Received January 16, 2022, accepted February 16, 2022, date of publication February 24, 2022, date of current

version March 7, 2022.

Digital Object Identifier 10.1109/ACCESS.2022.3154363

Maritime Traffic Route Detection Framework

Based on Statistical Density Analysis From AIS
Data Using a Clustering Algorithm
JEONG-SEOK LEE 1, HYEONG-TAK LEE 2, AND IK-SOON CHO 3
1 Graduate School, Korea Maritime & Ocean University, Busan 49112, South Korea
2 Ocean Science and Technology School, Korea Maritime & Ocean University, Busan 49112, South Korea
3 Division of Navigation Convergence Studies, Korea Maritime & Ocean University, Busan 49112, South Korea

Corresponding author: Ik-Soon Cho ([email protected])

ABSTRACT Maritime traffic routes by ships navigation vary according to country and geographic
characteristics, and they differ according to the characteristics of the ships. In ocean areas adjacent to coasts,
regulated routes are present, e.g., traffic separation scheme for ships entering and leaving; however, most
ocean areas do not have such routes. Maritime traffic route research has been conducted based on computer
engineering to create routes; however, ship characteristics were not considered. Thus, this article proposes a
framework to generate maritime traffic routes using statistical density analysis. Here, automatic identification
system (AIS) data are used to derive quantitative traffic routes. Preprocessing is applied to the AIS data, and
a similar ship trajectory pattern is decomposed into a matrix based on the Hausdorff–distance algorithm and
then stored in a database. A similar pattern makes the AIS trajectory simple using the Douglas–Peucker
algorithm. In addition, density-based spatial clustering of applications with noise (DBSCAN) is performed
to identify the waypoints of vessels then create routes by connecting waypoints. The width of maritime routes
created based on a similar ship trajectory is subjected to kernel density estimation analysis (KDE). Then,
waypoints evaluation of the main route is performed from the results of KDE 75% and 90% considering
the statistical in the total maritime traffic, and the results applied to the targeted ocean area are compared.
Finally, the result of KDE 90% of maritime traffic with framework analyzed the safety route, which can be
a basis for developing routes for maritime autonomous surface ships.

INDEX TERMS AIS data, clustering algorithm, DBSCAN, kernel density estimation, framework, maritime
traffic route.

I. INTRODUCTION for safe operation of ships and prevention of maritime

As a result of developments in information and communi- accidents [5]. Maritime traffic routes for ships navigation
cation technologies and cloud computing technologies, the vary by country and geographic characteristics. In ocean
amount of real-time data is increasing rapidly. In the maritime areas adjacent to the coast, there are prescribed routes that
transportation field, large amounts of automatic identification ships must use to enter and leave ports. In addition, the traffic
system (AIS) data are accumulated in real time, and research separation scheme is used to guide ship traffic. Moreover,
using such data is being conducted [1]. maritime autonomous surface ships (MASS) are expected in
The shipping industry is the most effective means of future [6]; thus, the need to create maritime traffic routes
transporting goods over long distances, and greater than 80% is increase, even in ocean areas where such routes do not
of the world’s merchandize ships use the oceans [2], [3]. currently exist.
Therefore, with the expansion of the shipping industry, the A ship’s trajectory data can identify the ship’s operating
demand for maritime transportation is increasing rapidly, environment from the ship’s past motions and patterns. AIS
which has resulted in increased ship sizes [4]. Thus, national data include various information, e.g., global positioning
organizations recommend or enforce route design standards system information, speed over ground (SOG) information,
course over ground (COG) information, ship type, length
The associate editor coordinating the review of this manuscript and of all, breadth, time, and place of arrival. In addition, a
approving it for publication was Haiyong Zheng . ship’s AIS data are exchanged data with adjacent ships and
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 10, 2022 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 23355
 J.-S. Lee et al.: Maritime Traffic Route Detection Framework Based on Statistical Density Analysis

