“Machine learning Megapodes: Our journey from shaved yaks to analyzed acoustic tracks ”

Abstract:The Commonwealth of the Mariana Islands (CNMI) is a small archipelago located in the Western Pacific. Like other island systems, the Marianas are home to several endemic and rare species. Our work primarily concerned the development of an avian acoustic monitoring system that will be used on Tinian Island to detect for the presence of the Micronesian megapode (Megapodius laperouse laperouse), a federally endangered ground bird. Tinian, a small island and home to a growing US military base, was likely occupied historically by megapodes, however, megapodes have not been detected there since 2020. Bioacoustic monitoring offers promising utility for the large-scale monitoring of rare species, but presents technical challenges for ecologists who lack robust programming backgrounds to implement post-processing algorithms to review large datasets of acoustic recordings. We had two specific aims while working with the Accelerator program: 1) to understand and produce a deep learning classifier capable of detecting for the presence of megapodes with a minimal false negative rate and 2) to learn best practices for and develop better data management and processing pipelines. At the onset of the program, we had already acquired and labeled over 800 clips representing the four different call types produced by megapodes, in addition to thousands of other clips representing the common biotic and abiotic sounds of Tinian. Through our collaboration with Spencer Wood and Vaughn Iverson, we successfully ground tested the development of megapode classifiers using OpenSoundscape, a Python utility library, and BirdNET, an open-source AI-powered bioacoustics analyzer. Our ability to successfully utilize these programs, however, required learning a number of support programs including Git, Docker, Linux, UW Hyak, and Python. Here we will discuss our experience navigating the growing field of AI-powered bioacoustics monitoring from the perspective of wildlife ecologists.

Speaker Bio:Kaeli Swift is a research scientist with the University of Washington. For the past 5 years she has lived and worked on the remote island of Tinian, where she studies the ecology of endemic forest birds. She received her PhD from the University of Washington where she studied crow funerals.