Ocean and Coastal Management 218 (2022) 106015

Contents lists available at ScienceDirect

Ocean and Coastal Management

journal homepage: www.elsevier.com/locate/ocecoaman

A semi-supervised deep learning approach for vessel trajectory classification

based on AIS data
Hongda Duan a,b , Fei Ma c , Lixin Miao b,c , Canrong Zhang a,b ,∗
a
Department of Industrial Engineering, Tsinghua University, Beijing 100084, China
b
Research Center for Modern Logistics, Shenzhen International Graduate School, Tsinghua University, Shenzhen 518055, China
c
Tsinghua-Berkeley Shenzhen Institute, Tsinghua University, Shenzhen 518055, China

ARTICLE INFO ABSTRACT

Keywords: Automatic identification system (AIS) refers to a new type of navigation aid system equipped in maritime
Automatic identification system vehicles to monitor ship performance. It provides trajectory information of vessels which can be used
Vessel trajectory classification for the classification task. The classification results facilitate ocean surveillance and conservation, vessel
Deep learning
management enhancement, fishery regulation, marine ecological sustainability, and nautical safety protection.
Semi-supervised learning
Some progresses have been made based on machine learning or deep learning strategies to perform supervised
Variational autoencoder
learning by assuming that during the training process, the category labels of historical trajectory data are
available. However, in reality, the label information may be difficult or expensive to obtain, resulting in that
only a small part of the training data is labeled, and most of the training data is unlabeled. To address this
issue, this paper proposes a semi-supervised deep learning approach to integrate the knowledge of unlabeled
data for vessel trajectory classification. Here, we call our approach SSL-VTC. Specifically, we first extract
vessel trajectories by integrating the kinematic and static information from historical AIS messages. Then, we
design convolutional neural networks to extract feature representations from the vessel trajectories. Finally, we
develop a semi-supervised learning algorithm based on the variational autoencoder to perform discriminative
learning and generative learning simultaneously. In this way, our SSL-VTC framework can fully leverage the
labeled data and unlabeled data during the training process. To the best of our knowledge, we are the first
to utilize semi-supervised learning for vessel trajectory classification. Experimental results on the public AIS
dataset show that our SSL-VTC can effectively extract feature representations from the AIS messages and its
performance is significantly better than the traditional supervised learning methods. The approach and findings
of this study provide important implications for researchers and stakeholders in the field of ocean management.

1. Introduction Vessel trajectory classification is an important research topic using

AIS data by predicting the category of ship trajectories. Ship types
Automatic identification system (AIS) is an automatic vessel self- are often used as the category labels for the classification task. The
reporting system used for navigation safety and maritime transporta- common ship types include fishing, cargo, passenger, and tanker. The
tion management (Zhang et al., 2018). The AIS system transmits the classification results play a crucial role in ocean surveillance and
vessel status information at a variable refresh rate, such as vessel conservation (Almpanidou et al., 2021), vessel management enhance-
name, position, speed, course, dimension, and draft, and those pieces ment (Zhang et al., 2021a), and navigation safety protection (Liu et al.,
of information can be received by other vessels or vessel traffic service
2021a). For example, by identifying and tracking illegal, unreported,
(VTS) centers (Harati-Mokhtari et al., 2007). With the widespread use
and unregulated (IUU) fishing vessels and overfishing activities, the
of AIS systems onboard, AIS data has arguably kickstarted the era of
classification results strengthen the fishery management and the ocean
digitization in marine governance and port and shipping operation
ecosystem conservation (Krüger, 2019; Arasteh et al., 2020; Alves,
and management (Yang et al., 2019; Cominelli et al., 2020) and has
attracted great attention from the academia recently (Le Tixerant et al., 2021; Cánovas-Molina et al., 2021; Garcia et al., 2021; Yu and Wang,
2018; Svanberg et al., 2019; Chen et al., 2020; Liu et al., 2021b; 2021). In addition, the classification results can also be used to assist
McWhinnie et al., 2021; Meyers et al., 2021; Tan et al., 2021). navigators or coastal authorities in detecting suspicious or unsafe vessel

∗ Corresponding author at: Research Center for Modern Logistics, Shenzhen International Graduate School, Tsinghua University, Shenzhen 518055, China.
E-mail address: [email protected] (C. Zhang).

https://doi.org/10.1016/j.ocecoaman.2021.106015
Received 26 June 2021; Received in revised form 17 November 2021; Accepted 14 December 2021
Available online 10 January 2022
0964-5691/© 2022 Elsevier Ltd. All rights reserved.
H. Duan et al. Ocean and Coastal Management 218 (2022) 106015

Fig. 1. Illustration of supervised learning and semi-supervised learning. Compared with supervised learning, semi-supervised learning can learn a more accurate decision boundary
to achieve better classification performance with the help of the unlabeled data.

behaviors and possible threats by comparing the classified ship types • To the best of our knowledge, this paper is the first work to
with the ones reported by vessels (Ljunggren, 2018; Nguyen et al., develop a semi-supervised deep learning approach based on the
2018). variational autoencoder for vessel trajectory classification using
To research the problem of vessel trajectory classification, differ- AIS data.
ent approaches (Hu et al., 2016; Sheng et al., 2018; Herrero et al., • We utilize, for the first time, both kinematic and static infor-
2019) based on machine learning have been proposed. Due to the mation of the AIS messages to extract vessel trajectories for the
great successes of deep learning in many domains (LeCun et al., 2015; classification task, which is more efficient than previous works.
Goodfellow et al., 2016), some researchers (Ljunggren, 2018; Mantecón • We conduct extensive experiments on the public AIS dataset to
et al., 2019; Arasteh et al., 2020; Kontopoulos et al., 2021) have show the effectiveness of our approach and further provide impor-
applied it for vessel trajectory classification, and have achieved higher tant implications of our study in the field of ocean management.
performance than traditional methods. Despite their progress, they are
limited to the following two issues. The rest of this paper is organized as follows. In Section 2, related
Firstly, previous works usually perform the task of vessel trajec- works are reviewed. In Section 3, this paper proposes the procedure of
tory classification with supervised learning by assuming that the label trajectory extraction. In Section 4, we present our SSL-VTC framework,
information is available during the training process. However, in re- which includes problem formulation, the semi-supervised learning algo-
ality, it may be expensive or time-consuming to obtain the labeled rithm, and the network. In Section 5, extensive numerical experiments
data (Chapelle et al., 2009; Zhu and Goldberg, 2009; Van Engelen and are conducted. Then, we propose some important implications of our
Hoos, 2020). As a result, during the training process, only a small study in Section 6. Finally, conclusions and some future works are
part of the data is labeled, and most of the data is unlabeled. Unla- shown in Section 7.
beled data usually still contains useful information for classification.
Secondly, most previous works only use the kinematic information 2. Related works
of AIS messages, such as latitude, longitude, speed over ground, and
course over ground, to extract vessel trajectories. Static information is In this section, we review previous works of vessel trajectory clas-
usually overlooked, including the length, width and draft of vessels. In sification, which are based on machine learning or deep learning
fact, static information can also be used effectively for vessel trajectory methods. In addition, we show two other related problems that also use
classification (Zhong et al., 2019). So far, there are few works which the AIS data: trajectory clustering and trajectory reconstruction. For a
take into account both kinematic and static information to classify comprehensive understanding of these works, please refer to the survey
vessel trajectories. papers (Tu et al., 2017; Le Tixerant et al., 2018; Svanberg et al., 2019;
To address these problems, this paper proposes an efficient ap- Yang et al., 2019).
proach for vessel trajectory classification with semi-supervised deep
learning. In the following, we call our approach SSL-VTC for brevity. 2.1. Vessel trajectory classification
Specifically, we first exploit both kinematic and static information
from historical AIS messages to extract vessel trajectories. Inspired Vessel trajectory classification refers to classifying the trajectories
by deep learning, we then design convolutional neural networks to into several categories based on the AIS spatio-temporal data. Consid-
learn feature representations from the extracted vessel trajectories. erable attention has been given by researchers to study this problem
Furthermore, in order to make full use of the AIS data which may be with machine learning or deep learning methods.
labeled or unlabeled, we resort to semi-supervised learning (Chapelle In terms of machine learning methods, different techniques have
et al., 2009; Zhu and Goldberg, 2009). It can use both labeled data and been used to address the classification problem. For example, Hu et al.
unlabeled data to help us learn a better classification model. Unlike (2016) present an approach to detect fishing activities from historical
semi-supervised learning, traditional supervised learning only uses the AIS data with conditional random fields. Sheng et al. (2018) use the
labeled data, so its classification performance is limited. The difference logistic regression algorithm to build a classifier with trajectory features
between supervised learning and semi-supervised learning is shown in to classify ship trajectories into fishing and cargo. Herrero et al. (2019)
Fig. 1. Here, we develop a semi-supervised learning algorithm based first filter AIS data using the interacting multiple model and then
on the variational autoencoder (Kingma and Welling, 2013; Kingma classify ship trajectories with a multiclass binary decision tree. Zhong
et al., 2014) for the classification task. It can perform discriminative et al. (2019) utilize a classifier based on random forest to classify ships
learning and generative learning simultaneously. In this way, our SSL- into fishing, cargo, and tanker.
VTC architecture can effectively take advantage of both labeled data Motivated by the great success of deep learning, researchers then
and unlabeled data for vessel trajectory classification. We conduct design different methods based on deep learning for vessel trajectory
experiments on the public AIS dataset to demonstrate that based on the classification. Jiang et al. (2016) focus on the detection of fishing
variational autoencoder, our semi-supervised deep learning framework activities in vessel trajectories using autoencoders. Ljunggren (2018)
can learn feature representations efficiently from the AIS data and employ a convolutional neural network to learn vessel-specific features
outperform traditional supervised learning approaches. and classify the ship types into fishing, passenger, cargo, and tanker
To sum up, our main contributions are three-fold: based on AIS kinematic data. Mantecón et al. (2019) propose a deep

