Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis

Summary

Hugging Face’s OlmoEarth Studio now supports on-demand computation and export of embedding vectors from its open-source foundation models. Users select an area of interest, time range, encoder variant, and input imagery—typically Sentinel-2 L2A composites—through the Studio interface or API, after which the platform returns a Cloud-Optimized GeoTIFF containing one band per embedding dimension. The vectors are stored as signed 8-bit integers ranging from –127 to +127, with –128 reserved for nodata, yielding compact files that remain compatible with standard geospatial tools such as QGIS, GDAL, or rasterio.

These embeddings capture surface characteristics in a continuous vector space, so locations with similar spectral and structural properties lie close together while dissimilar areas are separated. The studio examples demonstrate several practical uses without task-specific training. Cosine similarity between a query pixel or small region and every other pixel produces heatmaps that highlight matching land-cover types, such as irrigated fields or urban fabric. A linear probe—logistic regression trained on as few as 60 labeled pixels—can generate coherent land-cover maps, for instance distinguishing mangrove, water, and other classes in coastal Vietnam. Monthly embeddings allow direct comparison of two time slices via cosine distance, revealing changes such as the 2024 Park Fire burn scar in California. Unsupervised PCA reduction to three components, mapped to RGB, visualizes latent structure; in the reclaimed polder landscape of Flevoland the resulting false-color image delineates agricultural parcels, water bodies, and built-up areas without any labels.

The lightweight OlmoEarth-v1-Tiny encoder (192 dimensions) at 40 m resolution is used for most demonstrations, though larger variants offer richer representations at higher compute cost. Because embeddings are generated on demand rather than drawn from a static archive, users can produce seasonal or monthly time series tailored to specific conditions. When higher accuracy is required, Studio also supports supervised fine-tuning of a task-specific head. Performance remains dependent on input data quality, and users are advised to validate embeddings for their particular region and application before operational use.

Why it matters

This article provides ML engineers with actionable tooling and methodologies for leveraging geospatial foundation models in resource-constrained environments. The explicit demonstration of agricultural parcel mapping in the Netherlands makes it highly pertinent for Dutch AI practitioners in agritech and climate tech.