The Allen Institute for AI has added embedding export to OlmoEarth Studio, its platform for building Earth observation models. The update, described in an Ai2 blog post authored by Patrick Johnson, Favyen Bastani, Gabriel Tseng, Chris Wilhelm, Joseph Redmon, Hunter Pitelka, Patrick Beukema, Mike Jacobi, and Hadrien Sablon, lets users compute and export compact numerical representations of satellite and radar imagery derived from Ai2’s open-source OlmoEarth foundation models. Source code and model weights are publicly available, as is the accompanying research paper.
What embeddings are and what they support
Embedding vectors are numerical representations of input data — in this case, Earth-observation imagery — that position similar locations close together in vector space and dissimilar locations far apart. According to the post, OlmoEarth embeddings support similarity search, segmentation, and unsupervised exploration, and the exported files are delivered as Cloud-Optimized GeoTIFFs (COGs).
Users configure an embedding run by specifying an area of interest, a time span of one to twelve monthly periods, an encoder variant, a spatial resolution, and imagery sources. Three encoder variants are available: Nano (128 dimensions, 1.4 million parameters), Tiny (192 dimensions, 6.2 million parameters), and Base (768 dimensions, 89 million parameters). Spatial resolution options are 10, 20, 40, or 80 metres per pixel. Imagery sources are Sentinel-2 L2A optical imagery, Sentinel-1 RTC radar imagery, or both. Vectors are stored as signed 8-bit integers, with values ranging from -127 to +127 and -128 reserved for no-data.
Because embeddings are computed on demand rather than pulled from a pre-computed archive, the post notes that users can generate monthly embeddings to capture seasonal dynamics rather than relying on annual snapshots.
Similarity search
The post demonstrates similarity search by selecting a query pixel near the Merced urban centre in California and computing cosine similarity against every other pixel in the region. The resulting heatmap shows urban fabric and road corridors as highly similar to the query, while agricultural parcels are dissimilar. Switching to an agricultural query pixel and defining the query vector as the mean over a small window, the most similar patches — with cosine similarity of 0.89 and above — are all irrigated agricultural parcels. The least similar patches are an airport with bare ground, a reservoir with dry terrain, and arid rangeland. The post notes this is done without training data or labels.
Few-shot segmentation
The post describes a segmentation experiment over Ca Mau, Vietnam, a coastal mangrove region. Using 60 labelled pixels — 20 each for mangrove, water, and other land cover, drawn from ESA WorldCover 2021 — the team trained a logistic regression classifier on OlmoEarth-v1-Tiny embeddings and predicted labels for every pixel in the region. The classifier produced a weighted F1 score of 0.84. The post reports that increasing the number of training labels from 30 to 300 barely changed accuracy, and attributes this to the embeddings already encoding the relevant ecological distinctions during pretraining. The analysis used the 192-dimension Tiny encoder at 40-metre resolution.
The experiment is a standard linear probe evaluation for foundation models. The post notes that larger variants — Base, with 768 dimensions — encode richer representations at higher compute and storage cost.
Change detection
Because Studio can generate embeddings at any temporal resolution, the post demonstrates change detection by computing monthly embeddings for the same region in September 2023 and September 2024 and measuring per-pixel cosine distance. The Park Fire burn scar in Butte County, California, which burned from July to September 2024, is described as immediately visible in the result without labels or training.
Unsupervised exploration via PCA
The post also demonstrates principal component analysis applied to OlmoEarth embeddings. Reducing the embedding to three dimensions and mapping them to red, green, and blue channels produces a false-colour image. Applied to Flevoland in the Netherlands — a reclaimed polder landscape with a regular agricultural grid — the post states the PCA image reproduces parcel boundaries with high fidelity, with crop types, water bodies, and urban areas each receiving distinct colours.
Access
Custom-computed embeddings in Studio are available to users who contact Ai2. Instructions for computing embeddings using the publicly available OlmoEarth models directly are also available via the post. If an application requires higher performance than embedding similarity supports, Studio also offers supervised fine-tuning.
If the application requires higher performance, the post notes that Studio also supports supervised fine-tuning.