Spatiotemporal Heterogeneity of AI-Driven Traffic Flow Patterns and Land Use Interaction: A GeoAI-Based Analysis of Multimodal Urban Mobility

Olaf Yunus Laitinen Imanov

Source-linked AI summary

Spatiotemporal Heterogeneity of AI-Driven Traffic Flow Patterns and Land Use Interaction: A GeoAI-Based Analysis of Multimodal Urban Mobility

Olaf Yunus Laitinen Imanov

arXiv:2603.05581v1Mar 5, 2026cs.LGcs.AIeess.SP

TL;DR

Urban traffic models rarely integrate land-use heterogeneity, multimodal dynamics, and cross-city transferability in one framework. This paper combines MGWR, RF, and ST-GCN to model three mobility modes, achieving RMSE = 0.119 and R^2 = 0.891 for motor vehicle prediction while showing limited cross-morphology generalisability.

Problem
Existing studies rarely unify land-use heterogeneity, AI-driven traffic prediction, multimodal benchmarking, and cross-city transferability within one GeoAI framework.
Method
The study sequentially integrates MGWR-derived spatial features with an RF-GNN ensemble to model traffic flows across motor vehicle, public transit, and active transport.
Results
RMSE = 0.119 and R^2 = 0.891 for motor vehicle prediction, with analogous advantages across all mobility modes.
Takeaways & Limitations
Spatially differentiated land-use and transport interventions are supported, while GeoAI deployment should account for urban morphological context.
Takeaways & Limitations
The empirically calibrated dataset is synthetic rather than drawn directly from operational sensor networks, and six-hour aggregation suppresses sub-hourly dynamics.

Abstract

from arXiv · show

Urban traffic flow is governed by the complex, nonlinear interaction between land use configuration and spatiotemporally heterogeneous mobility demand. Conventional global regression and time-series models cannot simultaneously capture these multi-scale dynamics across multiple travel modes. This study proposes a GeoAI Hybrid analytical framework that sequentially integrates Multiscale Geographically Weighted Regression (MGWR), Random Forest (RF), and Spatio-Temporal Graph Convolutional Networks (ST-GCN) to model the spatiotemporal heterogeneity of traffic flow patterns and their interaction with land use across three mobility modes: motor vehicle, public transit, and active transport. Applying the framework to an empirically calibrated dataset of 350 traffic analysis zones across six cities spanning two contrasting urban morphologies, four key findings emerge: (i) the GeoAI Hybrid achieves a root mean squared error (RMSE) of 0.119 and an R^2 of 0.891, outperforming all benchmarks by 23-62%; (ii) SHAP analysis identifies land use mix as the strongest predictor for motor vehicle flows and transit stop density as the strongest predictor for public transit; (iii) DBSCAN clustering identifies five functionally distinct urban traffic typologies with a silhouette score of 0.71, and GeoAI Hybrid residuals exhibit Moran's I=0.218 (p<0.001), a 72% reduction relative to OLS baselines; and (iv) cross-city transfer experiments reveal moderate within-cluster transferability (R^2>=0.78) and limited cross-cluster generalisability, underscoring the primacy of urban morphological context. The framework offers planners and transportation engineers an interpretable, scalable toolkit for evidence-based multimodal mobility management and land use policy design.

1. Introduction

Urban traffic flow reflects dynamic interactions among land use, heterogeneous travel behaviour, network structure, and real-time data, while spatially varying effects challenge conventional global models. The study responds with a unified GeoAI framework integrating spatial modelling, machine learning, multimodal benchmarking, and explainability.

Motivation: Urban traffic flow emerges from interactions among land use configurations, sociodemographic heterogeneity, multimodal travel behaviour, and real-time AI-mediated dynamics.
Research problem: Land use effects on travel behaviour are spatially non-stationary, varying with urban morphology, network topology, and socioeconomic context.
Research gaps: Existing research rarely unifies land use heterogeneity with AI-driven prediction or systematically benchmarks spatial and deep-learning models across mobility modes.
Contributions: The GeoAI Hybrid sequentially embeds MGWR-derived spatial coefficient maps as features within an RF-GNN architecture.This design enables local spatial adaptability and global pattern generalisation.
Contributions: The study benchmarks OLS, GWR, MGWR, RF, GNN, and GeoAI Hybrid models across motor vehicle, public transit, and active transport.It also uses SHAP to quantify variable-level contributions from land use, network, and socioeconomic predictors.

