Scientist Develops Method For Monitoring Ecosystem Health Through Sound : Short Wave On a rapidly changing planet, there are many ways to measure the health of an ecosystem. Can sound be one of them? We dive into a new methodology that applies machine learning technology to audio soundscape recordings. Lead researcher Sarab Sethi explains how this method could be used to potentially predict ecosystem health around the world.

What Does A Healthy Rainforest Sound Like?

  • Download
  • <iframe src="https://www.npr.org/player/embed/905931181/906061379" width="100%" height="290" frameborder="0" scrolling="no" title="NPR embedded audio player">
  • Transcript

MADDIE SOFIA, HOST:

You're listening to SHORT WAVE from NPR.

EMILY KWONG, BYLINE: A couple of years ago, scientist Sarab Sethi found himself in a tropical rainforest in Borneo, an island in Southeast Asia. He was part of a team installing solar panels on the tops of trees above the tree canopy, a good place for those panels to soak up sunlight and power some very special devices - more on those in a moment.

How high are we talking?

SARAB SETHI: The highest I've been is up to 40 or 50 meters.

KWONG: Oh, my God (laughter).

SETHI: Yeah (laughter).

KWONG: Forty, 50 - I want to look that up - 164 feet?

SETHI: Yeah. Yeah. It's pretty hairy if you haven't got a head for heights (laughter).

KWONG: Thankfully, he does. But here's the thing you should know about Sarab - he's a postdoctoral researcher at Imperial College, London, and describes himself as, quote, "not a camper type of person." Having grown up in a city environment his whole life, he's more accustomed to honking cars and construction drills than the sounds of a tropical rainforest. So imagine Sarab settling in for his first evening at a remote campsite in Borneo when the camp's generator is finally turned off for the night.

SETHI: And then, suddenly, the sort of cacophony of the tropical forest at night hits you.

(SOUNDBITE OF AMBIENT RAINFORST NOISE)

SETHI: It's quite incredible, right? There's just so many insects, so many frogs. I think that was the first time it really hit me just how sort of loud these areas are and how much information there is really in the signals.

KWONG: And what do you remember thinking when you were hearing that?

SETHI: Honestly, I think I remember thinking I've - I'm in a bit too deep. And (laughter) I was in the middle of the forest, and it was all very loud. And it was all, like - you know, there's no electricity, and I was like, my God, is this a bad joke gone too far at this point? But (laughter)...

KWONG: Quite the opposite - see, all that forest sound is why Sarab was there. Those solar-powered devices from earlier are audio recorders that Sarab helped design. They're placed around the forest, continuously recording the sounds and then automatically transmitting that data. Sarab and his colleagues have potentially developed a new way to study ecosystem health using sound and AI. So today on the show, ecoacoustics - what can we learn about the health of a forest if we set up a recorder and listen? I'm Emily Kwong, and this is SHORT WAVE, the daily science podcast from NPR.

(SOUNDBITE OF MUSIC)

KWONG: Today, we're speaking with Sarab Sethi about ecosystem health monitoring using sound. But before we dig into it, let's first look at one traditional method for evaluating the health of an ecosystem. Say you're interested in measuring bird biodiversity, for instance. You might use the point count method, where you stand outside for hours on end with a lot of patience and a talented pair of ears.

SETHI: Every single bird you hear vocalizing or you see visually, you note it down, what species that was and at what time you saw it. You kind of repeat that thing over the 24 hours of the day at different hours, at different locations.

KWONG: It's a super thorough process for monitoring ecosystem health but incredibly tedious. So Sarab and his colleagues thought, you know, with all this modern technology we have - sensors, wireless networks, solar panels - there has to be a more efficient way to do this.

SETHI: Can we get something that's sort of approximately as good as this kind of data but with completely automated methods where your recorder is uploading audio to the Internet straight from the field?

KWONG: Allowing them to potentially track ecosystem health in real time. They've set up this acoustic monitoring network in Borneo, part of the SAFE Project, which records audio continuously, and it is a staggering amount of data.

