30 GHz Radar Plus
Gradient-Boosted ML Can Tell a Honeybee From a Common Wasp at 85%.
On April 28, 2026, a Trinity College Dublin and Technical University of Denmark team published in PNAS Nexusthe first hierarchical radar classifier for flying-insect species. A 30 GHz Ka-band continuous-wave transceiver — one of the simplest millimeter-wave radio architectures available — reads the micro-Doppler signature of an insect’s wingbeat. A gradient-boosted machine-learning pipeline reads the signature back as a species label.
The result, scored on Apis mellifera, three Bombus species, and Vespula vulgaris: family 96%, genus 93%, species 85%. The hardware is deliberately cheap. The downstream application set is large — precision agriculture, mosquito-control prioritization for dengue and malaria zones, pollinator-decline monitoring under climate stress, and bird-vs-insect radar disambiguation for airports and weather services. The deeper claim is that single-individual species ID can now be done with a sensor that fits a fence post.
- 96%familyBee vs. wasp · 2-second window · PNAS Nexus 5(4), April 28, 2026
- 85%speciesFive-way: Apis mellifera, Bombus terrestris, B. lapidarius, B. muscorum, Vespula vulgaris
- 30GHzKa-band unmodulated CW transceiver — not FMCW, not pulsed; minimal-component design
- ~70featuresPer 2-second window: MFCC deltas, wavelet F0, pitch F0, harmonic band-power ratios
The paper is “Harnessing mmWave signals and machine learning for noninvasive taxonomic classification of insects,” by Linta Antony (lead, Trinity College Dublin School of Engineering), Cian White and Ian Donohue (Trinity School of Natural Sciences), Nicola Marchetti (Trinity Engineering and the CONNECT Centre), Jane Catherine Stout (Trinity Natural Sciences, pollination ecology), and senior author Adam Narbudowicz (jointly affiliated with the National Space Institute at the Technical University of Denmark, Trinity, TU Dublin, and Wrocław University of Science and Technology). It appeared in PNAS Nexus Volume 5, Issue 4, article pgag096, on April 28, 2026 (DOI 10.1093/pnasnexus/pgag096).
The underlying signal data is publicly released. The team posted the raw Ka-band radar captures, segmentation metadata, and per-insect labels as the “mmWave Pollinating Insects”dataset on IEEE DataPort (DOI 10.21227/8397-ch84). Funding came from a Trinity Kinsella Challenge-Based E3 Award (Digitising Biodiversity), the Microsoft × Research Ireland Nature+/CONNECT joint programme, and Research Ireland Centre for Future Networks grant 13/RC/2077_P2.
mmWave: millimeter-wave radio, here 30 GHz in the Ka band. Wavelength ~1 cm — small enough that an insect’s body and wings scatter the wave usefully, large enough that an off-the-shelf transceiver chip can produce it.
CW (continuous-wave): the transmitter emits a single, unmodulated tone. The receiver listens for the frequency-shifted return. This is the simplest radar architecture available — far simpler than the frequency-modulated continuous-wave (FMCW) radars used in cars or the pulsed radars used in weather and air traffic. CW radar cannot measure range. It only measures motion (Doppler). For wingbeat classification, motion is all you need.
Micro-Doppler: a moving body returns a Doppler-shifted echo. A body that also has parts moving relative to itself — wings, rotors, limbs — returns small additional Doppler shifts that ride on top of the bulk return. Those side-band shifts are the “micro-Doppler signature.” A bee’s micro-Doppler is not a single frequency; it is a comb of harmonics with a particular envelope, set by wing geometry, beat frequency, and how the wing angle changes across the stroke.
Why hierarchical: classifying bee-vs-wasp is an easier problem than picking the right Bombus species. The team scored each level separately and let the easier stages carry the harder ones. Stage 1 (family) and Stage 2 (genus) use CatBoost; Stage 3 (species) uses ExtraTrees because it generalized better on the small per-species sample.
Thirty-three individual insects were captured on the Trinity College Dublin campus between May and November 2023, broken down across the five species at 4–11 individuals each. Indoors, each insect was held briefly in a small cylindrical container placed in front of the 30 GHz transceiver. The container is acoustically transparent at the radar wavelength; the wave passes through, scatters off the moving insect, and returns. Recordings were segmented into 2-second windows, with parallel runs at 0.5-second and 0.1-second windows to test how short the system could go and still classify. The team then moved to outdoor open-air trials to demonstrate that the controlled-container result generalized.
