The Influence of Input Representation and Major Soundscape Drivers on Acoustic Indices in the West Coast of Scotland
Hydrophone arrays such as the COMPASS array off the West coast of Scotland are now capable of multi-year deployments, with acoustic data providing information across a variety of spatial, temporal and ecological scales. Extracting information from massive data sets across all these scales remains a key challenge, with Deep Learning not always applicable. Acoustic indices are statistical and physical measures of acoustic data and have been applied to questions varying from event and species detection to summarizing soundscapes and identifying habitat health; but have yet to be applied extensively in the marine habitat and lack standardized methodologies. We use multi-label data from the COMPASS array to investigate changes in 4 common acoustic indices when applied to major soundscape drivers in two scenarios. The first considers delphinid whistles coincident with varying levels of shipping noise; the second considers coincident and isolated examples of similar sonotypes labelled as burst pulses, delphinid clicks and broadband clicks. Input representations to the indices are varied in Fourier Transform (FT) size and between linear and mel frequency scales. Results suggest input representation to acoustic indices is significant but signal dependent, with distributions between index values of different classes dependent on both the frequency scale and FT size used. Index distributions with minimal overlap in each scenario are used to construct False Color Spectrograms with indices as inputs to the RGB channel to demonstrate the potential utility of these indices in discriminating sound sources. These are contrasted to poorly chosen index and input representation combinations to highlight the potential importance of input representation in index performance.