The methodological framework is designed to support a single analytical objective: Identifying and characterising potential dockless e-bike integration with rail transit. Data sources, descriptive metrics, and modelling tools are therefore presented as complementary components of this unified pipeline. Given the limitations of shared micromobility datasets—namely, heterogeneous GPS accuracy, missing values, noisy distance records, and the absence of continuous trajectories—our approach combines reproducible data cleaning procedures, network-based routing, and complementary spatial modelling. The workflow begins with systematic validation of numeric, temporal, and geographic fields, followed by network-constrained route reconstruction and the detection of trips compatible with bike-and-ride behaviour. Bike-and-ride identification based on large-scale spatial data tends to overestimate actual intermodality relative to smart-card or trajectory-based methods. To address this limitation, we introduce the concept of ‘Inside Station Catchment Area’ (ISCA) trips, defined as DBS trips whose endpoints fall within station-accessible areas and occur during public transport operating hours. ISCA thus represents a proxy for potential intermodality rather than a direct observation of transfers. We first introduce our case study, then describe the data cleaning workflow and the reconstruction of network-constrained routes using OpenTripPlanner. We subsequently detail the spatiotemporal procedure used to extract ISCA trips. Finally, to investigate the influence of accessibility, built environment, and sociodemographic factors on shared e-bike usage, we estimate a set of regression models, including global specifications (GLM and OLS) and GWR models across temporal slices. The latter are used as a diagnostic tool to assess whether relationships exhibit meaningful spatial and temporal heterogeneity beyond global model assumptions, and to evaluate the added value of local modelling in this context.
Case study
The extent to which DBS services integrate with public transport depends on local governance, the spatial logic of deployment, and the availability of empirical data to observe usage at scale. The Lausanne–Morges agglomeration (PALM) provides a relevant setting: It combines a dense rail backbone with persistent short-distance car use, and has placed multimodality and station accessibility at the core of its strategic planning. In line with these objectives, Lausanne introduced its first dockless electric bikesharing (DBS) service in April 2024. Operated by Bird, the system runs 24/7 and relies on virtual parking zones. Since launch, coverage has expanded across most of the city and the University of Lausanne (UNIL) campus, offering an opportunity to examine whether DBS availability can support transfers to rail at the metropolitan scale. Within the canton of Vaud, the Lausanne–Morges metropolitan area has been structured since 2007 through the Lausanne–Morges Agglomeration Project (PALM), integrated into the Plan Directeur Cantonal (PDCn) under measure R1133. Because DBS flows are primarily concentrated within Lausanne and the Dorigny campus (UNIL), we adopt a more granular perimeter and focus on four PALM sectors (schémas directeurs): SDCL, SDOL, SDEL, and SDNL, excluding Région Morges to ensure a more focused analysis. PALM accounts for 286,000 inhabitants and 210,000 jobs (37% of the cantonal population and 48% of cantonal employment)33. Within our study area, a dense and integrated public transport network provides extensive regional and local coverage. The rail backbone is structured around the RER Vaud, operated by SBB CFF FFS since 1999, with Lausanne, Renens, and Prilly-Malley acting as major interchange hubs. At the local scale, the TL network complements rail services through two metro lines (M1, M2), trolleybus and bus lines, and the LEB railway line integrated into the TL system. The fare system is integrated under Mobilis, enabling seamless transfers across operators and modes. According to the Swiss Mobility and Transport Microcensus (MTMC 2021), private cars remain the dominant mode in Switzerland (69% of trips), while public transport accounts for 20%34. Car ownership is widespread (78% of households own at least one car), and bicycle ownership is also substantial (61% of households)34. Cycling trips, including bicycles and e-bikes, remain shorter than average daily travel distances, with marked regional differences across cantons. In Vaud, the public transport modal share remains lower than in leading Swiss cantons such as Basel-Stadt and Zurich (Fig. 1)34,35.
Fig. 1: Bivariate map of cycling and public transport commuting shares in Switzerland.
The alternative text for this image may have been generated using AI.
