The secret acoustic world of leopards: A paired camera trap and bioacoustics survey facilitates the individual identification of leopards via their roars
Editor: Vincent Lecours
Associate Editor: Mariane Kaizer
Abstract
Conservation requires accurate information about species occupancy, populations and behaviour. However, gathering these data for elusive, solitary species, such as leopards (Panthera pardus), is often challenging. Utilizing novel technologies that augment data collection by exploiting different species' traits could enable monitoring at larger spatiotemporal scales. Here, we conducted the first, large-scale (~450 km2) paired passive acoustic monitoring (n = 50) and camera trapping survey (n = 50), for large African carnivores, in Nyerere National Park, Tanzania. We tested whether leopards could be individually distinguished by their vocalizations. We identified individual leopards from camera trap images and then extracted their roaring bouts in the concurrent audio. We extracted leopard roar summary features and used 2-state Gaussian Hidden–Markov Models (HMMs) to model the temporal pattern of individual leopard roars. Using leopard roar summary features, individual vocal discrimination was achieved at a maximum accuracy of 46.6%. When using HMMs to evaluate the temporal pattern of a leopard's roar, individual identification was more successful, with an overall accuracy of 93.1% and macro-F1 score of 0.78. Our study shows that using multiple modes of technology, which record complementary data, can be used to discover species traits, such as, individual leopards can be identified from their vocalizations. Even though additional equipment, data management and analytical expertise are required, paired surveys are still a promising monitoring methodology which can exploit a wider variety of species traits, to monitor and inform species conservation more efficiently, than single technology studies alone.
Introduction
Globally, biodiversity is declining because of an increase in negatively impacting anthropogenic activities, habitat loss, prey depletion, pollution and climate change (Dickman et al., 2014; Craigie et al., 2010; Ceballos et al., 2015; Isbell et al., 2023; Jaureguiberry et al., 2022; Wolf & Ripple, 2016; Woodroofe & Frank, 2005). To reverse declines, conservation must be guided by reliable data. Collecting these data, however, is often challenging, particularly for species which are nocturnal, cryptic, elusive, solitary or endangered.
Leopards (Panthera pardus) are one such example of an elusive, solitary species. Whilst leopards harbour one of the largest geographic distributions of any big cat (Jacobson et al., 2016), this is still only at 20% of their historic range (Wolf & Ripple, 2017). Moreover, across Africa, there has been a continental decline in populations (Stein et al., 2020). Subsequently, according to the IUCN Red List of endangered species, leopards are classified as vulnerable to extinction (Stein et al., 2020).
Due to the threatened status of leopards, much research has been conducted into the threats impacting leopard occurrence, density and behaviour (Balme et al., 2014; Jacobson et al., 2016). Whilst these studies are important, they have mostly occurred in small, highly protected reserves (Balme et al., 2014; Henschel et al., 2011; Strampelli et al., 2018), even though African leopard range contractions are continental, and some sub-populations in North and West Africa have experienced such dramatic population reductions that they are now believed to be locally extinct (Stein et al., 2020). As such, only recently have comprehensive surveys been conducted on less studied populations (Searle et al., 2020, 2021). Considering that leopard research is limited by the presence of highly trained practitioners or good infrastructure, it is important to consider how other, easy-to-use remote sensing methods, may assist large-scale understanding of leopard persistence and how, as a species, they are affected by an assemblage of threats (Stein et al., 2020).
