Embedded Files
Research Group
Research Group
Department of Anthropology, Miami University
Department of Anthropology, Miami University
Ally Smith*, Ashley Sideri*, Kelsey Ellis
(*undergraduate presenters)Background
Background
Direct observation of early primate vocalizations can prove challenging at times, especially when species or individuals are difficult to locate, are unhabituated, or have the tendency to alter their behavior in human presence.
Most daily follows of primates begin at sunrise and end at sunset. However, in some taxa early morning vocalizations often occur prior to sunrise. Therefore, observers may inadvertently introduce bias into the data by failing to make contact before sunrise and miss some of the earliest vocalizations produced by these taxa.
One way to mitigate these challenges is for researchers to incorporate the use of passive acoustic monitoring into their data collection protocol.
However, this method also yields massive quantities of data that then need to be processed, either manually or through the application of machine learning.
Objective
Objective
Investigate the efficiency of utilizing BirdNet to identify and classify early morning vocalizations across multiple primate species collected from passive acoustic monitoring (PAM). Then, compare the timing of vocalizations captured by observers versus PAMs.
Methods
Methods
Study Site: Tiputini Biodiversity Station, Ecuador (0° 38’ N, 76° 9’ W)
Data Collection: PAMs: Song Meter SM2s recorded audio from 5:45 am to 8:10 am in January, February, and March 2013 through 2017. The recorders collected 10-minute recordings, with 5-minute breaks in between, creating a collection of ten recordings per device for each morning of deployment. The time of the first recorded early morning vocalization for each species on each day of data collection was noted. Direct observation: We identified the first recorded vocalizations (<2 hrs from sunrise) observed for target species each day entered into our long-term database. These data were recorded by observers from 2007 to 2018 primarily through focal follows for Ateles, Lagothrix, Plecturocebus, and opportunistically for Alouatta.
Analyses: To test the efficacy and efficiency of incorporating PAMs into our protocol, we used BirdNet to locate and identify species specific calls, then compared the times of primate vocalizations recorded by long-term observers to those picked up by the Song Meters deployed across TBS.
Results
Results
Creating custom classifiers and training BirdNet proved difficult for primate vocalizations, especially when using high quality species-specific recordings from outside the PAM dataset. At this time, Alouatta and Plecturocebus are the only species that have been completed.
The timing of first recorded vocalizations differed between observer collected data and the PAMs for Alouatta and Plecturocebus, with PAMs collecting vocalizations earlier than those recorded by observers, despite PAMs having truncated data for the earliest morning hours. However, these differences were only significant for Alouatta.
Table 1. Results of Wilcoxon Rank Sum Test comparing the median time of first morning vocalizations relative to sunrise (in mins) between observer recorded data and those collected from PAMs across several primate species at the Tiputini Biodiversity Station, Ecuador.
Discussion
Discussion
Passive acoustic monitors can be useful in complementing observer recorded data. However, the large amount of data produced can be time consuming to process.
BirdNet was developed to detect and classify the calls of North American birds using neural network machine learning.
Despite building our own classifiers, including a training set for noise, BirdNET had difficulty identifying primate vocalizations that were graded, emitted during periods of high ambient noise (e.g., insects, rainfall, etc.), or when two primate species vocalized simultaneously.
Alouatta was found to call 12 minutes earlier from data collected by PAMs than from observer recorded data, and Plecturocebus called 3.5 mins earlier.
The SongMeters were originally deployed to capture the morning chorus of birds, with avian activity often starting at or near sunrise.
We believe that if the SongMeters came on earlier closer to nautical twilight (i.e., 5:15 or 5:30 AM) we would have captured even earlier calls from Alouatta.
Our data highlight that PAMs can be useful to monitor species outside of human presence, however, they too should be programmed with species-specific questions in mind.
Future Research
Future Research
Continue this study by validating the first morning vocalizations of the other two target species - spider monkeys and woolly monkeys.
If successful, we hope to use BirdNet, or other ML applications in the future, to help detect primate individuals/groups and their activity patterns in the absence of observers from multiple PAMs deployed across TBS.
References
References
van Kuijk S. M. O’Brien S., Clink D. J., Blake J. G., & Di Fiore, A. (2023). Automated detection and detection range of primate duets: A case study of the red titi monkey (Plecturocebus discolor) using passive acoustic monitoring. Frontiers in Ecology and Evolution, 11. https://doi.org/10.3389/fevo.2023.1173722
Kahl S., Wood C.M., Eibl M., Klinck H.(2021). BirdNET: A deep learning solution for avian diversity monitoring. Ecological Informatics, 61.https://doi.org/10.1016/j.ecoinf.2021.101236.
Acknowledgments
Acknowledgments
Thanks to the Tiputini Biodiversity Station Management and Staff, the Ecuadorian government for permitting us to work in the Yasuní biosphere, and the many researchers, assistants, students, and sponsors that made this work possible. Special thanks to Silvy van Kujk (UT-Austin) who helped validate titi and howler monkey vocalizations, Jon Blake (University of Florida) for lending his SongMeter recordings to us, and Anthony Di Fiore (UT-Austin) for mentorship and guidance. Photos by Tim Laman and Kelsey Ellis.
Page updated
Report abuse