AllenAI's OlmoEarth Studio launches custom embeddings

Allen Institute for AI (AI2) has added custom embedding exports to OlmoEarth Studio, its platform for building Earth observation models. The update lets users compute numerical representations of Earth-observation data using open-source OlmoEarth foundation models. Embeddings can be exported as lightweight Cloud-Optimized GeoTIFFs (COGs) and used for downstream tasks like similarity search, segmentation, or unsupervised exploration. The system supports monthly time spans and multiple encoder variants (Nano: 128-dim, Tiny: 192-dim, Base: 768-dim), with resolutions ranging from 10m to 80m per pixel.

Users can configure parameters like area of interest, time range, encoder variant, resolution, and imagery sources (Sentinel-2 L2A, Sentinel-1 RTC, or both) via the Studio UI or API. The embeddings are computed on-demand, allowing for flexibility in capturing seasonal dynamics. For similarity search, users can extract an embedding from a query pixel and compute cosine similarity against every other pixel, generating heatmaps that highlight similar landscapes. The system distinguishes built-up surfaces from cropland without any labels, demonstrating its ability to generalize across unlabelled data.