Passive acoustic methods are in widespread use to detect and classify cetacean species; however, passive acoustic systems often suffer from large false detection rates resulting from numerous transient sources. To reduce the acoustic analyst workload, automatic recognition methods may be implemented in a two-stage process. First, a general automatic detector is implemented that produces many detections to ensure cetacean presence is noted. Then an automatic classifier is used to significantly reduce the number of false detections and classify the cetacean species. This process requires development of a robust classifier capable of performing inter-species classification. Because human analysts can aurally discriminate species, an automated aural classifier that uses perceptual signal features was tested on a cetacean data set. The classifier successfully discriminated between four species of cetaceans-bowhead, humpback, North Atlantic right, and sperm whales-with 85% accuracy. It also performed well (100% accuracy) for discriminating sperm whale clicks from right whale gunshots. An accuracy of 92% and area under the receiver operating characteristic curve of 0.97 were obtained for the relatively challenging bowhead and humpback recognition case. These results demonstrated that the perceptual features employed by the aural classifier provided powerful discrimination cues for inter-species classification of cetaceans.

Passive acoustic methods are widely used to detect and classify marine mammals; however, these passive sonar systems are often triggered by other transient sources, producing many false alarms. Additionally, to positively identify marine mammals, large volumes of data are collected that need to be processed by a trained analyst. To reduce acoustic analyst workload, an automatic detector can be implemented that produces many detections, which feed into an automatic classifier that significantly reduces the number of false detections. This requires development of a classifier capable of performing inter-species classification. A prototype aural classifier has been developed at Defence R&D Canada that uses perceptual signal features which model the features employed by the human auditory system. Previous effort has shown the aural classifier successfully discriminated cetacean vocalizations from five species: North Atlantic right, humpback, bowhead, minke, and sperm whales. This paper examines the effects of replacing principal component analysis (PCA) with discriminant analysis (DA) for feature space dimensionality reduction. PCA projects data onto a lower dimensional space so as to preserve the greatest scatter of data points, whereas DA projects the data to achieve the greatest separation of classes. Benefits of implementing DA and improvements to classification results will be discussed.

No abstract

Previous effort has shown that a prototype aural classifier developed at Defence R&D Canada can be used to reduce false alarm rates and successfully discriminate cetacean vocalizations from several species. The aural classifier achieves accurate results by using perceptual signal features that model the features employed by the human auditory system. Current work focuses on determining the robustness of the perceptual features to propagation effects for two of the cetacean species studied previously-bowhead and humpback whales. To this end, classification results are compared for the original vocalizations to classification results obtained after the vocalizations were re-transmitted underwater over ranges of 2 to 10 km. Additional insight into the propagation effects is gained from transmission of synthetic bowhead and humpback vocalizations, designed to have features similar to the most important aural features for classification of bowhead and humpback vocalizations. Each perceptual feature is examined individually to determine its robustness to propagation effects compared to the other aural features. To gain further understanding of propagation effects on the features, preliminary propagation modelling results are presented in addition to experimental data.

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; however, these systems often require tuning when deployed in different environments. Anecdotal evidence suggests that this requirement to adjust a PAM system's parameters is partially due to differences in the acoustic propagation characteristics. The environmentdependent propagation characteristics create variation in how a cetacean vocalization is distorted after it is emitted. If these difference are not accounted for it could reduce the performance of automated PAM systems. An aural classifier developed at Defence R&D Canada (DRDC) has been used successfully for inter-species discrimination of cetaceans. Accurate results are obtained by using perceptual signal features that model the features employed by the human auditory system. In this thesis, a combination of an at-sea experiment and simulations with modified bowhead and humpback whale vocalizations was conducted to investigate the robustness of the classifier performance to signal distortion as a function of propagation range. It was found that in many environments classification performance degraded with increasing range, largely due to decreased signal-to-noise ratio (SNR); however, in some environments as much as 40 % of the performance reduction was attributed to signal distortion resulting from environment-dependent propagation. It was found that sound speed profiles resulting in considerable boundary interaction were important for producing sufficient signal distortion to affect PAM performance, relative to the impacts of SNR. Therefore, in some environments the ocean acoustic properties should be taken into account when characterizing performance of automated PAM systems. For the environments in which signal-to-noise issues dominate, the use of multi-element arrays is expected to increase the performance of automated recognition systems beyond the minor improvements to be gained from adjusting a PAM system's parameters. Nonetheless, propagation modelling should be used to complement PAM experiments to account for bias in probability of detection estimates resulting from environment-dependent acoustic propagation.