AIS stations [7], [8]. The International Maritime Organi- value of the main route [15]. KDE is limited in that it gives
zation (IMO) requires cargo ships with gross tonnage of different results depending on the setting value and search
300 tons or more engaged in international voyages, ships range. However, KDE was used to quantitatively calculate
with gross tonnage of 500 tons or more not engaged in the route width and was applied as a method to evaluate
international voyages, and all passenger ships to be equipped the performance of the maritime route. KDE identifies an
with AIS [9]. These regulations form part of Chapter V of efficient maritime traffic route by setting the search range
the Safety of Life at Sea (SOLAS) 3rd convention [10]. This from an arbitrary trajectory and analyzing routes as 75%
AIS is connected to other navigation equipment, e.g., the and 90% of the total maritime traffic using a statistical
electronic navigation chart display and information system method [4], [5]. Therefore, it is possible to create a maritime
and the automatic radar plotting aid, to prevent collisions traffic route based on density analysis, and the result of
and facilitate safe navigation [1]. In addition, according density analysis can have a unique shipping area in a wide
to the literature [9], the large-scale AIS data are also ocean area. This is expected to be applicable to autonomous
used to extract maritime routes and analyze ship motion. ships based on a simplified main route, and it is applicable to
Thus, AIS data are used to research improved maritime anomaly detection based on routes [16]. In addition, create
routes by combining various ship information and computer a maritime traffic route to protect the ocean environment,
technologies. Son et al. [11] explained the width of sailing promote navigation safety of ships and crews, create marine
ships from a reliability-based statistical perspective and spatial planning (MSP), and apply to the MASS [17]–[19].
conducted a study within a harbor. However, to the best of
our knowledge, no previous study has considered density II. RELATED STUDY AND MOTIVATION
analysis based on statistics in maritime routes or maritime As large amounts of AIS data are collected, the data should
traffic networks using AIS data and route width information. maximize public interest by linking methodologies in other
Navigable areas differ depending on the size of the ship; research fields with the maritime industry. AIS data can
however, this problem has not been considered in previous be used to extract maritime traffic routes, and scientific
studies. Therefore, maritime traffic routes are extracted in techniques, e.g., deep learning, can be applied to real-time
wide ocean areas considering ship characteristics, and route maritime traffic networks [1], [20]. Moreover, AIS data are
width values and the main route are calculated according to applied to navigation technology for MASS by combining
statistical-based density analysis. the collision risk of ships with the International Regulations
In this study, the AIS data of ships passing through the for Preventing Collision at the Sea [21]. In addition,
Southern Sea of Korea were targeted. The Korean Ministry another study analyzed ship traffic in Singapore Strait using
of Oceans and Fisheries collects the AIS data of the ships spatial-temporal analysis. Here, dense areas were divided
passing through the ocean near Korea and provides these into hotspots [22], and maritime traffic was measured by
data for the General Information Center on Maritime Safety local hotspots and cold spots via spatial autocorrelation
and Security data. The raw AIS data are preprocessed analysis [8]. In addition, AIS data are a key element in
due to the data loss rate according to the distance and the analysis of a ship’s operation pattern. Thus, AIS data
inaccuracy of the data [12]. Trajectory clustering differs analyzes the ship’s COG, SOG, Rate of Turn data and Voyage
from point clustering; thus, the trajectory should not deviate Data Recorder information with a clustering algorithm. From
from generality and should include similar characteristics. the past data, it is possible to understand the ship’s engine
Using the Hausdorff–distance algorithm, which considers order and rudder use to prevent ship accidents [23]. Further,
the shape difference between two different sets of points, various machine learning techniques were used to predict
and the trajectory is clustered effectively [13]. Numerous the safety berthing velocity of ships. The algorithm with the
course adjustment movements are performed in an actual ship best performance was derived through the receiver operating
trajectory to proceed to the course set by a vessel’s autopilot, characteristic curve. This study aims to protect port facilities
which increases the complexity of a ship’s trajectory. Here, and the environment by presenting guidelines for large ships
the Douglas–Peucker (DP) algorithm is used to simplify the to dock safely at ports [24]. Therefore, AIS data can be fused
trajectory [14]. A simple trajectory effectively represents with various algorithms and obtain desired results. In this
waypoints that change a ship’s course. In other words, it is study, a method for detecting main maritime traffic routes
possible to create a vertex point and clusters the vertices based on vast trajectory data is proposed. After detecting the
generated in a specific part to select a waypoint for a maritime route, used the proposed a method to select the
maritime traffic route. Here, density-based spatial clustering width of the route, and furthermore, propose a framework to
(DBSCAN) is applied to point clustering, and its use has represent the safety of the route.
been proven in the traffic route extraction and anomaly
detection (TREAD) method [13]. A maritime traffic route in A. RESEARCH ON MARITIME ROUTE EXTRACTION
consideration of the characteristics of the trajectory’s motion Three methods are used to extract maritime routes, i.e.,
and ship size is created by connecting waypoints. Then, density-based analysis, clustering-based route extraction, and
the kernel density estimation (KDE) of the line is analyzed TREAD. Lee et al. [5] extracted density-based routes by
to calculate the statistical route width and the calculated analyzing 90% of maritime traffic with KDE based on a

23356 VOLUME 10, 2022

J.-S. Lee et al.: Maritime Traffic Route Detection Framework Based on Statistical Density Analysis

FIGURE 1. Maritime traffic route framework using the clustering algorithm and statistical density analysis.

