Novel class detection of underwater targets using Self-Organizing neural networks

An underwater target classifier can be trained only with the available limited instances of different ship and submarine emanations but in complex real world conditions, the classifier may encounter corrupted versions of the trained instances as well as novel occurrences such as targets belonging to an entirely different class. Most of the state-of-the art underwater target classifiers assign observed target data to the best matching class they have been trained with, even if the target data is an outlier, which leads to many strategic issues like security risks since a new target may be misclassified as an existing class. The unexpected or novel occurrences confuse the discriminating ability of a classifier trained only for situations of determinate meaning. Novel class or outlier detection of underwater targets, which aims to identify instances that deviate from the behavior of trained targets, assumes prime importance in the state-of-the art scenario since the ability to detect novel targets is crucial to take effective countermeasures in situations where security risks arise. The classifier must also be capable of distinguishing the interfering noise sources such as vocalizations of marine species from the targets of interest. This paper proposes a novel class detection scheme utilizing a clustering approach on an unsupervised neural network based Self-Organizing Map (SOM) provided with appropriate features. The results obtained from scatter plots, feature maps, unified distance matrix and cluster map exhibit good novel class detection accuracy for the chosen features of various target classes and selected map size of SOM.