SETHI: I think we've got about 17,000 hours so far from the network.

KWONG: Seventeen thousand? Sorry.

SETHI: Seventeen - one, seven, zero, zero, zero (laughter).

KWONG: What? Oh, God.

But it's not just background noise. Housed in those 17,000 hours is a treasure trove of information, impossible for us mere humans to listen through. But fortunately, the folks at Google have figured out a way to sort through all that audio. Sarab and his team turned to Google's AudioSet, a massive dataset of sounds that was developed using machine learning.

SETHI: While AudioSet has done as it has labeled data for kind of almost every type of sound that you can imagine there being. And so from that, it kind of knows that amongst all of dog barks, there is something that is consistent about all of dog barks that makes it a dog bark. And so it knows that this is one fingerprint. And then amongst all of glass smashing, you know, it knows that there's one consistent thing. So it's finding things that kind of we, as humans, perceptually see as consistent in types of sound and then fixing them down to one type of fingerprint.

KWONG: Taking Google's technology, they applied it to their forest recordings, training their machine to create an audio fingerprint, a way to kind of identify that forest through its sound. And the algorithm they've developed can potentially predict important indicators of a forest's health, like habitat quality and biodiversity, based on its soundscape alone. And it didn't just work in one particular kind of forest. Sarab and his co-authors analyzed the audio recordings of forests around the world. They published their findings this summer in Proceedings of the National Academy of Sciences.

What did you and your team show with these audio recordings, beyond the fact that, yes, the technology worked? What did it reveal about the character and what's happening in these forests?

SETHI: What you see quite nicely fall out from all of the sites we looked at - really clear, nice diurnal patterns. So that's sort of how day and night are different and how audio consistently follows this same kind of trajectory of fingerprints, and you can start to see where seasons change and where the day sort of - how it evolves through the day as the sun comes up and goes down and how that changes the species, communities that are vocalizing.

KWONG: That's lovely because we think about days and years and months mostly in relation to light.

SETHI: Yeah.

KWONG: Like the sun coming up, the sun coming down.

SETHI: Yeah.

KWONG: But you're saying there's, like, a rise and fall of sounds...

SETHI: Exactly.

KWONG: ...Throughout the 24-hour day that you measure.

SETHI: Yeah. Exactly - and to the point where you can - and we did this analysis within our papers - that you can just take a random piece of audio, and you can guess with pretty good accuracy what hour the audio was recorded at. Again, you know, it's questionable what's the point in that. I know what time I recorded my audio. But it kind of just shows you the amount of information that's, like, temporally encoded in this audio as well, and you can guess what month it was recorded from.

KWONG: So Sarab, we're going to actually listen to some of the sounds that your team has recorded from the SAFE acoustics website, acoustics.safeproject.net. So these recordings, they're all - are they all uploaded, like, wirelessly from these recorders?

SETHI: Yes. So actually...

KWONG: Wow.

SETHI: Due to COVID, they're not live right now, but they would normally be sort of recorded in real time and uploaded so you'd be able to listen to what the forest sounds like in all these different locations right now.

(SOUNDBITE OF RAINFALL)

KWONG: So this is very mood setting. This is the rain at night in an old growth forest in Borneo.

SETHI: Yeah. It's like we're there.

KWONG: So did you spend some nights under a tent in these conditions?

SETHI: Yeah. I mean, this kind of rain is like a godsend because most of the time I spend doing fieldwork is sweating. So when the rain comes in, it's nice and windy and cool, and, yeah, it's like music to my ears.

(SOUNDBITE OF RAINFALL)

KWONG: Here's 11 o'clock in a cleared forest, one that's been cleared of trees.

(SOUNDBITE OF AMBIENT FOREST NOISE)

SETHI: Yeah. It's pretty dead. It's kind of like two axes how the audio changes around the soundscape. There's the temporal patterns from nighttime to daytime through to nighttime again, and then there's the sort of forest gradient of old-growth forest to log forest to cleared forest.

KWONG: Let's see if I can find an old-growth that is loud by contrast.

