Species to Soundscapes: Build Ecoacoustic Biodiversity Monitoring Programs

Welcome and thanks for joining us today for a webinar on practical eco acoustic methods for monitoring biodiversity. My colleague Andy White will introduce our invited guests. >> Lynn Schwaskkov um distinguished professor and memorous college of science and engineering at JCU. So uh we've also got um Dr. Sharon Brody and Dr. Sebastian Hoer um speaking today. So I will hand over to Lynn to start the webinar. >> Hi everybody. Thanks very much for the opportunity to present this um webinar and thanks to the ARDC. Um what we will be talking about today is building eco acoustic monitoring programs and we're going to focus on um perhaps unusual parts of um of these kind of programs, unusual bits that people don't think about. I hope so. Anyway, so I I imagine everyone here knows that passive acoustic monitoring is where you just put automated sound recorders out in the bush and you capture the entire acoustic landscape of an ecosystem, just everything that happens. Um, and so that includes the geophony, which is the wind and rain and earthquakes and lightning and things like that. Anthrophony, which includes planes, trains, and automobiles. people talking, things like that. And the biophany, which is all the animal noises, birds, frogs, mammals, insects, everything that makes noise. Um, and of course, we know that biophany can be analyzed to measure biodiversity. And that's what we're going to be talking about today. So, what we're trying to achieve is something like this. We've got um I don't know, can people see? Maybe I'll get rid of him. Um, so what we're trying to achieve is this. We got biological reality or something like biological reality out there in the world and we want to make management decisions from the information we get from ego acoustics. So that we should all agree on that. Um but there should be a lot of steps in between biological reality and making a management decision. And today what I want to be talking about is these steps in here or what we are all going to be talking about is that I think what has happened in the eco acoustics community is that there's been an intense focus on the orange step here uh which is why I made this big circle orange. Um there's been an intense focus on recognizer or classifier building and the outputs from those and how good they are and people have become really really heavily focused on that and I think that in order to build e uh eco acoustic monitoring we need to think about these other steps a bit more and so that's what we're going to be focusing on today and I think the reason why people have focused so heavily on recognizers is because there's a general perception that vocalization detection algorithms rely on vast amounts of training data. And so therefore, training a model is really hard and difficult to do and difficult to achieve. You can only do one at a time and it takes ages and all that kind of stuff. And so I think that now we've got a few new methods that allow us to approach this. And I'm not going to this isn't going to be a methods talk, but you may have heard about fshot transfer learning or something like that. There's a paper on it here by Jacuzzi and Olden. If you're interested, that's where I got this sexy graphic. Um, and so what we have is we have models now which are kind of like large language models but for birds and they include bird net and perch. And they include an awful lot of species of birds now. So you can actually just often search for a call within these or get a call and get that call identified right away. So we can already do that for a lot of things. It's not perfect, but it does work reasonably well. It especially works well in Europe and North America. There are species missing from these models. So there are species of birds that are not present, especially in Australia. Um, but you can also use these models to detect novel signals. So you can get wind and rain, insects, amphibians, mammals, other stuff. And in my lab, we've been using large CNN's, these bird net and perch for detecting reptiles, mammals including deer and other things, frogs and insects. I've put whales here because other people have used them for whales. So the point is is that you can put a sound into these models and quickly create a recognizer for a model that is for a sound that is not already in the model. And it works like this. You take your sample call and you only need one individual sample call and it's usually not a bird anymore. Although it could be a bird that isn't in there. Could be an insect, could be a mammal, could be an amphibian. You put it in there into the birdet algorithm or the purge algorithm and it breaks it. You've got a 3se secondond sample call. It breaks it up into a feature set of 1,024 numbers that describes the sound that you've put in there. That's an example of a frog sound if you heard it. I'm not sure if you could hear it. Then you take all the rest of your audio and you chop it up into 3se secondond segments. You put it through birdn net or perch and it creates feature sets of 1,024 numbers or up to 1,024 numbers for all the rest of your sound samples. And then there's the frog again. And then you situate your sample call. So if you imagine all your audio looks like this, your sample call is here. Okay? So all your audio is described in a 1,024dimensional space and your call is here. And then calls that are close to it in that space tend to be that species or something like that species which allows you to build a model that looks like this. The distances here are ukitian distances. >> Okay? And so you can line up all the sounds um that you get on in your by distance and you can I guess I can leave that there. You can validate them and you can see that sounds that are close in uklidian distances between 16 and 17 in this case are yes they are your species and there's ones that no are not your species. Okay. And that allows you to create a um threshold distance and you can start thinking that everything less than that distance is your species. So you can build a a recognizer very quickly using a a logistic regression GLM and just do it fast. So we can do it fast. So you don't need to have millions of tons of of um data. You don't need to have millions of of sample calls. In a way, it's a way of finding a lot of sample calls in your data and allowing you to build a recognizer quickly. So, when we are doing monitoring, now what we need to think about is all these other steps. Okay. So, um you may be monitoring a rare threatened species. You may be monitoring a rare invasive species. Now, this one's especially rare. No one's ever heard of the sevenlegged gecko before. Um this is thanks to Chat GTP. But anyway, all invasive species are rare when they're first arriving. And so you may want to detect them um when they're first arriving. And so you may want to detect an invasive species when it's rare. You may want to detect common invasive species because you're trying to control them and you want to see if you're having any effect in their decline or you want to know which times of year are they most active? All kinds of things like that. Or you want to maybe detect the whole vocalizing community. And I think we can do that now using the technique I just told you about. So I think one thing we should think about in Australia is monitoring common native species because we can do that easily. Everyone who's got eco acoustic recordings has got lots of examples of native species and their activity on there and it would be worth thinking about noticing those species and seeing if they're declining. Excuse me. >> [clears throat] >> So once we know what we're doing, what we're trying to measure, we then have to decide what do we need to know and we need to decide whether we only want presence absence. Now people have focused in general on the fact that presence absence is what we