- Frog species are in decline worldwide due to habitat loss, disease, and climate change
- Machine learning improves frog call recognition to help identify migration patterns
- Undergraduates gain valuable experience with high-performance computing, using Jetstream cloud
Everyone knows that a frog says ribbit, but did you know that their calls have a deeper, more impactful meaning?
Frogs are considered “canaries of the coal mine,” for their ability to indicate ecological or environmental crises before humans notice a broader impact. Frogs have permeable skin, making them vulnerable to even small changes in their aquatic environment.
As a result, 32 percent of frogs are in decline due to a combination of habitat loss, disease, and climate change. This statistic worries scientists, so ecologists across the nation are eager to study declining frog populations.
“It is imperative that national frog surveyors are able to accurately identify where frogs are and where they are moving,” says Tenecious Underwood, a participant of the Jetstream Research Experience for Undergraduates (REU).
The Jetstream REU invites students to work with Indiana University (IU) researchers and staff on mentored research projects that involve the Jetstream platform. Led by the IU Pervasive Technology Institute (PTI), Jetstream is a national science and engineering cloud that provides on-demand high-performance computing (HPC) and data analysis resources for research and education.
Underwood, a senior at Livingstone College, and his cohort in the Jetstream REU are bringing machine learning and HPC to the problem of identifying frog migration patterns.
“The entire scope of the project is to make artificial intelligence useful in fields that make up the long-tail of science,” said Underwood.
However, before Underwood and his colleagues could start gathering these calls, they had to know what they were listening for.
From students to frog surveyors
The Jetstream REU team wanted to identify amphibians specifically from the Indiana region – Spring Peepers, Chorus Frogs, Green Frogs, and American Toads. They gathered sample calls from Cornell’s Macaulay Library archive of wildlife sounds - 85% of the calls were used as training data and the remaining 15% served as testing data.
Underwood and the team then created three neural networks to process their data: one audio-based recurrent neural network (RNN), and two convolutional neural networks (CNN)—one audio-based and one image-based. The latter analyzed spectrograms created from the audio files.
The result? Computer frog surveyors are much more accurate than their carbon-based counterparts. The CNN image-based model came out on top as being the most accurate and most efficient. It also took up less space and took less time to process.
“Human national surveyors accurately predict 80 percent of the frog calls species. The image-based CNN produced over 97 percent accuracy, with the four frog species,” said Underwood. “The CNN audio and the RNN audio both produced nearly 90 percent accuracy. All three are what you could call superhuman.”
Processing the over 5,000 audio files they had recorded would have taken up to twenty hours on a laptop computer. Thanks to Jetstream, it took only minutes.
“We couldn't have accomplished anything without Jetstream,” said Underwood. “It made the process more efficient and effective, and allowed us to be able to do a lot more.”
He hopes to improve the accuracy of the team’s findings, so ecologists can even more accurately identify frog calls. The solution? Add more data.
“I believe we had 1,500 samples per dataset, and the last one we had just 1,000.” Underwood said. “Those extra 500 segments would actually benefit the CNN and RNN audio because they're only taking this small proportion of data. If you give them more data, they would predict a more accurate rate.”
The man behind the data
Most college students don’t get the opportunity to work with high-performance computing tools until they reach graduate school. So how did Underwood land a chance to work with Jetstream while still an undergraduate?
Growing up in North Carolina, Underwood always had an affinity for math and computers. But it wasn’t until a lucky meeting with a fraternity brother from Indiana University at the Emerging Researchers National Conference in STEM, that he applied for the Jetstream REU program.
Underwood also acknowledges the strong influence of his alma mater, Livingstone College—'the black Harvard of the South.’
“[Livingstone] helped me grow and come out of my comfort zone to be able to go to an institution like Indiana University and be successful with the frog call project,” says Underwood. “You don't see many African-Americans in my field at all. So it's everything to do with being at an HBCU, and the fact that you are seeing people every day who empower you to be the best you that you can be.”
Next year, Underwood will study at Kentucky State University for a graduate degree in computer science with a specialization in cybersecurity. He hopes to open his own IT consulting firm and wants to do his part to make the tech industry more inclusive.
“I see myself bringing more cultures and more people into this field,” says Underwood. “I believe there needs to be more women, more cultures, and there definitely needs to be an increase in African-Americans in this field.”