Post provided by Peter H. Wrege

Heard but not seen, populations of forest elephants (Loxodonta cyclotis) are rapidly declining due to ivory poaching. As one of the largest land mammals in the world, this species is surprisingly difficult to observe in the dense forests of Central Africa, but their low frequency rumbles can be recorded. With the autonomous recording afforded by passive acoustic monitoring (PAM) though, we have a window onto forest elephant ecology and behaviour that’s providing data critical to their conservation and survival.

The diverse ways that PAM can contribute to conservation outcomes is growing and while still underappreciated, the availability of relatively inexpensive recorders, increased power efficiency, and powerful techniques to automate the detection of signals have led to an explosion in use. In 2007 there were only about 20 published papers using PAM techniques, but since then over 400 papers have appeared in peer-reviewed journals.

Essentially, PAM is the automatic recording of sounds in a given environment, often for long periods. The trick, and often greatest challenge, is to find the signals of interest (bird calls, elephant rumbles, gunshots) within the recordings. With these signals we can quantify abundance, occupancy and spatial or temporal patterns of activity. Particularly in landscapes or ecosystems where visual observation is difficult (e.g. oceans, rainforests, nocturnal environments) PAM may be uniquely capable of delivering informative and unbiased data. Because PAM is a relatively new method but of considerable interest across the disciplines of ecology, behaviour and conservation, there is huge interest in refining the sampling and statistical methods needed to deal with the peculiarities of acoustic data.

Landscape level insights

The ability to collect data continuously over large spatial scales, even in very remote locations, without observer presence or bias, offers unprecedented temporal detail on the occurrence of repeated events. For example, Astaras et al. 2017 monitored illegal gun-hunting in Korup National Park, Cameroon, over several years. It was known that poaching inside the national park was common (mostly for primates and small ungulates), but it was frightening that they recorded an average of one gunshot every two days in each 4.5 km² portion of the park year after year. These acoustic data provided details on the seasonal, diel, and spatial pattern of poaching that were unsuspected based on physical evidence obtained by anti-poaching patrols. For example, could the animation below shows how hunting activity changed through the week relative to bush meat market day in the nearby villages.

A study of forest elephants in Central Africa highlights the potential, but also the challenges, of using PAM at the landscape scale. The Elephant Listening Project (ELP) at Cornell University, in collaboration with the Wildlife Conservation Society, is establishing a grid of 50 acoustic units in the Nouablé-Ndoki National Park, Congo. With an average unit separation of 5km, the grid will cover 1300 km² of forest. Running continuously, this grid will record more than 8TB of sound every four months (and take the Congolese acoustic team 40 days negotiating the rivers, swamps, and spiny-palm tangles of the forest to refresh power and collect data). How do we find the gunshots and elephant rumbles in this vast matrix of forest sounds though? And how will this study inform and enhance conservation strategies?

Informing Conservation Strategies

PAM conducted on a landscape level can directly enhance the protection of the national park not only through a better understanding of how elephants move across the landscape, but by quantifying where the gun-hunting is happening. Currently the most common approach to counter poaching is patrols by eco-guards. But many protected areas are vast, posing the challenge of where to target patrol activity and how to design patrol strategies to effectively counter shifting poaching tactics. Patrol-based evidence of poaching activity and patrol impact is, at best, inherently biased and, at worst, misleading.

By recording actual gun-shots in the forest, PAM provides an unbiased assessment of where and when poaching is occurring. Given a baseline pattern of hunting frequency, this can be used to assess whether changing patrol tactics are actually translating into reduced poaching pressure. By recording elephant calls, we can discover patterns of landscape use and identify resource hot-spots where elephants accumulate at high densities. This can also inform when and where extra patrol attention might be fruitful.

Nearly all protected areas in Central Africa are bordered by forestry and/or mineral concessions that, if managed thoughtfully, could provide critical buffer zones of forest for the megafauna of tropical Africa (gorillas, chimpanzees, bongo antelope, forest buffalo, elephants). Acoustic monitoring can measure the impact of industrial activity in these concessions at meaningful spatial scales and identify movement corridors to critical resources outside of the protected areas.

