Every time a drone shuts an airport, the same explanation appears within hours: the things are simply too small for radar to see. It is a tidy story. It is also wrong, and the measurement literature has been saying so for a decade.

A DJI Phantom 3 and a pigeon were never measured against each other until a group at the University of St Andrews flew both past the same calibrated radars in 2019. The drone won. Not by a little: the Phantom’s modal radar cross-section came out around minus 20 decibels per square metre, a pigeon’s has been measured at minus 28.2 at X-band. The quadcopter is the bigger radar target.

So the drone is not invisible. Radar can see it perfectly well. The reason it keeps getting through is a far more uncomfortable answer: we spent fifty years deliberately teaching radar to throw away exactly this kind of target, and we were right to, until about 2015.

The drone is not invisible

Start with the numbers, because almost everything written about this subject gets them wrong by an order of magnitude or three.

The most systematic open dataset comes from a Czech team who put ten drone types on a turntable in an anechoic chamber at 9 GHz, calibrated against a 300 mm steel sphere. A DJI Phantom 4 returned a median of 0.020 m². A hexacopter-class DJI F550 managed 0.030 m². Even the tiny Yuneec Mantis, 190 mm across, came in at 0.004 to 0.006 m².

A separate group measured six DJI airframes in a compact-range chamber at Ku and K band in both polarisations. A Mavic Pro averaged minus 17 dBsm — about 0.02 m². A Phantom 4 Pro, minus 15 dBsm, around 0.032 m².

These are small targets. They are not vanishingly small. The figure of 0.0001 m² that circulates in trade coverage does not appear in the measurement literature as a drone value at all — it turns up as the extreme low tail of instantaneous fluctuation, as a back-scaled model, or as the measured cross-section of a bat. And the 0.01 m² that appears on radar datasheets as “the” small-drone figure is a convention the industry agreed on, not a measurement of any particular aircraft at any particular angle.

The bird is not a disguise. It is a twin

In 2019 Samiur Rahman and Duncan Robertson at St Andrews did the experiment everyone had been citing second-hand: they flew three DJI drones and four birds of prey past the same FMCW radars, at 24 GHz and 94 GHz, calibrated against a known trihedral. Not simulations, not chamber models — live targets, at 70 to 100 metres, with a falconry centre supplying the birds.

The results demolish the cliché. Average modal radar cross-section at 24 GHz: Phantom 3 Standard, minus 20 dBsm. Inspire 1, minus 15. S900 hexacopter, minus 8. Northern Hawk Owl, minus 30. Harris Hawk, minus 22. Indian Eagle Owl, minus 21. Tawny Eagle, minus 19.

Read the ordering. The Phantom sits above the hawk and far above the owl, and the largest bird in the test returns more than the smallest drone. More damaging still is the spread. The Phantom’s 10th-to-90th-percentile band runs from minus 29.5 to minus 12 dBsm. The Indian Eagle Owl’s runs from minus 30 to minus 12.5. Those are, to within rounding, the same interval.

A Chinese team reached the same conclusion from the other direction, calculating that the maximum detectable ranges of a pigeon and a drone on their radar were 12 km and 11 km respectively. And Polish radar engineers writing for NATO put the scale of the nuisance in one line: 95 per cent of birds have a radar cross-section smaller than 0.16 m², which means the bird population’s range of returns entirely contains the drone population’s. Assuming one bird per square kilometre, they note, there would be more than 1,200 birds within 20 km of a radar.

Every one of them is a candidate drone.

Why there is no single number

Here the physics gets interesting, and the popular explanation gets abandoned entirely.

The standard line is that a drone is too small relative to the radar wavelength — that it sits in the Rayleigh scattering region where return falls off with the fourth power of size. At X-band and above, this is simply false. Using the thresholds in Knott’s chapter of the Radar Handbook, a 35 cm quadcopter has a ka value of about 34 at 9.4 GHz, deep into the optical region, where anything above ka of 10 lies. Rahman and Robertson say it outright from their measurements: at K-band and W-band, drones and birds fall in the optical scattering region.

In the optical region, return scales roughly with projected area. The drone’s problem is not that it is electrically small. It is that it is physically small — a few hundred square centimetres of actual reflecting structure, most of it hidden inside a plastic shell.

And what reflects is not the shell. The armasuisse group modelling a Phantom 2 found significant differences when the battery pack was left out of the model, with the motors treated as solid copper and the cable harness contributing substantially. A millimetre-wave study of nine airframes found the frame dominates for carbon-fibre builds — carbon fibre is conductive, and plastic-framed drones measured about 7 dB below carbon-fibre ones. The propellers barely matter for detection: rotating elements are at least 20 to 25 dB weaker than the main body.

