Laura Figueroa Collaborates on New Tool for Monitoring Bee Populations
Content
Populations of many pollinators, including bumble bees, have been crashing over the last few decades, thanks to the widespread use of chemical pesticides, loss of habitat, and the acceleration of climate change. But it can be hard to know just how bad the problem is, or whether conservation strategies are helping, without relying on lethal traps—which can kill tens of thousands of bees in order to get an accurate count of their populations.
“I’ve been working with bees for over a decade and can tell a bumble bee from a honeybee in the field by sound. If my brain can easily distinguish between bee categories, then we should be able to train an AI model to do it, too,” said Laura Figueroa, assistant professor in the Department of Environmental Conservation and the lead author of a new paper, recently published in Approaches in Entomology: Methods, Philosophies, and Ethics.
The need is acute. Pollinators—such as bees—are responsible for ecosystem services estimated at well over $200 billion worldwide, and are critical for the reproduction of most plant species around the world.
The gold standard for monitoring insect populations has historically been to send trained entomologists out into the field and have them collect insects to process back in the lab. To do this, most use lethal traps, often nothing more than small, colored cups filled with soapy water.
The problem—beyond the obvious irony in killing thousands of bees whose population health you’re trying to monitor—is that such traps are indiscriminate in what they capture, with high collateral damage to non-target species, and demand a lot of processing time after collection.
“While lethal traps will continue to be important for monitoring pollinators—many bee species can only be identified under a microscope,” says Figueroa, “there has to be a better way to monitor bee activity.”
That’s where AI comes in. In 2024 Figueroa showed that some AI models—deep learning neural networks—are very successful at insect identification.
Figueroa and her colleagues used what they learned to actually build two different pollinator-focused AI models—one that detects overall bee activity in audio recordings taken from a site, and another, which can distinguish between honeybees and bumble bees using the same recordings. Their models are available on Raven Pro, an acoustic analysis tool frequently used by scientists to analyze species activity.
Figueroa and her team trained their models, which specialize in analyzing bee sounds, on 594 field recordings taken over the course of 5 years from 43 different sites across the northeastern U.S. Their models achieved very high precision rates, correctly identifying 93% - 97% of the bees in a sample, and were highly correlated with visual observations made by trained researchers in the field.
“Our model was not developed to replace traditional methods,” says Figueroa, “but instead it allows us to get precise estimates of bee activity in times and places that would be difficult for researchers to sample, for example in tree canopies or during heatwaves. It’s a complimentary tool in a diversified toolkit that we need to effectively monitor pollinators in the field.”
This story was originally published by the UMass News Office.