CartoVec: General-Purpose Location Embeddings from Open Map Data

CartoVec maps latitude and longitude to a 64-dimensional location embedding. It learns environmental context from structured Overture Maps attributes, without image inputs or downstream-task labels. Inference requires only coordinates, with no map lookup.

Paper: CartoVec: General-Purpose Location Embeddings from Open Map Data

Model Details

• Author: Furkan Gursoy
• Input: Latitude and longitude in degrees, ordered [latitude, longitude]
• Output: A 64-dimensional, L2-normalized embedding
• Model format: TensorFlow SavedModel, including trained weights and the inference graph
• Coverage: New York, New Jersey, Pennsylvania, Delaware, Maryland, Connecticut, Rhode Island, Massachusetts, Vermont, New Hampshire and the District of Columbia

Usage

Install the dependencies:

pip install "tensorflow==2.16.2" huggingface_hub

Download and load the encoder:

import tensorflow as tf
from huggingface_hub import snapshot_download

model_path = snapshot_download(
    repo_id="gursoyfurkan/CartoVec",
    allow_patterns=["saved_model.pb", "variables/*"],
)

model = tf.saved_model.load(model_path)
encode = model.signatures["serving_default"]

# Coordinates in degrees, ordered [latitude, longitude].
coordinates = tf.constant(
    [
        [40.7128, -74.0060],  # New York City
        [42.3601, -71.0589],  # Boston
    ],
    dtype=tf.float32,
)

embeddings = encode(latlon=coordinates)["embedding"].numpy()
print(embeddings.shape)  # (2, 64)

Each row is an L2-normalized location embedding. The model accepts batches of coordinates and requires no map data at inference.

Tested with TensorFlow 2.16.2.

Training

The encoder was trained on approximately 30.7 million sampled coordinates to predict land, water, land cover, land use, transportation and place attributes.

Intended Use

Use CartoVec embeddings as fixed location features in downstream machine-learning models. Individual coordinates can be embedded directly; embeddings sampled within a geographic area can be averaged to represent that area.

Evaluation

The accompanying study evaluates frozen embeddings on six Census outcomes, Airbnb prices and Philadelphia crime counts using five-fold spatial block cross-validation. CartoVec achieves the highest mean performance on five of six Census targets and the highest average across the eight tasks, with several close comparisons.

Scope and Limitations

This checkpoint covers the study region listed above; it is not a globally trained model. It reflects the map snapshot used for training, and its usefulness depends on map coverage and the downstream task. Applications involving consequential decisions require task-specific validation and fairness assessment.

Citation

If you use CartoVec in your research, please cite:

@misc{gursoy2026cartovec,
  title         = {{CartoVec}: General-Purpose Location Embeddings from Open Map Data},
  author        = {Furkan Gursoy},
  year          = {2026},
  eprint        = {2610.17190},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2610.17190}
}

License

CartoVec model weights are distributed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).

Noncommercial use, modification and redistribution are permitted subject to attribution and the other license terms. Commercial use is not licensed under these terms. Contact Furkan Gursoy to request separate written permission for commercial use.