2
H. Duan et al. Ocean and Coastal Management 218 (2022) 106015

learning framework with a convolutional neural network for vessel most previous works mainly use the vessel kinematic information,
monitoring as a basis for anomaly detection. Chen et al. (2020) present including LAT, LON, SOG, and COG. As shown in Zhong et al. (2019),
a convolutional neural network-ship movement mode classification the vessel static information, such as LEN, WID, and DRA, is also
algorithm. Arasteh et al. (2020) propose a simple yet effective convo- beneficial to the classification task. Inspired by this, we propose to
lutional neural network for trajectory classification, where the model extract vessel trajectories from the AIS data by combining the kinematic
is trained with a set of invariant spatio-temporal features extracted and static information. In the following, we show how to extract vessel
from the behavioral characteristics of vessel movements. Kontopoulos trajectories from the AIS data step by step.
et al. (2021) transform vessel trajectory patterns into images and
employ deep learning algorithms to classify vessel activities. These deep • Step 1 (Trajectory Division): The MMSI in Fig. 2 is a unique
learning methods have achieved higher performance than traditional nine-digit number for identifying a vessel. We use it to separate
machine learning methods. different vessels. For each vessel, we divide all AIS messages by
Although the above methods have made progresses on the prob- day. Each AIS message contains a timestamp, kinematic informa-
lem of vessel trajectory classification, they often assume that all data tion, and static information. If the timestamps of adjacent AIS
are labeled during the training process. However, in practice, it may messages differ by more than two hours, we further divide the
be time-consuming or expensive to obtain labeled data, resulting in obtained daily AIS data into vessel trajectories.
only a small part of the training data being labeled, and most of the • Step 2 (Trajectory Filtering): To ensure that the time length of
training data being unlabeled. These methods are unable to tackle this each trajectory is sufficient, we filter out trajectories with the time
challenge. To the best of our knowledge, we are the first to leverage length less than six hours. In addition, we remove trajectories
semi-supervised learning for the task of vessel trajectory classification, that have fewer than 160 AIS messages to make the extracted
which can efficiently combine the information of the labeled and trajectories have enough information for the classification task.
unlabeled data. In addition, most previous works only use the kinematic • Step 3 (Abnormal Trajectory Removal): After filtering, we find
information of the AIS data to extract ship trajectories. Here, we exploit that the speeds of some trajectories are very low. To ensure that
both kinematic and static information to make full use of the AIS data. the trajectory can be distinguished in the spatio-temporal domain,
we remove the following two types of abnormal trajectories. The
2.2. Vessel trajectory clustering and reconstruction
first is the trajectory with a maximum SOG of not greater than 1
knot per hour. The second is the trajectory where the number of
Apart from vessel trajectory classification, two other related prob-
AIS messages with SOGs greater than 2 knots per hour accounts
lems are also reviewed here: vessel trajectory clustering and vessel
for no more than 30% of the total number of AIS messages.
trajectory reconstruction. The former splits all trajectories into multiple
• Step 4 (Normalization and Seven-hot Encoding): After remov-
groups, so that the trajectories of the same group are very similar and
ing abnormal trajectories, we discard the timestamp attribute
the trajectories of different groups are different, while the latter is to
preprocess the original AIS messages because the AIS data collected and then normalize the seven attributes of kinematic and static
by sensors are often incomplete or noisy. Similar to vessel trajectory information of AIS messages in the trajectory, including LAT,
classification, researchers are increasingly using machine learning or LON, SOG, COG, WID, LEN, and DRA. Next, we consider how
deep learning methods to analyze the AIS data for vessel trajectory to represent these seven attributes so that they can be sent to
clustering. In addition, the AIS data preprocessing technique in vessel our deep learning model for classification. If we use their real
trajectory reconstruction also needs to be considered in vessel trajectory values directly, it is difficult for neural networks to disentangle
classification. the underlying meaning of these numbers. Inspired by Nguyen
For vessel trajectory clustering, Dobrkovic et al. (2016) use a genetic et al. (2018, 2021), we perform one-hot encoding for each at-
algorithm to cluster vessel position based on the AIS data. Yao et al. tribute and concatenate the one-hot encoding of these seven
(2017) propose a moving behavior feature extraction algorithm to attributes together, as shown in Fig. 3. Here, we call this pro-
extract moving behavior features that capture space and time invariant cedure seven-hot encoding. It is noted that one-hot encoding
characteristics of trajectories. Wang et al. (2020) present a ship tra- essentially converts the real value of each attribute into a set of bi-
jectory motion pattern extraction algorithm using the one-dimensional nary representations, in which only one bit is 1 and the others are
convolutional auto-encoder without spatio-temporal trajectory mea- 0. By discretizing the real values of different attributes, seven-hot
surement methods. For vessel trajectory reconstruction, Zhang et al. encoding helps our framework better learn the spatio-temporal
(2018) employ a multi-regime vessel trajectory reconstruction model information of the trajectories.
which includes three steps, namely (i) outliers removal, (ii) ship nav- • Step 5 (Trajectory Sampling): Finally, since the number of AIS
igation state estimation, and (iii) vessel trajectory fitting. Chen et al. messages per trajectory is different, the length of each trajectory
(2020) propose an ensemble ship trajectory reconstruction framework is different. We sample the same number of AIS messages from
combining a data quality control procedure and a prediction module. each trajectory to form fixed-length trajectories and send them to
Guo et al. (2021) present an improved kinematic interpolation for the our deep learning model for classification.
AIS trajectory reconstruction.
The above steps are summarized in Fig. 4. Through these proce-
3. Trajectory extraction dures, we transform the AIS data into vessel trajectories. Each trajectory
is used as the input to our framework for vessel trajectory classification,
For vessel trajectory classification, we need to extract trajectories which contains both kinematic and static information.
from the AIS data, which are further used as the input of our semi-
supervised learning algorithm. Therefore, this procedure provides an
4. Proposed framework
important basis for our SSL-VTC framework. In this section, we show
how to extract trajectories in detail.
In Fig. 2, we show some examples of the original AIS data, in which In this section, we first show the problem formulation and the
each row is an AIS message, including maritime mobile service identity overview of the proposed semi-supervised deep learning framework,
number (MMSI), basedatetime, latitude (LAT), longitude (LON), speed then explain how our SSL-VTC model is trained, and finally describe
over ground (SOG), course over ground (COG), length (LEN), width network structures, including the encoder, decoder, and classifier, to
(WID), and draft (DRA). To extract trajectories from the AIS messages, implement our semi-supervised learning algorithm.

