Skip to content

community

TESSERA And Why It Matters For Earth Observation

TESSERA is one of the most interesting pieces of open Earth observation infrastructure available today.

For the CoRE Stack community, it matters because it pushes in the same direction that this project cares about deeply: turning complex geospatial computation into reusable public infrastructure that more people can actually work with.

TESSERA is not just another model benchmark. It is an attempt to make Earth observation embeddings usable as a shared data product.

What TESSERA Is

TESSERA is a foundation model for Earth observation time series built around Sentinel-1 and Sentinel-2 data. Instead of treating each image as an isolated scene, it learns from the temporal behavior of each location across time.

That matters because land, water, vegetation, and human activity are not static. They change with:

  • seasons
  • rainfall
  • cultivation cycles
  • cloud cover
  • disturbance
  • recovery

So the model is designed to produce embeddings that represent how a place behaves through time, not only how it looks in one clean image.

The result is a compact representation that can later support tasks such as:

  • land-use and crop analysis
  • forest and vegetation monitoring
  • change detection
  • segmentation
  • biomass and canopy estimation
  • other downstream EO tasks with relatively small task-specific heads

The Problem TESSERA Was Built To Solve

A lot of Earth observation pipelines quietly depend on idealized data:

  • cloud-free images
  • well-behaved seasonal composites
  • dense observation sequences
  • heavily preprocessed inputs

But operational remote sensing is rarely that neat.

Real time series are messy. They are affected by:

  • cloud obstruction
  • missing observations
  • irregular revisit cycles
  • different sensor behavior
  • seasonal sparsity in usable optical data

TESSERA was built to handle that mess more honestly.

The core idea is simple: instead of throwing away imperfect observations and flattening time into a composite, the model learns to build stable representations from sparse, irregular, real-world time series.

That design choice is one of the most important reasons the project is meaningful.

How TESSERA Was Built

At a high level, TESSERA was built around a pixel-wise temporal learning setup.

1. The basic unit is a time series for a location

For a given spatial location, TESSERA gathers the available multi-temporal observations from:

  • Sentinel-2 optical bands
  • Sentinel-1 radar backscatter

This makes the representation multimodal from the start:

  • optical data helps with surface reflectance, vegetation, and land cover signals
  • radar helps where optical observations are limited, especially under cloud-heavy conditions

2. It learns from time, not only from static patches

TESSERA is built to learn from the sequence of valid observations across the year, with explicit attention to temporal position.

In practice, this means the model is not just asking:

“What does this place look like?”

It is asking:

“How does this place behave over time, given the observations we actually have?”

3. It uses separate encoders for the two modalities

The model processes Sentinel-1 and Sentinel-2 through separate branches before fusing them.

That is a sensible design because radar and optical measurements are not the same kind of signal. Letting each modality develop its own temporal representation before fusion helps preserve their strengths.

4. It is trained to be invariant to which valid observations were available

This is one of the key ideas in the system.

During training, TESSERA uses sparse random temporal sampling and a Barlow Twins self-supervised objective so that two different sampled views of the same place should still lead to consistent embeddings.

That means the model is encouraged to learn:

  • the stable underlying process
  • not accidental dependence on one exact set of timestamps

This is exactly the kind of invariance that operational EO systems need.

5. It adds regularization to avoid shortcut learning

The training setup also uses:

  • global shuffling, to reduce overfitting to local spatial neighborhoods
  • mix-based regularization, to make learning more stable under severe sparsity

These are important because Earth observation data has many easy shortcuts:

  • nearby pixels often look similar
  • region-specific patterns can dominate
  • sparsity can make models brittle

The TESSERA design tries to reduce those failure modes directly in training.

6. It produces compact embeddings as a usable data product

One of the strongest ideas in TESSERA is that the output is not only a trained model. It is also a reusable embedding layer that others can retrieve and use.

The released form emphasizes:

  • annual global coverage
  • 10 m resolution
  • pixel-wise embeddings
  • compact int8 storage

That last part matters a lot. If embeddings are supposed to function as public infrastructure, they cannot remain too expensive to store, move, and query.

Why This Is A Big Deal

TESSERA matters because it helps move Earth observation from:

  • expensive bespoke modeling
  • toward reusable shared representations

That shift can change who gets to build with EO data.

Instead of requiring every downstream team to:

  • preprocess raw satellite data
  • solve cloud issues again
  • design time-series models again
  • train large models again

TESSERA makes it more plausible to start from a robust shared representation and focus on the actual application question.

That is a powerful idea for:

  • students
  • researchers
  • nonprofits
  • public-interest technologists
  • geospatial developers

Why It Resonates With CoRE Stack

CoRE Stack and TESSERA are not the same thing, but they share an important instinct.

Both projects care about building common infrastructure instead of only producing one-off answers.

CoRE Stack does this through:

  • standardized geospatial data structures
  • watershed-linked outputs
  • public APIs
  • STAC delivery
  • reusable planning layers

TESSERA does this through:

  • robust temporal embeddings
  • open model weights and code
  • globally usable representation layers
  • practical tooling for retrieval and inference

That is why TESSERA is such a good fit for innovation calls and open collaboration. It gives people something real to build on.

Why The “Embeddings As Data” Idea Matters

This may be the most important conceptual contribution.

Usually, when people hear “foundation model,” they think of:

  • a large pretrained model
  • that must still be wrapped, fine-tuned, hosted, and adapted carefully

TESSERA pushes a stronger public-data idea:

  • embeddings themselves can be distributed as a reusable layer

That changes the workflow.

Instead of starting from raw imagery, many users can start from:

  • a compact, learned representation of place and time

That lowers the entry barrier for downstream experimentation and makes advanced EO analysis more accessible.

Why The Open Release Matters

TESSERA is especially valuable because it was not kept closed behind a paper.

The project has been released with:

  • an open repository
  • model and inference code
  • downstream task code
  • generated embeddings

That makes it much more useful to the broader community.

Open release matters because infrastructure becomes real only when others can actually inspect it, test it, adapt it, and build on it.

What People In This Community Should Take From It

TESSERA is worth paying attention to for at least three reasons.

1. It shows how EO models can be designed around reality, not idealized inputs

Clouds, sparse observations, and irregular time series are not edge cases. They are normal.

2. It shows that representations can be released as common infrastructure

This is the move from “here is our model” to “here is a usable public layer.”

3. It opens room for more application-layer experimentation

People can spend more time on:

  • ecological questions
  • climate questions
  • planning workflows
  • community tools
  • downstream adaptation

and less time rebuilding the same preprocessing and representation stack from scratch.

Closing Thought

Good open geospatial infrastructure does not only answer questions. It makes better questions easier to ask.

TESSERA is important because it helps make advanced Earth observation analysis more reusable, more open, and more realistic about the data conditions people actually work under.