Tessellating the Earth (TTE)

Learnable Spherical Voronoi Partitions for Location Encoding

_{Daniel Cher · Hamza Iqbal · Eric Xing · Brian Wei · Nathan Jacobs — Washington University in St. Louis · MVRL}

TTE is a location encoder that contrastively aligns a (lat, lon) with Sentinel-2 satellite imagery generating a dense embedding for downstream tasks. It uses a learnable Spherical Voronoi partition of S². A small set of global semantic tokens distills shared visual concepts from the imagery into a compact vocabulary the encoder references at inference, letting geographically distant sites covering similar environments share semantics.

🌍 See it learn: the interactive project page shows the Voronoi sites migrating and the location field forming during training.

Install

pip install -r requirements.txt          # inference only
# or, for training / evaluation:
conda env create -f environment.yaml && conda activate tte

Quick start

import torch
from tte import TTE

model = TTE.from_pretrained("MVRL/TTE").eval()      # ~14 MB, location encoder only

coords = torch.tensor([[37.77, -122.42],               # San Francisco — (lat, lon) in degrees
                       [-3.12,   60.02]])               # Amazon
emb = model.encode(coords)

load_tte_model accepts a HuggingFace repo id or a local checkpoint:

from tte import load_tte_model
model = load_tte_model("path/to/last.ckpt")            # or "MVRL/TTE"

Results

TTE sets a new state of the art among parametric location encoders on a variety of geospatial benchmarks (left) and as a geographic prior for iNaturalist-2018 species classification (right).

Global semantic tokens

_{Global semantic tokens learn coherent visual concepts. For each token: the Sentinel-2 imagery it most attends to (top) and its attention map across the globe (bottom).}

Citation

@inproceedings{cher2026tte,
  title     = {Tessellating the Earth: Learnable Spherical Voronoi Partitions for Location Encoding},
  author    = {Cher, Daniel and Iqbal, Hamza and Xing, Eric and Wei, Brian and Jacobs, Nathan},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

🔍 Additional Links

Check out our lab website for other interesting works on geospatial understanding and mapping:

• Multi-Modal Vision Research Lab (MVRL) - Link
• Related Works from MVRL - Link
• See our other location encoder work - Link

Acknowledgements

Frozen image backbone from SSL4EO-S12; evaluation uses the RANGE benchmark suite. Released under the terms in LICENSE.