Chinese report

Editor’s note: This is an English translation of the Chinese article “海外低空安全现状与发展趋势·中篇” , published by 海外低空安全 (Overseas Low-Altitude Security) on July 15, 2026. The technical claims, figures, assessments, and source notes below reflect the original article. Wording has been translated for clarity without changing the substance.

The technical boundaries of counter-UAS systems

Core question: What can today’s counter-UAS technologies do, and what can they not do? What are the operationally effective scenarios and known blind spots of each technical approach?

Chapter 5. Detection Systems: The Operational Performance of Radar, EO/IR, and RF

The first hurdle in counter-UAS is not “what to use to engage it,” but “knowing what is in the air.” Detection may not be the most exciting part of a technical paper, but in operations it is the starting point of the entire chain. Radar, electro-optical sensing, radio-frequency sensing, and acoustics each have their strengths—and each has its blind spots.

This chapter does not list equipment specifications. Instead, it addresses a practical question: Do these detection methods actually work in real battlefields and urban environments?

5.1 Radar: Long Reach, Limited Clarity on Small Objects

Radar is the backbone of the detection system. Long-range early-warning radar (S/Ku band) can detect drone targets at 30–50 km, medium-range radar (X band) covers 10–30 km, and short-range radar (Ka/W band) focuses on detailed detection within 5 km. But a drone is not a fighter aircraft: its radar cross-section (RCS) is generally 0.001–0.1 m², only one-thousandth to one-hundredth that of an F-35.

The micro-Doppler effect—the biggest breakthrough in counter-UAS radar. Conventional radar relies on RCS to detect a target. A drone’s RCS is so small that it can be indistinguishable from birds and airborne clutter. Micro-Doppler is valuable because the signatures produced by rotating propellers are entirely different from those produced by flapping wings. Machine-learning models (CNN + RNN) can distinguish “DJI Mavic vs. pigeon” within one to two seconds at greater than 90% accuracy. Over the past five years, this technology has moved from the laboratory into operational use, but it remains an expensive “massive-data + high-compute” technology.

Radar’s three main blind spots:

Slow-moving or hovering targets. Conventional radar Doppler filters normally remove extremely slow targets as ground clutter. To the radar, a drone hovering at an altitude of 50 m is “stationary” and is filtered out. The latest moving-target-indication (MTI) systems require a specially designed low-speed filter mode, but this increases the false-alarm rate.

Multipath effects in urban canyons. In a city, radar beams reflect repeatedly between buildings, so a single target may appear on the display as three to five false tracks. An FPV racing drone flying two metres above the ground at 20 km/h can disappear completely into ground clutter—not because the radar has stopped working, but because “the signal is submerged in the background.”

Low reflectivity of carbon-fibre airframes. Consumer drones make extensive use of carbon fibre (RCS 90%); below 10 × 10 pixels, only detection is possible because shape and texture are lost and bird-versus-drone discrimination becomes difficult; below 3 × 3 pixels, only coarse motion detection remains and AI is largely ineffective; below 1 × 1 pixel (sub-pixel) lies the detection limit.

The ISP pipeline—an underestimated layer. With the same optical lens and sensor, optimising the image-signal-processing (ISP) pipeline can change detection range by 30%–50%. Noise reduction (each additional 1 dB of noise reduction is approximately equivalent to an increase in detection range), non-uniformity correction (removing fixed-pattern noise), contrast optimisation, and digital detail enhancement (preserving sharp edges for AI feature extraction) are low-level engineering improvements whose importance has long been underestimated in industry discussions.

Sources: Inside Unmanned Systems , “The Edge of Visibility—EO/IR System Design Realities for Modern C-UAS,” March 12, 2026; JAPCC, Chapter 5; FLIR Boson/Neutrino product lines; Aaronia AARTOS fusion architecture.

Note: No public quantitative data on the specific scale of overseas EO/IR deployments was available for this section. Its sources are mainly analyses of technical principles and publicly available vendor material.

5.3 RF Reconnaissance and Passive Detection: Listening for Signals, but Vulnerable to Silence

RF reconnaissance transmits nothing. It only “listens” to the link signal between the drone and its controller. As long as a drone remains in communication with its controller during flight, it is effectively “broadcasting its own position.”

The only means of locating the pilot at the signal layer. Radar and EO/IR can detect only the airframe. Passive RF detection is the only method that can trace the controller’s location from the signal itself. In law-enforcement scenarios, “catching the pilot” is often more important than “engaging the drone.” Location methods progress from coarse to precise: RSSI (coarse signal-strength estimate, >100 m) → AOA (angle of arrival, antenna-array phase difference of ±1°–5°, triangulation with two stations) → TDOA (time difference of arrival, three-station location accuracy of ±10–50 m).

The largest blind spot: silent or autonomous flight. Passive RF detection depends on the drone actively transmitting a signal. When a drone uses autonomous —visual SLAM + RTK positioning + a pre-programmed route—and neither depends on a controller nor transmits video, RF reconnaissance fails completely. This is the underlying principle of the most dangerous swarm tactic: “electromagnetically silent penetration.”

