HF and NVIS radio intelligence: software for long-range SIGINT

Why HF and NVIS matter for SIGINT

High-frequency (HF) radio occupies the 3–30 MHz band. At those frequencies, signals can refract off the ionosphere and return to earth hundreds or thousands of kilometres from the transmitter – a phenomenon called ionospheric skip. That single physical fact gives HF a collection geometry that no line-of-sight sensor can replicate.

Near-vertical incidence skywave (NVIS) is the tactical variant. An antenna pointed steeply upward launches energy nearly straight into the F-layer, which reflects it back down into a roughly 300–600 km radius "footprint" around the transmitter. Insurgent nets, border-crossing coordinators, and rear-echelon logistics all exploit NVIS because it covers ground that VHF/UHF cannot reach without relay infrastructure. That makes NVIS a primary collection target in deny-area-access and counter-insurgency environments, where adversaries deliberately avoid UHF/VHF links that are easily geo-located by compact DF arrays.

Skip propagation also enables deep-reach collection. A receive site inside friendly territory can intercept signals originating 1,500–4,000 km away, well beyond any tactical UHF horizon, without overflight or forward basing. The tradeoff is variability: ionospheric conditions shift with solar flux, time of day, and season. Effective HF SIGINT software must model these dynamics – not treat every band as flat and static the way VHF processing does.

The combination of NVIS for regional area coverage and skip for long-range intercept makes HF an enduring collection discipline. Legacy thinking that "HF is old" ignores the fact that adversaries choose HF precisely because it is difficult for peer-competitor SIGINT systems to geo-locate at range. Understanding the full SIGINT platform component stack begins with treating HF as a first-class sensor domain, not an afterthought.

HF receiver hardware and digitizer requirements

A capable HF collection front-end must cover at least 1.5–30 MHz continuously, with optional low-band extension to 100 kHz for LF/MF maritime and strategic emitters. Wide-band HF front-ends from vendors such as Rohde & Schwarz, Ettus Research, and Epiq Solutions can digitize 1–32 MHz of instantaneous bandwidth in a single channel. Wider instantaneous bandwidth increases the probability of intercepting frequency-hopping waveforms and ALE handshakes that dwell on each frequency for only milliseconds.

Analog-to-digital converter (ADC) dynamic range is the central hardware constraint. HF bands are crowded: a strong broadcast station at –30 dBm can coexist with a weak tactical net at –110 dBm in the same 500 kHz slice. The receiver must handle that 80 dB spread without the strong signal blocking or intermodulating into the weak one. Practical systems require at least 14-bit ADCs with a spurious-free dynamic range (SFDR) above 90 dBc. Delta-sigma converters running at 250 MSPS with decimation achieve this in current COTS SDR hardware.

Pre-selector filtering matters equally. Without a tunable bandpass pre-selector or a switched filter bank, broadcast AM stations (520–1700 kHz) will saturate the front-end amplifier, consuming headroom needed for weak-signal tactical intercept above 3 MHz. Military-grade HF receivers add low-noise amplifiers with switchable attenuation in 10 dB steps, controlled by automatic gain control (AGC) loops that respond faster than a hop dwell period.

S-meter calibration – converting raw ADC counts to dBm at the antenna port – is mandatory for SIGINT, not optional as it is for amateur radio. Emitter power estimation, propagation modelling, and multi-site geo-location all depend on calibrated received signal strength (RSS). Calibration requires an injected reference signal at a known level, temperature-compensated gain tables per frequency, and periodic re-validation against a traceable RF standard. Uncalibrated RSS measurements produce geo-location errors of several hundred kilometres at HF ranges.

Software-defined HF processing

GNU Radio remains the dominant open-source framework for HF DSP prototyping. The gr-hf out-of-tree module provides ionospheric channel simulation, HF AGC, and SSB demodulation blocks. For production deployments, however, GNU Radio's Python scheduler introduces latency and throughput ceilings that matter when processing 10+ MHz of continuous HF spectrum on a multi-channel receive array. CUDA-accelerated pipelines using NVIDIA's cuSignal library can process the same load in a fraction of the CPU budget.

REDHAWK SDR, the US DoD-sponsored component framework, provides a higher-level integration model. Components communicate over CORBA-based ports; waveforms are assembled as XML-described component graphs. REDHAWK's HF waveform library includes pre-certified demodulators for several STANAG modes, which shortens ATO approval timelines for programmes with existing REDHAWK infrastructure. The cost is framework overhead: spinning up a REDHAWK component graph adds hundreds of milliseconds of initialisation latency relative to a native C++ pipeline.

