gc-vit-autonomous-radar-modulation-recognition

GC-ViT: AI-Based Global Context Vision Transformer for Radar Signal Modulation Recognition

An AI-based Global Context Vision Transformer (GC-ViT) that leverages the short-time Fourier transform (STFT) phase spectrum for feature extraction to identify phase-coded radar waveforms — a key capability for electronic warfare (EW) systems facing the growing use of low-probability-of-intercept (LPI) radars. Combines local and global self-attention to improve recognition robustness at low SNR.

Dataset: (https://www.kaggle.com/datasets/sidrabhatti/latestdataset-cnn)

Role & Attribution

Sidra Ghayour Bhatti — first author; conceived, implemented, and evaluated the GC-ViT model. Co-authored with Mohsin Ullah (Dept. of Electrical Engineering, Capital University of Science and Technology, Islamabad).

Method

Input representation. Each intercepted radar pulse is converted to its short-time Fourier transform (STFT) phase spectrum rather than the more commonly used magnitude spectrum — the deliberate choice is because phase-coded waveforms (Barker codes, P1–P4 polyphase codes, etc.) are defined by their intrapulse phase-modulation pattern, not by how their energy is distributed across frequency. Two different phase codes can have near-identical magnitude spectra while differing sharply in phase, so a magnitude-only spectrogram throws away the very information that distinguishes the waveform classes; the phase spectrum keeps it. The resulting 2D time-frequency phase map is cropped to its informative region and resized to 224×224×3, turning a 1D IQ waveform recognition problem into a 2D image classification problem that a vision architecture can exploit.

Model. A GC-ViT (Global Context Vision Transformer, Tiny variant) backbone, pretrained then fine-tuned with a 6-class softmax classification head for phase-coded waveform families. GC-ViT combines local window-based self-attention (fine-grained spectrogram texture) with a global query token (long-range structure across the full time-frequency map) in each block — the architectural reason it outperforms CNN baselines at low SNR, where discriminative structure is spread across the spectrogram rather than localized.

Figure 3: The suggested GC-ViT architecture — an input spectrogram passes through a stem, then alternating stages of local + global multi-head self-attention (MSA) blocks with a global token generator feeding each stage, downsampling and doubling channel depth between stages, ending in 2D average pooling and a classification head.

Training hyperparameters.

Results

Recognition accuracy (%) vs. SNR, per phase-coded waveform type:

All six waveform types reach 100% accuracy by −4 dB SNR; Barker and P4 codes are the weakest at −12 dB (most easily confused with other schemes at extreme noise), while P2 codes are correctly recognized at every tested SNR level. Comparable to or exceeding other transformer-based LPI radar recognition methods in the literature at similar SNR ranges.

Code

Global_Context_ViT_for_RadarSiganls.ipynb — the training/evaluation notebook used for this paper. It's Colab-specific (mounts Google Drive, expects a zipped private dataset and a locally-trained .h5 weights file that are not included here), so it won't run standalone, but it documents the exact preprocessing, model setup, and SNR-sweep evaluation protocol behind the reported results.

Published: Bhatti, S.G. & Ullah, M. (2024). "Radar signal modulation identification using global context vision transformer." Engineering Research Express, 6(4), 045331. doi.org/10.1088/2631-8695/ad8b96 (subscription required — not open access, so no PDF is included here; link only)

Citation

@article{bhatti2024radar,
  author  = {Bhatti, Sidra Ghayour and Ullah, Mohsin},
  title   = {Radar signal modulation identification using global context vision transformer},
  journal = {Engineering Research Express},
  volume  = {6},
  number  = {4},
  pages   = {045331},
  year    = {2024},
  doi     = {10.1088/2631-8695/ad8b96}
}