One area of leopard research which has received little attention, quite possibly because until now it has been too challenging to obtain the data, is vocalizations and communication. Leopard calls consist of a repeated pattern of strokes which produce the iconic ‘sawing roar’ (Eisenberg & Lockhart, 1972). These individual strokes, or roars, when viewed using a short-time Fourier transformation, have a low fundamental frequency which can be attributed to the species' long and heavy vocal folds and an elongated vocal tract (Weissengruber et al., 2002). This trait is independent of sex as both males and females are known to roar (Ulmer, 1966). However, unlike other felids such as lions (Panthera leo) or tigers (Panthera tigris), very little has been published in the scientific literature about leopard vocalizations. Initial leopard vocalization studies were confined to zoos where it is easier to observe leopards for prolonged periods of time (Desai, 1976; Eisenberg & Lockhart, 1972; Ulmer, 1966). More recently, in-situ studies have been conducted, but focus more on the reaction of prey to leopard playback experiments (León et al., 2023; Schel et al., 2010) or accidental, novel observations (Corredor-Ospina et al., 2021; Laman & Knott, 1997). Nevertheless, it is thought that leopards vocalize for similar evolutionary purposes as other large carnivores: mate attraction and territoriality-related activities such as defence or boundary discrimination (Grinnell & McComb, 2001; McComb et al., 1994; Stein & Hayssen, 2013). These evolutionary traits rely on the assumption that individuals can identify conspecifics via their vocalizations (Mccomb et al., 1993) and therefore know when, and when not, to communicate (McGregor et al., 1997), such as by eliciting vocalizations under mating (Laman & Knott, 1997). Indeed, within the Panthera genus, individual tigers and lions have been shown to possess unique roars (Ji et al., 2013; Wijers et al., 2020). The ability to identify individuals from their roars is also a useful ecological monitoring tool. Discriminating individuals within a species enables population densities to be estimated (Efford et al., 2009), or the spatiotemporal movement of an individual to be tracked (Wijers et al., 2020). Therefore, if leopards have distinct individual roars this could be used as an alternative method to monitor the species.
Producing a large enough dataset which can adequately demonstrate auditory discrimination in large African carnivore species is difficult. These species roam widely across often inaccessible areas which extend beyond protected areas (Ripple et al., 2014). Previous methods have relied upon fixing bioacoustics loggers to collars (Lehmann et al., 2022; Wijers et al., 2020) or recording individual vocalizations through handheld microphones (Hartwig, 1995; Lehmann et al., 2022). Whilst these methods were successful in showcasing that individuals of these lion, spotted hyaena (Crocuta crocuta) and African wild dog (Lycaon pictus) are vocally distinct; the approaches are not transferable across all species. Collaring is an expensive, invasive and risky procedure; observation studies are time-consuming and can only be used for individuals of species which are easy to rediscover so that a sufficiently large dataset can be gathered. Therefore, a different method is needed to study the vocalizations of elusive, solitary species like leopards.
Two of the most commonly used passive data collection technologies are camera traps and autonomous recording units (ARUs); however, there are few ex-situ studies which integrate both (Buxton et al., 2018). Combining passive non-invasive data capture technologies requires additional finances, time and human effort, which means that these modes of study are still limited. Nevertheless, the data which could be produced from multi-technology studies may be of key importance for species which are elusive, solitary or nocturnal and are therefore difficult to monitor through traditional methods alone.
Here, we investigate whether individual leopards produce unique calls based solely on the fundamental frequency of their roars. To achieve this, we pilot the first large-scale (~450 km2) paired camera trap and autonomous recording survey for large African carnivores which facilitated the synthesis of a sufficiently large dataset to test whether leopards exhibit unique vocalizations.
Materials and Methods
Study site
The study took place in the Matambwe sector of Nyerere National Park, which lies in southern Tanzania and between the latitudes of −7.505° and − 7.823° and longitudes of 37.728 and 38.308° (Fig. 1). Vegetation is dominated by floodplain grassland and Acacia woodland (Creel & Creel, 2002). There are two distinct seasons in southern Tanzania: a dry season from May to October and a wet season from November until April, with an average rainfall of ~1400 mm (Creel & Creel, 2002). Permission to conduct research in Nyerere National Park was granted by the Tanzania Wildlife Research Institute (TAWIRI), Tanzania National Parks Authority (TANAPA) and the Commission for Science and Technology (COSTECH), under research permits 2023-780-NA-2023-879 and 2023-665-ER-2021-287.
