Minimum Hellinger Distance Classification of Underwater Acoustic Signals

Open Access
Bissinger, Brett
Graduate Program:
Electrical Engineering
Master of Science
Document Type:
Master Thesis
Date of Defense:
October 09, 2009
Committee Members:
  • N K Bose, Thesis Advisor
  • Nirmal K Bose, Thesis Advisor
  • Richard Lee Culver, Thesis Advisor
  • underwater acoustics
  • hellinger distance
  • minimum distance
  • classification
Passive source classification in the underwater environment is a challenging problem in part because propagation through the space- and time-varying medium introduces variability and uncertainty in the signal. Acoustic propagation codes can predict received fields accurately but they are sensitive to input environmental parameters which cannot be known exactly. This uncertainty in environmental knowledge used in signal predictions results in imperfect statistical class models. Classifiers that rely on simulations of the environment must therefore be robust to imperfect environmental models. Maximum likelihood methods provide ideal performance when the class models are correct but their performance quickly deteriorates when class models are imperfect. Minimum distance methods generally can offer robustness to mismatches at the expense of performance, with that tradeoff governed by the distance metric used. Hellinger distance, when used as a distance metric, offers robustness to outliers while retaining the performance of a maximum likelihood method, properties that make it well-suited for classification of passive underwater acoustic signals. In the present work the robustness of the Minimum Hellinger Distance Classifier (MHDC) is quantified and its performance is compared to a Log-Likelihood Ratio Classifier(LLRC) with three different data sets: synthetic Gaussian data, synthetic acoustic data from propagation simulations and real acoustic data. In cases of acoustic data, class models are derived from Monte Carlo acoustic propagation simulations. In each case Receiver Operating Characteristic (ROC) curves show that the MHDC exhibits performance equivalent or superior to that of the LLRC, responding in a robust manner to imperfect class models.