National-Scale Multimodal Accessibility Fields from Open Geospatial Data

1 Introduction

• MAPPER combines OSM- and GTFS-derived multimodal travel-time graphs, harmonized opportunity layers, NKDE fields, and H3 sampling in an open-data pipeline.The workflow covers walk, bike, car, and public transport graphs and supports comparative assessment across modes.Source: §1 · p. 2

1.1 Contributions

• A native compute core combines contraction-hierarchy routing, C++/OpenMP accumulation, and DAG-style orchestration to make national-scale runs feasible on commodity hardware.The paper positions this architecture as enabling many graphs and a broad opportunity-layer set jointly.Source: §1.1 · p. 2
• The workflow generates continuous NKDE fields for more than 100 opportunity categories and exports dense, multimodal H3 feature stacks for reuse in urban and regional analytics.The output representation is intended for comparative diagnostics and planning applications such as service-provision assessment.Source: §1.1 · p. 3
• Germany-wide feasibility and routing diagnostics are demonstrated alongside a Bavaria case study that applies the state’s central-place planning travel-time targets to the resulting fields.Routing evaluation is treated as a check on impedance realism, not complete validation of accessibility inference.Source: §1 · p. 2, §1.1 · p. 3

2.1 Accessibility measures and operational trade-offs

• Cumulative-opportunity, gravity-style, and floating-catchment measures trade interpretability and continuous decay against catchment, threshold, and zonal sensitivities.MAPPER prioritizes continuous, place-based gravity-style surfaces over person-specific or capacity-equilibrium formulations.Source: §2.1 · p. 3

2.2 Network-constrained continuous accessibility fields

• Network kernel density estimation constructs continuous fields that respect network reachability, with results sensitive to bandwidth and kernel choices.Its continuous formulation allows thresholds and scenarios to be applied after field computation rather than built into the measure beforehand.Source: §2.2 · p. 4

2.3 Scale, multimodality, and reusable outputs

• Existing large-area products demonstrate gridded travel-time analysis, but often provide narrower indicator sets than broad, continuous multimodal opportunity fields.Hierarchical DGGS grids offer common multi-resolution indexing, although interoperability and full standards compliance remain incomplete in current implementations.Source: §2.3 · pp. 4–5

2.4 Open geospatial opportunity data and schema design

• OpenStreetMap POIs and building data support large-scale opportunity estimation, but noisy, heterogeneous categories require transparent harmonization and plausibility checks.Prior work also reports category-level variation in POI quality and completeness and remaining challenges in multi-activity assignment and regional data quality.Source: §2.4 · p. 5

2.5 Gap addressed

• The unmet infrastructure need spans multimodal travel-time graphs, harmonized opportunity inventories, continuous network fields, and reusable outputs for spatial modeling.Existing products rarely combine all four at country scale in an open-data pipeline oriented to urban and regional analysis.Source: §2.5 · p. 5
• Reproducibility must extend to routing and measure computation because tool internals can change accessibility results; MAPPER targets this computational gap on commodity hardware.The paper is positioned as a geocomputational workflow, not a digital twin, real-time system, or single-purpose case study model.Source: §2.5 · pp. 5–6

3 Data

• The Germany-wide production run contains 39,992,878 H3 cells and 972 accessibility columns, representing 108 opportunity layers across nine network and intermodal modes.Outputs are sampled to H3 resolution 10; the representative Parquet table occupies about 66 GB.Source: §3 · p. 6

3.1 Transport networks and temporal scope

• Walking, cycling, and car networks use Geofabrik’s Germany OSM extract, while public transport uses the Germany Full GTFS feed derived from DELFI NeTEx.Public transport layers represent scheduled Monday departures from 07:00 to 09:00 as a favorable planning baseline, not full-day or real-time service quality.Source: §3.1 · p. 6

3.2 Opportunity layers and harmonization

• OSM tags across amenity, shop, tourism, leisure, healthcare, office, and craft are harmonized into standardized opportunity classes rather than retaining the full tag vocabulary.Examples include merging doctor-related tags into doctor and pharmacy-related tags into pharmacy; some noisy or transport-infrastructure tags are excluded.Source: §3.2 · p. 7

3.3 Workplace proxies and census population

• Workplace layers approximate employment capacity from mapped usage, floor area, and activity type; 2022 Census population weights resident exposure and municipal fulfillment.Workplace proxies are not official employment counts, and the OSM taxonomy is an analytical harmonization rather than a complete or normatively correct service classification.Source: §3.3 · p. 7

4.1 Pipeline overview

• MAPPER combines open transport and opportunity data, computes continuous accessibility fields on transport networks, and samples node-level results to H3 cells for downstream analysis.The pipeline has three stages: ingest and harmonize inputs, compute fields using travel time as impedance, then sample and export the results.Source: §4.1 · p. 8, Figure 1 · p. 8