3
H. Duan et al. Ocean and Coastal Management 218 (2022) 106015

Fig. 2. Examples of the original AIS data.

Fig. 3. Seven-hot encoding.

4.1. Problem formulation

Suppose during the training process, we have two datasets: labeled

{( (𝑖) (𝑖) )}𝑛1
and unlabeled. We use 𝐷1 = 𝐱 ,𝑦 𝑖=1
to represent the labeled
dataset, where 𝐱(𝑖) is the 𝑖-th trajectory we extract based on the AIS
data using procedures in Section 3, 𝑦(𝑖) is the corresponding label, such
as fishing, cargo, tanker, and passenger, and 𝑛1 is the size of the dataset
{( )}𝑛1 +𝑛2
𝐷1 . We then use 𝐷2 = 𝐱(𝑖) 𝑖=𝑛 to represent the unlabeled dataset,
1 +1
where 𝑛2 is the size of 𝐷2 . For brevity, the superscript 𝑖 is omitted in
the following when referring to a single trajectory point. In addition, we
assume that each trajectory 𝐱 in 𝐷1 and 𝐷2 has a corresponding latent
variable 𝐳. Our target is to combine the information of the labeled and
unlabeled data based on the variational autoencoder to learn a classifier
that has good classification performance.

4.2. Overview

We propose the semi-supervised learning framework for vessel tra-

jectory classification, as shown in Fig. 5. Our semi-supervised learning
approach is achieved based on the variational autoencoder. It has three
modules: encoder, decoder, and classifier. We use the classifier for
Fig. 4. The procedure of trajectory extraction.
discriminative learning, and the encoder and decoder for generative
learning. The classifier is used to classify trajectories. The encoder is
used to extract latent variables from the labeled and unlabeled trajecto-
ries. If the input trajectory is labeled, the decoder reconstructs the input labeled, which makes the unlabeled data unusable. Our semi-supervised
trajectory combining the corresponding label embedding and the latent learning can efficiently use both labeled data and unlabeled data. Based
variable learned by the encoder. If the input trajectory is unlabeled, the on the variational autoencoder, the unlabeled data is exploited in our
decoder reconstructs the input trajectory combining the knowledge of deep learning model to help with the classification task. In this way,
the label embedding predicted by the classifier and the latent variable the classification performance of our method is better than that of the
learned by the encoder. The encoder, decoder, and classifier are a traditional supervised learning method.
unified whole, and are trained together for discriminative learning and
generative learning. The training of these three modules affects each 4.3. Semi-supervised learning algorithm
other. The good discrimination performance of the classifier promotes
In the above section, we discuss the problem formulation and the
the encoder and decoder to achieve good reconstruction results. The
overview of the proposed framework. In this section, we further present
good generation performance of the encoder and decoder can in turn
how to design a semi-supervised learning architecture to utilize the
promote training the classifier. In this process, the classification ability
labeled and unlabeled data for vessel trajectory classification. In the
of the classifier is enhanced by exploiting the knowledge of the labeled
deep learning model, the loss function plays a crucial role in updating
and unlabeled data.
the model parameters. Here, we utilize a loss function based on the
It is worth noting that our semi-supervised learning method is
variational autoencoder to combine the unlabeled data, as follows:
significantly different from traditional supervised learning methods.
For supervised learning, the data used for training is required to be 𝐿 = 𝐿1 + 𝐿2 + 𝛼 ⋅ 𝐿𝑐𝑙𝑓 (1)

4
H. Duan et al. Ocean and Coastal Management 218 (2022) 106015

Fig. 5. Our semi-supervised learning framework for vessel trajectory classification. It has three modules: encoder, decoder, and classifier. The encoder and decoder are used for
generative learning, and the classifier is used for discriminative learning.

It can be found that 𝐿 has three parts: 𝐿1 , 𝐿2 , and 𝐿𝑐𝑙𝑓 . We use We define the negative of the lower bound as  (𝐱):
𝐿1 for generative learning with the labeled data, 𝐿2 for generative ∑
 (𝐱) = − 𝑞(𝑦|𝐱)(−(𝐱, 𝑦)) + (𝑞(𝑦|𝐱)) (8)
learning with the unlabeled data, and 𝐿𝑐𝑙𝑓 for discriminative learning
𝑦
to measure the classification performance of our approach. In addition,
𝛼 is the coefficient to weigh generative learning and discriminative The calculation of  is partly the same as . In addition, we also
learning. Here, we set need to calculate 𝑞(𝑦|𝐱), which is implemented by the classifier, and the
corresponding entropy  (Cover and Thomas, 2006). Then, we define
𝑛
𝛼=𝛽⋅ 2 (2) 𝐿2 as the sum of all  on the unlabeled dataset:
𝑛1
∑
where 𝑛1 is the size of the labeled dataset, 𝑛2 is the size of the unlabeled 𝐿2 =  (𝐱) (9)
𝐱∈𝐷2
dataset, and 𝛽 is a hyperparameter. In the following, we show the
probability expressions of 𝐿1 , 𝐿2 and 𝐿𝑐𝑙𝑓 . In this way, we use 𝐿2 for generative learning on the unlabeled dataset.
Maximum likelihood estimation (Myung, 2003) is a common tech- Similar to 𝐿1 , it can also be implemented with the encoder and decoder.
nique in deep learning to design loss functions. It maximizes the In addition, the classifier also contributes to 𝐿2 .
likelihood function of observed samples so that the observed samples Next, we use 𝐿𝑐𝑙𝑓 for discriminative learning by measuring the per-
are the most probable. As shown in Kingma and Welling (2013), formance of the classifier on the labeled dataset with the cross-entropy
Kingma et al. (2014), the log-likelihood of a single labeled sample in loss:
the labeled dataset has a variational bound, as follows:
𝐿𝑐𝑙𝑓 = E𝐱∈𝐷1 [− log 𝑞(𝑦|𝐱)] (10)
log 𝑝(𝐱, 𝑦) ≥ E𝑞(𝐳|𝐱,𝑦) [log 𝑝(𝐱|𝑦, 𝐳) + log 𝑝(𝑦) + log 𝑝(𝐳) − log 𝑞(𝐳|𝐱, 𝑦)] (3)
To summarize, we use 𝐿1 , 𝐿2 and 𝐿𝑐𝑙𝑓 to perform generative learn-
where log 𝑝(𝐱|𝑦, 𝐳) is used to reconstruct the original input 𝐱 with the ing and discriminative learning simultaneously. By using unlabeled
corresponding label 𝑦 and the latent variable 𝐳, log 𝑝(𝑦) is used to data for generative learning, discriminative learning is promoted to
measure the label distribution, and log 𝑝(𝐳) − log 𝑞(𝐳|𝐱, 𝑦) is used to learn improve classification performance.
the latent variable 𝐳 from 𝐱 and the corresponding label 𝑦, and make
the latent variable 𝐳 follow a certain distribution, such as Gaussian 4.4. Network
distribution. By using the latent variable 𝐳, this variational bound can
effectively perform generative learning with the labeled data. To implement our algorithm for generative learning and discrimi-
Since the input data is high-dimensional, it is difficult to directly native learning, we design three modules: the classifier, the encoder,
optimize the likelihood function for deep learning. Here, we use the and the decoder, which are used to learn 𝑞(𝑦|𝐱), 𝑞(𝐳|𝐱, 𝑦), and 𝑝(𝐱|𝑦, 𝐳),
variational bound as our objective function. In addition, in deep learn- respectively. The network architectures of these three modules are all
ing, the loss function is in the minimization form. We define the based on the convolutional neural network (CNN) (LeCun et al., 1989,
negative of the variational bound as (𝐱, 𝑦): 1998, 2015), which is an artificial neural network using convolution
kernels to extract features from the input data. There are two main
(𝐱, 𝑦) = −E𝑞(𝐳|𝐱,𝑦) [log 𝑝(𝐱|𝑦, 𝐳) + log 𝑝(𝑦) + log 𝑝(𝐳) − log 𝑞(𝐳|𝐱, 𝑦)] (4) reasons for this setting: (1) The extracted trajectory data in this paper
is two-dimensional, and CNN can effectively extract feature repre-
Then, we define the sum of the variational bound of all the samples
sentations from the two-dimensional data. (2) Compared with other
in the labeled dataset as 𝐿1 :
feedforward neural networks, CNN requires fewer parameters, making
∑
𝐿1 = (𝐱, 𝑦) (5) it an attractive substructure for deep learning (Li et al., 2019). In the
(𝐱,𝑦)∈𝐷1 following, we describe the CNN structures of the classifier, the encoder,
In fact, we need to give the specific forms of the probability expressions and the decoder in detail.
in 𝐿1 to use 𝐿1 . In Section 4.4, we will introduce the structures of The classifier module is used to learn 𝑞(𝑦|𝐱) to classify vessel tra-
the encoder and decoder to effectively characterize 𝐿1 for generative jectories. Its input is the processed trajectory in the seven-hot form,
learning. and the output is the probability of each ship type. We firstly feed the
trajectory into five convolutional layers. The number of input channels
Similarly, we show how to perform generative learning using the
of each layer is 1, 5, 5, 5, and 5, and the corresponding kernel size of
unlabeled data. The log likelihood of a single unlabeled sample in 𝐷2 ,
each layer is 10, 10, 10, 5, and 3, respectively. Each convolutional layer
log 𝑝(𝐱), also has a variational bound (Kingma et al., 2014):
is followed by a ReLU function. Then, the output of those convolutional
log 𝑝(𝐱) ≥ E𝑞(𝑦,𝐳|𝐱) [log 𝑝(𝐱|𝑦, 𝐳) + log 𝑝(𝑦) + log 𝑝(𝐳) − log 𝑞(𝑦, 𝐳|𝐱)] (6) layers is flattened into a feature vector with size 250 and sent into a
∑ fully connected layer. Finally, we design a softmax layer after the fully
= 𝑞(𝑦|𝐱)(−(𝐱, 𝑦)) + (𝑞(𝑦|𝐱)) (7)
𝑦
connected layer to generate the probability of each class.