need for them for uh or all we get from acoustic monitoring. But that's not really true anymore [snorts] because now we can record for long periods of time. we can get vocal activity and vocal activity is often positively related to the numbers of animals. And so you may be able to get population size or an index of population size. You need to decide when you're going to record because you want to record when they're most active if you're looking at activity as a gauge of when of how many there are. Um and you may want to measure overall diver biodiversity. So you need to make a decision about what you're measuring or which things of those you're measuring. Then you need to go out and you need to sample. I'm I'm adding a comment here in that make sure you record metadata, especially where and when the recorders were placed because many recorders don't do a very good job uh recording that they've been moved. So, u make sure you record where and when you put them there. But think hard about your sites, okay? Include controls. Don't only put recorders where you think you'll hear animals. include places where you don't think you're going to hear them just in case you do hear them or if you're um c if you're creating an intervention. If you're controlling something, you want to know what's happening in the places you're not controlling them. Are you actually having an impact or are they just declining because of seasonality or some other thing, less food around? So, you need to think about where you're going to put your recorders with respect to what you're trying to achieve with management. Don't assume your animal only vocalizes at certain times. A lot of birds that are supposed to be dal vocalize at night. A lot of frogs that are supposed to be nocturnal vocalize during the day. We can you should really consider recording 24/7. And you should think about schedules. Schedules may not be necessary at all because we can store a lot more stuff than we used to be able to. Uh but if you are going to um sample, you may want to sample for instance 20 minutes an hour every second or third day rather than just during the time that you think your animal's going to be active. There's a bunch of advant advantages to that. And I'm introducing Dr. Sebastian Hoer who you saw at the beginning. He's a post-doal fellow in my lab. He's working with the Department of Environment, Tourism, Science, and Innovation in Queensland with indigenous ranger groups to help de develop monitoring programs that um do what the rangers require. So, he's helping them develop questions and designing sampling. And that's what he's going to be talking about today. Okay. Recognizer outputs. Those are the orange ones that I told you were not that important anymore or not not um not deserving of laser focus is what I would say. So classifiers and recognizers are of course important. I'm not trying to say they're not. Um but we need to think now about balancing precision and recall because for the most part for biological recognition we need to prioritize precision. So you need to think about how much precision and recall you need. For chorusing species that call a lot. You don't need precise recall. You don't need very high recall because you don't need every single call. you're going to hit a lot of the calls in a in a chorus and so you don't need every single call. And for presence absence, which some of us are still trying to do, then um you also don't need very high recall. You just need very precise um um well, you need precision. Okay. So, um then we need we have our our um our metrics that we're going to create. And so, we use detections for occupancy estimates over here. So that's where we need high precision. That's presence absence at a site level. And now in my lab, we're working more with activity. So what's the relationship between activity and population size? So if you have um surveys that tell you population size, you can compare them with activity and you can see if there's a relationship between activity and population size, that might give you an index of of um of numbers of individuals. And we can also get community composition from species lists. Of course, that's easier with just hits or precision. And that's very powerful because you can record for a long time and so you can get everybody in the community uh over the course of a year or a couple of years. And Sharon's going to be talking to you today. Um she's a post-docctoral fellow in my lab as well. She's been working with NESP and also Parks Australia building a monitoring program for the woodland birds on Christmas Island. and she'll talk to you about sampling and occupancy and activity and how we're using those to build a monitoring program. So the last thing we do of course once we've got our measures is we can start making ecological inferences. We can figure out whether occupied sites are increasing or decreasing. Are there fewer occupied sites because that's bad. Uh are there more occupied sites? Maybe if you're looking at an invasive species it's um it's it's bad too. Is activity increasing or decreasing? And what does it mean? We have to know the difference between seasonal activity and not seasonal activity. Um and when like yearly is there differences in how many how much activity there is. Are numbers of individuals or our abundance proxy increasing or decreasing? These are important things to know. And our interventions working and maybe we can start looking as I said at the abundances of um of common species. These are not common species. I don't even know what they are. But anyway, we can look at the abundances of these of different species that we haven't been thinking about at all because we have this data now. And the idea is that we can use information to make better management decisions, which is what we all want, I think, or what a lot of us want. Excuse me. The key takeaways from my talk are don't obsess about having loads and loads of training data. Good models can now be built fast for lots of things using only a few sample calls. Even just one works. And don't obsess about recall because it can be low for lots of the monitoring tasks that we need to do most of the time. But do think about what is it exactly that you want to know and how are you going to best place your recorders, record, do everything else to find that out. how best you're going to sample to answer your question to really answer your question. What inferences can you safely make from your data? Can you actually infer something about activity? Can you infer something about numbers of individuals? Things like that. And how will you use it to manage um how will you best use it to manage? I think we're moving on to Seb now. >> So yeah, thanks Lynn for that. Thank you all for tuning in and coming to this. So, as Lynn alluded to, monitoring is very different things to different people. And I wanted to start by giving more of a general overview of so what considerations are important when designing an eco acoustic monitoring program and then dive a little bit deeper into an acoustic monitoring project we established in collaboration with um First Nations Ranger groups across Queensland. So, I'll start with the more broad um considerations. So the most important thing when you're trying to uh come up with a with a design is just figuring out what the goal is of the whole thing because the survey design is entirely dependent on identifying a really clear survey or monitoring goal. So are you targeting single target species or is the whole community so biodiversity is what you're after? Are you looking at aquatic, terrestrial or boreal species or all of them? Are you interested in species richness, vocal activity, occupancy, habitat use? Um, also, is it a long-term monitoring project or is it more like a short-term biob blitz that maybe gets repeated every once in a while? Um, do you have any prior knowledge of the targets sounds