The summary covers all 17 splits, so it downloads the whole dataset (about 79 MB). Splits already opened in other tabs are reused.
Summary statistics by drone and polarisation (dBsm)
Mean σ (linear) is the true average reflecting area. Mean of dB averages the dB values directly; because of deep nulls it always comes out lower and is shown for comparison only.
Data completeness check
Single-frequency heatmaps are dominated by interference speckle, so check structural features (e.g. specular ridges) in the all-band linear mean first. Interpret the θ/φ axes (drone orientation during measurement) against the measurement setup in the original paper.
1. Data source
The Hugging Face dataset this Space reads, Goorm-AI-04/Drone_RCS_Measurement, has an empty dataset card and no license. Its contents, however, match the following open measurement dataset.
- Original: V. Semkin, J. Haarla, T. Pairon, C. Slezak, S. Rangan, V. Viikari, C. Oestges, Drone RCS measurements (26–40 GHz), IEEE DataPort, 2019. DOI: 10.21227/m8xk-dr55
- License: CC BY 4.0 (as listed on the Aalto University research portal). Reuse is free, but attribution is required.
- Paper: V. Semkin et al., "Analyzing Radar Cross Section Signatures of Diverse Drone Models at mmWave Frequencies," IEEE Access, vol. 8, pp. 48958–48969, 2020. (IEEE Xplore 9032332)
The mirror carries no license, but the original data is published under CC BY 4.0 and is citable, peer-reviewed research data. This Space reads the mirror and credits the original authors.
2. Measurement setup
| Item | Details |
|---|---|
| Environment | Anechoic chamber, quasi-monostatic (transmitter and receiver nearly co-located) |
| Frequency | 26–40 GHz in 1 GHz steps (mmWave, includes 5G FR2 bands) |
| Angles | θ −90° to 90°, φ 0° to 180° (as stored in the data) |
| Targets | DJI Phantom 4 Pro, F450, Mavic Pro, Matrice 100, Kyosho helicopter, Parrot AR.Drone, Walkera Voyager 4, custom hexacopter, Y600 tricopter frame, plus a 6S LiPo battery |
| Polarisation | HH, VV, HV/VH (all four for the Matrice 100) |
| Columns | f (GHz), theta, phi (deg), RCS (dBsm) |
3. Principles for reading the data
1. The unit is dBsm, not dB. RCS is expressed in decibels relative to 1 m². −14 dBsm ≈ 0.04 m².
2. Average in the linear domain. Averaging dB values directly gives the geometric mean, which for scattering data full of deep nulls comes out several dB below the true average reflecting area. The radar equation uses the arithmetic mean of σ (m²), so this Space computes every average linearly and shows the dB mean only for comparison.
3. Separate drones and polarisations. Pooling all 17 splits (drone × polarisation) into one histogram mixes distributions from different airframes and polarisations. Cross-polar returns (HV/VH) are much weaker than co-polar ones, so separating them matters. When showing a single distribution, state which split it comes from.
4. The left tail is not necessarily an "undetectable dead zone". Values below about −30 dBsm are mostly interference nulls, where returns from several scattering points cancel. These nulls move when the frequency changes by 1 GHz or the angle by 1°. The Deep nulls tab counts, for each angle cell, how many of the 15 frequencies fall below a threshold, separating nulls fixed to an angle (structural) from nulls that move with frequency (interference). Real-world detectability also depends on range, clutter, rotor micro-Doppler and integration.
5. Treat single-frequency "hotspots" with suspicion. A strong return in one (θ, φ) region at a single frequency may be an interference pattern. Check that it survives the all-band linear mean, and whether it appears in the same place for similar airframes, before attributing it to an antenna, battery or arm.
6. Interpret the coordinate system from the paper. Reading "θ = 0° means the radar is level with the drone" depends on how the drone was mounted and how the rotation axes were defined. Check the measurement setup in the original paper before translating θ/φ into elevation and azimuth.
7. Uses and limitations. As published, peer-reviewed measurements, the data is suitable for citation in research and teaching. State these limitations alongside it: - Measured in an anechoic chamber in a fixed pose; rotor spin, attitude changes and payload effects in flight are not captured - mmWave (26–40 GHz) only — results cannot be extrapolated directly to other radar bands such as X or Ku - One unit per model (no unit-to-unit variation)
4. Ideas for research and teaching
- Drone classification baseline: identify the model from (f, θ, φ) patterns. A model trained on this data already exists on Hugging Face (
Bombek1/DroneMamba-RCS) as a point of comparison. Splitting train/test by individual angle cells inflates scores through leakage between neighbouring angles, so split by angle blocks instead. - Dataset card exercise: writing a card for a mirror that lacks source, license and measurement details is a good teaching exercise in itself.
- dB statistics: a real-data example of why linear and logarithmic means differ.
5. Citation
@misc{semkin2019dronercs,
author = {Semkin, Vasilii and Haarla, Jaakko and Pairon, Thomas and Slezak, Christopher and
Rangan, Sundeep and Viikari, Ville and Oestges, Claude},
title = {Drone RCS measurements (26-40 GHz)},
year = {2019},
publisher = {IEEE DataPort},
doi = {10.21227/m8xk-dr55}
}
@article{semkin2020analyzing,
author = {Semkin, Vasilii and others},
title = {Analyzing Radar Cross Section Signatures of Diverse Drone Models at mmWave Frequencies},
journal = {IEEE Access},
volume = {8},
pages = {48958--48969},
year = {2020}
}
Related blog post (Korean): Small drone RCS analysis — Xdots in my habitus