Tessera AI model offers accessible way to view Earth

Tessera, an advanced artificial intelligence (AI) model, offers high-accuracy datasets that encode what the satellite ‘sees’ of Earth’s surface during the course of a year. This compressed data can be used by the scientific community to generate information-rich maps.

Crucially, the encoded datasets – called ’embeddings’ – use far less data than the pixellated images that are transmitted to Earth from satellites. A variety of applications are supported by the model, from monitoring agricultural crops to measuring areas burnt by fire and forest canopies.

A paper on Tessera was published at the 2026 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), held 3-7 June. The model itself was first launched in 2025 and the paper marks the first fully peer-reviewed announcement of Tessera to the scientific community.

The foundation model – Temporal Embeddings of Surface Spectra for Earth Representation and Analysis, or Tessera for short – was developed by researchers at the University of Cambridge in the UK, alongside global and European partners, including Aalto University in Finland.

Tessera’s processed datasets, or embeddings, offer several specific benefits to the Earth observation community. Because Tessera’s embeddings are pretrained, they capture patterns in the data and changes over time that other methods must learn from scratch. This means that non-AI experts can solve remote sensing problems at a global scale using only a fraction of the labelled data previously required. The embeddings are also lightweight enough to access from a laptop or even a mobile device, making them available to users without computational resources. And as an open-source project, it is freely modifiable, raising near limitless possibilities for using satellite datasets to study the Earth.

According to Nuno Miranda, Mission Manager for Sentinel-1 at the European Space Agency (ESA), this is an innovative and exciting step in the development and use of AI in the field of Earth observation. He said, “Foundation models are the new frontier of AI applied to remote-sensing data. Tessera demonstrates how data from the Sentinel-1 and Sentinel-2 missions can be applied in practice, helping users to analyse and understand the Earth system more efficiently.”

Srinivasan Keshav, professor at the University of Cambridge and co-lead on the Tessera project, noted, “With Tessera, we’ve addressed some of the challenges of working with the very large amounts of data provided by the Copernicus programme. Our embeddings make the data more accessible to users from traditionally unserved communities, especially those from ecology, conservation, plant science and zoology. We’ve also made these available without requiring registration and at no cost, opening the door to many new classes of critical problems.”

What is Tessera?

Tessera processes huge amounts of remote-sensing data from the Copernicus missions, Sentinel-1 and Sentinel-2. It combines two types of data: optical data from Sentinel-2, and advanced radar data, known as synthetic aperture radar (SAR) data, from Sentinel-1. The optical and radar datasets are fused by the foundation model and processed into global embeddings spanning each year from 2017-2025.

So, rather than the data-heavy and pixellated imagery from satellites, Tessera compresses data heavy, cloudy satellite imagery to create an embedding layer of Earth data. It does this at a resolution of 10 m, which is the same as the highest resolution captured by Sentinel-2.

Tessera’s embedding layers are basically compressed Earth observation data with missing values filled in. Each 10-m pixel contains a time series of what happened at that point over the year. This gives researchers a picture of change – rather than how a field, river or mountain looks at any given point in time – in a format that’s searchable.

Tessera is supported by tools that enable users to search and compare Earth imagery in a number of different ways. For example, users can search for geographic regions that are similar to each other and they can look for changes in landscapes. It is also possible to make predictions about vegetation health and urban growth.

Tracking habitat change

A UK-based project involving Tessera is developing ways to evaluate the UK government’s nature protection schemes using satellite data from Sentinels 1 and 2. Researchers used Tessera embeddings to track habitat change on land designated for protection across Cumbria, an area of northern England. The project, a partnership between Tessera, the Endangered Landscapes and Seascape Programme, and other UK partners, could eventually provide the government with a way to measure the effectiveness of investments in farming subsidies and nature conservation.

One of Tessera’s co-leads and a senior researcher on the Cumbria landscape monitoring project, Professor David Coomes, from the University of Cambridge, said, “Monitoring these environmental changes over vast scales is exactly the sort of problem that Tessera was designed to solve.”

How are foundation models changing Earth observation?

Tessera promotes transparency and reproducibility. It is open-source and aligned with FAIR principles (Findable, Accessible, Interoperable, Reusable) – a set of widely adopted guidelines developed by the international research community on the reusability of digital assets.

It offers an open and transparent alternative to systems such as AlphaEarth Foundations, an AI model by Google DeepMind, which also compresses complex satellite data from multiple sources to create embeddings using a closed model.

Moreover, Tessera facilitates access to Copernicus data and offers an efficient way to explore Earth observation data.

According to the University of Cambridge’s Srinivasan Keshav, “The adoption of embeddings represents a paradigm shift. Instead of distributing heavy imagery, models such as Tessera can now provide downstream users with compressed semantic representations of the information of Earth’s surface embedded in the original data.”

ESA partners on models for Earth observation

Several teams are working on foundation models for Earth observation, placing Europe at the forefront of this field. ESA has also pioneered the development of foundation models trained on Earth observation data through its open innovation laboratory, Φ-lab, a hub and catalyst for Earth observation innovation. Two foundation models to recently come out of Φ-lab are Thor, developed by the Norwegian Computing Centre, and TerraMind, developed with IBM Research Europe. Importantly, unlike models such as Tessera or AlphaEarth that aggregate information into long-term or annual embeddings, both Thor and TerraMind focus on learning from individual observations, preserving rich spatial contextual information within single imagery snapshots rather than encapsulating everything into a single compressed representation.

Thor (Transformer-based foundation model for Heterogeneous Observation and Resolution) is a versatile multi-modal foundation model, which combines different types of data and is designed to overcome both the challenges of varied inputs and rigid deployment constraints. While most current foundation models are architecturally rigid, Thor allows the user to adapt the internal resolution and optimise computational performance. It is trained on data from Sentinels 1, 2 and 3. This model was funded and supported through ESA’s Foundation Models for Climate and Society (FM4CS) project.

TerraMind, a foundation model released in April 2025, is also multi-modal and able to answer questions about climate and nature. Rather than focusing solely on downstream tasks, its core innovation lies in learning a unified representation space that aligns multiple sources of geospatial data – including satellite imagery, topography, land use/land cover, elevation and geolocation. This enables cross-modal reasoning and query-based interaction with Earth system data. By jointly embedding these diverse sources, TerraMind moves beyond traditional task-specific models toward a more general-purpose geospatial intelligence framework. It was trained on a dataset of more than nine million globally distributed samples spanning eight complementary data types, including radar from Copernicus Sentinel-1 and optical Sentinel-2 imagery.