Manifold-Based Classification of Underwater Unexploded Ordnance in Low-Frequency Sonar.

This paper addresses the problem of discriminating underwater unexploded ordnance (UXO) from non-UXO objects using manifold learning principles when applied to data collected from low-frequency sonar. Our classification hypothesis is that the sequence of measurements collected from an object lie in some low-dimensional subspace which is locally linear but globally nonlinear. These low-dimensional features and their behavior on the manifold can then be used to discriminate among various UXO and non-UXO objects that may be encountered in shallow water environments. With this goal in mind, techniques are developed to not only learn the manifold but also to provide an out-of-sample embedding for newly acquired data. The manifold features from the training set are then used to construct local linear subspaces for representing each newly embedded testing feature vector. A statistical-based technique is then used to select the most likely class label by finding the class which best represents the data. The ability of the classifier to discriminate among multiple object types is then demonstrated using a sonar data set collected from underwater objects in a controlled setting. Classification results are presented and compared with an alternative method that also relies on a set of features extracted using manifold learning principles. [ABSTRACT FROM AUTHOR]