Marine Acoustic Signature Recognition using Convolutional Neural Networks

07/19/2021

In a marine environment, there is a great diversity of sound sources: marine animals, natural phenomena and man-made activity. Differentiating these sources is an important response to ecological challenges and to ensure better control of the coastline. The current work aims to devise a model which analyses acoustic signals from hydrophones and classifies them according to the sound source. To achieve this, a convolution neural network (CNN) is proposed as a classifier, using the mel spectrogram representation divided into small intervals (windows) of the acoustic signal and its derivatives as input. Class scores are assigned to each window by the CNN. The developed methodology is applied to two different datasets composed of hydroacoustic data. The first comprises vessel noise data, where the objective is to detect the presence of vessels and distinguish the vessels according to the size. A classification accuracy of 83.2% and 88.8% is achieved using the mel spectrogram and the mel spectrogram plus its first and second derivatives as features, respectively. The second dataset complements the previous one by adding dolphin and whale vocalizations. Three data augmentation techniques (time stretching, pitch shifting and time shifting) are studied. Whichever of the three techniques is used individually outperform the model without data augmentation. This is equal true when all three techniques are used simultaneously. The impact of the window length is also studied, with five different models being creating where the window length varies between 0.22 s and 1.97 s. The classification accuracy ranges between 66.2% and 78.3%.

Authors in the community:

• 

  A

  Alexandre Martins Correia

  ist425833

Supervisors of this institution:

• 

  J

  João Miguel Da Costa Sousa

  ist12897

mechanical-engineering - Mechanical engineering

por - Portuguese

Instituto Superior Técnico