Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis
Hugging Face
OlmoEarth Studio can now export custom Earth-data embeddings. That makes labels optional for a lot of analysis: find similar places, map land cover, or spot change fast.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OlmoEarth Studio, AllenAI’s platform for Earth observation models, now lets users compute and export embedding vectors. These are compact numerical representations of Earth data from the open-source OlmoEarth foundation models, and the source code plus model weights are public alongside the paper, so people can inspect how the vectors are made.
The pitch is simple: embeddings are a fast, lower-cost way into the system. Studio can turn them into Cloud-Optimized GeoTIFFs, one band per embedding dimension, and users can choose an area of interest, a time span from 1 to 12 monthly periods, an encoder variant, spatial resolution, and imagery sources through the UI or API. The exported values are int8, with -127 to +127 used for data and -128 reserved for nodata.
AllenAI’s examples lean on OlmoEarth-v1-Tiny, a 192-dimensional model, though larger variants are available. With those frozen embeddings, a similarity search can highlight places that look like a query pixel; a few-shot classifier can draw a land-cover map from very few labels; and change detection can compare two dates directly. One example uses a logistic regression on just 60 labeled pixels over Ca Mau, Vietnam, and gets a weighted F1 of 0.84.
The other demos are just as blunt. A comparison between September 2023 and September 2024 shows the Park Fire burn scar in Butte County, California. PCA on embeddings from Flevoland in the Netherlands reproduces crop, water, and urban boundaries in false color. AllenAI says the same exported GeoTIFFs work with QGIS, GDAL, rasterio, or custom scripts, and that the workflow can feed into supervised fine-tuning if users need more performance than a linear probe can give.
My take — AI-written commentary, not fact-checked reporting
This is the rare AI feature that sounds boring and is actually useful. Open weights, exportable GeoTIFFs, and no-label workflows are the kind of unglamorous plumbing that makes a model matter outside a demo. The industry could use fewer grand claims and more tools that let people inspect the output and move on with their job.
Read more about this at: Hugging Face