A Novel Ship Trajectory Clustering Analysis and Anomaly Detection Method Based On AIS Data
A Novel Ship Trajectory Clustering Analysis and Anomaly Detection Method Based On AIS Data
Keywords: Ship trajectory; Minimum description length; Dynamic Time Warping; Transformer
1. Introduction
The rapid development of the automatic identification system (AIS) has enabled the acquisition
of extensive ship motion trajectory data, providing a foundational basis for predicting ship behavior
and monitoring abnormal ship behavior. In the realm of maritime research, ship trajectory clustering
and anomaly detection algorithms have become focal points, particularly with the development of
Internet technology and big data analysis (Yang et al., 2019). The significance of ship trajectory
clustering analysis and anomaly detection lies in advancing shipping management intelligence,
enhancing navigation safety, and enhancing the efficiency of the shipping industry.
Research of anomaly detection was initially conducted using statistical methods (Kowalska
and Peel, 2012), the form of normal trajectory model was a probability model of trajectory point
information, which primarily include Kernel Density Estimation (KDE) (Ristic et al., 2008; Dai et
al., 2020; Wang et al., 2022), Gaussian Mixture Model (GMM) (Laxhammar et al., 2009), Gaussian
Process (GP) (Smith et al., 2012), Hidden Markov Model (HMM) (Shahir et al., 2014), and Bayesian
Network (BN) (Mascaro et al., 2014; Zhen et al. 2017; d’Afflisio et al. 2021). The statistical method
of anomaly detection uses statistical testing to determine whether the behavior of ship matches a
statistical model representing conventional ship behavior. When the matching probability is low, it
is considered as abnormal behavior. The disadvantage of statistical method is that the matching
accuracy is related to historical data and ignores real-time scene. And this method is difficult to
combine multi-source information, expert knowledge, etc., and is not suitable for ship trajectory
anomaly detection in the increasingly complex marine traffic circumstance.
In recent years, much effort has been done to improve the performance of the existing
algorithms to make them applicable for ship trajectory anomaly detection in the increasingly
complex marine traffic circumstance. Clustering, as a tool for big data analysis, is an unsupervised
technique that does not depend on any prior knowledge. A trajectory clustering framework based on
AIS data was designed to analyze routes, which considered the geographic spatial information and
contextual features of ship trajectories, and thereby the density-based clustering algorithm
automatically classified different routes (Sheng and Yin, 2018). A new features of local fast ship
trajectories method was proposed to search for global and local features of ship trajectories (Tang
et al., 2021). Soares et al., (2015) proposed an unsupervised method to segment trajectories without
predetermined clear criteria. The DBSCAN algorithm (Ester et al., 1996) is a pioneer technique in
the context of density-based clustering. Lei (2016) developed a framework called MT-MAD
(maritime trajectory modeling and anomaly detection) to explore frequent movement behaviors and
established a single index for combining anomaly scores to determine the suspicious level of each
ship's trajectory. The DBSCAN approach is used to cluster course over ground (COG) and speed
over ground (SOG) in AIS data, considering both density points and those with similar COGs and
SOGs. To handle large datasets, an improved density-based spatial clustering of applications with
noise (DBSCAN) clustering algorithm was introduced (Nooshin and Hamid, 2022; Li et al., 2020).
Kontopoulos et al. (2021) clustered the track points and extracted the ship’s steering points based
on the DBSCAN algorithm. DBSCAN algorithm was reused to cluster track lines, in which the
Lagrange interpolation algorithm was used to fill the gaps between the steering points. Han et al.
(2021) proposed an optimized DBSCAN algorithm for abnormal ship trajectories detection by
clustering AIS ship trajectories with the same features, and ultimately validated the performance of
this method using data from the Gulf of Mexico. To speed up the computational efficiency of the
clustering algorithm. Yang et al. (2022) proposed Density based Trajectory Clustering of
Applications with Noise (DBTCAN) algorithm. This method uses Hausdorff distance as a similarity
measure to cluster trajectories of different lengths. Therefore, the DBTCAN algorithm can not only
recognize noise trajectories but also adaptively select its optimal input parameters, which can be
widely applied in ocean research.
