Your radar cannot tell a drone from a bird

A radar does not see shapes. It sees energy that came back, and it measures how much came back, how long it took, and how the frequency shifted on the way. Everything else is inference. So when a vendor says a system "detects drones," the honest expansion of that sentence is that it detects small slow objects, and then something downstream has to decide which of those objects is a drone.

That decision is the whole problem. I spent my service as a Counter-IED and Electronic Warfare Officer, and the failure mode that costs you credibility is never the target you missed. It is the forty alerts a day that turn out to be wildlife, because after the second week nobody looks at the screen anymore.

Why the two targets look alike

Start with the physical resemblance, because it is closer than most people expect. A group of researchers at the University of Ottawa and Canada's National Research Council, writing in a NATO Science and Technology Organization meeting proceedings on tracking and classifying the two target types, put it flatly: "The flying birds have similar radar cross section (RCS), same velocity range, similar signal fluctuation, and approximate signal amplitude to drones." Their conclusion follows directly: "The similarity between small drones and birds present major challenges in class separation."

Rahman and Robertson, in a 2018 Scientific Reports paper that measured both target classes on the same radars, make the same observation in one line: "Birds have similar physical size to drones and also will fly at similar altitudes and speeds."

Look at the two lists they actually flew. The drones were a DJI Phantom 3 Standard at 1.216 kg, a DJI Inspire 1 at 2.845 kg, and a DJI S900 hexacopter at 3.3 kg. The birds were four birds of prey: a Northern Hawk Owl at 0.26 kg, a Harris Hawk at 0.71 kg, an Indian Eagle Owl at 0.97 kg, and a Tawny Eagle at 1.84 kg. A Harris Hawk and a Phantom overlap in mass and in the size of the thing that reflects. Neither one is going to distinguish itself by being big.

So the amount of energy that comes back does not sort them. Neither does speed, and neither does altitude. If your classifier is built on range, velocity and return strength, you have built a small-slow-object detector and labeled it a drone detector.

What actually separates them is the parts that move

Here is where radar earns its keep. A target's bulk motion produces one Doppler shift, but the parts of the target that move relative to the body produce their own additional shifts. That is micro-Doppler, and it is the closest thing counter-drone radar has to a fingerprint.

The two targets modulate the return in genuinely different ways, and Rahman and Robertson measured both at K-band (24 GHz) and W-band (94 GHz) simultaneously with phase-coherent FMCW radars.

A drone's propellers turn continuously, at high tip speed, in a plane. That produces "a wide signal spread over both sidebands about the bulk-Doppler," with two distinctive features on top of it. Rotor blade flashes appear as vertical streaks, each one an individual blade passage. HERM lines, from helicopter rotor modulation, appear as "approximately horizontal lines roughly parallel to the bulk-Doppler trace," and they arise because multiple propellers are turning at slightly different speeds. A quadcopter is four rotors that never agree, and that disagreement is legible.

A bird does none of that. In the same measurements, "there is no such wide micro-Doppler spread and instead the characteristic signature takes the form of periodic flashes corresponding to the wingbeat frequency." The wingbeat frequencies they measured across the four species ran 4 to 6 Hz. That is a slow, periodic, low-frequency modulation, and it does not look like a rotor.

There is a second difference that matters more operationally than the first. The bird's wing signal is strong relative to its body, "between 0 and 10 dB below the bulk signal." The drone's blade signal is weak, "typically 20-40 dB lower than the main body return," and it varies with blade aspect angle. Read that pair again. The feature you need in order to positively identify a drone is the faint one. The feature that identifies the thing you do not care about is the loud one.

Micro-Doppler can carry more than target class, incidentally. A 2021 study in Frontiers in Signal Processing used S-band (2.44 GHz), C-band (5.80 GHz) and W-band (77 GHz) radar to classify the payload weight a DJI Spark or DJI Phantom 2 was carrying, over payloads spanning 0 to 40 percent of drone mass, reaching mean accuracies of 86.05 percent at W-band and 80.95 percent at S-band with a six-feature K-nearest-neighbours classifier. The signature really does encode a great deal. The catch is what it takes to read it.

Three ways the trick stops working

Range. This is the big one, and it is why detection-range figures and classification-range figures are different numbers that brochures like to blur. The NATO proceedings paper states the constraint plainly in its framing of prior work: "at distance more than 500 m radar micro-Doppler is not detectable." Rahman and Robertson's own systems are consistent with that order of magnitude. They report typical detection ranges at 10 dB signal-to-noise of roughly 700 m for the K-band radar and 380 m for the W-band radar. A system may well see a target at several kilometres. Whether it can extract blade flashes from that target at several kilometres is a separate question with a much smaller answer, because the discriminating feature sits 20 to 40 dB under a body return that is itself getting weaker with the fourth power of range.

