README.md

---
title: README
emoji: 🌍
colorFrom: green
colorTo: blue
sdk: static
pinned: false
---

<p align="center">
  <img src="https://huggingface.co/spaces/geotessera/README/resolve/main/tessera-v2-hero-lite.gif"
       alt="TESSERA — pixel-wise Earth observation foundation model" width="100%">
</p>

<h1 align="center">TESSERA</h1>

<p align="center">
  <b>A pixel-wise Earth observation foundation model</b><br>
  from the University of Cambridge · <a href="https://github.com/ucam-eo">ucam-eo</a>
</p>

<p align="center">
  <a href="https://geotessera.org"><img src="https://img.shields.io/badge/Website-geotessera.org-1f6feb?style=flat-square"></a>
  <a href="https://arxiv.org/abs/2607.03949"><img src="https://img.shields.io/badge/Preprint-TESSERA%20v2-b31b1b?style=flat-square"></a>
  <a href="https://github.com/ucam-eo/tessera"><img src="https://img.shields.io/badge/Code-ucam--eo%2Ftessera-181717?style=flat-square&logo=github"></a>
  <a href="https://pypi.org/project/geotessera/"><img src="https://img.shields.io/pypi/v/geotessera?style=flat-square&label=pip%20install%20geotessera&color=3775A9"></a>
</p>

---

**TESSERA** turns raw satellite time series into ready-to-use **embeddings**. For every
**10 m** pixel on Earth it reads a full year of **Sentinel-1 (SAR)** and
**Sentinel-2 (optical)** observations and compresses them into a compact,
task-agnostic vector — a learned summary of the land's spectral and temporal
behaviour. Trained self-supervised on billions of pixels, with **no labels
required**, these embeddings work as drop-in features for land cover and crop
mapping, biomass and carbon estimation, change detection, and more, and are far
cheaper to store and serve than raw imagery.

| | |
|---|---|
| 🛰️ **Inputs** | Sentinel-1 + Sentinel-2, one full year per pixel |
| 📐 **Resolution** | 10 m, global, annual embeddings for **2017–2025** |
| 🧩 **Output** | dense per-pixel embeddings — v2 adds nested **Matryoshka** vectors (use 16 / 32 / 64 / 128 dims) |
| 🧠 **Training** | self-supervised (Barlow Twins), no labels |

## Models

| Release | Description | Weights |
|---|---|---|
| **TESSERA v2** *(latest)* | Four compact pixel students — **Nano 1.07M · Small 7.11M · Medium 21.03M · Large 43.83M** — distilled from a **2B teacher**, with 128-d Matryoshka output. | [N](https://huggingface.co/geotessera/TESSERA-V-2.0-2B-N) · [S](https://huggingface.co/geotessera/TESSERA-V-2.0-2B-S) · [M](https://huggingface.co/geotessera/TESSERA-V-2.0-2B-M) · [L](https://huggingface.co/geotessera/TESSERA-V-2.0-2B-L) · [Teacher](https://huggingface.co/geotessera/TESSERA-V-2.0-2B-Teacher) |
| **TESSERA v1.1** | Wider QAT encoder, all-observation inference; MPC & AWS checkpoints (int8). | [TESSERA-V-1.1](https://huggingface.co/geotessera/TESSERA-V-1.1) |
| **TESSERA v1.0** | The original release (QAT / int8 and early fp32). | [TESSERA-V-1.0](https://huggingface.co/geotessera/TESSERA-V-1.0) |

👉 Everything in one place: the [**TESSERA collection**](https://huggingface.co/collections/geotessera/tessera-6a108c1a6618e155f551c065).

## Get started

**Just want embeddings?** Skip the model entirely — download ready-made global
embeddings with the [`geotessera`](https://github.com/ucam-eo/geotessera) Python library:

```bash
pip install geotessera
```

...or [request coverage](https://github.com/ucam-eo/geotessera#request-missing-embeddings)
for your region. New v2 embeddings can be [pre-requested here](https://github.com/ucam-eo/geotessera/issues/new?template=v2-embedding-prerequest.yml&labels=v2-embedding-prerequest).

**Want to run the model yourself?** Generate embeddings from your own Sentinel-1/2
tiles with the code and instructions in [`ucam-eo/tessera`](https://github.com/ucam-eo/tessera).

## Learn more

- 🌐 Website — https://geotessera.org
- 📦 Model & inference code — https://github.com/ucam-eo/tessera
- 🐍 Embeddings library — https://github.com/ucam-eo/geotessera
- 📄 Papers — [TESSERA v2: Scaling Pixel-wise Earth Foundation Models](https://arxiv.org/abs/2607.03949) · [TESSERA (v1)](https://arxiv.org/abs/2506.20380)

<sub>Model weights are released under CC0-1.0; the `geotessera` library under MIT. Use is additionally governed by the TESSERA <a href="https://github.com/ucam-eo/tessera/blob/master/aup.md">Acceptable Use Policy</a>.</sub>