A radar's value is shifting from how far it sees to how early, and how honestly, it judges. Five stages sit behind that shift: intent prediction, threat assessment, collision risk, manoeuvre assistance, and the predictive layer that binds them.
---
![yuq3nqXGMobgbdl.jpg]()
In December 2018, Gatwick Airport closed its runway for the better part of two days after reports of drones near the perimeter. Around a thousand flights were cancelled or diverted; well over a hundred thousand passengers were displaced. No one was ever charged. It is still not entirely settled how many aircraft were involved.
The uncomfortable question is not whether the radar saw them. It is: what would it have done with that? At a civil airport you cannot jam or shoot down a drone. The only lever available is closing the airspace — and one false alarm costs more than a hundred correct detections earn. What was missing was never detection. It was knowing, before the boundary was crossed, that it probably would be.
1. THE BOTTLENECK IS NOT THE EYE, IT IS THE JUDGEMENT
Radar has told the same story for a century: further, finer, clearer. More power, a bigger aperture, a cleverer waveform. That road is now physically expensive. The range equation buries distance under a fourth root — double the detection range and you need sixteen times the transmit power.
Meanwhile the problem has changed shape. A consumer quadrotor might present 0.01 m² of radar cross section; a large bird sits around 0.05 m². They fly at the same altitudes, at overlapping speeds, and both hover and turn abruptly. To the eye of a radar, they are less distinguishable than intuition suggests.
[Figure 1 — Detection range scales as the fourth root of cross section. The horizontal axis spans five orders of magnitude; the vertical axis barely moves past four. This is why "make the radar more sensitive" has diminishing returns against small targets.]
So the industry has slowly conceded something: better detection does not answer the judgement problem. You see more and can explain less. Tracks on the screen go from dozens to hundreds to thousands, while operator attention goes the other way.
Predictive situational awareness is the attempt to fix that. It is not another signal-processing trick. It is a chain of five stages — echo to track, track to intent, intent to threat, threat to risk, risk to an executable manoeuvre — and if any link breaks, the conclusion downstream is worthless.
[Figure 2 — The five stages, and the feedback loop that separates this from a conventional detect-and-display pipeline.]
2. INTENT: PREDICT THE PLAN, NOT THE POSITION
The naive forecast is a point: extrapolate the current velocity ten seconds ahead. It works until the target manoeuvres, at which point it fails completely — because extrapolation assumes nothing changes, and manoeuvring is precisely the act of changing.
The modern approach asks a different question. Not "where will it be" but "what are its plausible intentions, and how likely is each". Those are intent hypotheses — cross the boundary, deviate north, turn back, loiter and observe. Each hypothesis owns a family of trajectories, and each trajectory carries a weight. What reaches the operator is not a line but a bundle that forks and widens.
[Figure 3 — One instant, three hypotheses, three backbone paths, and the sample clouds behind them. Only hypothesis B enters the protected area.]
With that framing, the decisive question becomes computable: what fraction of the plausible paths enter the protected volume? That fraction is the crossing probability — not a binary verdict, but a number that updates with every new observation. After Gatwick, a Cambridge team built exactly this on real drone radar data, using a Bayesian model to predict whether a target would enter restricted airspace, and doing so before it reached its next waypoint.[1]
"One waypoint early" sounds modest. It is not. It moves the decision from "it is inside" to "it is not yet inside" — and only the second window still contains options.
Where intent can actually be read
Intent is not written on the target. It has to be inferred from three sources, in ascending order of reliability.
Kinematics. Heading, speed, acceleration, turn radius, climb rate. Straight track toward a boundary, no deceleration, steady altitude — those three together are worth more than any single feature. Loitering, orbiting, drifting in and out of speed usually means probing. Kinematics are what a radar gives you for free, and also the easiest layer to spoof deliberately.
Micro-motion. Rotor blades, propellers or wingbeats modulate the main Doppler return into sidebands, symmetric or not. That is the target's own motion fingerprint. Multi-rotor sidebands are continuous and symmetric; bird wingbeats sit at lower, less regular modulation rates, and the two look plainly different on a spectrogram.[2] This layer answers a prior question: is this a machine at all.