Data collection
Camera trap and passive acoustic monitoring array
In total, we deployed 64 stations (average spacing 2.54 km) comprised of paired xenon white-flash camera traps (Cuddeback Professional Colour Model 1347, Non Typical Inc., Wisconsin, USA) to maximize the chances of capturing both flanks of animals passing by. Cameras were mounted to trees on each side of the road or trail at a height of 30–40 cm, in protective cases, and secured with binding wire to prevent damage and loss by both humans and animals. Cameras were stationed along roads and trails to maximize the captures of carnivores (Cusack et al., 2015).
At 50 of the stations, we also installed autonomous recording units (CARACALs). CARACALs are custom-built, low-cost ARUs, that record in quadrophonic format via four active microphones (M1:M4) positioned on the printed circuit board at 90° intervals (Wijers et al., 2019). To produce directly comparable acoustic and photographic data, the locations of the CARACALs, as closely as possible, matched that of the cameras (Fig. 2). The largest distance between a CARACAL and the centre of a camera trap station was 74.5 m, with the mean (±SE) distance 13.1 ± 1.80 m. CARACALs were installed at heights between 2 and 4 m. Care was taken to ensure the M1 microphone was positioned facing northward and that the devices were parallel to the ground.
The entire camera trap and acoustic array was active for 62 days between 20 September and 20 November 2023. All batteries and the ARU's micro-SD-cards were changed, approximately, every 21 days. The acoustic array recorded audio data continuously at 16 kHz sampling rate and 32 bits/Sa(mple); data were saved in 1-hour files in .wav format. Cameras only recorded data when triggered.
Vocal individuality
Photographic events of leopards were first extracted, and images of both flanks were used to identify individuals by their rosette patterns (Henschel, 2008). The timing of these events was then used to investigate whether a photographically identified leopard roared in the concurrent audio recording. Manual processing of the audio recordings was made in Audacity 3.4.2 (Audacity Team, 2024) by visually inspecting spectrograms and annotating the timing of leopard roaring bouts. Only labelled leopard roaring bouts that occurred within 10 min either side of the photographic event were then extracted. Ten minutes was selected as a trade-off between maximizing dataset size and minimizing the chance of misattribution. Misattribution was tested for by conducting two sensitivity analyses that (a) identified the effect of purposeful mislabelling and (b) how model classification performance changed when including new data as time from the photographic event increased (Data S1). As an additional validation, each bout was also subjected to angle of arrival calculations facilitated by the quadrophonic hardware (Fig. 3); slight differences in signal arrival time at each microphone were used to estimate the direction of the bout to ensure it matched the direction in which the leopard arrived or departed the camera trap station (deduced from the leopard's direction of travel in the camera trap images). Roaring bouts that were affected by technological malfunction, or which contained geophonic or biophonic interference were excluded, as well as events where multiple leopards were photographed within the same hour at the same station.
Leopard roaring bouts are comprised of three stages (Fig. 4). An initial and concluding ruff are sandwiched between successive strokes/roars. Within a roar there are two actions which creates the distinctive ‘sawing’ noise. Only the second part of the vocalization was used for this analysis because it is of greater amplitude which creates a clearer fundamental frequency contour (Fig. 4).
Extraction of F0 summary features
First, each roar was subjected to a digital Butterworth bandpass filter between the frequencies of 150 and 350 Hz, using a 2048-point moving Hann window with 68% overlap. The extracted fundamental frequency contour (the F0 contour) was computed using a window size of 128 samples and by selecting the top 1% of the signal's amplitude. The extraction process was conducted using the autoc() function from the seewave package (Sueur et al., 2008). Following Ji et al. (2013) and Wijers et al. (2020) we aimed to extract four summary features from the F0 contour: minimum F0, maximum F0, mean F0 and roar duration.
To identify unique leopards, we analysed the performance of two different clustering algorithms: k-nearest neighbours (KNNs), which have previously been used to identify individual lions, and random forests, which have previously been used to identify spotted hyaena individuals (Lehmann et al., 2022; Wijers et al., 2020). We used the function knn() from the R package class version 7.3-20 (Venables & Ripley, 2002), and the random Forest() function from the randomForest version 4.7-1.1 package (Liaw & Wiener, 2002). To prevent bias resulting from temporal autocorrelation between roars within the same bout, we used leave-one out cross validation, whereby the test bout was kept separate from bouts included within the training data. All summary feature extraction and statistical analysis was conducted in R version 4.4.4 (R Core Team, 2022).
