Allen AI's OlmoEarth Studio now exports embedding vectors directly from its open-source geospatial foundation models, giving ML teams a compact, reusable representation layer for satellite and climate data without standing up a full inference pipeline. Source code, model weights, and a tech report dropped August 12, 2026. The embeddings ship as Cloud-Optimized GeoTIFFs (COGs), ready to plug into any downstream task: similarity search, segmentation, change detection, or unsupervised clustering.
Three encoder variants cover the compute-accuracy tradeoff: Nano at 128 dimensions and 1.4M parameters, Tiny at 192 dimensions and 6.2M parameters, and Base at 768 dimensions and 89M parameters. Each run is parameterized by area of interest (any polygon), time span (1 to 12 monthly periods), spatial resolution (10m, 20m, 40m, or 80m per pixel), and imagery source (Sentinel-2 L2A, Sentinel-1 RTC, or both). Embeddings are computed on demand rather than pulled from a cached global archive. Monthly granularity is supported — seasonal dynamics are captured, not averaged away into annual composites.
| Model | Embedding Dimensions | Parameters |
|---|---|---|
| Nano | 128 | 1.4 M |
| Tiny | 192 | 6.2 M |
| Base | 768 | 89 M |
Storage is int8: signed 8-bit integers ranging from -127 to +127, with -128 as the nodata sentinel. Floating-point vectors are recovered via `dequantize_embeddings` in the `olmoearth_pretrain` package. That quantization keeps COG file sizes manageable when exporting over large areas of interest.
| Parameter | Options / Range |
|---|---|
| Area of interest | Any polygon (user-defined) |
| Time span | 1 – 12 monthly periods |
| Spatial resolution | 10 m, 20 m, 40 m, 80 m per pixel |
| Imagery source | Sentinel-2 L2A, Sentinel-1 RTC, or both |
| Storage format | int8 COG (values −127 to +127; −128 = nodata) |
In a few-shot land-cover segmentation test over Ca Mau, Vietnam, a coastal mangrove region, Tiny embeddings at 40m resolution with annual Sentinel-2 L2A composites achieved a weighted F1 of 0.84 with just 60 labeled pixels (20 per class: mangrove, water, other). Increasing labeled training data from 30 to 300 pixels barely moved accuracy. The representations do the discrimination work before the classifier sees a label. The analysis requires a rasterio read, an array reshape, and an sklearn pipeline — no GPU needed at inference.
Global embedding structure was validated across 1.1M samples from seasonal Sentinel-2 imagery. Fifteen k-means clusters in a PCA-reduced embedding space align coherently with surface type. On similarity search, a query pixel near Merced, CA returns a cosine similarity heatmap where urban fabric and road corridors score high and agricultural parcels score low — with zero training data. The model has learned enough surface geometry from pretraining that a dot product handles the retrieval.
The release does not include a pre-computed global embedding archive. Every run is billed on compute. For teams needing wall-to-wall coverage at global scale, that cost adds up fast — especially at 10m resolution with Base (89M params). Allen AI offers supervised fine-tuning (SFT) in Studio for teams with labeled data, but access is gated; contact them directly.
For ML platform teams building geospatial pipelines, start with Tiny at 40m for exploration and similarity tasks, then promote to Base at finer resolution only when held-out eval F1 justifies the compute delta. The int8 COG format integrates cleanly with standard raster tooling — rasterio, GDAL, xarray — so no new runtime dependency is needed. The main architectural risk is vendor lock-in to OlmoEarth Studio for on-demand computation. Public weights let you self-host, but the Studio API handles tiling and imagery acquisition automatically — non-trivial to replicate yourself.
