An efficient acoustic classifier for high‐priority avian species in the southern Great Plains using convolutional neural networks

Abstract

AbstractPassive acoustic monitoring is a valuable ecological and conservation tool that allows researchers to collect data from vocal species across large geographic areas and temporal spans. Grassland bird populations, many of which are indicators of ecosystem health, have experienced precipitous declines over the past several decades. Acoustic monitoring of grassland bird populations provides opportunities to monitor declines and focus conservation practices, yet the ability to identify species efficiently and accurately from acoustic data is challenging. Therefore, development of automated classifiers such as convolutional neural networks (CNNs) are at the forefront of streamlining detection and identification of individual species. Here, we present a CNN classifier for 5 key grassland bird species across southcentral Oklahoma, a part of the southern Great Plains: northern bobwhite (Colinus virginianus), painted bunting (Passerina ciris), dickcissel (Spiza americana), eastern meadowlark (Sturnella magna), and Bell's vireo (Vireo bellii). We compiled a high‐quality training dataset consisting of 6,933 calls, built semiautonomously using template matching that can be expanded easily to any bird species of interest. Our trained multilabel CNN achieved a high level of classification accuracy (≥98%) for the 5 species using the library of test calls and field recordings played using a programmable game caller. The ability to conduct acoustic wildlife surveys across large spatial extents will allow for more efficient monitoring of wildlife to determine key population parameters and trends and effects of biotic and abiotic factors (e.g., vegetation, disturbance, weather) on these key species.