RANGE 🌎 — Multi-Resolution Geo-Embeddings for Any Location

Retrieval Augmented Neural Fields for Multi-Resolution Geo-Embeddings (CVPR 2025)

📄 Paper (arXiv:2502.19781) · 💻 Code (github.com/mvrl/RANGE) · 🤗 Models & Database (HuggingFace) · 🔬 MVRL Lab

RANGE turns geographic coordinates (longitude, latitude) into rich, general-purpose embedding vectors that you can drop into any downstream model — species distribution, climate, land cover, population, housing prices, or any task where where matters.

How it works

Location encoders like SatCLIP learn a smooth neural field over the globe, which captures low-frequency spatial patterns but washes out fine detail. RANGE is a retrieval-augmented location encoder: for a query location, it estimates the visual features a satellite image at that location would have, by retrieving from a pre-built database of (location, SatCLIP embedding, image embedding) triplets:

• RANGE retrieves database entries whose semantic (SatCLIP) embeddings are similar to the query's, and aggregates their high-resolution image embeddings.
• RANGE+ additionally retrieves by geographic proximity on the sphere and blends the two estimates with a weight β — moving β trades off between smooth, low-frequency embeddings (β→0) and sharp, high-frequency ones (β→1), giving you multi-scale control at inference time.

The final embedding concatenates this estimated high-resolution visual feature (1024-d) with the smooth SatCLIP location feature (256-d) into a 1280-d vector. Across 7 downstream benchmarks (biome, ecoregion, country, temperature, elevation, population, housing), RANGE outperforms state-of-the-art location encoders such as SatCLIP, GeoCLIP, CSP, and SINR — see the paper for full tables.

Installation

pip install rangegeo

Dependencies are minimal: torch, numpy, einops, huggingface_hub.

Usage

from rangegeo import RANGE

# downloads the SatCLIP backbone + RANGE database from HuggingFace on first use
model = RANGE("RANGE+", db="large", beta=0.5)   # or "RANGE"

# coordinates are (longitude, latitude) in degrees
embeddings = model.encode([
    [-90.19, 38.63],   # St. Louis
    [ 85.32, 27.72],   # Kathmandu
])
print(embeddings.shape)  # (2, 1280) numpy float32

That's it — encode() accepts any (N, 2) array-like of (lon, lat) degrees, batches internally (default batch size 10000), and runs on GPU automatically when available.

Options:

The model is a regular torch.nn.Module: inside a torch pipeline you can call model(coords_tensor) directly with an (N, 2) double tensor of (lon, lat).

For training code, evaluation benchmarks, database generation (generate_db.py), and baseline encoders, see the GitHub repository.

Citation

@article{dhakal2025range,
  title={RANGE: Retrieval Augmented Neural Fields for Multi-Resolution Geo-Embeddings},
  author={Dhakal, Aayush and Sastry, Srikumar and Khanal, Subash and Ahmad, Adeel and Xing, Eric and Jacobs, Nathan},
  booktitle={Computer Vision and Pattern Recognition},
  year={2025},
  organization={IEEE/CVF}
}

License

MIT — © the RANGE authors (MVRL, Washington University in St. Louis).