Hidden–Markov Model classification of F0 contour
A Hidden–Markov Model (HMM) allows probability distributions to be modelled over a sequence of observations. As such, we used 2-state Gaussian HMMs to model the temporal pattern of the F0 contour for each individual leopard. As before, we split our training and test data via leave-one out cross validation. Every bout was used once as a test sample and the remaining bouts formed the training data to build an HMM for each leopard. For each unique test bout, the HMM's were rebuilt using all the data with that bout excluded. This entire process was repeated over 1000 iterations with randomized parameter initializations. The test data were then compared with each HMM using log-likelihood to determine which HMM the unseen sample F0 sequences were most likely to be assigned to. The HMM which produced the highest log-likelihood value was determined to be the predicted class. The HMM classification was assessed using the performance metrics: accuracy, recall, precision and F1-score. We report these metrics for each leopard and the overall classification performance. HMM classification was performed in Python version 3.9.12 using the ‘hmmlearn’ library (Hmmlearn Development Team, 2024).
Results
Leopard captures
Photographic data were collected from 62 days, across 50 stations, totalling 3224 trap nights and 191 leopard capture events where 42 unique leopards were photographed and identified. There were 23 photographic events where a single leopard also roared within 10 minutes. From these events, there were 14 individual leopards (10 males, 4 females), 10 of which (8 males, 2 females) vocalized more than 1 bout. Of these 10 leopards, one leopard's (M1) recording suffered from technological malfunction, whilst recordings of leopards H1 and Z1 contained roaring bouts that had too low signal-to-noise ratio to extract individual roars. In total, across 7 leopards there were 26 useable bouts. From these, 217 individual roars were extracted (Table 1).
| Individual | B1 | D1 | F1 | G1 | H1a | J1 | K1 | M1a | N1 | R1 | S1 | T1 | X1 | Z1a |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Sex (M/F) | M | M | M | F | M | M | F | M | F | M | M | F | M | M |
| Number of bouts | 1 | 3 | 4 | 2 | 4 | 10 | 2 | NA | 1 | 1 | 2 | 1 | 3 | 3 |
| Total number of roars | 10 | 24 | 23 | 10 | NA | 84 | 16 | NA | 13 | 16 | 16 | 13 | 44 | NA |
- a Leopards for which recordings could not be used (see text).
Angle of arrival calculations
Of the 26 useable bouts, three were recorded in the same minute as one of their respective camera trap images and as such were discounted from angle of arrival calculations as they could be confidently assigned to the individual photographed. Of the remaining 23 bouts, 20 approximately (within 50°) matched the expected arrival or departure angle if the leopard was continuing along a road. Three bouts could not be validated with angle of arrival. This was because geophony, not in the frequency range of the roar, affected angle of arrival calculations but not the quality of the roar. However, all three bouts were within 10 minutes of the leopard being photographed so were still confidently assigned to the individual photographed.
Fundamental frequency summary features
Across all leopards, the max F0, min F0, mean F0 and roar duration were significantly different (all P < 0.05) (Table 2). KNN and random forest models built using each of the F0 features varied in their ability to accurately classify leopards. The performance ranged from an accuracy of 32.4–41.3% for KNN and 20.5–32.9% for random forest (Table 2). However, for both KNN and random forest, overall classification accuracy increased to 46.6 and 41.9% when all four features were combined.
| Feature | k | KNN | Random forest | χ2 | Kruskal Wallis | P value |
|---|---|---|---|---|---|---|
| Accuracy (%) | μ Accuracy (%) | Degrees of freedom | ||||
| Min F0 | 22 | 32.4 | 30.3 | 41.9 | 6 | <0.001 |
| Max F0 | 11 | 32.5 | 26.3 | 26.6 | 6 | <0.001 |
| Mean F0 | 30 | 33.6 | 20.5 | 48.2 | 6 | <0.001 |
| Duration | 15 | 41.3 | 32.9 | 123.4 | 6 | <0.001 |
| Combined | 8 | 46.6 | 41.9 | — | — | — |
HMM classification of F0 contour
Leopard F0 contours differed (Fig. 5). Modelling the temporal pattern of individual F0 sequences resulted in high classification performance with a macro: accuracy, recall, precision and F1-score of 93.1%, 80.7%, 76.4%, 0.78, respectively (Fig. 6). Performance varied across individual leopards (Table 3). The lowest and highest F1 score was 0.65 (S1) and 0.99 (G1).
