The role of spatial scales in assessing urban mobility models

AI summary

Overview

Research area: Urban mobility modelling and spatial analysis (physics.soc-ph), specifically the comparative evaluation of trip-flow models and the modifiable areal unit problem (MAUP) in a transport-planning context.

Technical level: Intermediate. The models themselves are described with accessible equations, but the paper assumes some familiarity with trip distribution modelling, R²-based evaluation, and geographic aggregation concepts such as Voronoi tessellation.

Scope: A systematic comparison of the gravity, radiation, and visitation models for public transport flows in Singapore, evaluated across spatial scales built from both administrative boundaries and data-driven distance-based clustering of transport nodes.

What This Paper Is About

Urban mobility models are usually judged on how well they reproduce observed travel flows, but the spatial units used for that comparison are often taken for granted. This paper asks how much model rankings and accuracy depend on the scale and shape of the spatial units used. Using Singapore public transport trip data, the authors test three well-known models across many aggregation levels and contrast planning-authority boundaries with spatial units derived from clustering transport nodes by distance.

Key Contributions

1. A comprehensive side-by-side comparison of the gravity, radiation, and visitation models using public transport flow data at transport-node level, extending prior work that typically examined a single model or limited datasets.
2. An explicit, systematic examination of spatial scale, providing empirical evidence that model performance is scale-dependent and that an intermediate scale exists where all models perform best.
3. A direct contrast between conventional administrative boundaries (subzone, planning area, region) and distance-based spatial units, showing the administrative units underperform relative to data-driven clustering.
4. A demonstration that model performance across scales can serve as a diagnostic of urban structure, revealing functional mobility zones and a performance dip around the 3,700–3,900 m scale.

Main Findings

• Visitation model leads overall: Across most spatial scales the visitation law achieves the best performance, followed by the gravity model and then the radiation model, consistent with earlier literature.
• Fine scales are noisy: At fine spatial scales both gravity and radiation perform poorly because of the large number of origin–destination pairs with small areas and populations susceptible to fluctuation; the visitation model retains relatively strong predictive power there.
• Performance improves with aggregation: All three models improve as spatial scale increases, indicating that moderate aggregation reduces noise while preserving meaningful mobility patterns.
• A common optimal scale around 3,000 m: All three models reach peak performance at an intermediate scale of approximately 3,000 m, which balances granularity against generalisation.
• Convergence at optimum: At their respective optimal scales the three models converge to similar adjusted R² values, so differences between them narrow considerably.
• Decline from over-aggregation: At larger spatial scales, performance declines for all models because over-aggregation obscures variability in travel behaviour between locations.
• A window where visitation suffers most: There are specific spatial windows where all models perform poorly, and within these the visitation law performs worse than the gravity and radiation models.
• A notable dip at 3,700–3,900 m: The explanatory power of all models drops noticeably around this threshold, suggesting clusters of transport nodes at roughly 4 km span multiple functional zones and mix flows from distinct mobility systems.
• Planning area is the best administrative level: Among subzone, planning area, and region, the planning area generally yields higher explanatory power, likely because subzones are too granular and regions too coarse.
• Administrative boundaries still underperform: Even at its best level (planning areas), all models underperform distance-based aggregation. At roughly 600 m (the scale closest to planning areas) and roughly 4,400 m (the scale closest to regions), models perform better than under the corresponding administrative boundaries.
• Functional regions emerge: Areas such as Jurong East and Bukit Batok in the west, and Woodlands, Yishun, and Sembawang in the north, appear as coherent mobility clusters that are not simply administrative units.

Methodology in Plain English

The researchers used Singapore public transport trip data recorded at the level of individual transport nodes — bus stops and train stations — giving trip counts between node pairs. They then built two families of spatial units to aggregate trips into zones.

The first family uses the Urban Redevelopment Authority's hierarchical planning boundaries: five planning regions, 55 planning areas, and subzones.

The second family is data-driven. The city is divided into Voronoi polygons, each tied to a single transport node, so that every point in the city is assigned to its nearest node. Nodes are then grouped into clusters using a distance threshold: small thresholds produce many small clusters, large thresholds produce fewer, larger clusters. The Voronoi cells of nodes in a cluster are merged to form the spatial unit for that cluster. The threshold distance therefore acts as a dial controlling spatial scale.

