Self-Supervised Radar Signal Recognition

This repo contains radar datasets accompanying the paper "Few-Shot Radar Signal Recognition through Self-Supervised Learning and Radio Frequency Domain Adaptation", published at the IEEE International Workshop on Machine Learning for Signal Processing (MLSP) 2025.

You can access our preprint 📄 here: https://arxiv.org/abs/2501.03461

You can download our dataset ⬇️ here: https://www.kaggle.com/datasets/abcxyzi/radcharssl-mlsp-2025

Quick Links

• RadChar (Radar Data)
• RadioML (Comm Data)
• DeepRadar (Radar Data)
• RadarComm (Comm & Radar Data)
• Citation

Dataset Overview

⚙️ Four datasets were considered in our paper:

1. RadChar (ours) - this contains radar signals only.
2. RadioML - this contains telecommunications signals only.
3. DeepRadar - this contains radar signals only.
4. RadarComm - this contains a mixture of telecommunications and radar signals.

RadChar contains several variants used to support self-supervised pre-training, few-shot fine-tuning, and model evaluation:

• RadChar-SSL - this is used for pre-training only.
• RadChar-nShot - this is used for few-shot fine-tuning only, where n = 1, 5, 10.
• RadChar-Eval - this is used for evaluation only.

Note, RadChar dataset variants were derived from our previous work.

RadChar (Radar Data)

RadChar is our own dataset. This was introduced in Multi-Task Learning For Radar Signal Characterisation, published in the 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW).

⚙️ RadChar contains the following characteristics:

• 5 classes
• 41 SNR levels (-20 to 20 dB with a 1 dB resolution)
• 512 I/Q samples per frame
• 0.3 μs in temporal resolution

RadChar is available in multiple sizes, provided as .h5 files. The baseline dataset, RadChar-Baseline (RadChar-Baseline.h5), can be downloaded from: https://github.com/abcxyzi/RadChar.

To support self-supervised pre-training, few-shot fine-tuning, and model evaluation in our paper, we introduce several RadChar variants:

1. RadChar-SSL - a newly introduced dataset used exclusively for pre-training.
2. RadChar-nShot - three small-scale datasets (n = 1, 5, 10) for few-shot fine-tuning, uniquely sampled from 90% of RadChar-Baseline.
3. RadChar-Eval - the evaluation dataset (i.e., test set), created from the remaining 10% of RadChar-Baseline.

RadChar-SSL and its parts can be downloaded from Kaggle.

RadioML (Comm Data)

RadioML 2018.01A (RadioML) is a dataset created by DeepSig Inc. This dataset was introduced in Over-the-Air Deep Learning Based Radio Signal Classification, published in the 2017 IEEE Journal of Selected Topics in Signal Processing.

⚙️ RadioML contains the following characteristics:

• 24 classes
• 26 SNR levels (-20 to 30 dB with a 2 dB resolution)
• 1,024 I/Q samples per frame
• 1 μs in temporal resolution

RadioML is provided as a single file GOLD_XYZ_OSC.0001_1024.hdf5 (21.45 GB), it can be downloaded from: https://www.kaggle.com/datasets/pinxau1000/radioml2018

10% of RadioML were used for self-supervised pre-training in our paper.

DeepRadar (Radar Data)

DeepRadar is a dataset created by the Radar and Microwave Group. This dataset was introduced in LSTM Framework for Classification of Radar and Communications Signals, published in the 2023 IEEE Radar Conference (RadarConf23).

⚙️ DeepRadar contains the following characteristics:

• 23 classes
• 17 SNR levels (-12 to 20 dB with a 2 dB resolution)
• 1,024 I/Q samples per frame
• 0.01 μs in temporal resolution

DeepRadar is provided in three parts (train, validation, and test). The individual .mat files can be downloaded from: https://www.kaggle.com/datasets/pinxau1000/radioml2018

The training set X_train.mat was used for self-supervised pre-training in our paper

RadarComm (Comm & Radar Data)

RadarComm is a dataset created by ANDRO Computational Solutions. This dataset was introduced in Multi-task Learning Approach for Automatic Modulation and Wireless Signal Classification, published in the 2021 IEEE International Conference on Communications (ICC).

⚙️ RadarComm contains the following characteristics:

• 6 modulation classes and 8 signal classes (each frame is dual-annotated with a modulation class and a signal class)
• 17 SNR levels (-12 to 20 dB with a 2 dB resolution)
• 128 I/Q samples per frame
• 0.1 μs in temporal resolution

RadarComm contains several dataset variants provided in .zip files. They can be downloaded from: https://www.kaggle.com/datasets/pinxau1000/radioml2018

The RadarComm dataset with dynamic channel effects RadComDynamic.zip was used for self-supervised pre-training in our paper. Note, we use only the modulation class labels for few-shot fine-tuning.

Citation

💡 Please cite both the dataset and the conference paper if you find them helpful for your research. Cheers.

@misc{huang2025fewshot,
  title        = {Few-Shot Radar Signal Recognition through Self-Supervised Learning and Radio Frequency Domain Adaptation},
  author       = {Zi Huang and Simon Denman and Akila Pemasiri and Clinton Fookes and Terrence Martin},
  year         = {2025},
  eprint       = {2501.03461},
  archivePrefix= {arXiv},
  primaryClass = {cs.LG},
  url          = {https://arxiv.org/abs/2501.03461}
}