Comparison of methods for automated identification of wild turkey gobbles. (14th February 2022)

• Record Type: Journal Article
• Authors: Wightman, Patrick H.; Henrichs, Darren W.; Collier, Bret A.; Chamberlain, Michael J.
• Is Part Of: Wildlife Society bulletin. Volume 46:Number 1(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-02-14
• Subjects: acoustic monitoring -- autonomous call recognition -- convolutional neural network -- gobbling -- Meleagris gallopavo -- Raven Pro -- wild turkey; Wildlife management -- Periodicals; Wildlife conservation -- Periodicals; 333.9540973
• DOI: 10.1002/wsb.1246
• Languages: English
• ISSNs: 0091-7648

Abstract

Abstract: Autonomous recording units (ARUs) allow for collection of extensive acoustic data sets, while reducing costs and time associated with traditional surveys used to determine gobbling chronology of male wild turkeys ( Meleagris gallopavo ). A challenge with ARUs is efficiently locating and identifying calls of interest, so autonomous call recognition (ACR) software such as Raven Pro have traditionally been used to identify wild turkey gobbles. However, ACR software often produces high false positive detections, requiring substantive time to verify selections as gobbles. We used ARUs across 3 study sites in the southeastern United States to collect 107, 580 hours of ambient sound. We developed a convolutional neural network (CNN) to autonomously identify wild turkey gobbles and compared results of our CNN to results gathered using the commercially available program Raven Pro. After processing of ambient sound, the CNN detected 15, 793 more gobbles than Raven Pro, and did so with 5, 716, 718 fewer selections. Collectively, our CNN improved precision from 0.01 to 0.32 relative to Raven Pro, while decreasing the time required for validation from 4, 452 hours to 219. We found precision of our CNN varied across ARUs primarily due to differences in occurrence of ambient sounds similar to gobbles. Thus, we recommend that additional site‐specific training data should be considered when developing CNNs. Our results suggest that researchers interested in describing gobbling activity by male wild turkeys should consider developing and applying CNNs for automated call recognition. Abstract : Our developed Convolutional Neural Network (CNN) used to autonomously identify wild turkey gobbles detected 15, 793 more gobbles than the commercially available program Raven Pro, and did so with 5, 716, 718 fewer selections. The CNN improved precision from 0.01 to 0.32 relative to Raven Pro, while decreasing the time required for validation from 4, 452 hours to 219. Our results suggest that researchers interested in describing gobbling activity by male wild turkeys should consider developing and applying CNNs for automated call recognition.