Evaluation of Acoustic Recording Devices to Survey Northern Bobwhite Populations

Abstract

Reliable methods for monitoring wildlife populations are paramount to effective conservation and management. There are a variety of available techniques that vary in cost and feasibility. We field‐tested a commercially available acoustic recording device (ARD) in July and August of 2018 with the goal of standardizing the process of surveying northern bobwhite (Colinus virginianus) populations using acoustic data. We projected bobwhite covey calls from programmable speakers, and analyzed recordings manually and with automatic recognition software. We manually detected 48% of projected covey calls with the furthest detection at 241 m. We developed an automatic classifier using Kaleidoscope Pro (Wildlife Acoustics Inc.), which detected 31% of the manually detected covey calls. The automatic classifier was ineffective, so we used manual analysis to acquire measurements for call strength and background noise at each visible covey call spectrogram signature. Detection probability declined as distance from the ARD increased and background noise increased. However, in our study system, there was no effect of vegetation structure on detection probability. Our results provide a blueprint for evaluating ARDs for application to wildlife monitoring. We believe ARDs can be used to evaluate northern bobwhite occupancy at fine scales and may potentially be useful in estimating autumn densities for setting harvest regulations. We encourage the evaluation of other software recognition programs and the relationship between covey call detection and landscape structure. © 2020 The Wildlife Society.

Broad‐scale applications of the Raspberry Pi: A review and guide for biologists

“…Finally, Wilhite et al. (2020) used a Raspberry Pi, an amplifier, and outdoor speakers for the playback of bird calls in the field to help assess how acoustic recording devices can help quantify bird populations.…”

The development of a convolutional neural network for the automatic detection of Northern Bobwhite Colinus virginianus covey calls

“…When reviewing the 600 training clips, the majority of the missed labels by both the technicians and the CNN were the result of either high levels of background noise masking the calls, coveys calling from far distances, multiple coveys calling at once, or a high density of calls within a short window of time. However, the ability of the CNN to detect more of these challenging calls presents better opportunities to obtain more accurate estimates of calling rates and bird abundance or density, all of which are key for effective bobwhite conservation (Wellendorf et al., 2004; Wilhite et al., 2020).…”

“…Additionally, other studies found that simultaneous calls from multiple target species can mask each other and be missed (Ruff et al, 2020b). To our knowledge, there are currently no CNNs developed purely for bobwhite, and our previous research (Wilhite et al, 2020) where we processed bobwhite acoustic data using automatic recognition software (Raven Pro and Kaleidoscope), found that despite developing advanced clustering classifiers (see Wilhite et al, 2020 for more information), these programs were ineffective for automatic covey call identification. Therefore, although several software programs and multispecies CNNs do exist for the detection of bobwhite covey calls, the results from these suggest that focal research using ARU monitoring of bobwhite could greatly benefit from the development of a single species bobwhite CNN.…”

Automated recognition of ruffed grouse drumming in field recordings

“…In general, open‐source automated recognition methods for automated acoustic monitoring are not currently available for most taxa, and face many challenges such as the lack of labeled audio recordings and lack of generalizability (Knight et al 2017, Stowell 2022). As a result, most ecological studies still rely on manual review of a limited subset of ARU recordings to extract species presence information (Venier et al 2012, Sidie‐Slettedahl et al 2015, Vold et al 2017, Sugai et al 2019, Wilhite et al 2020). The use of automated recognition methods to process large quantities of ARU data can greatly increase the number of detections relative to annotation effort even when recognizer precision is poor (Shonfield 2018, Knight et al 2022).…”