Tessera is an open-source, pixel-wise Earth-observation foundation model that compresses a calendar year of Sentinel-2 optical imagery (such as colour and infrared) and Sentinel-1 radar into a 128-dimensional embedding for every 10m pixel on Earth.
Since similar surfaces tend to have similar embeddings, nearest-neighbour search can find objects (such as airport runways) from a single example.
1 Project outline
This project will build and evaluate a query-by-example search index over Tessera embeddings. Given an example set of pixels, such as a port, the index returns the most similar set of pixels in a region.
The difficulty is the scale of the embeddings. The UK alone has around 3 billion pixels, and so a brute-force search is too slow for interactive use. The project will therefore explore two questions:
- Speed: how fast can query-by-example search be made with an inverted-file index?
- Size: how small can the index become through quantisation and truncation without losing too much quality?