Ecological Indicators (Sep 2022)
Using timelapse cameras and machine learning to enhance acoustic monitoring of small boat sound
Abstract
Small recreational boats are an omnipresent source of sound pollution in shallow coastal habitats, which can impact the behavior and physiology of a wide array of taxa. However, effective monitoring of this stressor is currently limited by a lack of tools. The present study coupled passive acoustic monitoring (PAM) with timelapse imagery to provide a comprehensive analysis of sound pollution at two coastal sites varying in habitat structure: Goat (rocky reef) and Kawau (sandy bay) Islands. A convolutional neural network (CNN) was used to automatically count boats in each image, and the relationship between the soundscape and number of boats present was analysed using power spectral density and adaptive threshold analyses. Small boat activity was positively correlated with third octave level (TOL) root mean squared sound pressure levels (SPLRMS 63 – 5011 Hz), and this effect was frequency dependent, at both Goat (F7,9704 = 5.665, p < 0.001) and Kawau (F7,42488 = 325.33, p < 0.001) Islands. However, at Goat Island this interaction effect was driven by a significant difference between 63 Hz and all other TOLs (p < 0.05), whereas at Kawau Island the interaction effect of TOL and boat number was more variable. Furthermore, low frequency (∼50 – 300 Hz) biophony was found to influence the likelihood of boat sound being detected at Goat Island. Small boat impacts are contextual, likely due to habitat specific propagation conditions and the presence/absence of vocalising animals. As such, monitoring of sound pollution in coastal habitats requires a tailored approach which accounts for the localised nature of shallow coastal soundscapes. These findings demonstrate the potential for timelapse imagery to elucidate variability in boat sound, which may be particularly useful for remote sites which are ecologically rich, yet have no acoustic protections, such as many marine protected areas.