Custom DSP pipelines built in C++17 with FFTW3 and Intel IPP achieve the lowest latency and the highest channel density per compute node. A typical architecture decomposes the wideband HF stream into 3 kHz sub-channels using a polyphase filter bank (PFB), then feeds each active sub-channel to a mode classifier and demodulator worker thread. The PFB approach eliminates the guard-band waste of classic channelisation and keeps channel edges clean enough for adjacent-channel rejection without per-channel tuning. Coupling this to an SDR platform with GPU-accelerated FFT offload gives a path to real-time processing of 30 MHz of HF spectrum on a 2U rack server.

Signal activity detection on HF requires energy detection thresholds that adapt to the noise floor per sub-channel, per frequency, per time-of-day. A static threshold tuned for quiet nighttime conditions will trigger thousands of false positives in midday band conditions, overwhelming analysts. Recursive least-squares noise floor trackers with a forgetting factor around 0.999 converge quickly to local conditions and keep the false-alarm rate manageable.

HF mode library

A defence-grade HF processing pipeline must decode a specific set of waveforms. The following are non-negotiable for a complete capability.

AM and SSB/DSB. Amplitude modulation (AM) and single-sideband (SSB, also called J3E in ITU notation) carry the majority of HF voice traffic – military, paramilitary, and commercial. Double-sideband (DSB) appears on legacy military nets. Demodulating these modes is straightforward but correct AGC and carrier-insertion oscillator (CIO) phase tracking are prerequisites for intelligible audio at low SNR.

STANAG 4285. The NATO serial tone modem standard for HF data. It defines a single-channel, 2400 bps serial tone waveform with optional rate reduction to 75, 150, 300, 600, or 1200 bps. STANAG 4285 uses a known 80-symbol preamble that allows coherent carrier and timing acquisition. Every NATO-affiliated force uses or has used 4285 for encrypted data links. A demodulator must output soft-decision bits, not hard decisions, to feed a downstream FEC decoder correctly.

STANAG 4539. The high-throughput NATO HF modem, supporting up to 9600 bps in 3 kHz bandwidth using PSK and QAM constellations with adaptive rate selection. It introduces a longer preamble and a channel-quality metric that drives rate adaptation. Decoding 4539 at low SNR requires a minimum-mean-square-error (MMSE) equaliser with a channel estimate length of at least 40 symbols to handle HF multipath spreads.

ALE (Automatic Link Establishment, MIL-STD-188-141B/C). ALE is the handshaking layer beneath HF voice and data. It uses 8-tone FSK to exchange station IDs, link-quality analysis (LQA) scores, and call requests. Intercepting ALE reveals order-of-battle information – which stations are active, which are calling which – without breaking any encryption. An ALE decoder is therefore a high-value collection tool independent of the ability to decrypt traffic.

HFDL (HF Data Link). Used by civil aviation over oceanic tracks. HFDL intercept reveals aircraft positions and routing – relevant to maritime patrol and ISR coordination in permissive and semi-permissive environments.

Beyond these, a complete library includes: FSK variants (RTTY, SITOR-B), OFDM waveforms such as STANAG 5066 appendix C, and military-specific frequency-hopping spread spectrum (FHSS) waveforms. Mode classification – automatically identifying which waveform is present before demodulation – requires a trained convolutional neural network or a cyclostationary feature analyser. Manual operator identification is too slow when collection spans thousands of simultaneous sub-channels.

NVIS direction finding: AOA with small-aperture HF arrays

Direction finding at HF using angle-of-arrival (AOA) methods faces a fundamental aperture problem. At 5 MHz, the wavelength is 60 metres. A classical interferometer baseline needs to be a significant fraction of a wavelength to produce unambiguous phase difference measurements, which means baselines of 10–30 metres are practical – a small array by HF standards.

The Wullenweber (circular direction finding, CDF) antenna, historically the gold standard for HF DF, uses a circular array of 40–120 elements spanning 100–200 metres in diameter. It delivers 1–2° RMS azimuth accuracy across the full HF band. Few forward-deployed units can carry or emplace such a structure. Compact alternatives include:

MUSIC and ESPRIT with small loop arrays. Deployed in a cross-loop or Adcock configuration (four or eight elements on a 5–15 m baseline), these subspace algorithms can resolve multiple simultaneous signals and deliver 3–5° azimuth accuracy in moderate SNR conditions. The key requirement is coherent multi-channel digitisation – all array elements must be sampled with phase-locked ADCs referenced to a common clock. Any interchannel phase mismatch directly degrades bearing accuracy.