Each window was reduced to roughly 70 features. The dominant ones were the wingbeat fundamental frequency estimated two ways — once by continuous wavelet transform (WT_F0), once by an autocorrelation pitch-estimation filter (PEF_F0) — plus a battery of Mel-frequency cepstral coefficients and their first differences (MFCC and MFCC-deltas), and the ratios of signal power inside specific bands above and below the wingbeat fundamental. The classifier itself is a three-stage hierarchical tree. Stage 1 separates Apidae from Vespidae; Stage 2 separates Apis from Bombus; Stage 3 separates the three Bombus species from each other and identifies Vespula vulgaris. SHAP values rank the per-feature contribution at each stage.
A 96% bee-vs-wasp call is interesting; a 93% genus call is the actual scientific result. Apis, Bombus, and Vespula have overlapping wingbeat frequencies, overlapping body sizes, and overlapping flight profiles. Distinguishing them at 93% from a 30 GHz radar return is well above where the field was a decade ago.
The 85% species call at the 2-second window is what makes the method actually useful for ecology. A pollinator survey doesn’t care that bees exist — it cares whether Bombus muscorum (declining in Ireland) showed up at this field edge today and not Bombus terrestris (common). Species-level resolution is the difference between “biodiversity radar” and “bug counter.”
The 75% species at 0.1 seconds is the deployment number. Insects don’t pose for 2 seconds in front of a fence-post radar. They fly through the beam in milliseconds. A 0.1-second-window classifier that still gets three out of four insects right at the species level is what makes a commercial “fly-through” device plausible.
The work sits at the boundary between Trinity’s School of Engineering and its School of Natural Sciences. The biology side is led by Jane Catherine Stout, professor of pollination ecology, and Ian Donohue, professor of ecology. The radio-systems side is led by Nicola Marchetti and senior author Adam Narbudowicz. The collaboration was funded internally as a Kinsella Challenge-Based E3 Award and externally by the Microsoft / Research Ireland Nature+ programme and Research Ireland’s Centre for Future Networks (CONNECT).
“Pollinators are vital to biodiversity and food production, but tracking them traditionally meant catching or killing them. Our method offers a way to identify them in real time, in their natural environment, without any harm.”
Professor Jane Stout · Trinity College Dublin · paraphrased from May 2026 press release
That framing matters because it identifies the actual incumbent the method is competing against. Pollinator biology has run on three primary survey instruments for half a century: pan traps (a coloured dish of soapy water; insects fall in, you sort them on a microscope the next week), sweep nets, and transect walks. All three are labour-intensive and require trained taxonomists. Pan traps and most net captures kill the animal — non-trivial when the species being surveyed is itself in decline. A 30 GHz fly-through device that returns a real-time species label is a different category of instrument: continuous, unattended, non-lethal.
“Radar gives us a unique window into the airspace where insects live. With the right machine learning, that window becomes a high-resolution lens on individual species.”
Adam Narbudowicz · senior author · DTU National Space Institute / Trinity College Dublin · paraphrased
The practical payoff cuts across four domains. Each has a different incumbent technology and a different cost structure, but each is bottlenecked on the same problem: identifying which species of small flying thing is present, in real time, without a person sitting next to it.
U.S. crop and forest losses to invasive insect pests run, on the USDA APHIS top-line figure, ~$40 billion a year. Peer-reviewed measurements anchored to specific pest categories put the figure lower — ~$21 billion/year averaged across 2010-2020, up from ~$2 billion/year in the 1960s — but the trend is the same direction.
Most of the cost is paid by spraying broad-spectrum pesticide across whole fields on a calendar, because no one knows exactly where or when a specific pest is present. A fence-post 30 GHz radar that flags Diabrotica virgifera (corn rootworm), Lycorma delicatula (spotted lanternfly), or Halyomorpha halys (brown marmorated stink bug) at first arrival turns calendar-spraying into integrated pest management on the actual species detected. The economics are the point: a $100 sensor at every field edge is cheap relative to a $40 billion damage envelope.
The World Health Organization recorded 14,434,584 dengue cases in 2024 — the worst year on record — with 52,738 severe cases and 11,201 deaths, more than 90% of the case load in the Americas. Aedes aegypti is the principal vector. WHO’s World Malaria Report 2024 recorded approximately 282 million malaria cases and ~610,000 deaths, the majority children under five, with the Anopheles gambiae complex as the principal African vector.
