Confounding effects of environment-dependent propagation on automated classification of cetaceans

Abstract

Significant effort has been made over the last few decades to develop automated passive acoustic monitoring (PAM) systems capable of classifying cetaceans at the species level. The utility of such systems depends on their ability to operate across a wide range of ocean acoustic environments; however, anecdotal evidence suggests that site-specific acoustic propagation characteristics impact the performance of PAM systems. This is because properties of the ocean acoustic environment can be markedly different between regions and seasons in which PAM is used to observe cetaceans. Variability in propagation characteristics leads to differences in how each cetacean vocalization is distorted as it propagates along the source-receiver path. Unless these differences are accounted for, the acoustic environment will bias estimates of a PAM system’s performance. The impact of environment-dependent propagation on an aural classifier was assessed using a pulse propagation model. Simulations were conducted that showed the sensitivity of classifier performance to sound speed profile type. Furthermore, it was found that classifier performance was range-dependent, largely due to changes in signal-to-noise ratio (SNR); however, in some environments only 60% of the performance reduction was attributed to decreasing SNR, indicating that signal distortion—resulting from environment-dependent propagation—had a large impact on performance.