Tokenization — breaking data into discrete, learnable units — is how data enters most modern AI models. For geographic data it has received far less attention, and that gap limits how well models can reason about space.
SPOK is a geospatial tokenization approach in which spatial relationships are encoded explicitly in the tokens themselves. Using urban space as the example, it treats parcels delimited by road links as the basic spatial tokens, and uses dynamic location referencing to encode the relative spatial relationships between them. This lets a Transformer-style GeoAI model query and model spatial relationships within its latent representation space.
In the paper’s comparisons on two tasks, explicit spatial tokenization combined with geographic semantic information captured complex spatial and temporal dynamics effectively. The road-network-based tokenization of 2D urban space is presented as a proof of concept, and the paper calls for more attention to geospatial tokenization as a foundation for geospatial foundation models and GeoAI.
Publication#
Published in the International Journal of Geographical Information Science, volume 39, issue 12, pages 2768–2808 (2025). Code and synthetic data supporting the model are available on figshare.
Cite#
@article{huang2025spok,
title = {{SPOK}: tokenizing geographic space for enhanced spatial reasoning in {GeoAI}},
author = {Huang, F. and Lv, J.-R. and Li, G.-L. and Yue, Y.},
journal = {International Journal of Geographical Information Science},
volume = {39},
number = {12},
pages = {2768--2808},
year = {2025},
doi = {10.1080/13658816.2025.2497810}
}