What happened?

OlmoEarth Studio, Ai2's platform for building Earth observation models, now lets users compute and export embedding vectors — compact numerical representations of Earth-observation data produced by the open source OlmoEarth foundation models.

The source code and model weights are publicly available alongside the research paper, so the community can inspect exactly how these embeddings are generated.

What embeddings are for

Embeddings are a fast, cost-effective entry point for leveraging OlmoEarth. They support a wide range of downstream tasks, from similarity search to segmentation to unsupervised exploration.

The working principle is simple: locations with similar surface characteristics end up with similar vectors, while locations that differ land far apart. For applications requiring higher performance, Studio also supports supervised fine-tuning.

How it is configured

  • Area of interest: draw or upload any polygon; Studio handles imagery acquisition and tiling.
  • Time span: 1–12 monthly periods
  • Encoder variant: Nano (128-dim, 1.4M params), Tiny (192-dim, 6.2M params) or Base (768-dim, 89M params)
  • Spatial resolution: 10, 20, 40 or 80 metres per pixel
  • Imagery sources: Sentinel-2 L2A, Sentinel-1 RTC, or both
  • Output: a COG file with one band per embedding dimension

The shape of the data

Studio delivers a COG with one band per embedding dimension. Vectors are stored as signed 8-bit integers (int8); values range from -127 to +127, with -128 reserved for nodata. To recover floating-point vectors, `dequantize_embeddings` in `olmoearth_pretrain` is used.

Because everything is computed on demand rather than pulled from a pre-computed global archive, the embeddings reflect exactly the conditions of interest. Monthly embeddings can be generated to capture seasonal dynamics, rather than being limited to annual snapshots.

Why does it matter?

The most expensive part of working with satellite imagery is usually not the imagery itself but getting it into a usable form. Embedding vectors deliver that step ready-made: instead of raw pixels you hold a numerical representation across which similarity can be computed.

That the weights and code are open is a separate matter. An embedding forms the ground for every analysis built on top of it; not being able to see how that ground was produced means not being able to see what the results rest on.

How the encoder is chosen

The difference between the three encoders is not only size but computational cost. Nano is the lightest at 1.4 million parameters; Base is the heaviest at 89 million and produces 768-dimensional vectors. Tiny sits between them and is the option used in most of the documented examples.

A larger vector dimension means a larger file for the same area — because one band is written per dimension, a 768-dimensional output takes six times the space of a 128-dimensional one. For anyone scanning a wide area in monthly periods, that is a direct storage and transfer cost.

What is not settled

The company says the embeddings have shown strong performance in its own benchmarking and in independent evaluations, but the report does not give numerical results for those evaluations. Custom-computed embeddings are for now available only to OlmoEarth Studio users, and access requires getting in touch with the company.