The figure presents the joint distribution of cycling and public transport commuting modal shares across the 26 Swiss cantons, based on data from the 2021 Mobility and Transport Microcensus (MTMC). The upper choropleth map encodes two variables simultaneously using a bivariate colour scheme. The transit modal share dimension is represented in blue, with three classes: Below 18.0% (light blue), 18.0–24.0% (medium blue), and above 24.0% (dark blue). The cycling modal share dimension is represented in red, with three classes: Below 6.9% (light red), 6.9–9.1% (medium red), and above 9.1% (dark red). Combinations of high values on both dimensions appear in dark purple-grey. Cantons are labelled with their official two-letter abbreviations. The lower scatter plot positions each canton according to its cycling modal share (horizontal axis, 2.5–22.5%) and transit modal share (vertical axis, 10.0–45.0%), with each canton represented by a filled circle coloured according to the same bivariate scheme as the map. Notable patterns include Basle-City (BS) as an outlier combining the highest cycling share (above 20%) with a high transit share, while Vaud (VD) and Neuchâtel (NE) exhibit high transit shares alongside low cycling shares.
Mobility is a key climate challenge in Vaud: Transport accounts for a substantial share of cantonal greenhouse gas emissions36. Despite a dense rail network, short car trips remain frequent in the Lausanne agglomeration37, indicating that access and egress conditions still constrain public transport use. At the same time, Lausanne has a comparatively low motorisation rate38, and multimodality is relatively common, with trips often involving more than one segment39. The rapid uptake of e-bikes has further reshaped cycling dynamics and motivates renewed attention to bicycle-based station accessibility37,40. To support modal shift, national and cantonal strategic documents stress the need to improve station accessibility and to facilitate intermodality between transit, walking, and cycling41. In Vaud, a large majority of residents and jobs are located within a few kilometres of a railway station, a range well suited to cycling as a feeder mode42,43. Mobility and land use policies in Switzerland are increasingly shaped by binding climate objectives at national, cantonal, and municipal levels. Switzerland’s second Nationally Determined Contribution commits to reducing greenhouse gas emissions by at least 65% by 2035 relative to 1990 levels44. This trajectory is reinforced by the Federal Act on Climate Protection, Innovation and Energy Security (2022), which sets sector-specific targets for transport, including a 57% reduction by 2040 and net-zero emissions by 205045. At the cantonal scale, the Perspectives Mobilité 2050 for the Canton of Vaud identifies only the most ambitious scenario (‘Sobriety’, S3) as compatible with long-term climate commitments, projecting a substantial increase in rail and cycling shares alongside a decline in car use46. These orientations are embedded in the Plan Directeur Cantonal, which aims to reduce the modal share of private cars to 50% by 205033, and in the cantonal climate strategy. At the municipal level, Lausanne has committed to achieving zero direct mobility-related emissions by 203047. Cycling plays a central role in this transition. The Federal Act on Cycling Infrastructure, in force since 2023, requires all cantons and municipalities to develop and implement comprehensive cycling networks by 204245,48. In Vaud, the Plan Directeur Cantonal and the Plan Climat Vaudois target a tripling of cycling distance travelled and a modal share of at least 10% by 203533,49. Lausanne’s climate strategy further reinforces this ambition, aiming for a major expansion of cycling activity by 203047. Rail transport constitutes the backbone of Switzerland’s low-carbon mobility strategy. The Programme de développement stratégique de l’infrastructure ferroviaire (PRODES) plans a substantial increase in service frequency by 2035, with half-hourly or quarter-hourly services on most main lines50. This expansion is integrated into the SBB CFF FFS 2030 Strategy, which explicitly emphasises seamless FLM access, notably through improved active mobility connections and bicycle parking around stations51. At the cantonal and metropolitan levels, rail-oriented strategies place strong emphasis on intermodality. The Stratégie ferroviaire vaudoise seeks to anchor rail at the core of daily mobility, supported by enhanced walking and cycling connectivity43. National and regional planning documents similarly highlight the Lausanne–Geneva metropolitan area as a priority for strengthening active modes as feeder services to mass transit37,41. Within this framework, intermodal interfaces and short-distance access to stations are identified as key levers for achieving modal shift33,52. Together, these strategic orientations underscore the central role assigned to cycling, rail transport, and their integration within Swiss climate and mobility policies. In this context, shared micromobility services are increasingly viewed as promising instruments to address FLM constraints and to operationalise intermodality objectives in dense and topographically constrained urban regions such as Lausanne.
Implementation of a semi-floating e-bike service in Lausanne
The Bird shared e-bike service was launched in Lausanne on April 22, 2024, as a pilot project supported by the municipal administration53,54,55. The initial fleet (100 e-bikes) expanded to 300 units by May 202453, and the service was extended to the UNIL campus in October 202456. Users can travel within the operational perimeter subject to parking rules and geofenced constraints (Fig. 2).