| Leopard | Accuracy (%) | Precision (%) | Recall (%) | F1-score |
|---|---|---|---|---|
| G1 | 99.8 | 100 | 97.4 | 0.99 |
| F1 | 93.4 | 74.4 | 86.5 | 0.80 |
| X1 | 95.0 | 75.5 | 83.5 | 0.79 |
| K1 | 96.4 | 77.9 | 74.7 | 0.76 |
| J1 | 83.6 | 89.3 | 65.1 | 0.75 |
| D1 | 91.2 | 57.9 | 87.5 | 0.70 |
| S1 | 94.1 | 60.0 | 70.2 | 0.65 |
Discussion
Leopards are a solitary, elusive, primarily nocturnal species which makes them challenging to study (Bailey, 1993). Currently, data collection relies upon camera traps (e.g., du Preez et al., 2014; Searle et al., 2021; Strampelli et al., 2018), telemetry studies (Nams et al., 2023) and spoor surveys (Searle et al., 2020; Stander, 1998), which are initially expensive, invasive and time-consuming, respectively. Technology such as ARUs, that can be cheap, have large detection areas, and be deployed for long periods of time without the need for regular servicing may improve leopard monitoring. Here, we make two major advances. First, we demonstrate that an inexpensive passive acoustic monitoring (PAM) device—CARACAL—can be used to acquire individual leopard vocalizations as an additional data source. Second, we demonstrate that using the fundamental frequency contour of the second part of an individual roar can be used to identify individual leopards with high accuracy. We provide further evidence that individuals within species can be differentiated using relatively simple methods, such as HMMs (Clemins et al., 2005; Ji et al., 2013; Wijers et al., 2020). Similar to lions (Wijers et al., 2020), tigers (Ji et al., 2013) and African elephants (Loxodonta africana) (Clemins et al., 2005), leopards have a simple, low frequency, vocal signature. We hypothesize that HMMs would be an appropriate tool to detect vocal individuality in other taxa such as cetaceans or anurans, that demonstrate (1) relatively simple vocal signatures and (2) inter-individual vocalization variation, which is sufficiently large and greater than the intra-individual vocalization variation. Below we outline the potential monitoring applications this creates for leopards and the wider applications of dual non-invasive passive technology studies.
Implications for leopard monitoring
This study provides the first evidence that leopards can be detected using ARUs. We demonstrate that leopard vocalizations can be recorded, and because not all leopard roaring bouts were recorded directly in front of the camera and ARU complex, leopard presence can be inferred at a greater detection range than that provided by camera traps. As leopards are a cryptic, elusive species, that can inhabit home-ranges as large as 800 km2 (Bothma & Le Riche, 1984), an increased detection range harnesses great potential for occupancy or behaviour studies. Future research avenues may be to investigate whether large-scale leopard occupancy studies can be conducted in less time using ARUs than with camera traps (Crunchant et al., 2020; Garland et al., 2020) or to investigate whether leopard vocal behaviour changes because of spatial or atmospheric conditions (Wijers et al., 2021).
Intraguild and anthropogenic factors may also alter the rate of leopard vocal activity. Lions and leopards coexist; however, leopards will, at a fine scale, temporally (Miller et al., 2018) and spatially avoid lions (Searle et al., 2021). Lions are known to attack and kill leopards (Balme, Pitman, et al., 2017) and will also steal successful hunts if the opportunity arises (Balme, Miller, et al., 2017). Given that leopards are vulnerable when in the proximity of lions, it may be that leopard communication is limited, to prevent discovery, in areas where lions exist in high densities. Similarly, leopard vocal behaviour may be impacted by anthropogenic pressures. Historically, targeted poaching of leopards has been a major driver in the decline of the species (Inskip & Zimmermann, 2009) which may have contributed to leopards fearing humans more than lions (Zanette et al., 2023). Thus, until further research has been conducted, PAM may not always be an appropriate single technology tool to monitor leopards across all habitat types and anthropogenic gradients.