that you're really interested in? Um, for example, do you know anything about the activity patterns or vocal activity? You know, what is it like during the day? What is it like over the whole year? Um, do you know anything about habitat preference or home range? Do your species or your target um sounds do they move around a lot or are they more stationary at a particular location? Um, also what kind of recording equipment do you have access to? Are you able to get sort of more long-term audio recorders or is it more of the smaller, cheaper um short-term battery powered audio recorders? Yeah. Similarly, you know, what kind of microphones do you have access to? you know, how many SD cards are you able to get and what is the storage capacity of those? And yeah, again, with the audio recorders, are you able to get solar powered devices or is it entirely dependent on battery? So, all of these are sort of yeah, come some of those really important considerations to begin with to even start thinking about designing it cuz when you survey or when you're thinking about the survey design, um the target sounds, they're crucial because they will determine all these other considerations for the survey design. um you know how many recorders do you actually need is is obviously often the starting point and then where are they placed? So the spatial considerations where you putting them um how many do you have to get where how do you replicate the placement of those recorders? Uh you got to think about the detection distance of each of those recorders and what might influence the detection distance because of the you know the habitat for example. Um again knowledge on on the habitat using the home range of those species. If they're stationary at a particular habitat um uh structure then maybe you need to only record there. Um and then there are temporal considerations. So how long can you actually deploy those audio recorders for or do you want to um what's the recording schedule like? Are you going to do continuous or you going to duty cycle your your recordings? What about the maintenance protocol? are you going to be able to maintain those audio recorders regularly and how often can you do that and do you have any help doing that? Um the sampling rate is really important obviously um to try to figure out what yeah what frequency you're going to be recording in. Uh and then there are considerations on the data storage you know how much data can you actually store and you have access to long-term data storage um how are you going to sort of wrangle or or um access that data and how you going to analyze the data. And to begin with, it could be really useful to start with a pilot study to understand your survey design. Um, where you also maybe get some information about the species presence in your audio recordings. And you might get some information about noise pollution. Maybe some anthropony with some, you know, human-made sounds that are really disturbing uh the recordings in the particular areas. You might have to actually move it or you might get some information about the habitat features that might impact your detection distance. you're not really getting any recordings or any sounds from the the place you're actually interested in because the the habitat features are are um are not yeah are sort of impacting that. Um so for the spatial consideration or for the recorder placement there are a couple of really sort of common ways of sampling. There's sort of your systematic sampling where you're sampling at sort of fixed intervals across a whole area. Uh it's probably the easiest design because it's you know those fixed distances you know the space between them so it can be easier to navigate that in the field and deploy them but you may require more samples and they may be less representative. As you can see in that little graphic there that yellow habitat doesn't actually get sampled because of those fixed distances. So you might miss out on some some um habitat. And then there's stratified random sampling. So that's when you are selecting the group. So in this case those four different habitat types and then you randomly sample within them. Um that sort of ensures representation of all of those relevant groups but it requires prior knowledge and you may not represent the groups very well because if it's random you might cluster some of those points and as you can see there some of those points are overlapping. And then there's uh stratified systematic sampling. So again, you're kind of designating particular groups beforehand um and you're putting a number of fixed samples within and you you are representing the groups and with a more even coverage that way. But again, it requires prior knowledge and maybe even coverage is not possible depending on what kind of um landscape it is and what what those habitat features look like. And then there's purpose of sampling. So that's where you're sort of selecting um the sampling based on assumptions. So you are um really highly efficient for a specific purpose. So let's say you wanted to sample creek lines. So you're just selecting the creek lines and you're sampling right along them. Um which obviously requires certain assumptions and prior knowledge and it introduces a bias because you're only sampling that particular habitat type or that particular um group. And then for the temporal considerations um particularly schedules for the deployment length uh that is really important to figure out because obviously will determine the type of recorder because you know some recorders are um going to be able to be deployed for longer periods of times and some of them will be more shortterm. So figuring out how long you can leave those audio recorders out there whether that's many many years or just a couple of weeks at a time um is really important to figure out what you can actually get. Uh yeah, as I said, maintenance protocols. Is it actually regularly accessible? Are you going to be able to go there and switch out batteries and SD cards and check on them for damages and and whether they're still functional? If you can't, then maybe a long more long-term um option with the solar panel may be maybe more suited in that case. Recording schedules, whether you're doing continuous duty cycles and again a pilot study could be super useful to understand that sort of temporal considerations as well. um you know identifying potentially seasonal calling windows uh looking at DL patterns of activity uh and then sort of weather dependent activity uh more specifically for the temporal or for the recording schedules uh the continuous recordings obviously you're recording all the time and therefore you have a higher chance to detect cryptic or rare species uh and you're getting a high resolution data for seasonal and deal activity patterns which he can then potentially link to environmental variables to get some um answer some interesting questions. But obviously it produces a large amount of data and there are higher costs for those recorders and for the maintenance cuz you have to swap that out all the time. Those SD cards for example. Now GD cycle is where you record a certain number of minutes per day and then a certain number of days per year. And that way you're lowering your chance of detecting cryptic or rare species and you're potentially getting a low resolution data for those um environmental links, but you're producing more manageable amounts of data and you might have a lower cost for recorders and maintenance and also longer battery life uh which enables also cheaper recorders. Now for the GT cycles aspect, um the best way you can do this is obviously as long as possible, but more importantly as frequently as possible. So doing half the time if you're recording 1 minute on, 1 minute off is better than 30 minutes on, 30 minutes off. And you can see this little um figure here from WID in 2024 where they looked at the