geographic information system (GIS). The Maritime route This method automatically generates a synthetic maritime
boundary was extracted using the Otsu binary analysis traffic network based on the historical positioning data of
method and the Canny algorithm, which is an edge extraction ships. Here, waypoint and route detections are based on
method, for the density analysis results [25], [26]. The the TREAD method and DBSCAN algorithm. In addition,
superiority of the previously set route was explained through TREAD uses the Hausdorff–distance algorithm for route
image processing. decomposition and the Douglas–Peucker algorithm to detect
Liang et al [27] aimed to achieve reliable mining results breakpoint detection. The created route is unsupervised, and
for massive vessel trajectories for efficiently computing the the learning result is displayed as performance evaluation.
similarities between different vessel trajectories. Liu et al [28] A maritime traffic network includes connected nodes that are
predicted the vessel trajectory for improving smart maritime created by extracting routes. Here, a node represents a ship’s
traffic services with deep learning. This work proposed an navigation waypoint, and the method used to extract the node
AIS data-driven trajectory prediction framework based long is approached in a scientific way.
short term memory network. Forti et al. [31] used AIS data to generate graph-based
Wang et al [29] used the Douglas–Peucker algorithm maritime traffic network, and they proposed an unsupervised
to simplify the ship trajectory, and the Hausdorff–distance approach to extract the patterns of ships [32], [33]. Here,
algorithm was used in the matrix. In addition, hierarchi- the Ornstein–Uhlenbeck mean-reverting stochastic process
cal density-based spatial clustering of applications with was applied to describe the application of graphs to detect
noise (HDBSCAN) was used to cluster trajectory according and connect waypoints via ship motion modeling [34], [35].
to hierarchical. Wang et al [29] compared results of the This creates a network and realizes maritime traffic pattern
K-means, spectral, DBSCAN, and HDBSCAN clustering detection and statistical characterization.
analysis methods and found that HDBSCAN had the best Yan et al. [9] proposed a maritime traffic network based
clustering performance. on the framework of the ship trip semantic object (STSO).
TREAD which is commonly used for the maritime traffic Here, STSO describe a network creation methodology that
extraction task, was proposed by Pallotta et al [13]. The AIS finds and connects stop point and waypoint using the
data were used to understand, classify, and predict maritime ordering points to identify the clustering structure clustering
traffic patterns in order to automatically detect anomalies in algorithm [36].
vessel operation patterns. Based on the study by Wang et al. [37], this study extracted
a maritime traffic network and proposed a different maritime
B. MARITIME TRAFFIC NETWORK traffic network creation method. The method included how
A method to create a maritime traffic network using the to create the lane boundaries, lane centerlines, and junctions.
TREAD method was proposed by Arguedas et al. [30]. The route was created by applying the image processing

VOLUME 10, 2022 23357

 J.-S. Lee et al.: Maritime Traffic Route Detection Framework Based on Statistical Density Analysis

technique based on the KDE density analysis. The outside

of the route was extracted as a lane, and the bumpy
lane was smoothed out. Further, by applying the Delaunay
triangulation model, attribute values according to adjacent
triangles were assigned. Finally, nodes and segments were
created and the centerline connected to them was presented
as a maritime traffic network.
III. MATERIALS AND METHODS
To generate a maritime traffic route, it is necessary to
understand ships’ movement patterns and operational factors.
Generally, ship trajectories are created to reduce travel
distance and maintain safety; thus, most merchant ships have
similar trajectories. This means that AIS data can be used
to generate maritime traffic routes [38]. In this study, route
width and the main route are also considered by applying
statistical density analysis to differentiate existing maritime
traffic networks. Therefore detecting the maritime traffic
route in the vast trajectory data is crucial [27]. In addition,
for a vessel to sail safely, an appropriate width of route is an
essential condition. The selecting appropriate width of route
was presented as a new method by performing a simplified
trajectory-based density analysis. Furthermore, safety of the
route is a key factor. To execute the sailing plan without
the ship grounding, the reliability of the detected maritime
traffic route must be guaranteed. To solve this problem,
FIGURE 2. AIS data plot from March 1, 2018 to March 7, 2018, (a) cargo
a novel statistical density analysis was applied. Further, the ships and (b) tanker ships, completely different from tanker ship
median center was used to express one waypoint that appears (c) towing ships and (d) passenger ships.
as clustering to detect maritime traffic routes. As a whole, During data preprocessing, missing values were removed,
a vast amount of trajectory data was concisely collated and and all inaccurate characters and symbols were deleted.
efficiently analyzed. The proposed maritime traffic route In addition, due to the characteristics of the ships, the
framework is illustrated in Fig. 1. SOG is nearly zero during berthing, anchoring, and drifting
A. DATA MINING AND PREPROCESSING AIS DATA conditions; thus, SOG is unsuitable for determining maritime
Currently, the AIS database (approximately 950 GB) stores traffic routes. Due to the characteristics of merchant ships,
data for all ships that passed through waters near Korea from there are no data with SOG of 25 kn or greater; thus, SOG was
2018 to 2020. However, the raw AIS data are unsuitable for selected as 3 kn or more and 25 kn or less. For COG, which
analysis. The AIS data comprise both static and dynamic indicates the ship’s moving direction, all values other than 0
◦ – 360 ◦ were deleted. As shown in Fig. 2, the operation
data. Thus, to understand and interpret the data accurately,
the data must be combined. Combining the data was is pattern differs depending on the type, length, and breadth of
performed based on the maritime mobile service identifier the vessel, and the route appears differently. Therefore, the
(MMSI), which has the same information. After being AIS data should be analyzed by categorizing ship types.
combined, the AIS data include static data comprising the
MMSI, name, type, IMO number, draft, tons, and length, B. CLUSTERING ALGORITHM
and the dynamic data include date (year-month-day-time), Ship trajectories are clustered to identified similar trajecto-
latitude, longitude, SOG, ship’s heading, and COG. The ries. Merchant ships exhibit similar trajectories according to
MMSI comprises nine numbers and attempts to make it easy the route selection purpose, and it is necessary to cluster
to identify a vessel. In addition, the ship type consists of appropriately [38]. The Hausdorff–distance algorithm can
numbers from 1 to 99 [39]. represent a similar ship trajectory as a matrix [29]. If the
After data combining, data preprocessing was performed. Hausdorff–distance is a subset of a non-empty metric space,
To perform data analysis efficiently, the data were divided where Ta represents trajectory a, and Tb represents trajectory
into daily data, and then the analysis area was divided into b, dH (Ta , Tb ) in the Hausdorff space is expressed as follows.
16 areas and stored in the database. In addition, data analysis
dH (Ta , Tb ) = max {d (Ta , Tb ) , d (Tb , Ta )} (1)
may yield different results depending on a vessel’s type and
characteristics. Therefore, the types of ships were divided into Here, T denotes the trajectory, Ta = [a1 , a2 , · · · , an−1 , an ]
cargo, tanker, passenger, and towing ships, and the length of denotes the trajectory of ship a, and Tb =
the ship was divided as > 60 m. [b1 , b2 , · · · , bn−1 , bn ] denotes the trajectory of ship b. Given

