•  
  •  
 

CAN YOU TRUST THE RESULTS OF YOUR AUTOMATED BAT CALL CLASSIFIER?

Abstract

Interest in bats and the vital ecosystem services they provide has increased substantially in recent years as the myriad threats bats face have drastically reduced many populations. This interest has resulted in increased survey efforts, many of which are acoustic in nature to minimize the detrimental effects surveys can have on already stressed populations. While the ability to assign species identity to bat calls has improved in recent decades, the process remains challenging and often lacks objectivity and repeatability among researchers. To address these issues and improve the speed of the identification process, several automated bat-call classifying software programs were developed in recent years and are becoming widely used. Unfortunately, these programs may not perform well in all areas depending on the quality of the recorded calls, the composition of the resident bat community, and a multitude of other factors. Our objective for this project was to determine if two classification programs, which are currently approved by the Fish and Wildlife Service as automated bat-call classifiers, produced similar results with the same dataset. To accomplish this objective, we collected bat calls with two Anabat Swift detectors in a forest landscape between July – August 2018 in Lumpkin co., GA, and classified the dataset using both classification programs. The dataset contained 148,563 total files and approximately 800 identifiable bat-call files. In general, we found little agreement between the two classification programs, consequently, we recommend caution when relying solely on software programs to determine presence/absence of species in a particular area. When species ID is not critical, using acoustic data and automated software to determine general indices of activity and indicate possible residents may be more appropriate than attempting to ID each call to species and conclusively determine residents because of the high degree of uncertainty that remains when attempting to ID calls.

This document is currently not available here.

Share

COinS