Vehicle speed dataset for the major European road network derived from Sentinel-2 imagery, 2022-2026
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Vehicle speed dataset for the major European road network derived from Sentinel-2 imagery, 2022-2026".
Jane: Here's the gist of the paper. 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.
Tom: Data Acquisition Methodology The 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.
Jane: A moving vehicle appears at slightly displaced positions in the three bands, forming a moving echo.
Tom: The detected displaced intensity peaks are linked into per-vehicle trajectories through a prediction-and-matching procedure.
Paper summary: Tom: So, to wrap up, the paper "Vehicle speed dataset for the major European road network derived from Sentinel-two imagery, two thousand twenty-two-two thousand twenty-six" describes a system that uses satellite imagery to capture individual vehicle speeds on major European roads over four years> The authors show that by exploiting the time differences between blue, green, and red bands of Sentinel-two data, they can link moving echoes into trajectories and estimate ground speeds> <ref:2608.22116#pg1,Vehicle speed dataset for the major European road network derived from Sentinel>
Jane: It’s essentially providing a way to get real-time traffic information at a continental scale without needing sensors on every single stretch of road or relying on drivers to report their travel> The authors are making the case that this satellite remote sensing method offers an independent observation channel for speed data>
Lu: What they really highlight is the ability to do temporal analyses, like looking at congestion patterns because the road networks they use are collected continuously throughout those four years> They also discuss how you can use this data for calibrating attributes of the road network itself>
Meng: From a deployment standpoint, I still worry about the noise mentioned in their validation section. They mention that speed uncertainty can be between thirteen kilometers per hour and twenty five kilometers per hour depending on the localization noise> That’s a significant margin if you need high precision for something like autonomous vehicle path planning>
Lalam: The authors also state that even with this uncertainty, the estimates are effectively unbiased by the filter up to about one hundred thirty kilometers per hour> So, at high speeds, the estimation seems pretty solid despite the noise in the system>
Tom: It’s a useful dataset because it bridges a gap between ground-based sensor limitations and raw satellite data> It gives us a large sample of speed behavior across different road types like motorways and trunk roads over four years>
Jane: The implication is that we can build much more accurate models for transport flow than we could with just traditional surveys or even with driver-reported data alone> This moves us toward understanding traffic patterns in a way that’s truly continental.
Conclusion: Tom: So, we’re wrapping up our look at this new dataset from the paper "Vehicle speed dataset for the major European road network derived from Sentinel-two imagery, two thousand twenty-two to two thousand twenty-six."
Jane: It basically takes satellite pictures from those Sentinel-two satellites and turns them into a way to measure how fast vehicles are going on big roads across Europe over those four years.
Tom: Exactly. The authors are using the specific timing differences between the blue, green, and red bands of that imagery to track moving cars and turn those visual echoes into actual speed measurements for every single vehicle they find.
Jane: That’s a pretty clever way to do it without needing cameras on every bridge or highway exit. So what does this mean for us?
Tom: It means we have a much bigger, continuous record of traffic patterns than we ever had before, covering motorways, trunk roads, and all those other major European routes.
Jane: And the authors stress that they’ve done a lot of checks on it. They looked at things like timing models and internal quality scores to make sure the speeds aren't just random noise.
Tom: Yeah, they found that the speed uncertainty can be between thirteen kilometers per hour and twenty five kilometers per hour depending on how noisy the location is. But up to about one hundred thirty kilometers per hour, those estimates seem pretty solid.
Jane: That’s a big thing because it gives us a real way to study traffic behavior across different road types over time, which could really help in building better transport models.
Tom: It’s useful for calibration too; you can use this data to check if your maps and road network attributes are accurate at a continental scale.
Jane: So we have this new source for speed data that is derived from existing satellite imagery, and it’s been collected over a significant time period across all these major European roads.
Tom: Right. And the next thing we need to look at is how they actually built that echo detection pipeline—how they link those moving peaks into smooth trajectories.
