RadarML: A Multi-Channel Over-the-Air Dataset of Common Radar Waveforms for ES Evaluation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: We're looking at the paper: "RadarML: A Multi-Channel Over-the-Air Dataset of Common Radar Waveforms for ES Evaluation".
Dev: Here's the gist of the paper. RadarML introduces a publicly available dataset of modulated radar pulses recorded over-the-air across multiple semi-independent receiver channels to enable comparative analysis of radar modulation classification techniques and stimulate research into multi-channel signal detection and modulation classification methods Dataset Scope The dataset contains a library of around two million waveforms with varying centre frequencies, bandwidths, pulse widths and modulations, with 1,000 examples of each signal variation per receiver channel The goal of this dataset is to enable comparative analysis of radar modulation classification techniques on real data and stimulate research into multi-channel signal detection and modulation classification methods Waveform Diversity The waveforms vary in centre frequency, time duration or bandwidth (if applicable) Some example spectrograms generated from the dataset are shown in Fig.
Rosa: 1 Table 2 summarizes the structure of each waveform dataset file The samples for each pulse are provided unprocessed, as digitised by the ADCs, in the np.int16 format with no synthetic noise added Hardware Implementation The dataset was generated using the UCL ARESTOR platform which is a Radio Frequency System on a Chip (RFSoC) based Software Defined Radio (SDR) A single DAC (Digital-to-Analogue) channel was used as the transmitter and 3 semi-independent ADC channels on different RF-Tiles were used for receive The receivers on the RFSoC operate at a sample rate of 3.84 GS/s, which is then decimated by a factor of 32 in two stages to achieve a final sample rate of 120 MS/s Data Augmentation and Evaluation A code repository is provided which includes a PyTorch dataloader for viewing the raw data, and a wrapper dataloader that supports augmentations in SNR, frequency, phase and time to improve generalisation A spectrogram transform and a baseline VGG13-style classifier are also included Training uses pulses degraded from 30 dB to 0 dB SNR, with random frequency, phase and time shifts The trained model is evaluated from 15 dB to −15 dB SNR, with per-SNR confusion matrices and macro F1 plotted against SNR Data Availability The full dataset as described in this paper is available at https://doi.org/10.5522/04/30752767 The dataset is released under the CC BY 4.0 license and the code is released under the GPL-3.0 license This paper has presented RadarML, a multi-channel dataset of representative radar waveforms captured over-the-air to support the evaluation of electronic surveillance and modulation classification techniques It is hoped that RadarML will enable reproducible comparison of algorithms and stimulate further research into multi-channel signal processing and fusion approaches Acknowledgments This dataset was generated under an iCASE PhD Studentship funded by the Engineering and Physical Sciences Research Council (EPSRC) and the Defence Science and Technology Laboratory (Dstl) Prof.
Title and authors: Rosa: We're talking about the "RadarML: A Multi-Channel Over-the-Air Dataset of Common Radar Waveforms for ES Evaluation" paper, and it seems the core focus is on providing a standardized source for evaluating electronic surveillance techniques.
Dev: Right, so beyond just having a dataset, they are providing the structure for comparison across different modulation types and pulse parameters, which is what makes it useful for testing.
Rosa: The authors built this using the UCL ARESTOR system as an SDR platform to generate these signals over-the-air in the two point four gigahertz industrial scientific and medical band.
Dev: They used a single node acting as both transmitter and receiver(s), specifically three semi independent receiver channels on different rf tiles, which is key for their multi-channel setup.
Taro: I wonder about the limitations they mention regarding the data generation; how long did they run these experiments to ensure this dataset is truly representative of real operational conditions?
Rosa: The paper doesn't explicitly state a fixed duration for the collection, but it emphasizes that the goal is to capture common waveforms across varying center frequencies, bandwidths and pulse widths.
Dev: They specify a structure where there are one thousand examples of each signal variation per receiver channel, which gives us a defined scope for what we can test on.
Taro: From an autonomy research standpoint, I'm curious about the failure modes they tested; how gracefully does the system handle signals that don't fit those common waveform categories?
Rosa: The paper mentions that they include various pulse modulations like linear chirps and barker pulses in their examples shown in figure one.