Other limitations include an inability to detect millimetre-wave or terahertz communications when the receiver’s frequency coverage is limited to 6 GHz or below; high false-positive rates amid dense urban Wi-Fi, Bluetooth, 4G, and 5G activity; frequency hopping across more than 100 channels, which requires simultaneous wideband reception; and complete signal blockage by buildings or mountains (the original article gives an approximate minimum line-of-sight distance of 5–7 km).

Product landscape. The AARTOS series (Germany) includes the ultra-mobile X2, with 3–8 km coverage, and the X9 for high-density RF environments, with 5–15 km coverage. Sentrycs (Israel, acquired by Ondas in 2025) provides passive RF detection, identification, tracking, and takeover; deployment by one person in five minutes; an Intel + Docker software-defined architecture; and a “Horizon Engine” that continuously analyses the RF spectrum with AI and automatically identifies new drone protocols. It has reportedly been operationally validated in 25 countries. Its principal advantage is that it neither jams nor spoofs, reducing the regulatory barrier to deployment.

Sources: Inside Unmanned Systems , “Aaronia Showcases C-UAS & EW Tech at Milipol Paris 2025,” November 25, 2025; Inside Unmanned Systems , “Ondas to Acquire Sentrycs,” November 5, 2025; JAPCC, Chapter 5.

Note: Some references in this section to Russian counter-UAS manuals come from secondary Russian-language sources. No independently verified public English or Chinese version was found.

5.4 Acoustic Detection: Not Ready Yet

Acoustic detection has a unique theoretical value: it is unaffected by radar stealth, GPS spoofing, or RF silence. If a drone is flying, its motors and propellers produce sound. In practice, however, the signal-to-noise ratio deteriorates rapidly beyond 50 m; urban background noise from traffic, air-conditioning, and crowds completely masks a drone’s acoustic signature; and separating sound sources becomes difficult when several drones fly at the same time.

Acoustics currently serve as “auxiliary confirmation” rather than the backbone of a counter-UAS detection system. They can corroborate other methods, but a single acoustic station cannot independently carry the detection mission.

Sources: CALL 26-1115 (operational assessment of acoustic detection); JAPCC, Chapter 5.

Chapter 6. Soft Kill and Hard Kill: From Spoofing to Physical Interception

Once a target has been detected, the question is “what do we do?” A common view in the counter-UAS community is that more than 80% of missions should first be addressed with electronic means. But electronic methods are not universal: autonomous swarms are resistant to them, and electromagnetic energy cannot be transmitted indiscriminately in dense urban areas.

This chapter examines the main countermeasure methods in turn: what each one can do, what it cannot do, and when it should be used.

6.1 Spoofing: The Most “Elegant” Counter-UAS Method

Near the ground, GNSS signal power is −130 dBm (on the order of 10⁻¹⁶ W), 20–30 dB below ambient thermal noise. A signal generator costing only a few hundred yuan and transmitting at milliwatt-level power can create a “false GPS signal” near the ground that is more than 100 times stronger than the real satellite signal. A drone’s GNSS receiver follows a “strongest signal first” priority rule, so the false signal naturally takes control. This is the underlying physics of spoofing.

The fundamental difference between spoofing and jamming. Jamming transmits noise that overwhelms the GNSS band. Once GPS fails, the drone switches to a failsafe mode—hovering or returning —and the pilot immediately knows the signal has been lost. Spoofing transmits a stronger false satellite signal, alters the reported position, and causes the drone to fly according to new coordinates while the pilot remains unaware of the attack: no warning, no abnormal indication.

“Push” spoofing directly replays a recorded false GPS signal. It is simple, but can easily cause the receiver to lose lock. “Pull” spoofing first synchronises with real GPS, then gradually increases signal strength and shifts the position, creating a smooth transition that the drone does not detect. “Pull” spoofing is the method most commonly used in operations.

“Virtual no-fly-zone” defence is the most mature tactical use of spoofing. A device deployed around a protected site generates “virtual GNSS coordinates” several kilometres away from the true position. After entering the area, the drone’s own GNSS shows that it has entered a no-fly zone—for example, the no-fly zone around a capital-city airport—and automatically triggers return-to- or landing. Parameters given in the original article are a spoofing radius of 500 m–5 km and an effect within 5–30 seconds.

A trapping case from the Russia–Ukraine battlefield—a textbook operational application. The original article states that Russian forces deployed large-scale false GPS signals “behind” known drone operating areas, causing Ukrainian FPV and reconnaissance drones to misjudge the return direction and fly toward Russian positions rather than Ukrainian positions. When the battery ran low, automatic return-to- reportedly brought them down in Russian positions. Tactical points given in the article are: the effect is best triggered below 20% battery; a 5–15 km displacement is large enough to change the landing point without triggering an abnormality warning; and reported success rates are above 70% against non-professional pilots and below 40% against professional pilots.