5
H. Duan et al. Ocean and Coastal Management 218 (2022) 106015

Algorithm 1 Training Process of Our Semi-supervised Learning Table 1

Algorithm for Vessel Trajectory Classification (SSL-VTC). The numbers of trajectories of different categories of the processed dataset.
Vessel Type Fishing Passenger Cargo Tanker
Input: 𝐸: The number of epochs.
Initialize the parameters of the classifier, encoder, and decoder. Training Set 1436 13678 37842 17335
Validation Set 479 4349 11808 5291
for 𝑖 = 1 to 𝐸 do
Test Set 786 5030 12263 5223
Draw a mini-batch from the labeled dataset 𝐷1 .
Calculate the loss function 𝐿1 using Eq. (5).
Calculate the loss function 𝐿𝑐𝑙𝑓 using Eq. (10).
Draw a mini-batch from the unlabeled dataset 𝐷2 . gov/ais/), covering the coastal waters of Canada, the United States, and
Calculate the loss function 𝐿2 using Eq. (9). Mexico. We analyze the data from January to June in 2019. Inspired
Calculate the overall loss function 𝐿 using Eq. (1). by previous works (Krüger, 2019; Wang et al., 2019; Zhang et al., 2019;
Perform gradient descent to update the parameters of the classifier, Li et al., 2020; Nguyen et al., 2021), we split all the original AIS data
encoder, and decoder. into three parts: training set, validation set, and test set, corresponding
end for to the data collected from January to April, May and June, respectively.
The reason behind these settings is that the historical AIS data is used
as the training set to train our model, which can help us perform
the predictive classification task based on the current or future data.
The encoder module is used to learn 𝑞(𝐳|𝐱, 𝑦), i.e., capture the latent The validation set is used to select the suitable model trained on the
variable 𝐳 for each trajectory 𝐱. If the input trajectory 𝐱 is labeled, training set, while the test set is used to evaluate the performance of
we send it to the encoder together with the corresponding label 𝑦 the final model trained on the training set. By using the validation
to enhance the expression ability of 𝐳. If the input trajectory 𝐱 is set, the evaluation results are ensured to be unbiased (Tennenholtz
unlabeled, we pair it with every possible label and send them to the et al., 2018). Finally, the sizes of the training set, validation set and
encoder. Note that the input label is in the one-hot format. For the input test set are 168.3 GB, 52.6 GB, and 56.0 GB, respectively. We classify
trajectory, we feed it into five convolutional layers, which have the trajectories into four categories, including fishing, passenger, cargo,
same configuration as the convolutional block in the above classifier. and tanker. The dataset is processed as shown in Section 3. The regions
Then, we flatten the features produced from the convolutional layers covered by the AIS dataset and processed AIS trajectories are shown
into a feature vector with a size of 250. For the input label, we send in Fig. 6. The number of trajectories in each category is displayed in
it into a fully connected layer to generate a feature vector with 50 Table 1. It can be observed that the number of cargo vessels is the most,
nodes. Then, we concatenate these two vectors and send them into the number of fishing vessels is the least, and the number of tankers is
two fully connected layers to obtain variables 𝝁 and 𝝈. By using the slightly greater than the number of passenger vessels.
reparameterization technique (Kingma and Welling, 2013; Doersch,
2016; Blei et al., 2017) with 𝝁 and 𝝈, we can obtain the latent variable 5.2. Experimental setting
𝐳.
The decoder module is used to learn 𝑝(𝐱|𝑦, 𝐳), i.e., generate the In each experiment, we use the training set to train our framework,
trajectory 𝐱 with the latent variable 𝐳. Similar to the encoder, the label and use the validation set to select the best model and report its
information is also used to improve the generation performance of the accuracy on the test set as the final result. The model is trained using
decoder. Firstly, we concatenate the latent variable 𝐳 and the label 𝑦 the Adam optimizer (Kingma and Ba, 2014). In addition, we set the
in one-hot form. Then, we send them into two fully connected layers learning rate, batch size, and number of epochs to 0.0001, 100, and 50,
to generate a feature vector with size 250. After reshaping, we feed respectively. We perform a series of experiments with PyTorch (Paszke
it into five deconvolutional layers to generate the original trajectory. et al., 2019) on an NVIDIA TITAN V GPU card.
The number of output channels in each layer is 5, 5, 5, 5, and 1, and
the corresponding kernel size in each layer is 3, 5, 10, 10, and 10, 5.3. Experimental results
respectively. Each of the first four deconvolutional layers is followed
by a ReLU function, and the last deconvolutional layer is followed by In this section, we first show that our framework learns feature
a sigmoid function. representations effectively for vessel trajectory classification, then show
In this way, we show how to use the classifier, encoder, and decoder the necessity of combining static information, next show that our
to implement our whole architecture for semi-supervised learning. The method can efficiently leverage the knowledge of the unlabeled data in
algorithm of our approach is shown in Algorithm 1. Based on the different semi-supervised settings, and then investigate the influence
variational autoencoder, our semi-supervised learning algorithm can of hyperparameter 𝛽 in our semi-supervised learning algorithm, and
fully leverage the labeled and unlabeled data to perform discrimina- finally study the scenario where static information may be missing.
tive learning and generative learning simultaneously. The classification
performance of the proposed framework will be verified in the next 5.3.1. Validation of the proposed neural network and seven-hot encoding
section. In order to show the effectiveness of the convolutional neural
network and seven-hot encoding, we compare our classifier with other
5. Experiments machine learning and deep learning methods, including support vector
machine (SVM) (Cortes and Vapnik, 1995), decision tree (DT) (Wu
In this section, we first illustrate the public AIS dataset used in this et al., 2008), K-nearest network (KNN) (Altman, 1992), multilayer
study and the experimental setting, and then display and analyze the perceptron (Haykin, 1994), convolutional neural network (LeCun et al.,
extensive experimental results to show that our SSL-VTC approach can 1989, 1998, 2015) which directly use the real values of kinematic
learn good feature representations from trajectories of different ship and static information, and multilayer perceptron using the seven-hot
categories for classification. method. Note that SVM, DT, and KNN are all traditional machine
learning methods that can perform the classification task. Besides,
5.1. Dataset the multilayer perceptron is also a deep learning-based approach that
consists of several fully connected layers. Unlike the convolutional
The original AIS dataset used in this study is provided by the U.S. neural network, the multilayer perceptron has no convolutional layers.
Coast Guard Navigation Center (available at https://marinecadastre. The classification accuracies of different methods are shown in Table 2.

