MoRA: Mobility as the Backbone for Geospatial Representation Learning at Scale

Details

Geospatial representation learning has split between Earth observation paradigms (physical states) and urban computing (socio-economic context). MoRA argues that a location's functional meaning is better captured through human movement — the mobility graph encodes where people go, in what patterns, and with what regularity. The framework uses a human mobility graph as its core backbone and fuses three auxiliary modalities via asymmetric contrastive learning: 100M+ Points of Interest (POIs), large-scale remote sensing imagery, and structured demographic statistics. All auxiliary modalities are interpreted through the lens of mobility dynamics. Key technical components include spatial tokenization and graph neural networks. The compact 128-dimensional embeddings are evaluated on 9 downstream prediction tasks across social and economic domains (crime, housing prices, foot traffic, and others). MoRA achieves average 12.9% improvement over SOTA models across these tasks. The paper also demonstrates geospatial scaling laws analogous to those observed in language modeling. Code and pretrained models are open-sourced. For enterprise applications in urban planning, real estate analytics, transportation optimization, or retail site selection, MoRA provides richer location context than existing approaches and the scaling result suggests further gains with more data.