2. Literature Review

The literature links traffic heterogeneity and travel behaviour to spatial context, land use, and network structure, while advancing toward graph-based, hybrid, and GeoAI methods. However, no prior study simultaneously integrates six specified dimensions within one unified GeoAI framework.

Traffic heterogeneity: Traffic heterogeneity varies systematically across space and time, while MFD validity depends on spatial homogeneity within a defined perimeter.Event-driven changes, congestion spillovers, and urban context further underscore the networked and context-sensitive character of traffic heterogeneity.
Land use and mobility: The D-variables framework connects land use to travel behaviour through density, diversity, design, destination accessibility, and distance to transit.Evidence indicates compact, mixed-use development reduces vehicle kilometres travelled and promotes transit and active-mode uptake, although effects vary by city size and morphology.
Land use and mobility: LUTI models formalise bidirectional transport-accessibility and land-use change, while newer research develops spatially explicit land-use mix measures.These measures capture dimensions including diversity, accessibility, interuse compatibility, and block-level adjacency.
GeoAI and traffic prediction: Traffic prediction has progressed from feedforward and recurrent networks to graph-based spatiotemporal models that exploit road-network topology.Hybrid statistical-neural models seek a balance between accuracy and interpretability, with SHAP used to explain prediction attribution.
GeoAI and traffic prediction: GeoAI research increasingly combines spatial heterogeneity, geographic knowledge, foundation models, transfer learning, and multimodal spatiotemporal fusion.The central challenge is incorporating spatial heterogeneity and geographic knowledge into aspatial deep-learning pipelines.
Research gap: No prior study among nine representative works simultaneously addresses all six dimensions—GWR/MGWR, GNN, XAI, land use mix, multimodal prediction, and cross-city transferability—within one unified GeoAI framework.Table 1 positions the present study against these principal gaps.

3. GeoAI Hybrid Framework

The GeoAI Hybrid sequentially combines spatiotemporal feature engineering, MGWR-based local modelling, RF-GNN global learning, and SHAP interpretability. MGWR coefficient maps augment Random Forest inputs, while an ST-GCN models network-linked temporal flows and is combined with RF-Spatial predictions using a validated mixing weight.

Sequential architecture: The framework proceeds through four stages: feature engineering, MGWR local modelling, RF-GNN global pattern learning, and SHAP interpretability.MGWR outputs spatial feature maps that are passed to the RF-GNN stage as auxiliary inputs.
Feature engineering: Each zone–time feature vector combines land use, network topology, sociodemographic, and lagged-flow variables.Spatiotemporal lag features use first-order spatial-neighbourhood averaging.
Local spatial modelling: MGWR assigns each covariate its own optimal bandwidth, producing spatially varying coefficient maps that encode locally calibrated land-use sensitivity.Bandwidths are estimated through back-fitting that minimises corrected AICc.
Global pattern learning: An ST-GCN uses a directed, inverse-travel-time-weighted road graph with L=3 convolutional layers and T′=12 input time steps for one-step-ahead flow prediction.The Random Forest is trained on the original features augmented with stacked MGWR coefficient maps, using five-fold spatial cross-validation.
Hybrid prediction and interpretation: The final prediction averages RF-Spatial and ST-GCN outputs, with the validation-RMSE-minimising mixing weight α*=0.42 consistent across mobility modes.The selected weight indicates a moderate advantage for the GNN component in capturing network topology.

4. Data and Study Context

The study uses 350 traffic analysis zones across six cities with contrasting Turkish and Nordic urban morphologies. Multimodal flows, land use, and network data form a temporally ordered one-year panel for model estimation.

