Original research brief
OlmoEarth Studio Adds Custom Embedding Exports for Geospatial Analysis
OlmoEarth Studio now lets users compute and export custom embedding vectors as Cloud-Optimized GeoTIFFs for similarity search, segmentation, change…
OMNIASSIST / FIELD NOTESnews · source-led editorial
What this piece is grounded in
01
OlmoEarth Studio now lets you compute and export embedding vectors
02
The source code and model weights are publicly available alongside the research paper
03
Studio delivers a COG with one band per embedding dimension
04
Vectors are stored as signed 8-bit integers (int8)
01 / FIELD NOTE
What changed
OlmoEarth Studio, a platform for building Earth observation models, now lets users compute and export embedding vectors—compact numerical representations of Earth-observation data produced by the open-source OlmoEarth foundation models. The source code and model weights are publicly available alongside the research paper, so the community can inspect exactly how these embeddings are generated. Users can choose their area of interest, time range, encoder variant, resolution, and imagery sources via the Studio UI or API, and receive a Cloud-Optimized GeoTIFF (COG) with one band per embedding dimension. Vectors are stored as signed 8-bit integers, with values ranging from -127 to +127 and -128 reserved for nodata.
02 / FIELD NOTE
Why it matters for practical AI work
Embeddings are described as a fast, cost-effective entry point for leveraging OlmoEarth, supporting tasks from similarity search to segmentation to unsupervised exploration. Locations with similar surface characteristics end up with similar vectors, enabling operations like finding 'more like this' pixels, few-shot segmentation with a simple linear classifier, change detection by comparing monthly embeddings, and PCA-based visualization. The exported COGs are lightweight and easy to share, working with any geospatial tool such as QGIS, GDAL, or rasterio. Because embeddings are computed on demand rather than pulled from a pre-computed global archive, they reflect exactly the conditions a user cares about, such as seasonal dynamics.
03 / FIELD NOTE
What a reader can verify
The post includes examples using OlmoEarth-v1-Tiny (192-dim) embeddings at 40-meter resolution with Sentinel-2 L2A composites. A few-shot segmentation test over Ca Mau, Vietnam, labeled just 60 pixels (20 per class) and produced a coherent land-cover map with weighted F1 = 0.84. Change detection using monthly embeddings from September 2023 and September 2024 highlighted the Park Fire burn scar in Butte County, California. The post links to a tech report, documentation, an embeddings tutorial with working code, and a Colab notebook for hands-on testing without local setup. Users can also compute their own embeddings using the publicly available OlmoEarth models.
04 / FIELD NOTE
What remains unknown
The post does not specify pricing, availability limits, or a timeline for general access to custom embedding exports in Studio; it says to 'reach out if you're interested in gaining access.' Performance claims are based on the authors' own benchmarking and independent evaluations, but the post advises checking embedding quality for your use case. It also notes that performance depends on input imagery quality—persistent cloud cover, atmospheric artifacts, or missing observations can affect the resulting vectors. The post does not detail the exact compute or storage costs of using larger encoder variants.
Questions readers ask
What are OlmoEarth embeddings?
OlmoEarth embeddings are compact numerical representations of Earth-observation data produced by the open-source OlmoEarth foundation models. They are computed on demand in OlmoEarth Studio and exported as Cloud-Optimized GeoTIFFs, with one band per embedding dimension. They support downstream tasks like similarity search, few-shot segmentation, change detection, and unsupervised exploration.
How can I use the exported embeddings?
The exported GeoTIFF works with any geospatial tool, including QGIS, GDAL, rasterio, or your own scripts. The post includes an embeddings tutorial with working code for similarity search, few-shot segmentation, change detection, and PCA visualization, plus a Colab notebook for hands-on use without local setup. You can also fine-tune the models with supervised fine-tuning in Studio for higher performance.