60 labeled pixels map Vietnam's mangroves with 84% accuracy via OlmoEarth

OlmoEarth Studio now lets you export embedding vectors from its open-source Earth observation models. These compact numerical representations turn satellite imagery into something you can search, cluster, and classify without training a full model. The workflow mirrors any other prediction in Studio: draw a polygon, pick a time span of 1-12 months, choose an encoder variant (Nano at 128 dimensions up to Base at 768), set spatial resolution from 10 to 80 meters, and select Sentinel-2, Sentinel-1, or both. You get back a Cloud-Optimized GeoTIFF with one band per embedding dimension, stored as int8 values ranging from -127 to +127, with -128 reserved for nodata.

The company’s own benchmarking and independent evaluations both point to strong performance, but the real proof is in the demos. Query a pixel near Merced, California, and the cosine similarity heatmap lights up urban corridors while agricultural parcels stay dark — no labels involved. Switch the query to an irrigated field and the most similar patches (0.89 and above) are all agricultural, while airports and dry rangeland score near zero. That kind of semantic separation is what makes embeddings useful.

The few-shot segmentation test is the most striking result. Researchers labeled just 60 pixels — 20 each for mangrove, water, and other — over Ca Mau, Vietnam’s coastal mangrove region. A logistic regression trained on those 60 samples produced a wall-to-wall map with weighted F1 of 0.84. For context, that’s a competitive result from a classifier that took minutes to train, not weeks. The embeddings are doing the heavy lifting.

Because Studio computes embeddings on demand rather than serving from a pre-computed archive, you can generate monthly vectors to track seasonal dynamics. That’s a meaningful upgrade over annual snapshots for anyone watching crop cycles, water levels, or vegetation shifts. Access is currently gated — you need to reach out for Studio access — but the model weights and source code are public, so technically inclined users can compute embeddings themselves. The int8 quantization keeps file sizes manageable, and a dequantization function is provided in the olmoearth_pretrain repository.