A K-State entomology researcher says an artificial intelligence program he developed to identify bumble bees could help conservation efforts.
Brian Spiesman, research assistant professor of pollinator ecology, said he took a software program intended to identify or classify everyday objects and altered it to pick out the differences between bumble bee species.
“We didn’t start completely from scratch; we basically wiped this algorithm’s memory and used the same architecture of the program but retooled it to look for different types of bees,” Spiesman said.
Spiesman said he and his fellow researchers have submitted their findings to the online journal Scientific Reports. They are using a form of artificial intelligence programming called Computer Vision, which uses an artificial neural network that allows researchers to train the software to pick out bee species by comparing an image of the bee to thousands of other stored images Spiesman has fed it.
“With all of that we can basically train the network to learn what makes species different,” Spiesman said. “We don’t tell it things like, ‘Look for the shape and color of the wings;’ we just say, ‘Here’s all the images, figure out what makes them different,’ and it figures out on its own what makes them different.”
The program, which is available through the website BeeMachine.AI, has catalogued 36 of the approximately 46 species of bumble bees in North America. People upload a photo of a bee, and the website will identify it.
“If we feed it a photo of a completely different species it hasn’t seen yet, it’ll try to give an answer of one of the 36 it already knows,” Spiesman said. “It can’t learn new species without seeing them first.”
Spiesman said one of the cool things to come from this research is that he has not told the model what is important with identifying different bee species.
“Taxonomists have a particular set of features they look at to identify bees by known characteristics,” Spiesman said. “The algorithm doesn’t do that at all, it just looks at pixels, at ones and zeroes, and comes up with its own set of characteristics that may have nothing to do with what people are looking at.”
Spiesman said if he gives the AI a photo of a known species of bumble bee, the program correctly identifies it about 92% of the time.
“It’ll give us a list and sort of rank the likelihood of each species appearing,” Spiesman said. “The model will say, ‘I think it’s this one, but here are other possibilities.’ If we take into consideration those possibilities it’s correct about 97% of the time.”
Spiesman said he was not expecting the program to be so accurate. The researchers started with bumble bees because of the large amount of good image data available, he said, and now they are having conversations about how to introduce rarer bee species into the data set.
Spiesman said one of his ultimate goals is to make the program capable of using automated sampling methods. He wants to be able to load the AI software onto a mini-computer and aim a camera at a flower so the software can identify what kinds of bees are attracted to the flower using the live camera. He said this will help improve conservation efforts and potentially allow researchers to catalogue even more insects beyond bees.
“Instead of having to trap bees, we get the data instantaneously,” Spiesman said. “We don’t have to use lethal sampling methods, and we don’t have to wait months to get the data and begin analyzing.”
He said these methods are important for being able to study some of the more endangered bee species.
“At the end of the day, we’re out there to conserve bees,” Spiesman said.