Article Preview
TopIntroduction
Since the invention of sonar, a growing interest was given to underwater sounds that radiated from boats. Although the first applications were related to military purposes, several studies aim at using the underwater sounds for non-military uses, such as maritime traffic management (Fillinger, 2009; Zwemer, 2018), underwater surveillance (Fillinger et al., 2010), assessment of the impact of noise pollution on marine life (Codarin, 2009; Holles, 2013), among others.
Since the acoustic signal produced by a ship has many sources (propeller, machinery, hydrodynamic, vibrations, etc.) that produced tones at different frequencies, most developed methods to detect, classify or track boats are applied in the frequency domain or time-frequency domain. Several approaches were developed to detect the harmonic frequencies in the signals and to extract the acoustic signature of the ships. In many cases, these methods are based on the spectrum (Guo et al., 2020), DEMON spectrum (Chung et al., 2011) and Cepstrum (Das, 2013; Santos-Domínguez, 2016) and the automatic detection is usually performed by detecting the peaks with a threshold (Reis et al., 2019).
In recent years, deep neural networks have seen a lot of successful applications in many different domains. One successful deep learning architecture used in computer vision is a convolutional neural network (CNN). This architecture is known to automatically learn complex feature representations using its convolutional layers and has led to impressive results in many problems such as in image classification (Krizhevsky et al., 2012), speech recognition (Palaz et al., 2015) or time series classification (Cui, 2016; Guennec, 2016; Zhao, 2017). More recently, several studies (Li, 2019; Yamaguchi, 2019) aimed at detecting boats have used machine learning methods, which have the advantage of automatically extracting characteristic attributes to classify the data. These methods are often more robust to noise and do not require hand designed features. However, all the proposed methods performed the detection on a 2D-signal in the frequency domain of the recorded sound. In this article, the researchers aimed to show than the detection itself can be performed with a simpler 1D signal that does not required the computation of the spectrum or the use of a 2 dimensional CNN. With a 1D signal, the detection can be performed with a time series classifier.