2-3. Detection and classification of leopard seal vocalizations
During the analysis of the underwater soundscape, we confirmed that
leopard seal vocalizations were prominent and persistent throughout the
recording period, as evidenced by spectrograms from randomly selected
WAV files. Consequently, we focused on detecting and classifying these
vocalizations. All acoustic data files, recorded in 10-minute durations,
were subjected to the detection process on both a waveform and a
spectrogram in 1-minute intervals. Call signals, which are stronger than
the background noise level and whose features could be clearly
distinguished in the spectrogram, were manually detected, and then
classified into HDT, MST, LDT, HST, DT, and AT based on the call types
in previous studies (Stirling and Siniff, 1979; Rogers, 2007; Klinck,
2008). The call patterns of the five types were mostly consistent with
those reported in previous studies except for AT, and the acoustic
characteristics of each call type were extracted by a process that
considered the detailed features of the signals. For low-frequency
vocalizations, uncertainty was accounted for by re-detecting calls
within 1-hour segments, counting the number of missed and false
detections, and adjusting these values relative to the total recording
times. Similarly, for HDT and MST, uncertainty was estimated by randomly
extracting detected signals, counting the number of false detections,
and adjusting these values relative to the corresponding time period.