Skip to content
MoRA: What Human Mobility Reveals Beyond Satellite Imagery for Urban AI

MoRA: What Human Mobility Reveals Beyond Satellite Imagery for Urban AI

The ICLR 2026 paper MoRA combines human mobility networks from transaction data with POI, satellite imagery, and demographics to learn 128-dimensional region embeddings, achieving top R² on 8 of 9 socioeconomic prediction tasks across China, showing mobility links provide information beyond geographic proximity.

Data Party THU
Data Party THU
Data Party THU

Introduction

Geospatial representation learning aims to generate reusable digital features for regions to support diverse prediction tasks. Satellite imagery captures buildings and roads, but human activity links between places add a complementary signal about a region's socioeconomic role. The ICLR 2026 paper MoRA: Mobility as the Backbone for Geospatial Representation Learning at Scale investigates whether integrating inter-location human mobility with points of interest (POIs), satellite imagery, and demographic data improves region-level representations.

The work is a collaboration among Hong Kong University, Tencent WeChat, and City University of Hong Kong. Ya Wen and Jixuan Cai are co-first authors; Yulun Zhou is corresponding author. The paper, code, and a distilled model are publicly available ( https://github.com/ylzhouchris/MoRA).

Methodology: Three-Stage Alignment

MoRA produces a 128-dimensional vector for each region through three main steps:

Mobility network encoding: Offline stores are mapped to H3 grid cells (average area ~36 km²). A mobility graph is built from shared transaction activity between grids over 54 weeks, yielding ~200,000 nodes. The graph is pre-encoded with LINE and then processed by a LightGCN graph neural network to aggregate information from connected regions.

Auxiliary modality encoding: POI names and categories are encoded with the BGE-m3 text model (over 100 million POIs from Tencent Map). Satellite imagery is encoded with RemoteCLIP. Demographic statistics are processed by a multilayer perceptron.

Cross-modal contrastive alignment: Inspired by CLIP, a contrastive loss aligns the mobility-derived representation with the three auxiliary modalities for the same region, reinforcing their correspondence.

MoRA method flow
MoRA method flow

Experiments: Nine Socioeconomic Prediction Tasks

The learned region embeddings are fed into downstream regression models (not zero-shot) for nine tasks: population density, average years of schooling, elderly proportion, per-capita housing area, energy consumption, offline consumption, and others. Each model is trained ten times; mean R² and standard deviation are reported.

National-scale results: MoRA achieves the highest average R² on eight of nine tasks. The average relative improvement over the strongest baseline (AlphaEarth) is 12.9% (not 12.9 percentage points). Nighttime light prediction is the sole exception where MoRA does not lead.

Ablation study: Replacing the mobility graph with a geographic-adjacency graph (connecting only neighboring grids) degrades performance, confirming that long-range activity links provide information beyond spatial proximity.

Scaling Analysis: Pre-training Scope vs. Performance

Three nested pre-training scopes — Jiangsu province, East China, and nationwide — are evaluated on Jiangsu test sets. Expanding from Jiangsu to East China yields clear gains; further expansion to nationwide gives diminishing returns and some task-specific drops. The authors caution that this does not establish a universal "geospatial scaling law" nor validate zero-shot transfer to unseen countries or cities.

Practical Considerations and Limitations

Distilled model: A lightweight model approximates MoRA embeddings from latitude/longitude alone, avoiding the need for the four raw data modalities at inference. However, its performance differs from the full model on multiple tasks; users must re-validate on their own data.

Data representativeness: Transaction data reflects platform-specific consumption scenarios and participating populations, not all residents.

Evaluation independence: Some prediction targets share sources or statistical overlap with pre-training inputs; more independent data and stricter cross-region tests are needed.

Privacy: The paper describes anonymization, grid aggregation, and model distillation as mitigation measures, but these do not constitute guarantees against all attacks.

Causality: Predictive performance on aggregate indicators cannot infer individual attributes or prove that human flow causes socioeconomic differences.

Conclusion

MoRA demonstrates a concrete pathway to incorporate human activity networks into geospatial representation learning. It makes a previously speculative question testable: when an AI "recognizes" a place, how much extra information do its connections to other places provide beyond local attributes?

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

Contrastive LearningGraph Neural NetworksUrban ComputingHuman MobilityICLR 2026Geospatial Representation LearningMobility NetworksSocioeconomic Prediction
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.