What are OlmoEarth embeddings?

OlmoEarth embeddings have just been rolled out as a brand new feature in OlmoEarth Studio - a platform that's all about building Earth observation models. This is a big deal, as it lets users compute and export embedding vectors that can be used for all sorts of downstream tasks. The embeddings themselves are generated using the open-source OlmoEarth foundation models, and the best part is that the source code and model weights are out there for everyone to see. This means the community can take a closer look at how the embeddings are generated and use them for their own projects.

The embeddings are super versatile, supporting a whole range of tasks - we're talking similarity search, few-shot segmentation, change detection, and unsupervised exploration. Take similarity search, for instance - it can be used to track down locations with similar surface characteristics. And then there's few-shot segmentation, which can be used to label the landscape with discrete labels. The embeddings have already shown some serious promise in benchmarking and independent evaluations, making them a really valuable tool for users. Plus, the exported Cloud-Optimized GeoTIFFs (COGs) are nice and lightweight, easy to share, and users can pick and choose their area of interest, time range, encoder variant, resolution, and imagery sources via the Studio UI or API.

So, how do you get started with OlmoEarth embeddings? Well, users can reach out to get access to custom-computed embeddings or use the publicly available OlmoEarth models to compute their own embeddings. There are instructions available for using the models, and users can tailor the output to their specific needs by tweaking parameters like area of interest, time span, encoder variant, spatial resolution, and imagery sources. With OlmoEarth embeddings, users can unlock some really valuable insights into Earth observation data and use them for all sorts of applications.

Source: Hugging Face