Fig. 2: Parked Bird DBS around two main multimodal hubs.
The alternative text for this image may have been generated using AI.
Pictures of parked Bird DBS around Gare de Lausanne and Gare Prilly-Malley multimodal hubs. The two photographs illustrate the physical presence of Bird dockless electric bikesharing (DBS) vehicles at two multimodal hubs in the Lausanne metropolitan area. A Two Bird electric bicycles are parked in front of one of the entrances of Gare de Lausanne railway station. B Several Bird electric bicycles are parked at Malley metro station, near Gare Prilly-Malley, representing a multimodal hub located in the western part of the city. Both photographs show the ‘semi-floating’ operational model of the Bird service, in which bicycles may be parked within designated zones.
Trips are made on app-unlocked, GPS-enabled electric-assist bicycles with geofenced parking. The operation follows a semi-floating model: Users can start and end trips without physical docks, but parking is restricted to predefined zones (Fig. 3), some of which require locking to designated racks57. This hybrid configuration aims to combine the flexibility of dockless systems with improved management of public space and safety. The service complements the existing station-based system in the Lausanne–Morges area (PubliBike Velospot)58,59. The geofencing scheme distinguishes functional zones (operation, mandatory parking, low-speed areas, and prohibited areas) to ensure orderly integration into the urban environment56,58.
Fig. 3: Spatial distribution of DBS parking spots within the study area.
The alternative text for this image may have been generated using AI.
The figure presents the location of dockless bikesharing (DBS) designated parking spots across the main part of the Lausanne city area. The upper panel displays the spatial distribution of DBS parking spots across the Lausanne administrative area. Yellow-filled circles indicate individual DBS parking spots. The dark grey detached polygon to the left indicates the University of Lausanne (UNIL) campus of Dorigny. Parking spots are visibly concentrated in the central city and along major transit corridors, with sparser coverage in peripheral and residential areas. The lower-left inset shows the location of the Canton of Vaud (VD) within Switzerland, highlighted in light grey, providing national geographic context. The lower-right inset displays the Lausanne municipal and metropolitan areas in the Canton de Vaud.
Pricing follows a standard pay-per-use model (unlock fee and per-minute fare), complemented by subscription options and occasional promotions60.
Dockless bikesharing data collection
The dataset analysed in this study originates from the Bird DBS system operating in Lausanne and the UNIL campus and was made available through a research collaboration between the University of Lausanne (HEC Lausanne) and the operator. Bird granted secure access to anonymised operational data extracted from its backend infrastructure. Data access, confidentiality, and use were governed by formal agreements, and data handling complied with applicable data protection requirements.
The dataset covers 459 consecutive days (April 22, 2024–July 24, 2025) and comprises 63,322 ride records (Supplementary Fig. 2). It includes anonymised user identifiers, trip metadata, and spatiotemporal information (Table 1). Each record corresponds to a completed ride and provides start and end coordinates (WGS84), start and end timestamps (ISO 8601), and a distance variable (cleaned_distance) produced after operator preprocessing. Descriptive exploration indicates 7364 unique users over the observation period, with heterogeneous usage intensity (median use markedly lower than the mean). The dataset provides sufficiently granular origin and destination information to support routing and network-based analyses, while the absence of continuous GPS trajectories limits analyses to start and end locations and inferred routes. This description motivates the analytical framework, which details the data cleaning, routing procedures, and the identification of segments potentially connected to public transport networks.
Table 1 Raw data descriptionData cleaning workflow
All cleaning and validation steps were performed using reproducible Python workflows, combining trip-level exclusions with geo-enabled consistency checks. Ten sequential exclusion criteria were applied to remove incomplete, inconsistent, or implausible records (Supplementary Fig. 1).
