DATA AUGMENTATION AND PREPROCESSING TECHNIQUES FOR ENHANCED UNDERWATER DETECTION AND CLASSIFICATION

The opaqueness of the ocean provides opportunities for illicit trade enterprises, acts of terrorism, and covert military operations. Compact submarines or divers can infiltrate crucial infrastructure, such as harbors, to either cause destruction or transport illegal goods within the secure expanse of the ocean. Autonomous vehicles (AXVs) emerge as potential disruptive technologies that could effectively counter these threats. Monitoring the underwater realm using multiple AXVs is sensor-intensive, resulting in collected data amounts that would quickly overwhelm human operators. Artificial Intelligence (AI) methods provide means of real-time assessment of sensor data, but this requires large and realistic passive sonar data sets for training and validation. The soundscape in a bustling harbor is characterized by a cacophony of various anthropogenic sound sources, creating a challenging environment for the interpretation of received passive sonar data. Any recorded training data set should include this wide mix of sound sources. In this study, we investigate data preprocessing methods aimed at isolating specific targets of interest, to improve detection in sounds collected from hydrophones. The primary objective of preprocessing is to mitigate the influence of undesired targets and accentuate the acoustic signatures of the intended targets. We employ the ShipsEar dataset, recorded by the University of Vigo, as an illustrative example. Additionally, augmenting techniques are applied to expand the dataset, generating more intricate soundscapes that simulate the complexity of a harbor environment. Finally, we deploy different machine/deep learning solutions to perform detection and classification. We conduct this analysis both with and without preprocessing and data augmentation to assess the influence these measures have on the overall detection performance. This comparative approach allows us to systematically evaluate the efficacy of preprocessing and augmentation techniques in enhancing the accuracy and robustness of the binary detection algorithms.