UniGeoCLIP: Unified Geospatial Contrastive Learning

EarthVision Workshop @ CVPR 2026

Authors: Guillaume Astruc · Eduard Trulls · Jan Hosang · Loïc Landrieu · Paul-Edouard Sarlin

LASTIG, Univ Gustave Eiffel · IGN · ENSG · CNES · LIGM, École des Ponts ParisTech · Google Switzerland

---

Overview

UniGeoCLIP is a multimodal contrastive framework that jointly aligns five geospatial modalities — aerial imagery, street-level imagery, Digital Surface Models (DSM), text descriptions, and geographic coordinates — into a single unified embedding space.

This repository releases the location encoder, a key architectural contribution of the paper.

---

Updates

• [Code released] The full training and inference code for UniGeoCLIP is now available at google-research/scaling-geoloc.

---

Location Encoder

The location encoder maps raw latitude/longitude coordinates to a rich D-dimensional embedding by combining multi-scale Random Fourier Features with self-attention.

Key design choices:

• K Fourier projections with increasing spectral bandwidths are computed and treated as tokens
• A transformer with register tokens processes them jointly, enabling cross-scale interactions
• The parameter count is independent of the number of frequency scales

This contrasts with GeoCLIP's design, where each scale is processed by a separate MLP and aggregated by averaging — a design that misses inter-scale dependencies and grows in parameters with K.

Usage

To use the location encoder, import and instantiate the LocationEncoder from location_encoder.py. This module maps raw latitude/longitude coordinates (in degrees) to a fixed-dimensional embedding suitable for downstream models.

Example

import torch
from location_encoder import LocationEncoder

# Example: batch of 3 locations [latitude, longitude] in degrees
locations = torch.tensor([
    [48.8584, 2.2945],   # Paris
    [40.6892, -74.0445], # New York
    [35.6586, 139.7454], # Tokyo
])

# Instantiate the encoder
encoder = LocationEncoder(
    sigma=[2**0, 2**4, 2**8, 2**12],  # frequency scales (default)
    embed_dim=768,                    # embedding dimension (default)
    n_registers=4,                    # number of transformer register tokens
    num_heads=12,                     # transformer heads
    mlp_ratio=4,                      # dimensionality multiplier for MLP layers
    depth=12                          # number of transformer blocks
)

# Forward pass
# Output: [batch_size, embedding_dim] tensor
embeddings = encoder(locations)
print(embeddings.shape)  # torch.Size([3, 768])

• Input tensor shape: [batch_size, 2] ([latitude, longitude] in degrees)
• Output tensor shape: [batch_size, embed_dim] (default: 768)

Notes

• Coordinates are internally projected using the Equal Earth map projection before encoding.
• Parameters can be adjusted for model size/speed tradeoffs.

Pretrained weights

Pretrained weights are not available in this release.

---

Results

UniGeoCLIP's location encoder achieves a mean R² = 57.0 on the PDFM (Population Dynamics Foundation Model) benchmark across 27 downstream regression tasks (health, socio-economic, and environmental indicators), outperforming all contrastive baselines:

Model	Mean R²
SatCLIP	30.1
GeoCLIP	49.8
† GeoCLIP (retrained)	51.6
UniGeoCLIP (ours)	57.0

† retrained on our dataset under identical conditions.

Increasing encoder depth yields consistent gains across all retrieval tasks:

Depth	SV → GPS	Multimodal Ens.	OOD
0 (fixed RFF)	55.0	10.2	13.6
4	73.1	40.7	30.0
8	73.1	44.0	27.8
12	74.4	47.0	29.2

See the paper and project page for full results across all tasks and modalities.

---

Citation

If you use this code, please cite:

@inproceedings{astruc2026unigeoclip,
  title     = {UniGeoCLIP: Unified Geospatial Contrastive Learning},
  author    = {Astruc, Guillaume and Trulls, Eduard and Hosang, Jan
               and Landrieu, Lo{\"i}c and Sarlin, Paul-Edouard},
  booktitle = {EarthVision Workshop, CVPR},
  year      = {2026}
}

---

License

The code in this repository is released under the MIT License.

---

Acknowledgements

This code builds on top of prior work and code from GeoCLIP