tessera-v1.1-mpc-encoder

Tessera v1.1 encoder (MPC preprocessing) — unofficial mirror

This is an unofficial mirror of the Tessera v1.1 QAT encoder-only checkpoint, originally distributed via Google Drive by the Tessera team at the University of Cambridge. Mirrored here to make it auto-downloadable via huggingface_hub (Google Drive has no equivalent to hf_hub_download), for use with torchgeo-bench.

The original authors have stated they plan to publish an official HuggingFace mirror once the Drive links stabilize — once that exists, prefer it over this one.

Provenance

• Original source: ucam-eo/tessera, v1.1 branch, checkpoint table under "v1.1 QAT Model Weight (latest, recommended)"
• Original file: tessera_v1_1_mpc_encoder.pt
• Preprocessing source: Microsoft Planetary Computer (Sentinel-2 L2A + Sentinel-1 RTC)
• License: CC0-1.0, per the Tessera README — "Embeddings and model weights are released under the CC0 license: essentially, they can be freely used for both commercial and non-commercial purposes." Attribution is not legally required but is requested.

Model

Tessera is a per-pixel self-supervised foundation model producing 128-dimensional embeddings from Sentinel-1/Sentinel-2 time series (10m resolution). This checkpoint is the encoder-only inference graph (S2 + S1 transformer backbones

• dim-reducer), trained with Barlow Twins on MPC-preprocessed data.

Citation

If you use this checkpoint, please cite the original paper:

@misc{feng2025tesseratemporalembeddingssurface,
      title={TESSERA: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis},
      author={Zhengpeng Feng and Clement Atzberger and Sadiq Jaffer and Jovana Knezevic and Silja Sormunen and Robin Young and Madeline C Lisaius and Markus Immitzer and David A. Coomes and Anil Madhavapeddy and Andrew Blake and Srinivasan Keshav},
      year={2025},
      eprint={2506.20380},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2506.20380},
}