Uncovering latent urban mobility patterns via smart-card and survey data fusion Authors: Khoa D. Vo, Seung Woo Ham, Mousumi Roy, Swapnil Mishra, Prateek Bansal Published: 2026 ABSTRACT Multiday, multimodal, time-dependent origin-destination (TD-OD) flows describe when, where, and how urban travel occurs. However, existing approaches are typically single-mode or rely on dense multimodal observations that are rarely available at scale. We show that multimodal TD-OD flows can be recovered by integrating household travel surveys with smart-card transit data. The proposed framework estimates cross-modal flow ratios from survey data and applies them to time-varying transit flows to recover private-vehicle and walking demand at hourly and day-of-week resolution. Validation against independent datasets in Singapore and Seoul shows strong agreement (common part of commuters > 0.70; R-squared > 0.60). The recovered flows support policy-relevant analyses, showing that transit is most competitive for intermediate distances (11–16 km) and transit-only data can underestimate peak epidemic infections by up to 50%. These findings demonstrate the importance of a scalable data fusion for multimodal mobility analysis in sustainable and resilient urban planning. Introduction Quantifying when, where, and how people travel is central to managing modern cities. Multiday, multimodal, time-dependent origin–destination (TD–OD) flows provide a unified representation of urban mobility. They capture how travel demand evolves across days, varies over the course of the day, and shifts across transport modes. Origin–destination (OD) flows underpin a wide range of urban and transport applications1,2, including congestion pricing3, ride-sourcing and demand prediction4,5, freight systems6, and activity-based modeling7,8. Beyond transport, they are essential for epidemiological modeling9–11, evacuation planning12,13, and environmental exposure and inequality analysis14–16. Together, these applications establish TD–OD flows as core inputs for forecasting system-level dynamics, evaluating policies, and assessing equity in urban systems. Research on OD flow estimation has progressed through several methodological stages. Early work was dominated by gravity models inspired by physical laws17, later extended in various forms18–21. These models establish interpretable relationships between OD flows and measures of attractiveness, such as distance or impedance, while requiring relatively limited data. Entropy-maximization approaches derive gravity-type trip distributions as the most probable OD matrix under marginal and travel-cost constraints22, while generalized least-squares methods reconcile survey-based prior OD matrices with traffic counts to obtain statistically consistent flow estimates23,24. The increasing availability of large-scale passive mobility data25–27 has substantially expanded the methodological toolbox for modeling urban flows. These data sources fall into two broad paradigms. The first comprises point or network observations, including loop detectors and traffic counts28–30, license-plate recognition31, and floating car data32, which provide partial coverage of network flows. OD matrices must therefore be inferred by inverting assignment models or through statistical reconstruction against a prior33–35. The second comprises individual-trajectory and transaction data, including GPS probe trajectories36, probe vehicle data37, internet-of-things (IoT) sensors38, call detail records (CDRs)39, mobile-phone data40, location-based service data5, and transit smart-card transactions41–43, from which OD flows can be directly inferred or aggregated at high spatial and temporal resolution. Building on these data sources, recent studies increasingly adopt deep learning models to estimate or forecast OD flows from historical observations5,30,40–42. Parallel work explores static flow generation using geographic context (e.g., OpenStreetMap and point-of-interest data) to enhance representation learning20,44–47. Recent generative approaches further extend this line by employing graph-based, adversarial, and diffusion models to synthesize dynamic OD flows48–53. When dense historical OD flows are available for all target modes, these approaches achieve strong performance in OD estimation and forecasting. Despite these advances, most existing OD estimation frameworks remain oriented toward single-mode settings with dense historical observations. In practice, they are designed to estimate or forecast OD flows within the same observable mode, such as car OD flows derived from traffic counts or GPS probes29,36, transit OD flows inferred from smart-card records41–43, or active-mode flows estimated from surveys or fitness applications38. As a result, multimodal TD–OD recovery under heterogeneous and incomplete data availability remains comparatively underexplored. More broadly, the challenge of multimodal TD–OD recovery at scale is driven by fundamental limitations of available mobility data rather than the choice of modeling paradigm. Three persistent constraints arise, regardless of whether physics-based18–21, statistical23,24, or deep learning models5,30,40–42 are used. First, while large-scale passive mobility datasets offer extensive spatiotemporal coverage and high temporal resolution, they often lack reliable travel-mode labels at the population scale. This limitation becomes particularly acute when modal shares vary across time of day and day of week54, rendering multimodal OD recovery underdetermined without additional constraints. Second, proxy-based mode inference approaches that rely on trajectory features, such as travel speed, acceleration, or dwell time55,56, are inherently ambiguous, as different transport modes frequently exhibit overlapping spatiotemporal signatures, especially at the city scale and during off-peak periods. Third, passive mobility datasets are subject to systematic representation bias arising from user self-selection and platform-specific coverage57–59. Without explicit calibration, such biases limit the ability of passive mobility data alone to faithfully represent population-level multimodal travel patterns. Here, we show that multimodal TD–OD flows can be recovered under partial observability by integrating household travel surveys with transit smart-card data. We focus on recovery rather than forecasting: instead of inferring TD–OD flows for each mode from observations of that same mode (Fig. 1a), the approach reconstructs flows for unobserved modes by combining cross-modal relationships with a reference mode that has dense coverage (Fig. 1b). Household travel surveys, though sparse and noisy, provide cross-modal structure through mode-share relationships, while smart-card data provide near-census transit flows. The proposed framework estimates stable cross-modal flow ratios from survey data and applies them to time-varying transit flows to recover private-vehicle and walking demand at hourly and day-of-week resolution without direct observations. By separating cross-modal structure from the anchoring of near-census absolute flows, the framework addresses key data limitations, including missing mode labels in passive data, ambiguity in trajectory-based mode inference, and representation bias. The contribution lies in the data-fusion framework, which combines cross-modal information from survey data with marginally consistent anchoring of absolute flows using a reference mode, rather than relying on end-to-end forecasting from mode-specific datasets. We validate the approach at city scale using independent datasets in Singapore and Seoul (0.73 million and 1.46 million OD–time records, respectively), and demonstrate applications through case studies on multiday multi-modal flow recovery, transit competitiveness, and epidemic spread. Together, these results show the importance of the scalable data-fusion approach for multimodal mobility analysis in planning sustainable and resilient cities.Fig. 1Cross-modal data-fusion framework for multimodal time-dependent origin–destination (TD–OD) flow recovery.Conceptual comparison between a conventional mode-specific OD estimation or forecasting and b the proposed cross-modal recovery framework. Conventional approaches estimate TD–OD flows independently for each mode using mode-specific observations. In contrast, the proposed framework separates cross-modal inference from flow