Context and identity. Remote ID, AIS, flight plans, historical behaviour, terrain, boundary geometry. The most neglected layer, and often the most decisive — because a compliant drone following a filed route and a silent drone skirting a boundary can be kinematically identical.
Intent prediction means putting all three into one probabilistic frame. Kinematics-only systems fail on targets that imitate normal tracks. Identity-only architectures go blind the moment something stops broadcasting. The value of three layers is not that there are three, but that they fail differently.
[Pull quote] One distinction gets blurred constantly. Intent prediction asks what the target means to do. Threat assessment asks how much harm that would cause to you. The first is an inference about the target; the second is a judgement about yourself. Compressing both into a single number is why many systems behave erratically in the field.
3. THREAT: A FILTER CHAIN, NOT A NUMBER
"Threat level: high" reads as professional and carries almost no information. High in what sense — it can reach us, or it is about to? Is it dangerous, or merely badly placed?
The robust engineering answer is a staged filter chain. Each layer answers one question; only what passes goes forward. The practical benefit is diagnosability: when the conclusion is wrong, you can name the layer that failed.
[Figure 4 — Existence, capability, geometry, intent. Lower layers rest on measurement, higher layers on inference.]
Existence asks whether this is real. Birds, breaking waves, turbine blades, tarpaulin in the wind and ionospheric returns have all played the intruder. The governing metric here is false-alarm rate, and its cost is chronically underestimated: every false alert spends operator trust, and trust is not refundable.
Capability asks what it could do. Speed, agility, endurance, payload, reach. A twenty-minute quadrotor heading straight at you has a bounded worst case. A long-endurance platform with payload capacity and autonomous routing deserves tracking even from far away. Capability raises the ceiling on the worst case.
Geometry asks where it is relative to you — range, bearing, height difference, closure rate, time remaining, available margin. This layer is almost entirely a function of measurement accuracy, and it is where precision earns its keep: a hundred metres of margin error can turn "easily avoidable" into "manoeuvre now".
Intent asks what it intends, with what probability — and inherits directly from the previous section.
Only after four layers does a level make sense. And the correct use of that level is to allocate resources, not to decide whether to feel alarmed: which tracks get handed to EO for confirmation, which raise alerts, which stage a response, which are simply watched. Published figures for low-altitude surveillance radars show this constraint clearly — hundreds of tracks in search, a dozen or so tracked simultaneously, a three-second revisit, roughly 8 km against a 0.01 m² target.[3] Tracking capacity is finite, so something must rank; and the ranking instrument is threat assessment.
4. COLLISION RISK: FROM "WILL IT HIT" TO "HOW LIKELY, HOW BAD"
The traditional approach computes closest point of approach and time to CPA. Below threshold distance and below threshold time, alarm. Simple, explainable, cheap — and the well-known cause of alarm fatigue.
Because CPA describes geometry and ignores uncertainty. Two tracks 800 m apart with a CPA of 150 m look comfortable, until you learn the position error is ±200 m. The true relative position is a probability cloud; CPA is one sample from it, not the truth.
[Figure 5 — Conflict cone and closest approach. Neither trajectory actually collides, but one breaches the protected volume.]
So the criterion gets replaced. Instead of "will it hit", ask "will it breach the protected volume" — a volume around the target that carries a time dimension, so that predicted penetration constitutes a conflict. NASA's DAIDALUS writes this out cleanly: the volume is defined by a horizontal distance threshold, a vertical threshold and a look-ahead time, shaped like a capped cylinder with an extension along the closure direction equal to the distance covered in that time.[4] Detection does not fire on a fixed threshold; it continuously computes time remaining to violation and grades severity from "remain well clear" up to "urgent".
DAIDALUS also makes a deliberate simplification worth noting: it extrapolates the intruder as non-manoeuvring. That looks crude and is actually a design stance — the system answers "if it continues like this", and leaves the rest to the operator and to the manoeuvre guidance. It is a restrained piece of engineering philosophy.
