The title of Rachel Carson’s 1962 book “Silent Spring,” the seminal account of the environmental toll of the pesticide DDT, hinged on what the sounds of nature (or lack of them) can tell us about an ecosystem’s health.
Scientists are now using this insight to gauge the constellation of species living in a tropical forest and nearby farmland, a key benchmark of how biodiversity is faring as jungles fall to the plow or grow back into forests.
The combination of small, durable microphones and cutting-edge artificial intelligence holds out the possibility that the squawks, peeps and groans that fill a jungle might one day enable conservationists to track the state of vast expanses of land with minimal effort.
“I think the time of field ornithologists to map vocalizing birds in the field is over. I say this with regret because I am one of these field ornithologists myself,” Jörg Müller, an ecologist at Germany’s University of Würzburg involved in the new research, said in an email. “The next few years I expect a boost of retrained AI models by many organizations and boost in using sound recorders globally.”
The work by Müller and a team of German and Ecuadorian scientists is just part of a much broader effort to automate ecosystem monitoring, as scientists struggle to track the rapid changes sweeping the planet. Some researchers are studying whether air quality stations could double as ecosystem monitors by sucking up bits of DNA shed by nearby species. Some are looking at the potential for monitoring the health of coral reefs by wiring them for sound. Others are tapping images posted to social media to track species such as zebras and whale sharks.
Müller’s group wanted to see if audio recordings of jungle sounds could serve as a relatively simple barometer of the state of species in Ecuador’s Choco, a cloudforest that hugs the western flanks of the Andes. In 2021 he and collaborators installed audio recorders at 43 sites in the forest. The condition of the land varied, including untouched old-growth; livestock pastures and cacao plantations; and abandoned farms reverting to forests. The devices were programmed to record 2-minute snippets of sound every 15 minutes throughout the day and night for two weeks.
The scientists then used a variety of tools to sift through the nearly 2,000 hours of audio recordings to tease out who was living in these different places. As a gold standard, human experts listened to audio samples from prime times for animal noises—dawn and dusk for birds and mammals such as howler monkeys, and the middle of the night for amphibians such as frogs. They counted each species heard in each recording.
Then the computers got their turn. In one case, the recordings were scanned using a program that measures overall characteristics of the soundscape, rather than looking for individual animals. For instance, it can gauge how broad and dense the range of sound frequencies are, or the number of different sounds made every second.
The researchers also tried out some of the latest tools in the field of artificial intelligence. Known as convolutional neural networks, or CNN’s, these programs are built to learn how to distinguish between different visual images and, over time, correctly identify pictures of, say, a cat or a dog. In the case of sounds, noises are converted into visual representations called spectrograms. After being trained to identify tropical bird species using data sets from another forest, the CNN was turned loose on the Choco forest data.
The computers—particularly the CNN—proved adept at quickly finding meaningful patterns. The AI system had a 69% match with the time-consuming human experts, even though the computer was trained only for bird sounds—a subset of all the forest noises. When the expert selections were limited to just birds heard by the experts and that the CNN was trained to identify, it was an 85% match, Müller and colleagues reported Tuesday in Nature Communications.
Though less accurate, a combination of different soundscape indicators got as high as a 62% match for the analysis by people.
While not every creature in the forest makes a racket, the researchers found evidence that the audio recordings can act as surrogates even for the silent ones. To check, they set up traps that use lights to lure nighttime insects, many of which don’t make sounds that standard audio recorders would detect. The biodiversity of these quiet bugs mirrored the biodiversity patterns detected by the computers.
The results hold out the promise that automated systems “would allow conservation managers to assess forest recovery cost-effectively and to better quantify the conservation value of their protected areas,” the scientists wrote.
Of course, there’s a lot of work to be done before such an approach could be deployed on a large scale around the world. Besides the need for a lot of hardware, the power and accuracy of the CNN’s hinges on how much data is available to train the program. In this traning data, different sounds have already been identified, enabling the computer program to check how accurate it is and then tweak future analyses, growing more accurate over time.
The authors plead with their scientific brethren to help gather more recordings, particularly of species besides birds. “We urge the conservation community to prioritize the creation of global sound repositories for taxa beyond birds, based on which machine learning models can be rapidly improved and extended,” they write.
Get ready to see a lot more microphones in the woods.
Müller, et. al. “Soundscapes and deep learning enable tracking biodiversity recovery in tropical forests.” Nature Communications. Oct. 17, 2023.
Image: ©Anthropocene Magazine