Missing values (df_cleaning_01). We removed all records containing at least one missing field across identifiers, timestamps, coordinates, the city label, or the distance variable. This step eliminated 1385 trips (2.19%) and 44 users (0.60%), leaving 61,935 trips and 7320 users. Missingness was primarily concentrated in the distance-related timestamp ride_completed_at. Duplicate trip identifiers (df_cleaning_02). To preserve a strict one-to-one mapping between rows and trips, we screened the dataset for repeated ride_id values. No duplicates were detected, leaving counts unchanged at 61,935 trips and 7320 users. Ensuring identifier uniqueness safeguards valid aggregation and supports reliable standard-error estimation. Date-format validation (df_cleaning_03). We checked that all timestamp fields—dt, ride_started_at, and ride_completed_at—were well-formed and parsable, adhering to ISO-compliant formats and valid calendar values. No records required removal, confirming consistent timestamp formatting. Counts therefore remained 61,935 trips and 7320 users. Temporal consistency (df_cleaning_04). We assessed each trip for plausibility by verifying that timestamps fell within the study window (April 22, 2024 to July 24, 2025) and that trip durations were neither unrealistically short (below a minimal operational threshold) nor excessively long (beyond a conservative upper bound). No additional records were removed at this stage. Textual geographic scope (df_cleaning_05). We restricted the dataset to trips whose start_city field was tagged ‘Lausanne’. This step removed 214 trips (0.35%) and 103 users (1.41%), resulting in 61,721 trips and 7217 users. Coordinate validity (df_cleaning_06). We validated the start and end coordinates for numerical range and formatting, ensuring that longitudes fell within [−180, 180] and latitudes within [−90, 90], and that all coordinate pairs in start_longitude, end_longitude, start_latitude, and end_latitude (WGS84) were non-zero. No records were removed at this stage, indicating full syntactic validity of the geospatial fields. Counts therefore remained 61,721 trips and 7217 users. Identical origin and destination (df_cleaning_07). We excluded trips whose departure and arrival coordinates were exactly identical, as these records most likely reflect logging artefacts or negligible repositioning events. Removing these cases avoids inflating near-zero distances. This step removed 184 trips (0.30%) and 15 users (0.21%), leaving 61,537 trips and 7202 users. After these seven steps, the trip-level data frame can be confidently converted into a geospatial object suitable for routing. This allows the cleaned dataset to be spatialized (Supplementary Fig. 1) and provides a solid basis for generating network-constrained itineraries for all subsequent modelling stages.
ErroneousOpenTripPlannerbicycle routes (gdf_cleaning_08). We cross-checked all trips against OpenTripPlanner (OTP) routing results and removed those flagged as implausible, such as clearly infeasible. This step excluded 16 trips (0.03%) and 3 users (0.04%). Additional details on the routing procedure are provided in the subsection ‘Network-constrained route computation using OpenTripPlanner’ (p. 7). Coordinate and location consistency (gdf_cleaning_09). We performed a second check to ensure internal consistency across all geospatial fields and verified that each trip was indeed located within the Lausanne agglomeration boundary, even when start_city was labelled ‘Lausanne’. No additional removals were required, leaving 61,521 trips and 7199 users.
Distance outliers (gdf_cleaning_10). We removed rides shorter than 100 m and rides at or above 10,000 m to discard probable GPS noise, stationary events, or mislogged long-distance trips falling outside the system’s operational profile61,62. This threshold-based step excluded 1992 trips (3.24%) and 221 users (3.07%). The resulting cleaned analysis dataset, gdf_cleaned_df, contains 59,529 trips and 6978 users, corresponding to net reductions of 5.99% and 5.24%.
Network-constrained route computation using OpenTripPlanner
To obtain realistic distances and travel times, we reconstructed cycling routes between recorded origins and destinations using OpenTripPlanner (OTP). Service-reported distances and durations often reflect GPS imprecision and include non-riding components such as unlocking or parking time. OTP routing allows us to generate geometrically valid shortest-path itineraries constrained by the actual street network, providing a consistent basis for subsequent analyses. OTP was deployed locally and configured exclusively in BICYCLE mode using an OpenStreetMap (OSM) extract of Switzerland. Elevation data were deliberately excluded, as slope effects are largely mitigated by electric assistance in shared e-bike systems. Each trip was routed by submitting origin and destination coordinates to the OTP API. Valid routes were stored as LineString geometries along with routed distance and travel time attributes (Fig. 4). Trips for which routing failed were removed at this stage.
Fig. 4: Flow map of all cleaned DBS segments across the metropolitan area.
The alternative text for this image may have been generated using AI.
Flow map of all cleaned DBS trips and access and egress segments across the Lausanne metropolitan area. The figure displays network-constrained flow maps of routed dockless bikesharing (DBS) trips across the Lausanne metropolitan area, with routes aggregated onto road network edges using 10-m buffers and coloured using a linear scale according to the level of trip aggregation, ranging from dark purple (low aggregation, minimum: 1 trip) through red to light yellow (high aggregation, maximum: 5094 trips). The dark polygon delineates the Lausanne city boundary. A The upper panel shows the full cleaned DBS sample of 59,529 trips, routed on a graph of 11,991,485 points and 55,123 edges. The accompanying circular inset provides a zoomed view within a 3-km buffer around the urban centre station. B The lower panel shows the Inside Station Catchment Areas (ISCA) subsample of 16,335 first-mile and last-mile trips, routed on a graph of 3,280,836 points and 40,479 edges. The accompanying circular inset similarly zooms in on the city centre.