Rong et al. (2020) employed the DP algorithm to identify turning points based on ship type,
size, final destination, and other maritime traffic patterns. The DBSCAN method clusters turning
points, and then combines the density region of these points by KDE. DBSCAN and the Kernel
Density Estimation-based Outlier Factor processing algorithm was introduced to calculate the
abnormal probability distribution value of trajectory points, eliminating low-probability distribution
edge points (Jin et al., 2023). Bai et al. (2023) designed an adaptive threshold fast DBSCAN
algorithm for vessel trajectory clustering. The fast DTW algorithm is used to reduce the
computational complexity and ensure the accuracy of trajectory similarity, and the DBSCAN
parameters are adaptively determined by combining the similarity distribution of trajectories with
an improved K-adaptive nearest neighbor. However, the method is distance-based trajectory
clustering algorithm and does not consider the speed and acceleration of ship, and thereby attracting
the attention of more and more scholars in the field of artificial intelligence technology.
Detecting anomalies in ship trajectories is a complex and challenging task due to the influence
of the surrounding environment on the navigation statuses of both the own ship and other ships (Liu
et al., 2022). Thus, artificial intelligence technology is increasingly applied in ship trajectory
anomaly detection research. Traditional supervised learning techniques were employed to predict
arrival time and reorganize data based on spatial grids. Deep learning architectures based on long
short-term memory (LSTM) are also explored to address the next position prediction problem
(Gözde et al., 2021). Park et al. (2021) employed the spectral clustering method to cluster ship
trajectories and established a ship trajectory prediction model using bi-directional LSTM (Bi-
LSTM). The support vector machine is employed to predict the ship's next course at the exit of the
traffic route in Tokyo Bay based on dynamic historical AIS data. Nevertheless, greater emphasis
should be placed on enhancing prediction accuracy by increasing efforts (Nishizaki et al., 2018).
Venskus et al. (2017) presented a self-learning adaptive classification method based on the self-
organizing map and virtual pheromone for ship trajectory anomaly detection. By utilizing the
gradient descent algorithm, the verification dataset is used to calculate the pheromone intensity
threshold for trajectory anomaly detection. Venskus et al. (2019) extended their previous work and
studied a data batch processing strategy for neural network retraining to detect anomalies in
streaming maritime traffic data. The results demonstrate that it is possible to reduce the retraining
time while keeping the accuracy relatively unchanged. Mantecón et al. (2019) proposed a supervised
deep learning framework for ship anomaly detection, utilizing a convolutional neural network to
infer navigation states from the ship trajectory based on AIS information. Zhao and Shi (2019)
combined DBSCAN and recurrent neural networks (RNN) to obtain the clustering ship trajectories
and prediction of large-scale ship trajectories. Wen et al. (2020) clustered the ship trajectories based
on DBSCAN algorithm to identify key regions, and then applied artificial neural networks to learn
the relationships between key regions to generate reasonable routes for different ships. However,
these algorithms face challenges related to parameter setting, noise recognition, and sensitivity to
density distribution in datasets. With the development of deep learning technologies, Li et al. (2023)
divided multiple different subsequences by quantifying the similarity of time distribution and
matching the time distribution, and then constructed the adaptive transformer model based on
transfer learning to conduct the accurate prediction of the future trajectory. However, limited by the
long training and learning time of the mentioned models above, the time complexity of parameter
calculation is still relatively high and cannot be used for trajectory prediction of a large amount of
ship data.
Given this background, analyzing and studying ship trajectory clustering analysis and anomaly
detection based on machine learning becomes crucial. The main contributions of this paper are as
follows:
(1) The minimum description length (MDL) criterion is designed based on the model
description length and the data description length using original AIS ship trajectory data, which
ensures to extract ship trajectory features while well maintaining the trajectory features and shape.
(2) The DBSCAN clustering algorithm using the dynamic time warping (DTW) trajectory
similarity measurement method is proposed according to the position, speed and course of ship,
which essentially solves the problem of slow accuracy and trajectory similarity between ship
trajectories and clustering effectiveness.
(3) A ship trajectory prediction model based on the transformer model is constructed for the
anomaly detection of ship by using the transformer's self-attention mechanism to capture the long-
term dependency relationship between ship trajectory information. The results demonstrate that this
method is superior to other algorithms in terms of effectiveness and reliability.
