Jump to navigation
Automatic detection and classification of toothed whale echolocations clicks in diverse long term recordings
Lindeneau, Scott M.
Tarokh, MahmoudLevine, Richard
We present results of classification of toothed whale echolocation clicks to species for 6 TB of recordings in the development and evaluation data of 7th Intl. Detection, Classification, Localization, and Density Estimation Workshop. The data span multiple seasons, years, locations, and instruments. Five species were acoustically identified by analysts and a sixth category was assigned to echolocation clicks that could not be identified to species by analysts. Methods were developed to identify periods of echolocation activity taking into account sporadic false positives that occur throughout the data, and to reduce false positives from both anthropogenic and biologic sources. Dense echolocation activity presented particular challenges for noise removal and long term recordings permitted the targeting of regions for noise estimation outside of echolocation encounters. Extracted features consisted of noise normalized cepstral characterizations of spectra. These features are were classified with a Gaussian mixture model. A robust scoring method was introduced to reduce outlier influence by using a per-click voting scheme within each encounter as opposed to joint likelihood scores. The error rate across 300 Monte Carlo trials on the development data set is was 15.5% when encounters from the unspecified toothed whales are were removed.
Master of Science (M.S.) San Diego State University, 2017
© 2015 SDSU Library & Information Access. All Rights Reserved.