Skip to content
Deep dive · TESSERA embeddings

How a satellite learns what a forest is.

Every 10-metre square of land now has a short fingerprint made of numbers. It sums up how that patch of ground behaved across a whole year. Here is how the fingerprints are made, why they are useful, and where to get them for free.

Start the story
The Sacramento–San Joaquin Delta drawn from TESSERA embeddings: rivers in deep teal, farm plots in greens, reds and pinks.
Sacramento–San Joaquin Delta, 2025 · TESSERA embedding visualisation
01

A year, in 128 numbers.

TESSERA is an AI model that reads a full year of satellite passes and condenses every patch of land into a fingerprint. Six steps, no maths required.

Start with one place.

This is farmland in Kansas, seen from orbit. Each green circle is a field watered by a rotating sprinkler arm.

One picture shows what the fields looked like on one day. It can't tell you what they do over a year.

Image
Landsat · NASA/USGS

Now watch it for a whole year.

Europe's Sentinel satellites pass over the same spot every few days. Sentinel-2 takes optical pictures, like a camera. Sentinel-1 sends radar pulses that see through cloud and darkness.

Stack a year of passes and the place shows its rhythm: when fields green up, when they're harvested, when the ground lies bare or wet.

Inputs
Sentinel-2 optical · Sentinel-1 radar

Squeeze the year into 128 numbers.

TESSERA reads the whole stack, pixel by pixel, and boils each 10 × 10 metre square down to a list of 128 numbers. That list is called an embedding.

Think of it as a fingerprint for how that patch of ground behaves over a year. No single number means "wheat" or "forest". Together, they capture the pattern.

We draw the numbers as coloured stripes. The values on this page are illustrative, not real model output.

Similar places get similar fingerprints.

A rainforest in Brazil and one in Borneo are about 19,000 km apart, almost on opposite sides of the planet. Their fingerprints still look alike. A crop field or a city block looks completely different.

That is what makes embeddings useful: you can compare places by their numbers alone, before anyone has labelled them.

Ask the map: where else looks like this?

Pick one spot and look for every pixel with a similar fingerprint. The matches light up in pink. Tap the map to try another spot, and use the slider to make the match stricter or looser.

This demo compares the colours in the picture, a simplified stand-in for the full 128 numbers.

Now do it for every pixel on Earth.

There is one embedding for every 10 m pixel of land on Earth, for each year from 2017 to 2025, with a new year added annually.

These pictures show real 2025 embeddings, turned into colour. Rivers, fields, cities and canyons stand out on their own, without anyone drawing them in.

02

Open, for anyone.

The embeddings are free to download, and so is the code that makes them.

Per pixel

128

numbers in each fingerprint

Resolution

10 m

one fingerprint for every 10 × 10 m square of land

Coverage

2017–25

one embedding per year, every land pixel worldwide

Store

1.6 PB

in one cloud-native Icechunk store

The model

Built at Cambridge, with Arbol.

TESSERA was developed by researchers at the University of Cambridge together with Arbol's Chief Scientist, Dr Clement Atzberger. The paper is peer-reviewed and appears at CVPR 2026.

Read the paper
The data

Free on AWS.

The embeddings sit on the AWS Registry of Open Data through the Amazon Sustainability Data Initiative. They are free to use, with no download fees.

Open the registry
The pipeline

Re-engineered by dClimate.

We rebuilt the inference pipeline as an open-source, cloud-native library. A year of embeddings for an area the size of a US state went from days to under an hour.

How we did it
03

See the planet the way the model does.

CYCLOPS, our land management app, reads farms and forests with these embeddings. You can also pull them for your own region or read the research behind them.

Explore the data Read the paper Get the code Talk to the team