The Vienna 4G/5G Drive-Test Dataset

This paper introduces the Vienna 4G/5G Drive-Test Dataset, a comprehensive open resource containing georeferenced LTE and 5G measurements, inferred base station descriptors, and high-resolution city models to facilitate machine learning applications in mobile network analysis, planning, and optimization.

The Gist

Imagine you are trying to build a perfect, digital "twin" of the city of Vienna—a virtual replica where you can test how mobile phone signals travel through streets, around buildings, and into basements. To do this, you need two things: a map of the city's physical layout and a logbook of how signals actually behave in the real world.

Until now, researchers had the map, or they had the logbook, but rarely both together in high detail. This paper introduces the Vienna 4G/5G Drive-Test Dataset, which is like handing researchers a complete, high-definition "survival kit" for understanding mobile networks.

Here is a breakdown of what this paper is about, using some everyday analogies:

1. The Problem: The "Blind Spot" in Network Planning

Imagine you are a city planner trying to design a new subway system. You have a map of the city, but you don't know where the trains actually stop, how fast they go, or where the tunnels are blocked. You'd be guessing.

In the world of mobile networks (4G and 5G), companies use AI to plan where to put antennas. But they often lack real-world data that is paired with detailed 3D maps. They have the "what" (the signal strength) but not the "where" (the exact building height or antenna angle) to explain why the signal is weak. This dataset fills that gap.

2. The Solution: A "Dual-Lens" Camera

The researchers didn't just take one type of picture; they took two simultaneous views of the same city:

• The "Passive" Lens (The Wide-Angle Scanner): Think of this as a super-sensitive radio telescope mounted on a car roof. It doesn't talk to the network; it just listens. It hears every signal from every cell tower, 24/7, like a birdwatcher recording every bird song in a forest. This gives a neutral, unbiased view of the whole network.
• The "Active" Lens (The Smartphone): This is the data from actual smartphones mounted on the dashboard. These phones are "talking" to the network, downloading data, and reporting back: "Hey, the signal is strong here," or "I'm moving fast, and the connection is lagging."

By combining these two, the dataset gives you both the network's perspective (what the towers are broadcasting) and the user's perspective (what the phone actually experiences).

3. The "Digital Twin" Ingredients

To make a realistic simulation (a Digital Twin), you need more than just signal numbers. You need context. This dataset provides:

• The "Ghost" Towers: For many cell towers, the researchers didn't just guess where they were; they used math and signal timing to estimate their exact location, how high the antenna is, and which way it is pointing (like guessing where a lighthouse is by looking at the light beam).
• The 3D City Model: They included a high-resolution digital version of Vienna's buildings and terrain. Imagine a video game map where every building is 3D, not just a flat square. This allows researchers to teach AI how signals bounce off glass or get blocked by a skyscraper.

4. Why This Matters (The "Why Should I Care?")

Think of this dataset as a training gym for AI.

• For Engineers: It helps them calibrate their "ray tracing" tools. Ray tracing is like a video game engine that simulates how light (or radio waves) bounces around. This dataset lets them check: "Does my simulation match reality?" If not, they can fix the math.
• For AI Researchers: It allows them to train AI models to predict coverage. Instead of guessing, the AI can learn: "Oh, when a signal hits a 30-story building at this angle, it drops by 50%."
• For the Future: It helps prepare for 5G and future 6G networks, ensuring that when we get new phones, the network is already optimized to handle them without dropping calls in the middle of the street.

5. How They Did It (The "Road Trip")

Between March 2024 and March 2025, the team drove around Vienna (about 100 square kilometers) in a car equipped with this high-tech gear.

• They drove through busy city centers and quiet suburbs.
• They recorded millions of data points.
• They cleaned up the data (removing "noise" like GPS glitches) and organized it into neat folders: one for the scanner data, one for the phone data, one for the estimated tower locations, and one for the city map.

The Bottom Line

This paper isn't just about sharing a list of numbers. It's about opening the door for anyone to build better, smarter mobile networks. It's like giving a chef not just a recipe, but the actual ingredients, the kitchen layout, and a video of the cooking process, so they can learn to cook the perfect meal every time.

By making this data open and free, the authors hope that researchers worldwide can stop guessing how networks work and start proving it, leading to faster internet and fewer dropped calls for everyone.

Technical Summary

1. Problem Statement

Machine learning (ML) for mobile network analysis, planning, and optimization is currently hindered by a lack of large-scale, comprehensive real-world datasets. Existing open datasets typically focus on Key Performance Indicators (KPIs) and measurement traces but lack explicit deployment descriptors (e.g., base station locations, antenna orientations) and high-resolution 3D environmental context (building and terrain models). This gap prevents robust benchmarking of environment-aware learning, propagation modeling, and Digital Twin (DT) construction, particularly for geometry-conditioned problems like ray-tracing calibration and radio map estimation.