SETHI: An old-growth loud - you want to go, like, 5 a.m., 6 a.m.? Here we are.

(SOUNDBITE OF AMBIENT FOREST NOISE)

SETHI: So this is a dawn chorus, and you're just hearing loads and loads and loads of birds all calling at once as they all wake up.

KWONG: It's so rich.

SETHI: Yeah. And it's, like, it's very loud. These birds are completely surrounding you, and they're up in the canopy. And you can hear big branches moving as bigger animals move around the forest, either through the floor or the - you know, monkeys move through the canopy.

KWONG: Yeah. There's just this sense you're in a place that humans just haven't touched.

SETHI: Yeah. Once you've got these fingerprints - and I just look at the fingerprints, and I have no - none of this biodiversity data that I was talking about before - one thing you can still do is sort of look at what are the anomalous sounds you got there? Like, what are the unusual sounds, and what are the usual sounds? So you say what sounds appeared today that we just completely weren't expecting? And we tested how you could actually use that to detect illegal logging or poaching, which is, like, a particularly big issue in protected areas.

KWONG: What are the sounds of illegal logging and poaching?

SETHI: Well, gunshots and chainsaws, really. They're the main ones, but, you know, it's the cars sort of being driven into the forest. It's the humans talking. It's the humans using machetes. You know, if you go into the middle of a tropical forest, you wouldn't expect to hear humans. It's an unexpected sound, so these come out as anomalous events.

KWONG: So is this one of the practical applications of this tool - the ability to pick out anomalous sounds and know what human activity is happening there?

SETHI: Absolutely. Yeah. But, you know, it's one thing to say this in theory and another to test it. So we were like - you know, we took speakers out into the middle of the forest, and we played sounds of gunshots, sounds of chainsaws, sounds of people talking and ran it through our algorithm and how we thought we'd automatically pick it out, and it did work.

KWONG: Got it. Finally, I want to ask about the future future, about the possible predictive power of this tool. You know, when I think about it, it reminds me of a stethoscope, almost. Like, you lean in, and you listen to the pulse of an ecosystem. When you do that on a human heart, sometimes, you know, you can hear, like, an indication that, like, a murmur that a patient might have a heart disease, right?

SETHI: Yeah.

KWONG: And this tool - could this tool be used to listen to ecosystems like that, to detect a moment when an ecosystem has a problem and maybe catch it before it gets sick?

SETHI: Absolutely. I think that's a really lovely analogy. So I mean, I think we're all quite comfortable at this point - we've heard so much about sort of tipping points with climate change, for example, or, you know, the idea that we might get to a point where there's an irreversible change in the natural world, and we can't - we get mass extinctions and this idea. But it's - if we can, you know, listen with this very high-quality stethoscope, as you put it, to the ecosystem, can we start to hear, you know, the murmur before the heart just gives out and stops? Or can we hear sort of - can we start to predict these collapses before they happen and then, you know, actually take active management interventions to try and stop them happening? We can, you know, even out the effort and not just kind of overexploit it to the point - you know, more sensibly direct your efforts.

KWONG: What gets you excited about it when you think about how it could be used in the world of ecosystem health?

SETHI: So one part that really excites me is it's a link back to theory. You know, we can actually start to explode the amount of data we're collecting. We can then link that back to mechanistic models of how our ecosystems might actually be responding to climate change or responding to land use change or agriculture, and we can start to say, you know, we're understanding more deeply how sort of humans are impacting the ecosystems, and then, hopefully, that kind of understanding could feed into better targeted management practices or better targeted sort of policy decisions.

KWONG: This episode was produced by Abby Wendle, edited by Viet Le and fact-checked by me, Emily Kwong. You're listening to SHORT WAVE, the daily science podcast from NPR.

(SOUNDBITE OF MUSIC)

Copyright © 2020 NPR. All rights reserved. Visit our website terms of use and permissions pages at www.npr.org for further information.

NPR transcripts are created on a rush deadline by an NPR contractor. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.