anchoring, enabling recovery of flows for transport modes without direct observations. c, d Household travel surveys provide sparse cross-modal trip observations, covering only a subset of OD–time tuples and exhibiting sampling noise. e Cross-modal ratios between target and reference modes are estimated from survey data and generalized using adaptive representation learning, enabling consistent coverage across OD–time tuples. f Passive observations for a reference mode (e.g., smart-card transit data) provide near-census TD–OD flows, defining the absolute flow scale. g Multimodal TD–OD flows are recovered by combining cross-modal ratios with reference-mode flows, producing population-scale flow estimates across all modes and OD–time tuples. Cross-modal data-fusion framework for multimodal time-dependent origin–destination (TD–OD) flow recovery. Conceptual comparison between a conventional mode-specific OD estimation or forecasting and b the proposed cross-modal recovery framework. Conventional approaches estimate TD–OD flows independently for each mode using mode-specific observations. In contrast, the proposed framework separates cross-modal inference from flow anchoring, enabling recovery of flows for transport modes without direct observations. c, d Household travel surveys provide sparse cross-modal trip observations, covering only a subset of OD–time tuples and exhibiting sampling noise. e Cross-modal ratios between target and reference modes are estimated from survey data and generalized using adaptive representation learning, enabling consistent coverage across OD–time tuples. f Passive observations for a reference mode (e.g., smart-card transit data) provide near-census TD–OD flows, defining the absolute flow scale. g Multimodal TD–OD flows are recovered by combining cross-modal ratios with reference-mode flows, producing population-scale flow estimates across all modes and OD–time tuples. Results We recover multimodal TD–OD flows for transport modes that lack direct large-scale passive observations across the full set of OD–time tuples (o, d, t), where o and d denote origin and destination zones and t an hourly time-of-day bin. Consistent with the data-fusion framework introduced in Fig. 1b, recovery is achieved by combining survey-derived cross-modal relationships with smart-card reference flows. Household travel surveys provide cross-modal information in the form of target-to-reference flow ratios between transport modes (e.g., private vehicles or walking relative to transit). These ratios are observed only for a sparse subset of OD–time tuples and are subject to sampling noise (Fig. 1c, d). To extend this cross-modal information across the full spatiotemporal domain, these ratios are stabilized and generalized using adaptive representation learning (ARL) (Methods), yielding estimated cross-modal ratios \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{R}}_{odt}^{ij}$$\end{document}R^odtij for all (o, d, t) (Fig. 1e). In parallel, smart-card transactions provide near-census TD–OD flows for the transit reference mode, denoted by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bar{N}}_{odt}^{\,j}$$\end{document}N¯odtj (Fig. 1f). Combining these components yields recovered target-mode flows (Fig. 1g): 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{N}}_{odt}^{i}={\widehat{R}}_{odt}^{ij}\,{\bar{N}}_{odt}^{\,j}.$$\end{document}N^odti=R^odtijN¯odtj. Multiday TD–OD flows are obtained by applying time-of-day cross-modal ratios to day-specific smart-card observations (e.g., Mon–Wed, Thu–Fri, Sat–Sun). We apply the framework to one week in Singapore (18–24 July 2022) and Seoul (18–24 October 2021), selected to avoid public holidays and to align with the availability of independent validation data, focusing on private-vehicle and walking modes. Smart-card data provide day-specific transit TD–OD flows for these periods, capturing variation in both total demand and time-of-day allocation. In contrast, household travel surveys are collected over extended field periods (Singapore 2022; Seoul 2021) and are not tied to specific calendar days. As a result, survey-derived cross-modal ratios reflect aggregated modal composition rather than day-specific dynamics. We therefore treat these ratios as approximately stable across weekday groups at the time-of-day resolution, while allowing flow magnitude and temporal variation to be governed by smart-card observations. This separation enables survey data to inform cross-modal structure, while passive data determine when and how much people travel. Summary statistics for the input datasets and recovered TD–OD matrices are reported in Supplementary Tables 1 and 2. We evaluate the recovered flows against independent passive datasets using two complementary metrics: the common part of commuters (CPC) and the coefficient of determination (R2) (Methods). In Singapore, validation uses TomTom vehicle-probe data from July 2022, aggregated by weekday groups (Mon–Wed, Thu–Fri, Sat–Sun). In Seoul, validation uses cellular signaling data from October 2021, providing aggregate TD–OD flows across multiple Mondays. Figure 2 summarizes the ARL for estimating cross-modal ratios (see Methods). The approach combines adaptive aggregation-controlled by an aggregation level parameter k-to stabilize sparse survey-derived ratios (Fig. 2c–e), supervised learning to generalize across all OD–time tuples (Fig. 2a), and passive data to anchor absolute flow magnitude (Fig. 2b). The optimal aggregation level k* is selected using an internal robust objective defined over k (Fig. 2f), yielding a well-defined intermediate optimum that balances fidelity to raw survey observations with stability induced by aggregation, with a consistent value of k* = 1 across both cities. This internal criterion closely aligns with external validation (using TomTom data for Singapore and cellular data for Seoul), supporting its use when independent benchmarks are unavailable (Supplementary Table 3). ARL also provides interpretable insights into cross-modal structure. SHAP (SHapley Additive exPlanations) analysis60 (Fig. 2g) identifies flow-related features-particularly TD–OD flows of the reference transit mode and OD distance as dominant predictors, with secondary contributions from temporal and land-use variables (Supplementary Table 4). These results extend classical gravity-based formulations by showing that observed flow structure, when available at near-census scale, provides a strong complementary signal for cross-modal inference (see Supplementary Note for details).Fig. 2Adaptive representation learning (ARL) for cross-modal ratio estimation.a ARL stabilizes and generalizes cross-modal ratios. Sparse survey-derived ratios are progressively smoothed through adaptive aggregation, followed by supervised learning to extend estimates across all origin–destination–time (OD–time) tuples, with the aggregation level selected via an internal robust objective. b Passive observations for the reference mode define the flow anchor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bar{N}}_{odt}^{j}$$\end{document}N¯odtj, providing the absolute scale for multimodal flow recovery. c Survey observations for target and reference modes are used to construct raw cross-modal ratios and aggregated marginal statistics at origin and destination levels, capturing spatial structure under sparse sampling. d Adaptive aggregation combines raw ratios \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{odt}^{ij}$$\end{document}Rodtij with aggregated priors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widetilde{R}}_{odt}^{ij}$$\end{document}R~odtij derived from marginal counts, enabling a smooth transition between local observations and aggregated structure. e Aggregation weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w}_{odt}^{ij}(k)$$\end{document}wodtij(k) increase with effective sample size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{odt}^{ij}$$\end{document}Nodtij and are normalized by the 98th percentile to reduce sensitivity to extreme values, controlling the degree of smoothing across aggregation levels. f The aggregation level is selected by minimizing an internal robust error as a function of parameter k, with an intermediate optimum balancing fidelity to raw survey observations and stability from aggregation. The selected value is consistent with external validation based on the coefficient of