OTP-derived distances and durations were compared with service-reported values. Routed distances were on average 9.05% shorter than reported distances, while routed travel times differed by 1.58%. These discrepancies reflect known GPS and reporting artefacts and confirm that OTP provides a reliable approximation of network-constrained travel behaviour.
Extraction and detection of ‘Inside Station Catchment Areas’ segments
From the cleaned dataset, we identified trips potentially connected to public transport using a three-step procedure combining spatial, origin–destination, and temporal criteria (Fig. 5).
Fig. 5: Spatiotemporal detection of ‘Inside Station Catchment Areas’ DBS segments (ISCA) within the dataset.
The alternative text for this image may have been generated using AI.
The diagram illustrates the three-stage filtering pipeline applied to the cleaned dockless bikesharing (DBS) database to identify trips classified as Inside Station Catchment Areas (ISCA) segments. The input database (gdf_cleaned_df) contained 59,529 trips made by 6978 users. Detection 1 applied a spatial filter retaining only trips whose origin or destination fell within the merged 100-m station catchment areas, reducing the dataset to 17,253 trips and 3096 users, corresponding to a removal of 42,227 trips (71.00%) and 3882 users (55.63%). Detection 2 applied an origin–destination matrix filter to ensure that retained trips exhibited a valid intermodal origin–destination pair, resulting in no additional data loss. Detection 3 applied a temporal filter restricting trips to the operational window of train and metro services, removing a further 918 trips (5.32%) and 63 users (2.03%). The final ISCA database (gdf_cleaned_isca_df) comprised 16,335 trips made by 3033 potentially intermodal users, representing an overall reduction of 43,194 trips (72.56%) and 3945 users (56.55%) relative to the cleaned database.
Identifying trips occurring inside influence areas for DBS rental (gdf_isca_01). The first and most decisive exclusion step consisted of defining pedestrian catchment areas around each station entrance and exit, thereby delimiting the ‘socially acceptable’ area within which users are willing to walk to access a DBS. This definition relies on a rule-based spatial threshold to identify trips compatible with rail access or egress, operationalised here using buffers of 100 m or 1 min. These spatiotemporal thresholds are consistent with the radius-based sensitivity analysis conducted by Ju et al.63 across four Californian cities, as well as with other empirical studies employing similar parameters32. Our work also builds on the study by Rieder22, which uses a 60-m walking radius around public transport stations to capture dockless bike and e-scooter trips in Zurich, a threshold they identify as the most commonly applied in the literature.
To ensure that this choice does not drive the results, we assess the sensitivity of key ISCA indicators to alternative cutoffs between 50 and 200 m. Figure 6 reports the relative position of several ISCA indicators with respect to DBS benchmarks across thresholds, including peak-hour incidence, proximity to multimodal hubs, and the share of trips below a 2.4 km reference distance (see the subsection ‘Spatial and temporal distances’, p. 16, for the determination of this threshold). Across this range, these indicators vary only moderately. An auxiliary marginal analysis further shows that trips added when expanding the catchment area beyond this threshold progressively move closer to DBS reference values. We therefore retain this value as the reference threshold in the remainder of the analysis.
Fig. 6: Conditional stability of ISCA identification.
The alternative text for this image may have been generated using AI.
The four panels display the sensitivity of key Inside Station Catchment Areas (ISCA) characteristics to variations in the spatial proximity threshold, ranging from 50 to 200 m. The reference threshold used in the analysis is 100 m. ISCA modal share (%): The proportion of the cleaned DBS sample classified as ISCA trips increases monotonically with the threshold, from 16.29% at 50 m to 49.25% at 200 m, reflecting the mechanical expansion of the catchment area perimeter. ISCA in multimodal hubs (%): the share of ISCA trips associated with multimodal hubs decreases as the threshold increases, from 76.25% at 50 m to 58.59% at 200 m, suggesting that hub-proximate trips are spatially concentrated close to station entries. ISCA during peak hours (%): The share of ISCA trips occurring during peak hours remains stable across all thresholds, ranging narrowly between 42.65 and 43.90%, indicating that the temporal composition of detected intermodal trips is robust to the choice of spatial threshold. ISCA below 2413 m (%): The share of ISCA trips with a routed distance below the 2413-m reference value remains stable across thresholds, varying between 84.07 and 87.01%, confirming that the distance distribution of detected trips is insensitive to moderate changes in the spatial criterion. In all panels, the horizontal axis represents the spatial threshold in metres and the vertical axis represents the share in percentage. Annotated values indicate the exact share at each threshold.
