electricsheepafrica/radar-cuas-signatures

---
license: apache-2.0
task_categories:
- tabular-classification
language:
- en
tags:
- radar
- counter-uas
- drone-detection
- micro-doppler
- synthetic-data
- signal-processing
- range-doppler
- electronic-warfare
pretty_name: "Counter-UAS Radar Signatures"
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: train
path: data/features.parquet
---

# Counter-UAS Radar Signatures (C-UAS)

## Dataset Description

Physics-based synthetic radar signature dataset for counter-unmanned aerial system (C-UAS) classification. Contains 5,500 labeled samples across 11 target classes including small commercial drones, tactical UAVs, loitering munitions, birds, and ground clutter sources. All signals are generated from first principles using coherent pulse-Doppler radar simulation with realistic noise, environmental propagation, and electronic warfare models.

### Dataset Summary

| Property | Value |
|----------|-------|
| **Total samples** | 5,500 (500 per class, perfectly balanced) |
| **Classes** | 11 |
| **Features** | 35 columns (31 numeric, 4 categorical/ID) |
| **Radar system** | X-band, 9.5 GHz, 50 MHz bandwidth, 20 kHz PRF |
| **Range coverage** | 100 m to 10 km |
| **SNR range** | 6 to 55 dB |
| **CPI** | 256 pulses |
| **Environments** | Clear, light rain, heavy rain, urban multipath, coastal |
| **Jamming** | None (50%), barrage noise (20%), repeater (15%), DRFM (15%) |
| **Format** | Apache Parquet |

## Target Classes

| Class | Group | Category | RCS (m^2) | Max Speed (m/s) | Micro-Doppler |
|-------|-------|----------|-----------|-----------------|---------------|
| `fpv_racing` | 1 | Threat | 0.008 | 30 | 4-rotor, 35k RPM |
| `loitering_munition` | 1 | Threat | 0.025 | 50 | Pusher prop, 8k RPM |
| `small_tactical_uav` | 2 | Threat | 0.5 | 45 | 2-rotor, 5k RPM |
| `medium_uav` | 3 | Threat | 5.0 | 60 | Pusher prop, 3k RPM |
| `commercial_quad` | 1 | Threat | 0.012 | 15 | 4-rotor, 8k RPM |
| `commercial_hexa` | 1 | Threat | 0.035 | 18 | 6-rotor, 7k RPM |
| `small_bird` | 0 | Clutter | 0.0002 | 15 | Wing flap, 5 Hz |
| `large_bird` | 0 | Clutter | 0.003 | 25 | Wing flap, 2.8 Hz |
| `flock` | 0 | Clutter | 0.015 | 20 | Multi-wing, 4 Hz |
| `ground_vehicle` | 0 | Clutter | 15.0 | 30 | Vibration, 30 Hz |
| `small_aircraft` | 3 | Clutter | 25.0 | 80 | Propeller, 80 Hz |

## Signal Generation

Each sample is generated through a full coherent pulse-Doppler simulation:

1. **Pulse matrix construction** (256 pulses x 64 range bins): Target body return is injected at the appropriate range bin with Doppler phase progression `exp(j * 2pi * f_d * t)` where `f_d = 2v/lambda`.

2. **Micro-Doppler injection**: Class-specific modulation sidebands are added at `body_doppler +/- spread_hz` with amplitude-modulated envelopes at the blade/wing rate. Amplitude fractions range from 0.3 (ground vehicle vibration) to 0.9 (FPV racing drone rotors).

3. **Thermal noise**: Complex Gaussian noise with power `N = kTBF` where `B = 50 MHz` bandwidth and `F = 4 dB` noise figure.

4. **Environmental effects**: Two-way atmospheric and rain attenuation, Rician multipath fading (K-factor = 10/multipath_factor), and ground clutter with environment-dependent Doppler spread.

5. **Electronic warfare**: Barrage noise, repeater, or DRFM jamming applied stochastically to ~50% of samples.

6. **Range-Doppler map**: FFT along pulse axis with fftshift to produce the (256 Doppler x 64 Range) power map.