4.2 Graph construction and source harmonization

• Walk, bike, car, and scheduled public transport graphs use mode-specific travel times, while harmonized opportunity sources keep source placement fixed across modes.Car times are free-flow estimates without congestion; public transport times omit initial waiting before the first boarded service and depend on timetable and transfer assumptions.Source: §4.2 · p. 9

4.3 Accessibility model

• Accessibility is a network-based potential that weights opportunities by shortest-path travel time, preserving barriers and network topology that Euclidean smoothing can blur.The Gaussian kernel bandwidth is fixed at 850 s, balancing local contrast against fragmented fields and excessive metropolitan spillover.Source: Figure 2 · p. 9, §4.3 · p. 10

4.4 H3 sampling and output structure

• Node-level values are assigned to resolution-10 H3 cells and exported as mode- and layer-indexed feature tables with run metadata for reuse.The sampler uses in-cell graph nodes first, then falls back to a graph node within 100 m or an eligible edge within 150 m; unresolved cells receive a configured default.Source: §4.4 · pp. 10–11, Figure 3 · p. 11, Figure 4 · p. 11

• The validation targets travel-time impedance realism rather than validating downstream accessibility field values.The stated rationale is that systematic travel-time offsets translate directly into shifts in the accessibility field.Source: §5 · pp. 11–12
• Predicted travel times are compared with observed trip durations and mode information from smartphone-based travel-diary studies in Germany.Trip origins and destinations are snapped to the corresponding mode graphs using production-workflow rules.Source: §5 · p. 12
• Bland–Altman plots assess mode-specific bias across observed trip durations, while aggregation analysis examines mean error for pooled accessibility surfaces.The aggregation diagnostic reflects MAPPER’s intended use for pooling information across opportunities rather than predicting individual route choice.Source: §5 · p. 12

6.1 Travel-time evaluation

• Walking, cycling, and car show small mean bias, while public transport has negative bias because predictions exclude initial waiting included in observed door-to-door durations.The public-transport pattern is interpreted as an optimistic upper-bound effect under modeled timetable conditions, not high timetable-routing fidelity.Source: §6.1 · p. 12, Figure 5 · p. 12

6.2 Workplace proxy plausibility

• The OSM-derived workplace model recovers 91% of total employment, with ratios of 1.08 for industry and manufacturing, 0.85 for services, and 0.68 for agriculture.The authors treat this as an order-of-magnitude plausibility check, not full validation, given expected OSM coverage biases and classification gaps.Source: §6.2 · p. 13

6.3 Bavaria case study operationalization

• The Bavaria case study evaluates LEP service bundles by thresholding continuous NKDE fields, translating policy travel-time targets into activation thresholds rather than hard routing cutoffs.Population-weighted mean scores aggregate subcategory fulfillment across inhabited H3 cells while preserving cross-border opportunity influence.Source: §6.3 · p. 14
• Under T1, only 4.4% of municipalities and 10.2% of the population meet the basic LEP provisioning bundle on foot, while car-based fulfillment is near-universal.Walking remains the limiting mode across the three thresholds, and the modal ordering is stable as overall coverage rises under T2 and T3.Source: §6.3 · p. 16, Table 1 · p. 16
• At T1, walking exposure to undersupply rises from 46.5% under the basic unmet interpretation to 68.7% under municipality-specific target requirements.Public-transport exposure falls from 6.5–6.6% at T1 to around 1.6–3.6% at T2–T3; walking remains an outlier at T2.Source: §6.3 · pp. 16–17, Table 2 · p. 17
• For bike, car, and public transport, the difference between basic unmet and target unmet is much smaller than for walking.The comparison uses the same accessibility field to assess alternative interpretations of service adequacy.Source: §6.3 · pp. 16–17

What the case study shows

• Across thresholds, walking is most restrictive, car least restrictive, and cycling and public transport intermediate; planned service levels are easier to reach by car than by sustainable modes.The size of the gap varies across space, supporting comparison rather than an absolute judgment that the central-place system is working or failing.Source: §7 · p. 17

What MAPPER enables

• Continuous, harmonized multimodal fields support map diagnostics, municipal and population-weighted summaries, alternative planning thresholds, and downstream statistical learning without rebuilding the data foundation.The aligned output structure also supports scenario-oriented geospatial modeling on a common spatial support.Source: §7 · pp. 17–18

Uncertainty and limits of interpretation

• Bavaria’s results classify continuous accessibility fields using NKDE thresholds rather than testing direct nearest-destination compliance, and depend on bandwidth, threshold, and opportunity-schema choices.Public transport estimates are optimistic and schedule-based, car accessibility may be overstated without congestion, and the harmonized open-data schema is not an error-free service census.Source: §7 · p. 18

