A simple categorization method for frequency notched echolocation

The effectiveness and power of passive acoustic monitoring (PAM) of wildlife in ecosystems depends on our ability to discern between species' vocal repertoires. For dolphin echolocation clicks there is heavy overlap within the bioacoustic niche, which presents challenges in signal processing. Furthermore, because echolocation is a highly directional signal, it is imperative to identify a robust categorizing characteristic that is consistent across different methods, animal position, and behavior. We applied a computationally simple method to identify instances of frequency peaks and notches within the click bandwidth; a spectral characteristic associated with white-beaked dolphins. In studies with visual species confirmation 75–90 % of events contained spectral banding in white-beaked dolphin echolocation and burst pulses. A pilot study between May–August 2021 placed 4 autonomous acoustic recorders in the Skagerrak and was reviewed for white-beaked dolphin using freeware PAMGuard and R. A decision tree process automatically reviewed the PAMGuard output to identify instances of possible dolphin activity. Those timeframes were manually audited to find true white-beaked dolphin events (total 23). We split data into testing and training datasets and used a nonlinear least squares model to fit a spectral template within the bandwidth where spectral peaks and notches are most prevalent. We tested our modelled template against the mean spectrum of every event in the testing dataset, and against noise timeframes which contained no white-beaked dolphin signals as a control. Using a Pearson Correlation test, we found that 87 % of the true white-beaked dolphin events and 3 % of the noise events were significantly correlated with the modelled template, aligning with studies using visual species confirmation. Our results show that using basic machine learning techniques can be used to reliably categorize white-beaked dolphin signals in PAM data from other odontocetes in Scandinavian waters. Categorization methods such as these which focus on robust signal characteristics are key for applying autonomous PAM data to broader ecology studies. • Autonomous broadband recordings can be used to monitor white-beaked dolphins without visual confirmation in Danish waters. • We detail a low computational and accessible methods using freely available software. • The categorization of stable spectral peaks and notches in all recorded echolocation events, regardless of acoustic axes. • We achieved a correct event categorization rate conforming with visually confirmed studies.