mean difference in vocal activity rate between continuous uh recordings and the GT cycle recordings. So on the y- axis here, you see if it's at zero, it means that there's basically no difference between duty cycled and continuous. And if it's up all the way to 12 here, it means that um the duty cycle there's there's a huge difference. Um and on the x-axis the coverage. So half the time, a third of the time, a sixth of the time, and so on. And within those brackets is sort of the frequency. So 1 minute on, 1 minute off, 10 [snorts] minutes on, 10 minutes off, that sort of thing. And you can see that um even if you're only sampling a sixth of the time, if you're doing more frequently, so 1 minute on, 5 minutes off, you're actually getting a better result than if you're recording half of the time, but 30 minutes on, 30 minutes off. So more frequently is is really important. Um but ideally, obviously, also as long as possible. Now, for the sampling rates, um it's kind of obvious, but we want to make sure that, you know, we're we're getting the frequency of the things that we're interested in. And a really important thing to consider here is that sort of Nyquist Shannon theorem. So that sort of determines that the sampling rate must be at least twice as high as the highest frequency of interest. And that's because you need two sampling points per cycle to catch that crest and the trough um to accurately represent each of the sound waves. And that means if you're interested in let's say a bird at 3 kHz, you want to make sure that you're recording at at least 6 kHz to capture that bird. Now, for the sampling rates, um, if you're recording at around 22 kHz, which is that green box here, you will be getting, you know, most of the birds, amphibians, reptiles, most mammals, uh, most invertebrates, and a few bats. If you then up it to 44 kHz, you're getting additional bats and you're getting additional invertebrates. And then above 44 kHz, um, you're in that ultrasonic, uh, frequency range, and that's where you're getting a lot of those really high pitch vocalizing bats that you can't hear. Um but obviously the higher your sampling rate is the more data storage is required. So that's that's an important thing to consider. Okay. So that's sort of the more general considerations or important considerations when you're trying to come up with a um eocic monitoring project. And I wanted to sort of dive into um what we've been working on which is that uh listening to country uh project which is a collaborative partnership designed to empower indigenous ranger groups by sort of bridging the gap between acoustic data collection and the actionable actionable ecological management outcomes. Um and what we're trying to do is we're building capacity for acoustic monitoring. So sort of enabling the rangers or teaching the rangers how to use uh acoustic devices and think about what needs to be in place for this to happen and add this as a tool to their toolkit um in order to look after and care for country. And it's all based within that brightway science approach which um if you haven't heard of it is defined as a collaborative process of bringing indigenous and western scientific knowledge and methods together to create an ethical, productive and mutually beneficial research. And as the mach people put it, it's an exchange of knowledge and a chance to learn based on mutual respect, responsibility, and connectedness. It's a collaboration that recognizes the value each group brings to a project when there's a shared interest in the outcomes. And um yeah um we want to make sure that we work in close partnerships with each of those groups and co-design um and execute monitoring projects that meet the specific objectives of those groups. And the foundation is really a commitment to a transparent and adaptive and deeply collaborative process to ensure that the rangers needs um yeah ranges needs are are met and they prioritize sort of every step that prioritizes every step of the way. Um and at the moment we've got these four groups here, the Junja Norman B Yman and Tagalaka that we're working with really closely. um as [clears throat] well as we're in convers conversations with um a bunch of other groups and hopefully have uh many more in the future. Uh there are a bunch of different things that a bunch of different objectives that the groups are interested in. Um there's an interest in threatened species detection particularly ghoulian finches and golden shoulder parrots are really really important and of interest to the groups. uh feral animal control and management and in particularly what the impact of that is once you start controlling those uh fereral species um the impact of preservation and sort of restoration of water bodies on biodiversity. They've been uh really badly impacted by angulates and so you know once you actually protect those water bodies what does that mean for biodiversity? uh gathering evidence for the success of cultural burning practices. Uh as well as assessing biodiversity in areas that are really difficult to access regularly like on top of mountains and and yeah really far um hard to access places. And then as a public engagement tool as well um to get sort of the younger generation interested uh ongoing and going back out onto country and getting involved. So you can take those sound clips and potentially go into classrooms and then show them what healthy country looks sounds like um to get them wanting to come out and then help. Now, so far um we've had u the chance to deliver sort of hands-on training to the ranger groups uh on sort of the principles on acoustic monitoring and recorder deployment, equipment maintenance and the data management workflows uh all in order to establish a really an ongoing acoustic monitoring projects where the ranger groups are able to do everything from start to finish by themselves. Um it's really important that we make sure that the sort of um data transfer analysis is really transparent and um can really easily be followed by the groups and we utilizing um EOS that open acoustics um platform to sort of store the audio data have it accessible and then analyze it as well. And ultimately the goal is to build a useful tool to visualize the audio data um really simply so that the groups can really see potentially when their target species uh are active and where they're active. So get information about sort of space and time um when things are are yeah are around. And again this is all based um on the rightway science approach. So it's really important that um we're very transparent. there's a lot of communication and collaboration going on and a knowledge exchange. Um, it's really important to build these relationships based on trust and mutual respect and ultimately the ranger groups are conceptualizing the research ideas and they're there to collect data and they're currently possessing the skills to successfully deploy and maintain audio recorders but sort of the specialized partners required to help realize the full potential of the data. And that's where we come in where we support with the data management and the analysis um process. And then ultimately we want to report those scientific findings together and make sure that the research um as well as management outcomes are met. Now if we're running through the conceptual conceptualization uh using one of those objectives I mentioned, it kind of could look like this. So you think about the goal. The goal in this case is to figure out um if biodiversity is potentially increasing at those water bodies after you're controlling and excluding excluding angulates from uh from water bodies. And so in this case you might be interested in a whole community target with various life histories and um potentially looking at species richness and vocal activity to begin with. Now for that you probably want to get a sort of long-term audio recorder option. So something