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

OlmoEarth Studio, a platform for building Earth observation models, has introduced custom embedding exports. Embeddings are compact numerical representations from the open-source OlmoEarth foundation models, designed as a fast, cost-effective way to apply the models to downstream tasks. Similar locations produce similar vectors, and the embeddings have shown strong performance in both OlmoEarth's benchmarking and independent evaluations.

Users can compute embeddings through the Studio UI or API by configuring a model run. Parameters include a drawn or uploaded polygon for the area of interest (Studio handles imagery acquisition and tiling), a time span of 1-12 monthly periods, an encoder variant (Nano at 128-dim/1.4M params, Tiny at 192-dim/6.2M params, or Base at 768-dim/89M params), spatial resolution (10, 20, 40, or 80 meters per pixel), and imagery sources (Sentinel-2 L2A, Sentinel-1 RTC, or both). The output is a Cloud-Optimized GeoTIFF (COG) with one band per embedding dimension, stored as signed 8-bit integers (values -127 to +127, -128 for nodata); floating-point vectors can be recovered via the dequantize_embeddings function in olmoearth_pretrain.

The embeddings support a range of tasks: similarity search, few-shot segmentation, change detection, and unsupervised exploration. The article notes a visualization of global structure in embeddings from 1.1M seasonal Sentinel-2 samples, with 15 k-means clusters in a PCA-reduced space. For applications requiring higher performance, Studio also supports supervised fine-tuning (SFT).

Custom-computed embeddings are available now for Studio users, and instructions for computing embeddings with the publicly available models are provided. The lightweight COG format makes results easy to share. The source code and model weights are public, so the community can inspect exactly how the embeddings are generated.