23358 VOLUME 10, 2022

J.-S. Lee et al.: Maritime Traffic Route Detection Framework Based on Statistical Density Analysis

FIGURE 3. Conceptual diagram of the Hausdorff–distance and process to

calculate the similar trajectory.

two sets of points Ta and Tb , the Hausdorff–distance in one

direction is obtained as follows.
d(Ta , Tb ) = max{min{ ai − bj }} (2)
d(Tb , Ta ) = max{min{ bj − ai }} (3)
Here, kai −bj and bj −ai k use the Euclidean–distance to
FIGURE 4. Simplifying the ship trajectory using the Douglas–Peucker
define two trajectory points as the Hausdorff–distance [40]. algorithm.
An example of the Hausdorff–distance based on the ship
trajectory is shown in Fig. 3.
Fig. 3 shows that the shortest distance from Ta to Tb
is d (a3 , b3 ), and the shortest distance from Tb to Ta is
d (b4 , a4 ). In other words, the Hausdorff–distance calculates
the largest distance among all distances from a point on one
trajectory to the nearest point on another trajectory.
These connected distances can be clustered by the matrix
of the past trajectory. This study approaches from a statistical
perspective and extract the main route in consideration of
route width. Therefore, to cluster similar trajectories, a value
of 0.1 or less for each MMSI is used as one clustering. FIGURE 5. Extracted to the vertex point of the ships trajectory.
After clustering, the vessel changes points to lines in over
time based on the MMSI number. Thus, the trajectory line is the Douglas–Peucker algorithm first creates trend lines of
performed as a line-based analysis rather than a point-based P1 , P2 , · · · , Pn−1 , Pn and sets the threshold value d. Then,
analysis. dmax exceeding threshold d is selected, and the maximum
value is expressed as Pmax . Therefore, when dmax > d, the
C. SIMPLIFICATION OF SHIP TRAJECTORY procedure is divided into two parts with the corresponding
Owing to the complexity of a ship’s operation, the ships do boundary point Pmax and then repeated (Fig. 4).
not continuously maintain a certain course. Ships operation
includes continuously resisting the external force conditions D. WAYPOINT DETECTION IN MARITIME TRAFFIC ROUTE
of the ship, e.g., wind, waves, and dynamic factors, and A clear waypoint can be calculated by simplifying the
continuously steadying the course based on the course setting trajectory of each ship. However, even if the trajectories of all
value set by the autopilot system. Therefore, clustering ships are similar, the waypoints will differ slightly. Therefore,
trajectories that maintaining natural motion is performed to extract the waypoints of all trajectory lines, all vertex points
first, and trajectory simplification is performed on the (where one trajectory line meets another trajectory line and
clustered result. Here, the Douglas–Peucker algorithm is creates a certain angle) are extracted. An example is shown
used [29], [30], [41]. This algorithm maintains the threshold Fig. 5.
points to maintain the essential shape of the line. As the ships have a similar route owing to the operational
The Douglas–Peucker algorithm begins by connecting one characteristics of the ship, the DBSCAN algorithm is used
end of the line to the trend line, and the distance between to find the position where many vertex points (representing
the trend line and the trajectory is measured vertically. Here, a waypoint) are derived. The DBSCAN algorithm is an
vertices that are closer to the trend line than the threshold unsupervised machine learning technique that is widely
are removed. This procedure is performed until all vertices used in point information-based research [23]. For waypoint
are removed while dividing the trend line [42]. In this study, detection in a maritime traffic network, waypoints have been

VOLUME 10, 2022 23359

 J.-S. Lee et al.: Maritime Traffic Route Detection Framework Based on Statistical Density Analysis

FIGURE 7. Median center selection based on the DBSCAN cluster results

of vertex point.

