Modelling reindeer rut activity using on-animal acoustic recorders and machine learning
Researchers have been using sound to study the biology of wildlife to understand their ecology and behaviour for decades. By gathering audio from free-ranging species using on-animal recorders, their vocalizations can be used to describe their behaviour and ecology through signal processing. Unfortu...
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2023
|
| Subjects: | |
| Online Access: | https://spectrum.library.concordia.ca/id/eprint/992735/ https://spectrum.library.concordia.ca/id/eprint/992735/2/Boucher_MSc_F2023.pdf |
| Summary: | Researchers have been using sound to study the biology of wildlife to understand their ecology and behaviour for decades. By gathering audio from free-ranging species using on-animal recorders, their vocalizations can be used to describe their behaviour and ecology through signal processing. Unfortunately, processing hours of recordings is incredibly time-consuming. By applying machine learning to audio recordings, researchers have used neural networks to decrease the processing time of acoustic data. However, until now, most of this research has focused on analyzing the data of stationary recorders. To show the utility of on-animal recorders in combination with machine learning, we recorded the vocalizations of reindeer (Rangifer tarandus) during their rut at the Kutuharju research station in Kaamanen, Finland. We used vocalizations as an activity index to describe the rut activity of male reindeer. In 2019 and 2020, we placed recorders around the necks of seven reindeer during their rut. We trained convolutional neural networks to identify reindeer grunts, which were then used to classify their vocalizations. Of the networks’ vocalization classifications, around 95% of them were correct. With such high metrics, we could reliably explore the males' activity patterns using a neural network. We then analyzed the reindeers’ vocalization using generalized additive models. The patterns suggested heavier, older males vocalized more than lighter, younger males and, overall, were more active during the day than night. Overall, on-animal acoustic recorders, in tandem with machine learning, proved to be effective tools, and with more attention, they could prove valuable tools for other researchers. |
|---|