NEAFWA 2018

AUTHORS: Cathleen Balantic, Vermont Cooperative Fish and Wildlife Research Unit, University of Vermont; Therese Donovan, U.S. Geological Survey, Vermont Cooperative Fish and Wildlife Research Unit, University of Vermont; Jonathan Katz, Vermont Cooperative Fish and Wildlife Research Unit; Mark Massar, U.S. Bureau of Land Management

ABSTRACT. Audio recordings of the environment can provide long-term, landscape-scale presence-absence data to model populations of sound-producing wildlife. Automated detection algorithms allow researchers to avoid manually searching through terabytes of recordings, but often produce unacceptably high false positive rates, in which an event was detected wherein the target species did not actually vocalize. In the R package AMMonitoR, we developed a method that allows researchers to pair template-based automated detection with a suite of statistical learning algorithms trained to predict whether a detected event is a true or false positive. We introduced a novel ensemble classification approach that explicitly captures a research program’s monitoring values as they relate to classification performance. To test our method, we acquired 675 total hours of recordings in the Sonoran Desert, California between March 2016 and May 2017, and created vocalization templates for three target avian species: Eurasian Collared-Dove (Streptopelia decaocto), Gambel’s Quail (Callipepla gambelii), and Verdin (Auriparus flaviceps). We verified a subset of automated detections as true and false positives, and trained and tested five classification algorithms and four performance-weighted ensemble classifier methods. We then selected a high-performing ensemble classifier from the train/test phase to predict the class of new detections, and assessed its overall performance. For three target species, our ensemble classifier was able to identify 98% (Eurasian Collared-Dove), 85% (Gambel’s Quail), and 99% (Verdin) more false positives than the baseline detection system, and comparative positive predictive values improved from 6% to 75% (Eurasian Collared-Dove), 87% to 97% (Gambel’s Quail), and 2% to 69% (Verdin). Statistical learning approaches can thus be implemented and customized to mitigate false detections acquired within the context of automated acoustic wildlife monitoring. Furthermore, performance-weighted ensemble methods afford researchers the opportunity to employ a classification system customized to reflect a program’s species monitoring needs.