The DeuteroNoise dataset: An initial exploration of an underwater vessel sound dataset within the framework of a JPI-oceans project

Abstract

The vast and largely unexplored underwater environment is a rich source of diverse sound events. These sounds, which range from the calls of marine life to the noise generated by human activities, create a complex acoustic environment. Research has been conducted on gathering and categorizing this type of data. However, only a few databases have been recently openly published focusing on anthropogenic sounds. This paper outlines the preliminary steps towards the creation of a comprehensive dataset for the detection of underwater sound events, with an initial emphasis on the sounds of vessels and boats. Within the framework of the JPI Oceans project, DeuteroNoise, a wide spectrum of vessel sounds under varying conditions has been captured and annotated with the necessary metadata for sound event detection tasks. Moreover, the proposed dataset will facilitate the development of more accurate and adaptable vessel sound event detection models and encourage further research in this area. The dataset contains raw audio files and the respective initial analysis based on labeled events, duration of the events, signal-to-noise ratio (SNR) and impact measurements. The retrieved data can be grouped into noisy and non-noisy spots. The noisy locations include the Port of Barcelona (Spain), the Port of Constant,a (Romania), and the Lagoon of Venice (Italy). In contrast, non-noisy data was collected during two measurement campaigns at Pont del Petroli in Badalona (Spain). In addition to the dataset, to illustrate its potential application, a classifier for vessel/boat sound events is proposed. The classifier uses mel spectrograms as input data and is built on a pre-trained model that leverages a residual neural network. This system is capable of classifying vessel/boat related events from background sound environment.