Marine Acoustic Signature Recognition Using Convolutional Neural Networks

Abstract

In a marine environment, there is a great diversity of sound sources, such as marine animals, natural phenomena and man-made activity. Differentiating these sound sources using passive biological-harmless systems, is an important step towards the decrease of underwater acoustic pollution and ocean sustainability. This paper presents a framework based on machine learning algorithms for recognizing underwater acoustic signals. A CNN architecture is proposed as a classifier, using the mel spectrogram of the signal as input. Firstly, this methodology is tested using a marine dataset of real underwater recordings of vessel noise, to detect the presence or absence of vessels and distinguish them according to their size. Afterwards, the database is extended with a higher diversity of sound sources: dolphins, humpback whales, small vessels, large vessels and background noise. Three data augmentation techniques are proposed for overcoming the lack of marine animal data. Recognition results indicate that the proposed framework and associated models can be a reliable tool in real-time classification, achieving an accuracy for the two datasets of 88.8% and 78.3%, respectively, with a spectrogram window shorter than 2 s. This methodology shows to be competitive in terms of accuracy, and simplicity, when compared to other recognition techniques.

Keywords: underwater noise, Hydroacoustic signal recognition, Convolutional neural networks, Mel spectrogram, Vessel noise, Marine animal vocalizations