The next step brings probability in. A usable risk quantity looks like this:
R = Σᵢ P(intentᵢ) × P(conflict | intentᵢ) × C(intentᵢ)
The first factor comes from intent prediction, the second from integrating trajectory uncertainty across the volume boundary, the third from severity — the same incursion over empty farmland and over a pipeline are not the same event. Only the product is a number you can rank with, set thresholds on, and defend to management.
None of this is new in principle. TCAS has used volumes, time thresholds and graded alerts for decades, and ACAS X goes further by formulating the decision as a Markov decision process, solved offline into a policy table that is simply looked up in flight.[5] The gain is that alerting is no longer stitched together from hand-written rules; the balance between missed alerts and nuisance alerts is optimised against a cost function instead of tuned by intuition.
5. MANOEUVRE ASSISTANCE: SHOW THE FORBIDDEN BANDS, NOT JUST THE PANIC
An alert that says "conflict" invites one immediate question: so what do I do? If the system cannot answer, half the value is gone — and the other half is just anxiety.
The productive formulation turns advice into the boundary of a feasible set. DAIDALUS sweeps heading, speed and vertical speed, marks every value that would produce a conflict as a forbidden band, and what remains is the feasible band.[4] What reaches the operator is a map, not an order: these headings are out, these flight levels are out.
That is better than a single commanded vector for one reason: it returns the decision to whoever knows the situation. The system does not know fuel state, task priority, or how familiar the operator is with that corridor. It knows geometry. Forbidden bands are constraints. A single instruction is a decision made on someone else's behalf.
[Figure 6 — The time budget of one response. Every stage from detection to execution consumes the total; the later a decision falls, the less room is left.]
This is why the work in the previous sections has to happen early. At three kilometres and two minutes out, you have nearly a hundred seconds to confirm, judge, decide and communicate. At three hundred metres and twenty seconds, the only option left is the most drastic one — which is usually also the most expensive and the most disruptive. Early judgement is not about hearing bad news sooner. It is about still being allowed to choose a gentle response.
One boundary worth stating plainly
None of this is an autopilot. It outputs probabilities, constraints and recommendations; the action is still taken by a person. That has to be a hard guarantee, for three reasons. Models trained on historical data fail quietly when behaviour leaves the training distribution. Sensor fusion raises confidence and also consensus-ises error, making it harder to spot. And in contested settings, spoofed signals, injected data and deliberately imitated flight profiles are real attack surfaces.
The international discussion is fairly explicit about the residual risks: data quality, interpretability, and operator over-reliance.[6] A responsible design therefore shows confidence and evidence chain together. Not "threat: high", but: heading held into the boundary for 47 s; no Remote ID broadcast; micro-Doppler consistent with a multi-rotor; 2.1 km from the boundary; entry predicted in 96 s. A judgement you can trace is a judgement that can be argued with; only a judgement that can be argued with deserves trust.
6. WHERE IT IS ALREADY LANDING — AND WHAT IS STILL UNSOLVED
Low altitude. Airport perimeters, energy sites, large events, urban corridors. The requirement has moved from "is something there" to "which few of these need action". Two Chinese mandatory national standards released in late 2025 — one on real-name registration and activation, one on operational identification — answer "who may fly and who is flying", and thereby supply the identity layer that intent inference was missing.[7] A companion test-method standard for low-altitude drone detection radar now fixes a common vocabulary for range, false-alarm rate, revisit and track capacity.[8] The parallel consensus is that no single sensor closes the judgement loop: radar supplies all-weather continuous tracking and direct Doppler, RF detection supplies control-link and identity clues, EO supplies visual confirmation, acoustics fill the near field. Fusion earns its keep by letting dissimilar sensors check each other's failure modes.
Inland waterways. An underrated case. A vessel has enormous inertia and almost no stopping distance, and a bridge pier offers no second chance. One university team combines millimetre-wave radar, AIS and vision for active bridge protection: radar for centimetre-level tracks, AIS for identity and tonnage, vision to verify manoeuvre, and a platform continuously computing heading offset from the bridge axis and real-time air draught.[9] A reported case involved a cargo vessel drifting off course with a fatigued helmsman, warned five minutes early. Five minutes is what makes a VHF call, a siren and a bridge loudspeaker possible at all — the gentle response of section 5, bought by early judgement rather than by a louder alarm.
