The Classifier Is the Bottleneck
Millimeter-wave radar can identify pollinators from their wingbeat signatures, but the model has to be trained on the species mix a particular grower actually has in the field. That problem is fundamentally regional—and the institutional machinery built to solve regional agricultural problems already exists. It is called Cooperative Extension.
Bottom Line Up Front
Why the Training Problem Will Not Solve Itself in the Market
The Antony et al. demonstration used three European species—Apis mellifera, Bombus terrestris, and Vespula vulgaris—deliberately chosen because they are common, behaviorally distinct, and amenable to controlled lab data collection. Even within that small set, the classifier required several thousand wingbeat samples per species to converge.1 A working U.S. system needs reference signatures for, at a minimum:
- The Western honeybee (Apis mellifera) and any commercially relevant managed alternatives—blue orchard mason bee (Osmia lignaria), alfalfa leafcutter (Megachile rotundata), bumblebee species being commercially propagated for greenhouse and tunnel crops.
- The native bumblebees—roughly 46 Bombus species in North America, with regional dominance shifting from B. impatiens in the East to B. vosnesenskii in California to B. occidentalis in the Pacific Northwest.
- The solitary bees that do disproportionate work on specific crops—Peponapis on cucurbits, Andrena on early tree fruit, Xylocopa on passion fruit and a range of legumes.
- The non-bee pollinators that increasingly matter under managed-honeybee scarcity—syrphid hoverflies, beetles, certain moths and butterflies.
- The non-target insects that produce confusable wingbeat signatures—yellowjackets, paper wasps, bee-mimic flies, and certain beetles.
That is, conservatively, a couple of hundred species before the classifier is operationally useful, and the species mix that matters in any given orchard might be only a dozen of them. Each species needs controlled data collection with positive identification by a trained taxonomist, which is the binding labor constraint.
The reason a private firm will not build this library on its own is that the economics do not close. A single grower cannot justify funding species training that benefits competitors; a single sensor company has no incentive to release reference signatures into the public domain because they are differentiating intellectual property; and the reference data is jointly produced—it requires field entomologists, taxonomic specialists, and radar engineers working together. This is the canonical structure of an applied-research public good, and the canonical institutional answer in American agriculture is the land-grant university extension system.
Why the Land-Grant Model Fits This Problem Exactly
Three features of the U.S. agricultural research system make it the right institutional vehicle for regional radar-classifier training, and all three are direct consequences of the 19th- and early-20th-century statutes that created it.
The Hatch Act of 1887 established State Agricultural Experiment Stations at every land-grant institution, with a specific mandate for applied research keyed to local crops, climates, and pests. The Smith-Lever Act of 1914 created the Cooperative Extension Service, which tied federal, state, and county funding together to push research findings out to working farmers through county extension agents. The Morrill Acts created the land-grant universities themselves; every U.S. state has at least one (some have two or three, including 1890 institutions and tribal colleges in the system).
The pollinator-monitoring problem maps onto this structure with almost no friction. The State Agricultural Experiment Station produces the reference data through controlled field collection and lab radar measurements. The Cooperative Extension specialist translates the resulting classifier into a tool the working grower can actually use. The county extension agent, who already visits farms in the district, distributes and supports the deployment. None of this is hypothetical—exactly this institutional pipeline already operates for soil testing, varietal trials, integrated pest management thresholds, and pesticide applicator certification.
It also already operates for pollinators specifically. EPA's Managed Pollinator Protection Plan framework has, since 2017, routed federal funding to state Departments of Agriculture for pollinator stewardship; in Michigan, for example, the Department of Agriculture and Rural Development has secured annual EPA funding flowing to MSU Extension to implement the state's MP3, with MSU producing crop-specific stewardship guides for blueberries and vegetables and providing pesticide applicator training that includes pollinator-protection certification.2 USDA-NIFA's Crop Protection and Pest Management Program supports this work through grants such as 2021-70006-35450 to MSU and the North Central IPM Center.2 The National Managed Pollinator Protection Plans Working Group hosted by MSU Extension coordinates extension entomologists across at least a dozen states.3
The radar classifier training program would slot into this existing structure as one more funded activity under the same MP3 / NIFA umbrella, not as a new program requiring its own statutory authority.
