OlmoEarth embeddings export

OlmoEarth Studio added a feature to compute and export embedding vectors. It takes the output of the publicly released OlmoEarth foundation model and delivers it in Cloud-Optimized GeoTIFF (COG) format, ready to use directly for downstream analysis.

Users can select the region (AOI), time range (1-12 months), encoder variant (Nano, Tiny, Base), spatial resolution (10/20/40/80m), and imagery source (Sentinel-2 L2A, Sentinel-1 RTC, or both). The storage format is int8, with a value range of -127 to 127, and -128 reserved for nodata.

There are four core use cases.

• Similarity search: find terrain on the map similar to a query pixel.
• Few-shot segmentation: with just 60 pixels, produced a wall classification map in Ca Mau, Vietnam with weighted F1 0.84.
• Change detection: compared embeddings from September 2023 and September 2024 to capture the Park Fire burn scar in Butte County.
• Unsupervised exploration: used PCA to reduce embeddings to 3-dimensional color, visualizing farmland patterns in Flevoland.

This approach's strength is that it can be used quickly and cheaply even with little or no labeling. However, if higher performance is needed, Studio's supervised fine-tuning (SFT) can be used, and the impact of input imagery quality or cloud/atmospheric artifacts should be checked.