TEE: Tessera Embeddings Explorer

v1.9.0 | User Guide | Docker Hub

A web-based tool for exploring and classifying land cover from Sentinel-2 satellite imagery using Tessera embeddings.

What can TEE do?

• Explore any 5km x 5km area on Earth using 128-dimensional Tessera embeddings (2018-2025)
• Find similar pixels instantly — double-click anywhere to highlight similar locations
• Label habitats using K-means clustering (with a settable seed for reproducible clusters), manual pins, polygon drawing, and standard schemas (UKHab, EUNIS, HOTW)
• Export hand-drawn or classified pixels as GeoJSON / Shapefile / KML polygons, not just points
• Evaluate classifiers (k-NN, Random Forest, XGBoost, MLP, Deep MLP, Spatial MLP, U-Net) for classification or regression, with learning curves or k-fold cross-validation, on ground-truth shapefiles at any scale — with a spatial hold-out, a geographic split for k-fold, a field-grouped split (no within-field leakage), separate train/test years, or a separate held-out test file, all keyed off one random seed, and PNG/CSV export of every panel
• Generate classification / regression maps as GeoTIFFs (native UTM, optional value clamping) for use in GIS, with an in-browser preview
• Compare years side by side to detect land-use change

Privacy by design: Similarity searches and labelling run entirely in your browser. ML evaluation runs on your own compute server. Ground-truth data never leaves your machine.

Quick Start

Hosted version

Open tee.cl.cam.ac.uk to explore existing viewports without an account. To create your own viewports, ask a TEE enroller for an account.

Docker (self-hosted)

docker pull sk818/tee:stable
docker run -d --name tee --restart unless-stopped \
    -p 8001:8001 -v /data:/data -v /data/viewports:/app/viewports \
    sk818/tee:stable

Open http://localhost:8001.

ML Evaluation

Evaluation requires a compute server (tee-compute). See the Compute Server Setup section of the User Guide for full instructions.

# Everything on your laptop (no GPU server needed)
./scripts/deploy-compute.sh --local

# Or offload ML to a GPU server via SSH tunnel
./scripts/deploy-compute.sh gpu-box

Then open http://localhost:8001 and go to Validation > Evaluate.

Documentation

The User Guide covers everything:

• Creating and managing viewports (and the "Add Years" flow / no-data years)
• Similarity search and labelling workflows
• Classification schemas (UKHab, EUNIS, HOTW)
• Auto-labelling with K-means, including the clustering seed
• Compute server setup (local, GPU, all-local modes)
• Validation: learning curves, k-fold cross-validation (mechanics and fold selection), confusion matrices, regression metrics and scatter
• Classification vs regression, the task-type override, and one settable random seed
• Classifier parameters and hyperparameter variants; the Deep MLP, Spatial MLP, and U-Net models
• Spatial train/test splits, spatial k-fold, group-by-field splits, separate train/test years, and a separate held-out test file
• PNG / CSV export of the validation panels (hi-res)
• Exporting labels (points or classified-pixel polygons) and generating classification / regression maps, with projection guidance and value clamping
• Sharing labels with other users
• CLI for headless batch evaluation

Community

Join the TEE discussion channel at eeg.zulipchat.com for help, feedback, and announcements.

License

MIT License — see LICENSE file for details.

Authors

• S. Keshav — Primary development and design
• Claude Opus 4.7 — AI-assisted development

Acknowledgements

Thanks to Julia Jones (Bangor), David Coomes (Cambridge), Anil Madhavapeddy (Cambridge), and Sadiq Jaffer (Cambridge) for their insightful feedback.

Citation

@software{tee2025,
  title={TEE: Tessera Embeddings Explorer},
  author={Keshav, S. and Claude Opus 4.7},
  year={2026},
  url={https://github.com/ucam-eo/TEE}
}