Bees and other pollinating insects play vital roles in food webs and crop pollination, yet monitoring them has proved difficult. That’s why researchers have developed a radar system that could lead to a cost-effective, noninvasive way to track pollinators.

Traditionally, identifying pollinators has proven tricky and time-consuming, and typically requires capturing and killing insects to get a close look at them. To find a better way to monitor pollinators, scientists are developing vision systems that use machine learning to automatically classify insects.

However, these machine learning systems face a major limitation when acquiring usable images, because of issues such as variable lighting, poor weather, and cluttered backgrounds—not to mention that many insects can just fly away when approached.

That’s why researchers based in Europe instead analyzed radar scans of insects. It’s not a new idea—scientists have used radar for decades to study migratory insects. However, that research was mostly focused on insects flying in large numbers at high altitudes, instead of lone insects flying near the ground, as pollinators do when visiting flowers.

“Typically, the radar reflection from single insects is very weak,” says Adam Narbudowicz, an associate professor of space research and technology at the Technological University of Denmark. “It’s probably impossible to detect them just by looking at a single point in time.”

Instead, “we hoped to be able to detect insects by integrating signals over longer durations,” Narbudowicz explains. Specifically, they focused on how insect wingbeats generate micro-Doppler signatures—distinctive time-varying patterns in radar reflections that arise from tiny motions, like wobbles. Micro-Doppler signatures allow radar systems to identify more subtle distinctions between objects, which can help the system distinguish between birds and drones, for instance.

Millimeter-Wave Radar for Pollinators

The scientists opted for millimeter waves for their radar system, because those wavelengths better match insect sizes than other portions of the radio-frequency spectrum. Millimeter waves have also found use in recent generations of cellular networks.

Narbudowicz and his colleagues trained a machine learning model on five species of pollinator insects, including honeybees and common wasps, captured on the campus of Trinity College Dublin. Several members of each species were individually placed in small plastic cylinders on top of a millimeter-wave antenna, which recorded their radar signatures. The insects were then released.

“In the beginning, we really weren’t sure it would work, as the insects are really small and the micro-Doppler signals we worked with were very weak,” Narbudowicz says.

The scientists had their model analyze more than 70 different features of the radar reflections from the insects. These included not only wingbeat frequencies but also the speed at which insect wing movements changed and the strength of the reflections. The researchers detailed their findings on 28 April in the journal PNAS Nexus.

Tiny differences in radar reflections, called micro-Doppler signatures, can be enough to distinguish between different insect species.Linta Antony, Cian White, et al.

“It’s fascinating to see how different species use their wings in different ways, and also that this can be observable in radar signals,” Narbudowicz says. “When looking at raw signals, it’s difficult to capture all the subtle details, but with sufficient machine learning you can distinguish those.”

The model was able to classify the five kinds of insects to the species level with 85 percent accuracy. When it came to more broadly distinguishing between the four bee species and the one wasp species—two different families of insects—it was able to do so with 96 percent accuracy.

The researchers found that species-identification accuracy improved the longer the insects remained within the radar system’s beam—for instance, accuracy went from 75 percent for 0.1 seconds to 84 percent for 1 second. They suggest developing traplike structures that insects can fly inside while the system examines them, releasing them unharmed after the analysis is complete.

“The power levels we use are below the levels that could harm insects,” Narbudowicz says. Comparatively, “a traditional insect trap relies on drowning the insect in poisonous cyanide liquid.”

Although this research focused only on pollinators, the scientists note that it could also help track pests and invasive species. Narbudowicz and his colleagues now want to develop a portable, field-deployable version of their technology. Ultimately, the scientists would like to collect more insect radar signatures to develop a global database enabling the instant classification of any insect within the system’s range.

The database could also include environmental data such as temperature and humidity to see how such factors might affect wingbeat frequencies and other pollinator features. The researchers suggest this database could one day also help monitor shifts in behavior through unusual changes in wingbeat patterns.