Improving automated detection of frog calls in noisy urban habitats using narrow-banded recognizers

Tracking behavioral and demographic changes of anuran populations in urban landscapes presents difficulties due to the high amount of noise interference from anthropogenic sources. In this study, we used Song Scope software to build narrow-banded recognizers that only cover a limited portion of the full spectral range of a call and tested if these recognizers can improve automated call-detection capabilities at noisy sites. We built recognizers for two species with naturally broad-spectrum calls, the Green Frog (Lithobates clamitans) and American Bullfrog (L. catesbeianus) and tested them at five noisy ponds in the suburbs of Chicago, Illinois, USA. Narrow-banded recognizers had greater percentages of true positives compared to full-spectrum recognizers. Classification indices used to assess call recognition efficacy showed that narrow-banded recognizers were more effective at all sites for the Green Frog, and at two sites for the American Bullfrog. High-frequency recognizers had 13% fewer errors caused by anthropogenic noise (P < 0.010) than other recognizers. Finally, for every recognizer, true positives standardized by the maximum daily value was highly and significantly correlated with the number of calls identified manually, indicating that automated detection data is an accurate proxy for the actual number of calls at noisy sites. For acoustic taxa, we recommend that scientists consider identifying broad-spectrum calls using narrow-banded recognizers to reduce detection problems associated with noise interference between anthropogenic noises and biotic acoustic signals.