This paper is organized as follows: Section 2 presents the proposed method for ship trajectory
clustering analysis and anomaly detection. Section 3 presents the experiments setup include the
datasets used in the experiments, evaluation results of the proposed method. At the end of this
section, the proposed method is compared with other previous recent studies. Finally, the paper is
concluded and some suggestions are given for future studies.
2. Methodology
To address the challenges related to long-term jumps and missing ship trajectories, a method is
proposed to divide sub-trajectories and ensure the availability of ship trajectory data. The MDL
criterion is utilized to extract ship trajectory feature points and derive ship trajectory features. In
water traffic scenarios, the lengths of AIS ship trajectories cannot be made consistent for all
trajectories. Therefore, in view of the lack of consideration for water traffic scenario factors when
measuring ship trajectory similarity, a ship trajectory feature extension method is introduced,
incorporating ship position, speed, and course. This extended feature set is used in conjunction with
the DTW trajectory similarity measurement distance and the DBSCAN algorithm to achieve ship
trajectory clustering. This approach enhances the accuracy of similarity measurement between ship
trajectories and enables the generation of normalized ship motion trajectories. To address ship
trajectory anomaly detection, a ship trajectory prediction method based on a transformer model is
proposed. The normalized ship motion trajectory obtained from clustering serves as the training data
for the network. This enables the transformer model to predict trajectories and serve as a tool for
anomaly detection, completing the detection of ship trajectory anomalies. The flowchart depicting
ship trajectory clustering and anomaly detection is illustrated in Fig. 1.
Dynamic data Static data
Time
MMSI
Extraction
Data Position
Source AIS dataset AIS data
SOG
COG NAME
MDL method
Ship
Data Data
trajectory
preprocess clean
feature
ing ing
extraction
Trajectory Course
clustering
Position
Transformer model
Trajectory Predicting Speed anomaly
prediction the Position
and trajectory anomaly
anomaly of the target detection
Detection ship Course anomaly
n −1 k −1
L( D / H ) = log 2 (d ( Px Px +1 , Py Py +1 )) + log 2 (d ( Px Px +1 , Py Py +1 )) (2)
x −1 y −1
where Len(P) represents the length of sequence P, and dα and dθ are the vertical distance and angular
distance between sequences, respectively.
Pc1 Pc2
P1 P4
P5
P2
P3
L( D / H ) = log 2( d ⊥( PP 1 2 ) +d ⊥( PP
1 4 ,PP 1 4 ,P2 P3 ) +d ⊥( PP
1 4 ,P3 P4 ) ) +
(4)
log(
2 d( PP 1 2 ) +d( P
1 4 ,PP 1 P4 ,P2 P3 ) +d( PP
1 4 ,P3 P4 ) )
Data input
Determine
No MDLpar> MDLnopar
Yes
Calculate the encoding
length L(H) of trajectory T
Construct a new trajectory
description, relative to the
T ={P1,P2,...} by connecting in
compressed encoding
sequence
length L(D/H) of the
original trajectory
MDLpar=L(H)+L(D/H) Determine
MDLnopar=L(T ) MDLpar< MDLnopar
Yes No
Select MDLpar as the Select MDLnopar as
description the description
encoding length for encoding length for
the final trajectory the final trajectory
Output final
trajectory
0 m=n=0
m = 0orn = 0
DTW (T1 , T2 ) = DTW ( Rest (T1 ), Rest (T2 )) (5)
dist (a , b ) + min DTW ( Rest ( A), B)
1 1
DTW ( A, Rest ( B))
p11
p5
p4 p10
p6
p1 p3
p2 p7 p9
T1
p8
Mij
q10
q6
q2 q9
T2 q5 q7
q1 q3
q4 q8
p11
p10
p9
p8
p7
p6
p5
p4
p3
p2
p1
T1
T2 q1 q2 q3 q4 q5 q6 q7 q8 q9 q10
X T =[xt , yt , st , ct ] (6)
Each row of the expanded feature matrix represents the feature vector of a trajectory point,
where xt and yt represent the longitude and latitude, respectively. st indicates the speed, and ct
indicates the course. DBSCAN is a density-based clustering algorithm that automatically discovers
clusters of any shape in data and identifies noise points. In this study, the improved DTW distance
algorithm is utilized as a replacement for the ε-field in DBSCAN clustering.