6
H. Duan et al. Ocean and Coastal Management 218 (2022) 106015

Fig. 6. The region covered by the used AIS dataset, in which the blue lines represent the processed AIS trajectories.

Table 2 Table 3
Performance comparison of our approach with different methods. Ablation study of the static information.
Method Accuracy(%) Method Accuracy(%)
Support Vector Machine 78.50 Without Length, Width, Draft (WO-LWD) 71.43
Decision Tree 77.62 Without Length (WO-LEN) 80.54
K-Nearest Network 81.25 Without Width (WO-WID) 88.48
Multilayer Perceptron 81.80 Without Draft (WO-DRA) 87.96
Convolutional Neural Network 85.08 Full (SSL-VTC) 92.22
Multilayer Perceptron with Seven-hot 89.61
Convolutional Neural Network with Seven-hot (SSL-VTC) 92.22
Table 4
Performance comparison between our method and the baseline method in different
semi-supervised learning setting.
From Table 2, we make the following observations: (1) Our pro- Percentage of labeled data 5% 20% 40% 60%
posed convolutional neural network with seven-hot encoding achieves Baseline (WO-LWD) 56.36 65.02 67.15 69.88
the highest accuracy. (2) In the scenario with or without seven-hot Baseline 72.60 80.59 86.13 89.35
encoding, the convolutional neural network performs better than the SSL-VTC (WO-LWD) 59.74 66.83 68.08 70.25
SSL-VTC 77.93 84.51 88.96 90.01
multi-layer perceptron, which shows that the convolutional neural
network has a stronger discrimination ability by learning better feature
representations from the AIS trajectory data. (3) Compared with
deep learning methods without seven-hot technique, the accuracy of the information of length, width, and draft significantly improves the
the seven-hot technique is higher, which demonstrates the benefits of classification performance. In addition, by comparing the performance
seven-hot encoding by enabling our architecture to better capture the of the WO-LEN, WO-WID, and WO-DRA methods, we find that WO-LEN
spatio-temporal information of trajectories. (4) The accuracies of the has the lowest accuracy and WO-WID has the highest accuracy, which
traditional machine learning methods (SVM, DT, KNN) are significantly indicates that different static information contributes differently to the
lower than those of deep learning methods, which verifies the superior- classification performance. The length information has the greatest
ity of deep learning. To summarize, by comparison with other methods, contribution, which implies that the length information may be more
we not only show the benefit of seven-hot encoding, but also present distinguishable than the width and draft (Zhong et al., 2019).
that the feature representation using our convolutional neural network
has sufficient discriminant ability.
5.3.3. Performance of our semi-supervised deep learning approach
5.3.2. Ablation study of the static information In the above, we discuss the effectiveness of our network and the
To illustrate the necessity and importance of combining the static necessity of incorporating static information, which lay the foundation
information for vessel trajectory classification, we compare our classi- for our whole semi-supervised learning architecture. In the following,
fier with the classifier that does not combine the length information we conduct a series of semi-supervised learning experiments to verify
(WO-LEN), the classifier that does not combine the width information that our framework can effectively combine the information of the
(WO-WID), the classifier that does not combine the draft information unlabeled data for vessel trajectory classification. We set 5%, 20%,
(WO-DRA), and the classifier that does not combine the information of 40%, and 60% of the training set as the labeled dataset and the
length, width, and draft (WO-LWD). For a fair comparison, we make corresponding remaining training data as the unlabeled dataset. In each
different methods have the same network structure. The accuracy of semi-supervised learning setting, we compare the performance of our
different methods is shown in Table 3. method with the baseline method which only uses the labeled data. In
It can be found that the accuracy of the method without static infor- addition, we also show the performance of these two methods without
mation is lower than that of our method with static information. The combining static information of length, width, and draft (WO-LWD).
WO-LWD method has the lowest accuracy, indicating that combining The experimental results are shown in Table 4.

7
H. Duan et al. Ocean and Coastal Management 218 (2022) 106015

Fig. 7. Trends of test accuracies of our SSL-VTC method and the baseline method.

It can be observed that: (1) For each percentage of labeled data, Table 5
The values of 𝛽 for different percentage of labeled data.
the classification accuracy using our semi-supervised learning method
Percentage of labeled data 5% 20% 40% 60%
is higher than that using the baseline method, which shows the ef-
fectiveness of our semi-supervised learning framework based on the 𝛽 10 10 100 1000

variational autoencoder. (2) With the increase in the percentage of

unlabeled data, the accuracies of our method and baseline method Table 6
decrease, but the gap between them increases. This indicates that Influence of 𝛽 in our semi-supervised learning algorithm when only 40% of training
data is labeled.
the scarcer the labeled data, the more effective our semi-supervised
𝛽 1 10 100 1000
learning method improves the classification performance by exploit-
ing the unlabeled data, which is consistent with other works about Accuracy 75.02 84.44 88.96 88.94

semi-supervised learning (Berthelot et al., 2019; Ma et al., 2020).

(3) Regardless of our method or the baseline method, the accuracy
without combining the static information is lower than that with com- 5.3.4. Influence of 𝛽 in our semi-supervised learning algorithm
bining the static information, which further demonstrates that combin- In Section 5.3.3, we show the performance of our approach when
ing the static information is beneficial to improving the classification the percentage of the labeled data is set to 5%, 20%, 40%, and 60%. It is
performance. worth noting that to implement the semi-supervised algorithm, we first
We next show trends of test accuracies of our method and the need to determine the value of 𝛼, as mentioned in Eq. (2) in Section 4.3.
𝑛
baseline method during the training process in two scenarios: 20% of 𝛼 equals 𝛽 ⋅ 𝑛2 , where 𝑛1 is the size of the labeled dataset, 𝑛2 is the
1
the training data is labeled and 40% of the training data is labeled. size of the unlabeled dataset, and 𝛽 is a hyperparameter. As we already
The results are displayed in Fig. 7. We can observe that in these two know the percentage of the labeled data, that is, the ratio of 𝑛2 to 𝑛1 ,
scenarios, the test accuracy of our method increases faster than the determining the value of 𝛼 is equivalent to determining the value of
test accuracy of the baseline method, which shows that our method 𝛽. Here, we use grid search (LaValle et al., 2004; Syarif et al., 2016)
can significantly improve the classification performance by effectively to determine 𝛽 for each percentage of the labeled data, as shown in
using the unlabeled data. By comparing the performance when 20% Table 5. It can be found that different semi-supervised settings have
of the training data is labeled with the performance when 40% of the different appropriate values of 𝛽.
training data is labeled, we again find that when the proportion of To explain this phenomenon, we investigate the impact of 𝛽 in our
labeled data becomes lower, our method has more obvious advantages semi-supervised learning algorithm for vessel trajectory classification
than the baseline method. when 40% of the training data is labeled. We set the value of 𝛽 to 1,
10, 100, and 1000, and make other settings the same as in Section 5.3.3.
Then, in the scenarios where 20% of the training data is labeled and
The experimental results are shown in Table 6.
40% of the training data is labeled, we display the confusion matrices
It can be seen that the classification accuracy increases rapidly as 𝛽
of our method and the baseline method respectively, as shown in Figs. 8
increases from 1 to 100. However, when 𝛽 increases from 100 to 1000,
and 9. We can see that when 20% of the training data is labeled, the
the classification accuracy does not change significantly, indicating that
classification accuracy of each class using our method is higher than
the value 100 of 𝛽 is appropriate for vessel trajectory classification
that of the corresponding class using the baseline method. When 40%
when only 40% of training data is labeled. These observations explain
of the training data is labeled, the classification accuracy of fishing that for a certain percentage of labeled data, we need to select an
using our method is lower than that using the baseline method, but appropriate value for 𝛽 to achieve good classification performance.
the overall classification accuracy of our method is still higher than They are also what we are trying to do by using grid search to select 𝛽.
that of the baseline method. This demonstrates that our approach can
effectively capture clues of most classes for vessel trajectory classifi- 5.3.5. Performance with missing static information
cation, but the baseline method cannot. In addition, we can further In the above, we study the label missing of the training data. At
find that the classification accuracy of different categories is different, the same time, we also show that the static information is useful
which shows that in our deep learning model, different categories have for vessel trajectory classification. However, from the original AIS
different discrimination abilities. It is worth noting that with the help of messages, we find that the vessel static information is not always
deep learning, we do not need to manually set parameters for different available. This phenomenon is called data missing, which often exists
categories, but let our model automatically update the corresponding in raw data (Scheffer, 2002; Enders, 2010; Little and Rubin, 2019).
parameters for each category. Motivated by this, in this section, we study the scenario in which the

8
H. Duan et al. Ocean and Coastal Management 218 (2022) 106015

Fig. 8. The confusion matrices of the baseline method and our SSL-VTC method when only 20% of the training data is labeled.

