AI 相机系统以低成本监测大黄蜂

By Sean Nealon

Researchers have developed a low-cost, semi-automated, AI-driven method that uses remote cameras to survey bumblebees and potentially other insects.

The new tool could have important implications for efforts to conserve declining insect populations . This includes bumblebee species, several of which have been petitioned to be listed under the Endangered Species Act.

Researchers also say the technology could benefit agriculture, given that many crops depend on insects as pollinators.

“Insects are vitally important, and we need methods to better understand their populations,” said Michael Getz, a data scientist at Biodiversity Research Institute in Maine, who led this research as a master’s student at Oregon State University. “This tool should help us to do that in a low-cost, easily scalable manner.”

The findings were recently published in the journal Remote Sensing in Ecology and Conservation.

Traditional insect surveys typically rely on one of two methods, both of which have drawbacks. Traps are often used, but they are lethal to the insects. Direct observation, often involving insect nets, is labor-intensive, time-consuming and costly.

The Oregon State research team built on recent studies showing trail cameras can capture images of insects in a noninvasive manner. Most previous work, however, focused on nocturnal insects such as moths, which can easily be attracted to cameras using light.

Less was known about how to attract and monitor diurnal insects, which are active in daylight. The researchers sought to determine the most effective way to lure bumblebees to camera stations, identify them automatically once photographed and compare camera-based monitoring with conventional survey techniques.

To accomplish this, the researchers built an inexpensive camera system using commercially available components and deployed it during bloom in a red clover seed field at Oregon State’s Hyslop Field Research Laboratory in Corvallis.

Alongside the cameras, researchers surveyed bumblebees using hand-netting and blue vane traps, a tool used by researchers to attract insects, which is deadly to bees. The camera system recorded six bumblebee species, closely matching the diversity of the species documented through netting and traps.

To identify the bumblebee images, they put the images into two types of custom-built deep learning image analysis models. They found that tiled models, which divide images into smaller sections and analyze each segment individually, outperformed models that evaluated entire images at once.

The researchers also tested several visual patterns placed behind the cameras to attract bumblebees and found bullseye patterns drew significantly more bee visits compared to uniform-colored backgrounds.

“One of the biggest challenges in understanding pollinator health and ecology is collecting enough information across large landscapes,” said Tim Warren, an assistant professor in the Oregon State Department of Horticulture. “These results show that we can gather reliable data with camera systems like this and begin to understand population and movement trends at a scale that hasn’t been practical before.”

The ability to generate large-scale monitoring data is especially valuable as more bumblebee species are evaluated for federal protection, said Andony Melathopoulos, a co-author of the paper and an associate professor of pollinator health at Oregon State.

Data gaps often complicate decisions about whether species should be listed under the Endangered Species Act and how critical habitat should be designated, he said.

“This technology opens up the possibility for much more richer data so that we can be more precise when it comes to decisions on listing a species and characterizing critical habitat,” Melathopoulos said.

Other co-authors of the paper are Lincoln Best, a bee taxonomist at Oregon State, and Oksana Ostroverkhova, a physics professor at Oregon State.