2.3. Ship trajectory prediction and anomaly detection
A ship trajectory prediction model based on the transformer architecture is developed using the
normal trajectory model obtained from DBSCAN clustering as the foundation for data.
Simultaneously, the criteria for trajectory anomaly detection are translated into feature deviation
values, encompassing position, speed, and course. Real-time ship trajectory anomalies are detected
by assessing the predictability of normal trajectories and the unpredictability of abnormal
trajectories.
2.3.1 Transformer model
The proposed transformer model in this study incorporates several key components, including
position coding, a multihead attention mechanism, residual connections, and normalization (Zhang
et al, 2023).
(1) Position coding
The transformer architecture, being a parallel input model that lacks the sequential iteration
advantage of RNN structures, necessitates the introduction of position coding to incorporate
positional information. The input position information is expressed as follows:
Q = XW Q
(9)
K = XW
K
V = XW V
The attention score matrix (Fig. 7) composed of α is calculated using Q and K. Subsequently,
the final output matrix of attention layer comprising b is calculated using V and the attention score
matrix. The output result is recorded as Attention(Q, K, V):
QK T
Attention(Q, K , V ) = softmax( )V (10)
dk
where Q is the query matrix, K is the key matrix, V is the value matrix, and X is the input matrix; W,
Q, K, and V are linear transformation weight matrices, and softmax is used to calculate the weight.
dk is the dimension of input data.
Fig. 7 Calculation process of attention score.
The calculation of the sub-attention mechanism involves applying linear transformations to the
K, Q, and V matrices. These linear transformations serve to capture different aspects of the input
and enable the model to focus on multiple levels of information. After each self-attention operation,
the outputs from multiple attention heads are spliced together to form the final output. This allows
the model to integrate information and enhance its overall performance. The multiple-head attention
mechanism, denoted as MultiHeads, can be represented as follows:
where Wi Q , Wi K , and WiV are the linear transformation weight matrices of the i-th head of Q,
K, and V, respectively, and Wo is the linear transformation weight matrix after the multihead
attention matrix is concatenated, Concat is a feature concatenation function.
(3) Residual connection
The submodule of the transformer model primarily consists of a multihead attention
mechanism layer and a feedforward layer. Between these layers and the input layer, as well as
between the attention layer and the feedforward layer, there exist residual connections and data
normalization. Residual connections allow for the connection of input and output data, addressing
the issues of gradient vanishing and weight matrix degradation. The dimensions of the input and
output data from the attention layer are consistent, facilitating the residual connection. The residual
connection formula for the multihead attention mechanism (H) and the residual connection formula
of feedforward (H’) are as follows:
H = H + Feedforward ( H ) (14)
Start
Sort by MMSI
identification code Delete the error data of MMSI
and receiving time
37.62
Latitude(°)
37.60
37.58
37.56
37.66
37.64
Latitude(°)
37.62
37.64
Latitude(°/N)
37.62
37.60
37.58
37.56
400 373.54
350 315.8
300
255.25
point(m)
250 230.06
200
145.51 141.28 139.54
150 133.54 128.71
118.04 117.1
107.12 112.42 113.46 111.01
101.24
91.06 89.5 96.37
100 82.32 79.5 83.55
66.35 72.09 69.42
64.55 68.92 72.57
54.32
52.46 61.05
48.66 50.88 46.51 45.58
50
0
1 2 3 4 5 6 7 8 9 1011121314151617181920212223242526272829303132333435
Sequence of points
Fig. 25 Schematic of the distance difference between the real trajectory and the predicted
trajectory of ship steering.
Course difference from predicted
14 12.34
12 10.96
10
point(°)
8.13
8 6.81
6 4.684.564.89 4.58
4.29
4 3.13
2.34
2 0.75
1.280.96 0.861.010.79 1.060.96 1.241.49
0.10.110.470.120.190.4 0.320.190.220.38 0.20.320.06 0.38
0
1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536
-2
Sequence of points
Fig. 26 Schematic of the course difference between the real trajectory and the predicted
trajectory of ship steering.