The consequence is that the echo is an interference pattern between a handful of small metal objects, and it swings violently with aspect. That Phantom 4 with its tidy 0.020 m² median ranges from 0.004 to 0.27 m² across viewing angles at a single frequency — a factor of about 70, or 18 dB. Any datasheet quoting one number for a drone is quoting a convention.

The underlying signal-processing concepts — Doppler, MTI and clutter rejection — that the rest of this article turns on.

The fourth root is not the villain

Radar detection range scales with the fourth root of radar cross-section. This is usually presented as the reason small drones defeat radar, and it is worth doing the arithmetic because it argues the opposite.

A thousand-fold reduction in cross-section — 30 dB — costs a factor of 1,000 to the power of one quarter, which is 5.62. A peer-reviewed study of an L-band air-surveillance radar published relative range factors that agree exactly: a minus 30 dBsm target is detected at 0.178 of the range of a 1 m² target.

Turn that around. A radar that holds a 1 m² target at 50 km still holds a 0.001 m² target at 8.9 km. For defending an airfield perimeter, that is not merely adequate, it is generous. Raw sensitivity is not what is failing.

Geometry is the villain

The same study, examining an L-band AESA air-surveillance radar with a published instrumented range of 470 km, produced the most startling number in this entire subject.

With the antenna 30 metres up and the beam on the horizon, the radar horizon against a target at 60 m altitude is 54.5 km. Tilt the beam up by 2.5 degrees — as you must, to look at the airspace a surveillance radar exists to watch — and the beam axis now intersects 60 m altitude at 690 metres. At 20 degrees of elevation, 80 metres.

A national air-defence radar built to see a fighter 470 km away cannot be pointed at a quadcopter creeping along at rooftop height more than a few hundred metres from its own mast. Not for want of power. The beam is simply somewhere else.

And the filter is the real villain

Point the beam low and you buy the other problem. Clutter — terrain, buildings, vegetation, wind turbines, vehicles, rain — can return thousands of times more energy than the target hidden inside it.

Radar solved this decades ago with Moving Target Indication: reject everything whose radial velocity is near zero, keep what is moving. It works beautifully against hills and car parks. It also happens to describe a hovering quadcopter exactly. At 1.5 GHz, one metre per second of radial velocity produces ten hertz of Doppler shift. A drone hovering over a runway threshold, or flying a crosswise track so its radial velocity is near zero, produces single-digit hertz. It lands inside the clutter notch — the band of Doppler frequencies the radar has been deliberately engineered to discard.

Radar engineers at Poland’s PIT-RADWAR put the trap in a single sentence to a NATO audience, and it is the thesis of this whole article.

For fifty years, slow and low and small meant birds, lorries, trees and weather. Suppressing them was not an oversight; it was the single most valuable thing a radar could do. Their same paper warns that upgrading an existing radar to find low, slow, small targets may be very difficult and require substantial tradeoffs. The installed base of air-surveillance radar across Europe was optimised, correctly, for a world that ended around 2015.

An academic at the Oxford Thermofluids Institute, asked last November why European airports kept closing, chose his words with more precision than most of the coverage that quoted him.

Not configured to. Not incapable of.

Micro-Doppler, and what it cannot do

There is a real answer, and it is elegant. A drone’s rotors spin at roughly 100 revolutions per second, and each blade modulates the return into sidebands and into a ladder of horizontal stripes on a spectrogram known as HERM lines. A bird flaps at 4 to 6 hertz. Two orders of magnitude apart.

Better still, the HERM line spacing equals rotation rate times blade count, and is independent of the radar frequency — so a radar can read a drone’s rotor speed and number of blades straight off the spectrogram. Convolutional neural networks trained on this have reported testing accuracies of 94.4 per cent across four classes and 98.3 per cent for the simple drone or not-a-drone question.

The cleanest demonstration of the micro-Doppler principle: a vibrating tuning fork written into a radar spectrogram.

Now the caveats, because the 99 per cent figures come from laboratories.

The micro-Doppler signature sits 20 to 30 dB below a bulk return that is already around minus 20 dBsm. It needs dwell time a surveillance radar sharing its attention across a full scan may not have. The UCL group that has done much of this work is candid that their results came from ideal conditions with a hovering drone of the same type at short ranges to maximise signal-to-noise, and that an unknown model flying erratically to evade detection will likely have a much reduced classification success rate.

Aspect ruins it: at middling elevation angles the body shadows the rear rotors and the HERM lines collapse. Hexacopters produce a confused pattern as six rotors’ contributions overlap. You cannot optimise pulse repetition frequency for blade flash and HERM lines simultaneously. And a gliding bird produces no wingbeat modulation at all — which makes it, to a classifier, indistinguishable from a drone that has stopped its rotors.

