Pollinating insects form a vital part of any ecosystem, enabling the biodiversity that we see on Earth today. However, biodiversity is in rapid decline around the world, and monitoring insect species is a difficult task that often requires some insects to be killed. To support the conservation of biodiversity, which is critical to ensure the sustainability of human civilization, more robust monitoring is required. In a study published in PNAS Nexus, researchers have developed a new method to identify and classify individual insects, based on radar imaging and machine learning.
Radar has long been used to study migrating insects that fly at high altitudes and in large numbers, but such systems typically perform wide-area, long-range monitoring. However, thanks to a combination of millimetre-wave radar and machine learning, narrow focused identification is now possible, by detecting changes in the radar reflection of insects caused by the flapping of their wings.
“Having a background in antenna engineering, there was always the question of whether this technology can be used to address some of the environmental challenges that we’re facing,” says co-lead author Adam Narbudowicz from the Technical University of Denmark. “Some five or six years ago, we started talking with [co-author] Ian [Donohue] about those possibilities, and eventually the idea of micro-Doppler emerged, which seemed feasible from an engineering point of view and could provide some useful data on biodiversity.”
The approach taken in this study doesn’t focus on morphological features of the insects, as these are difficult to detect with radar. Instead, it uses the harmonic patterns generated by the micro-Doppler effect of an insect beating its wings as a detection strategy. Millimetre-wave radar can provide insight into biomechanical characteristics not visible with cameras, and these characteristics are encoded in the harmonic patterns of the wingbeat.
The team used machine learning to improve the accuracy of the identification and incorporated a SHAP (SHapley Additive exPlanations) analysis – an explainable AI tool that interprets and explains key outputs and prioritizes key features – to identify which signal features are the most critical for differentiating insect species. The SHAP analysed each insect across the full spectrum of micro-Doppler harmonics, extracting key features including fundamental wingbeat frequency, energy distributions, cepstral coefficients (sound signals) and how quickly an insect’s wing movement change. These data were then used to train the machine learning model.
The actual process of obtaining this data from the insects involved capturing insects at the Trinity College Dublin campus and placing them in a plastic box on top of a millimetre-wave antenna that recorded their radar signatures. The researchers then released the insects back into the wild. After data capture, the relevant micro-Doppler features were extracted from the data for model training.
The model allowed non-invasive monitoring of different insects and could distinguish between bees and wasps with 96% accuracy. The model also classified five key pollinating insect species – red-tailed bumblebee, buff-tailed bumblebee, moss carder bumblebee, western honeybee and common wasp – with an accuracy of 85%.
“I think the most impressive thing is that we can detect and classify them with such an accuracy. From a biological point of view, it’s impressive how different species beat their wings in different manner, and from an engineering point of view it’s fascinating how different wingbeats affect harmonics of radar micro-Doppler reflections,” says Narbudowicz. “Those differences are of course impossible to see just by looking at spectrograms, but it appears that a sufficiently trained machine learning algorithm can see them.”
Lidar tracks mosquito behaviour by monitoring wingbeats
Narbudowicz points out that the current study used precise lab-grade transceivers and a relatively controlled set-up, and that the natural next step is to move this technology to outdoor field deployment. “This requires a number of steps,” he explains. “Firstly, the device needs to be miniaturized, and battery operated; the transceiver will be less accurate than the one used in the lab, but a big problem is the ground truth verification, since in the field it can be difficult to verify exactly which species flew over the sensor.”
Despite the greater challenge with deploying the technology in the field today, the researchers suggest that this radar reflection approach could be utilized in the future in a fly-through device, which would make it much easier and cheaper to achieve non-lethal monitoring of insect biodiversity in different environments.