4
3.5
Speed difference from predicted 3 3.2
2.5
point(kn)
2 2.27 2.232.32
2
1.5
1
0.5
0.46
0
0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.10.050.050.050.050.050.050.160.150.150.150.150.140.140.140.140.140.140.130.130.13 0.12
1 2 3 4 5 6 7 8 9 1011121314151617181920212223242526272829303132333435
-0.5
Sequence of points
Fig. 27 Schematic of the speed difference between the real trajectory and the predicted
trajectory of ship steering.
37.63
37.62
Latitude(°/N)
37.61
37.60
37.59
37.58
150 113.55
110.46 110.06
95.98 95.98 95.56 91.91 86.84 87.38 85.85 99.17 99.75
94.16
84.56 80.73
75.04 75.14 81.82 81.03 85.34
83.32 77.74
100 66.6
64.32 68.26
58.11
74.37 72.72
54.26 46.62 54.16
38.16 31.32 33.42 38.67
38.07
50
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41
sequence of points
Fig. 29 Schematic of the distance difference between the real trajectory and the predicted
trajectory of ship acceleration.
3
2.47
Course difference from predicted
2.5
1.91
2 1.71 1.73 1.66 1.7
point(°)
1.5
1.16
1.04
0.89
1 0.75
0.71
0.65 0.67
0.5 0.57 0.52
0.47
0.5 0.31
0.24 0.27 0.29
0.19 0.2 0.26 0.28
0.15
0.05 0.13 0.15
0.04
0.04 0.11 0.16 0.09 0.09
0.01 0 0.03 0.01
0.01
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41
-0.5
Sequence of points
Fig. 30 Schematic of the course difference between the real trajectory and the predicted
trajectory of ship acceleration.
3.5 3.2
Speed difference from predicted
3
2.5 2.32
2.27 2.23
2
2
point(kn)
1.5
1
0.46
0.5 0.16 0.15
0.10.10.10.10.10.10.10.10.05
0.05
0.05
0.05 0.05 0.15
0.05 0.15
0.15
0.14
0.14
0.14
0.14
0.14
0.14
0.13
0.13
0.13 0.120.10.10.10.140.1
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
-0.5
Sequence of points
Fig. 31 Schematic of the speed difference between the real trajectory and the predicted
trajectory of ship acceleration.
4. Conclusion
In this research, a ship trajectory anomaly detection model is developed, taking into account
the characteristics of AIS trajectory data in maritime traffic research. The proposed model includes
an improved DTW algorithm that considers local trend characteristics and ship motion information,
enhancing the accuracy of trajectory similarity measurement. Additionally, a transformer trajectory
prediction model is constructed, incorporating a threshold-based anomaly detection method. The
transformer model is trained using normalized motion trajectories obtained through DBSCAN
cluster analysis, enabling it to predict ship trajectories based on historical data. The effectiveness of
the proposed method is demonstrated through AIS data from the Port of Yantai, highlighting its high
accuracy in detecting abnormal ship trajectories. Future research endeavors may involve considering
the impact of meteorological environments, as such information can influence ship clustering and
anomaly analysis. In addition, enhancing the adaptability of method parameters is one of the
important research interests in the future.
Disclosure statement
No potential conflict of interest was reported by the authors.
Acknowledgements
This work was supported by the Applied Basic Research Plan of Liaoning Province in 2022
(2022JH2/101300265); the Fundamental Research Funds for the Dalian maritime university
(017231034); 2023 DMU navigation college first-class interdisciplinary research project
(2023JXB14).
References
Bai, Xiangen, Xie, Zhexin, Xu Xiaofeng, Xiao, Yingjie, 2023. An adaptive threshold fast DBSCAN
algorithm with preserved trajectory feature points for vessel trajectory clustering. Ocean
Engineering. 280, 114930.
Dahlbom, Anders, and Lars, Niklasson, 2007. Trajectory clustering for coastal surveillance. 2007
10th international Conference on Information Fusion. IEEE, 1-8.
Dai, Jianhua, Liu, Ye, Chen, Jiaolong, Liu, Xiaofeng, 2020. Fast feature selection for interval-valued
data through kernel density estimation entropy. International Journal of Machine Learning and
Cybernetics. 11(12), 2607-2624.