The ability to collect acoustic data of leopards without the need for invasive (collaring) or time-consuming methods (human observations) is highly promising. Vocal discrimination, alongside a sufficiently high enough rate of vocalization, could facilitate the tracking of the spatiotemporal movement of individuals across a landscape (Wijers et al., 2020). Furthermore, given the species' vocal individuality, leopard population density could be assessed using a combination of PAM and spatial capture-recapture approaches (Lucas et al., 2015; Stevenson et al., 2015; Efford et al., 2009). However, in this specific study, only one third of all unique, photographed leopards roared 10 min from their photograph being taken. Thus, here, to conduct an acoustic population density estimate, we would need to analyse all available acoustic data and not just the very small subset of audio that occurred alongside leopard camera trap events.
Advantages and limitations of integrating PAM with camera traps
This study confirms the applicability and potential of using a paired camera trap and ARU approach to demonstrate vocal individuality in a species, which, would not have been possible through the deployment of just one technology. We demonstrate that using multiple technologies, which record complementary modes of data, can be used to discover species traits. Broadening our data collection methods will improve our ability to assess a species' population status and how they inhabit their landscapes. This is particularly important given global threats to biodiversity (Midgley & Bond, 2015). Improvements in biomonitoring techniques are altering how we study biodiversity (Hodgson et al., 2018), identify hotspots, organize protective patrols (Hambrecht et al., 2019; Katsis et al., 2022), and monitor the impacts of anthropogenic disturbance (Buxton et al., 2018; Pardo et al., 2022; Vélez et al., 2023). We add to the growing body of literature that suggests using PAM, in conjunction with camera traps, can, when the research questions permit, be a promising, complementary monitoring tool which exploits a wider variety of species traits, to monitor single species and whole communities more efficiently, than single technology studies alone (Crunchant et al., 2020; Francomano et al., 2024; Garland et al., 2020; Vélez et al., 2023; Zwerts et al., 2021).
When integrating multiple technologies for the same survey, researchers do need to consider how the additional data will be handled and whether a paired survey, and the additional technology, is worth the increased financial costs (e.g., equipment, batteries, data storage) and labour. In the field, we did not find that any additional time was required to service both technologies. However, this was primarily because one team member serviced the camera traps and another the ARUs. Had only one person been responsible for checking both technologies, whilst this would have resulted in only one salary, time in the field would have increased, and with it, additional extra financial costs (e.g., ranger fees). To alleviate some of the additional costs associated with dual technology studies, future development of a technology which has the capability to continuously record audio whilst only photographing animals when triggered could be of high value to ecologists and conservationists.
