Listening to hope: Using bioacoustics to monitor Hawaiian honeycreepers during the application of Incompatible Insect Technique

Abstract

In Hawaiʻi, more than two-thirds of native bird species have gone extinct since the arrival of humans. The invasive mosquito Culex quinquefasciatus, as a vector for introduced avian disease, has contributed to approximately 50% of these extinctions, with four more species poised to disappear in as little as 10 years. To suppress C. quinquefasciatus and disease transmission, conservation partners are employing the Incompatible Insect Technique (IIT) on a landscape scale on the islands of Maui and Kaua'i. For my Master's thesis, I developed a novel technique to monitor the response of native Hawaiian honeycreepers to IIT in a relatively rapid and repeatable design. I deployed 54 autonomous recording units (ARUs) to record the sounds of honeycreepers during the 2024 and 2025 breeding seasons in The Nature Conservancy’s Waikamoi Preserve and bordering private lands owned by Mahi Pono on Maui. I placed ARUs 150m apart within and above the IIT treatment area (Treatment), as well as in an area to the west that did not receive IIT (Control), in a Control-Impact (CI) design. To process these acoustic data, I used the machine learning classifier Perch and a new analytical method co-developed by the UH Hilo Listening Observatory for Hawaiian Ecosystems (LOHE) lab and Google Research. Perch predicts whether species are vocalizing within acoustic data, which I used to calculate the proportion of detection windows containing the target species vocalizations, or call densities, through manual validation. I compared the maximum likelihood estimates of call densities for six native species within the 2024 and 2025 breeding seasons across our CI design. I also used these data to model the relationships between call density, forest structure, and elevation to understand the use of habitat across the study site. Models were fitted using generalized linear mixed models (GLMM) and a generalized additive modeling framework (GAM). I found increased call densities in 2025 across both the Control and Treatment sites (p < 0.05) for ʻapapane (Himatione sanguinea), Hawaiʻi ʻamakihi (Chlorodrepanis virens wilsoni), and ʻiʻiwi (Drepanis coccinea). ʻIʻiwi also showed a significant negative effect in the 2025*Treatment interaction term, showing that the increase in the Control site for ʻiʻiwi was greater than in the Treatment site (β = -0.484, p < 0.001). The elevation-smoothed GAMs revealed two different patterns: 'akohekohe (Palmeria dolei), Hawai'i 'amakihi, and 'i'iwi had curvilinear relationships with elevation, while 'apapane, kiwikiu (Pseudonestor xanthophrys), and Maui 'alauahio (Paroreomyza montana) had increasing relationships with elevation for all habitat types. I conclude that there has not been sufficient time to detect a change in call densities since IIT implementation, but my study design is sensitive enough to detect changes between years and can reveal important patterns for honeycreeper conservation.