A Concrete Proposal: Federated Regional Classifiers
The right architecture is federated. Each land-grant entomology program, in partnership with its state Department of Agriculture, builds a local classifier covering the species mix that matters in its state's principal pollinator-dependent crops. Each contributes its labeled training data to a national reference library. The classifier becomes a public artifact distributed at no cost to growers in the region, with annual model updates as new species are added.
| Function | Performing Institution | Existing Analogue |
|---|---|---|
| Strategic priorities, crop ranking, regulatory hooks | State Department of Agriculture | State Pollinator Protection Plan; pesticide registration authority |
| Field collection of reference specimens, taxonomic identification | Land-grant entomology department + State Agricultural Experiment Station | Variety trials; insect taxonomic reference collections |
| Radar data collection on identified specimens, classifier training | Land-grant electrical engineering / agricultural engineering programs in partnership with entomology | UAV imaging programs at most ag schools; precision agriculture labs |
| National reference library, model federation, cross-state validation | USDA-ARS regional bee labs (Beltsville MD, Logan UT, Tucson AZ, Baton Rouge LA, Stoneville MS, Davis CA) | USDA-ARS National Plant Germplasm System |
| Distribution to growers, deployment support, training | Cooperative Extension Service, county extension agents | NRCS Web Soil Survey delivery; IPM threshold tables |
| Funding | EPA MP3 grants; USDA-NIFA SCRI and CPPM; state Specialty Crop Block Grants | MSU Extension MP3 funding model (NIFA grant 2021-70006-35450) |
The output the grower sees is a downloadable, regularly updated classifier model that runs on whatever sensor platform the grower has bought (or rented). The classifier itself—the list of species it can recognize, and how well—is a public artifact in the same sense that USDA Plant Hardiness Zones, the NRCS Web Soil Survey, USDA-NIFA crop variety trials, and Cornell's eBird are public artifacts. Sensor manufacturers compete on hardware and dashboard; they do not compete on whether Bombus impatiens versus Bombus griseocollis is correctly classified, any more than they compete on whether the soil under the orchard is a Hanford sandy loam.
Why a Public-Goods Approach Beats the Commercial-Only Path
A purely commercial path produces three failures the federated public-goods architecture avoids.
First, it fragments the data. Each sensor vendor builds a proprietary species library; libraries do not interoperate; cross-vendor data comparison becomes impossible; researchers studying pollinator decline cannot pool data across deployments. This is the IoT-platform-fragmentation problem that has plagued precision agriculture more broadly, and it slows everyone down.
Second, it under-serves the long tail of species. A vendor optimizes its classifier for the species that drive sales—honeybees in almond country, bumblebees in greenhouse tomato—and ignores wild and rare species because there is no margin in identifying them. But the wild species are precisely the ones that ecologists, regulators, and the growers themselves need data on, both because of their ecosystem-service contribution and because they are the indicators of habitat health.
Third, it locks growers into single-vendor stacks. A grower who buys a sensor from Vendor A loses access to its classifier when switching to Vendor B; switching costs are high; competition is reduced; long-run prices are higher. This is the same dynamic that has driven the open-data and right-to-repair conversations across the rest of agriculture.
The public-goods classifier path fixes all three: data is federated, the long tail of species gets covered because the marginal cost of adding a species to a public library is paid once and amortized across all users, and growers can take their classifier with them across hardware vendors. None of this prevents the sensor companies from competing vigorously on hardware and software around the classifier; it simply takes the species library off the table as a moat.
Forward Action Items
Three things should happen this funding cycle if the institutional pipeline is to be ready when the radar hardware is fielded.
- State Departments of Agriculture should add radar-classifier training as an explicit deliverable in their next MP3 cycle. The MP3 framework already accommodates research and monitoring activities; New York's Pollinator Protection Plan, for example, lists "research and monitoring of pollinators to better understand, prevent and recover from pollinator losses" as a core pillar.4 Adding a radar-classifier line item is a marginal change.
- Land-grant entomology programs should partner with their on-campus electrical and agricultural engineering departments to stand up reference-data collection capability. The hardware—a 60–94 GHz mmWave radar sensor, a clean-room cylinder enclosure for individual specimens, and the analysis pipeline—is well within the capability of any university with an active radar or precision-agriculture lab. The Trinity College team's published methodology and dataset provide a working starting point.1
- USDA-ARS should designate one of its existing bee research labs as the national reference repository. Beltsville, Maryland is the natural choice given its central role in the 2025 colony-loss forensics; Logan, Utah is the alternative given its specialization in non-Apis pollinators and its existing relationship with Western land-grant institutions.5
None of this requires new statutory authority, new federal programs, or new funding lines. It requires reading the existing program language broadly enough to recognize that training a radar classifier is the same kind of activity as running a variety trial: place-based, applied, publicly funded, publicly distributed, and indispensable to a working agriculture sector that needs to know more about its pollinators than it currently does.