Study geography and sampling: 350 traffic analysis zones span six cities in Turkey and the Nordic cluster, enabling transferability assessment across contrasting urban morphologies.The cities are Istanbul, Ankara, Izmir, Copenhagen, Helsinki, and Oslo.
Multimodal mobility data: 3 mobility-mode datasets combine loop-detector motor-vehicle counts, public-transit APC data, and pedestrian and cyclist counts from surveys and smartphone GPS samples.All series cover 52 weeks, and values are independently min-max normalised to [0, 1] by mode and city.
Land-use and network variables: 6 land-use categories derive from cadastral, OSM, and remote-sensing data, while road-network topology metrics come from routable OSM graphs.The categories are residential, commercial, industrial, institutional, open space, and mixed use; the LUM index is computed at TAZ level.
Panel construction and temporal resolution: 8,760 hourly time steps across 350 spatial units are aggregated into 1,460 six-hourly intervals, yielding 511,000 observations per mode.The aggregation balances temporal resolution against computational feasibility.
Panel construction and temporal resolution: 85 % of the temporally ordered data are used for training, with 7.7 % each reserved for validation and testing to prevent data leakage.Training covers weeks 1–44, validation weeks 45–48, and testing weeks 49–52.

5. Results

Results show strong multimodal spatiotemporal heterogeneity, spatially varying land-use effects, and superior GeoAI Hybrid performance. Clustering, SHAP diagnostics, residual analysis, and transfer experiments further demonstrate interpretable typologies, mode-specific predictors, reduced spatial autocorrelation, and morphology-constrained generalisability.

Spatiotemporal patterns: Evening motor-vehicle peaks exceed morning peaks by 6–12% across zones, while transit has a sharper morning peak and active modes show three weekday peaks.Traffic intensity varies substantially within days and weeks across all three modes.
Spatial heterogeneity: MGWR reveals non-stationary land-use effects, with land-use mix strongest in commercial cores and transit corridors and population-density effects reversing in some edge-city zones.Land-use mix has the narrowest bandwidth (ℎ∗= 0.18), whereas employment accessibility has the broadest (ℎ∗= 0.61).
Model performance: 61.9%, 61.2%, and 58.7% RMSE reductions relative to OLS occur for motor vehicle, public transit, and active modes, respectively, with all GeoAI Hybrid pairwise comparisons significant at p< 0.01.The GeoAI Hybrid maintains MAPE below 8% throughout the diurnal cycle, whereas GWR exceeds 20% during morning transition hours.
Land-use interactions: The active-mode land-use-mix slope is steepest (β̂= 0.82, r= 0.74), while motor-vehicle flows have the shallowest slope (β̂= 0.60, r= 0.61).All three modes show positive associations with land-use mix.
Traffic typologies: k= 5 clusters achieve a silhouette score of 0.71, identifying CBD Peak, Mixed Commercial, Suburban, Residential, and Commercial Periphery typologies.CBD Peak zones exceed three times the citywide average during morning peaks, while Residential zones maintain uniformly low flows.
Interpretability and transferability: Land-use mix leads SHAP importance for motor vehicle (|̄ϕ| = 0.184) and active modes (|̄ϕ| = 0.178), while transit stop density ranks first for public transit (|̄ϕ| = 0.201).The GeoAI Hybrid reaches Moran’s I=0.218, a 72.1% reduction from OLS; within-cluster transfer reaches R^2≥ 0.784 for Turkish cities and R^2≥0.873 for Nordic cities, versus R^2=0.631 cross-cluster.

6. Discussion

The discussion shows that traffic–land use relationships are spatially heterogeneous, supporting the GeoAI Hybrid’s multimodal performance and interpretable policy applications. It also identifies residual autocorrelation, morphology-dependent transferability, and data and temporal limitations that constrain deployment.

