Summary: | Thesis (Ph.D.)--University of Washington, 2016-04 Transient source detection and classification is a particularly challenging problem. Marine mammal vocalizations are a well-known example of these non-stationary sources, including a variety of clicks, pulse bursts and frequency sweeps. There are both environmental and legal needs to improve remote marine mammal monitoring, which can be done efficiently using passive acoustic monitoring (PAM), and public data are available to test new methods. In this thesis, I propose the use of non-negative matrix factorization (NMF) based feature representation for detection and classification of marine mammals for the following reasons: NMF can learn non-stationary signals, training requires less detailed annotations than existing species classification techniques, it can capture species-specific information from non-stereotyped vocalizations, and some NMF-based methodologies incorporate noise removal and session effect compensation. In particular, co-occurrence constraints in NMF analysis were helpful in addressing session effects in species classification. An additional direction of the research was to minimize the need for strictly labeled training data, which is arduous to create and thereby limits performance. I investigated weakly supervised learning techniques to leverage data with incomplete annotations. In these trials, recordings were made in the visual presence of a single species, but there were no annotations to indicate when vocalizations occurred. Automated detection algorithms identified potential vocalizations and then confidence-based selection methods filtered the best examples in an iterative training procedure. This method was particularly beneficial for species classification from clicks, which is very sensitive to on- and off-axis variations. Changes in orientation make the signal more variable and interfere with establishing consistent features for species classification. Weakly supervised species classification from clicks automatically identified ...
|