points can be extracted using the same method, and the final
maritime traffic route is created when the start point, end
point, and all waypoints are connected.
FIGURE 6. DBSCAN works process and create the cluster.
E. KERNEL DENSITY ESTIMATION OF MARITIME TRAFFIC
extracted using DBSCAN from TREAD and the STSO, ROUTE WIDTH
which is also considered in this this study [8], [14]. Two When the extraction of maritime traffic routes is completed
parameters are required to use DBSCAN, i.e., epsilon (ε) is by connecting the waypoints to each other, KDE analysis
a factor that determines the minimum distance between all is performed on the trajectory the vessel has followed
data points, and the minimum point (minPts), which is the to create the route width. Lee et al. [5] analyzed the
minimum number of data that must be included within the ε KDE analysis to calculate route width with trajectory point
minimum distance [43]. As shown in Fig. 6, when there are data. However, this study is based on the trajectory line.
many minPts inside ε, the DBSCAN algorithm is initiated Because when the ships were docked in a port for a long
and expands in the direction of the data that satisfies the same period of time, there is a limit to an accurate analysis
conditions adjacent to each other. because there are many trajectory point data. The density
Here, if the dataset is defined as D, the vertex point is analysis method estimates the characteristics of random
represented by p and the ε neighborhood area of point is data variables. The most effective method to achieve this
defined as N and follows Eq. (4). involves drawing a histogram. A histogram is parametric and
suitable for visualizing only simple frequencies. However,
N (p) = {q ∈ D |dist (p, q) ≤ ε (4)
such a method is discontinuous at the boundary of each
where p follows Eq. (5), which constitutes the minPts of the histogram bin and has different disadvantages depending
DBSCAN algorithm. on the width of the bin. In addition, KDE is effective
for estimating nonparametric density functions for data
N(p) ≥ minPts (5) following nonparametric distributions. The KDE equation is
In addition, a boundary point is a point that is close to the core given in Eq. (7).
point but less than minPts from the ε neighborhood, and noise 1 Xn
is defined as a point other than the core and boundary points. fˆ (x, ) = K (x − xi ) (7)
n i=1
In waypoint detection using the DBSCAN algorithm, several Here, x is a random variable, and xi is an observation.
points appear as individual clusters, and the exact center point In addition, means bandwidth and kernel width; smoothing
must be found. As a measure of the central tendency of the parameter as a parameter for adjusting the kernel. The
dataset, DBSCAN is suitable for detecting outliers and has the kurtosis of the kernel is determined according to the size of
characteristic of minimizing multiple waypoints to only one . Here, K is divided by the sum of the observations with the
waypoint. Therefore, an example of extracting the DBSCAN kernel function applied to the number of observations [44].
median center is shown in Fig. 7, and the equation is given as In this study, trajectory line data were used to perform KDE
follows. analysis. The grid cell visually representing the analysis result
q
was set to 100 m. The search radius used to calculate the
di = (Xi − X t )2 + (Yi − Y t )2 + (Zi − Z t )2
t
(6)
density of adjacent lines in one random trajectory line was set
Here, Xi , Yi , and Zi are the coordinates for feature i, to 500 m. According to Article 60, paragraph 5 of the United
and n is equal to the total number of features. The median Nations Convention on the Law of the Sea (UNCLOS),
center is used to search for effective and integrated waypoints an artificial island and a maritime facility installed in the
of different merchant ships, and weights can be assigned exclusive economic zone is determined as a safety zone
according to the attribute values of these points. Therefore, at least 500 m from the maritime structure [45]. Finally,
when waypoints 1 and 2 are connected, a leg of a single as a determining factor of route width, 75% and 90% of
maritime route is created. In addition, the start and end the total maritime traffic are represented using the quantile

23360 VOLUME 10, 2022

J.-S. Lee et al.: Maritime Traffic Route Detection Framework Based on Statistical Density Analysis

FIGURE 8. KDE analysis conceptual diagram applying 75% and 90% of

the total maritime traffic.

method. Due to recent developments in offshore wind farms,

research is being conducted to define the minimum separation
distance to ensure the safe passage of ships. Here, 90% of
the maritime traffic is arranged by configuring the safety
distance between the offshore wind farms and ships [46].
In addition to ocean energy development, 90% of marine
traffic is applied to MSP, which considers various entities FIGURE 9. Area of analysis; Southern Sea of Korea with 3D result of
using different ocean spaces [47]. However, density analysis passing ships.

when 90% of maritime traffic is applied represented that TABLE 1. Counting analysis point data and traffic data according to ship
even if only one vessel passes through the space of the type.

