TESSERA — A pixel-wise earth observation foundation model

What is TESSERA?

The coloured overlay is not a photograph. Each square is a 10m×10m pixel, but instead of three RGB numbers, TESSERA stores 128 numbers that encode a full year of satellite observations. Developed at the University of Cambridge, the system compresses a year's worth of Sentinel-1 (radar) and Sentinel-2 (optical) passes into dense, per‑pixel embeddings — a thousand tiny time-series portraits of the planet. Look at it long enough and seasonal rhythms, not single snapshots, start to sing.

How it works

Millions of satellite passes feed a self‑supervised encoder that learns by comparing random temporal views of the same pixel. No human labels are required during training. A field that greens in spring and turns golden in autumn gets a different embedding from an evergreen stand, even if one still image could fool you. It’s the foundation‑model playbook — the transfer learning trick that powered modern NLP and vision — applied pixel by pixel. Stick a small task‑specific head on top and you can fine‑tune for land‑cover maps, solar‑panel detection, crop monitoring, and more.

Claims, openness and backing

It has been reported that TESSERA can recover over 90% of final task performance using less than 1% of labelled data — a bold efficiency claim that, if borne out across use cases, would be a big deal for organisations with scarce annotation budgets. The project is advertised as globally covering terrestrial areas at 10m resolution and is being built in the open, with every layer customisable. The team acknowledges computational support from DAWN (Cambridge’s AI supercomputer), AMD, the UKRI AIRR Isambard facility, Vultr, and donations from Jane Street; satellite data come from the Copernicus programme.

Why it matters

Why should anyone care? Because encoding time into each pixel could turn a noisy satellite snapshot into a reliable story about land change, agriculture, and infrastructure — faster and cheaper. For climate scientists, NGOs, and local planners, that’s not just a neat demo; it could be a practical tool. Skepticism is healthy — real‑world performance and governance questions remain — but if the early claims hold, TESSERA may be the sort of open infrastructure that democratizes high‑resolution Earth observation.

Sources: geotessera.org, Lobsters