Performance evaluation of decision trees and multilayer perceptrons in seabed classification

Abstract

Accurate knowledge of the oceanic propagation medium, and, thus, seabed properties, is of paramount importance in source localization, especially of weak targets. In this work, we employ machine learning techniques to perform seabed identification and classification using impulse responses of different media; supervised learning is employed. We first design a decision tree architecture that relies on features selected from the impulse responses corresponding to the different media. Examples of these features are kurtosis, skewness, and strength of the signal. Training sets are created by identifying features from noisy signals and testing follows after extracting corresponding features from a different data set. Performance is evaluated as a function of Signal-to-Noise Ratio. A principal component analysis is also implemented for the investigation of the potential for dimensionality reduction. Subsequently, multilayer perceptrons are employed using identical data both for training and testing and the two machine learning techniques are compared; advantages and disadvantages of each are identified and discussed in this work. [Work supported by ONR.]