The development of a kernel to detect Ziphius cavirostris vocalizations and a performance assessment of an automated passive acoustic detection scheme

Abstract

An ensemble consisting of 150 Ziphius cavirostris vocalizations was compiled from acoustic data recorded at two High-frequency Acoustic Recording Package (HARP) locations: the Naval Postgraduate School (NPS)'s Point Sur HARP and Scripps Institution of Oceanography (SIO)'s site H HARP. The ensemble was analyzed via a principal component analysis (PCA). The results of the PCA verified the statistical robustness of the signal and yielded one dominant mode which accounted for 73% of the variance. The dominant mode was used to create a kernel for a matched filter detection scheme. The subsequent detector output was statistically evaluated against a ground truth. The ground truth identified 28,434 Ziphius clicks by visually inspecting over 170 minutes of data recorded by NPS's Data Acquisition System (DAS) at the Southern California Offshore Range (SCORE). The inability to visually discriminate a signal embedded in noise created a conservatively biased ground truth estimate which increased the detector's false alarm rate. At an acceptable 0.1% false alarm rate, the detector had an overall 44% probability of detection. A further assessment of the detector's performance divided the data into two categories: cluttered and uncluttered. At a false alarm rate of 0.1%, the probability of detection was 26% and 61%, respectively.