electricsheepafrica/africa-synth-radar-swarm-intent-all

⚠️ Synthetic dataset — Parameterized from published SSA literature, not real observations. Not suitable for empirical analysis or policy inference.

Synthetic Swarm Intent Recognition

Dataset Description

Synthetic radar dataset for classifying coordinated drone swarm behavior and inferring tactical intent from aggregate radar observables. Contains 3,000 labeled samples across 6 swarm behavior classes, each with both a class label and an intent label. Features capture formation geometry, velocity distributions, behavioral dynamics, and aggregate Doppler statistics from multi-target range-Doppler maps.

Dataset Summary

Property	Value
Total samples	3,000 (500 per class, perfectly balanced)
Classes	6
Intent labels	6 (one-to-one with classes)
Features	29 columns (24 numeric, 5 categorical/ID)
Radar system	X-band, 9.5 GHz
Range	500 m to 8 km
SNR range	10 to 29 dB
Swarm sizes	1 to 60 members
Format	Apache Parquet

Swarm Behavior Classes

Class	Intent Label	Formation	Targets	Speed (m/s)	Behavior	Maneuver Rate
recon	reconnaissance	spread	3-12	20 +/- 5	grid_search	0.1 Hz
attack	attack	wedge	8-30	40 +/- 10	converge	0.5 Hz
jamming	electronic_warfare	ring	2-8	15 +/- 5	orbit	0.3 Hz
decoy	deception	dispersed	10-40	25 +/- 8	erratic	1.0 Hz
evasion	evasion	scatter	4-15	35 +/- 10	evade	2.0 Hz
individual	individual	none	1	25 +/- 10	straight	0.1 Hz

Signal Generation

Each sample generates a multi-target range-Doppler map:

1. Formation geometry: Targets are placed according to the formation type:

   • Spread/dispersed: Random positions within formation radius
   • Wedge: Angular spread with range tapering (attack formation)
   • Ring: Targets equally spaced on a circle (jamming standoff)
   • Scatter: Random positions at 0.5-1.5x formation radius (evasion)

2. Per-target Doppler: Velocity varies by behavior type:

   • Converge: Members accelerate toward a common point
   • Scatter: Members diverge with high velocity spread
   • Orbit: Sinusoidal velocity variation (circular path)
   • Evade: Large random velocity offsets
   • Straight: Tight velocity clustering

3. Multi-target pulse matrix: Each swarm member contributes amp * exp(j * phase) at its range bin and Doppler frequency. The per-target RCS follows log-normal statistics.

4. Environment: Ground clutter with environment-dependent amplitude and Doppler spread. Two-way atmospheric and rain attenuation applied to signal bins. Multipath scintillation on detected targets.

5. Feature extraction: CFAR-like detection (10 dB above median noise), noise-thresholded Doppler profiles (median + 3*MAD), and behavioral feature derivation from RD map dynamics.

Features

Aggregate Radar (5)

• rd_total_power_db: Total integrated power in RD map (dB)
• rd_num_detected_targets: CFAR-detected target count
• rd_range_extent_m: Range extent from profile standard deviation
• rd_doppler_extent_hz: Doppler extent from CFAR-detected bins (Hz)
• rd_mean_rcs_m2: Mean per-target RCS from config (m^2)

Swarm Formation (7)

• sw_num_targets: Number of swarm members
• sw_formation_radius_m: Formation radius (m)
• sw_formation_type: Formation geometry (spread/wedge/ring/scatter/dispersed/none)
• sw_centroid_range_m: Swarm centroid range (m)
• sw_centroid_velocity_ms: Swarm centroid velocity (m/s)
• sw_velocity_spread_ms: Velocity dispersion across members (m/s)
• sw_range_spread_m: Range spread of detected targets (m)

Aggregate Micro-Doppler (6)

• md_mean_doppler_hz: Mean Doppler from thresholded profile (Hz)
• md_std_doppler_hz: Doppler spread from thresholded profile (Hz)
• md_bandwidth_hz: 90th-10th percentile Doppler bandwidth (Hz)
• md_energy_spread: Peak energy concentration
• st_md_periodicity_hz: Dominant periodicity from envelope FFT (Hz)
• st_md_contrast: Peak-to-median power contrast

Behavioral (4)

• bh_maneuver_rate_hz: Rate of formation changes (Hz) — derived from behavior type with noise
• bh_altitude_var_m: Altitude variance across swarm (m) — derived from range extent
• bh_speed_var_ms: Speed variance across swarm (m/s) — derived from Doppler std
• bh_heading_var_deg: Heading variance across swarm (deg) — behavior-dependent

Scenario (3)

• snr_db: Signal-to-noise ratio from actual RD map (dB)
• range_m: Scenario range (m)
• environment: Environment condition

Key Discriminating Features

Feature	Recon	Attack	Jamming	Decoy	Evasion	Individual
Targets	3-12	8-30	2-8	10-40	4-15	1
Speed	Low	High	Low	Medium	High	Medium
Heading var	Low	Medium	Low	High	Very high	Very low
Maneuver rate	Very low	Medium	Low	High	Very high	Very low
Formation	Spread	Wedge	Ring	Dispersed	Scatter	None

Usage

from datasets import load_dataset

ds = load_dataset("electricsheepafrica/radar-swarm-intent")
df = ds['train'].to_pandas()

# Dual-label classification
print(df[['class_label', 'intent_label']].drop_duplicates())

# Formation type distribution
print(df['sw_formation_type'].value_counts())

# Behavioral analysis
behavioral = ['bh_maneuver_rate_hz', 'bh_altitude_var_m',
              'bh_speed_var_ms', 'bh_heading_var_deg']
print(df.groupby('class_label')[behavioral].mean())

Data Quality

• Perfectly balanced: 500 samples per class
• No NaN or Inf values
• SNR derived from actual RD map (not random)
• Swarm target counts consistent with config ranges
• Doppler features use noise-thresholded profiles (median + 3*MAD)
• Behavioral features show expected class separation (evasion has highest heading variance, individual has lowest maneuver rate)

Citation

@misc{electricsheep2025swarm,
  title={Synthetic Swarm Intent Recognition: Radar-Based Behavioral Classification of Coordinated Drone Formations},
  author={Electric Sheep Africa},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/electricsheepafrica/radar-swarm-intent}
}