Spatial heterogeneity: MGWR reveals substantial spatial heterogeneity: land use mix operates at neighbourhood scale (ℎ∗= 0.18), whereas employment accessibility acts regionally (ℎ∗= 0.61).Global models underestimate land use effects in high-density cores and overestimate them in peripheral zones.
Model interpretation: The GeoAI Hybrid uses MGWR coefficient maps as auxiliary RF-GNN features, with mixing weight 𝛼∗= 0.42 indicating a moderate GNN advantage.Road network topology captures spatial dependencies that coefficient maps do not encode.
Residual dependence: 0.218 is the GeoAI Hybrid residual Moran’s 𝐼, lower than OLS (0.782) and GWR (0.521) baselines but indicating remaining unmodelled spatial processes.All reported values are significant at 𝑝 < 0.001; CAR priors within a Bayesian GNN are proposed for future mitigation.
Policy implications: 0.201 is the public-transit SHAP value for transit stop density, while green space ratio contributes 0.121 for active modes.Feature-importance rankings remain broadly stable across seasons, supporting SHAP-guided intervention prioritisation.
Transferability: 0.784–0.851 are within-cluster transfer 𝑅2 values, compared with 0.631 for cross-cluster transfers, confirming morphology-dependent generalisability.Deployment should transfer from a morphologically similar source city and fine-tune with local data; 4–8 weeks of flow observations may suffice.
Limitations: The study is limited by synthetic rather than operationally sourced data, 6-hourly aggregation, partly crowdsourced OSM land-use classification, and a non-endogenised AI route-guidance feedback loop.Future work should replicate the analysis using publicly available traffic repositories and address finer temporal dynamics.

7. Conclusion

The GeoAI Hybrid framework models multimodal spatiotemporal traffic heterogeneity while revealing spatially varying land use–traffic relationships and interpretable predictors. Its results support differentiated planning and establish transfer limits and future research priorities for GeoAI deployment.

Framework performance: RMSE = 0.119 and 𝑅2 = 0.891 for motor vehicle prediction, with analogous advantages across all modes and strongest gains during diurnal transitions.The framework combines MGWR-derived spatial features with an RF-GNN ensemble.
Spatial heterogeneity: MGWR reveals systematic spatial non-stationarity, with land use mix operating at the narrowest spatial bandwidth.The findings support spatially differentiated land use mix targets rather than uniform planning prescriptions.
Interpretability and transferability: Land use mix dominates motor vehicle and active mode flows, whereas transit stop density dominates public transit, according to SHAP-based interpretability.These rankings provide a data-driven basis for prioritising transport and land use interventions.
Interpretability and transferability: Cross-city transfer experiments demonstrate moderate within-morphology and limited cross-morphology generalisability, establishing a deployment protocol for data-sparse contexts.The conclusion positions urban morphology as a key condition for GeoAI transfer.
Future research: Future research should endogenise AI navigation feedback, develop a Bayesian spatiotemporal model for zone-level uncertainty, and replicate transfer analysis in African, South Asian, and Latin American cities.These settings combine rapid urbanisation and data scarcity, creating an urgent need for GeoAI transfer methods.

CRediT Author Statement

Olaf Yunus Laitinen Imanov contributed across the study’s conceptual, analytical, methodological, technical, visualization, and writing activities.

Olaf Yunus Laitinen Imanov contributed to conceptualization, data curation, formal analysis, methodology, software, visualization, and writing the original draft and revisions.

Declaration of Generative AI and AI-Assisted Technologies in the Writing Process

The author states that no generative AI or AI-assisted technologies were used during preparation of the work and accepts full responsibility for its content.

Declaration of Generative AI and AI-Assisted Technologies in the Writing Process: The author reports no use of generative AI or AI-assisted technologies during the work’s preparation.The author also takes full responsibility for all content of the publication.

Funding

The research received no specific grant from public, commercial, or not-for-profit funding agencies.

Funding: The research received no specific grant from any public, commercial, or not-for-profit funding agency.

Data and Code Availability

The data and analysis code supporting the study’s results are available from the corresponding author upon reasonable request to facilitate replication and reproducibility.

Data and Code Availability: Data and analysis code are available from the corresponding author upon reasonable request.The authors encourage replication requests and support reproducibility of the reported findings.