ocean, there is a limit to representing all the space of the

ocean as the passage of ships, and to make other marine
uses sustainable to reduce the width of the route. Therefore,
in this study, 75%, i.e., range of 2σ , was also analyzed
according to Chebyshev’s inequality. Chebyshev’s inequality TABLE 2. Statistical analysis result of cargo ships.
is more general, stating that a minimum of 75% of values
must lie within two standard deviations of the mean and
88.89% within three standard deviations for a broad range of
probability distributions [48]. The route width created in this
study from a statistical perspective is shown in Fig. 8.
IV. CASE STUDY
A. DATA MINING AND PREPROCESSING FOR AIS
In this study, the analyzed area was the Southern Sea of
Korea, where many cargo and tanker ships pass through. TABLE 3. Statistical analysis results of tanker ships.
In addition, many passenger ships are present in these waters
due to the nearby islands. The analysis area is from latitude
33.70 ◦ north to 35.05 ◦ north, and longitude 127.60 ◦ east to
128.75 ◦ east. The analysis area is shown in Fig. 9.
The data used for analysis were for cargo ships and tanker
ships that have passed through the target area. It is impossible
to target all ship types to select a maritime traffic route
because many types of ship, e.g., towing ships, passenger
ships, and fishing ships, exhibit irregular and nonuniform
traffic patterns. Thus, towing ships, passenger ships and
fishing ships were omitted from the analysis because of SOG was 22.30 kn. The mean vessel length was 168.28 m,
industrial activities, e.g., maritime work, fishing activity, and the longest vessel length was 399.00 m. The mean
and cable installation, rather than maintaining a constant breadth of the vessel was 23.72 m, with a maximum breadth
track according to vessel characteristics. Table 1 shows the of 62.00 m.
data collected from March 1, 2018 to March 7, 2018, with The 36 tanker ships passed through the mean SOG at
319 cargo ships and 136 tanker ships passing through the 10.96 kn, and the maximum was 20.10 kn. The mean vessel
target area. length was 141.42 m, and the largest vessel was 336.00 m.
In addition, the detailed data statistics of the ships are In addition, the mean breadth of the vessels was 25.71 m,
shown in Tables 2 and 3. Maritime traffic routes were created and the maximum breadth was 60.00 m. There was a slight
based on the data of 319 cargo ships and 136 tanker ships. difference in COG between cargo and tanker ships because
The 319 cargo ships that passed through the target area the purpose of operation is different according to the type of
were navigating at an mean of 11.85 kn, and the maximum ships.

VOLUME 10, 2022 23361

 J.-S. Lee et al.: Maritime Traffic Route Detection Framework Based on Statistical Density Analysis

FIGURE 10. Hausdorff-distance calculation result of cargo ships. FIGURE 12. Similar ship trajectory clustering results (23 clusters) with 3D
result of passing ships.

which simplifies complex trajectories, was applied to the

trajectory created by clustering, and the results are shown in
Fig. 13(b). Fig. 13(c) shows the detection of vertex points
where one trajectory line meets another trajectory line, and
Fig. 13(d) illustrated the results of DBSCAN algorithm based
on the extracted vertex points. Here, the parameter values of
DBSCAN were ε = 1000 m and minPts = 50 pieces.
The vertex points of individual ships represent waypoints;
thus, the median center value was extracted (Fig. 13(e)) to
extract a single accurate waypoint. If these waypoints are
connected with lines, they form a maritime traffic route. As a
result, as the data before applying the proposed framework,
871 waypoints were extracted, but performing the proposed
method, 43 waypoints were extracted. In addition, as a result
FIGURE 11. Hausdorff-distance calculation result for tanker ships. of analysis based on the framework, it was reduced from
457 ships to 195 ships, and the number of vertex points
B. RESULTS OF SHIP TRAJECTORY CLUSTERING AND decreased from 271,626 to 8,683, as shown in Table 4.
WAYPOINT DETECTION A large amount of data went through the proposed
Clustering similar ship trajectories was selected according to framework, and the result show that the number of vertex
the Hausdorff–distance algorithm. It calculates the distance points was reduced to 96.9%. These points are selected as
between ships MMSI, and as the numerical value of distance waypoints by DBSCAN, and the waypoints can be decreased
appears as closer to zero, each trajectory means similar. up to 95.2%. Based on these results, KDE analysis can be
Conversely, dissimilar trajectories have large numerical performed, and the route width can be accessed from a
value. statistical perspective. Thus, the KDE analysis result of the
In this study, the range of ± 2σ was calculated from main route can be obtained.
the mean value, and a similar trajectory was selected as
0.1, which is less than −2σ , as shown in Fig. 10 and C. APPLYING KDE ANALYSIS RESULT AND CALCULATING
Fig. 11. Fig. 10 and Fig. 11 shows the Hausdorff–distance ROUTE WIDTH BASED ON STATISTICS
calculations of a cargo and tanker ships, respectively. Similar tracks were collected by clustering ship trajectories,
The calculations average values of cargo and tanker ships and waypoints were extracted by clustering vertex points.
are 0.72 and 0.77, respectively, and the −2σ values are If data preprocessing and clustering of similar tracks are not
0.13 and 0.15, respectively. Thus, a value of 0.1 was selected performed, detected routes can be in the all ocean space.
as a reference value by rounding down. Fig. 12 shows the Note that merchant ships sail to the next port, and related
clustering of similar trajectories of cargo and tanker ships, economic and safety factors are included; thus, they have
where the number of clusters of similar tracks is 23. similar appearances. Therefore, a high density (red area) is
Fig. 13 shows the process of extracting the waypoints shown in a lot of traffic, and the low density (green area) is
of similar ship trajectories. Here, Fig. 13(a) shows the ship shown with little traffic. Here, Initial maritime traffic route
trajectory clustering results. The Douglas–Peucker algorithm, (IMTR)-KDE-90% means maritime traffic route of initial

23362 VOLUME 10, 2022

J.-S. Lee et al.: Maritime Traffic Route Detection Framework Based on Statistical Density Analysis

FIGURE 13. Extracting waypoints from similar ship trajectory clusters.