Maciej Adamiak, Sascha Fendrich, Julian Psotta, Alexander Zipf
Heidelberg Institute for Geoinformation Technology Heidelberg Institute for Geoinformation Technology University of Heidelberg Faculty of Geographical Sciences University of Lodz
cs.CV
Submitted: 2026-08-22
Updated: 2026-10-31
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 90/100
The gist: 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.
Key concepts
- Sentinel-2 Level-2A Imagery
- This is high-quality satellite data from the Copernicus program, used as the input image source. It provides surface reflectance information across different spectral bands (blue, green, red) of the Earth's surface. This imagery is processed to capture how light interacts with surfaces like roads and vehicles.
- Echo Detection Pipeline
- This is a multi-stage process that finds moving vehicles by looking for 'echoes'—slight shifts in intensity across the three Sentinel-2 bands. The pipeline segments road geometry, detects these echoes, links them into continuous vehicle trajectories using prediction and matching, and finally converts those pixel movements into estimated ground speeds.
- Per-record Quality Indicators
- These are internal metrics used to assess the reliability of each detected speed measurement. Examples include 'score' (how much the prediction error accumulated), 'std_shift' (disagreement between successive displacement measurements), and 'bearing_consistency' (checking if the calculated directions are consistent). These indicators help determine if a speed estimate is trustworthy.
- Displacement Range and Speed Estimates
- The algorithm restricts detection to pixel displacements between 0.1 and 3.0 pixels per gap, which corresponds to speeds roughly between 7 km/h and 205 km/h. The study found that the speed estimates are generally unbiased up to about 130 km/h, providing a reliable range for traffic analysis.
Terminology
Summary
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. p. 0 Vehicle speed dataset for the major European road network
derived from Sentinel-2 imagery, 2022–2026
Maciej Adamiak ∗1,3, Sascha Fendrich1
, Julian Psotta1,2, and Alexander Zipf1,2
1Heidelberg Institute for Geoinformation Technology (HeiGIT), Germany
2GIScience Research Group, Institute of Geography, University of Heidelberg, Germany
3Faculty of Geographical Sciences, University of Lodz, Poland
Abstract
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.Where the paper says it
Data Acquisition Methodology
The 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. p. 0 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. 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 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. The resulting displacements are converted into ground speeds
using publicly accessible inter-band time delays. Each record contains the trajectory geometry, perchannel displacements and headings, internal quality indicators, the estimated speed, the acquisition
timestamp, and the source Sentinel-2 product identifier. Each record contains the trajectory geometry, perchannel 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.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Data Sources and Network Definition
The input data for the dataset involves two primary components: OpenStreetMap road network geometries and Sentinel-2 Level-2A surface reflectance imagery. Road geometries are retrieved via the OSM API version 0. p. 2 Road geometries (Fig. 1) are retrieved via the OSM API version
0. Additionally, relation member ways
are preserved only if they have one of the tags motorway, motorway_link, trunk, trunk_link,
primary, primary_link, secondary, or secondary_link. Such an approach ensures that the dataset
contains only major roads with a total length of 209 437 km. For applications that require conservative subsets, filtering on
score, std_shift, and bearing_consistency is recommended. Data Availability
The dataset is hosted at Zenodo [14] and conforms to the following licensing rules: (i) The kinematics
GPKG file is available under the CC BY license as derived from Copernicus imagery; (ii) The GPKG
files created from road geometries (segments and centerlines) are available from OpenStreetMap
under the Open Database License (ODbL); (iii) The referenced Sentinel-2 Level-2A products are
freely available from the Copernicus Data Space Ecosystem. In OpenStreetMap nomenclature, a motorway is a restricted-access divided highway, a trunk road is the
most important non-motorway road in a country’s network, a primary road is the next tier of major
road linking large towns, and a secondary road is one that is not part of the major routes but forms
the national road network [9]. The dataset covers these four classes, including their associated link