Most soberingly, one peer-reviewed field study simply failed to find the effect: identifiable micro-Doppler signatures cannot be obtained in our radar data, its authors reported, despite expecting them from both wingbeat and blade rotation.

A fidget spinner stood in for a rotor: the same physics that lets a radar count a drone’s blades.

There is a nice irony in who solved this best. The Dutch company whose radars now sit on military bases across Europe began as a bird-radar firm, built to keep flocks away from airports. Its founder has been direct about where the advantage came from.

One more thing worth noticing about the counter-UAS radar market: essentially no manufacturer publishes a detection range against a stated radar cross-section. They publish instrumented range, or range against named drone classes. Blighter is a rare exception and its numbers are instructive — a nano-quadcopter out to 3 km, a walking person out to 5 km. A human being on foot is an easier radar target than a drone.

So they stopped fighting the physics

Faced with all this, the practical answer for a decade was to give up on seeing the drone and listen for its pilot instead. Radio-frequency detection picks up the control link, and a US Department of Transportation audit of the FAA’s airport trials found radar and radio frequency were the two most common methods, with electro-optics and acoustics trailing.

That audit also found the limits. Airport radio spectrum is congested, which made detection difficult and, in some instances, not possible. Sensor coverage distance limited how well a drone could be located. Operators spent a lot of manpower discerning false positives. Technical readiness was low. This was a government auditor describing detection — not interception — as the binding constraint, and it was published in 2022.

Then two things broke the radio-frequency half of the answer.

The first was competence. When drones were repeatedly flown over Belgian military sites in late 2025, the defence minister described the failure on the record.

The second was fibre optics. Since emerging at scale in Russia’s Kursk region in August 2024, spooled fibre-optic control cables thinner than fishing line — standard models beyond 30 km, some Ukrainian units reaching around 40 — have spread across the front. A drone trailing a glass thread has no radio link to jam, and, by the same token, nothing for a radio-frequency sensor to hear. The countermeasures being improvised against them are telling: shotguns, physical barriers, radar tripwires, acoustic sensors. The problem goes straight back to radar and optics.

Which is how we ended up shooting rockets at toys

If you cannot reliably detect the thing, and you cannot reliably jam it, what is left is making sure that the moment you do see it, killing it is cheap.

Hence the APKWS laser-guided rocket, a guidance kit bolted onto an unguided Hydra-70. The Pentagon placed an initial order worth 322 million dollars against a 1.7 billion dollar ceiling in December 2025 for tens of thousands of rounds, at a per-shot cost reported at 24,900 dollars — against an AIM-9 that can exceed half a million. Hence high-power microwave, where one manufacturer reported disabling all 61 drones flown at a live-fire demonstration, including a 49-drone swarm.

But cheap interceptors are a reply to the economics, not to the detection problem. And even a cheap interceptor runs into the same wall the Belgian minister described in his next breath: over a military base you can shoot a drone down, but nearby you have to be careful, because it can fall on a house, a car, a person.

So the honest summary is not that drones are invisible to radar. A Phantom returns more energy than a pigeon and rather more than most owls. Radar can see it.

The trouble is that we spent half a century building a sensor network whose beams point at airliner altitudes and whose filters are tuned to discard everything slow, low and small — because for fifty years, everything slow, low and small was a bird. That was the correct engineering decision for every year until roughly the one in which a consumer quadcopter became something you could buy for the price of a mobile phone. Unlearning it is turning out to be expensive.

Sources: Rahman and Robertson, In-flight RCS measurements of drones and birds at K-band and W-band, IET Radar Sonar & Navigation 13(11), 2019, and Radar micro-Doppler signatures of drones and birds, Scientific Reports 8:17396, 2018; Sedivy and Nemec, Drone RCS Statistical Behaviour, NATO STO-MP-MSG-SET-183; Ezuma et al., Radar Cross Section Based Statistical Recognition of UAVs at Microwave Frequencies, 2021; Schröder et al., IEEE Radar Conference 2015; Semkin et al., IEEE Access, 2020; Patel, Fioranelli and Anderson, IET Radar Sonar & Navigation 12(9), 2018; Gong et al., Progress In Electromagnetics Research M 81, 2019; Meller and Rutkowski, NATO STO-MP-SET-311; Braun and Hegyi, Sensors 26(13), 2026; Knott, Radar Cross Section, in Skolnik (ed.), Radar Handbook; Ritchie, Horne and Peters, UCL; US DOT Office of Inspector General report AV2022026, 30 March 2022; Euronews, 3 and 20 November 2025; Atlantic Council UkraineAlert, 24 February 2026; Air & Space Forces Magazine, 10 December 2025; Blighter A800 Mk 2 datasheet; Robin Radar Systems product literature.