Toward an unsupervised and automated analysis framework for large acoustic datasets to assess animal biodiversity and ecosystem health

Passive acoustic monitoring is a promising and non-invasive method to assess the biodiversity and potentially health of terrestrial and marine ecosystems. Over the last decade, various methods have been proposed to extract information on the animal biodiversity primarily based on acoustics indices. Several recent studies have shown that ecological relevance and effectiveness of these indices remain uncertain. We propose a new, multi-step method to estimate animal biodiversity from acoustic datasets by applying an unsupervised detection and classification technique. Our semi-automated framework extracts every acoustic event with a pre-defined signal-to-noise ratio. In a second step, the detected events are grouped into classes based on the similarity of acoustic features. The number of resultant classes are linked to animal biodiversity in an area by applying a transfer function, which is established using manually/expert reviewed class labels. Our framework provides diel and seasonal changes in the overall number of sound classes as well as number of acoustic events in each class. We will demonstrate its performance by application to three datasets collected in the Chukchi Sea, Alaska, Sapsucker Woods Sanctuary, Ithaca, NY, and Abel Tasman National Park, New Zealand.