Development of new bioacoustics tools to monitor and better understand populations of two critically endangered bird species on Maui

Abstract

Bioacoustic monitoring using automated recording units (ARUs) offers an alternative method to monitor vocalizing wildlife compared to traditional methods. Some benefits of using ARUs include flexible scheduling, infinite replayability, and the ability to collect large datasets with minimal effort. Still, a drawback is the time-consuming nature of reviewing recordings for target species. Automatic detection algorithms can expedite ARU data analysis; however, automated detection algorithms are not well-developed for Hawaiian forest bird species. Here, I focused on using automated tools to improve monitoring protocols for two critically endangered Hawaiian forest birds: the kiwikiu (Pseudonestor xanthophrys) and the ʻākohekohe (Palmeria dolei). My first objective was to gather training data for the automatic detection algorithm (BirdNET) and use it to examine the differences in vocalization characteristics in my target species in three distinct areas of their population. I deployed ARUs at The Nature Conservancy’s Waikamoi Preserve, Manawainui, and upper Kīpahulu Valley in Haleakalā National Park. For kiwikiu, I found significant differences in the mean length and frequency variability in songs between locations. For ʻākohekohe, I found a significant difference in the mean length of their “squirtle” calls. However, I did not find differences in repertoire sizes or diversity among locations in either species. For kiwikiu songs, I observed that shared syllables occurred most frequently in the same locations, suggesting variation in songs between the sub-populations of kiwikiu. ʻĀkohekohe vocalizations may also vary between the sub-populations. These results demonstrate that some vocalizations vary by location and may indicate the presence of dialects in these species. My second objective was to train the BirdNET algorithm to detect kiwikiu and ʻākohekohe and to test whether the newly trained algorithm could accurately detect kiwikiu and ʻākohekohe in soundscape recordings. This testing occurred in two phases, an efficacy phase, where the identifications BirdNET produced were compared to previously annotated files, and an expansion phase, where BirdNET identified vocalizations in unannotated files. For BirdNET identifications over a 0.50 confidence threshold, BirdNET accurately identified ʻākohekohe 65.8% of the time, and kiwikiu 27.4% of the time. My results demonstrate the potential for using BirdNET to analyze recording data for kiwikiu and ʻākohekohe and improve the tools available to conduct future bioacoustic research for these species.