An additional consideration of integrating PAM with camera traps is the data processing costs. Large-scale PAM produces large amounts of data, which, manually, requires a longer time to analyse than camera trap images. For our acoustic survey, we accumulated ~72,000 h of data which totalled ~35 terabytes. To analyse this volume of data quickly and efficiently, automated species classifiers are recommended (Stowell, 2022). The production of species specific automated acoustic recognition algorithms is now common (Stowell, 2022). Additionally, more user-friendly GUI's such as Arbimon (https://rfcx.org/ecoacoustics ), Kaleidoscope ( https://wildlifeacoustics.com ) or PAMGuard (https://pamguard.org ) can perform automated detection on audio originating from a wide variety of environmental samples. However, for many species, classifiers do not exist either due to a lack of research interest or training data. Nevertheless, in this instance, we were able to bypass the requirement of an automated species classifier because the camera trap images provided timestamps to be targeted in the audio recordings which meant that data extraction was quick. However, for PAM to be an efficient monitoring tool, one of two options must be available: (1) the existence of an automated classifier which accurately extracts timestamps of vocalizations from the acoustic data for the desired species; (2) the volume of data collected is reduced permitting manual analysis. Some acoustic surveys are specifically designed to record data in intervals, (e.g., 1 min every hour) and this can provide a smaller, but sufficiently informative, dataset to answer many questions (Metcalf et al., 2023). However, if the ARU has been programmed to record at intervals and a researcher is interested in individual identification, there is the potential that an incomplete vocalization (e.g., a leopard roaring bout) is recorded, which may lead to insufficient data collection. Instead, using a posteriori knowledge of the target species' behaviour should be considered. For example, lions and leopard's vocalize primarily at night, or, at dusk and dawn (Bailey, 1993; Hamilton, 1976; Pfefferle et al., 2007; Turnbull-Kemp, 1967; Wijers et al., 2021). When targeting these species, it could be recommended to only record between dusk and dawn (a 12-h cycle). This would not only save data storage, but also ARU battery life, thus, extending the time between deployment and servicing. As such, this could then reduce the number of field checks, batteries, SD cards and external hard drives required for a full survey, whilst also ensuring minimal impact on the camera set-up. However, checking devices fewer times over a long survey period does reduce a researcher's ability to correct mistakes or solve field issues e.g., replace broken or stolen equipment, change corrupt memory cards or swap dead batteries.
Wider conservation applications
PAM has mostly been used as a terrestrial monitoring tool for birds, bats, primates and anurans (Gibb et al., 2019; Sugai et al., 2019), whereas camera traps studies primarily focus on medium and large terrestrial mammals (Burton et al., 2015; Chen et al., 2022; Cordier et al., 2022). However, paired camera trap and ARU studies have been used across different taxonomic groups (Buxton et al., 2018). The capability for paired surveys to discover fundamental knowledge is of huge value. Novel discoveries can inform monitoring by providing alternative methods for how best to acquire data for a target species. Whilst good conservation is routed in evidence-based decision making, resources for acquiring necessary information are limited. Thus, it is imperative that we use the most efficient methods to gain data that can inform practitioners and decision makers. Only by testing innovative methods will we know whether current guidelines and practices are appropriate for the species that they are designed to monitor. It is our expectation that non-invasive, multi-technology studies will continue to produce complementary datasets that facilitate novel discoveries which previously would have only been remotely possible through a combination of highly invasive, expensive and time-consuming methods.
Author contributions
JG, MW & BIS conceived the ideas and designed methodology; JG & CS collected the data; JG analysed the data; JG led the writing of the manuscript. AM & MW designed audio devices and provided remote field assistance. AM wrote code for the audio devices and for audio analysis pipeline. All authors contributed critically to the drafts and gave final approval for publication.
Acknowledgements
We appreciate the support from the Government of Tanzania, Tanzania Wildlife Research Institute (TAWIRI), Commission for Science and Technology (COSTECH) and Tanzania National Parks Authority (TANAPA) for permitting this research in Tanzania. This work was supported by the Frankfurt Zoological Society and Lion Landscapes. Particular thanks are extended to Lion Landscapes field assistants Isaya Kachira and Joseph Francis. JG was funded via a doctoral training grant awarded as part of the UKRI AI Centre for Doctoral Training in Environmental Intelligence (UKRI grant number EP/S022074/1). The camera trap survey was funded by Wildlife Conservation Network's Lion Recovery Fund (TZ-RC-02), the Darwin Initiative Capability and Capacity fund (DARCC009), and WWF Germany (213/10143411).
Conflict of Interest Statement
The authors have no conflicts of interest to declare.
Statement on Inclusion
Our study brings together authors from a number of different countries, including the country in which this study was based. Authors were engaged early on and before any fieldwork commenced. Whenever possible, our research was discussed with local ecologists to build capacity for conservation technology in the landscape.
Open Research
Data Availability Statement
Data will be uploaded upon acceptance and be made publicly available from the Figshare Repository. Code will be published upon publication with all necessary licencing and attributions that cannot currently be included due to the anonymous peer review process.
