A fully automated framework for acoustic identification and localization of terrestrial wildlife at scale.

Autonomous acoustic recorders are increasingly important tools for monitoring sound-producing taxa. Coupled with advances in machine learning, they enable greater sampling effort and larger spatial and temporal scales than in-person surveys. With synchronized recorders, acoustic localization enables collection of fine-grained spatial data on animals' positions. Spatial data has important applications in abundance estimation, territory delineation, studies of movement and microhabitat use. Applications of acoustic localization have however been limited in scale, due to the technical and cost challenges of hardware and effort-intensive analysis steps.. We present a fully automated framework for identification and acoustic localization of terrestrial wildlife. Using low-cost GPS-Audiomoth recorders, we provide an open source software pipeline for detection, time-delay estimation, localization, error rejection and resolution of multiple simultaneous sound-sources. We achieve high spatial accuracy on a loudspeaker test, where 99% of broadcast calls localized were accurate to within 5 m. We demonstrate its utility for surveying birds at large spatial scales with a localization array of over 60 recorders at a forested site. Using a CNNfor automated detection, our localization framework produces spatial patterns of species observations similar to those found by in-person spot mapping surveys. Our framework enables applications of acoustic localization at large scales.