Ai2 released OlmoEarth v1.3, satellite embeddings for the whole planet

The Allen Institute for AI, known as Ai2, has released OlmoEarth v1.3, a satellite embedding model, along with precomputed annual embeddings for the entire planet and a new browser tool for exploring them called the OlmoEarth Embeddings Explorer.

An embedding compresses what a model learned from millions of images into a compact numerical fingerprint. OlmoEarth v1.3 reads a year of radar data from Sentinel-1 and multispectral imagery from Sentinel-2 and Landsat, and produces one 128 dimensional vector for every 10 meter pixel on Earth. The first 64 dimensions also work on their own as a smaller embedding. Ai2 has computed these for the whole planet and is sharing them in both 64 and 128 dimension versions.

That changes how people can use the model. Until now, the main path was fine-tuning in OlmoEarth Studio: training a model for each task and running it over satellite imagery to make a map, which needs GPUs and the ability to download and store large amounts of imagery. With precomputed embeddings, the expensive step is done once and the results can be reused cheaply for many tasks, such as tracking change over time or finding places that look similar. Ai2 says the explorer requires no sign-in and no advanced technical knowledge.

Ai2 describes v1.3 as state of the art for satellite embeddings and has published a technical report, model weights, data and code alongside the release.

Why it matters for developers: planet scale embeddings turn earth observation into something closer to a vector search problem. Teams working on agriculture, climate, insurance, mapping or supply chains can build classifiers or similarity search on top of ready made features instead of running their own imagery pipelines.

What to try: open the Embeddings Explorer on a region you know, then pull the embeddings for a small area and test a simple classifier against your own labels.

Building with AI? See this week's hot AI repos and how to use each one.