Machine Listening to the Seas: A Deep Dive into Acoustic-Based Vessel Classification

Abstract

The acoustic classification of ships at sea is a crucial factor for the monitoring of maritime activities and safeguarding navigation safety in the maritime environment. To utilize the advanced AI-powered methodologies that employ deep learning techniques, to improve ship classification accuracy and the efficiency is need of the hour. The work at hand proposes the ShipsEar Audio Spectrogram Transformer (SE-AST) and its selfsupervised variant that is ShipsEar-Self Supervised Audio Spectrogram Transformer (SE-SSAST) to classify different types of ships based on their acoustic signatures, which leverages the ShipsEar dataset. The model captures the complex patterns locally as well as globally through spectrogram representations of audio signals. Our approach was evaluated on the ShipsEar dataset, which contains the acoustic signatures from 5 different classes and our SE-AST algorithm achieved an accuracy of 98% demonstrating significant improvements in the classification performance. Similarly, the SE-SSAST algorithm achieved an accuracy of 99%. This methodology also improves the ship classification accuracy as well as reduces the reliance on large, labelled data sets which makes it a cost-effective solution for real-world applications. The results of this research contribute to the advancement of maritime acoustic monitoring systems, maritime safety, environmental monitoring and maritime traffic management.