2. Methodology

The authors collected and processed a city-scale dataset covering approximately 100 km² of Vienna, Austria, between March 2024 and March 2025. The methodology involves a dual-modality data collection approach and a rigorous post-processing pipeline:

• Data Collection:

  • Passive Measurements: A PCTEL IBflex wideband scanner equipped with two OmniLOG PRO omnidirectional antennas and a GPS receiver captured network-side signals. This provided an unbiased view of the radio environment, including Multi-Cell and Beam-level (for 5G NR) observations.
  • Active Measurements: Three 5G-capable smartphones mounted on a vehicle captured user-side logs (RSRP, RSRQ, RSSI, SINR, throughput, and Timing Advance) via Nemo Outdoor software.
  • Environmental Context: High-resolution Building Models (BKM) and Ground-Level Models (GLM) were obtained from the City of Vienna (Stadtvermessung Wien) at 1m spatial resolution.

• Data Processing & Inference:

  • Harmonization: Raw Nemo logs were parsed to extract Master Information Blocks (MIB), System Information Blocks (SIB), and reference signals (CRS/SSB). Data was cleaned, exploded into individual entries, and aligned temporally and spatially.
  • Base Station (BS) Estimation: For a representative subset of cells, the authors inferred deployment descriptors:
    • LTE: Used network-reported cell IDs to group measurements, then estimated locations using public registers and Timing Advance (TA) circles.
    • 5G NR: Since NSA mode handsets do not expose physical gNB IDs, a PCI-based grouping heuristic was used. TA circles were constructed around UE locations to validate spatial consistency before estimating BS positions.
    • Parameters: Estimated locations, sector azimuths (using RSRP-weighted centroids), and antenna heights (via Google Earth Pro) were derived.

3. Key Contributions

The paper introduces the Vienna 4G/5G Drive-Test Dataset, a unique open-access resource with the following distinct features:

• Dual Modality: Combines passive scanner data (network-side, multi-cell/beam view) with active handset logs (user-side, throughput/TA view).
• Deployment Descriptors: Provides estimated metadata for a subset of Base Stations, including coordinates, sector azimuths, and antenna heights, which are rarely available in open datasets.
• 3D Environmental Context: Includes high-resolution Building and Terrain models (GeoTIFF rasters) aligned with the measurement data, enabling geometry-conditioned learning.
• Comprehensive Scope: Covers both LTE and 5G NR (NSA) across diverse urban and suburban environments, including beam-level measurements for 5G.
• Standardized Format: Data is organized into four components (Scanner, Handset, Estimated Cell Info, City Model) in CSV/Parquet and GeoTIFF formats, facilitating immediate use without complex preprocessing.

4. Results and Validation

The authors performed extensive technical validation to ensure data reliability:

• Georeferencing Accuracy: GPS samples were snapped to OpenStreetMap road segments, yielding a Root Mean Square Error (RMSE) of ~4 meters, confirming sufficient precision for drive-test analysis.
• Cell Assignment Consistency: Over 2.6 million CRS/SSB records were matched to MIB/SIB events. While ~18% were unmatched (due to asynchronous decoding), the spatial residuals for matched records were tight (Median |∆x| = 0 m, P90 = 145 m).
• KPI Statistics: The dataset captures realistic propagation characteristics. Low-band LTE (800 MHz) showed the strongest RSRP, while NR 3500 MHz (C-band) showed the weakest due to path loss, consistent with physical expectations.
• BS Estimation Accuracy:
  • LTE: Validated against public registers, showing a localization RMSE of 38.55 m and a sector orientation error of 27.95°.
  • 5G NR: Validated internally via Timing Advance consistency, showing a median residual of 69.6 m, confirming the inferred locations are physically plausible.

5. Significance and Use Cases

This dataset bridges the gap between empirical measurements and simulation-based modeling, enabling several critical research workflows:

• Ray-Tracing Calibration: Allows researchers to calibrate deterministic simulators (e.g., NVIDIA Sionna RT) by comparing simulated paths against real-world measurements in a known 3D environment.
• Environment-Aware Learning: Facilitates the training of ML models for radio map estimation and coverage prediction that are conditioned on building and terrain features.
• Digital Twin (DT) Construction: Provides the foundational data (measurements + deployment metadata + 3D city model) required to build and calibrate city-scale Digital Twins.
• Beam Management: Offers beam-level measurements for studying 5G NR beam selection and tracking under mobility.
• Benchmarking: Enables fair comparison between AI-based methods and traditional model-based approaches in a unified, reproducible setting.

In summary, the Vienna 4G/5G Drive-Test Dataset is a significant step forward in mobile network research, offering the first large-scale, open resource that jointly integrates real-world drive-test data, inferred deployment metadata, and high-resolution 3D city models.