Accounting for both automated recording unit detection space and signal recognition performance in acoustic surveys: A protocol applied to the cryptic and critically endangered Night Parrot (Pezoporus occidentalis)

Abstract

Research into the suitability of autonomous recording units (ARUs) when surveying for vocal species is increasing. Simultaneously, there has been extensive research into methods for efficiently extracting signals of interest from the acoustic data sets that accrue from the deployment of ARUs. For some species, bioacoustic monitoring supported by computerised signal detection offers the only effective and efficient method for widespread survey. In these circumstances, the detection space of both the ARU and the performance of the signal detection process must be considered concurrently, but typically, these two elements have been considered separately. Here, using the Night Parrot (Pezoporus occidentalis) as a case study, we consider both ARU detection space and the signal detection process to develop a robust and repeatable survey protocol for the species. After developing a call recogniser for the Night Parrot, we test its performance on a data set of Night Parrot calls given at a known distance from an array of ARUs. Having established a relationship between ARU type, recogniser performance and distance, we determine the sampling radius of an ARU for a given recogniser score cut-off, and the associated probability of detecting a Night Parrot that calls within that sampling radius. Using these data, we outline how to develop a robust and repeatable survey protocol for the Night Parrot, with a defined probability of detection. This protocol could be adapted for other scenarios where deployment of ARUs is necessary to determine a species' status and distribution.

Passive acoustic surveys reveal interactions between frugivorous birds and fruiting trees on a large forest dynamics plot

“…Thus, ecoacoustics on long-term vegetation plots has the potential to provide essential information about plant-animal interactions, a subject not only important theoretically to ecology, but also to the predictions of how biodiversity responds to global change. This adds to the developing literature about how ecoacoustics broadly, and the use of ARUs and vocal identification software more specifically, can expand scientists' ability to monitor and protect biodiversity (Burivalova et al, 2019;Leseberg et al, 2022; Shonfield & Bayne, 2017; Teixeira et al, 2019;Zwart et al, 2014). Below we talk about parts of our analysis that may have confounded our ability to detect relationships between animals and plants, and then about possible future research directions for ecoacoustics that could integrate with plot data.…”

What factors influence the rediscovery of lost tetrapod species?

“…Molecular analysis has detected the presence of Archey's frog ( Leiopelma archeyi ) (CR) in the diet of ship rats ( Rattus rattus ) (Egeter et al., 2019). Audio recordings have been used to develop a survey protocol for a nocturnal bird species (the night parrot, Pezoporus occidentalis [CR]) (Leseberg et al., 2022). Indigenous ecological knowledge has been used to improve the understanding of mammal population declines in Australia (Ziembicki et al., 2013).…”

Automated recognition of ruffed grouse drumming in field recordings

“…Measuring the area surveyed by an ARU is important for calibrating results of monitoring between ARUs and human surveyors (Van Wilgenburg et al 2017, Yip et al 2017, Darras et al 2018) and can aid in calculating species distribution (Leseberg et al 2022) or density if site-level heterogeneity is modeled (Darras et al 2016). Future work should experimentally quantify the sampling area covered by ARU surveys and field-based human point counts by measuring the effective detection radius of ruffed grouse drumming for each detection method (Darras et al 2018), potentially also modeling detection as a function of covariates.…”

“…We note, however, that both signal processing and deep learning approaches are likely to decrease in accuracy or fail completely when the data analyzed is substantially different from the data used for training and validation. Finally, deep learning approaches can provide continuous outputs rather than discretized or binary detections, which could be useful for interpreting and statistically analyzing the outputs of a model (Knight et al 2017, Rhinehart et al 2020, Leseberg et al 2022). Alternatives to supervised deep learning and signal processing include supervised shallow learning (relying on hand‐crafted features; Priyadarshani et al 2018) and unsupervised learning (Stowell and Plumbley 2014) which is often used in tandem with supervised learning.…”