CNNs have proved to be an important tool for automatic data classification. However, they are data hungry, as a large amount of labelled data points is needed to properly train and validate the models. This is a critical issue in the classification of underwater acoustic data, since most datasets are not publicly available, owing to the financial and technical complexity in obtaining such data and also to their potential defence-sensitive information. Therefore, much work in this area is conducted using synthetic data only (DENOS et al., 2017), or on a limited set of real data augmented with semi-synthetic examples for training (HUO; WU; LI, 2020; BERG;
HJELMERVIK, 2018; LE et al., 2020). There are, however, a few datasets commonly used in the literature, that are summarised in Table 4.
One common criticism of the synthetic or augmented datasets usage in machine learning is that these might be created tofitthe model rather than the other way around, thus generating a biased process. Therefore, there is a need for a complete dataset that comprises the information needed for this research area. In this way, the present work proposes a dataset for the vessel-type classification using underwater acoustic signals, which will be publicly available, containing the Table 4 – A summary of some datasets available.
Dataset Description
Ocean Network
Canada (HEESEMANN et al., 2014)
This (non-annotated) dataset is curated by the University of Victoria, Canada and contains a continuous monitoring of the west and east coasts of Canada and the Arctic. It is available fromhttps:
//oceannetworks.ca DeepShip: An
Underwa-ter Acoustic Benchmark Dataset (IRFAN et al., 2021)
DeepShip is a benchmark dataset for underwater ship classification which consists of 47 hours and 4 min of recordings of 265 different ships belonging to four different classes (background sound was not available). This dataset is available for download athttps:
//github.com/irfankamboh/DeepShip ShipsEar: An
under-water vessel noise database (SANTOS-DOMÍNGUEZ et al., 2016)
ShipsEar is a database containing underwater recordings of ship and boat sounds, which has 90 recordings of 11 different vessel types. It also has some useful information about the recordings, such as channel depth, wind, distance, location, to cite a few. This dataset is available for download athttp://atlanttic.uvigo.
es/underwaternoise/
Five-element acoustic dataset (PERRINE et al., 2009)
The main purpose of this dataset is to facilitate research on Doppler correction techniques for underwater acoustic transmissions. The dataset is composed of 360 communication packets with dura-tion of 0.5 sec generated by a transceiver and captured by 5 hy-drophones at nine different positions, and is available for down-load athttp://users.ece.utexas.edu/~bevans/projects/
underwater/datasets/
Fish classification with Dual-Frequency Iden-tification Sonar (DID-SON) (MCCANN et al., 2018)
Fishery acoustic observation data was collected using Dual-Frequency Identification Sonar (DIDSON) with the purpose of classifying fish species. From 100 hours of data, 524 clips were extracted with eight species labelled. The dataset is available for download athttps://osf.io/sxek6/
Passive sonar spectrogram images derived from time-series (LAMPERT;
O’KEEFE, 2013)
The main purpose of this dataset is to facilitate solutions for the problem of detecting tracks in a spectrogram. It contains 4142 spectrograms generated from synthetic and also real small-boat data.
The dataset is available for download athttps://sites.google.
com/site/tomalampert/data-sets?authuser=0 MDT
dataset
(VALDENEGRO-TORO;
PRECIADO-GRIJALVA; WEHBE,
2021)
This dataset was obtained from a forward-looking sonar (ARIS Explorer 300) placed in a water tank in which a rotating turntable was used to allow various poses for the objects observed. The MDT dataset contains 2471 images with 12 classes of object, including bottle, pipe, platform and propeller and it is available fromhttps://
github.com/mvaldenegro/marine-debris-fls-datasets
acoustic data and the physical environment variables that can impact the signal representation.
The next chapter will describe the development of the proposed dataset and the pipeline used in the proposal of a method to approach the underwater acoustic classification problem using DL.
4 A VESSEL-TYPE CLASSIFICATION METHOD
Underwater acoustic classification is a challenging task, as the sound propagation in the water has a complex behaviour and demands the consideration of various environmental features.
As DL gained attention showing the potential to solve problems unfeasible by humans alone, their application in this area is increasing and presenting promising results. Besides that, this advance is still in the early stages and needs more attention on tools, methodologies, and datasets development.
This chapter will present the gap found in the literature that this work aims to fulfil and the motivation behind this objective. Also, the development steps conducted in this way will be described, focusing on three main stages: the initial approach involving the usage of a variation of the most common DL architecture found in the literature, the VGG, applied to one of the existing datasets, aiming to establish a baseline to be used as the base on the subsequent developments;
the formulation and proposal of a new annotated dataset, which seeks to fill the gaps stated on the Chapter 3; and the development of a complete method, starting from the baseline, to the development of an underwater acoustic classification approach, considering the distance from target to sensor as a significant factor, and optimising the preprocessing of the audio signals.