Fig. 9. The confusion matrices of the baseline method and our SSL-VTC method when only 40% of the training data is labeled.

static information is missing when only 20% of the training data is Table 7
The performance comparison with different percentages of available static
labeled. We assume that only a small part of the static information is
information.
available, and the remaining information is unknowable. We consider
Percentage of available static information 5% 20% 60%
two methods for dealing with the missing static information: one is to
Operation Zero Mean Zero Mean Zero Mean
directly set the static information to zero for the trajectory that cannot
access the static information, which is a common way to handle data Baseline 66.02 68.63 68.64 70.12 80.04 80.76
SSL-VTC 67.39 70.37 76.08 77.68 81.59 82.66
missing. Another method is to fill the unknown static information with
the mean of the available static information. This method leverages
the knowledge of the available static information. We set the static
information of 5%, 20%, and 60% of the training data to be available. missing static information plays a lesser role. Therefore, regardless of
In each setting, we compare the performance of our method and the which operation is used, the accuracy gap between these approaches
baseline method. The experimental results are shown in Table 7. decreases with the increase in available static information. (3) The
From Table 7, we make the following observations: (1) For each trend of the baseline method is consistent with that of our method. But
proportion of available static information, the classification accuracy for a certain setting, the accuracy of the baseline method is lower than
using our method with mean operation is higher than that with zero that of our method, which further verifies our observations above and
operation, which shows the effectiveness of our proposed mean op- the superiority of our framework.
eration. (2) For our approach, the accuracy with the zero operation
or the mean operation increases as the proportion of available static 6. Implications
information increases, while the accuracy gap between these two op-
erations decreases. This can be explained by the fact that when the The approach and findings of this study provide important im-
missing static information decreases, the existing static information plications for researchers and stakeholders in the maritime field. For
may already be sufficient for vessel trajectory classification, and the researchers, our work shows the following two highlights: (1) The

9
H. Duan et al. Ocean and Coastal Management 218 (2022) 106015

approach in this study mainly consists of two parts: vessel trajectory 7. Conclusions
extraction and semi-supervised learning algorithm, which can empower
relevant research directions using the AIS data, such as trajectory In this paper, we propose SSL-VTC, a semi-supervised deep learning
prediction and analysis, anomaly detection, and collision avoidance. approach for vessel trajectory classification based on AIS historical
Besides, it is noteworthy that this study is the first to consider the data for the first time. Specifically, we extract vessel trajectories by
combination of static and kinematic information for vessel trajectory leveraging the kinematic and static information of AIS messages, and
extraction. In our study, the classification accuracy considering both feed them into our convolutional neural network to obtain feature
static and kinematic information is 29.1% higher than that considering representations. Then, we use these feature representations for semi-
only kinematic information. This illustrates the significance of ves- supervised learning based on the variational autoencoder to perform
sel static information for trajectory classification, and further inspires discriminative learning and generative learning simultaneously, which
researchers to apply the static information to other AIS trajectory enables our architecture to make full use of the unlabeled data for
analysis studies. (2) The classification results of ship types in this paper vessel trajectory classification. In the extensive experiments, based on
the public AIS dataset, we first verify the effectiveness of the proposed
can provide support for studies related to vessel identification. Some
neural network and seven-hot encoding by comparing with other ma-
previous works reveal that the analysis using the AIS data can be
chine learning and deep learning methods. Then, through the ablation
enhanced by considering ship types, such as (a) evaluating the coastal
study of static information, we illustrate the necessity and significance
emission policies for vessel evasion behaviors (Tan et al., 2021), (b)
of combining the AIS data static information. Next, we verify the
tracing illegal oil discharges from vessels (Liu et al., 2021b), and (c)
superiority and robustness of the proposed semi-supervised approach
identifying whether to use clean energy or green technology (Bai et al.,
for vessel trajectory classification by varying the proportion of the
2021). By using the classification results of our approach, this study
labeled data, and perform sensitivity analysis on the hyperparameter 𝛽
can provide reference and assistance for the above studies.
to investigate its impact on vessel trajectory classification. In addition,
For stakeholders in the maritime industry, this study provides the we also consider the missing static information in the raw dataset and
following four management implications: (1) This work contributes to propose zero operation and mean operation to further verify the superi-
the management of fishing activities and the conservation of marine ority of our approach. Finally, this study provides important managerial
ecosystems. Illegal, unreported and unregulated (IUU) fishing is one of implications for researchers and stakeholders in the maritime field.
the most serious threats to the sustainability of fisheries worldwide, In future research, the following directions will be considered.
and the stability and balance of marine ecosystems (Arasteh et al., Firstly, in addition to the AIS system, synthetic aperture radar (SAR) is
2020; Cánovas-Molina et al., 2021). Maritime regulatory authorities also a widely used message transmission system for marine supervision
and researchers have devoted increasing efforts to regulating fish- and fishery management (Hou et al., 2020; Barreto et al., 2021; Liu
ing activities, and fisheries management is one of the latest research et al., 2021b). SAR has the capability of all-time imaging in all-weather
hotspots (Jiang et al., 2016; Krüger, 2019; Garcia et al., 2021; Neto conditions. It will be a promising and attractive research direction to
et al., 2021; Sultan, 2021; Yu and Wang, 2021). By identifying the better carry out vessel trajectory classification tasks by combining the
fishing vessels using our trajectory classification framework, this work advantages of the AIS data and the SAR data. Secondly, we investigate
can contribute to monitoring illegal fishing activities to protect the the scenario of missing static information in the actual AIS dataset and
ecosystem (Kularatne, 2020; Warren and Steenbergen, 2021). More- propose two operations to address this problem in this paper. We can
over, by extracting, analyzing and mapping the footprint of fishing design more machine learning or deep learning techniques to address
vessels (i.e., fishing area and intensity), this study can provide guid- the static information missing issue to further improve the performance
ance and suggestions for policy-makers in fishery management. (2) of our proposed approach. Thirdly, in the actual navigation environ-
Apart from identifying fishing vessels, this work can also assist in ment, vessel trajectories are affected not only by their own factors
the identification and management of other ship types. By identifying (e.g., speed, course) recorded in the AIS data, but also by external
tankers, it is possible to estimate the environmental risk of oil spill factors (e.g., wind, current) (Yang et al., 2020). Inspired by this, we
accidents and drift grounding accidents for oil tankers (Yang et al., can comprehensively analyze the AIS data and the meteorological and
2019; Barreto et al., 2021; Liu et al., 2021b; Zhang et al., 2021b). By hydrological data to investigate the interaction between internal and
external factors, and provide managerial insights for vessel trajectory
identifying cargo ships, this work can be applied to track cargo flows
analysis and safe navigation.
and provide a reference for port trade analysis (Jia et al., 2019; Xu
et al., 2021). By identifying passenger ships, this work is capable of
Declaration of competing interest
offering information for maritime search and rescue operations. (3) This
study helps to identify abnormal ship behaviors and detect dangerous
The authors declare that they have no known competing finan-
navigation areas to safeguard shipping safety. For example, if a vessel
cial interests or personal relationships that could have appeared to
declares itself as type ‘‘A’’ but performs a maneuver of type ‘‘B’’, it
influence the work reported in this paper.
is likely that it may be carrying out illegal activities, which is also
known as the false ship effect (Ljunggren, 2018; Nguyen et al., 2018). In Acknowledgments
our architecture, ship types are identified by vessel trajectories, which
enables our framework to be utilized as an important tool to detect The authors thank the editor and referees for their helpful comments
illegal, suspicious or unsafe behaviors of vessels. Furthermore, it can be and suggestions. This work is supported by Natural Science Foundation
used to alert the surrounding area of anomalous vessels to emphasize of Guangdong Province, China, under Grant No. 2021B1515020059
the navigation safety of neighboring vessels. (4) This study can provide and National Natural Science Foundation of China under Grant No.
predictive guidance to shipping regulators on maritime surveillance. 71872092.
In this paper, our model is trained using five months of historical AIS
data and tested on the sixth month of AIS data, which is essentially a References
predictive classification task. By using our framework to make predic-
tions for future AIS data, our work can assist in the predictive analysis Almpanidou, V., Doxa, A., Mazaris, A.D., 2021. Combining a cumulative risk index
and species distribution data to identify priority areas for marine biodiversity
of future ocean and coastal developments, thus providing predictive
conservation in the black sea. Ocean Coastal Manag. 213, 105877.
planning and management guidance (Herrero et al., 2019; Yang et al., Altman, N.S., 1992. An introduction to kernel and nearest-neighbor nonparametric
2019; McWhinnie et al., 2021; Meyers et al., 2021). regression. Amer. Statist. 46 (3), 175–185.