with a sonar panel that can sit out there and record audio data um continuously particularly in regions where you might not be able to access it maybe due to um weather like the wet season where you might have to have these devices out there for months on end and no one is able to check on them and make sure that they're still still functioning and recording then um so yeah continuous recordings in this case we're interested in want to make sure that there's some regular maintenance schedule uh set up where um groups are able to go out there and and two um checks on the recorders and the SD cards. And in this case, we're using purpose of sampling because we are interested in water bodies, particular water bodies. Um want to make sure that we have multiple copies and that long-term data storage is uh is um uh given. So we create you know local copies as well as in our case we're using echosounds for cloud storage and then we're analyzing the audio data and machine learning um and have uh extensive uh expert validation to make sure that those detections are actually true detections and then yeah potentially develop uh call recognizers for particular species to make sure that you know future audio data that we're getting is going to be more effectively and um quickly analyzed. Thanks, Seb. So, yeah, I'm going to be um conscious of time, so I might give a a very overview at the start, but yeah, I'm Sharon in in at JCU in Townsville. And in my post-doal research, we we've been working with uh Parks Australia and Christmas Island National Park to explore what role acoustic monitoring uh might might have for the species conservation and management. Um using Christmas Island National Park as a test case for national parks generally. Uh this is a a NESP um project supported by the um Australian government. Um and I'll give you a brief overview background of the project and the steps we've taken to build a monitoring program and some opportunities and challenges we faced along the way. So um if you don't know Christmas Island um it's Australian island territory in the Indian Ocean about uh 350 km south of um Java. It is renowned for its high species endemism and um of course its um famous red crabs and the red crab migration. Um and Christmas Island National Park covers about 63% of of the island managed by Parks Australia. Um and it's a good test case for establishing an acoustic monitoring program because it's an isolated island. There are relatively few species there. There are about 22 resident breeding species of bird and as and one species of mammal at Christmas Island fly blinding fox. Um so building this program I'll just introduce the our target species. So we have six um six forest bird species that we're targeting initially in this acoustic monitoring program and the flying fox. And I guess it's just important to point out why we've chosen these species. So they're all endemic to Christmas Island either at the species or subspecies level. Um so they're being in it being in an island habitat, their species of of concern being endemic and four of the bird species are um are currently vulnerable or threatened. The Christmas Island flying fox is um classified as critically um endangered. Um and so you'll meet these birds a bit more later on. Um and so I came up with my own hierarchy of what were the what are the questions that what do we want to know in our um acoustic monitoring program and um starting from a broader um overall question of what we really want to know is in monitoring are the species population stable or are they increasing or are they declining? But to get there we need to using acoustic monitoring data we need to know what are we going to measure. Um so um can we get metrics like sight occurrence, presence or absence out of the acoustic data and vocal activity so using detection rate but then we have further more detailed questions from there. Okay, if we can measure um we have some questions about how reliable is this method going to be? Can we reliably detect these species using acoustic data? When should we record? How many sites should we record? And is vocal activity related to is it is it a good reflection of um abundance? Um and so just an overview of our um actual recording um regime. We recorded at about 80 80 sites give or take you know depending on the project's been going for over two years now. So sometimes you know recorders are in are in and out out of commission. Um uh we commenced back in May 2023 after a pilot program. Um the number of sites generally became restricted by the number of recorders that were available um and working. Uh I we started with a few more in the in the pilot site but that was whittleled down. We used barlt recorders from Frontier Labs down in the left hand leftand bottom picture. Um there we put a um a mesh over the microphone because we had problems with something, we're not sure, chewing the microphone um phone. Um and we also had at the start five solar bars. So that's a solar powered bar um there which we actually found worked well only if it was out in an open habitat. It didn't they weren't reliable in forest because the you know power was quite intermittent. Initially we started recording 24/7. Um but we found that um after a couple of months that uh we the project just didn't have the resources to be able to go out and service and change the batteries um regularly enough. Uh running the bar LTS on battery power. Um they ran out of power after about a month of continuous 24-hour um recording. So we changed that um schedule down to 20 minutes per hour on the hour all day. so that um we could then get 3 months um recording out of each recorder. And we wanted to because we're recording multiple species and the hawk being a nocturnal species, we wanted to record across the whole day. Um and then later in the project again because of resourcing logistics, um we changed that to every second day just because of being able to get out to the island, go around to all the sites and change batteries. So we still had that temporal coverage um over the over the year. And the map there just shows you the the generally the the um sites recording sites that were used across the island. So we wanted an islandwide coverage. Um and just to mention that what we do with our audio when we collect it off SD cards u it's uploaded to the ecoounds repository which is uh was created and is run by Qout eco acoustics research Paul Rose Lab and um Anthony Trusinger. Um and we upload it um from JCU using a nice fast internet connection which is um necessary. It's not much luck if you don't have one. Um and the ecos repository is brilliant. It's good at um you know safe storage. It organizes Anony's written it so that it organizes your audio file metadata for you. It does have playback and annotation download functions. We don't use that in this project because it's a it's a big project. Um and now it also has um in into it analysis functions. So uh with the perch or birdnet um classifiers, it will um generate your embeddings for you and it will soon be generating your um classifier detections for you. So just a note on that. Um and so the question reliable automated call detection, can we do it? Yes, we can. And so I'm going to sort of skim over this because Lynn did touch on um the the state of classifiers where we're at at the moment. But um all this great intense work was done by Phil Iinsky at Qout at the start of the project to build our classifiers and we use the perch um CNN model which is developed by Google research. Um and yeah as Lynn talked about it's it uses that fshot transfer learning. So we only needed to start with uh sometimes one um training example of each species. So we did collect some voucher calls. So an actual you know um when you go out and identify a bird and say yes I'm recording you know the emerald dove. This is it. And that's our example training call. Um or for the gosh hawk we