Scientists Built a $100 AI Camera Trap That Photographs Bees Instead of Killing Them

Researchers in Oregon have built a low-cost camera trap that uses AI to spot and identify wild bumblebees without netting or killing a single one. The system matched the species diversity that lethal traps and hand-netting turned up in the same field, at a fraction of the effort.

The story, reported by Digital Camera World, centers on a team at Oregon State University who wanted a way to survey pollinators that did not involve the two standard options. The university lays out the problem plainly. While trapping is widely used, it has the drawback of killing the very insects under study, and direct observation with nets is labor-intensive, time-consuming, and costly. Neither scales well when the goal is watching bees across large areas over long stretches of time.

The build itself is deliberately cheap. The scientists put together their traps from commercially available parts, including a Raspberry Pi, and housed each one in a weatherproof case with a solar panel and battery, at a cost of roughly $100 to $200 per unit. They deployed the cameras during bloom in a red clover seed field at Oregon State's Hyslop Field Research Laboratory in Corvallis, aimed at colored platforms that acted as visual lures. Over 18 days, the units shot continuous time-lapse frames through daylight hours. The AI acts as "an automated insect detector," as horticulture assistant professor Tim Warren put it, letting the camera save power when nothing is on the sensor and flagging the insects worth a closer look so researchers do not have to sort every frame by hand.

Two findings stand out. The team tested several visual patterns and found that a fluorescent bullseye drew significantly more bee visits than plain, uniform-colored backgrounds, which tracks with the way many flowers use concentric UV signals to advertise pollen. On the software side, they fed the images into custom deep-learning models and found that a tiling approach, which splits each image into smaller sections and analyzes them one at a time, beat models that judged the whole frame at once. The payoff: the camera trap recorded six bumblebee species and produced diversity estimates comparable to aerial netting. The findings appeared in the journal Remote Sensing in Ecology and Conservation.

Two points are worth keeping clear here. This is a research prototype that performed well in one field over roughly two and a half weeks, not a finished product you can buy. And the AI is doing detection and species identification, not some magic count of every bee in the county. The models were pretrained on bee images tied to species, then tested against the ground truth of traditional surveys. 

Several bumblebee species have been petitioned for federal protection, and the rusty patched bumble bee became the first bee listed as endangered under the U.S. Endangered Species Act back in 2017 after losing an estimated 87% of its historic range. Michael Getz, now a data scientist at the Biodiversity Research Institute in Maine who led the work as a master's student, framed the goal simply: "Insects are vitally important, and we need methods to better understand their populations," he said, adding that the tool should help do that in a low-cost, easily scalable way. Professor Tim Warren made the scaling argument too, noting that the biggest challenge in understanding pollinator health is gathering enough information across large landscapes. Cheap open-source hardware and models that run on it are how you get there. Regulators deciding which species deserve protection need population and movement data, and until now, that data has been expensive and thin.

Most of the AI conversation in photography right now is about generated pictures of things that never happened. This is the opposite use of the same underlying technology: cameras and machine vision pointed at the real world to document something small, fast, and genuinely hard to count. The next test is whether these traps hold up across more sites, more seasons, and more species than a single clover field in Corvallis.

Lead image by Tony Wills, CC BY-SA 3.0, via Wikimedia Commons.