New AI Framework Learns Geography Through Human Movement Patterns

Understanding what a geographic location truly represents—its economic significance, social role, and connection to other places—remains a key challenge in building AI systems that comprehend the world. A new approach shifts focus away from traditional satellite imagery toward the patterns of human movement itself.

Rethinking Location Meaning#

Rather than relying primarily on Earth observation data, researchers behind MoRA argue that human activity and mobility patterns reveal far more about a location's functional purpose and socio-economic character. The framework treats human movement as the foundation for interpreting other data sources—including point-of-interest catalogs, remote sensing imagery, and demographic information—all aligned through the lens of how people actually move between places.

How the System Works#

MoRA combines several machine learning techniques to fuse its diverse inputs:

• Spatial tokenization breaks geographic space into discrete, analyzable units
• Graph neural networks process a massive mobility graph containing over a billion edges connecting more than 100 million points of interest
• Asymmetric contrastive learning ensures that satellite imagery, POI data, and statistical information are interpreted consistently with observed human movement patterns

The result is a compact 128-dimensional representation that encodes a location's socio-economic context and functional relationships.

Validation and Performance#

To demonstrate its effectiveness, the team built a benchmark comprising nine prediction tasks spanning social and economic domains. Across these downstream applications, MoRA outperformed existing state-of-the-art models by an average of 12.9%, suggesting that mobility-grounded representations capture meaningful geographic knowledge.

The work also echoes scaling laws observed in large language models, suggesting that geospatial representation learning improves systematically with larger datasets and model capacity.

The researchers have released code and pretrained models for public use.