Air traffic and unmanned systems. Civil aviation is furthest along, because the ACAS X transition was, at heart, the migration from deterministic threshold alerting to probabilistic state estimation. Unmanned systems, being more crowded and more cost-sensitive, are the better testbed for what comes next.
Four things remain genuinely unresolved. Data is scarcer than algorithms — labelled anomalous behaviour is by definition rare, and simulation-trained models inherit their simulator's biases. Interpretability fights performance — Bayesian networks explain but scale poorly, deep networks scale but hide their reasoning; hybrids that keep an explainable spine and use learned models at the edges are the honest compromise. Adversarial robustness — spoofing, ghost tracks and imitated profiles attack the premise that behaviour is learnable, not the model that learned it. And the human factor — the more accurate the system, the more relaxed the operator, and the heavier the consequence when it errs.
7. CLOSING
Radar's line has been consistent: "is anything there", then "where", then "what is it". What is now happening extends the third step into a fourth: "what will it probably do, and what should we do about it".
None of that is a single clever algorithm. It is whether the five stages can actually be joined — intent probabilities feeding the threat layers, threat layers feeding quantified risk, risk feeding the manoeuvre bands, and every conclusion traceable back to evidence. Break one link and the system degrades into an expensive display.
For the radar industry that shifts where competition happens. It used to be power, aperture, bandwidth, sensitivity. Increasingly it will be how much earlier the judgement arrives, how far it can be trusted, and how honestly the uncertainty is shown to the person who has to act on it.
At Gatwick that December, a system that could have said, ninety seconds out, "this one enters the boundary with probability 0.78, here are the four reasons, suggest contacting the operator first" would have changed the shape of the night. Those ninety seconds are the whole point of the chain.
TERMS USED
CPA / TCPA — closest point of approach, and the time until it occurs. The classical basis of collision alerting.
Well-clear volume — a three-dimensional safety boundary around the intruder, usually with a time dimension; breaching it is a conflict, well before physical contact.
Bands — the ranges of heading, speed and vertical speed that would produce a conflict, computed by sweeping each variable.
Micro-Doppler — modulation of the main Doppler return by local motion: rotors, propellers, wingbeats. The principal physical discriminator between machines and birds.
Remote ID (RID) — a drone's broadcast identity, position and operator information; effectively an electronic licence plate.
REFERENCES
[1] Liang, Ahmad, Godsill et al., Bayesian intent prediction for drones using real radar data, University of Cambridge, 2021 (presented at the Sensor Signal Processing for Defence Conference).
[2] On the use and limits of micro-Doppler for drone/bird discrimination, see the Everything RF technical review and the signal-processing literature it cites.
[3] Figures quoted from a publicly listed low-altitude surveillance radar (X-band, 2-D AESA, pulse-Doppler), used only to illustrate the capacity constraint.
[4] Muñoz et al., DAIDALUS: Detect and Avoid Alerting Logic for Unmanned Systems, NASA Langley Research Center; and associated technical reports on sensor uncertainty mitigation and dynamic well-clear volumes.
[5] On the TCAS II to ACAS X shift from deterministic thresholds to probabilistic state estimation, see RTCA DO-385 / EUROCAE ED-256 material and the Skybrary ACAS X entry.
[6] Joint Air Power Competence Centre, Faster Than the Threat — on AI/ML in threat detection, trajectory prediction and intent modelling, and the associated risks.
[7] GB 46761–2025, Real-name registration and activation requirements for civil unmanned aircraft; GB 46750–2025, Operational identification specification for civil unmanned aircraft systems. Released December 2025, effective 1 May 2026.
[8] T/CIE 321–2025, Test methods for low-altitude UAV detection radar, Chinese Institute of Electronics, effective June 2025.
[9] Public reporting on active bridge-collision monitoring at the School of Navigation, Wuhan University of Technology, Changjiang Daily, April 2025.
Product figures cited here come from public sources and are used solely to illustrate technical principles and engineering constraints, not to evaluate any specific product. All diagrams are schematic, not measured data.