7. **Feature extraction**: 31 numeric features extracted from the RD map using peak-centered windowing, robust noise estimation, and radar-equation RCS calibration.

## Features

### Range-Doppler
- `rd_peak_power_db`: Peak power in RD map (dB)
- `rd_peak_range_bin`, `rd_peak_doppler_bin`: Peak location indices
- `rd_num_range_bins`, `rd_num_doppler_bins`: Map dimensions (64, 256)

### Micro-Doppler
- `md_mean_doppler_hz`: Mean Doppler within peak-centered signal window (Hz)
- `md_std_doppler_hz`: Spectral spread of signal structure, second moment (Hz)
- `md_skewness`, `md_kurtosis`: Higher-order spectral shape statistics
- `md_energy_spread`: Fraction of energy in top 20% of signal bins
- `md_dominant_freq_hz`: Peak Doppler frequency (Hz)
- `md_bandwidth_hz`: 3-dB bandwidth of Doppler peak (Hz)
- `md_num_harmonics_detected`: Count of harmonic peaks at 2x, 3x, 4x fundamental

### Blade Flash / Modulation
- `bf_flash_rate_hz`: Dominant periodicity in amplitude envelope (Hz)
- `bf_max_doppler_spread_hz`: Maximum instantaneous frequency spread (Hz)
- `bf_modulation_index`: Amplitude modulation depth (std/mean)

### Slow-Time
- `st_md_bandwidth_hz`: 90% energy bandwidth in peak-centered window (Hz)
- `st_md_periodicity_hz`: Dominant periodicity from envelope FFT (Hz)
- `st_md_contrast`: Peak-to-median power contrast in signal region

### Temporal
- `t_duration_s`: CPI duration with ~1% timing jitter (s)
- `t_num_significant_returns`: Samples exceeding 3x noise floor
- `t_duty_cycle`: Fraction of pulses with significant returns

### Quality & Metadata
- `snr_db`: Signal-to-noise ratio, 0-55 dB
- `noise_floor_db`: Measured noise floor with receiver variation (dB)
- `range_m`: Target range (m)
- `velocity_ms`: Radial velocity (m/s)
- `aspect_deg`: Aspect angle (deg)
- `elevation_deg`: Elevation angle (deg)
- `environment`: Environmental condition (clear/light_rain/heavy_rain/urban_multipath/coastal)
- `jamming_type`: Active EW type (none/noise_jamming/repeater_jamming/drfm_jamming)
- `rcs_estimated_m2`: Radar-equation calibrated RCS estimate (m^2)
- `rcs_true_m2`: Ground-truth RCS from target model (m^2)

## Usage

```python
from datasets import load_dataset
import pandas as pd

# Load from HuggingFace
ds = load_dataset("electricsheepafrica/radar-cuas-signatures")
df = ds['train'].to_pandas()

# Quick classification
from xgboost import XGBClassifier
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelEncoder

feature_cols = [c for c in df.columns if c not in
['sample_id', 'class_label', 'class_group', 'environment', 'jamming_type']]
X = df[feature_cols].values
y = LabelEncoder().fit_transform(df['class_label'])

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, stratify=y)
clf = XGBClassifier(n_estimators=200, max_depth=6)
clf.fit(X_train, y_train)
print(f"Accuracy: {clf.score(X_test, y_test):.3f}")
```

## Data Quality

- Perfectly balanced: 500 samples per class
- No NaN or Inf values
- RCS estimation correlates with ground truth (r ~ 0.72)
- Per-class velocity ranges match kinematic envelopes
- SNR clamped to physically realistic [0, 55] dB range
- Micro-Doppler features use peak-centered windowing to avoid noise artifacts

## Citation

```bibtex
@misc{electricsheep2025cuas,
title={Counter-UAS Radar Signatures: Synthetic Micro-Doppler and Range-Doppler Features},
author={Electric Sheep Africa},
year={2025},
publisher={HuggingFace},
url={https://huggingface.co/datasets/electricsheepafrica/radar-cuas-signatures}
}
```