had to go to xenoanto. We didn't get any voucher calls on the island but the gosh hawk is uh sometimes considered a subspecies of the brown gosh hawk. So we use an example call of CNO Kanto for that and that starts the process that's um building the classifiers improving them through iterations. So you can use those initial train examples to find more training examples in your actual field data to build really good um classifiers. Um and just some technical notes before I go on it will it will help with my next slide. The way the perch classifier model works is it it segments the audio into 5-second segments, analyzes that and for each species model, it gives each segment 5-second segment a prediction score as to how confident it is that that 5-second segment has your target species and then what we do with that. So Lynn also um touched on this before. uh we have all these detections um having I think at last count we have over 380,000 hours of audio and so if you work try to work out you know how many 5-second segments that is it's millions and millions and so obviously not all of those segments are going to contain your target species but they're all analyzed and they all get a score by the perch model. Um but you want to count detections. How many of these um call se audio segments have your um actual species in them calling in them. And so we use this logistic regression approach to determine a score threshold. So we want to basically only you know we want to chuck out all the segments that aren't our species um and um count how many detections of our species we've got. So you validate a sample of of audio segments across a range of scores. do a logistic regression um that gives you curves like this. And the one on the left is the uh uh call um dete classifier for the thrush. And ideally you want it to look like this. So when you have um uh the data dots down the bottom are negative examples. So so segments that don't have the species. Um you want that nice curve where it's you know can ideally almost get separation between your ne your negative and positive examples. Um and the hawk there shows an example of where um we can get a score threshold that gives us um high confidence of true positive calls but we're still missing a lot of of calls. Um but um yeah so talking about um what I did with these detection calls firstly I took an occupancy modeling approach and an assumption of occup occupancy models is that there are no false positives. So we I had to choose a score threshold that gave me a very high confidence that I only had um true detections in there. Um and so that's often, you know, the first question that you can address with um this type of data. Um I've applied the score threshold only looked at call detections above the threshold and it can give us a nice picture of where where we're detecting our species. So where do our um species occur, what sites do they use? Um, and the nice thing with acoustic data is if you have long-term acoustic data, it's very um it's a very rich source of of data. You can aggregate it in different ways. So, you can choose different periods. You can aggregate, you know, months and days. This is just um over the our first year of monitoring um the sites that we um detected the species at. And you can see that even um just a couple of examples that these birds are quite widespread across the island. and the emerald dub wasn't detected at some sites, but the pigeon and a few of the other species are detected just about everywhere across the island. [clears throat] Um, so uh, one other question we wanted to address was how does this acoustic um, detection method compare with traditional field surveys? Um, and I did this using the occupancy uh, modeling. So this is the um detection probability estimates using occupancy models um in that I did in R and I compared it with um point count surveys that we did um uh at at our sites. So we did these surveys at our recorder sites um over two years except for the hawk which was a nocturnal survey done in the second year. Um and I calculated the um the occupancy and detection probability in that same period. And I the takeaway to this is that um except for the hawkowl we actually get better or as good as detection probability um using acoustic modeling uh acoustic monitoring data um and with acoustic data we get better estimates so lower um smaller confidence intervals which is um which is better for occupancy modeling. We want um good estimates. Um and then so stepping on from this what we did if you if you'll uh sort of notice back here that the detection probabilities vary. So detection of the dove was and the goss are quite low um but the other species are quite high. So we wanted to focus on what were the times of year that we could detect better detect the emerald dove and the gohawk because their detection probability is generally low whereas the other species are quite high most of the year. Um so to be able to um you know m maximize um the the monitoring capability. So this is a temporal um plot of the monthly detection probability just from the acoustic data to show us that you know there are better times of year than in July and when we did the surveys when the doves were hardly calling they're very seasonal um to give us an idea of when when to record. Um and so from that what I did was I did a power analysis. So I used an approach where we can get uh use an approximation equation to calculate the the statistical power uh and the aim here is to detect a decline in occupancy from one monitoring period to another. So say from this year to next year u what's the what's the power and what's the effect size that we can detect. And this was to um inform you know how many sites do we need to record at uh how many days do we need to record for um and is it feasible and this is the example for the emerald dove which with its low probab probability of detection is the most challenging one. I found in all the cases that the advantage of acoustic monitoring is because we can record for extended periods of time. Um you get better better power than just doing a few surveys. But beyond 30 days, you don't get much much you know extra advantage out of it. Um but it's quite um straightforward. If you're going to put recorders out for a week, there's no sort of extra effort just to leave them out for another few weeks. But what this is uh really showing here is that there's some decisions to be made um depending on you know how many what's what what feasible as far as the number of sites that you're able to monitor um and what's uh what's an what's a meaningful um effect size so occupancy decline that you're wanting to detect. So, say for example, if you had um um 100 um recorders and you had the capability to record at 100 sites for a month, um you could detect, you know, let's just say a bit more than a um 0.25 decline in occupancy with 80% power. So um it gives a real sort of guideline to parks Australia for what kind of effort they need and what uh what um size decline are we able to um detect even for this um uh dub that's difficult to detect sometimes. Um and so yeah moving on. So moving on from occupancy um analysis looking at using that vocal activity. So rather than just reducing the data to presence absence um actually looking at calculating the detection rate. So number of detections this is a plot example of a number of detections um each day standardized for the number of hours recording. So recording effort um and this is a nice example of the seasonality in the emerald dub and the low detection rate. Um the gray box there shows a period where we missed data because a lot of the recorders ran out of power. So we were didn't have enough um recorders uh going there. Um yes. So that's just just to show a use there of how you get really nice um temporal activity patterns. Um and moving on from that, what we also wanted to know was does vocal activity reflect