Entropy-based analysis of influential factors for underwater acoustic target recognition in passive sonar data

Abstract

Underwater acoustic target recognition (UATR) using passive sonar frequently demonstrates significant performance degradation with unseen data, owing to discrepancies in data distributions between the training and test sets. In this study, four selected influential factors that meaningfully contribute to cluster formation within the dataset, namely, ship type, ship size, individual location, and combined location category distinguishing between port and canal, are systematically analyzed. Unsupervised deep embedded clustering is repeatedly applied to the ShipsEar dataset using four different numbers of clusters, each corresponding to the number of classes associated with a specific influential factor. We propose an entropy-based metric for evaluating the alignment between the results clusters and ground-truth classes in each case. Measurement location, which determines sound propagation conditions, is a dominant influential factor along with ship type and ship size, which are related to the inherent characteristics of sound sources. To enhance the influence of ship type or ship size in the corresponding classification tasks, we mitigate the effect of measurement location by training two separate classifiers (one for ports and one for canals) using the respective subsets of the dataset. With the location constraint, classification accuracy improves across various data-splitting methods during testing.