Verified Sources
- Antony, L., White, C., Marchetti, N., Donohue, I., Stout,
J. C., and Narbudowicz, A., "Harnessing mmWave signals and machine
learning for noninvasive taxonomic classification of insects," PNAS Nexus, vol. 5, no. 4, pgag096, April 2026. DOI: 10.1093/pnasnexus/pgag096.
https://academic.oup.com/pnasnexus/article/5/4/pgag096/8662959 - Michigan State University Extension, "Michigan Pollinator
Protection Plan Resources" (MSU implements Michigan MP3 with EPA funding
via Michigan Department of Agriculture and Rural Development; supported
by USDA-NIFA Crop Protection and Pest Management Program grant
2021-70006-35450 and the North Central IPM Center, NIFA grants
2018-70006-28883 and 2022-70006-38001).
https://www.canr.msu.edu/resources/michigan-managed-pollinator-protection-plan - National Managed Pollinator Protection Plans Working
Group, hosted by Michigan State University Extension. Membership
includes extension entomologists at MSU, Cornell, Penn State, Oregon
State, University of Nebraska-Lincoln, University of Minnesota, and the
Institute of American Indian Arts Land-Grant Programs.
https://www.canr.msu.edu/resources/national-mp3-working-group - New York State Department of Agriculture and Markets,
"Pollinator Protection" (state Pollinator Protection Plan including
research and monitoring as a core pillar).
https://agriculture.ny.gov/plant-industry/pollinator-protection - USDA-ARS regional bee research laboratories, Beltsville MD
(lead lab for the 2025 colony-collapse forensic investigation), Logan
UT (Pollinating Insect Biology, Management, and Systematics Research
Unit), Tucson AZ, Baton Rouge LA, Stoneville MS, and Davis CA.
https://www.ars.usda.gov/news-events/news/research-news/2025/usda-researchers-find-viruses-from-miticide-resistant-parasitic-mites-are-cause-of-recent-honey-bee-colony-collapses/ - USDA, "Pollinators" (overview of NIFA grants to land-grant
institutions and Cooperative Extension System for pollinator
information transfer).
https://www.usda.gov/farming-and-ranching/plants-and-crops/plant-breeding/pollinators - "Bees Are Gold: Almond Pollination in 2026," AgNet West, February 2026 (BeeHero interview at World Ag Expo on tight bee supplies after the 2024–25 commercial colony collapse).
https://agnetwest.com/almond-pollination-2026-bee-hero/ - Humboldt County Beekeepers event listing describing the
Harry H. Laidlaw Jr. Honey Bee Research Facility at UC Davis as "the
largest and most comprehensive state-supported apiculture facility in
North America and the only one in California." UC Davis Department of
Entomology and Nematology, Bee Biology Program, "People."
https://beebiology.ucdavis.edu/people/ - E. L. Niño Bee Lab, UC Davis, "Dr. Elina L. Niño" (UCCE
Apiculture Specialist since 2014; founded California Master Beekeeper
Program in 2016; lab portfolio includes precision beekeeping research).
https://elninobeelab.ucdavis.edu/people/elina-nino - "Historic Occasion: USDA-ARS Bee Lab Opens on UC Davis Campus," Entomology & Nematology News, UC ANR (federal-state co-location of USDA-ARS bee research with UC Davis bee biology faculty).
https://ucanr.edu/blog/entomology-nematology-news/article/historic-occasion-usda-ars-bee-lab-opens-uc-davis-campus - California Department of Fish and Wildlife, "State and Federally Listed Endangered and Threatened Animals of California" (Bombus crotchii
reinstated as CESA candidate species 30 September 2022 following Third
District Court of Appeal reversal). California survey protocols: CDFW,
"Survey Considerations for California Endangered Species Act (CESA)
Candidate Bumble Bee Species," June 2023.
https://nrm.dfg.ca.gov/FileHandler.ashx?DocumentID=109405 - California Grants Portal, "Pollinator Habitat Program"
(CDFA, $15 million allocation under Budget Act of 2021, SB 170 Skinner;
eligible recipients include Resource Conservation Districts and the
University of California system).
https://www.grants.ca.gov/grants/pollinator-habitat-program/ - Environmental Science Associates, "California Advances
Species Protections to Counter Potential Federal Withdrawal," 17
December 2025 (Governor Newsom signed AB 1319 on 11 October 2025
establishing the "provisional candidate species" category under CESA,
effective through 31 December 2031).
https://esassoc.com/news-and-ideas/2025/10/california-advances-species-protections-to-counter-potential-federal-withdrawal/
No comments:
Post a Comment