Dev: And they also detail the structure of each dataset file, showing dimensions like three adc channels and thirty-three thousand six hundred pulses acquired per waveform configuration, which gives us concrete scale information <ref:2611.16786#pg2,pulses acquired per waveform configuration>.
Taro: That scale matters because it tells us how much data we're dealing with when we try to train a model to recognize these signals.
Rosa: The paper also provides tools like the wrapper dataloader that lets you augment in snr, frequency, phase and time to help improve generalization of the models.
Dev: So while the dataset itself is fixed, it's not just raw data; it comes with tools designed to make the AI systems trained on it more robust across different signal qualities.
Taro: That augmentation capability sounds like a necessary step for deployment, because real radar environments are never perfectly clean or perfectly tuned.
Rosa: So in this segment, we've seen how the authors structured the collection and what kind of variety they included in the RadarML dataset. Next up, we're going to look at what improvements they suggest for using this data.
The paper's summary: Dev: Now let's talk about what the paper actually summarizes—the main contribution of this work beyond just presenting the dataset itself.
Rosa: The summary stresses that the primary goal is to enable comparative analysis of radar modulation classification techniques on actual data, moving away from reliance on purely synthetic definitions.
Dev: Essentially, they want to create a standard reference set so researchers can test their algorithms against real-world data and see how well they perform across different modulation types.
Rosa: They also explicitly aim to stimulate research into multi-channel signal detection and modulation classification methods by utilizing the multi-channel nature of the data.
Dev: This means the paper isn't just a data dump; it’s a framework for testing how combining multiple semi independent receive channels helps improve both detection and classification performance.
Taro: From my view, this shifts the research focus toward fusion approaches where you leverage the information from several receivers simultaneously to get better results than any single channel alone.
Rosa: That's right, it encourages looking at signal processing techniques that specifically exploit the multi-channel nature of these measurements.
Dev: The authors are laying out a path for researchers to use this dataset as a standard reference point, which should lead to more reproducible comparisons across different modulation types.
Taro: I think if we can get this dataset widely used, it will force the community to develop better ways of handling those multi-channel signals in complex electromagnetic environments.
Rosa: So, the summary really boils down to providing a real data foundation for testing both classification methods and multi-channel signal processing techniques.
Dev: And it sets up a clear direction for future work in how we combine these channels effectively.
The paper's improvements: Rosa: Moving on to the improvements suggested by the authors, they focus heavily on making the usage of this dataset more practical and robust for real-world application.
Dev: They provide a code repository that includes a PyTorch dataloader for viewing raw data and that wrapper dataloader which supports augmentations in snr, frequency, phase and time.
Taro: That wrapper is key because it allows you to test how well your model generalizes when you intentionally degrade the signal quality in controlled ways before training.
Rosa: Exactly; they show that these augmentations can significantly improve the generalization of the trained model across different operational conditions.
Dev: They also include a spectrogram transform and a baseline VGG13 style classifier right within the code, which serves as a ready-to-use demonstration of how these components fit together.
Taro: So this isn't just theory; they are giving us the practical tools to actually build and evaluate these systems using this data pipeline.
Rosa: And they also detail their training process where pulses are degraded from thirty decibels down to zero decibels signal to noise ratio with random frequency, phase, and time shifts.
Dev: Then, when evaluating, they use a wider range of degradation from fifteen decibels down to negative fifteen decibel snr and plot confusion matrices and macro f one against that snr.
Taro: That SNR range for evaluation is important because it shows how the performance of the classification method changes depending on how clean the signal actually is in practice.
Rosa: It really shows that you can use this dataset not just to get a single accuracy number, but to see how stable that performance is across varying signal qualities.
Dev: The authors are providing a complete setup, from data loading and augmentation through training and evaluation metrics like per-snr confusion matrices and macro f one plotted against snr.
Conclusion: Rosa: So wrapping up the discussion on the RadarML paper, the main implications are that we now have a standardized, multi-channel dataset to push modulation classification techniques against real radar signals.
Dev: And it strongly encourages research into how we can use those multiple channels effectively for detection and classification in complex electromagnetic environments.
Taro: For me, this means having a concrete way to test fusion methods by using the multi-channel structure, which is what we need to move beyond single-sensor thinking.