determination (R2) in both Singapore and Seoul, using TomTom and cellular data, respectively. g SHapley Additive exPlanations (SHAP)-based analysis identifies flow-based features and OD distance as primary contributors to cross-modal ratios, with additional contributions from temporal and land-use features. Adaptive representation learning (ARL) for cross-modal ratio estimation. a ARL stabilizes and generalizes cross-modal ratios. Sparse survey-derived ratios are progressively smoothed through adaptive aggregation, followed by supervised learning to extend estimates across all origin–destination–time (OD–time) tuples, with the aggregation level selected via an internal robust objective. b Passive observations for the reference mode define the flow anchor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bar{N}}_{odt}^{j}$$\end{document}N¯odtj, providing the absolute scale for multimodal flow recovery. c Survey observations for target and reference modes are used to construct raw cross-modal ratios and aggregated marginal statistics at origin and destination levels, capturing spatial structure under sparse sampling. d Adaptive aggregation combines raw ratios \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{odt}^{ij}$$\end{document}Rodtij with aggregated priors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widetilde{R}}_{odt}^{ij}$$\end{document}R~odtij derived from marginal counts, enabling a smooth transition between local observations and aggregated structure. e Aggregation weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w}_{odt}^{ij}(k)$$\end{document}wodtij(k) increase with effective sample size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{odt}^{ij}$$\end{document}Nodtij and are normalized by the 98th percentile to reduce sensitivity to extreme values, controlling the degree of smoothing across aggregation levels. f The aggregation level is selected by minimizing an internal robust error as a function of parameter k, with an intermediate optimum balancing fidelity to raw survey observations and stability from aggregation. The selected value is consistent with external validation based on the coefficient of determination (R2) in both Singapore and Seoul, using TomTom and cellular data, respectively. g SHapley Additive exPlanations (SHAP)-based analysis identifies flow-based features and OD distance as primary contributors to cross-modal ratios, with additional contributions from temporal and land-use features. City-scale validation with external datasets We validate the recovered TD–OD flows against independent passive benchmarks in each city (see Methods for details of the validation setup). At the OD–time level, log–log scatter plots show strong agreement with external references in both Singapore and Seoul (Fig. 3a,b), with CPC values of 0.72 and 0.71 and corresponding R2 values of 0.62 and 0.67. Deviations are primarily concentrated in low-flow OD–time cells, where survey-derived ratios are least supported (including zero-valued cells), and in a limited number of high-flow intra-zone OD pairs, where aggregation effects and zone-boundary definitions introduce mild dispersion.Fig. 3Validation of recovered multimodal time-dependent origin–destination (TD–OD) flows against independent passive datasets.a Singapore (Mon–Wed): comparison of recovered private-mode TD–OD flows with TomTom vehicle probes at zone–hour resolution. Left: log–log scatter of TD–OD flows, distinguishing intra- and inter-zone pairs, with agreement quantified by the common part of commuters (CPC) and the coefficient of determination (R2). Right: hourly OD totals aggregated across all zone pairs, showing close alignment between recovered flows (line) and TomTom observations (bars). b Seoul (Mon): analogous comparison for recovered all-mode TD–OD flows (combining inferred non-transit modes with observed transit) against cellular signaling data. Left: scatter-based comparison; right: aggregated hourly totals, showing consistency across temporal scales. c, d Singapore spatial validation. Intra-zone flow intensities (c) and inter-zone corridor flows (d) are compared between recovered OD flows (left) and TomTom observations (right), showing consistent spatial patterns in both local intensity and network-level connectivity. e, f Seoul spatial validation. Intra-zone flow intensities (e) and inter-zone corridor flows (f) are compared between recovered OD flows (left) and cellular signaling data (right), showing consistent spatial patterns in both local intensity and inter-zone connectivity. Validation of recovered multimodal time-dependent origin–destination (TD–OD) flows against independent passive datasets. a Singapore (Mon–Wed): comparison of recovered private-mode TD–OD flows with TomTom vehicle probes at zone–hour resolution. Left: log–log scatter of TD–OD flows, distinguishing intra- and inter-zone pairs, with agreement quantified by the common part of commuters (CPC) and the coefficient of determination (R2). Right: hourly OD totals aggregated across all zone pairs, showing close alignment between recovered flows (line) and TomTom observations (bars). b Seoul (Mon): analogous comparison for recovered all-mode TD–OD flows (combining inferred non-transit modes with observed transit) against cellular signaling data. Left: scatter-based comparison; right: aggregated hourly totals, showing consistency across temporal scales. c, d Singapore spatial validation. Intra-zone flow intensities (c) and inter-zone corridor flows (d) are compared between recovered OD flows (left) and TomTom observations (right), showing consistent spatial patterns in both local intensity and network-level connectivity. e, f Seoul spatial validation. Intra-zone flow intensities (e) and inter-zone corridor flows (f) are compared between recovered OD flows (left) and cellular signaling data (right), showing consistent spatial patterns in both local intensity and inter-zone connectivity. Temporal dynamics are accurately reproduced in both cities. Recovered hourly flow profiles closely track external benchmarks (Fig. 3a,b, right panels), capturing the morning peak, midday plateau, and evening peak, with hourly profile R2 values of 0.97 for Singapore and 1.00 for Seoul. Spatially, the recovered maps reproduce both localized flow intensities and major inter-zone corridors. In Singapore, agreement is slightly stronger for inter-zone flows (CPC = 0.75; R2 = 0.69; Fig. 3d) than for intra-zone flows (CPC = 0.69; R2 = 0.65; Fig. 3c). In Seoul, intra-zone flows are recovered particularly well (CPC = 0.81; R2 = 0.72; Fig. 3e), while inter-zone patterns remain well aligned despite the broader and noisier all-mode cellular benchmark (CPC = 0.72; R2 = 0.60; Fig. 3f). These results are interpreted in the context of the external benchmarks. TomTom vehicle probes in Singapore and cellular signaling data in Seoul cover only a subset of the population (approximately 25%–30%) and may exhibit representation bias59. Accordingly, they serve as independent reference datasets to assess the plausibility of spatial and temporal flow structure, rather than definitive ground truth. Differences in performance across intra- and inter-zone flows are consistent with known data coverage characteristics. TomTom probes tend to overrepresent short-distance local circulation trips, which reflect vehicle movements that do not always correspond to distinct activity destinations and are therefore more readily captured by vehicle-based probe data than by activity-oriented travel surveys. In contrast, cellular signaling data provide a broader representation of short-distance and walk-related travel. Additional validations confirm the robustness of the proposed framework across multiple dimensions. Multiday benchmarks for Singapore and time-of-day dynamics at the OD, origin, and destination levels for both cities show consistent agreement with reference data (Supplementary Figs. 2–5). Robustness to survey noise and limited sample size is further supported by progressive downsampling of the household travel survey, with performance remaining stable across sampling fractions (Supplementary Table 5). Importantly, recovery accuracy is largely insensitive to the choice of cross-modal ratio learning model. Comparable performance is observed across polynomial regression, random forest, extreme gradient boosting (XGBoost), and multilayer perceptron architectures (Supplementary Table 6). This consistency