MoRAX adds human mobility data to geospatial AI, boosting predictions
A lightweight framework that helps geospatial models understand human activity, zero-shot.
Geospatial Foundation Models (GFMs) excel at learning visual and physical representations from Earth observation data, but they largely ignore human activity. Human mobility data reveals how regions functionally connect—information missing from satellite imagery. However, mobility data is usually only available for specific cities, limiting its use in transferable urban modeling. MoRAX, developed by Ya Wen and colleagues, solves this by augmenting geospatial embeddings with functional structure derived from mobility, while preserving the GFM's coverage and consistency.
MoRAX uses a teacher-student setup: the teacher observes mobility flows during training and learns enhanced representations, while the student never sees mobility data but mimics the teacher's outputs. In experiments across four cities in two countries, the teacher consistently outperformed GFMs and strong baselines across eight tasks, including socioeconomic and environmental predictions. The student approach performance on most tasks without needing mobility input. Cross-country transfer results confirm that mobility-conditioned modulation provides a general mechanism for grounding geospatial foundation models in the human dimension of cities, enabling zero-shot deployment in unseen locations.
- MoRAX teacher model outperforms standard GFMs in 8 prediction tasks across 4 cities in 2 countries
- Student model, trained without mobility data, approaches teacher performance on most tasks
- Enables zero-shot urban representation learning in cities with no available mobility data
Why It Matters
Bridges the gap between satellite data and human activity, making urban AI more accurate and transferable.