VHF vs HF DF accuracy tradeoffs. At VHF (100–500 MHz), wavelengths are short enough that a 1-metre aperture produces many phase cycles of differential path length, giving sub-degree bearing resolution. At HF, the same physical aperture produces a fraction of a phase cycle, making the bearing estimate sensitive to noise. A VHF DF system with a 2-metre array achieves better absolute angular resolution than an HF system with a 20-metre array. The advantage of HF DF is not angular precision – it is range. A single HF DF site can fix a bearing to an emitter 1,500 km away. No VHF system does that without satellite relay.

Multi-site HF DF is essential for geo-location. Two or three sites separated by 300–800 km, each contributing a bearing line, produce a fix by intersection. Time-difference-of-arrival (TDOA) at HF is practical only when the signal has enough bandwidth for sub-symbol timing resolution – narrowband HF voice (3 kHz) produces TDOA geo-location errors of tens of kilometres even with synchronised clocks. Wider-band waveforms, ALE preambles, and FHSS sync bursts produce better TDOA accuracy. Combining AOA and TDOA in a weighted least-squares estimator improves fix quality over either method alone. The full multi-site architecture is described in the direction-finding network architecture guide.

Integration: HF tracks into the common operational picture

HF SIGINT collection generates a different data type than VHF/UHF collection. VHF/UHF bearings are typically short-range, high-update-rate, and geometrically well-conditioned. HF bearings are long-range, update slowly (ionospheric conditions require revalidation), and carry larger geometric uncertainty ellipses. Fusing these into a single track picture requires a sensor model that encodes each measurement's accuracy as a function of frequency, propagation mode, and SNR – not a single covariance matrix applied uniformly.

The standard integration path outputs SIGINT tracks as ASTERIX or STANAG 4607 GMTI-format records, or as CURSOR-ON-TARGET (CoT) events over XMPP/TCP for TAK-compatible C2. Each track carries a signal descriptor (frequency, mode, estimated emitter class), an estimated position with uncertainty ellipse, and a timestamp of last activity. The receiving COP system merges these with VHF/UHF SIGINT tracks, radar tracks, and blue-force positions using a common data fusion engine.

Temporal alignment is the first integration challenge. An HF geo-location fix may have taken 10–20 seconds to accumulate enough bearing samples for a stable estimate. The fix timestamp must reflect the centre of that collection window, not the output time, or the fused track will show an apparent velocity artefact. Propagation delay from emitter to receiver – up to 10 ms at 3,000 km – is small relative to collection window duration and usually ignored, but at very high geo-location accuracy requirements it must be modelled.

The second challenge is emitter identity correlation. The same physical emitter may appear as separate tracks in HF, VHF, and UHF SIGINT, and in radar, depending on which systems are collecting. Associating these into a single entity record requires a multi-hypothesis tracker (MHT) that considers frequency, emission type, location overlap, and temporal coincidence simultaneously. Poorly tuned association logic produces track proliferation – the single emitter appears as four separate entities in the COP, misleading the analyst. The defence data fusion guide covers the association architecture in depth.

Operator interface design for HF SIGINT must surface propagation context that VHF displays do not need. A bearing line on an HF display should carry a visible skip zone annotation – the region close to the collection site where the ionospheric return cannot illuminate. An analyst who does not see the skip zone may incorrectly rule out nearby emitters. Similarly, multipath bearings – where two refraction paths from the same emitter arrive at different azimuths – must be flagged rather than silently discarded or presented as two separate emitters.

Discipline at the edge matters for long-range collection

HF and NVIS collection is unforgiving of engineering shortcuts that VHF systems tolerate. A poorly calibrated ADC, an uncorrected interchannel phase error, a noise floor tracker with the wrong time constant, or a skip zone that goes unannotated – any one of these degrades the collection picture in ways that only become visible when the intelligence product is checked against ground truth days later.

The hardware and software decisions described here are not independent. ADC dynamic range determines how wide a sub-channel bank is practical. The sub-channel width determines which waveforms are receivable in a single demodulator instance. The demodulator output feeds both the mode library and the DF pipeline, and both feed the fusion engine that produces COP tracks. A weakness anywhere in that chain propagates forward. Architects who treat HF as a straightforward port of their VHF processing stack consistently underestimate the differences and deliver systems that miss collection requirements in operational conditions.

Building a production-quality HF SIGINT pipeline – from wideband digitiser through polyphase channeliser, mode classifier, STANAG demodulator, multi-site DF correlator, and COP integration – takes deliberate engineering at each layer. The discipline to get the details right at the receive edge is what separates a system that works in a lab from one that produces actionable intelligence at operational range.