Calibrated multi-modal transport networks for Switzerland

Abstract

Cleaned OpenStreetMap-derived walk, bike, and car networks for Switzerland (plus a cross-border buffer for realistic frontier-cell routing), with per-edge calibrated travel durations. These are the networks used to compute the companion accessibility metrics dataset: https://doi.org/10.5281/zenodo.21410968. Contents 1) Networks in .graphml representation (read e.g. using python OSMnx) and in node/edge .gpkg representation.2) Per-edge calibrated durations for every routing profile in the companion dataset (.csv): Mode Profile ID Description Walk rwalk Average pedestrian Walk walk_prm Person with reduced mobility — slower baseline, higher slope penalty, hard-exclusion of steps Bike rbike Regular bicycle Bike ebike25 E-bike, 25 km/h class (pedelec) Bike ebike45 E-bike, 45 km/h class (s-pedelec) Car car_night Late-evening / night traffic Car car_base Baseline daytime hours Car car_peak Peak traffic hours (Mon–Fri, 7:30–8:30 and 15:30–18:30) Usage To attach calibrated durations to the network, join edges_ _calibrated.csv on the edge_id column (format u:v:key) into whichever representation you prefer: networkx — load with nx.read_graphml(...) (or ox.load_graphml(...)), split each edge_id back into (u, v, key), and apply with nx.set_edge_attributes. GeoPandas / QGIS — load _edges.gpkg and merge (or add a table join) on edge_id. Attach only the profiles you need — each duration_calibrated_ column is independent, so a car-peak-only analysis doesn't need the other columns loaded. Methodology Networks extracted from OSM via the aperta-atlas preparation pipeline (PBF clipping, degree-2 chain collapse, snap-eligibility flagging, per-edge attribute decoration). Per-edge durations for walk, bike, and each car profile are fitted against Swiss travel survey (MTMC 2015/2021) and MOBIS (https://mobis.ethz.ch) trips. Calibration effectiveness The improvement over raw / imputed OSM speed limits (for car travel) and a fixed default speed (for walking / cycling), named "uncalibrated", is substantial: Mode Qualifier Comparison Trip distance Mean biasUncalibrated Mean biasCalibrated R²Uncalibrated R²Calibrated Walk Typical MTMC (door-to-door) all -20.9% +6.5% 0.602 0.657 Bike Typical MTMC (door-to-door) all -33.9% +2.9% 0.579 0.697 E-Bike 25 km/h MTMC (door-to-door) all -37.7% -6.6% 0.512 0.667 Car Night MTMC (door-to-door) all -39.8% +0.4% 0.527 0.726 MOBIS (node-to-node) all -32.0% -7.7% 0.559 0.706 MOBIS (node-to-node) < 5 km -41.0% -12.7% -0.002 0.370 MOBIS (node-to-node) > 25 km -22.8% -2.5% 0.294 0.484 Peak hours MTMC (door-to-door) all -47.4% +3.3% 0.429 0.738 MOBIS (node-to-node) all -45.2% -9.7% 0.383 0.733 MOBIS (node-to-node) < 5 km -51.4% -11.4% -0.122 0.441 MOBIS (node-to-node) > 25 km -34.9% -3.5% -0.126 0.557 R² is the coefficient of determination relative to the 1:1 line (predicted = observed): R² = 1 − Σ(predicted − observed)² / Σ(observed − mean observed)². Unlike the squared Pearson correlation, it also penalises systematic bias and scale errors, so it measures how well the predicted times can be used as they are, not only whether they correlate with observed times. R² = 1 is a perfect fit; R² < 0 means the prediction is worse than simply returning the mean observed travel time. R² is not expected to approach 1: individual travel times vary (route choice, driving style, traffic, stops) in ways that even a perfect, unbiased estimate of expected travel time cannot capture. Within a distance band, the spread of observed times is smaller, so R² is lower there than over all trips for the same prediction quality. In addition, MTMC times are self-reported and rounded, mostly to 5 minutes, which caps the attainable R² for these trips regardless of model quality. Reproduction Produced with the aperta accessibility library (https://github.com/mmiotti/aperta) via the aperta-atlas pipeline (https://github.com/mmiotti/aperta-atlas), scenario switzerland-default. Attribution and license This dataset is a Derivative Database of OpenStreetMap and is released under the Open Database License (ODbL). Contains data © OpenStreetMap contributors; see https://www.openstreetmap.org/copyright.