indicates that performance is governed primarily by the structural design of the data-fusion framework, rather than by the specific regression model used. To further isolate the role of structural constraints, we evaluate large language model baselines61 under zero-shot and transit-anchored configurations (see Supplementary Note for details). Zero-shot predictions show weak alignment with reference flows in both cities. Incorporating observed transit counts as an hourly calibration signal yields partial improvement, but substantial dispersion persists. In contrast, the proposed framework achieves consistently higher fidelity across the full dynamic range in both cities (Supplementary Fig. 6). These results demonstrate that accurate recovery is not explained by model capacity or contextual priors alone, but by the explicit structural constraints imposed by cross-modal ratio estimation and marginal flow consistency. Finally, we assess robustness to the choice of anchoring mode by treating private-vehicle flows as the reference to recover transit and walking flows (Supplementary Fig. 7). The framework maintains strong performance under this alternative configuration, indicating that the formulation is not tied to a specific observed mode. Together, these complementary evaluations demonstrate that the proposed data-fusion framework reliably captures city-scale temporal dynamics and spatial flow structure that cannot be recovered from sparse surveys or transit-only smart-card data in isolation. Estimating multiday, multimodal TD–OD flows We apply the framework to reconstruct 1 week of multimodal TD–OD flows in Singapore and Seoul, highlighting both multiday variation and cross-city contrasts. Figure 4 summarizes total trip volumes and time-of-day profiles across three weekday groups (Mon–Wed, Thu–Fri, Sat–Sun). In Singapore (Fig. 4a), total flows are stable from Mon–Wed to Thu–Fri (7.46 vs. 7.67 million trips; +2.7%) and decrease on weekends (6.38 million; −14.7%). Seoul shows the same qualitative pattern with a larger weekend contraction (Fig. 4c): total flows increase slightly from Mon–Wed to Thu–Fri (19.9 vs. 20.6 million; +3.5%) and drop sharply on Sat–Sun (12.9 million; −35%).Fig. 4Multiday multimodal time-dependent origin–destination (TD–OD) flows by weekday group.Singapore (a, b) and Seoul (c, d). a, c Total daily trips aggregated over all OD pairs for three weekday groups (Mon–Wed, Thu–Fri, Sat–Sun). Both cities exhibit stable weekday demand (Mon–Wed and Thu–Fri) and a pronounced reduction on weekends (Sat–Sun), reflecting consistent day-to-day mobility levels with systematic weekend contraction. b, d Aggregate hourly time-of-day profiles of TD–OD flows for transit, private vehicle, and walking, obtained by summing flows across all OD pairs within each weekday group. Weekday profiles show characteristic morning and evening peaks dominated by transit and private modes, whereas weekend profiles display flatter temporal distributions with reduced peak intensity and relatively higher daytime activity. Within each city, a common vertical scale is used to enable direct comparison of flow magnitude and temporal structure across weekday groups. Multiday multimodal time-dependent origin–destination (TD–OD) flows by weekday group. Singapore (a, b) and Seoul (c, d). a, c Total daily trips aggregated over all OD pairs for three weekday groups (Mon–Wed, Thu–Fri, Sat–Sun). Both cities exhibit stable weekday demand (Mon–Wed and Thu–Fri) and a pronounced reduction on weekends (Sat–Sun), reflecting consistent day-to-day mobility levels with systematic weekend contraction. b, d Aggregate hourly time-of-day profiles of TD–OD flows for transit, private vehicle, and walking, obtained by summing flows across all OD pairs within each weekday group. Weekday profiles show characteristic morning and evening peaks dominated by transit and private modes, whereas weekend profiles display flatter temporal distributions with reduced peak intensity and relatively higher daytime activity. Within each city, a common vertical scale is used to enable direct comparison of flow magnitude and temporal structure across weekday groups. The time-of-day profiles (Fig. 4b,d) reveal consistent weekday commuting structure in both cities, with pronounced bimodal peaks in transit flows. In contrast, private-vehicle and walking flows exhibit weaker morning peaks and relatively stronger midday activity. On weekends, activity shifts later and becomes more dispersed. Transit flows broaden into a midday plateau, private-vehicle flows form a wider midday peak, and walking flows increase substantially. This increase is most prominent in Seoul, where midday walking becomes the largest mode over several hours. These patterns highlight that cross-city differences emerge not only in total trip volumes but also in mode-specific temporal structure. Seoul exhibits a stronger midday walking surge, whereas Singapore retains more pronounced transit-dominated commuting peaks. These contrasts likely reflect a combination of network coverage, urban form, and survey design. In Singapore, walk-only trips are recorded only above a duration threshold; short walk segments used to access transit are excluded, which lowers the reported walk share relative to everyday experience62. Additional multiday spatial summaries are provided in Supplementary Fig. 8. Across days, the recovered patterns remain stable and interpretable across modes and time-of-day. Differences between weekday groups are primarily expressed as changes in flow magnitude and temporal allocation, while the overall modal structure remains broadly consistent. This behavior reflects the use of shared cross-modal relationships combined with day-specific transit observations to reconstruct TD–OD flows. As a result, weekday–weekend differences and holiday effects are captured through shifts in total demand and timing rather than through re-estimated mode-specific behavior (Supplementary Fig. 9). These results correspond to a typical study week without major system-wide disruptions (e.g., pandemics or large policy interventions). When day-specific cross-modal information is available, the formulation can be extended to account for both day-of-week variation in demand and modal composition. Assessing transit competitiveness We use the recovered multimodal TD–OD flows to examine how transit competitiveness varies with OD distance, using private-to-transit flow ratios as a diagnostic. Deviations from a distance-conditioned baseline are interpreted as relative losses or gains in transit flows generated by each origin zone (see Methods for details). In both cities, the baseline relationship exhibits a clear non-monotonic pattern: transit is most competitive at intermediate distances, while very short and long trips are more likely to be dominated by private modes. In Singapore (Fig. 5a,b), the private-to-transit flow ratio declines with distance up to approximately 16 km before increasing again. In Seoul (Fig. 5c,d), the corresponding turning point occurs earlier, at around 11 km, indicating that private modes begin to regain an advantage at shorter distances than in Singapore.Fig. 5Transit competitiveness inferred from distance-based benchmarking of private-to-transit flow ratios.a, c Private-to-transit flow ratios for origin–destination (OD) pairs as a function of OD distance in Singapore (a) and Seoul (c). Point colour indicates OD-pair density, and the locally weighted scatterplot smoothing (LOWESS) curve represents the distance-based baseline relating travel distance to relative private-versus-transit use. b, d LOWESS-smoothed baseline curves for Singapore (b) and Seoul (d), highlighting the distance at which the private-to-transit ratio reaches its minimum, corresponding to the range where transit is most competitive relative to private travel (Singapore: ~16 km; Seoul: ~11 km). e, f Origin-level transit loss/gain patterns derived from deviations between observed private-to-transit flow ratios and the distance-based baseline in Singapore (e) and Seoul (f). Positive values indicate relative transit gains, where transit use exceeds the baseline for trips generated from that origin, whereas negative values indicate relative transit losses, where private travel is more prevalent than expected given OD distance. Values are aggregated over destinations and time of day, and selected origins with pronounced gains or losses