D’Afflisio, Enrica, Braca, Paolo, Willett, Peter, 2021. Malicious AIS spoofing and abnormal stealth
deviations: a comprehensive statistical framework for maritime anomaly detection. IEEE
Transactions on Aerospace Electronic Systems. 57(4), 2093-2108.
Ester, Martin, Kriegel, Hans-Peter, Sander, Jorg, Xu, Xiaowei, 1996. A density-based algorithm for
discovering clusters in large spatial databases with noise. In: Kdd. Vol. 96. No. 34. pp. 226-231.
Gözde, Boztepe Karataş, Pinar, Karagoz, Orhan, Ayran, 2021. Trajectory pattern extraction and
anomaly detection for maritime vessels. Internet of Things. 16, 100436.
Han, Xuyang, Armenakis, Costas, Jadidi, Mojgan, 2021. Modeling vessel behaviours by clustering
AIS data using optimized DBSCAN. Sustainability. 13(15), 8162.
Jin, Chen, Hao, Chen, Quan, Chen, Xin, Song, Wang, Hongdong, 2023. Vessel sailing route
extraction and analysis from satellite-based AIS data using density clustering and probability
algorithms. Ocean Engineering. 280, 114627.
Kontopoulos, Ioannis., Varlamis, Iraklis, Tserpes, Konstantinos, 2021. A distributed framework for
extracting maritime traffic patterns. International Journal of Geographical Information Science.
35(4), 767-792.
Kowalska, Kira, Peel, Leto, 2012. Maritime anomaly detection using Gaussian process active
learning. In: 2012 15th International Conference on Information Fusion, pp. 1164-1171.
Laxhammar, Rikard, Falkman, Gӧran, Sviestins, Egils, 2009. Anomaly detection in sea traffic-a
comparison of the Gaussian mixture model and the kernel density estimator. In: 2009 12th
International Conference on Information Fusion, pp. 756-763.
Lei, Po Ruey, 2016. A framework for anomaly detection in maritime trajectory behavior.
Knowledge and Information Systems. 47(1), 189-214.
Li, Kexin, Guo, Jian, Li, Ranchong, Wang, Yujun, Li, Zongming, Miu, Kun, Chen, Hui. 2023. The
Abnormal Detection Method of Ship Trajectory with Adaptive Transformer Model Based on
Migration Learning. In: 2023 4th International Conference on Spatial Data and Intelligence, pp.
204-220.
Li, Huanhuan, Liu, Liu, Jingxian, Yang, Zaili, Liu, Ryanwen, Wu, Kefeng, Wan, Yuan, 2020.
Adaptively constrained dynamic time warping for time series classification and clustering.
Information Sciences. 534, 97-116.
Liu, Ryanwen, Liang, Maohan, Nie, Jiangtian, Yuan, Yanli, Xiong, Zehui, Yu, Han, Guizani, Nadra,
2022. STMGCN: Mobile edge computing-oempowered vessel trajectory prediction using spatio-
temporal multigraph convolutional network. IEEE Transactions on Industrial Informatics. 18(11),
7977-7987.
Mantecón, Tomas, Casals, David, Navarro-Corcuera, Juanjosé, del-Blanco, Carlos. R., Jaureguizar,
Fernando. 2019. Deep learning to enhance maritime situation awareness. In 2019 20th International
radar symposium. IEEE, 1-8.
Mascaro, Steven, Nicholso, Ann, Korb, Kevin, 2014. Anomaly detection in vessel tracks using
Bayesian networks. International Journal of Approximate Reasoning. 55(1), 84-98.
Nooshin, Hanafi, Hamid, Saadatfar, 2022. A fast DBSCAN algorithm for big data based on efficient
density calculation. Expert Systems with Applications. 203, 117501.
Park, Jinwan, Jeong, Ju-kwon, Park, Youngsoo, 2021. Ship trajectory prediction based on Bi-LSTM
using spectral-clustered AIS data. Journal of Marine Science and Engineering. 9(9), 1037.
Ristic, Bronko, La Scala, B., Morelande, M., Gordon, N., 2008. Statistical analysis of motion
patterns in AIS data: anomaly detection and motion prediction. In: 2008 11th International
Conference on Information Fusion, pp. 1-7.
