Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis
Hugging Face’s OlmoEarth Studio now allows users to compute and export custom embedding vectors from Earth observation data, enabling downstream analysis such as similarity search, segmentation, and change detection without requiring labeled datasets.
OlmoEarth Studio, developed by Hugging Face in collaboration with Allen Institute for AI, now supports the computation and export of embedding vectors derived from Earth observation data using open-source OlmoEarth foundation models. Users can generate these embeddings on demand through the Studio UI or API, selecting parameters such as area of interest, time range, encoder variant, resolution, and imagery sources. The exported embeddings are provided as Cloud-Optimized GeoTIFFs (COGs), which are lightweight and designed for easy sharing and further analysis.
The embeddings enable a range of downstream tasks, including similarity search to identify locations with comparable surface characteristics, few-shot segmentation for land-cover mapping with minimal labeled data, and change detection to highlight shifts between time periods. Performance benchmarks, including internal and independent evaluations, indicate strong results for these applications. For applications requiring higher accuracy, Studio also supports supervised fine-tuning, allowing users to refine models based on their specific needs.
Similarity search leverages cosine similarity to compare embedding vectors, producing heatmaps that visually indicate areas with landscape characteristics similar to a user-defined query pixel. For example, queries near urban centers highlight built-up surfaces while agricultural areas remain distinct, demonstrating the model’s ability to differentiate land cover without explicit training. Few-shot segmentation demonstrates the potential of embeddings for land-cover classification with minimal labeled data, achieving a weighted F1 score of 0.84 using just 60 labeled pixels in a coastal mangrove region.
Change detection compares embeddings from different time periods to identify surface condition shifts, such as burn scars from wildfires, without requiring labels or training. Unsupervised exploration techniques, such as PCA-based false-color visualization, reveal inherent structural patterns in the embeddings, automatically grouping similar landscapes like agricultural parcels or urban areas. These capabilities underscore the utility of OlmoEarth embeddings for applications ranging from ecological monitoring to disaster assessment.