ucam-eo/tessera

[CVPR26] TESSERA is a foundation model that can process time-series satellite imagery for applications such as land classification and canopy height prediction. Developed at the University of Cambridge, it enables efficient extraction of temporal patterns from Earth observation

What it solves

TESSERA addresses the challenges of processing voluminous and cloud-corrupted satellite time series data. Traditionally, extracting insights from this data is difficult due to a lack of labeled training datasets and the high computational cost of processing temporal signals. TESSERA provides a way to summarize petabytes of Earth observation data into compact, analysis-ready embeddings that preserve spectral-temporal signals, even when observations are missing due to cloud cover.

How it works

TESSERA is an open foundation model that uses self-supervised learning to create 128-dimensional latent representations (embeddings) of Earth's surface at a 10-meter resolution. It employs a technique based on Barlow Twins, where auto-encoder embeddings from two cloud-free random samples of a satellite time series are forced to align. This process creates an embedding that represents the entire time series, including missing observations. The model processes Sentinel-1 and Sentinel-2 satellite data, stacking them along the time dimension to generate composites for inference.

Who it’s for

It is designed for researchers and practitioners in remote sensing, ecology, and environmental science who need to perform downstream tasks such as habitat mapping, carbon accounting, crop type classification, and environmental change detection without needing massive labeled datasets or extreme local compute for every project.

Highlights

  • Global Coverage: Provides analysis-ready, 10m resolution embeddings for the entire globe for 2024, with historical data for some regions (US and Europe) dating back to 2017.
  • Self-Supervised: Uses self-supervised learning to summarize petabytes of data, reducing the need for labeled training sets.
  • Temporal Preservation: Preserves phenological signals typically lost in conventional approaches, enabling better insights into ecosystem dynamics.
  • Open Access: Software is MIT licensed, and model weights and embeddings are released under CC0 for free commercial and non-commercial use.

Related

  • Project
    ucam-eo/geotessera

    A Python library for accessing and working with Tessera geospatial foundation model embeddings, providing tools to download, export, and visualize 128-channel representation maps of the Earth.

  • Dispatch
    AlphaEarth Foundations: Unified Geospatial Embeddings for Global Mapping

    Google DeepMind has introduced AlphaEarth Foundations, an AI model that integrates petabytes of multimodal Earth observation data into unified digital embeddings to revolutionize global mapping and environmental monitoring.

  • Dispatch
    OlmoEarth Embeddings: Custom Vector Exports for Earth Observation

    Hugging Face and AllenAI have introduced custom embedding exports in OlmoEarth Studio, allowing users to generate compact numerical representations of Earth-observation data for downstream geospatial analysis.

  • Project
    NVlabs/timeloop

    An infrastructure for modeling, mapping, and code-generation for dense and sparse tensor algebra workloads on accelerator architectures.

  • Project
    e-sensing/sits

    SITS is an open‑source R package (with a Python wrapper) for building, regularising, and classifying satellite‑image time‑series data cubes. It connects to many public EO collections, offers a suite of ML/DL models (including GPU‑accelerated CNNs and attention encoders), and provides end‑to‑end tools for training, classification, smoothing, uncertainty estimation, and active learning.