Rong, Hao, Teixeira, A. P., Soares, Carlos Guedes, 2020. Data mining approach to shipping route
characterization and anomaly detection based on AIS data. Ocean Engineering. 198, 106936.
Shahir, Hamed Yaghoubi, Glässer, Uwe, Nalbandyan, Narek, 2014. Maritime situation analysis: a
multi-vessel interaction and anomaly detection framework. In: 2014 IEEE Joint Intelligence and
Security Informatics Conference, pp. 192-199.
Sheng, Pan, Yin, Jingbo, 2018. Extracting shipping route patterns by trajectory clustering model
based on automatic identification system data. Sustainability 10(7), 2327.
Smith, Mark, Reece, Steven, Roberts, Stphen, Rezek, Iead., 2012. Online maritime abnormality
detection using Gaussian processes and extreme value theory. In: 2012 IEEE 12th International
Conference on Data Mining, pp. 645-654.
Soares Júnior, A., Moreno, Moreno, B. N., Times, V. C., Matwin, S., Formiga Cabral, L. D. A.,
2015. GRASP-UTS: an algorithm for unsupervised trajectory segmentation. International Journal
of Geographical Information Science, 29(1-2), 46-68.
Tang, Chunhua, Chen, Meiyue, Zhao, Jiahuan, Liu, Tao, Liu, Kang, Yan, Huaran, Xiao, Yingjie,
2021. A novel ship trajectory clustering method for finding overall and local features of ship
trajectories. Ocean Engineering. 241, 110108.
Venskus, Julius, Treigys, Povilas, Bernatavičienė, Jolita, Medvedev, Viktor, Voznak, Miroslav,
Kurmis, Mindaugas, Bulbenkiene, Violeta, 2017. Integration of a self-organizing map and a virtual
pheromone for real-time abnormal movement detection in marine traffic. Informatica. 28(2), 359-
374.
Venskus, Julius, Treigys, Povilas, Bernatavičienė, Jolita, Tamulevičius, Gintautas, Medvedev, Viktor,
2019. Real-time maritime traffic anomaly detection based on sensors and history data embedding.
Sensors. 19(17), 3782.
Wang, Jiecheng, Liu, Yantong, Chang, Jincai, 2022. An improved model for kernel density
estimation based on quadtree and quasi-interpolation. Mathematics. 10(14), 2402.
Wang, Siwen, Li, Ying, Hu, Xing, 2023. A novel method for ship trajectory prediction in complex
scenarios based on spatio-temporal features extraction of AIS data. Ocean Engineering. 281, 114846.
Wen, Yuanqiao, Sui, Zhongyi, Zhou, Chunhui, Xiao, Changshi, Chen, Qianqian, Dong, Han, Zhang,
Yimeng, 2020. Automatic ship route design between two ports: a data-driven method. Applied
Ocean Research. 96, 102049.
Yang, Dong, Wu, Lingxiao, Wang, Shuaian, Jia, Haiying, Li, Kevin X, 2019. How big data enriches
maritime research–a critical review of Automatic Identification System (AIS) data applications.
Transport Reviews. 39(6), 755-773.
Yang, Jiaxuan, Liu, Yuan, Ma, Lingqi, Ji, Chengtao, 2022. Maritime traffic flow clustering analysis
by density based trajectory clustering with noise. Ocean Engineering. 249, 111001.
Zhang, Yuqi, Li, Chaofeng, Shang, Shaopeng, Chen, Xinqiang, 2023. SwinSeg: Swin transformer
and MLP hybrid network for ship segmentation in maritime surveillance system. Ocean Engineering.
281, 114885.
Zhao, Liangbin, Shi, Guoyou, 2019. Maritime anomaly detection using density-based clustering and
recurrent neural network. Journal of Navigation. 72(4), 894-916.
Zhen, Rong, Jin, Yongxing, Hu, Qinyou, Shao, Zheping, & Nikitakos, Nikitas, 2017. Maritime
anomaly detection within coastal waters based on vessel trajectory clustering and naїve bayes
classifier. Journal of Navigation. 70(3), 1-23.