survey counts. So again, we use the survey data that we did in those two two years, the um four four survey visits at each site doing point counts to see if our detection rate was a good um uh you know, I guess if if it related to the survey counts. Um and what we found was there, you know, reasonably good correlations between we've got acoustic detection rate and the average survey count in each in each year. So each of these points is one site in one year. Um and I guess it's important to point out there that um even though you know the correlations aren't perfect that point count surveys um the counts from those surveys they're you know not a true reflection of abundance either. So they're um an index of abundance as well. So here we can see that the acoustic detection rate as an index correlates well with survey counts um as an index and I think I've got yeah got the other examples in there as well um uh just showing that we can correlate um uh the detection rate and so it gives us a good um a good uh foundation that we can use activity vocal activity as uh a metric for monitoring to be able to monitor if there's changes going on. Um how do we go? Okay. And um yeah, so from here what we're doing, we're still exploring the vocal activity metrics. Um I want to get into seeing if we can reduce the false negatives. So using more of the data at the moment, using that, you know, really blunt threshold. Um assures us we have positives, but it also throws out a lot of that um uh you know, um false negatives, I guess we call them. So, um, detections that we're missing because we're using that high threshold. And, um, so there are ways that we can, um, you know, use use detection probability to use more of that data. And then looking at, um, power analysis again to be able to detect changes in that vocal activity is where we want to go with that. Um, and I just want to acknowledge just some acknowledgements there of who's been involved, lots of people involved in this project. Um, and that's me. I think that's my last slide. So, I'm hoping we got time for questions now. [laughter] >> Oops. We all talked for too long. Well, yeah. >> Oh, um, you can go on for another hour in my opinion, but I guess I guess I'm not the boss of it. [laughter] >> Liz, there's only a couple of questions that haven't been, um, asked, but um, if anyone else has any questions, feel free to ask away. Um, can I also just say that we can we can last beyond the hour. So if you if people want to stay online, I understand people are busy, but if people want to stay online and are desperate for an answer, or you can just email us and we'll be happy to answer you. >> Yeah. Brilliant. So if you if you need to run now, please um give us a question and then and take off. Um but yeah, I'm very happy to host people on for another 10 minutes, 10, 15 minutes to um follow up on these. Um well, I'm going to start with the one of the last questions which um was uh which was about the standardizing effort between point counts and acoustic monitoring to do the occupancy modeling and compare them. Sharon, I think you started getting towards that and you you you explored that in your later slides. So, is there anything you'd like to you could say about the standardizing the effort between those two? >> So, I guess when you say standardize, so we wanted to compare the two methods. So, the effectiveness of the two methods. So, we did the surveys were done within a five or six week period. So we actually had a burer go out and do surveys at all the sites four times which is the general approach in occupancy modeling. You you do repeat surveys. So I just looked at at the detections in that same period. So in the so in the five week five to six week survey period. So I got daily detections um daily presence absence out of the acoustic detections in that same period. So we we wanted to do a comparison as you know what's the results if you were to use repeat surveys for occupancy modeling or if you were to use acoustic recording for occupancy modeling what's the outcomes of using either of those approaches so um uh so I guess you know they are different methods so I don't know if standardizing the rights is the right word [clears throat] >> um uh but it's yeah it's comparing the methods to that >> like my assessment of your answer is correct. Um is that that's that's great. Another question. Um I'm going to move over to um one from Clara. For a project tracking site level changes in the population, sorry, for a project tracking site level changes in population composition across seasons. Would it be sufficient to leave the recorders out for a week at a time at each site for each season, or would they be better staying out for longer? Um, I can probably take that unless you want to answer it, Seb. Um, it depends on your animal [laughter] or whatever you try to follow. Okay. So, um, a week at a time might be good if it's a Christmas Island thrush and it never shuts up. But if it has a seasonal time that it calls or something like that, then you need to have your your um your recorders out at the right time. Um, [clears throat] sometimes what you need to do is record for a while to be able to figure out what the right time period is. So, when is that animal most active? And then that way you can um you can detect its activity each season and you know when to record to do that if you can only do it. So, we always we've been recommending more and more to people record 24/7 or do what Seb suggested and record in chunks of time. Um, but record for long periods so that you at least initially get an idea when your animals are active. Does that answer your question, Clara? And I'll Yes. Yes. We got a yep in the chat. >> Oh, great. >> Awesome. Thanks, Lynn. Um, Andrew had an answer for Sharon about um talking about the Christmas Island pipistral insectto virus in insecttoous bat. Um, Andrew, would you like to speak to that comment that you put in the chat? >> Yeah. Oh, well, um, it's just something that's, you know, talking about Christmas Island. And I thought, "Oh, there was this little bat that went extinct a long time ago." Well, not a long time ago. 2009. >> Yeah. Recently. And look, it won't be the first species that's been presumed extinct and then excitedly been rediscovered sometime down the track. Um, Bronato walabe is a classic example of that. Um, but yeah, I just thought if you're out there sticking monitors around Christmas Island, um, and I know you'd have probably have to change the frequency um, settings on your devices, would it be also worth running a recognizer for that if you could create one? There's got to be there's got to be spectrographs because they were monitoring for those bats using anabats. Um, and yeah, so yeah, I just thought you might >> Yeah, there's lots of lots of considerations. Yeah, because bat monitor monitoring in in ultrasound is is um you know yeah has a different kind of approach. Um and I know they did back at the time it went extinct. There was some intensive monitoring going on and that's generally how they would monitor for um microbats is using the um the specialized um ultrasound recorders. Um and I do know that it's the topic comes up every now and then about whether it should just be um tried again to see if they can be um detected. But I do know at the time it was when um the working group was raising the alarm bells. They were intensively monitoring and they basically watched it go extinct pretty much in the monitoring data. They could see it going. Yeah. But yeah, it it's not um you're not the only one who um yeah brings brings that up. It is an ongoing discussion, but it it's kind of um yeah, it recording in ultrasound um uses more power and more data storage. um it's got a specialized so it's it's possible but we have to factor in you know that the file size is going to get so much bigger and and um but it's uh yeah and building a recognizer yeah definitely