2017 Volume 57 Issue 1 Pages 19-33
Acoustic monitoring of the echolocation calls of bats has recently attracted attention as a means of evaluating environmental conditions. However the ultrasound calls of bats have a large variety depending on the species’ characteristics, activities and surrounding environments. Therefore it is necessary to develop classifiers that can identify bat species by various echolocation calls precisely. In this study, we developed a bat species classifier and applied it to acoustic monitoring. A dataset of 6,348 echolocation calls consisting of 3 families, 7 genera and 11 species of Japanese bats was developed by extracting 75 dimensional feature vectors from each echolocation call with a bat acoustic analyzing tool. A classifier was developed combining the Random Forest and Support Vector Machine algorithms. The overall accuracy of classification at the genus level was 96.3%, and the species level was 94.0%. We applied the classifier to a site in Osaka, Japan and we estimated the existence of Pipistrellus abramus and Miniopterus fuliginosus successfully and Rhinolophus ferrumequinum, Myotis ikonnikovi and Myotis macrodactylus incorrectly in the field. Based on this result, we discuss the improvement of classifiers and applicability to real world situations.
