By Alexander Boucher, MSc Candidate at Concordia University

Have you ever listened to the grunts of deer, moose and reindeer during their rut and wondered how researchers go about studying their vocalizations? Probably not. But over the decades, a few Cervidae enthusiasts have.

In 1859, Charles Darwin noted that sexual selection is a driving force in species evolution. He stated that traits would be sexually selected for or against as part of an individual’s reproduction cycle.  In turn, it has been shown that vocalizations play an important role in these selections for a plethora of species, and many animals have evolved a menagerie of different vocalizations to attract the opposite sex. 

Beyond mating rituals, individuals additionally  use acoustic signals to convey information and regulate multiple types of interactions. For example, the emitter can share diverse information for danger avoidance, peer recognition, and social learning. Identifying and understanding the contents of these vocalizations has been a central focus of bioacoustics. 

Figure 1. Picture of an adult male reindeer from our study site in Finland wearing a collar with a GPS locator and audio recorder.

In the past, researchers studied vocalizations by conducting in-person observations and the context for which these vocalizations occurred (think of the who, what, where, when and why of these vocalizations). However, many researchers have now begun to record these vocalizations and conduct computer analyses. With the appropriate acoustic analysis software, researchers can quantify several different aspects of vocalization such as vocal range, frequency, and call rate. Although finding the vocalizations within hours of recordings, especially if researchers continuously record the animals’ environment, is challenging.

In recent years, researchers have collaborated with computer scientists to apply machine learning techniques to the field of bioacoustics. With machine learning networks, an algorithm can be trained to identify vocalizations within a series of recordings. For my research, we used convolutional neural networks to predict vocalizations within recordings; that is to say, our algorithms learned to identify grunts within images. To do so, we transformed the sounds from our recordings into spectrograms  and taught the network what a reindeer vocalization did and didn’t look like. Once we had our trained network, we began analyzing our recordings to describe the vocal activity of reindeer over several age and weight classes throughout their rut.

Figure 2. Spectrograms of sounds collected from our on-animal recorders. A) and B) represent the vocalizations of reindeer, while C) is the sound captured by the recorder as it scratched against some branches and D) is the sound of wind.

Figure 3. Picture of a collar with an audio recorder and GPS locator.

Overall, we found fine-scale differences between the different age classes. We found older, heavier males typically spend more of their time grunting.  This doesn’t mean younger males don’t grunt, but rather that their vocalization activity is opportunistic, and they grunt only when competition from older males is diminished. We also found that young and old males spend little time resting during the peak, as described by their movement patterns, as we listened back through their recordings Some days, males might spend as little as one hour at rest, but on average spend about four and a half hours at rest a day, and many of their breaks during the day are thirty minutes or less at a time. To put things into perspective, a full-grown, male reindeer can lose 30% or more of his body weight during a given rut, meaning that grunting and attracting a mate becomes a main priority. Male reindeer put a lot of time and effort into reproduction, mainly during the rutting season. 

Figure 4. Picture of a group of female reindeer with some young.

Although vocalizations of reindeer are quite intricate and may present more fine-scale differences than our study sought to understand, research must start from somewhere and with the rapid advancements of technology today, many novel approaches for studying the behaviour of animals, amongst many other disciplines, continue to be inundated with new techniques every year. As researchers, we must determine the efficacy and utility of these different and new approaches. To date,  recorders have been rarely attached to animals, and even more seldom have researchers used machine learning to predict the vocalizations of animals using these animal-borne recorders. We were the first researchers to try and map the rutting activity of a mammal during its rut using these devices and techniques. While animal-borne recorders may sound trivial initially, developing a recorder that withstands the tortures of a reindeer’s environment is no easy feat. Of the recorders we deployed, one failed after only three days, and only a couple lasted more than fifteen days (a typical rut lasts about one month). Despite these challenges, with our deployed recorders and machine learning model, we demonstrate the utility of these novel methods to further the understanding of male reindeer vocalizations during the rut.

About the author: Alex Boucher is an MSc candidate at the Weladji Lab in the Biology Department at Concordia University. Working under Dr. Robert Weladji, he is using machine learning and animal-borne acoustic recorders to study the mating activity of semi-domesticated male reindeer in Kaamanen, Finland using their vocalizations during rutting season.