ASA 125th Meeting Ottawa 1993 May

5aAO13. Application of neural and statistical classifiers to the problem of seafloor characterization.

Zoi-Heleni Michalopoulou

Dimitri Alexandrou

Duke Univ., Dept. of Elec. Eng., Box 90291, Durham, NC 27708-0291

Simulated seafloor backscatter is obtained by employing the Kirchhoff approximation and the statistical properties of bottom reverberation. Such data are presented to multilayer perceptrons for training and testing, aiming at the development of a neural processor that discriminates among signal returns coming from seafloors with different roughness parameters. Experiments on the same data sets are performed with optimum Bayesian classifiers as well; a comparison of the results indicates suboptimum performance for the perceptrons. The same procedure is followed with real data collected by the bathymetric system Sea Beam over Horizon Guyot and Magellan Rise. In this case, the perceptron performance is comparable to that of the Bayesian classifier, which is no longer optimum, since no prior knowledge of the probability distribution parameters is available. In addition, self-organizing maps have been applied to both synthetic and real data sets and resulted in a successful separation of the output space into distinct regions corresponding to the existing classes. [Work supported by the Office of Naval Research, Code 1125GG, through Contract No. N00014-93-I-0049.]