are annotated. Transit competitiveness inferred from distance-based benchmarking of private-to-transit flow ratios. a, c Private-to-transit flow ratios for origin–destination (OD) pairs as a function of OD distance in Singapore (a) and Seoul (c). Point colour indicates OD-pair density, and the locally weighted scatterplot smoothing (LOWESS) curve represents the distance-based baseline relating travel distance to relative private-versus-transit use. b, d LOWESS-smoothed baseline curves for Singapore (b) and Seoul (d), highlighting the distance at which the private-to-transit ratio reaches its minimum, corresponding to the range where transit is most competitive relative to private travel (Singapore: ~16 km; Seoul: ~11 km). e, f Origin-level transit loss/gain patterns derived from deviations between observed private-to-transit flow ratios and the distance-based baseline in Singapore (e) and Seoul (f). Positive values indicate relative transit gains, where transit use exceeds the baseline for trips generated from that origin, whereas negative values indicate relative transit losses, where private travel is more prevalent than expected given OD distance. Values are aggregated over destinations and time of day, and selected origins with pronounced gains or losses are annotated. Mapping deviations from the distance-based baseline reveals pronounced spatial heterogeneity across origin zones. In Singapore (Fig. 5e), transit losses are concentrated in eastern and parts of the central subzones, including Changi, Bedok, and Novena, whereas gains are observed in major hubs and outer areas, such as the Downtown Core, Woodlands, and Jurong West. In Seoul (Fig. 5f), losses are more prevalent in peripheral districts (e.g., Gayang1-dong, Seocho3-dong, and Nonhyeon2-dong), while central neighborhoods, such as Jongno-dong, Hoehyeon-dong, and Myeong-dong, exhibit substantial gains. These patterns point to two complementary areas for policy intervention. In outer districts, improving feeder services, express connections, and corridor capacity may help shift trips from private modes toward transit. In central and gateway zones, maintaining and extending high-performing transit networks is important for sustaining existing advantages. More generally, benchmarking private-to-transit flow ratios against a distance-based baseline provides a transferable and spatially explicit diagnostic for identifying where transit underperforms or exceeds expectations. Such insights cannot be obtained from transit smart-card data alone (Supplementary Fig. 10), which lack information on competing modes, nor from household travel surveys, which do not provide OD-complete private-to-transit flow ratios at comparable spatial and temporal resolution. Integrating OD flows into an epidemic spread model We integrate the estimated OD matrices into a spatially extended transmission model63, to quantify how population mobility and mode-specific OD structures shape epidemic dynamics (see results in Fig. 6). We parameterize transmission heterogeneity based on the ratio of total mobility flow to resident population. Zones where mobility flow exceeds resident population by factors of 50 (Singapore) or 15 (Seoul) are assigned Rt ∈ [1.5, 3], reflecting a larger mobile population in commercial and business districts increase contact opportunities and consequently intensify transmission, while residential-dominated zones maintain Rt ∈ [0.8, 1]. Here, the transmission during travel (e.g., while riding in public transport) is not explicitly modelled; instead, we consider that individuals may contract infections within their residential subzone or the subzone they visit.Fig. 6Integration of origin–destination (OD) flows from multiple travel modes into an epidemic model.a, d Citywide daily new infections for Singapore (a) and Seoul (d), comparing epidemic simulations based on all-mode and transit-only mobility matrices. Shaded bands indicate 95% confidence intervals (CI) around the median. b, e Subzone-level differences in total infections between all-mode and transit-only mobility scenarios. Positive values indicate higher infections under all-mode mobility, whereas negative values indicate lower infections. c, f Differences in peak timing (in days) for each subzone relative to the citywide peak under the all-mode scenario. Positive values indicate subzones peaking after the citywide maximum, whereas negative values indicate earlier peaks. Integration of origin–destination (OD) flows from multiple travel modes into an epidemic model. a, d Citywide daily new infections for Singapore (a) and Seoul (d), comparing epidemic simulations based on all-mode and transit-only mobility matrices. Shaded bands indicate 95% confidence intervals (CI) around the median. b, e Subzone-level differences in total infections between all-mode and transit-only mobility scenarios. Positive values indicate higher infections under all-mode mobility, whereas negative values indicate lower infections. c, f Differences in peak timing (in days) for each subzone relative to the citywide peak under the all-mode scenario. Positive values indicate subzones peaking after the citywide maximum, whereas negative values indicate earlier peaks. In Fig. 6a,d, we compare citywide daily incidence using mobility matrices constructed from all travel modes versus transit-only OD. The divergence is substantial: transit-only mobility consistently produces at least a 50% lower peak and delays peak timing by more than 50 days in both cities. This systematic underestimation arises because transit-only data fail to capture private-mode movements into high-flow destinations where contact opportunities—and thus infection risk—are maximized. The epidemiological significance extends beyond academic interest: pandemic preparedness based on transit-only mobility would critically underestimate healthcare surge requirements and mistime intervention deployment, potentially overwhelming hospital systems during unexpected peaks. Figure 6b, e reveals the spatial distribution of cumulative infection differences between all-modes and transit-only mobility matrices. Locations where multimodal data identify substantially higher transmission than transit-only projections correspond to centers accessed primarily by private vehicles. These systematically missed hotspots represent blind spots in surveillance and intervention targeting when mobility proxies are incomplete. The magnitude of these differences (up to 15,000 additional infections per subzone) demonstrates that mode-specific travel patterns fundamentally alter spatial epidemic dynamics, not merely scale overall transmission. Figure 6c,f displays the temporal asynchrony between subzone-level and citywide epidemic peaks when using all-modes mobility. Subzones peaking before the citywide maximum are distinguished from those with delayed peaks. This 7–10 day spread in peak timing across subzones implies that uniform citywide interventions-the standard pandemic response-are temporally misaligned with local transmission dynamics. Early-peaking zones require preemptive measures while late-peaking areas need sustained interventions beyond citywide relaxation, insights invisible without multimodal mobility data. Discussion This study proposes and validates a city-scale data-fusion framework for recovering multiday, multimodal TD–OD flows at planning-relevant resolutions (hourly and zone scale) by integrating survey-based cross-modal information with near-census transit smart-card observations. Rather than replacing gravity-based or deep-learning OD estimation/forecasting models designed for data-rich, single-mode settings, the proposed framework complements existing approaches by addressing a distinct and practically important gap: multimodal TD–OD recovery under partial observability, where some transport modes lack direct passive observations. The framework offers three practical advantages. First, the framework is lightweight. Multimodal TD–OD recovery relies only on a dense passive dataset for a single reference mode, together with limited cross-modal calibration information. The household travel survey component uses solely basic trip attributes, such as origin, destination, departure time, and travel mode, to estimate aggregate cross-modal ratios. No detailed activity diaries, sociodemographic variables, or lengthy survey instruments are required. The low-dimensional survey input (cross-modal ratios) allows robust recovery of multimodal mobility flows despite limited sample sizes (Supplementary Table 5). Second, it is interpretable: cross-modal relationships are learned explicitly in ratio space, and the adaptive aggregation mechanism provides transparent control over the trade-off between stabilizing noisy survey ratios and preserving OD-time heterogeneity under sparse observations. Feature-importance analysis further indicates that recovered relationships are primarily driven by observed flow structure, supporting diagnostic analysis and policy-facing interpretation (Fig. 2g). Third, it is modular: absolute flows are anchored using a high-coverage passive dataset for a reference mode, while cross-modal structure is supplied by surveys. This separation makes the framework extensible beyond smart-card data; other passive datasets (e.g., cellular signaling or GPS-based flows) can serve as reference modes when their coverage characteristics and biases are explicitly accounted for. The modularity of the framework is further illustrated by an additional validation in which private-vehicle demand is treated as the reference mode to recover transit and walking flows (Supplementary Fig. 7). Importantly, the recovered TD–OD flows are validated at the city scale (i.e., Singapore and Seoul), rather than at isolated corridors or subareas, demonstrating strong spatial and temporal agreement across OD, origin, destination, and time dimensions. With this validation established, the subsequent applications focus on the implications of the recovered multimodal flow patterns rather than on further model assessment. The applications illustrate how recovering multimodal TD–OD flows expands the scope of policy-relevant analysis beyond what can be learned from single-day surveys or single-mode passive datasets. The first application (Fig. 4) shows how multiday, multimodal TD–OD flows reveal policy-relevant temporal dynamics that cannot be observed from single-day surveys or single-mode passive datasets alone. Systematic contrasts between weekdays and weekends, peak spreading across the week (e.g., stronger Thu–Fri evening activity), and flattened weekend morning profiles with pronounced midday plateaus point to opportunities for adaptive service planning. These include reallocating vehicle and staffing resources toward late-evening flows or shifting capacity from early peaks to midday periods on weekends. Divergent temporal signatures across modes-bimodal transit peaks, extended private-vehicle plateaus, and pronounced midday walking surges-further motivate cross-modal coordination, such as aligning transit headways with first/last-mile services and micromobility availability. Linking flow patterns to zonal attributes also enables evaluation of equity by time and mode, identifying areas that are well served on weekdays but underserved on weekends, or vice versa. The second application (Fig. 5) translates recovered multimodal flows into interpretable measures of transit competitiveness. Benchmarking private-to-transit ratios against a distance-based baseline reveals a non-monotonic relationship, with transit most competitive at intermediate distances and more vulnerable to private modes for very short and very long trips. Spatial patterns show that transit losses concentrate in peripheral districts-where first/last-mile access, feeder services, and express corridors could yield the greatest gains-while central areas exhibit sustained or excess transit capture, underscoring the importance of maintaining high-performing networks. This distance-based benchmarking provides a generalizable diagnostic lens for identifying where transit underperforms or exceeds expectations-patterns that are not observable from transit smart-card data alone and cannot be derived from surveys lacking OD-complete private-to-transit ratios. These insights support performance-based investment, allowing agencies to prioritize corridors by substitutable trips and expected returns. The third application (Fig. 6) demonstrates that incomplete mobility observations introduce systematic biases that amplify through transmission dynamics. Using transit-only mobility proxies leads to substantial underestimation of peak exposure and temporal misalignment, resulting in underprediction of epidemic spread, particularly in districts accessible primarily by private modes. These errors propagate nonlinearly through epidemic models, causing missed transmission events and cascading projection failures at the city scale. Cities relying solely on transit-based mobility proxies, therefore, face elevated risk during future outbreaks. The proposed multimodal TD–OD recovery framework mitigates these risks by enabling epidemiological models to account for the full mobility system. Several limitations motivate future work. First, the empirical analysis is conducted on two cities using datasets that are not fully publicly accessible, including smart-card transactions and external passive benchmarks. This limits the immediate reproducibility of the full pipeline and may constrain the generalizability of the results to settings with similar data availability and urban structure. Second, survey-derived ratios remain uncertain in low-flow contexts; explicit uncertainty quantification and hierarchical pooling across space and time would enable principled propagation of this uncertainty into recovered TD–OD matrices. Third, empirical validation is conducted on a representative week in Singapore and a representative weekday in Seoul, consistent with the availability of independent passive benchmarks. The validation is therefore regime-conditioned, corresponding to regular operations with typical school and work schedules and no major disruptions. Extending validation to structurally different regimes (e.g., holidays, major policy changes, or extreme events) would require comparable passive benchmarks and likely regime-aware or time-varying cross-modal ratios to capture shifts in modal composition. Fourth, in the current implementation, cross-modal ratios vary by time of day but not by day of week, as the available household travel surveys do not support reliable day-specific stratification. Surveys that oversample day-of-week strata or employ longitudinal diary designs would enable this extension. Methods Data and preprocessing We analyze one week of smart-card transactions for Singapore (18–24 July 2022) and Seoul (18–24 October 2021). Trip records are preprocessed by merging consecutive boarding segments with dwell times under 15 min into complete journeys and then aggregated from stop–minute to zone–hour resolution. The study areas comprise 303 subzones in Singapore and 420 traffic analysis zones in Seoul (Supplementary Fig. 1), with time discretized into 24 1-h bins per day. The resulting smart-card datasets provide high spatiotemporal coverage, spanning millions of trips and more than 0.73 million OD–time tuples in Singapore and 1.46 million in Seoul. In contrast, household travel surveys-trip-diary records reporting origin, destination, departure time, travel mode, and trip purpose-provide substantially fewer observations at the same zone–hour resolution: 40.3 thousand OD–time tuples in Singapore and 117.6 thousand in Seoul. IRB approval was not sought because we received de-identified and anonymized survey and smart card data from relevant transport planning agencies under a data-sharing agreement. As cross-modal ratios are computed only for OD–time tuples with positive counts in both modes, the subset usable for cross-modal ratio estimation is further limited. In Singapore, only 2.1 thousand OD–time tuples support private-to-transit ratios and 1.0 thousand support walk-to-transit ratios; in Seoul, the corresponding counts are 8.2 and 7.0 thousand, respectively (Supplementary Table 1). This imbalance between smart-card observations and survey-derived ratios is addressed using the adaptive representation learning (ARL) framework described below, which stabilizes and generalizes these estimates. Within the framework, survey data are used exclusively to estimate cross-modal ratios and do not define TD–OD flow patterns or disaggregated behavioral models. Trip-purpose information is used solely to derive land-use mixture features, and sociodemographic attributes are not required, enabling estimation from readily available inputs. Adaptive representation learning (ARL) for