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

Ai2 just made its Earth-observation foundation model a lot more useful to people who aren't Ai2. OlmoEarth Studio can now compute and export embedding vectors — compressed numerical fingerprints of satellite imagery — as Cloud-Optimized GeoTIFFs you can drag into QGIS or load with rasterio. The underlying OlmoEarth models are open source, weights and all, so nobody has to take Ai2's word for how these vectors get made.

The pitch is that embeddings are the cheap, fast way in. Pick an area of interest, a time window from one to twelve months, an encoder size (Nano at 128 dimensions, Tiny at 192, or Base at 768), a resolution between 10 and 80 meters, and whether you want Sentinel-2 optical data, Sentinel-1 radar, or both. Studio hands back an 8-bit quantized GeoTIFF with one band per dimension. Because it's computed on demand rather than pulled from some static archive, you get embeddings tuned to your exact conditions — including monthly snapshots if you care about seasonal change rather than one annual blob.

What's genuinely striking is how little labeled data you need once you have these vectors. Ai2's team ran a similarity search over Merced, California, and a single query pixel lit up every road and building in the region while farmland stayed dark — no training involved, just cosine similarity. For land-cover mapping over mangrove forests in Ca Mau, Vietnam, they trained a logistic regression on just 60 labeled pixels — 20 each for mangrove, water, and other — and got a weighted F1 score of 0.84 across the whole region. Bumping that to 300 labels barely moved the needle, which says the Tiny encoder, with only 6.2 million parameters, had already sorted out the ecological boundaries during pretraining.

The change-detection example is the one that'll grab attention from anyone doing disaster response. Comparing September 2023 and September 2024 embeddings over Butte County, California, the Park Fire's burn scar shows up instantly as a spike in cosine distance — no training, no labels, just two GeoTIFFs and some Python. And a PCA-based false-color visualization over Flevoland in the Netherlands reproduced the polder region's agricultural grid with enough fidelity to distinguish crop types and water bodies, purely from unsupervised structure in the embeddings.

Ai2 is upfront that frozen embeddings aren't the ceiling — Studio still supports supervised fine-tuning for people who need more accuracy than a linear probe delivers, and cloud cover or missing imagery can degrade the vectors regardless of model quality. But the bigger story here is accessibility: a research lab is handing out a geospatial foundation model's internals in a format that works with tools most GIS analysts already have installed.