As an original contribution of this paper, we further generated isochrones of 1 min walking distance to better reflect the urban reality shaped by physical barriers and potential detours. The resulting isochrones were then merged for each public transport stop. Among the 119,058 origin and destination points projected within the Lausanne metropolitan area, and given the precision limitations inherent to GPS-based data, we complemented the initial spatial detection (gdf_isca_01) with a spatial point-clustering procedure. The objective was to account for spatial groupings of points in order to compensate for GPS inaccuracies and thus introduce greater spatial tolerance into the detection process. To this end, we applied Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN), an unsupervised clustering algorithm grounded in point density and hierarchical structure, which automatically identifies the most stable e-bike clusters64,65. This step yielded 789 clusters of origin and destination points, which were subsequently overlaid with the merged entrance–exit isochrones of each public transport node (Fig. 7).
Fig. 7: Spatial clustering of DBS trips within Gare de Lausanne, Lausanne-Flon, and Grancy 100-m Station Isochrones.
The alternative text for this image may have been generated using AI.
The map displays the spatial distribution of dockless bikesharing (DBS) origin or destination points around three transit stations in central Lausanne: Gare de Lausanne-Flon (top, multimodal hub), Gare de Lausanne (centre, multimodal hub), and Grancy (bottom, metro station). Purple-filled circles indicate trips classified as Inside Station Catchment Areas (ISCA) trips, whose points fall within the merged station catchment area. Orange-filled circles indicate trips classified as Outside Station Catchment Areas (OSCA) trips, which do not meet the spatial proximity criterion. Blue-filled polygons represent merged isochrones of 100 m around each station entry and exit, derived from the pedestrian network. Blue open circles represent merged circular buffers of 100 m. Black-filled circles indicate individual station entry points. Transit infrastructure visible on the map includes Metro Line 2 (M2), Metro Line 1 (M1), and the Lausanne–Échallens–Bercher regional railway (LEB).
From the multiple isochrones generated for each of the 43 stations under study, we retained only those trip origins and destinations located within their boundaries. This spatial exclusion step eliminated 42,277 DBS trips and 3882 unique users, leaving a total of 17,253 trips and 3096 users (Fig. 5).
Adopting a trip-chain logic (gdf_isca_02). The second criterion used to detect ISCA segments rests on the exclusion of trips that, although spatially located within clusters intersecting the walk-accessible isochrones around rail stations, have both their origin and destination falling within the same observation. In other words, when the start and end points of a shared e-bike trip lie within the same isochrone or within two different isochrones, the trip is excluded as it is highly likely to be non-intermodal and instead to substitute for a metro or train journey. This exclusion criterion did not indicate any such trip in our dataset (Fig. 8). Nevertheless, to the authors’ knowledge, this logic has not been applied in previous research and may prove relevant when implemented on larger datasets. Identifying trips occurring during transit schedules (gdf_isca_03). A further innovative aspect of this potential bike-and-ride trip extraction method consisted of excluding trips that took place outside the operating hours of the public transport system. For this purpose, we examined the Swiss GTFS data (July 21, 2025), retaining only trips occurring between the first morning departure or arrival and the last evening service. In general, this resulted in a continuous service window between 4 AM and 1 AM. By applying this procedure to all 43 stations, we excluded 918 trips that, although spatially connected to station catchment areas, occurred outside the train and metro service window, an exclusion proportion almost identical to that reported by Ju, Martin, and Shaheen63. This corresponded to 63 fewer users attributable to the temporal criterion (Fig. 5). This final filtering step refines the spatiotemporal detection of ‘Inside Station Catchment Areas’ (ISCA) trips, ultimately yielding a subsample of 16,335 trips made by 3033 potentially intermodal users (Fig. 4B).
Fig. 8: Advancing bike-and-ride detection through ISCA spatiotemporal filtering.
The alternative text for this image may have been generated using AI.