estimating cross-modal ratios The central challenge is that survey-based cross-modal ratios-the number of trips by one mode relative to another for a given OD pair and time period-are unstable when computed from small samples. ARL addresses this by blending each raw ratio with a spatially aggregated prior, with the degree of blending determined by local sample size. A supervised model then learns to predict these stabilized ratios across all OD–time tuples, while a robust selection procedure identifies the aggregation level that yields the most consistent flow estimates. These components-adaptive aggregation, supervised learning, and robust selection-form a unified pipeline (Fig. 2a). Together, they produce stable estimates of cross-modal ratios \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{R}}_{odt}^{ij}$$\end{document}R^odtij and the corresponding recovered flows \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{N}}_{odt}^{i}$$\end{document}N^odti for all OD–time tuples. Adaptive representation through aggregation For each OD–time tuple (o, d, t), the raw survey-based cross-modal ratio is defined as 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{odt}^{ij}=\frac{{N}_{odt}^{i}}{{N}_{odt}^{j}},$$\end{document}Rodtij=NodtiNodtj, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{odt}^{i}$$\end{document}Nodti and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{odt}^{j}$$\end{document}Nodtj denote the survey trip counts for the target mode i and reference mode j, respectively (Fig. 2c). These ratios are often unstable under sparse sampling. To improve robustness, we construct an adaptively smoothed ratio (Fig. 2d) 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{odt}^{ij}(k)={w}_{odt}^{ij}(k)\,{R}_{odt}^{ij}+\left[1-{w}_{odt}^{ij}(k)\right]{\widetilde{R}}_{odt}^{ij},$$\end{document}Rodtij(k)=wodtij(k)Rodtij+1−wodtij(k)R~odtij,which interpolates between the raw ratio and an aggregated prior \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widetilde{R}}_{odt}^{ij}$$\end{document}R~odtij through the weighting function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w}_{odt}^{ij}(k)$$\end{document}wodtij(k). The aggregated prior is constructed from origin- and destination-level marginal ratios (Fig. 2c–d): 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widetilde{R}}_{odt}^{ij}={u}_{ot}^{i}\,{\widetilde{R}}_{ot}^{ij}+{u}_{dt}^{i}\,{\widetilde{R}}_{dt}^{ij},$$\end{document}R~odtij=uotiR~otij+udtiR~dtij, where 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widetilde{R}}_{ot}^{ij}=\frac{{N}_{ot}^{i}}{{N}_{ot}^{j}},\,{\widetilde{R}}_{dt}^{ij}=\frac{{N}_{dt}^{i}}{{N}_{dt}^{j}},$$\end{document}R~otij=NotiNotj,R~dtij=NdtiNdtj, and the mixing weights reflect trip counts at each end: 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${u}_{ot}^{i}=\frac{{N}_{ot}^{i}}{{N}_{ot}^{i}+{N}_{dt}^{i}},\,{u}_{dt}^{i}=\frac{{N}_{dt}^{i}}{{N}_{ot}^{i}+{N}_{dt}^{i}}.$$\end{document}uoti=NotiNoti+Ndti,udti=NdtiNoti+Ndti. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{ot}^{i}$$\end{document}Noti and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{dt}^{i}$$\end{document}Ndti denote marginal counts aggregated over destinations and origins, respectively: 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{ot}^{i}={\sum }_{d}{N}_{odt}^{i},\,{N}_{ot}^{j}={\sum }_{d}{N}_{odt}^{j},$$\end{document}Noti=∑dNodti,Notj=∑dNodtj,8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{dt}^{i}={\sum }_{o}{N}_{odt}^{i},\,{N}_{dt}^{j}={\sum }_{o}{N}_{odt}^{j}.$$\end{document}Ndti=∑oNodti,Ndtj=∑oNodtj. Thus, the prior combines origin- and destination-level ratios, weighted by trip counts so that higher-flow zones contribute more. The aggregation weight depends on the effective sample size (Fig. 2e) 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{odt}^{ij}={N}_{odt}^{i}+{N}_{odt}^{j},$$\end{document}Nodtij=Nodti+Nodtj, and is defined as 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w}_{odt}^{ij}(k)=\min \left\{1,{\left(\frac{{{N}^{ij}}_{odt}}{{Q}_{0.98}}\right)}^{k}\right\},$$\end{document}wodtij(k)=min1,NijodtQ0.98k, where Q0.98 is the 98th percentile of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{{N}_{odt}^{ij}\}$$\end{document}{Nodtij}, used to reduce sensitivity to extreme values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{odt}^{ij}$$\end{document}Nodtij. This weighting scheme in Eq. (3) assigns higher weight to the raw ratio when the sample size is large, and shifts weight toward the aggregated prior when observations are sparse. The aggregation parameter k controls the rate of this transition: larger values of k increase reliance on aggregated information, whereas smaller values retain more of the raw, data-driven signal for well-supported OD–time tuples. Supervised learning on aggregated representations For each aggregation level k, the smoothed ratios \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{odt}^{ij}(k)$$\end{document}Rodtij(k) are used as training targets for supervised models, which predict \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{R}}_{odt}^{ij}(k)$$\end{document}R^odtij(k) across all OD–time tuples (Fig. 2a). The models leverage flow-based, spatiotemporal, and land-use features derived from survey and passive datasets (Fig. 2g). Given the predicted ratios and reference-mode flows \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bar{N}}_{odt}^{j}$$\end{document}N¯odtj (Fig. 2b), target-mode flows are recovered as 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{N}}_{odt}^{i}(k)={\widehat{R}}_{odt}^{ij}(k)\,{\bar{N}}_{odt}^{j},$$\end{document}N^odti(k)=R^odtij(k)N¯odtj, and total multimodal flows are obtained by summation: 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{N}}_{odt}(k)={\sum }_{i}{\widehat{N}}_{odt}^{i}(k).$$\end{document}N^odt(k)=∑iN^odti(k). Robust selection of aggregation level The aggregation parameter k is selected based on the stability of recovered flows across candidate values (Fig. 2f). For each k, we define 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{RobustError}}}}(k)=\sqrt{\frac{1}{| \Omega | }{\sum }_{(o,d,t)\in \Omega }{\left({\widehat{N}}_{odt}(k)-{{{{\rm{Median}}}}}_{k}\{{\widehat{N}}_{odt}(k)\}\right)}^{2}},$$\end{document}RobustError(k)=1∣Ω∣∑(o,d,t)∈ΩN^odt(k)−Mediank{N^odt(k)}2, which quantifies the deviation of flows at a given k from the median estimate across all candidate values. The optimal value is 14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}^{\star }=\arg {\min }_{k}{{\rm{RobustError}}}(k),$$\end{document}k⋆=argminkRobustError(k), selecting the aggregation level that yields the most stable and internally consistent flow estimates. Marginal constraints via iterative proportional fitting (IPF) The ARL provides cross-modal ratio estimates at the OD–time level, but these alone do not guarantee consistency with aggregate flow constraints. To enforce coherence, we apply IPF, which rescales the ARL-based OD–time estimates to match observed totals at the origin, destination, and time-of-day levels, while preserving the relative flow structure learned by the ARL. The procedure requires three sets of marginal constraints-origin, destination, and time-of-day-constructed as follows. From passive data for the reference mode j, we compute origin and destination marginals: 15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bar{N}}_{o}^{j-}={\sum }_{d,t}{\bar{N}}_{odt}^{j},\,{\bar{N}}_{d}^{j+}={\sum }_{o,t}{\bar{N}}_{odt}^{j}.$$\end{document}N¯oj−=∑d,tN¯odtj,N¯dj+=∑o,tN¯odtj. From survey data, we obtain corresponding marginals for both target and reference modes 16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{o}^{i-}={\sum }_{d,t}{N}_{odt}^{i},\,{N}_{o}^{j-}={\sum }_{d,t}{N}_{odt}^{j},$$\end{document}Noi−=∑d,tNodti,Noj−=∑d,tNodtj,17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{d}^{i+}={\sum }_{o,t}{N}_{odt}^{i},\,{N}_{d}^{j+}={\sum }_{o,t}{N}_{odt}^{j}.$$\end{document}Ndi+=∑o,tNodti,Ndj+=∑o,tNodtj. These define origin- and destination-level cross-modal ratios 18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{o}^{ij-}=\frac{{N}_{o}^{i-}}{{N}_{o}^{j-}},\,{R}_{d}^{ij+}=\frac{{N}_{d}^{i+}}{{N}_{d}^{j+}},$$\end{document}Roij−=Noi−Noj−,Rdij+=Ndi+Ndj+, which are used to construct target-mode marginal constraints by anchoring to passive observations: 19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{G}}_{o}^{i-}={\widehat{R}}_{o}^{ij-}\,{\bar{N}}_{o}^{j-},\,{\widehat{A}}_{d}^{i+}={\widehat{R}}_{d}^{ij+}\,{\bar{N}}_{d}^{j+}.