feasible >> yeah and actually following on from that um I'll read out another um question or couple set of questions from a reflection from Clayton who says logistics power data storage access and labor for maintenance has been repeatedly flagged tagged as a challenge. Is this a real limitation impacting research or management or is it just a frustration that needs to be accepted and contingencies put in place? So Lynn started off answering they're not a frustration. You just need to optimize them so you get the answers you're actually after. But uh Lynn and others would you like to expand um on answers to this [snorts] question? >> I'm happy to Seb. Yeah, unless you unless you want to. Yeah, like I guess it is a real limitation, you know. It's it's how many recorders you have. It's how much money you have to drive, you know, hire cars and drive around and put the recorders out. Um uh and um you know, some analysis shows that yeah, if you had 300 sites, you could really, you know, detect and monitor um really well, but it's it's got to be it it's within limitations of what's what's feasible. That's just a practical, you know, um, consideration. And some people might be frustrated by that. I don't know. Or some people might just accept it. [laughter] >> And often that's where like pilot studies come in, right? Like you give it a try, you record for a certain amount of time, maybe with a certain number of recorders, and then you realize, oo, I actually need more or maybe I don't actually need that many. Maybe I can cut down. you know, like just going out there and then as Lynn mentioned, starting off with continuously recording is probably your best bet just so that you can kind of understand when things are. >> I'm after that. Once you really understand when things are active, you don't want to start off potentially missing stuff. >> Yeah, definitely. >> I can answer that one. says, "Sorry if this is off topic for the webinar, but I was wondering if you can speak to how well-developed the Acoustics Observatory of Australia is. Would it be feasible to use the existing recordings on there to do occupancy modeling in well-developed survey survey areas instead of conducting new acoustic surveys?" Yes, it would be brilliant if somebody used the acoustic observatory to do that. So, we have four recorders uh at each site in the in the um observatory. Um, and it's got and it has two that are close to water and two that are further away. These may not be perfectly useful for you. They not may not be in the right place for whatever it is you want to look at, but it's certainly feasible to check out the locations. Um, my postoc has just published a paper a different postoc slate Alan Anderson has published a paper looking for every threatened species that we could find, threatened, vulnerable, all those sorts of things that we could find on the acoustic observatory. So, if you're specifically interested in a threatened species, you could look at that paper and see whether that species is in there and whether he found it. Um, if he didn't find it, you could talk to him about how hard he looked. Um, because he looked reasonably hard, but he had 73 species to look for, so he didn't look all that hard. So, um, hopefully, um, that is helpful. Uh yes, you can absolutely access all the data there and you could look through it and you we can help you look through it and facilitate your use of the A2O. Absolutely. Does that answer questions? >> Awesome. >> Might help you with what do you call it? Um pilot study for example. >> Yeah. >> Somebody called Tom has his hand up. >> Yeah. Go for it Tom. You'll need to unmute. >> Thanks. Yeah, I was just trying to figure out how to unmute that. Um, I have two questions. Um just looking at some of the data that you guys were presenting and I um if with the vocalizations, it was kind of making me think of like how how do you deal with um the same species um say a a bird um making different types of coals. So a territorial coal versus um some sort of mating attracting coal. Um and and that goes for different species too, but uh yeah, primary interested in that. And second, this is a little bit more of a I don't know if this has ever happened yet, but it's sort of maybe an interesting hypothetical, but um what do you think would happen if uh something like a liar bird came along and um made a bunch of different vocalizations in front of one of these um devices, listening devices? Would that would you be able to tell if it's a live bird versus the real thing? >> Um, I'll answer I can answer those questions. >> First of all, um, we don't know anything about broods. We work on frogs and mammals and [laughter] and most of them only have one call. No, that's true, but it's also not not really correct. So, what we tend to use for a lot of the birds is things like contact calls that they make all the time um and that are are useful, but you can actually build a recognizer with several different calls. Um, and it does it doesn't matter. It learns to recognize them just the way a person does in a way. And so, it's fine. You can use all the calls of a bird and find it. It's nice to keep them separate though because you might want to use, for example, territorial sounds at some seasons and you might want to use calls at other seasons. that kind of thing. Um, but [clears throat] we do not specialize on that in our lab. In our lab, we're focused on identifying, finding things, and looking at levels of activity. So, we haven't done that. As far as the liar bird goes, it's my understanding that people can tell the difference between a liar bird mimic of a call and an actual call by that bird. Is is that true? I've seen someone do that, so I'm thinking that that's true. Um I just wanted to say also that I just shared the link in here um because that's a question that we quite often get with the li birds. Um there's a paper um on the vocimicry of libertirds and if you look at the spectrograms in there there's a slight difference between the vocalization of a li bird and the animal that it's trying to copy you know whether that's a kabara or something. So the thing is more about how much data do you have. If you have enough training data, then you can distinguish those two for sure. At the start, you might not be and it might actually pull out the same species, but then it just attempt it depends on how much data you have. But if you have enough training data, if you have enough of the liar bird vocalization copying the cookabara and the actual cucara, you should be able to tell them apart because there is if you look at those spectrograms in that paper for example, there is a difference in the vocalizations. >> But it's not easy. And when we say enough training data, it means you need to have examples of the liar bird copy and you need to have examples of the not liar bird copy and you need to be able to chuck them into your data and find both of them and look at the difference. So you yes, you should quickly be able to build recognizers that tell the difference, but it'll be harder than building a recognizer for two different species. >> Wow, great. Well, those were awesome questions, Tom. And I feel like now's probably a good time to probably wrap up. I would love to go into this another time. So, please um uh if you are keen to revisit this conversation um or extend it on in any particular direction, please get in touch with Lynn, Sharon, Sebastian, myself, Andy. Um I'm not an ecologist, sorry. just butting and like talking about this stuff cuz my kids are bur now and so I'm so excited. >> Um, >> thank you for being part of this. It's my >> would like to thank all of our presenters today for your your time and diligence and sharing your expertise and we just would like to thank uh the audience and the great questions you've given today and um we hope that we get to see you in one of these series in the future. So thank you very much everyone.