The figure contrasts the conventional bike-and-ride spatial detection approach with the proposed Inside Station Catchment Areas (ISCA) spatiotemporal detection method through two schematic diagrams. The upper panel illustrates the existing bike-and-ride detection approach, which relies solely on a circular buffer of 100 m around a station centroid (large light blue filled circle). DBS points falling within the buffer are indiscriminately classified as intermodal regardless of the actual pedestrian network geometry, transit service availability, or origin–destination logic, resulting in the inclusion of spatially proximate but functionally unrelated trips. The lower panel illustrates our own ISCA spatiotemporal detection method, which combines three successive filters applied within a 100-m perimeter. First, a spatial filter restricts eligible trip points to areas covered by network-derived isochrones (star-shaped light blue filled polygon), which reflect true pedestrian accessibility from station entries rather than a uniform radial buffer (light blue open circle). Second, an origin–destination matrix filter excludes trips whose OD pair does not conform to an intermodal logic. Third, a transit service consideration filter removes trips occurring outside the operational hours of train and metro services. Spatial clusters of retained trip endpoints are indicated by purple open circles. Purple-filled circles indicate ISCA trips, while orange-filled circles indicate OSCA trips.
Regression models
To address our final research objective, namely, identifying and quantifying the influence of accessibility, built environment, and sociodemographic factors on shared e-bike usage, we estimate a set of regression models on DBS trips classified as ISCA and OSCA (‘Outside Station Catchment Areas’). These include global specifications (GLM and OLS) as well as GWR models estimated across temporal slices, the latter used as a diagnostic tool to assess potential spatial and temporal heterogeneity. The following subsections present the selection of dependent and independent variables, detail the model specification and diagnostic tests, and finally compare the performance of global and local models.
Three dependent variables are considered, capturing complementary dimensions of shared e-bike usage: Usage intensity (Y1), defined as the number of trips per spatial unit, travel distance (Y2), and travel time (Y3). All variables are aggregated on a regular hexagonal grid (hex100) and standardised prior to estimation (Table 2).
Table 2 Specification of dependent variables Y
Explanatory variables are grouped into three analytical dimensions—accessibility, land use, and sociodemographic characteristics (Table 3)—reflecting complementary determinants of spatial behaviour. Each dimension integrates several thematic categories derived from official statistics and open geodata, allowing us to examine how spatial accessibility, functional land use, and social or material resources are associated with Bird mobility behaviours:
Accessibility (D1), capturing the physical and network-based potential for movement within and towards multimodal nodes, including proximity to stations, road and cycling infrastructure, parking availability, and multimodal services;
Land use (D2), representing the functional and morphological characteristics of urban space that influence trip generation, destination attractiveness, and activity density;
Mobility tools and sociodemographic patterns (D3), summarising resident resources, mobility equipment, travel pass ownership, and household attributes that condition modal choice and trip frequency.
Table 3 Independent variables and associated spatial layers used in the regression models (hex100 and constant across τ)
This variable framework builds on existing research, both regarding the components associated with these three analytical dimensions and mobility behaviour more broadly. For instance, Thao and Ohnmacht66 draw on Swiss public surveys (STATPOP and MTMC) to examine how densities, POIs, transit accessibility, mobility tools, and sociodemographic attributes influence trip frequency and daily distance travelled across modes. Similarly, Beza, Demissie, and Kattan67 investigate the spatiotemporal influence of integrating shared micromobility into first- and last-mile public transport journeys. Their regression model incorporates variables related to cycling infrastructure, the built environment, employment, POIs, and social groups. By applying GWR to explore the links between DBS and public transport, Li, Shang, Zhao, and Yang25 rely on the well-established ‘5Ds’, which are also reflected in our indicators. By contrast, Zhang, Cui, Liu, Jia, Shi, and Yu32 extend this framework by integrating additional variables, such as housing prices or metro station passenger volumes, which are not available in our context.
Before estimating the regression models, we performed diagnostic checks to ensure the stability and interpretability of the coefficients. All predictors were standardised to z-scores to facilitate comparability and avoid scale-dependent effects. We then screened the predictor set for redundant information using pairwise Pearson and Spearman correlations.
Only one pair—distance to Lausanne-Flon (X1B) and to Lausanne-Gare (X1C)—exceeded ∣r∣ > 0.90, reflecting their close geographic proximity (Fig. 9); these were merged into a single measure of central-station accessibility. Variance Inflation Factors (VIFs) were subsequently computed, with all values remaining well below conventional thresholds (max VIF = 2.65), indicating negligible multicollinearity. These diagnostics confirm that the final predictor set is well-conditioned for regression analysis.