It’s not all easy listening: facing the challenges

While advances in technology have vastly simplified the operational side of PAM, the challenges posed to analyse and interpret the resulting data are considerable. Typical PAM applications quickly produce terabytes of sound files that need to be managed and the data we’re looking for are individual signals buried within these long files.

What many new users of PAM don’t realize is that to go much beyond simple ‘presence-absence’ in organismal studies, you need to know some detail about the communication system and behavior of the study species. If you put in the hours, you can find the target signals in sound files, but what does one elephant rumble, or 15 rumbles, mean in terms of number of animals? In other words, what is the cue production rate and how might this vary with sex, social context, season, whether it’s day or night? These are often not easy metrics to obtain.

We estimated the calling rate of forest elephants by visually counting individuals in forest clearings where an acoustic array allowed us to determine how many of the recorded calls were produced by the observed animals. These clearings are about the only place we can directly observe forest elephants for any length of time. But why would we expect a similar rate of elephants walking the wider forested landscape where the social context is very different? For a nearly invisible and dangerous animal this is not an easy challenge to solve. Researchers trying to obtaining call rates (and call context) for birds and primates where observer presence often alters calling behaviour might have similar problems.

One of the most exciting and enabling developments is that automatic detection algorithms are making landscape level PAM easier. Developing high-performance detectors often requires considerable expertise, but there are some programs that offer user interfaces for simple detector design (e.g. Raven, Kaleidoscope). However, the proper implementation of detection algorithms requires repeated verification of performance. No detector is perfect and performance characteristics are likely to change when moving to new environments. Users must commit considerable effort to ‘hand browsing’ sufficient samples of raw sounds to verify how many signals are missed (to verify or adjust recall rate) and monitor the ‘false positive’ rate when detection results are used directly. A frequently used ELP elephant detector had an average recall of 75-80% in the forest clearings where it was trained, but this fell to a recall of less than 60% at a new site in Cameroon (recall is the percent of actual recorded calls that the detector successfully identifies).

Protected area managers are increasingly interested in the ‘real time’ detection of events because these could trigger intervention activities. For example, the monitoring of North Atlantic right whales in shipping lanes off of Boston Massachusetts, USA, has been used for a number of years to reduce ship-strikes by alerting captains to alter course or slow ship speed. Such application requires detectors with particularly high performance (to minimize false alerts) and these often need a deep understanding of signal processing to design.

Future visions for Passive Acoustic Monitoring

By linking specific acoustic signals with specific behaviours or social interactions we can vastly increase the value of PAM. When subunits of forest elephant families re-unite after a separation, they give a unique sequence of overlapping calls. With recordings like these we can gain information about social dynamics at the scale of landscapes. Acoustically, forest elephants seem to engage in a different set of social interactions at night. By combining an acoustic array (to localise the source of calls not always audible to an observer) with thermal imaging (see video below), we are beginning to understand a formerly unknown aspect of elephant behaviour, further enriching what we can extract from PAM systems.

There are few tools available in ecology and conservation that can gather data across an extensive spatial scale simultaneously and without hands-on operation. Sound recorders and camera traps are two of these tools. In both cases perhaps the biggest challenge ahead is to design ways to move data out of difficult and remote environments automatically and affordably, allowing near-real-time monitoring. The component technology to build a system like this exists but has not yet been assembled into something affordable enough to deploy widely in rainforests where there is no cell-phone cover, and even satellite access can be dicey. Often the biggest impediment to successful conservation is a lack of knowledge about a species’ ecology. PAM might help us to jump this hurdle.

To find out more about Passive Acoustic Monitoring and Conservation read our Methods in Ecology and Evolution article ‘Acoustic monitoring for conservation in tropical forests: examples from forest elephants‘.