Rosa: And for AI systems specifically, it shows how you can train them to perform automatic radar waveform detection and classification directly from raw digitized signals.
Dev: The authors have given us a system that supports training with data augmentation techniques that improve generalization under varying signal quality conditions.
Taro: I think the main thing is reproducibility; having this standardized dataset means we can compare algorithms reliably across different modulation types without relying on synthetic definitions for everything.
Rosa: So, to recap, the RadarML paper provides a publicly available set of over two million waveforms to support evaluation of electronic surveillance and modulation classification techniques.
Dev: It’s hoped that this dataset will enable reproducible comparisons of algorithms and stimulate further research into multi-channel signal processing and fusion approaches in this area.
Taro: And I think it sets a good foundation for testing how systems recover when the world misbehaves by providing these diverse, real data samples.
Ryan White, Adam Hosford, Matthew A. Ritchie
University College London · Defence Science and Technology Laboratory
eess.SP
Submitted: 2026-11-12
Updated: 2026-11-12
Comments: 4 pages, 5 figures
Code: https://github.com/UCLRadarGroup/Radar_ML
License: http://creativecommons.org/licenses/by-sa/4.0/
Importance score: 79/100
The gist: RadarML introduces a publicly available dataset of modulated radar pulses recorded over-the-air across multiple semi-independent receiver channels to enable comparative analysis of radar modulation…
Key concepts
- RadarML Dataset
- This is a large, real-world collection of over two million radar waveforms captured over the air using multiple receiver channels. It includes variations in frequency, bandwidth, and modulation type for thorough testing.
- Multi-channel Analysis
- The dataset uses signals recorded across several semi-independent receiver channels simultaneously. This setup is designed to stimulate research into improving signal detection and classification performance in complex electromagnetic environments.
- Data Augmentation
- A code repository provides tools to augment the training data by randomly changing parameters like Signal-to-Noise Ratio (SNR), frequency, phase, and time. This helps train AI models that are more robust and generalize better to real-world signal variations.
Terminology
Summary
RadarML introduces a publicly available dataset of modulated radar pulses recorded over-the-air across multiple semi-independent receiver channels to enable comparative analysis of radar modulation classification techniques and stimulate research into multi-channel signal detection and modulation classification methods
Dataset Scope
The dataset contains a library of around two million waveforms with varying centre frequencies, bandwidths, pulse widths and modulations, with 1,000 examples of each signal variation per receiver channel The goal of this dataset is to enable comparative analysis of radar modulation classification techniques on real data and stimulate research into multi-channel signal detection and modulation classification methods
Waveform Diversity
The waveforms vary in centre frequency, time duration or bandwidth (if applicable) Some example spectrograms generated from the dataset are shown in Fig. 1 Table 2 summarizes the structure of each waveform dataset file The samples for each pulse are provided unprocessed, as digitised by the ADCs, in the np. p. 0 Synthetic Waveform Library
Modulation
Baseband
Centre
Frequencies
(MHz)
Bandwidths
(MHz)
Pulse
Widths
(µs)
CW -30, -10, 10, 30 N/A 10, 50, 100
Linear FM -30, -10, 0, 10, 30 +/- 10, 20, 40, 80 10, 50, 100
Exponential FM -30, -10, 0, 10, 30 +/- 10, 20, 40, 80 10, 50, 100
Quadratic FM -30, -10, 0, 10, 30 +/- 10, 20, 40, 80 10, 50, 100
Tangential FM -30, -10, 0, 10, 30 +/- 10, 20, 40, 80 10, 50, 100
4 CP-FSK -30, -10, 0, 10, 30 10, 20, 40, 80 10, 50, 100
Barker 5 -30, -10, 10, 30 N/A 10, 50, 100
Barker 7 -30, -10, 10, 30 N/A 10, 50, 100
Barker 11 -30, -10, 10, 30 N/A 10, 50, 100
Barker 13 -30, -10, 10, 30 N/A 10, 50, 100
Barker 5 ⊗ 11 -30, -10, 10, 30 N/A 10, 50, 100
Barker 5 ⊗ 13 -30, -10, 10, 30 N/A 10, 50, 100
For every waveform combination, the dataset contains a
NumPy file with dimensions [3, 1000, 33600], the dimension
definitions are summarised in Table 2. The samples for each pulse are provided unprocessed, as
digitised by the ADCs, in the np.Where the paper says it
Hardware Implementation
