Research map / Image generation / Autoregressive and latent generation

Satellites Reveal Mobility: A Commuting Origin-destination Flow Generator for Global Cities

NeurIPS 2025 · Image generation › Autoregressive and latent generation · approach: Retrieval and matching with generative features · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GlODGen, which generates commuting OD flows for any city from public satellite imagery (via frozen RemoteCLIP features) plus population data, using a WEDAN graph diffusion model trained on US data, and releases it as a tool.

Problem

Commuting OD flows are costly or privacy-sensitive to obtain, and existing computational generators need hard-to-collect urban data (sociodemographics, POIs, land use), limiting them to data-rich cities.

Approach

Replace hard-to-collect urban features with region-level satellite image embeddings from a frozen vision-language geo-foundation model (RemoteCLIP), concatenated with population, as node conditions for a graph denoising diffusion OD generator (WEDAN) trained on US data and transferred globally.

Key results

• On US urban areas, satellite plus population inputs reach 98.3% of CPC for WEDAN (0.623 vs 0.634) and over 90% for all data-driven baselines.
• Trained on US data and transferred to the UK, GlODGen achieves CPC 0.485 vs 0.362 for the best baseline (GMEL), a 34% relative gain, and RMSE 72.68 vs 149.1.
• Rank-aligned flow curves show generated OD flows correlate with real data in case studies across Beijing, Shanghai, London, Paris, Rio de Janeiro and Senegal.

Measured effects reported in the paper

Surprising findings

• Public satellite imagery plus population data achieves nearly the same performance (98.3% of CPC) as expensive sociodemographic, POI and land-use features for WEDAN on US data.

Limitations

• OD flow generation model is trained primarily on US data, limiting geographic robustness.
• Generates only static commuting flows, not time-dependent mobility.
• Real-world OD reference data is noisy and biased, so evaluation is only approximate; cross-city correlation values are not directly comparable.

Experimental setup

Datasets: US LODES commuting OD data with ACS socioeconomic indicators ONS UK commuting flows Paris OD synthesized from INSEE 2015 census Beijing internet location service OD Shanghai CDR-derived OD Rio de Janeiro CDR-derived OD Senegal D4D CDR OD

Benchmarks: US LODES commuting OD (3,333 urban areas) UK ONS commuting flows (326 LADs) Beijing, Shanghai, London, Paris, Rio de Janeiro, Senegal OD data

Metrics: CPC RMSE NRMSE Spearman rank correlation

Baselines: Gravity Model Radiation Model Random Forest DeepGravity GMEL NetGAN WEDAN

Builds on: WEDAN RemoteCLIP graph denoising diffusion DeepGravity NetGAN

Closest papers by meaning

• SeoulMMOD: A Large-Scale Multimodal Origin-Destination Flow Benchmark (NeurIPS 2026 · similarity 0.88)
• Explainable Origin-Destination Crowd Flow Interpolation via Variational Multi-Modal Recurrent Graph Auto-Encoder (AAAI 2024 · similarity 0.83)
• CityLens: Evaluating Large Vision-Language Models for Urban Socioeconomic Sensing (ICLR 2026 · similarity 0.83)
• UrbanVLP: Multi-Granularity Vision-Language Pretraining for Urban Socioeconomic Indicator Prediction (AAAI 2025 · similarity 0.83)
• GeoRouter: Dynamic Paradigm Routing for Worldwide Image Geolocalization (NeurIPS 2026 · similarity 0.83)
• Diffusion Transformers as Open-World Spatiotemporal Foundation Models (NeurIPS 2025 · similarity 0.83)