Acoustic detection of cetaceans is challenged by the variability of calls and the presence of variable background noise. One detection method is to start with frequency band-limited signal and noise estimates, and apply a likelihood ratio test (LRT). These detectors suffer from false alarms when broadband signals overlap the band of interest, triggering detection. Some detectors only consider previous samples, causing further false alarms. The authors propose a method of reducing false alarms by defining a guard band that is not expected to contain energy from the species of interest. A second LRT is performed by testing the ratio of the signal estimate in the signal band with the signal estimate of the guard band. This method is shown to reduce false alarms with a small reduction in detection performance. A detection method is also presented that can be optimized for high processing efficiency, while improving false-alarm rejection from signals that are longer in duration than the signal of interest. Performance is demonstrated on real cetacean recordings and ocean noise. The detection algorithm is implemented in PAMGUARD, an open source Java application designed for passive acoustic monitoring (PAM) of cetaceans.

Link To Publication