1.4 million parameters across 128 dimensions. The Tiny encoder uses 6.2 million parameters across 192 dimensions, while the Base encoder uses 89 million parameters across 768 dimensions.

Downstream applications

In benchmarking examples published by Hugging Face, the 192-dimensional Tiny encoder supported several analytical workflows without task-specific training. Cosine similarity search on Sentinel-2 imagery distinguished urban structures in Merced, California, and separated irrigated agricultural land from bare ground and reservoirs. Unsupervised Principal Component Analysis mapped embedding values to three-dimensional color spaces, revealing field boundaries across polder land in Flevoland, Netherlands.

For land-cover classification, a logistic regression classifier trained on 60 labeled pixels—20 each for mangrove, water, and other classes—achieved a weighted F1 score of 0.84 across Ca Mau, Vietnam, using ESA WorldCover 2021 as a reference. Monthly embedding comparisons measured per-pixel cosine distance between September 2023 and September 2024 to map the burn scar of the Park Fire in Butte County, California.

While frozen embeddings handle similarity search and few-shot segmentation, OlmoEarth Studio also supports supervised fine-tuning by training a task-specific model head directly on user labels. Input imagery for the platform relies on European Space Agency Sentinel-2 L2A data accessed through the Microsoft Planetary Computer.

Xentir Media
Xentir Media NewsroomSource-backed AI and technology coverage, drafted by Xentir's automated editorial system under fixed human-set rules. See our editorial policy and AI usage policy.
J
Jomon · Founder & EditorFounder and editor of Xentir Media. Sets the editorial rules the newsroom system runs under, and is accountable for its corrections. About Jomon · [email protected]
The Xentir Brief
The developments worth knowing — one useful email.
Get the Brief →