TABLE 4. The compared results for number of waypoints, total ships, and TABLE 5. Results of ship’s route width calculation.
vertex points.

AIS data set. In the initial AIS data, where the clustering ocean space and avoid friction with other maritime users.
algorithm is not applied, 90% of the total maritime traffic In addition, this proposed framework can be applied to
volume is expressed by KDE. Clustered maritime traffic route the waypoint and cross track limit of MASS because the
(CMTR)-KDE-75% represents the maritime traffic route of framework has the advantage of being clearly extracted as a
clustered AIS data set. Based on the data applied with the maritime traffic route.
framework of this study, 75% of the total maritime traffic
volume is expressed as KDE. In addition, CMTR-KDE-90% D. EVALUATION BASED ON KDE ANALYSIS
is the result of KDE representing 90% of the total maritime The density-based traffic route creation can visually identify
traffic volume based on the data applied with the framework areas where maritime traffic is and is not concentrated. The
of this study. area with the highest density, i.e., the largest legend (red area),
Fig. 14 shows the KDE analysis results obtained with is the main traffic route traveled by many ships. However,
and without the proposed framework, where KDE analysis it is difficult to understand the main flow because the passage
was performed based on the 90 percentage of total maritime pattern is broken or dispersed in various parts of the main
traffic. Fig. 14(a) shows the KDE results for 90% total route.
maritime traffic without the proposed framework, and it If the ocean area set as a waypoint shows a large
appears that most of the analyzed area was traversed by ships. difference according to the KDE calculation result, the
Fig. 14(b) shows the KDE results for 75% total maritime reliability of the route of a vessel moving in a certain
traffic with the proposed framework, and Fig. 14(c) shows direction will be reduced. Therefore, the KDE calculation
the KDE results applied to 90% of the total maritime traffic result can be described as a main traffic route when it is
volume. In addition, the total length of the maritime traffic kept constant without significant differences depending on
route is approximately 930 km, and the total number of legs is the route. In other words, if the KDE calculation results with
17. The biggest reason for applying density analysis is that the waypoints show a lot of difference, these results should not
route width can be calculated quantitatively from a statistical be used to create a route. Therefore, the density calculation
perspective; therefore, to calculate a navigable area, a random results of the waypoint ocean area were compared and
route is selected. The result of the route width in south-eastern evaluated. Fig. 15 shows the KDE calculation result of
bound WayPoint = 5 → 4 → 3 → 2 → 10 is shown in waypoint–detection. IMTR-KDE-90% showed the largest
Table 5. value at 8,584.6 (trajectory line/km2 ) and the smallest value at
The south-eastern route has the mean width of 2,061.6 m 588.54, with an average of 1,893.83. CMTR-KDE-75% was
when the proposed framework was applied with 90% of the 3,037.1 at maximum, 714.45 at minimum, and 1,656.24 on
total maritime traffic. This demonstrate a lot of difference average. In addition, CMTR-KDE-90% was 2,698.76 at
compared to the when the proposed framework was not maximum, 740.5 at minimum, and 1,382.30 on average.
used. This method can prevent excessive navigation area in Therefore, the KDE analysis result of 90% of maritime traffic

VOLUME 10, 2022 23363

 J.-S. Lee et al.: Maritime Traffic Route Detection Framework Based on Statistical Density Analysis

FIGURE 14. Comparison of the general KDE analysis results without the proposed framework.

FIGURE 15. Waterfall graph visualization results according to the KDE calculation result obtained with the proposed framework.

obtained using the proposed framework with the smallest V. DISCUSSION

difference between the maximum and minimum values was Unlike urban road traffic, maritime traffic involves larger
found to be the best. The density-based evaluation was spaces, various environmental factors, and irregular maritime
performed to select the main route. Generally, this area was traffic characteristics. To solve this problem, the ship’s
investigated to prove that it is unique in terms of ship passage. operation pattern should be considered for extracting mar-
In addition, there is a limit to selecting potentials all areas as itime traffic routes based on ship type and geographic
traffic routes; thus, 75% and 90% of the total maritime traffic characteristics. Density-based extraction analysis to detect
volume were applied, and the route width was indicated. maritime traffic routes and networks, automatic anomaly

