Downloadable ContentDownload PDF
Automated acoustic identification of nine bat species of the eastern United States
Increased sophistication in methodology of echolocation monitoring has enabled the pursuit of a widened array of research and management objectives for bats. I developed and tested an automated program for detecting, measuring and classifying species of bat echolocation calls. I analyzed 584 echolocation sequences from nine bat species of the eastern United States (Lasiurus borealis, Myotis austroriparius, M. grisescens, M. leibii, M. lucifugus, M. septentrionalis, M. sodalis, Nycticieus humeralis, and Pipistrellus subflavus). A base set of 11 variables was measured automatically and manually, and an additional 28 variables were measured automatically as a part of the full variable set. Species were classified using Discriminant Function Analysis (DFA) using each variable set and measurement method with both a single DFA and a hierarchical set of DFAs that classified genus level before species level. Classification rates were also summarized for a subset of the sequences with Discriminant probabilities > 0.75. Overall classification rates were 67.6% for manually measured base variable set, 63.0% and 68.3% for automatically measured base and full variable sets, respectively, and 84.4% for the 62.5% of calls with Discriminant probabilities > 0.75 using the automatically measured full variable set. Hierarchical classification improved overall classification rates by < 2% for base variable set classifications and 6.8% and 6.9% for the automatically-measured full variable set with and without the Discriminant probability cutoff, respectively. Hierarchical classification worked synergistically with a large variable set to increase classification rates. Six of nine species had overall classifications of 93.3-100% with the best classifier. Also, six of nine species had higher classification rates than that achieved in previous studies. Automated call processing is a fully objective and repeatable method requiring less cost, time, and user expertise than manual methods, especially for large datasets. Further sophistication will likely continue to offer improvements to echolocation monitoring methodology.