Three trip-prediction models are then applied at each configuration: the gravity model (flows proportional to origin and destination "masses" and decaying with distance, with calibrated parameters), the radiation model (a parameter-free model based on intervening opportunities, requiring only population distribution), and the visitation model (a scaling law relating visitor numbers to the product of travel distance and visit frequency, extended to aggregate origin–destination flows using population density).

Performance is measured with adjusted R², which penalises model complexity; only the gravity model is affected by this adjustment because the radiation and visitation models are parameter-free. For each spatial configuration, the data are split 50/50 into training and testing portions, and the model is run 100 times with randomised splits; the mean and standard deviation of adjusted R² over those 100 runs are reported.

Why This Matters

Impact on research: The paper shows that comparing mobility models without controlling for spatial scale can misjudge their relative strengths, since rankings and accuracy shift with aggregation. It also connects mobility modelling directly to the modifiable areal unit problem, arguing that spatial scale should be treated as a substantive analytical choice rather than a technical detail. Model performance curves themselves become a diagnostic tool for uncovering urban spatial organisation.

Real-world applications:

• Transport planning: Identifying mobility-defined functional clusters (for example Jurong East and Bukit Batok, or Woodlands, Yishun, and Sembawang) can guide where transit connectivity investments would have the greatest impact, rather than following administrative borders.
• Land-use planning: Recognising mobility-defined zones helps distribute housing, employment centres, and amenities in line with how residents already move through the city.
• Model selection and validation: Planners and analysts can choose an appropriate aggregation level — around 3,000 m in this Singapore case — before judging which mobility model to deploy.
• Resource allocation across regions: Understanding which areas form a coherent mobility system supports coordinated sharing of services and infrastructure across statutory boundaries.

Industry relevance: Transport consultancies, urban analytics firms, ride-hailing and mobility platform operators, and public agencies that build demand forecasts all rely on spatial aggregation choices. The finding that data-driven units outperform administrative ones is directly actionable for anyone calibrating trip distribution models on public transport or digital trace data.

Future Directions

• Replacing population as a proxy for activity with employment density or land-use mix, since areas such as Singapore's central business district have low residential population but generate disproportionately high traffic flows.
• Replacing Euclidean distance with network distance or travel time, which would incorporate congestion, service frequency, and connectivity and better match how commuters perceive mobility.
• Incorporating additional determinants of mobility such as land-use patterns, employment distribution, and structural properties of the transport network to improve explanatory power.
• Investigating the spatial window where all models perform poorly, particularly the dip around 3,700–3,900 m, to understand the deeper urban organisational structure constraining mobility dynamics.

Target Audience

Researchers in urban mobility modelling, transport geography, and computational social science who compare trip distribution models or work with spatially aggregated mobility data will benefit most. The paper is also valuable for transport and urban planners, GIS analysts, and data scientists in public agencies or consultancies who must choose spatial units before building or validating demand models. Readers interested in the modifiable areal unit problem and in using model performance as a diagnostic of urban structure will find the scale-by-scale comparison particularly relevant.

Authors’ abstract

Urban mobility models are essential tools for understanding and forecasting how people and goods move within cities, which is vital for transportation planning. The spatial scale at which urban mobility is analysed is a crucial determinant of the insights gained from any model as it can affect models' performance. It is, therefore, important that urban mobility models should be assessed at appropriate spatial scales to reflect the underlying dynamics. In this study, we systematically evaluate the performance of three popular urban mobility models, namely gravity, radiation, and visitation models across spatial scales. The results show that while the visitation model consistently performs better than its gravity and radiation counterparts, their performance does not differ much when being assessed at some appropriate spatial scale common to all of them. Interestingly, at scales where all models perform badly, the visitation model suffers the most. Furthermore, results based on the conventional admin boundary may not perform so well as compared to distance-based clustering. The cross examination of urban mobility models across spatial scales also reveals the spatial organisation of the urban structure.