Core limitation—visual SLAM is the “natural enemy” of spoofing. If a drone uses cameras to “see” its route—visual SLAM + IMU + a pre-set route—and does not depend on GNSS at all, spoofing is completely ineffective. In addition, the P(Y) code (military GPS) and BeiDou B2a (encrypted) cannot be spoofed, according to the original article. Signal leakage may affect legitimate nearby GNSS users, including aircraft , phone positioning, and financial timing. Professional pilots may detect spoofing characteristics with a GNSS signal monitor and switch to pure inertial .

Sources: JAPCC, Chapter 6; Russian counter-UAS manual (Russian, 2023), chapter on GPS trapping tactics; a Chinese monograph on new-domain and new-quality operational theory, chapter on confrontation.

Note: Some references in this section are based on secondary Russian-language material and a Chinese academic monograph, without cross-verification from public English-language sources.

6.2 Electronic Jamming: The Oldest and Most “Primitive” Method

Electronic jamming has existed since the birth of radio. Omnidirectional jamming—360° coverage over 1–5 km with indiscriminate electromagnetic pollution—and directional jamming—a narrow 10°–30° beam over 3–20 km, with areas outside the beam remaining safe—represent two approaches with very different regulatory costs.

The central problem with omnidirectional jamming has never been the technology; it is electromagnetic pollution. Every wireless device using the same or adjacent bands within 5 km of the jammer may be affected: Wi-Fi interruption, Bluetooth disconnection, degraded mobile service, loss of GPS positioning, and radio failure. Near an airport, harmonics and spurious emissions may affect aeronautical bands even when the primary frequencies do not overlap.

Effectiveness declines against encrypted and frequency-hopping signals: fixed-frequency analogue >95% → frequency hopping (FHSS, such as DJI LightBridge) 60%–80% → frequency-hopping spread spectrum + encryption (DJI O3/O4) 30%–50% → military frequency hopping + encryption + spread spectrum 95% in test data; end-to-end latency from detection to recommendation 100 targets/s in DARPA OFFSET; and a 3–5× improvement in operator decision speed in semi-automatic mode, according to JAPCC.

Sources: JAPCC, Chapter 4; DARPA OFFSET AI module report; technical material on Israeli AI counter-UAS systems; US military C-RAM decision-support cases; counter-UAS command-and-control system architecture documents.

Note: Deployment cases for counter-UAS command-and-control systems come mainly from public US and NATO material.

7.6 AI Counter-UAS: Three Major Obstacles and the End-State Assessment

AI is not a silver bullet. Its weaknesses become clearer under closer examination.

First obstacle: false alarms. AI can only reduce false alarms from “extremely high” to “acceptable,” not to zero. Examples include bird flocks classified as drones, balloons and windborne objects classified as drones, missed detections in extreme weather, and low-flying helicopters or light aircraft classified as drones in airport environments. The original text says multimodal fusion can keep the false-alarm rate within a range ending at 5%, but the lower bound is missing from the source. It describes this as JAPCC best practice.

Second obstacle: adversarial attack. Examples include visual adversarial patterns that prevent AI recognition; RF camouflage that moves a control signal into bands not normally used by drones; radar-stealth design; and deceptive signals that imitate a drone and cause the counter-UAS system to track a false target. This is not science fiction: adversarial examples in image recognition have been shown to evade YOLOv8 detection consistently and reduce accuracy below 30%.

Third obstacle: data freshness. New models always appear before they enter the database. Counter-UAS AI is therefore always fighting with an “information delay.”

End-state assessment—AI vs. AI is not a compute race, but a data-operations contest. Whoever has the more complete and current dataset of drone signatures wins. The decisive capability of a counter-UAS system does not lie in laser power or radar range, but in having “more data, faster iteration, and a stronger closed loop.” Every engagement expands the dataset; every false alarm is an opportunity to correct the model. The system must “get smarter with every engagement,” rather than remain forever at its factory version.

Part II Summary: The Central Contradiction in Counter-UAS

From detection through engagement to AI, the three chapters in Part II define the technical boundaries of counter-UAS. There is one central contradiction: the attacker’s technology-iteration cycle is getting shorter while the defender’s equipment-deployment cycle is getting longer.

FPV took less than 18 months to evolve from a consumer toy into a battlefield weapon. Swarms moved from laboratory concept to tactical application in three years. Autonomous AI killing moved from idea to operational use in only one year. Yet selecting, tendering, deploying, and training on a counter-UAS system commonly takes more than two years.

This requires a fundamental change in counter-UAS design logic. Organisations can no longer expect to “buy one system and use it for ten years.” They need continuously updatable infrastructure in which sensors, AI models, decision engines, and countermeasure equipment are decoupled and can be upgraded independently. This is why software-defined systems, open architecture, and continuous data-based learning recur throughout the white-paper series. They are not mere technical preferences; they are directions imposed by operational pressure.

Translation source: 海外低空安全 (Overseas Low-Altitude Security), “Overseas Low-Altitude Security: Current State and Development Trends · Part II,” July 15, 2026. Translation prepared for publication by N-TET.