Passive Acoustic Monitoring Data from Hawaiʻi Island Used in Direct Call Density Estimate Study

This Hawaiian passive acoustic monitoring (PAM) dataset consists of 17.52 hours collected using Song Meters (models 2, 4, or Mini, Wildlife Acoustics Inc., Maynard, MA) in 16-bit .wav format at a sampling rate of 44.1 kHz and default gain from five sites on Hawai‘i Island: Hakalau, Hāmākua, Mauna Kea, Mauna Loa, and Pu‘u Lā‘au. The file names contain the identifier for the automated recording device (ARU) used, as well as the date and time the recording was collected (e.g., ARU_YYYYMMDD_HHMMSS.wav). These recordings were compiled from various research projects and were annotated by members of the Listening Observatory for Hawaiian Ecosystems (LOHE Lab) at the University of Hawai‘i at Hilo. Using Raven Pro software (Cornell Lab of Ornithology, Ithaca, NY), annotators were asked to create a selection box that captured both time and frequency around every bird call they could either acoustically or visually recognize, ignoring those that were unidentifiable at a spectrogram window size of 700 points. Annotators were allowed to combine multiple consecutive calls of the same species into one bounding box label if pauses between calls were shorter than 0.5 seconds.

The majority of recordings were collected at Hakalau Forest National Wildlife Refuge on the eastern slope of Mauna Kea, totalling 11.25 hours collected in 2021 and 2022. Hakalau is one of the largest (13,240 ha) intact, disease-free, native wet forests in the Hawaiian archipelago, and as such it is widely viewed as having the most intact and stable forest bird community remaining in Hawaiʻi. Hakalau provides habitat for eleven native Hawaiian bird species (including five federally listed endangered species), as well as many introduced bird species. The next largest contribution of data came from the high-elevation dry forests of Pu‘u Lā‘au on the southwest slope of Mauna Kea with 5.2 hours of audio recorded in 2017. Pu‘u Lā‘au is within Ka‘ohe Game Management Area, a mixed management area open to the public for activities such as hiking and hunting, and a site with ongoing native vegetation restoration efforts intended to preserve and restore habitat for the few remaining native bird species that live there. The remaining recordings were collected in high-elevation open habitat on the southern slopes of Mauna Loa (0.55 hours) in 2018, similar habitat on the eastern slopes near the summit of Mauna Kea (0.17 hours) in 2019, and at a low-elevation site in Hāmākua (0.25 hours), an anthropogenically degraded habitat, in 2022. The recording locations on Mauna Loa and Mauna Kea are potential nesting sites for native endangered seabirds that build burrows in lava flow crevices. The Hāmākua site has low potential to harbor native bird species, but is densely populated by introduced bird species, and was included to assess how well our computational per-site analysis would handle absent species in an acoustically active environment. Feature embeddings were extracted from the recordings using the pre-trained Google Perch model. We then trained a linear classifier over the pre-computed embeddings using examples from 7 native bird species, and 6 common non-native bird species, with variable numbers of training samples. None of the training examples were sourced from the PAM recordings used in this study. The classifier was then run over the embedded PAM data and a logit score was generated for each 5 second segment within the dataset.