10
H. Duan et al. Ocean and Coastal Management 218 (2022) 106015

Alves, C., 2021. Marine resource management and fisheries governance in Belize exhibit LaValle, S.M., Branicky, M.S., Lindemann, S.R., 2004. On the relationship between
a polycentric, decentralized, and nested institutional structure. Ocean Coastal classical grid search and probabilistic roadmaps. Int. J. Robot. Res. 23 (7–8),
Manag. 211, 105742. 673–692.
Arasteh, S., Tayebi, M.A., Zohrevand, Z., Glässer, U., Shahir, A.Y., Saeedi, P., Wehn, H., Le Tixerant, M., Le Guyader, D., Gourmelon, F., Queffelec, B., 2018. How can automatic
2020. Fishing vessels activity detection from longitudinal AIS data. In: Proceedings identification system (AIS) data be used for maritime spatial planning? Ocean
of the 28th International Conference on Advances in Geographic Information Coastal Manag. 166, 18–30.
Systems. pp. 347–356. LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521 (7553), 436–444.
Bai, X., Hou, Y., Yang, D., 2021. Choose clean energy or green technology? Empirical LeCun, Y., Boser, B., Denker, J.S., Henderson, D., Howard, R.E., Hubbard, W.,
evidence from global ships. Transp. Res. Part E: Logist. Transp. Rev. 151, 102364. Jackel, L.D., 1989. Backpropagation applied to handwritten zip code recognition.
Barreto, F.T., Dammann, D.O., Tessarolo, L.F., Skancke, J., Keghouche, I., Inno- Neural Comput. 1 (4), 541–551.
centini, V., Winther-Kaland, N., Marton, L., 2021. Comparison of the coupled LeCun, Y., Bottou, L., Bengio, Y., Haffner, P., 1998. Gradient-based learning applied to
model for oil spill prediction (CMOP) and the oil spill contingency and response document recognition. Proc. IEEE 86 (11), 2278–2324.
model (OSCAR) during the DeepSpill field experiment. Ocean Coastal Manag. 204, Li, C., Bai, L., Liu, W., Yao, L., Waller, S.T., 2020. Graph neural network for robust
105552. public transit demand prediction. IEEE Trans. Intell. Transp. Syst..
Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A., 2019. Li, H., Hu, M., Huang, Y., 2019. Automatic identification of overpass structures: A
MixMatch: A holistic approach to semi-supervised learning. Adv. Neural Inf. method of deep learning. ISPRS Int. J. Geo-Inf. 8 (9), 421.
Process. Syst. 32. Little, R.J., Rubin, D.B., 2019. Statistical Analysis with Missing Data, Vol. 793. John
Blei, D.M., Kucukelbir, A., McAuliffe, J.D., 2017. Variational inference: A review for Wiley & Sons.
statisticians. J. Amer. Statist. Assoc. 112 (518), 859–877. Liu, K., Yu, Q., Yuan, Z., Yang, Z., Shu, Y., 2021a. A systematic analysis for maritime
Cánovas-Molina, A., García-Charton, J.A., García-Frapolli, E., 2021. Assessing the accidents causation in Chinese coastal waters using machine learning approaches.
contribution to overfishing of small-and large-scale fisheries in two marine regions Ocean Coastal Manag. 213, 105859.
as determined by the weight of evidence approach. Ocean Coastal Manag. 213, Liu, B., Zhang, W., Han, J., Li, Y., 2021b. Tracing illegal oil discharges from vessels
105911. using SAR and AIS in bohai sea of China. Ocean Coastal Manag. 211, 105783.
Chapelle, O., Scholkopf, B., Zien, A., 2009. Semi-supervised learning (chapelle, o. et Ljunggren, H., 2018. Using deep learning for classifying ship trajectories. In: 2018 21st
al., eds.; 2006)[book reviews]. IEEE Trans. Neural Netw. 20 (3), 542. International Conference on Information Fusion. FUSION, IEEE, pp. 2158–2164.
Chen, X., Liu, Y., Achuthan, K., Zhang, X., 2020. A ship movement classification based Ma, F., Zhang, W., Li, Y., Huang, S.-L., Zhang, L., 2020. Learning better representations
on automatic identification system (AIS) data using convolutional neural network. for audio-visual emotion recognition with common information. Appl. Sci. 10 (20),
Ocean Eng. 218, 108182. 7239.
Cominelli, S., Halliday, W.D., Pine, M.K., Hilliard, R.C., Lawson, J.W., Duman, N.I., Mantecón, T., Casals, D., Navarro-Corcuera, J.J., del Blanco, C.R., Jaureguizar, F., 2019.
Devillers, R., 2020. Vessel noise in spatially constricted areas: Modeling acoustic Deep learning to enhance maritime situation awareness. In: 2019 20th International
footprints of large vessels in the Cabot Strait, Eastern Canada. Ocean Coastal Radar Symposium. IRS, IEEE, pp. 1–8.
Manag. 194, 105255. McWhinnie, L.H., O’Hara, P.D., Hilliard, C., Le Baron, N., Smallshaw, L., Pelot, R.,
Cortes, C., Vapnik, V., 1995. Support-vector networks. Mach. Learn. 20 (3), 273–297. Canessa, R., 2021. Assessing vessel traffic in the salish sea using satellite AIS: An
Cover, T.M., Thomas, J.A., 2006. Elements of Information Theory (Wiley Series in important contribution for planning, management and conservation in southern
Telecommunications and Signal Processing). Wiley-Interscience, USA. resident killer whale critical habitat. Ocean Coastal Manag. 200, 105479.
Dobrkovic, A., Iacob, M.-E., Van Hillegersberg, J., 2016. Maritime pattern extraction Meyers, S.D., Luther, M.E., Ringuet, S., Raulerson, G., Sherwood, E., Conrad, K.,
from AIS data using a genetic algorithm. In: 2016 IEEE International Conference Basili, G., 2021. Ship wakes and their potential shoreline impact in Tampa Bay.
on Data Science and Advanced Analytics. DSAA, IEEE, pp. 642–651. Ocean Coastal Manag. 211, 105749.
Doersch, C., 2016. Tutorial on variational autoencoders. arXiv preprint arXiv:1606. Myung, I.J., 2003. Tutorial on maximum likelihood estimation. J. Math. Psych. 47 (1),
05908. 90–100.
Enders, C.K., 2010. Applied Missing Data Analysis. Guilford Press. Neto, J.B.G., de Alencar Goyanna, F.A., Feitosa, C.V., Soares, M.O., 2021. A sleeping
Garcia, S.G., Barclay, K., Nicholls, R., 2021. Can anti-illegal, unreported, and unregu- giant: the historically neglected Brazilian fishing sector. Ocean Coastal Manag. 209,
lated (IUU) fishing trade measures spread internationally? Case study of Australia. 105699.
Ocean Coastal Manag. 202, 105494. Nguyen, D., Vadaine, R., Hajduch, G., Garello, R., Fablet, R., 2018. A multi-task deep
Goodfellow, I., Bengio, Y., Courville, A., Bengio, Y., 2016. Deep learning, Vol. 1. (2), learning architecture for maritime surveillance using ais data streams. In: 2018
MIT press Cambridge. IEEE 5th International Conference on Data Science and Advanced Analytics. DSAA,
Guo, S., Mou, J., Chen, L., Chen, P., 2021. Improved kinematic interpolation for AIS IEEE, pp. 331–340.
trajectory reconstruction. Ocean Eng. 234, 109256. Nguyen, D., Vadaine, R., Hajduch, G., Garello, R., Fablet, R., 2021. GeoTrackNet–a
Harati-Mokhtari, A., Wall, A., Brooks, P., Wang, J., 2007. Automatic identification maritime anomaly detector using probabilistic neural network representation of
system (AIS): data reliability and human error implications. J. Navig. 60 (3), 373. AIS tracks and a contrario detection. IEEE Trans. Intell. Transp. Syst..
Haykin, S., 1994. Neural Networks: A Comprehensive Foundation. Prentice Hall PTR. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T.,
Herrero, D.A., Pedroche, D.S., Herrero, J.G., López, J.M.M., 2019. Ais trajectory Lin, Z., Gimelshein, N., Antiga, L., et al., 2019. Pytorch: An imperative style,
classification based on IMM data. In: 2019 22th International Conference on high-performance deep learning library. arXiv preprint arXiv:1912.01703.
Information Fusion. FUSION, IEEE, pp. 1–8. Scheffer, J., 2002. Dealing with missing data. Massey University.
Hou, X., Ao, W., Song, Q., Lai, J., Wang, H., Xu, F., 2020. FUSAr-ship: building a high- Sheng, K., Liu, Z., Zhou, D., He, A., Feng, C., 2018. Research on ship classification
resolution SAR-ais matchup dataset of gaofen-3 for ship detection and recognition. based on trajectory features. J. Navig. 71 (1), 100.
Sci. China Inf. Sci. 63 (4), 1–19. Sultan, R., 2021. Fisheries benefits of a marine protected area with endogenous fishing
Hu, B., Jiang, X., de Souza, E.N., Pelot, R., Matwin, S., 2016. Identifying fishing efforts–a bioeconomic analysis. Ocean Coastal Manag. 207, 105594.
activities from AIS data with conditional random fields. In: 2016 Federated Svanberg, M., Santén, V., Hörteborn, A., Holm, H., Finnsgård, C., 2019. AIS in maritime
Conference on Computer Science and Information Systems. FedCSIS, IEEE, pp. research. Mar. Policy 106, 103520.
47–52. Syarif, I., Prugel-Bennett, A., Wills, G., 2016. SVM parameter optimization using grid
Jia, H., Prakash, V., Smith, T., 2019. Estimating vessel payloads in bulk shipping using search and genetic algorithm to improve classification performance. Telkomnika 14
AIS data. Int. J. Shipp. Transp. Logist. 11 (1), 25–40. (4), 1502.
Jiang, X., Silver, D.L., Hu, B., de Souza, E.N., Matwin, S., 2016. Fishing activity Tan, Z., Liu, H., Shao, S., Liu, J., Chen, J., 2021. Efficiency of Chinese ECA policy on
detection from ais data using autoencoders. In: Canadian Conference on Artificial the coastal emission with evasion behavior of ships. Ocean Coastal Manag. 208,
Intelligence. Springer, pp. 33–39. 105635.
Kingma, D.P., Ba, J., 2014. Adam: A method for stochastic optimization. arXiv preprint Tennenholtz, G., Zahavy, T., Mannor, S., 2018. Train on validation: squeezing the data
arXiv:1412.6980. lemon. arXiv preprint arXiv:1802.05846.
Kingma, D.P., Rezende, D.J., Mohamed, S., Welling, M., 2014. Semi-supervised learning Tu, E., Zhang, G., Rachmawati, L., Rajabally, E., Huang, G.-B., 2017. Exploiting AIS
with deep generative models. arXiv preprint arXiv:1406.5298. data for intelligent maritime navigation: A comprehensive survey from data to
Kingma, D.P., Welling, M., 2013. Auto-encoding variational bayes. arXiv preprint methodology. IEEE Trans. Intell. Transp. Syst. 19 (5), 1559–1582.
arXiv:1312.6114. Van Engelen, J.E., Hoos, H.H., 2020. A survey on semi-supervised learning. Mach.
Kontopoulos, I., Makris, A., Tserpes, K., 2021. A deep learning streaming methodology Learn. 109 (2), 373–440.
for trajectory classification. ISPRS Int. J. Geo-Inf. 10 (4), 250. Wang, X., Li, J., Zhang, T., 2019. A machine-learning model for zonal ship flow
Krüger, M., 2019. Detection of AIS spoofing in Fishery scenarios. In: 2019 22th prediction using AIS data: a case study in the south atlantic states region. J. Mar.
International Conference on Information Fusion. FUSION, IEEE, pp. 1–7. Sci. Eng. 7 (12), 463.
Kularatne, R.K.A., 2020. Unregulated and illegal fishing by foreign fishing boats in Sri Wang, T., Ye, C., Zhou, H., Ou, M., Cheng, B., 2020. AIS ship trajectory clustering
Lankan waters with special reference to bottom trawling in northern Sri Lanka: A based on convolutional auto-encoder. In: Proceedings of SAI Intelligent Systems
critical analysis of the Sri Lankan legislation. Ocean Coastal Manag. 185, 105012. Conference. Springer, pp. 529–546.

