Bat species identification from zero crossing and full spectrum echolocation calls using hidden Markov models, fisher scores, unsupervised clustering and balanced winnow pairwise classifiers

Abstract

A new classification technique for the identification of bats to species from their echolocation calls is presented. Data is compiled and split in half for training and testing classifiers including 8, 782 recording files (bat passes) with 223, 123 candidate calls (pulses or extraneous noise) representing 17 species of bats found in North America including 10 species of Myotis. Some files are of high quality consisting of hand-selected search phase calls of tagged free flying bats while others are from a variety of field conditions including both active (attended) and passive (unattended) recordings made with a variety of zero crossing and full spectrum recording equipment from multiple vendors. Average correct classification rates are 88.8% with an average of 39.9% of all files identified to species. Most importantly, classifiers for two species of U.S. endangered bats, Myotis sodalis and Myotis grisescens have a positive predictive value of 100% and 98.6% respectively and a true positive rate of 67.4% and 93.8% respectively suggesting that the classifiers may be well suited to the accurate detection of these endangered bats. © 2013 Acoustical Society of America.