Fig. 9: Pearson Correlation Matrix among Independent Indicators.
The alternative text for this image may have been generated using AI.
The matrix displays pairwise Pearson correlation coefficients (r) between all 33 independent indicators considered in the regression models, grouped into three thematic domains along both axes: Accessibility (D1), Land Use (D2), and Social Factors (D3). Each cell is colour-coded according to the correlation value, ranging from dark blue (strong negative correlation, r = −1.00) through white (no correlation, r = 0.00) to dark red (strong positive correlation, r = 1.00). Indicators retained for inclusion in the models are displayed with coloured cells. Indicators excluded due to multicollinearity are displayed with greyed-out cells. The matrix is based on 442 valid pairs, with a mean absolute correlation of \(| \bar{r}| =0.03\), indicating overall low collinearity among retained predictors. Accessibility indicators (D1) include distance to nearest station entrance, distance to Lausanne-Flon, distance to Gare de Lausanne, cycling network density, intersection density, slope, bicycle parking, car parking, bus stops, and public bikesharing stations. Land use indicators (D2) include population, sector 1–3 jobs, local, intermediate, and superior points of interest, and land use classes including residential, public services, commercial, industrial, farmland, and green space. Social factor indicators (D3) include driving licence ownership, motorisation rate, bicycle ownership, e-bike ownership, transit pass, public bikesharing pass, age, household size, and income (X12).
On this basis, we estimate a set of global regression models to quantify the relationships between explanatory variables and shared e-bike usage. For count outcomes (Y1), we estimate a Poisson GLM, while for continuous outcomes (Y2 and Y3), we estimate global OLS models.
To assess whether these relationships vary across space and time, we complement these global specifications with GWR models estimated across temporal slices. GWR is used here as a diagnostic tool to evaluate the extent of spatial and temporal non-stationarity relative to global models.
Temporal-slice GWR estimation relies on spatial and temporal bandwidths controlling the influence of neighbouring observations. Smaller bandwidths enhance local sensitivity, while larger values yield smoother coefficient surfaces. Model calibration is performed separately for five time periods corresponding to daily peaks, off-peak intervals, and weekends.
Based on the general form of a basic GWR model68,69, we define temporal-slice GWR as follows (Formula (1)):
$${y}_{i}={\beta }_{0}({u}_{i},{v}_{i},{\tau }_{i})+{\sum }_{k}{\beta }_{k}({u}_{i},{v}_{i},{\tau }_{i}){x}_{ik}+{\varepsilon }_{i}.$$
(1)
where:
i is the index of an observation located at (ui, vi) and time τi;
yi is the dependent variable at observation i;
xik is the value of the kth independent variable for observation i;
β0(ui, vi, τi) is the local intercept at (ui, vi, τi);
βk(ui, vi, τi) is the local coefficient of the kth predictor at (ui, vi, τi);
εi is the random error term.
The temporal dimension is discretized into five periods τ ∈ τ1, …, τ5 corresponding to morning and evening peaks, midday, night, and weekend intervals (Table 4). Bandwidths are selected through cross-validation by minimising the corrected Akaike Information Criterion (AICc) (Supplementary Fig. 3).
Table 4 Definition of periods τ
We evaluated the performance of global and spatially varying specifications across the three behavioural outcomes (Y1–Y3) and five temporal periods (τ1–τ5), following Hassam, Alpalhão, and Neto70 (Table 5).
Table 5 Model performance comparison between global and local regression models per τ (Y1–Y3)
For Y1, GWR bandwidths ranged from 70 to 100% of active hexagons across periods, indicating convergence toward the global GLM. For Y2 and Y3, bandwidths covered 60–93% and 68–84% of observations, respectively. Although apparent R2 gains are observed (ΔR2 ≈ 0.22), the magnitude of these bandwidths indicates that improvements reflect broad spatial smoothing rather than locally concentrated heterogeneity.
Temporal comparisons indicate moderate variation in model fit across periods, with slightly higher fit during peak periods for Y1, and during off-peak periods for Y2 and Y3. Residual diagnostics further indicate modest reductions in spatial autocorrelation between global and GWR specifications (Supplementary Fig. 3).
Taken together, these results indicate that spatial non-stationarity remains limited in this context. Global GLM and OLS specifications are therefore retained as the primary analytical framework, while GWR results are reported in the Supplementary Information as a diagnostic and sensitivity analysis (Supplementary Tables 4–6).