The dataset was generated using the UCL ARESTOR platform which is a Radio Frequency System on a Chip (RFSoC) based Software Defined Radio (SDR) A single DAC (Digital-to-Analogue) channel was used as the transmitter and 3 semi-independent ADC channels on different RF-Tiles were used for receive The receivers on the RFSoC operate at a sample rate of 3.84 GS/s, which is then decimated by a factor of 32 in two stages to achieve a final sample rate of 120 MS/s
Data Augmentation and Evaluation
A code repository is provided which includes a PyTorch dataloader for viewing the raw data, and a wrapper dataloader that supports augmentations in SNR, frequency, phase and time to improve generalisation A spectrogram transform and a baseline VGG13-style classifier are also included Training uses pulses degraded from 30 dB to 0 dB SNR, with random frequency, phase and time shifts The trained model is evaluated from 15 dB to −15 dB SNR, with per-SNR confusion matrices and macro F1 plotted against SNR
Data Availability
The full dataset as described in this paper is available at https://doi.org/10.5522/04/30752767 The dataset is released under the CC BY 4. p. 3 DATA AND CODE AVAILABILITY
The dataset and some source code to get started are
publicly available, the links for which are provided below:
• Dataset: The full dataset as described in this paper is
available at https://doi. 5522/04/30752767. The dataset is released under the CC BY 4. 0 license and
the code is released under the GPL-3.Where the paper says it
Where the paper says it
Conclusion
This paper has presented RadarML, a multi-channel dataset of representative radar waveforms captured over-the-air to support the evaluation of electronic surveillance and modulation classification techniques It is hoped that RadarML will enable reproducible comparison of algorithms and stimulate further research into multi-channel signal processing and fusion approaches
Acknowledgments
This dataset was generated under an iCASE PhD Studentship funded by the Engineering and Physical Sciences Research Council (EPSRC) and the Defence Science and Technology Laboratory (Dstl) Prof.
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
-
Radar modulation classification performance can be enhanced by using RadarML as a standard reference dataset to enable comparative analysis of techniques on real data. This allows for reproducible comparison of algorithms across different modulation types and pulse parameters without relying solely on synthetic data definitions.
-
Multi-channel signal detection and modulation classification methods can be stimulated by utilizing the multi-channel nature of RadarML. This research direction focuses on using multiple semi-independent receive channels to improve both detection and classification performance in complex electromagnetic environments.
-
AI systems can be trained to perform automatic radar waveform detection and classification, as demonstrated by the VGG13 classifier mentioned in the paper. The improved AI system can identify specific radar modulation types from raw digitized signals based on features extracted from spectrograms, such as those shown in Fig. 2 and Fig. 1.
-
AI systems can be augmented with data processing techniques that improve generalization across different operational conditions, as shown by the wrapper dataloader supporting augmentations in SNR, frequency, phase and time. This allows the trained model to perform reliably even when deployed in environments with varying signal quality or parameter shifts.
Abstract
This paper introduces "RadarML", a publicly available dataset of modulated radar pulses recorded over-the-air across multiple semi-independent receiver channels. The dataset was generated using the UCL ARESTOR platform which is a Radio Frequency System on a Chip (RFSoC) based Software Defined Radio (SDR). The dataset contains a library of around two million waveforms with varying centre frequencies, bandwidths, pulse widths and modulations, with 1,000 examples of each signal variation per receiver channel. The goal of this dataset is to enable comparative analysis of radar modulation classification techniques on real data and stimulate research into multi-channel signal detection and modulation classification methods.
Related papers
- Open-H-RF: A Large-Scale Ultrasound Raw Channel Dataset Initiative
- Scalable Attention for 5G NR Channel Estimation
- Low-Altitude Wireless Networks: A Comprehensive Survey
- Sensing-Assisted OFDM with Index Modulation: A Waveform for Optical Wireless Integrated Communication and Localisation
- Iteration-Free Candidate Pruning and an Index-Domain VLSI Architecture for Sort-Free MIMO Box Decoding up to 1024-QAM
- A Foundation Model for Large-Scale Wireless Network Planning, Operation and Optimization