Using intrinsic and contextual information associated with automated signal detections to improve call recognizer performance: A case study using the cryptic and critically endangered Night Parrot Pezoporus occidentalis. Issue 11 (25th September 2020)

• Record Type: Journal Article
• Authors: Leseberg, Nicholas P.; Venables, William N.; Murphy, Stephen A.; Watson, James E. M.
• Editors: Zamora‐Gutierrez, Veronica
• Is Part Of: Methods in ecology and evolution. Volume 11:Issue 11(2020)
• Page Start: 1520
• Page End: 1530
• Publication Date: 2020-09-25
• Subjects: acoustic monitoring -- bioacoustics -- call recognizer -- night parrot -- rare species; Ecology -- Periodicals; Evolution -- Periodicals; 577
• DOI: 10.1111/2041-210X.13475
• Languages: English
• ISSNs: 2041-210X

Abstract

Abstract: Rapid expansion in the collection of large acoustic datasets to answer ecological questions has generated a parallel requirement for techniques that streamline analysis of these datasets. In many cases, automated signal recognition algorithms, often termed 'call recognizers', are the only feasible option for doing this. To date, most research has focused on what types of recognizers perform best, and how to train these recognizers to optimize performance. We demonstrate that once recognizer construction is complete and the data processed, further improvements are possible using intrinsic and contextual information associated with each detection. We initially construct a call recognizer for the Night Parrot Pezoporus occidentalis using the r package monitoR, and scan a test dataset. We then examine a number of intrinsic variables associated with each detection generated by the recognizer, and several contextual variables associated with the species' environment and ecology, to determine if they might help predict whether a given detection is a true positive (target signal) or false positive (non‐target signal). We test several logistic regression models incorporating different combinations of intrinsic and contextual variables, selecting the best‐performing model for application. We train the model, using it to calculate the probability each detection is a true or false positive. Substituting this model‐derived probability for raw recognizer score improved the recognizer's performance, reducing the number of detections requiring proofing by 60% to achieve a recall of 90%, and by 76% to achieve a recall of 75%. This technique is applicable to any recognizer output, regardless of the underlying algorithm. Application requires an understanding of how the recognizer algorithm determines matches, and knowledge of a species' ecology and environment. Because advanced programming skills and expertise are not required to apply this technique, it will be particularly relevant to field ecologists for whom building and operating call recognizers is an element of their research toolbox, but not necessarily a focus.