Saving every tweet: Meet researchers using bird calls, AI to study biodiversity
Exploring the Anamalai Hills of Tamil Nadu, Vijay Ramesh, 30, would halt in his tracks every time he heard a white-bellied blue flycatcher. The bird remained elusive. He rarely saw it.
This is a species endemic to the Western Ghats, and he assumed there just weren’t that many. It was only after he set up audio recorders at some of the same sites that he realised the bird wasn’t nearly as rare as he had thought.
“In a tropical forest, you’re often just hearing a bird instead of seeing it. The white-bellied blue flycatcher is very shy, and the recorders showed it to be more common than we thought. That’s what really stands out when you study these birds from an acoustic point of view,” says Ramesh.
Bioacoustics, a field devoted to the study of animal sounds, is the focus of PhD candidates Ramesh, Sarika Khanwilkar, 29, and Pooja Choksi, 31, and their research collaboration, Project Dhvani. Since 2018, the initiative launched by the three students (they’re pursuing doctorates in ecology, evolution and environmental biology at Columbia University) has been collecting soundscape recordings from sites across central India and the Western Ghats. This includes sounds from birds, bats, insects and, of course, humans.
In addition to understanding the patterns of presence or absence of a species, the project uses artificial intelligence and machine learning to analyse the data collected and study how various species interact with their environment and react to changes in their habitat.
Passive acoustic monitoring is a major component of Project Dhvani’s ground research, usually conducted over the summer and winter, when they can capture vocalisations from a large number of resident and migratory bird species. Project Dhvani’s team of ecologists set up audio recorders at selected spots. These recorders continually capture the different frequencies of this ecosystem (usually 0 to 24,000Hz), some of which are inaudible to humans.
“There are many traditional methods that ecologists use to understand how species respond to changes in their habitat. But the auditory dimension is one that hasn’t been explored much,” Choksi says.
Bioacoustics, as a broader discipline, can be traced back to the early 1970s, when acoustic communication between animals first began to be studied. The term “soundscape ecology” and the idea that one could study sounds from an entire landscape is more recent, with studies dating back to 2011.
In India, the Indian Institutes of Science Education and Research as well as the Indian Institute of Science, among others, are using sound to study ecology. Long-term acoustic research is underway in Australia, Costa Rica, Peru, the UK and US. And bioacoustics has been a game-changer in monitoring hunting and poaching activity in real time in countries across Africa.
The soundscapes Project Dhvani studies range from protected areas to unprotected human-dominated landscapes. “Sites around towns and villages hold a lot of biodiversity and it’s important to study these lesser-explored landscapes,” Choksi says.
Once recordings have been collected, the effort moves to a computer lab, where raw data is first manually sifted and bird calls are manually annotated by naturalists and ecologists.
Recordings are then fed into a deep-learning model, where calls are logged by frequency, intensity and other such metrics. Over time, as more recordings are made at each site, the data begins to offer a picture of the biodiversity, and ways in which it may be changing.
“It’s a low-cost way to monitor a large landscape continuously or over long periods of time, and to understand the health of an ecosystem,” Ramesh says.
Under the radar
The challenges for such a project are unusual. Some birds, for instance, consistently drown out others. As many as 20 species may be calling in a span of 10 seconds, and only a few dominant ones will be apparent (like the white-cheeked barbet), while others (like the crimson-backed sunbird) are lost, says Keshav Bhandari, a research affiliate working on the project.
Species of the same kind may call at slightly different frequencies in different regions, so when logging the data, one could at first miss them altogether.
Background noise (leaves, wind, rain, traffic, footfalls) can make some bird calls harder to identify. Added to which, some birds mimic others, and some mimic sounds they’ve heard that aren’t bird calls at all, like horns and sirens.
“We are interested in identifying all birds that are heard in the recording, especially the rare ones. These factors hinder the ability of the model to capture some of the faint patterns that may be present in the spectrogram,” Bhandari says.
In areas around the Kanha National Park in Madhya Pradesh, for instance, Choksi says, five birds accounted for approximately 45% of the calls on a particular set of recordings. “What we are trying to do is to create an algorithm that can pick out the birds in the background that are completely sidelined,” Choksi says.
To listen in yourself, go to projectdhvani.weebly.com.