OlmoEarth Studio Now Exports Custom Embedding Vectors as COGs

Allen Institute for AI has extended OlmoEarth Studio with the ability to compute and export custom embedding vectors derived from its open-source OlmoEarth foundation models. Published August 12, 2026, the feature lets practitioners define an area of interest, select a time range, choose an encoder variant, and receive a Cloud-Optimized GeoTIFF (COG) they can feed directly into any geospatial toolchain—QGIS, GDAL, rasterio, or bespoke scripts. Source code and model weights remain publicly available alongside the research paper, and a Colab notebook covers all four example workflows without requiring local setup.

For ML engineers already wrestling with embedding pipeline logistics—an increasingly crowded problem space as agentic workflows push models toward direct corpus interaction—the release matters because it collapses imagery acquisition, tiling, and vector computation into a single Studio job rather than a multi-step DIY pipeline against raw Sentinel archives.

Encoder Variants and Output Format

Three encoder sizes cover the compute-versus-fidelity tradeoff:

The exported COG stores one band per embedding dimension. Vectors are quantized to signed 8-bit integers, with values ranging from -127 to +127 and -128 reserved as a nodata sentinel. Floating-point vectors can be recovered via the dequantize_embeddings utility in olmoearth_pretrain. Spatial resolution options are 10 m, 20 m, 40 m, or 80 m per pixel, and the time span parameter accepts 1 to 12 monthly periods. Imagery sources are Sentinel-2 L2A, Sentinel-1 RTC, or a fusion of both. Because embeddings are computed on demand rather than drawn from a precomputed global archive, monthly granularity is fully supported—relevant for workflows that need to capture seasonal dynamics rather than annual snapshots.

Four Downstream Patterns Demonstrated

The Ai2 post walks through all four workflows using OlmoEarth-v1-Tiny (192-dim) at 40-meter resolution with Sentinel-2 L2A composites as the baseline configuration.

Similarity search extracts a query pixel's embedding and computes cosine similarity against every other pixel, producing a ranked heatmap. In the Merced, California example, built-up surfaces and road corridors score high similarity while agricultural parcels score low—without any labels. The most similar agricultural patches return cosine similarity scores of 0.89 and above against a crop-field query.

Few-shot segmentation demonstrates that a logistic regression trained on just 60 labeled pixels (20 per class) over Ca Mau, Vietnam achieves a weighted F1 of 0.84 for three land-cover classes—mangrove, water, and other—using ESA WorldCover 2021 as the label source. The accuracy curve saturates rapidly: increasing the label count from 30 to 300 produces negligible gains, which the team attributes to the encoder's pretraining having already organized ecologically meaningful distinctions. The classifier is a scikit-learn pipeline: StandardScaler followed by LogisticRegression(max_iter=2000) over the 192-band embedding stack.

Change detection compares monthly embeddings from September 2023 and September 2024 over Butte County, California using per-pixel cosine distance. The 2024 Park Fire burn scar surfaces immediately without labels or training—just two COGs and a distance computation.

PCA visualization reduces the embedding space to three dimensions, maps them to RGB, and renders a false-color image. Over Flevoland in the Netherlands, individual agricultural parcels, water bodies, and urban zones receive distinct hues that closely match the actual landscape grid.

Availability and the Fine-Tuning Path

Custom embedding exports are live for OlmoEarth Studio users as of the announcement date; access requires reaching out to the Ai2 team directly. For practitioners who need higher task-specific accuracy, Studio also supports supervised fine-tuning (SFT), which trains a task-specific model head on user-supplied labels and typically outperforms linear probes over frozen features. Ai2 notes that embedding quality remains sensitive to input imagery fidelity—persistent cloud cover, atmospheric artifacts, or gaps in the composite period can degrade the resulting vectors.

The 89M-parameter Base variant sits at the top of the OlmoEarth stack, yet the 60-label segmentation result at F1 0.84 illustrates how much representational work happens during pretraining. For teams building geospatial ML products, the COG-native export path removes one more reason to maintain custom inference infrastructure.