How AI helps detect low-observable and low-flying aerial targets

A radar sees a target by the energy that comes back. Radar cross-section (RCS) is the measure of how much a target reflects toward the radar, and it is not the same as physical size: shape, material and orientation all matter. A small object made largely of plastic and composites, or a shaped airframe designed to scatter energy away from the source, returns far less than a conventional aircraft. Lower RCS means a weaker echo, and a weak echo is easier to lose in receiver noise and clutter.

Geometry works against low flyers as well. Radar energy travels in roughly straight lines while the Earth curves away beneath it, so there is a radar horizon: below a certain height, a target at long range is simply not illuminated. Hills, buildings and tree lines add terrain masking, and over water, waves and multipath can hide or distort a low contact. Flying low is one of the oldest ways to shorten the time a radar has to see you — which is why low-flying aircraft detection depends so heavily on sensor placement and on combining sensors rather than on any one radar alone.

Slow speed is the third problem. Doppler processing separates movers from the static background, but a hovering or creeping target has a radial velocity close to that of the clutter itself, so the filter that removes the ground can remove the target with it. We cover that trade in detail in understanding radar clutter, noise and false targets.

Small drone detection is the sharpest form of the problem. A multirotor is physically small, often lightweight and plastic, can hover, and flies low. A flock of birds is also small, slow and low, and birds are far more numerous than drones. A system that raises an alert for every bird is quickly ignored; one that filters birds too aggressively can miss the one contact that matters.

The key is micro-Doppler. A target's body moves at one velocity, but its moving parts add their own small frequency shifts around it. Spinning rotor blades produce a rapid, regular, periodic signature; a bird's wings beat slowly and irregularly, and the body glides between beats. Seen as a time-frequency picture, those signatures look different, and that difference is exactly the kind of pattern a trained classifier is good at reading — more reliably than a hand-set rule on speed or amplitude.

The honest caveats matter. Micro-Doppler needs a sufficient dwell on the target and enough signal to resolve the moving parts, so it works best as one input among several, with a confidence the operator can see rather than a bare label. Because the goal here is counter-drone detection and identification, the output is an alert and a classification for a person to review.

Learned detection. A fixed CFAR threshold looks at one cell against its neighbours. A learned detector is trained on real examples of targets and real examples of clutter, and judges texture, shape, Doppler spread and persistence together. That lets it retain a faint, slow return while rejecting a clutter spike of similar strength, improving detection of hard targets and reducing false alarms together rather than trading one for the other. The broader argument is in the air-surveillance pillar.

Micro-Doppler classification. Once a candidate return is detected, a classifier trained on rotor and wing signatures can label it as likely multirotor, likely bird, or unknown, and say how confident it is. This is the step that turns 'something small is out there' into a usable alert, while keeping genuine uncertainty visible instead of hiding it.

Multi-sensor cueing. No single sensor sees every low target. A radar with a poor view of a valley can hand a rough position to an optical or infrared camera, and an acoustic or radio-frequency sensor can add a bearing. A system that fuses these into one track is far harder to blind with terrain or clutter than any sensor alone. We explain how conflicting sensors become one trusted track in multi-sensor fusion in air surveillance.

Low-flying and small-drone detection is also a latency and trust problem. Contacts can appear and vanish within seconds, the sensors are often at remote or exposed sites, and the data from them is sensitive. That argues for inference at the edge, close to the sensor, inside the operator's own perimeter, rather than a round trip to a distant cloud, and for models the operator can inspect and control. Our sovereign edge-AI family for sensor telemetry is Sanjaya; you can see the wider sovereign model portfolio.

The posture is decision support and observe-and-assist. A detection system should surface a contact, its classification and its confidence, flag when its inputs look degraded, and leave every consequential decision to a person who is in command. Detection and identification are the scope here; anything beyond that belongs to the operator and their chain of authority, not to the model.

AI in radar & air surveillance — the pillar: how AI turns raw returns into a recognised air picture.

Radar signal processing — what happens to the echo — detection, Doppler, CFAR and tracking.

Clutter, noise & false targets — why radar sees things that aren't there, and how it is cleaned up.

Multi-sensor fusion — turning radar, IFF and EO/IR that disagree into one trusted track.

Radar + EO/IR fusion — how a radar cue and a camera confirmation work as one picture.

AI in signal processing — learned detection, recognition and ECCM beyond fixed thresholds.