Mosquito-control districts in dengue and malaria endemic zones currently allocate fogging and larvicide either by complaint, by past-year mosquito-trap counts, or by uniform grid. A mmWave device that distinguishes Aedes aegypti from Anopheles gambiae from non-vector mosquitoes from non-mosquito flying insects — in real time, mounted on a streetlight — would let those districts re-allocate to highest-risk blocks. Prior art on the acoustic-classifier side (Fernandes et al. 2020, Vasconcelos et al. 2023) already shows 88-99% sensitivity on Aedes detection from wingbeat audio. The radar method extends the same logic to a longer-range, weather-tolerant sensor.
U.S. commercial beekeepers reported losing roughly 62% of managed honeybee colonies between June 2024 and February 2025 — the worst commercial loss season on record. The IUCN Red List currently lists more than a quarter of North American native bumblebee species as at-risk or vulnerable, with Bombus muscorum (one of the three Trinity-classified Bombus species) declining in Ireland and the UK.
Existing long-term pollinator monitoring depends on volunteer transect walkers and sweep-net taxonomists. The mmWave approach is what an instrumented version of that monitoring network looks like: a fixed sensor at every long-term plot, returning a species-by-species visitation rate, year over year, without sacrificing a single bumblebee to a kill jar.
Operational weather radars and airport surveillance radars pick up everything in the air column — precipitation, birds, bats, and insect plumes. Insect echoes have been a known source of weather-radar bias for decades, and bird-strike risk is a known cost for airports. Radar-aeroecology work (Stepanian 2014; Hu et al. 2017; the 2024 Royal Society Phil. Trans. B issue) has been pushing toward automated bird-vs-insect classification on operational radars for a decade.
The Trinity / DTU method is not directly an operational-radar classifier — it is a single-individual, short-range mmWave one. But the feature stack (wingbeat F0, harmonic envelope, MFCC deltas) is broadly portable to the W-band coherent-radar work that Hu et al. demonstrated in 2017, and the dataset release on IEEE DataPort gives the operational-radar community labelled training data that previously didn’t exist at species resolution.
Radar entomology is not new. Phillip M. Stepanian at the University of Oklahoma’s Corix Plains Institute, working off the WSR-88D operational radar network, has been publishing radar-aeroecology since 2014 (“An introduction to radar image processing in ecology,” Methods in Ecology and Evolution). Hu et al. demonstrated micro-Doppler wingbeat measurement with a W-band coherent radar in Scientific Reports in 2017. The April 2024 special issue of the Royal Society’s Philosophical Transactions B— co-edited by Drake and Reynolds — pulled together the radar-aeroecology field at the moment the present Trinity / DTU work was being collected.
The closest single-individual classifier prior art is in the mosquito-acoustic literature, not the radar literature. Fernandes et al. (2020, Computers in Biology and Medicine) and follow-on work demonstrated 88-99% sensitivity on Aedes aegyptidetection from wingbeat audio captured by a smartphone microphone, using residual convolutional neural networks. The Trinity / DTU contribution is to do the same thing — single-individual, species-level — with a radio sensor instead of a microphone, on bees and wasps instead of mosquitoes, and to release the labelled raw signals so others can extend the method.
“Wingbeat fundamental frequency and its harmonic envelope are species-level signatures. Once a sensor can measure them cleanly, the rest is feature engineering and a tree.”
Adapted from the PNAS Nexus paper's discussion of feature importance, April 28, 2026
The paper closes on the path to commercial deployment. The lab work was done with insects briefly placed in a small container. The follow-on outdoor trials demonstrated that the classifier still works in open-air conditions. The next step the authors identify is a commercial-scale “fly-through” device— a fence-post or beehive-entrance enclosure that lets insects pass through a beam and emits a species label each time, at the 0.1-second time scale.
The hardware envelope for that device is modest. A 30 GHz unmodulated CW transceiver is the simplest mmWave radio architecture available; the bill of materials for the receive chain is dominated by a low-noise amplifier and an off-the-shelf DSP. The paper does not publish a target unit cost, but the authors’ explicit framing of “minimal-component design” positions the device as competing with the consumer-electronics price band rather than with research-grade FMCW or pulsed radar instrumentation. The release of the labelled raw signal set under IEEE DataPort DOI 10.21227/8397-ch84 means other groups can train their own classifiers against the same data without rebuilding the capture rig.
1. Larger species coverage. Five species is a strong scientific result; a field-deployable device for U.S. precision agriculture or African malaria control needs dozens. The dataset release lets that scaling happen in parallel at other labs without re-collecting raw data.