$$\end{document}G^oi−=R^oij−N¯oj−,A^di+=R^dij+N¯dj+. From passive data, the reference-mode time-of-day marginal is 20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bar{N}}_{t}^{j}={\sum }_{o,d}{\bar{N}}_{odt}^{j},$$\end{document}N¯tj=∑o,dN¯odtj, and from survey data, the corresponding time-of-day cross-modal ratios are 21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{t}^{ij}=\frac{{N}_{t}^{i}}{{N}_{t}^{\,j}},$$\end{document}Rtij=NtiNtj, where 22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{t}^{i}={\sum }_{o,d}{N}_{odt}^{i},\,{N}_{t}^{j}={\sum }_{o,d}{N}_{odt}^{j}.$$\end{document}Nti=∑o,dNodti,Ntj=∑o,dNodtj. These yield the target-mode time-of-day marginal 23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{P}}_{t}^{i}={\widehat{R}}_{t}^{ij}\,{\bar{N}}_{t}^{j}.$$\end{document}P^ti=R^tijN¯tj. Finally, the three sets of marginal constraints—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{G}}_{o}^{i-}$$\end{document}G^oi− (origin), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{A}}_{d}^{i+}$$\end{document}A^di+ (destination), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{P}}_{t}^{i}$$\end{document}P^ti (time of day)—jointly define the feasible solution space for target-mode flows. Starting from the ARL-based OD–time estimates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{N}}_{odt}^{i}$$\end{document}N^odti, IPF iteratively rescales flows to match these marginals, enforcing consistency with observed reference-mode totals while stabilizing estimates under sparse survey conditions. Supplementary Fig. 11 illustrates these constraints through the estimated origin-level marginals and time-of-day profiles for private vehicle and walking modes, shown for representative weekdays and weekends. Validation setup Recovered TD–OD flows are validated against independent passive datasets: TomTom vehicle-probe data in Singapore and cellular signaling data in Seoul. Because these data sources differ in sampling coverage and penetration rates, their marginal distributions are not directly comparable (Supplementary Fig. 12). To isolate differences in spatial and temporal patterns from differences in absolute scale, benchmark flows are aligned to the same marginal constraints as the recovered flows using IPF, matching trip-generation, trip-attraction, and time-of-day profiles. This alignment removes differences in absolute scale and enables validation to focus on relative variation across origins, destinations, and time of day, without introducing additional information into the recovered TD–OD flows. Validation is performed over the set of OD–time tuples defined by the recovered dataset, ensuring a consistent basis for comparison. Agreement is quantified using two complementary metrics. The common part of commuters (CPC)18–20 measures the overlap between recovered and benchmark flow distributions and emphasizes agreement in relative spatial and temporal allocation: 24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{CPC}}}}=\frac{2{\sum }_{(o,d,t)}\min \left({{\widehat{N}}^{i}}_{odt},\,{N}_{odt}^{i\star }\right)}{{\sum }_{(o,d,t)}\left({\widehat{N}}_{odt}^{i}+{N}_{odt}^{i\star }\right)}.$$\end{document}CPC=2∑(o,d,t)minN^iodt,Nodti⋆∑(o,d,t)N^odti+Nodti⋆. The coefficient of determination (R2) captures agreement in flow magnitudes, reflecting the ability to reproduce variance in absolute demand levels: 25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}^{2}=1-\frac{{\sum }_{(o,d,t)}{\left({\widehat{N}}_{odt}^{i}-{N}_{odt}^{i\star }\right)}^{2}}{{\sum }_{(o,d,t)}{\left({N}_{odt}^{i\star }-{\bar{N}}^{i\star }\right)}^{2}}.$$\end{document}R2=1−∑(o,d,t)N^odti−Nodti⋆2∑(o,d,t)Nodti⋆−N¯i⋆2. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{N}}_{odt}^{i}$$\end{document}N^odti denotes the recovered flows, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{odt}^{i\star }$$\end{document}Nodti⋆ the benchmark flows, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bar{N}}^{i\star }$$\end{document}N¯i⋆ the mean benchmark flow across all OD–time tuples. Together, CPC and R2 provide complementary assessments of structural consistency and volumetric accuracy. Transit competitiveness Transit competitiveness is assessed by comparing the estimated cross-modal ratio \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{R}}_{od}$$\end{document}R^od for each OD pair (o, d) with a distance-based baseline \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{od}^{\star }$$\end{document}Rod⋆. The baseline is obtained using locally weighted scatterplot smoothing (LOWESS)64, which relates the private-to-transit ratio to travel distance through a smooth nonparametric fit. Given the total OD flow Nod across transit and private modes, the baseline-expected and estimated numbers of transit trips are 26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{N}}_{od}^{\star }=\frac{{N}_{od}}{1+{R}_{od}^{\star }},\,{\widehat{N}}_{od}=\frac{{N}_{od}}{1+{\widehat{R}}_{od}}.$$\end{document}N^od⋆=Nod1+Rod⋆,N^od=Nod1+R^od. The deviation from the baseline is then 27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {\widehat{N}}_{od}={\widehat{N}}_{od}-{\widehat{N}}_{od}^{\star },$$\end{document}ΔN^od=N^od−N^od⋆, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {\widehat{N}}_{od}$$\end{document}ΔN^od measures the difference between observed and distance-expected transit usage. Positive values indicate that transit captures more riders than expected for trips of that distance, while negative values indicate underperformance. For example, a negative value on a particular corridor suggests that transit is less attractive there than on other corridors of comparable length, potentially signaling opportunities for service improvement. Aggregating \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {\widehat{N}}_{od}$$\end{document}ΔN^od across origins or destinations yields zone-level indicators of transit competitiveness, highlighting locations where transit performs below or above distance-based expectations. Mobility-driven epidemic model To illustrate the epidemiological implications of reconstructed multimodal mobility, we adopt a mobility-driven transmission model63 in which infections arise through three pathways: (i) local transmission within a region, (ii) exposure to infected individuals arriving from other regions, and (iii) exposure that occurs when susceptible individuals travel outward to regions with active infections. The number of new infections in region i at time t is given by 28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I}_{t,i}={\sum }_{j=1}^{M}\frac{{m}_{ji}(t){S}_{t,i}}{\left[{\sum }_{l=1}^{M}\,{m}_{jl}(t){N}_{l}\right]}\cdot {R}_{t,j}\left[{\sum }_{k=1}^{M}{m}_{jk}(t){\sum }_{\tau=0}^{t-1}{I}_{\tau,k}{g}_{t-\tau }\right],$$\end{document}It,i=∑j=1Mmji(t)St,i∑l=1Mmjl(t)Nl⋅Rt,j∑k=1Mmjk(t)∑τ=0t−1Iτ,kgt−τ, where mji(t) denotes mobility flows from region j to region i, Nl is the population of region l, and Rt,j is the time-varying reproduction number in region j. The denominator normalizes by the effective population present in region j at time t, accounting for the inflow and outflow of individuals. The susceptible population evolves as 29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S}_{t,i}={N}_{i}-{\sum }_{s=1}^{t-1}{I}_{s,i},$$\end{document}St,i=Ni−∑s=1t−1Is,i, where Ni is the total population of region i. The generation time distribution g is obtained by discretizing a Gamma distribution with mean 6.5 days and coefficient of variation 0.62, consistent with SARS-CoV-2 transmission dynamics. To capture heterogeneity in contact intensity across regions, we parameterize transmissibility based on the ratio of total mobility flow to resident population—a proxy for the density of transient interpersonal contacts. In Singapore, subzones where mobility flows exceed 50 times the resident population—typically commercial and transport hub areas—are assigned a higher transmissibility range (Rt ∈ [1.5, 3]), while all others are assigned Rt ∈ [0.8, 1]. In Seoul, where contrasts between resident and transient populations are less pronounced, a lower threshold of 15 is used. Details on the construction of mobility matrices from OD flow data are provided in Supplementary Notes. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Supplementary Information Reporting Summary Transparent Peer Review file