Birds stop flapping. The wingbeat signature only exists while the wings are beating. The authors are candid about this in their own data collection: "the birds used did not always flap their wings. Typically, they flapped their wings during take-off and landing and glide in the middle to conserve energy." A gliding raptor is a smooth, quiet, unmodulated return moving at drone-like speed. It has, in that moment, deleted the feature that would have exonerated it.

Aspect angle. Micro-Doppler measures motion along the radar's line of sight, so the geometry of the encounter changes what you can see. The paper notes that "due to the radial flight path, only the frontal aspect of the bird was seen by the radar. It is expected that the main body/wing signal strength ratio will change if the birds fly across the radar beam." The same dependence applies on the drone side, where blade signal strength varies with blade aspect angle. A signature characterised in one geometry is not a signature guaranteed in another.

The accuracy number, and what it was measured on

Because micro-Doppler runs out at a few hundred metres, the interesting research has moved to classifying targets by how they fly rather than by how they modulate. Birds and drones manoeuvre differently, and a track is available long before a blade flash is.

The NATO proceedings paper does exactly this. It feeds tracks from interacting-multiple-model filters into Bayesian, kernel SVM and decision-tree classifiers with 5-fold cross validation, and reports that "The decision tree classifier has a 97% true positive rate and a 3% false-negative rate."

Ninety-seven percent is the number that would end up on a slide. Here is what sits underneath it, from the same paper's own methods section, and this is the part worth internalising.

The real flight data amounted to 11 free flights of flocking pigeons recorded by miniature GPS devices at 5 samples per second, and 11 drone flights, also GPS. Then: "Due to limited amount of GPS data, we have synthetically generated trajectories comparable to drones and birds for the experiment." The 730 tracks that were split 584 for training and 146 for testing are largely generated from 3D kinematic models whose parameters the authors chose, with bird velocity drawn as N(11,5) m/s against drone N(16,4), and bird turn rate as N(0.3,0.15) rad/s against drone N(0.05,0.02).

That is a legitimate way to do early-stage work, and the authors state it openly rather than hiding it. But the honest reading of the 97 percent is that a classifier separated two synthetic distributions the experimenters specified, using GPS-derived trajectories rather than radar detections. It is a promising result. It is not a field false-alarm rate, and if it appears in a procurement deck as one, that is the deck's dishonesty and not the paper's.

The authors also describe their own failure mode, which is the realistic one: "a bird track with low manoeuvrability similar to that of UAV causes false alarm and sometimes a UAV track gets miss classified as bird due to its unstable flight because of environmental conditions." A soaring bird flying straight looks like a drone on a waypoint mission. A drone in gusty wind looks like a bird. The confusions are not random noise, they are structured, and they concentrate exactly in the weather and behaviour you will actually encounter.

What I could not confirm

Everything above comes from three sources I opened and read: the Rahman and Robertson Scientific Reports paper, the Mehta et al. NATO STO meeting proceedings paper, and the Dhulashia et al. Frontiers in Signal Processing paper. Where I quote, the words are theirs.

I did not obtain radar cross section measurements for these targets. Rahman and Robertson explicitly deferred that analysis, noting that while their data contain absolute RCS information, "the detailed analysis of their RCS is reported in a separate publication," which I did not open. My statements about RCS similarity rest on the NATO paper's characterisation and on the size and mass figures, not on a measurement I have seen.

I have not tested or benchmarked any commercial counter-drone radar, and nothing here is a claim about a specific product. Vendor systems are multi-sensor, they fuse radar with radio-frequency, optical and sometimes acoustic channels, and a fused system can perform better than any single channel described here. I have not evaluated whether any particular one does. I am also not claiming these three papers represent the current state of the art; classification research moves quickly and I read three papers, not a field.

Finally, the 500 m micro-Doppler figure is the NATO authors' statement about the technique in general. It is not a universal constant. It depends on transmit power, aperture, frequency, integration time and target size, and a purpose-built high-power system will do better than the research radars discussed here.

The signal

Three things to carry out of this.

First, ask for the classification range, not the detection range. They are different numbers and the gap between them is where your false alarms live. A system that detects at 5 km and classifies at 500 m is a perfectly useful system, but it means that for the outer 4.5 km you have a small-object tracker and a human making the call.

Second, ask what the accuracy figure was measured on. Real radar returns or GPS tracks. Field data or synthetic trajectories. Birds actually present or a bird class assembled from parameters. All four combinations exist in the literature, they produce very different numbers, and only one of them tells you anything about a Tuesday afternoon at your site.

Third, and this is the piece that transfers well beyond drones: the discriminating feature is usually the weak one. Anything can tell you something is there. Telling you what it is means reading a signal 20 to 40 dB down from the one that got your attention, and that is the signal that disappears first when range, weather or geometry turn against you. Detection is easy and it is what gets demonstrated. Classification is hard and it is what you actually bought.