11
H. Duan et al. Ocean and Coastal Management 218 (2022) 106015

Warren, C., Steenbergen, D.J., 2021. Fisheries decline, local livelihoods and conflicted Yu, J.-K., Wang, H.-X., 2021. Evolution of distant water fisheries policies in China:
governance: An Indonesian case. Ocean Coastal Manag. 202, 105498. Overview, characteristics and proposals. Ocean Coastal Manag. 207, 105592.
Wu, X., Kumar, V., Quinlan, J.R., Ghosh, J., Yang, Q., Motoda, H., McLachlan, G.J., Zhang, S., Chen, J., Wan, Z., Yu, M., Shu, Y., Tan, Z., Liu, J., 2021a. Challenges
Ng, A., Liu, B., Philip, S.Y., et al., 2008. Top 10 algorithms in data mining. Knowl. and countermeasures for international ship waste management: IMO, China, United
Inf. Syst. 14 (1), 1–37. States, and EU. Ocean Coastal Manag. 213, 105836.
Xu, L., Xie, F., Wang, C., 2021. Passive or proactive capacity sharing? A perspective Zhang, W., Li, C., Chen, J., Wan, Z., Shu, Y., Song, L., Xu, L., Di, Z., 2021b. Governance
of cooperation and competition between two regional ports. Marit Policy Manag. of global vessel-source marine oil spills: Characteristics and refreshed strategies.
1–18. Ocean Coastal Manag. 213, 105874.
Yang, Z., Daamen, W., Vellinga, T., Hoogendoorn, S.P., 2020. Impacts of wind and Zhang, L., Meng, Q., Xiao, Z., Fu, X., 2018. A novel ship trajectory reconstruction
current on ship behavior in ports and waterways: A quantitative analysis based on approach using AIS data. Ocean Eng. 159, 165–174.
AIS data. Ocean Eng. 213, 107774. Zhang, R., Xie, P., Wang, C., Liu, G., Wan, S., 2019. Classifying transportation mode
Yang, D., Wu, L., Wang, S., Jia, H., Li, K.X., 2019. How big data enriches maritime and speed from trajectory data via deep multi-scale learning. Comput. Netw. 162,
research–a critical review of automatic identification system (AIS) data applications. 106861.
Transp. Rev. 39 (6), 755–773. Zhong, H., Song, X., Yang, L., 2019. Vessel classification from space-based ais data using
Yao, D., Zhang, C., Zhu, Z., Huang, J., Bi, J., 2017. Trajectory clustering via random forest. In: 2019 5th International Conference on Big Data and Information
deep representation learning. In: 2017 International Joint Conference on Neural Analytics. BigDIA, IEEE, pp. 9–12.
Networks. IJCNN, IEEE, pp. 3880–3887. Zhu, X., Goldberg, A.B., 2009. Introduction to semi-supervised learning. Synth. Lect.
Artif. Intell. Mach. Learn. 3 (1), 1–130.

12