roads such as slip roads and ramps.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Echo Detection Pipeline
The core of the pipeline is the echo detection algorithm which operates on the three-band image stack from one scene clipped to a single road segment. This process involves four stages: (i) retrieval and segmentation of road geometries from OpenStreetMap, (ii) retrieval of Sentinel-2 Level-2A imagery clipped to those road segments, (iii) detection of vehicle echoes and their linkage into per-vehicle trajectories, and (iv) conversion of trajectories into displacement, heading, and speed estimates. p. 1 0, a Python package developed by the
authors. The pipeline consists of four stages: (i) retrieval and segmentation of road geometries
from OpenStreetMap; (ii) retrieval of Sentinel-2 Level-2A imagery clipped to those road segments;
(iii) detection of vehicle echoes and their linkage into per-vehicle trajectories; and (iv) conversion of
2 The analysis is restricted to
the pixels of the road segment that carry valid (non-zero) data in all three bands, so that peaks are
never sought where any band lacks coverage. For each band, the steps are as follows: (i) Scale the
band to 0–1 using the band’s minimum and maximum pixel values; (ii) Remove the background
- a white top-hat transform with a disk-shaped structuring element of radius 1 pixel suppresses the
road surface, retaining only small bright objects; (iii) The filtered band is then renormalized by
its new maximum over the valid pixels; (iv) Find the interior - the valid mask is eroded by
the same disk to obtain the interior pixels, away from the edges. For the B04 band, the position is predicted by advancing the current
trajectory tip along the heading, scaled by the gap ratio of that step. The gap ratio is approximately
0. Each completed candidate is assigned a matching cost equal
to the sum of its prediction errors in pixels, modified by a turn penalty; the minimum cost is 0, so a
low cost indicates a smooth, predictable trajectory. Candidate trajectories are sorted in ascending
order of cost and accepted greedily, subject to the constraint that each detected peak may belong
to at most one trajectory.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Validation and Quality Indicators
Technical validation rests on four elements: the characterization of the timing model, deterministic tests of the data processing code, internal per-record quality indicators, and plausibility checks of the resulting speed distributions. Internal quality indicators include score (accumulated prediction error), std shift (disagreement between per-gap displacements), and bearing consistency (collinearity of the two gap azimuths). p. 8 std_shift measures the disagreement between the
two per-gap displacements, thereby flagging violations of the constant-speed assumption within
the 1. 005 s window, and bearing_consistency measures the collinearity of the two gap azimuths. Since vehicle dynamics over one second are effectively constant, low bearing_consistency or high
std_shift predominantly indicate localization noise or mis-association rather than true maneuvers,
which justifies their use as rejection criteria. Table 2: Monte-Carlo characterization of the speed and direction estimators (2 × 105
simulated
trajectories per row; Gaussian per-peak localization noise; admissible displacement bounds of 0. admissible displacement range of 0. 0 px per gap restricts detection to approximately 7 km/h
to 205 km/h. 0 px per gap restricts detection to approximately 7 km/h
to 205 km/h. Within the displacement range, estimates are effectively unbiased by the filter up
to about 130 km/h (0.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Deep learning enhanced road traffic analysis can be improved by incorporating a method that detects moving echoes of road vehicles in Sentinel-2 imagery to estimate velocities, as described in previous work [5]. This allows AI systems to estimate vehicle speeds on European E-roads without relying on fixed ground infrastructure or user-reported data.
-
The system can perform realistic travel speed and travel time estimations for infrastructure planning, transport modeling, road safety analysis, and humanitarian aid logistics by using the derived per-vehicle trajectories.
-
AI systems can model emissions from traffic flow and plan safe evacuation routes or optimize relief logistics in crisis zones by utilizing the estimated vehicle speeds.
-
The system supports tourism applications by estimating concentration patterns near reception sites based on observed traffic behavior along road networks.
Abstract
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.
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs