Source-linked AI summary
Vehicle speed dataset for the major European road network derived from Sentinel-2 imagery, 2022-2026
Maciej Adamiak, Sascha Fendrich, Julian Psotta, Alexander Zipf
TL;DR
Existing speed data are limited by heuristic road-speed assumptions, restricted access, or dependence on reporting vehicles and ground infrastructure. This paper constructs a reproducible Sentinel-2 vehicle-speed dataset and finds that production imagery corroborates its estimated speed uncertainty through heading-residual analysis.
Problem
Mobility analyses lack broadly representative, accessible vehicle-speed measurements beyond heuristic road-type speed limits, limiting realistic travel-speed and travel-time estimation.
Method
The paper transfers moving-echo speed estimation to freely redistributable Sentinel-2 imagery and links displaced band peaks into vehicle trajectories.
Results
A sample of 1,345,000 records has a 6.91° median absolute heading residual, corroborating the reported per-record speed uncertainty.
Takeaways & Limitations
The open-source pipeline and product-level traceability make the dataset independently regenerable and the whole dataset reproducible.
Takeaways & Limitations
The dataset captures only late-morning traffic, not peak-hour conditions, and individual vehicles are not tracked across scenes or acquisition days.
Abstract
from arXiv · showhide
The dataset provides individual vehicle speed observations on European E-roads: motorways, trunk roads, primary and secondary roads, as tagged in OpenStreetMap as e-road, for the years 2022-2026. Speeds are derived from Copernicus Sentinel-2 Level-2A satellite optical imagery using a processing pipeline that exploits the short, well-characterized acquisition delays between the blue (B02_10m), green (B03_10m), and red bands (B04_10m) of the Sentinel-2 push-broom instrument. A moving vehicle appears at slightly displaced positions in the three bands, forming a moving echo. The detected displaced intensity peaks are linked into per-vehicle trajectories through a prediction-and-matching procedure. The resulting displacements are converted into ground speeds using publicly accessible inter-band time delays. Each record contains the trajectory geometry, per-channel displacements and headings, internal quality indicators, the estimated speed, the acquisition timestamp, and the source Sentinel-2 product identifier. The dataset is distributed as GeoPackage files, with one record per detected vehicle, and can support studies of traffic patterns, speed behavior, transport modeling, and the calibration of road network attributes at a continental scale.
Background & Summary
The section motivates vehicle-speed data for mobility analysis, explains Sentinel-2’s inter-band timing as the basis for detecting vehicle motion, and introduces an open, reproducible dataset generated with NibbleRGB from OSM and Sentinel-2 imagery.
- Motivation: Vehicle speed measurements underpin mobility analysis, but estimating realistic travel speeds and travel times from linked traffic data remains unresolved.Applications include infrastructure planning, transport modeling, road safety analysis, road traffic management, and human…
- Motivation: Continuously collected floating-car, navigation-device, and connected-vehicle data support temporal traffic analyses but can misrepresent conditions when reporting coverage is low or skewed.Access to these data is typically restricted by commercial licensing.
- Measurement principle: Sentinel-2’s separate detector lines image the same ground location at slightly different times, so moving objects appear at displaced positions across spectral bands.The interval between band acquisitions is a fraction of a second, enabling motion detection for fast-moving objects.
- Prior work: The authors’ earlier deep-learning method detected vehicle moving echoes and estimated velocities in PlanetScope imagery, but relied on commercially licensed imagery.That method was validated against drone video and GPS trajectories and supported traffic-flow analysis and modeling.
- Dataset production: NibbleRGB combines OpenStreetMap road geometries with Copernicus Sentinel-2 Level-2A surface-reflectance imagery to produce the dataset.The passage defines OSM road classes including motorways, trunk roads, and primary roads.
- Reproducibility: Freely available Copernicus imagery and open-source processing code make each record traceable to a Sentinel-2 product, independently regenerable, and reproducible as a whole.Related resources include Sentinel-2 truck detections, generic motion detection, and the authors’ PlanetScope speed estimates.
Methods
The dataset was generated with NibbleRGB 1.0.0 through a four-stage pipeline combining OpenStreetMap road retrieval, Sentinel-2 imagery processing, vehicle-echo trajectory construction, and speed estimation. The released GitLab source code directly implements the described processing steps.
- Pipeline: NibbleRGB 1.0.0 implements four stages: road retrieval, Sentinel-2 imagery retrieval, vehicle-echo detection and trajectory linkage, and displacement, heading, and speed estimation.The package was developed by the authors, and all described processing steps correspond directly to the released GitLab source code.
- Road network: Road geometries were retrieved from OpenStreetMap through API 0.6 relation endpoints using unique relation identifiers.The repository includes configurations for individual roads, regions, and the identifiers used for the dataset period.
- Sentinel-2 imagery: Sentinel-2 Level-2A surface-reflectance scenes intersecting each road segment’s bounding box were queried from two STAC catalogs between 01.06.2022 and 30.06.2026.The sources were the Copernicus Data Space Ecosystem and Element84 Earth Search.
- Speed conversion: The speed conversion used inter-band delays of 0.527 s for B02–B03 and 0.478 s for B03–B04, yielding a 1.005 s observation window per scene.These were mean temporal offsets of the Sentinel-2 multispectral instrument and may exhibit static and dynamic variation.
- Echo detection and trajectories: Trajectory construction seeded B02–B03 peak pairs separated by 0.1–3.0 px, predicted B04 positions with gap ratio 0.907, and accepted matches within a 2 px radius and 35° turn-angle limit.Candidates were scored by summed prediction error with a turn penalty, sorted by cost, and accepted greedily without reusing detected peaks.
Data Records
The dataset is hosted on Zenodo as 7.35 GB of GeoPackage files, with yearly vehicle-trajectory records supported by road centerlines, computation units, and production configuration files.
- Dataset organization: 7.35 GB of GeoPackage files are hosted on Zenodo and are compatible with common GIS software and libraries.GeoPackage is an open, SQLite-based geospatial format compatible with QGIS, geopandas, and GDAL.
- Dataset organization: The yearly eu_kinematics_<year>.gpkg files contain trajectory geometry and data totaling 22188448 records.There is one kinematics file per year.
- Dataset organization: The eu_centerlines.gpkg file contains estimated road centerlines totaling 44490 records, while eu_segments.gpkg contains 44488 computation work-unit boundaries.These files provide road-centerline and spatial-segmentation data for the dataset.
- Dataset organization: The deposit includes default.yaml as the NibbleRGB production baseline and eu_<year>.yaml files containing yearly production configuration and OpenStreetMap relation identifiers.The configuration files accompany the GeoPackage data in the Zenodo deposit.
Technical Validation
Technical validation combines timing-model characterization, deterministic pipeline tests, per-record quality indicators, and plausibility checks of speed distributions and headings. Results quantify timing perturbations, estimator limitations, and close agreement between detected headings and road-centerline bearings.
- Validation framework: Validation covers timing-model characterization, deterministic processing-code tests, internal record-level quality indicators, and plausibility checks of speed distributions.These four elements form the stated validation framework.
- Timing model: Fixed inter-band offsets of 0.527 s and 0.478 s total 1.005 s, while optical distortion and satellite altitude contribute at most ±2% and attitude adds 0.15% peak-to-peak modulation.Speed is computed from trajectory ground displacement divided by the fixed mean inter-band offsets.
- Unit tests: 6 test suites containing 98 tests achieve 91% coverage across speed computation, echo detection, and road segmentation functionality.End-to-end echo detection is checked against a reference Sentinel-2 image with a known number of echoes, and road segmentation against a reference road relation and area of interest.
- Internal quality indicators: Every record carries three internal indicators that support conservative subset selection without reprocessing imagery.score measures accumulated trajectory prediction error, while std_shift measures disagreement between per-gap displacements and flags violations of the constant-speed assumption within the 1.005 s window.
- Plausibility and coverage checks: The 0.1–3.0 px displacement bounds detect approximately 7 km/h to 205 km/h, remain effectively unbiased up to about 130 km/h, and above approximately 160 km/h can reject 9–84% of vehicles and bias survivors low by up to 25 km/h.Low-speed estimates are overestimated by approximately +9 to +46 km/h at and below 40 km/h because localization noise lengthens the summed path.
- Heading agreement with the road centerline: Across 1345000 sampled records, the median absolute heading residual is 6.91°, with 1σ dispersion of 10.24°, decreasing from 16.49° below 40 km/h to 8.18° above 100 km/h.The observed dispersion yields effective localization noise of σp ≈0.24px and corroborates the simulated behavior.
Usage Notes
Each record is an independent, single-scene observation of one vehicle, so vehicles are not tracked across scenes or acquisition days. Tile-grid overlap can cause the same vehicle to appear in different scenes, while temporal aggregation of overlapping trajectories may support approach traffic-volume estimation.
- Usage Notes: Each record represents one independent observation of a single vehicle in one scene.The record does not represent a vehicle tracked across multiple observations.
- Usage Notes: The same physical vehicle is not tracked across scenes or imagery acquisition days.This limits interpreting records as continuous vehicle histories.
- Usage Notes: Sentinel-2 tile-grid overlap can place the same vehicle in different scenes.Duplicate appearances may therefore occur across overlapping scenes.
- Usage Notes: Aggregating overlapping trajectories over time could support approach traffic-volume estimation.The passage presents this aggregation as a feasible approach rather than a completed estimate.
Code Availability
NibbleRGB version 1.0.1 is openly available under the AGPL-3.0 license with production configurations, tests, and reference data, and can run on any internet-connected machine.
- Code Availability: NibbleRGB version 1.0.1 is available under the AGPL-3.0 license with production configuration files, a test suite, and reference test data.The source code is hosted at the stated GitLab repository.
- Code Availability: The code runs on any machine with internet access without additional setup, although deployment near the Sentinel-2 data catalog is preferred.The preferred AWS region is us-west-2 (Oregon).
Funding
The work was supported by core funding from the Klaus Tschira Stiftung (KTS, Germany), without a specific grant number.
- Core funding came from the Klaus Tschira Stiftung (KTS, Germany), and no specific grant number applies.