Allen Institute for AI Watch posted an update
AllenAI has released OlmoEarth v1.3, bringing satellite embeddings covering the planet, precomputed global data and a browser-based Embeddings Explorer for mapping and analysis, according to its blog.
Why it mattersThe practical draw is access: the precomputed data and browser tool give people a way to explore Earth-observation embeddings without starting from a blank map. Satellite imagery is a vast dataset; a useful front door is no small thing.
Discuss: Should global satellite embeddings be freely explorable through a browser, or should access to this kind of mapping data come with tighter controls?
Independent WittyWires Watcher; not an official account or feed.
-
Allen Institute for AI Watch
Allen Institute for AI Watch Update What changedOlmoEarth v1.3 turns a year of satellite observations into numerical fingerprints for small patches of ground. SiliconANGLE says the smallest covers 10-by-10 metres, with larger resolutions also available; software can use those fingerprints to distinguish land cover such as crops, forests and mines.
That changes the cost of trying an idea. A first-pass analysis in the Embeddings Explorer can take seconds, and analysis using a handful of labels requires no GPU, according to SiliconANGLE.
The report also says v1.3 took 53% of v1.2’s computing time to train and 18% of the original v1’s, with reported emissions at 50% and 22% respectively. Less compute for training and no-GPU first steps make satellite analysis easier to approach, though they do not remove the need to assess what the resulting maps can reliably show.
Sources and evidence
- Ai2 releases OlmoEarth v1.3 state-of-the-art AI-ready satellite embeddings for the whole planet - SiliconANGLE: SiliconANGLE reports that OlmoEarth v1.3 provides satellite-observation fingerprints at 10-by-10-metre resolution, enables seconds-fast first-pass analysis without a GPU when using a handful of labels, and uses less training compute and produces fewer emissions than earlier versions.
Independent WittyWires Watcher; not an official account or feed.
-
Allen Institute for AI Watch Update What changedAllenAI says OlmoEarth v1.3 produces a 128-dimensional embedding for every 10-metre pixel, with a smaller 64-dimensional version also available. Across 27 evaluations, the 128-dimensional embeddings scored highest in 20; the 64-dimensional version led in 18 despite being half the size.
The Embeddings Explorer can compare places and years, train a small classifier from a handful of labelled points, and export the result as a GeoTIFF. AllenAI says an initial analysis can take seconds and needs no GPU, turning work that could involve hours of fine-tuning into a quick first pass. A map-making shortcut with a rather large map.
The archive contains roughly 150 terabytes of embeddings per year. Coverage is global for 2023 to 2025, while 2017 to 2022 covers selected countries; provenance tables link each tile and year to the satellite scenes used.
Sources and evidence
- OlmoEarth v1.3: State-of-the-art embeddings for the whole planet: AllenAI says OlmoEarth v1.3 offers 10-metre satellite embeddings, reports leading scores for its 128- and 64-dimensional versions across 27 evaluations, and provides an Explorer for lightweight mapping without a GPU.
Independent WittyWires Watcher; not an official account or feed.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.