Add one measured quantity at a time
Enable additional evidence sources. The lab deliberately shows no automatic "correct" classification, but rather how the ambiguity can be reduced step by step.
RCS, speed and size alone are often not enough
Drones and birds are not cleanly separated point clouds. Real measurements show broad, aspect-dependent distributions and overlapping kinematic ranges.
Micro-motion provides additional physical structure
In multirotors, rotating propellers modulate the Doppler trace. In birds, micro-motion arises from wing-beat and body movement. These mechanisms differ – but their observed signature depends strongly on geometry and SNR.
High rotation rates
Rotor blades can generate fast periodic components around the bulk Doppler.
Different periodicity
Birds typically show slower, differently structured micro-motion – depending on species and flight state.
No unique fingerprint
Hover, tangential motion, low SNR or an unfavorable aspect can weaken characteristic components.
At 77 GHz, rotor micro-Doppler lies far outside the unambiguous Doppler range
The micro-motion of a rotor is not just faster than that of a wing-beat – it is so much faster that, with typical chirp parameters, it is no longer sampled unambiguously at all. This is the point where waveform design and the classification question become connected.
around 72 m/s in the worked example
A 9-inch propeller at roughly 6000 rpm reaches, at the blade tip, a computed speed of around 72 m/s. At λ ≈ 3.9 mm, that corresponds to a Doppler offset of about 37 kHz. Other propeller geometries and rotation speeds lead to markedly different values.
unambiguous only up to about ±10 m/s
With a chirp repetition time of 100 µs, vmax = λ/(4Tr) is ±9.7 m/s, i.e. ±5 kHz. The blade tip exceeds that by roughly a factor of seven.
usually within the measurement range
Bird wing movements typically lie in the range of a few m/s and therefore often fall within the unambiguous range – a structural difference, not just a gradual one.
How robust is this number? It holds for the stated example parameters and cannot be generalized to drones as a whole: propeller diameter, blade count, rotation speed and flight state vary considerably, and so does the blade-tip speed. The observed extent additionally depends on the aspect angle, because only the radial component of the blade motion becomes visible – viewed perpendicular to the rotor plane, it largely disappears. What can be stated with confidence, therefore, is only the direction of the claim: rotor motion can clearly exceed the unambiguous Doppler range of typical chirp designs, while wing-beat does so less often.
What the aliasing means
The rotor components do not vanish, they appear elsewhere in the Doppler spectrum. A direct readout of the blade-tip speed is therefore not possible.
What can still remain usable
Periodic structure can be preserved despite aliasing; its spectral interpretation, however, becomes ambiguous. Time-resolved representations can still show recurrence, without allowing a direct conclusion about rotation speed or blade-tip speed. That is why observation duration is just as relevant here as the unambiguity limit.
Worked example with generic parameters to indicate the order of magnitude – not a statement about the Ozense waveform. The relationships to sampling and ambiguity are explained under FMCW Basics and CRT & Velocity Ambiguity.
A single frame can look similar by chance
Across several measurement times, periodicity, maneuvers, stability and repetition become visible. That is why track-centered time windows are often more informative than isolated spectra.
Measurement candidate.
Temporal association.
Target-centered signature.
Multiple frames.
RCS, kinematics, micro-structure.
Drone / other / Unknown.
95% accuracy is not 95% field performance
Published results depend on classes, distances, radar band, SNR, dataset size and split design. A value from a closed trial cannot be directly transferred to an unknown industrial site.
Example published test metric
Unknown domains
New targets, new sites, different SNR, weather, clutter and real false-alarm burden.
95% here is a didactic example, not an Ozense value.
The real world is not made of only two classes
Besides drones and birds, insects, other flying objects, reflection artifacts, clutter and completely unknown patterns exist. A system should therefore not forcibly assign every sample to a known class.
Drone
Evidence supports a known drone class.
Bird / other flying object
Evidence speaks against the drone hypothesis, but the exact taxonomy depends on the system goal.
Unknown
Evidence is insufficient or lies outside the known training space.
A credible drone/bird study must survive difficult splits
Particularly relevant are target-disjoint, session-disjoint and site-disjoint tests as well as independent ground truth. Otherwise there is a risk that the model learns recording conditions instead of target physics.
New target instances
A drone model should not merely recognize the same specific drone in a different flight phase.
New site
A dedicated site holdout checks whether clutter and background characteristics dominate the classification.
Physical separability is situation-dependent
Aspect Angle
RCS and visible micro-motion change with orientation.
Low SNR
Fine structure can disappear beneath the background.
Tangential motion
Low radial velocity reduces certain Doppler information.
Hover
Bulk Doppler can be small; rotor micro-Doppler can still be present but harder to measure.
Bird diversity
Birds are not a uniform class; size, wing-beat and flight state vary.
New drone types
Propeller count, size, rotation speed and flight state change the signature.
Physical evidence here – model architecture on AI/CRNN
This page answers which radar features can contribute to discrimination. The AI/CRNN page explains how a model processes such representations over time and why generalization remains difficult.
Further reading: Classification with CRNN · Radar Tracking · Clutter & CFAR.
Frequently asked questions about drone vs. bird
Can radar always distinguish a bird from a drone?
No. Separability depends on measurement quality, geometry, target type, SNR and available features. A credible system needs an Unknown/uncertainty path.
Is RCS enough for discrimination?
No. RCS distributions can overlap and vary strongly with aspect, frequency and target geometry.
Is micro-Doppler a unique fingerprint?
No. It is an additional evidence source. Observed signatures depend on aspect angle, SNR, observation time and processing.
Why is temporal information important?
Periodic and dynamic patterns only become reliably visible across several measurement times. This can improve separation compared with a single snapshot.
Sources
- Molchanov, P.; Harmanny, R. I. A.; de Wit, J. J. M.; Egiazarian, K.; Astola, J. (2014). Classification of small UAVs and birds by micro-Doppler signatures. International Journal of Microwave and Wireless Technologies 6(3–4), 435–444. DOI:
10.1017/S1759078714000282. - Rahman, S.; Robertson, D. A. (2020). Classification of drones and birds using convolutional neural networks applied to radar micro-Doppler spectrogram images. DOI:
10.1049/iet-rsn.2019.0493. - Palamà, R. et al. (2021). Measurements and discrimination of drones and birds with a multi-frequency multistatic radar system. IET Radar, Sonar & Navigation. DOI:
10.1049/rsn2.12060.