Multimodal GeoAI can diagnose transport inequalities across cities but current studies are thin, uneven and poorly validated; the field needs equity-aware models, participatory workflows and stronger data governance such as urban data trusts to translate technical advances into inclusive mobility outcomes.

Main Finding

A systematic survey of 18 multimodal GeoAI studies (screened with PRISMA-ScR from 57 candidate publications, 2019–2025) shows that multimodal geospatial data fusion combined with modern AI (especially deep learning) can produce high-resolution diagnostics and forecasts for urban mobility tasks—accessibility mapping, demand forecasting, and origin–destination (OD) flow prediction—with explicit potential to support equity-oriented planning. However, persistent methodological, data-coverage, explainability, validation, and governance gaps limit the ability of current GeoAI work to reliably produce inclusive, equitable, and policy-actionable outcomes. The paper proposes a research roadmap (equity-aware loss functions, adaptive fusion pipelines, human-in-the-loop workflows, and urban data trusts) to close those gaps.

Key Points

• Scope and sample

  • Surveyed 18 multimodal GeoAI papers selected via PRISMA-ScR out of 57 candidates (2019–2025).
  • Primary urban mobility tasks reviewed: accessibility mapping, demand forecasting, OD flow prediction.
  • Focus on inclusive and equity-oriented applications.

• Methodological trends

  • Multimodal fusion approaches grouped into:
    • Data-level fusion (combining raw heterogeneous inputs, e.g., satellite + GPS),
    • Feature-level fusion (extracting and concatenating learned features),
    • Decision-level fusion (ensemble or model-combination outputs).
  • Rising use of deep learning architectures (CNNs, RNNs, graph neural networks).
  • Emerging techniques: knowledge graphs for relational structure, federated learning for privacy-preserving distributed training, and explainable AI (XAI) to make models interpretable for stakeholders.

• Equity and inclusion emphases

  • Several studies explicitly target underserved populations or spatially marginalized neighborhoods.
  • Few studies systematically optimize for equity (most report disparities post-hoc rather than incorporating equity into objective functions).

• Major gaps identified

  • Population coverage: under-representation of non-smartphone users, informal transit users, and low-income neighborhoods in data sources.
  • Multimodal integration: inconsistent or shallow fusion strategies; limited cross-modal uncertainty quantification.
  • Equity optimization: lack of equity-aware loss functions and formal trade-off analysis (efficiency vs. equity).
  • Explainability & validation: insufficient causal validation, external replication, and stakeholder-centered explanations.
  • Governance & ethics: weak data governance, privacy protections, and accountability mechanisms—raising risks for marginalized groups.

• Proposed directions (roadmap)

  • Methodological: equity-aware objectives, adaptive fusion pipelines that weight inputs by coverage/uncertainty, uncertainty-aware and causal models, federated and privacy-preserving architectures.
  • Socio-technical: participatory/human-in-the-loop workflows to validate and contextualize model outputs; urban data trusts to manage access and governance.
  • Evaluation: standardized equity-sensitive metrics, out-of-sample and cross-city validation, and transparent reporting of population coverage.

Data & Methods

• Data modalities commonly fused

  • Satellite and aerial imagery for built environment features (land use, road networks).
  • GPS / mobile phone trajectories for trip patterns and speeds.
  • Transit smart-card / AVL data for public transport usage and schedules.
  • Volunteered geographic information (VGI) and social sensing (e.g., social media) for points of interest and crowd-sourced signals.
  • Administrative data (census, land parcels) for demographic/contextual attributes.

• Typical analytic pipelines

  • Preprocessing: geospatial alignment, tiling, aggregation to grid or zone units, bias correction for uneven sampling.
  • Feature extraction: CNNs for imagery, sequence models for trajectories, graph embeddings for networks.
  • Fusion strategies:
    • Data-level: stack inputs or co-register raw modalities before feature extraction.
    • Feature-level: concatenate or jointly learn modality-specific features via shared layers.
    • Decision-level: ensemble predictions or model stacking with meta-learners.
  • Training & evaluation: supervised learning on labeled outcomes (flows, demand), cross-validation, and standard metrics (RMSE, MAE, R^2), with few incorporating equity metrics (e.g., error disparity across subpopulations).

• Methodological innovations highlighted

  • Graph Neural Networks (GNNs) and spatio-temporal architectures to model networks and flows.
  • Knowledge graphs to encode multimodal relationships (e.g., POI–transport–demography).
  • Federated learning approaches to combine models across jurisdictions without centralizing raw data.
  • Explainable AI tools (saliency maps, feature importance, counterfactual explanations) to interpret predictions for planning decisions.

Implications for AI Economics

• Distributional impacts and welfare

  • GeoAI outputs influence infrastructure investments, pricing, and service allocation—decisions with direct distributional effects. Economists should quantify welfare changes across groups, not only aggregate accuracy gains.
  • Equity-aware loss functions and multimodal fusion can alter allocation of scarce transport resources; understanding efficiency–equity trade-offs is essential.

• Value of data and returns to information

  • Multimodal data fusion changes the marginal value of different data types (e.g., satellite imagery vs. high-frequency GPS). Economic research can estimate returns to acquiring additional modalities or improving coverage of underrepresented populations.
  • Federated learning and data trusts alter data market structures and bargaining power; there are implications for data pricing, contracts between public and private actors, and incentives to share.

• Policy design, regulation, and market failures

  • Biases and coverage gaps create negative externalities (e.g., systematically under-served neighborhoods). Policies (subsidies, regulation, procurement rules) may be needed to correct market-driven misallocation.
  • Data governance mechanisms (urban data trusts, access rules) act as institutional interventions; economists can model their effects on social welfare, innovation incentives, and privacy.

• Measurement and causal inference

  • GeoAI models produce high-resolution counterfactuals that economists can combine with causal methods to evaluate interventions (e.g., congestion pricing, route subsidies, new transit lines).
  • Suggested methodological fusion: combine GeoAI outputs with microdata (household travel surveys, administrative records, synthetic populations) to estimate demand elasticities, distributional incidence, and spillovers.

• Practical economic research directions

  • Quantify the efficiency–equity frontier: simulate how incorporating equity-aware objectives affects system-wide congestion, travel times, and welfare.
  • Compute the value of marginal data: experimental acquisition of modalities (e.g., subsidizing GPS devices in low-income areas) and estimate marginal improvement in predictive accuracy and welfare.
  • Cost-benefit analyses enriched with GeoAI: use high-resolution accessibility and flow predictions to refine benefit estimates for transport projects, explicitly accounting for distributional impacts.
  • Market design for mobility services: use fused GeoAI signals to design pricing/subsidy schemes that balance efficiency and fairness (e.g., demand-responsive pricing with equity constraints).
  • Institutional economics of data sharing: model incentives for private mobility platforms to participate in federated learning or data trusts, and the welfare implications of various governance regimes.

• Evaluation metrics and empirical strategies to adopt

  • Use distributional welfare metrics (e.g., changes in travel time by income quintile, concentration indices) alongside traditional predictive accuracy.
  • Pre-register counterfactual simulations and combine GeoAI with randomized or quasi-experimental designs where possible.
  • Validate across contexts: cross-city and cross-income-group out-of-sample tests to assess generalizability and fairness.

Overall, the surveyed GeoAI advances provide powerful tools for economists interested in urban mobility policy, but realizing their potential requires explicit incorporation of distributional objectives, rigorous validation, and attention to data governance and incentives.