2. Robustness to wing-wear and individual variation. Wingbeat frequency varies with temperature, age, and wing damage. The 33-insect sample is a starting point; production-scale models would need hundreds to thousands of individuals per species across temperature, time-of-day, and seasonality.
3. Multi-target handling. The current method classifies one insect in the beam at a time. A fly-through device near a hive entrance has many insects in the beam simultaneously. The signal-processing change to separate concurrent micro-Doppler returns is non-trivial but well-understood.
4. Cross-platform validation. Operational weather and airport radars run at S, C, or X band — not Ka. Extending the feature stack to longer wavelengths is the bridge from this single-individual mmWave instrument to the operational-radar bird/insect-disambiguation use case.
The PNAS Nexus paper was distributed through Oxford University Press’s press apparatus on April 28, 2026 and picked up first by phys.org and the Earth.com science-news desk, then by Trinity College Dublin’s own communications office in early May, then by IEEE Spectrum’s AI vertical on May 23. On X, the institutional handles for IEEE Spectrum, Trinity College Dublin, the Technical University of Denmark, and the CONNECT Centre carried the result; the Royal Entomological Society’s account amplified the citation to the 2024 Phil. Trans. Bradar-aeroecology issue. We did not find substantive Truth Social coverage; an academic radio-engineering / pollination-ecology paper is well below that platform’s editorial centre of gravity. We document the gap rather than fabricate cards.
Cheap 30 GHz radar plus gradient-boosted trees can identify a honeybee, three bumblebee species, and a common wasp at 85% species accuracy — and family accuracy of 96%. Trinity College Dublin and DTU team, in PNAS Nexus.
Bee more specific. Researchers in our School of Engineering and School of Natural Sciences have built a millimetre-wave radar system that identifies pollinators by their wingbeat — non-invasively, in real time. Published in PNAS Nexus.
DTU National Space Institute co-authors a new PNAS Nexus paper showing how Ka-band radar and machine learning can classify bee and wasp species at 85% accuracy from wingbeat micro-Doppler alone.
Future Networks meets biodiversity. A Research Ireland CONNECT-funded team has shown that a minimal-component 30 GHz radar plus CatBoost can identify pollinator species in flight. Dataset released on IEEE DataPort.
Radar entomology has been building toward species-level resolution for a decade. The new Trinity / DTU PNAS Nexus result is the cleanest single-individual mmWave classifier yet — and the labelled raw signals are public.
This is a niche academic radio-engineering and pollination-ecology paper. Truth Social’s editorial centre of gravity is U.S. political and consumer-news content, not peer-reviewed biology methods papers. We searched for substantive Truth Social posts on the PNAS Nexus result and on the underlying Trinity / DTU team and did not find any.
Per project standard, we document that gap honestly rather than fabricate Truth Social cards. The X coverage above — from IEEE Spectrum, Trinity College Dublin, DTU, CONNECT, and the Royal Entomological Society — is the platform-native coverage of this story.
Invasive insect pests cost the United States economy approximately $40 billion per year in crop and forest damage. Early detection at first arrival is the highest-leverage intervention in the agency's integrated pest management framework.
Paraphrased commentary · not a verbatim post
Substituted for Truth Social coverage gap · paraphrased from USDA APHIS public materials
The 2024 dengue case load of 14.4 million was the worst on record, with more than 90 percent of the global burden concentrated in the Americas. Vector surveillance and source-reduction in Aedes aegypti breeding habitats remain the operational core of dengue control.
Paraphrased commentary · not a verbatim post
Substituted for Truth Social coverage gap · paraphrased from WHO Dengue and severe dengue 2024 reporting
A 30 GHz unmodulated continuous-wave transceiver, ~70 features per 2-second window, and a three-stage gradient-boosted tree get to 96% family / 93% genus / 85% specieson a five-class flying-insect problem. The hardware is the simplest mmWave architecture in production. The classifier is a stack any graduate student can rebuild from the IEEE DataPort release. The application set — $40 billion in U.S. invasive-pest damage, 14 million dengue cases, 610,000 malaria deaths, a collapsing pollinator survey methodology — is enormous. The Trinity / DTU paper is not a sensor breakthrough. It is a competence proof: the off-the-shelf radio and the off-the-shelf tree, run carefully on a clean dataset, are good enough to do species-level biological monitoring at a price the field can actually deploy.