Quantitative steps for refining passive acoustic monitoring detector libraries: A southern right whale (Eubalaena australis) case study.

Passive acoustic monitoring is widely used to detect sound-producing species, but the massive datasets generated require tools and detectors for efficient processing. These detectors are most effective when their target templates or call types accurately represent the variation within and between these call types. However, the process of creating a library of call types for detectors has not been standardized and is often subject to qualitative classification that does not translate to the quantitative detector. This article presents a case study creating and testing a quantitative call library for southern right whales (Eubalaena australis) in a Brazilian calving ground to be used with the low-frequency detection classification system (LFDCS). Call attributes were extracted from their LFDCS pitch tracks, dimensionally reduced via uniform manifold approximation projection, and grouped by a K-means algorithm into updated call types. A call library was created using exemplars of each call type. Using the updated call library increased the true detection rate from 58.5% to 83.3%, demonstrating the adaptability and efficiency of the detector following call library adjustment. This project aims to assist future acoustic studies by developing a protocol for quantifying call libraries for the acoustic detection of a variety of species.