Application of neural and statistical classifiers to the problem of seafloor characterization

In this paper neural and statistical classifiers are applied to the problem of seafloor classification. The feature vectors used consist of acoustic backscatter as a function of angle of incidence. Simulated seafloor backscatter is obtained by employing the Helmholtz-Kirchhoff approximation and the statistical properties of bottom reverberation. These synthetic data are used initially to train multilayer perceptrons and then to test them for their ability to discriminate among signal returns produced by seafloors with different roughness parameters. The same data are also processed with optimum Bayesian classifiers. A comparison of the results indicates a suboptimum performance for the perceptrons. The same procedures are applied to real data collected by the Sea Beam bathymetric system over two Central North Pacific seamounts. In this case, the perceptron performance is similar to that of the statistical classifier, which is no longer optimum, since no prior knowledge of the probability distribution parameters is available. In addition, Self Organizing Maps are applied to both synthetic and real data and are shown to result in a successful separation of the output space into distinct regions corresponding to different seafloor classes