Conclusion

• MAPPER combines OSM- and GTFS-derived travel-time graphs, harmonized opportunity layers, NKDE inference, and H3-aligned outputs into a reusable accessibility feature stack.The framework is intended for urban and regional analysis tasks rather than rebuilding data for each case study.Source: §8 · p. 18
• The Germany-wide run demonstrates country-scale computational feasibility, and the Bavaria application supports comparative planning diagnostics across modes, thresholds, and summary representations.Routing validation supports routed graphs as an impedance baseline for aggregated accessibility analysis; a workplace plausibility check provides an additional scale check for one proxy input family.Source: §8 · p. 18
• Bavaria’s interpretation remains conditional on timetable assumptions, threshold choices, schema decisions, and the absence of congestion; future work includes robustness and temporal diagnostics.The conclusion also identifies land-use and service-planning scenarios, and stronger integration with scenario and optimization workflows, as future directions.Source: §8 · p. 18

Code and data availability

• MAPPER source code is openly available under Apache License 2.0 as release v2.0.0, with configuration examples, installation instructions, and reproduction materials included.The data are described in Section 3 and Appendix A sufficiently to identify and obtain the same inputs.Source: Code and data availability · pp. 18–19

A Selected Workflow Parameters

• Table A.1 documents production settings for input selection, graph construction, public-transport routing, NKDE, H3 sampling, and output generation to support reproducibility.The table includes selected parameters for both the production pipeline and the Bavaria case study.Source: A Selected Workflow Parameters · p. 23, Table A.1 · pp. 23–24

B OSM Opportunity Schema

• Raw OSM tags are grouped into standardized classes to stabilize noisy taxonomies; priority rules and selected exclusions shape the final opportunity inventory.Tables B.1 and B.2 summarize schema families and representative class mappings; the harmonization is not intended to reproduce the full OSM ontology.Source: B OSM Opportunity Schema · p. 24, Table B.1 · p. 25, Table B.2 · p. 25

C Workplace Proxy Plausibility Check

• The workplace proxy check aggregates OSM-derived estimates and official Bavarian employment totals into broad WZ2008 sector groups.The comparison is used to assess whether proxy layers are strongly mis-scaled before accessibility analysis.Source: Table C.1 · p. 26, C Workplace Proxy Plausibility Check · p. 26
• Proxy deviations are largest for agriculture, while industry is close to parity and services are somewhat underestimated, supporting cautious use as opportunity fields rather than calibrated employment counts.Greenhouse footprints incompletely proxy agricultural employment; mixed-use or generically tagged buildings contribute to service underestimation.Source: C Workplace Proxy Plausibility Check · p. 26, Table C.1 · p. 26

D LEP Threshold Translation for the Bavaria Case Study

• The appendix lists LEP travel-time targets and their mapped NKDE activation thresholds, which are used only in the Bavaria case study, not core MAPPER field generation.The values are derived using the policy-activation mapping defined in the main text.Source: D LEP Threshold Translation for the… · p. 26, Table D.1 · p. 27

E Threshold Sensitivity of Population Exposure

• Walking remains the most restrictive mode, car the least restrictive, and cycling and public transport occupy intermediate positions across the threshold range.The comparison is stable across the continuous threshold spectrum.Source: E Threshold Sensitivity of Population… · p. 27
• The target-unmet curves are generally above the basic-unmet curves, especially for walking at high thresholds.Target-unmet evaluates cells against the score implied by the municipality’s assigned central-place level; basic-unmet uses the basic-level score for all cells.Source: E Threshold Sensitivity of Population… · p. 27, Figure E.1 · p. 28
• T3 = 0.30, T2 = 0.60, and T1 = 0.90 are scenario thresholds on a continuous spectrum, not uniquely privileged cut-offs.The figure plots classification threshold on the x-axis and population below threshold (%) on the y-axis.Source: E Threshold Sensitivity of Population… · p. 27, Figure E.1 · p. 28

F Kernel Sensitivity of the Central-Place Fulfillment

• Near-universal upper-level fulfillment for car and public transport persists under the exponential kernel, supporting broad motorised reachability rather than a Gaussian-threshold artefact.Walking remains by far the most restrictive mode, while motorised and cycling access is broadly available and coverage rises from the strict to the lenient scenario.Source: F Kernel Sensitivity of the… · p. 28
• The 90 min public-transport target maps to θ = 4.5 × 10−3 and the 60 min car target to θ = 2.7 × 10−2 under the exponential kernel.These thresholds are six and two orders of magnitude above their Gaussian counterparts, respectively.Source: F Kernel Sensitivity of the… · p. 28, Table F.1 · p. 29
• The exponential tail can saturate fulfillment tests for high-catchment modes, whereas the faster-decaying Gaussian preserves more granularity and is retained for the main analysis.For a representative Bavarian cell, median non-zero supermarket fields were ≈7.5 by public transport and ≈118 by car, against thresholds of 0.17 and 0.30, respectively.Source: F Kernel Sensitivity of the… · p. 29