23364 VOLUME 10, 2022

J.-S. Lee et al.: Maritime Traffic Route Detection Framework Based on Statistical Density Analysis

detection in ship navigation patterns, graph-based extraction size of the ship. To represent the main traffic route, proposed
analysis, application of the Ornstein–Uhlenbeck process, framework can effectively represent the navigational area
method of STSO, and detected centerline [8], [15], [31], [38]. of a ship, and it can be applied to route design and safety
In common, the studies related to maritime route network diagnosis. In addition, the framework can build a maritime
were a method that uses a vast amount of AIS data from transport network at the national level beyond a specific
large ocean space, detects maritime routes, finds waypoints, port. In future, when generating a maritime traffic network,
and finds networks that connect them. The maritime route it will be necessary to subdivide the network according to
connected by lines plays an important role in the design of the purpose, type, and size of each vessel, and it will be
routes and safety diagnosis by finding networks. However, necessary to create guidelines for selecting the correct route
as ships increase in size, they cannot use all routes and width. Further, combining the extraction of maritime traffic
similar routes. Therefore, the safety route should be based on route and density analysis will contribute to the development
the appropriate water depth and route width, and the routes of the MASS technology.
connected by waypoints should be connected without a large REFERENCES
difference in the KDE calculated values. [1] B. Murray and L. P. Perera, ‘‘An AIS-based deep learning framework
This study deviated from the method of detecting an overall for regional ship behavior prediction,’’ Rel. Eng. Syst. Saf., vol. 215,
network in large ocean space and focused on detecting a Nov. 2021, Art. no. 107819.
[2] E. Tu, G. Zhang, L. Rachmawati, E. Rajabally, and G.-B. Huang,
traffic route near a specific large port. In addition, it evaluated ‘‘Exploiting AIS data for intelligent maritime navigation: A comprehensive
whether the extraction routes were well connected. As shown survey from data to methodology,’’ IEEE Trans. Intell. Transp. Syst.,
in Fig. 14(a), the KDE analysis of initial AIS data appears to vol. 19, no. 5, pp. 1559–1582, May 2018.
[3] UNCTAD, Review of Maritime Transport 2019, United Nation Publication,
have almost passed in the ocean area. Generally, it is possible New York, NY, USA, 2019.
to determine which ocean area ships can pass; however, it is [4] J. Weng and D. Yang, ‘‘Investigation of shipping accident injury severity
difficult to find the main route between dispersed routes. and mortality,’’ Accident Anal. Prevention, vol. 76, pp. 92–101, Mar. 2015.
[5] J.-S. Lee, W.-J. Son, H.-T. Lee, and I.-S. Cho, ‘‘Verification of novel
Therefore, Fig. 14(b) shows the KDE analysis of 75% of maritime route extraction using kernel density estimation analysis with
maritime traffic with the proposed framework, and Fig. 14(c) automatic identification system data,’’ J. Mar. Sci. Eng., vol. 8, no. 5,
is the KDE analysis result of 90% of the maritime traffic. The p. 375, May 2020.
[6] J.-S. Lee and I.-S. Cho, ‘‘Extracting the maritime traffic route in Korea
main route was extracted by clustering similar tracks, and based on probabilistic approach using automatic identification system big
waypoint detection via DBCSAN obtained a more concise data,’’ Appl. Sci., vol. 12, no. 2, p. 635, Jan. 2022.
and accurate route. From a route width perspective, the larger [7] C. B. Moorthy, M. K. Deshmukh, and D. Mukherejee, ‘‘Optimal location
of wind turbines in a wind farm using genetic algorithm,’’ TELKOMNIKA
the width, the better; however, in coastal waters, if the route Indonesian J. Electr. Eng., vol. 12, no. 8, pp. 658–665, Aug. 2014.
width is excessive, it may overlap with navigational obstacles, [8] J.-S. Lee, W.-J. Son, H.-T. Lee, and I.-S. Cho, ‘‘A study on the factors
which induce unsafe navigation. affecting optimal site of offshore wind farm from the perspective of
maritime traffic using spatial analysis,’’ Korea Soc. Coastal Disaster
In addition, if the routes connecting waypoints are broken Prevention, vol. 7, no. 2, pp. 85–96, Apr. 2020.
or disconnected, route reliability will be reduced. Therefore, [9] Z. Yan, Y. Xiao, L. Cheng, R. He, X. Ruan, X. Zhou, M. Li, and R. Bin,
as shown in Fig. 15, the most stable route is evaluated as ‘‘Exploring AIS data for intelligent maritime routes extraction,’’ Appl.
Ocean Res., vol. 101, Aug. 2020, Art. no. 102271.
the difference between the maximum and minimum density [10] International Convention for the Safety of Life at Sea (SOLAS).
values is smaller based on the KDE calculation results. (Nov. 1974). [Online]. Available: http://www.imo.org/About/Conventions/
ListOfConventions/Pages/International-Convention-for-the-Safety-of-
Therefore, this study selected the result of
Life-at-Sea-%28SOLAS%29,-1974.aspx
CMTR-KDE-90% with the smallest difference. The ocean [11] W.-J. Son, J.-S. Lee, H.-T. Lee, and I.-S. Cho, ‘‘An investigation of the ship
space is used by not only merchant ships, but also various safety distance for bridges across waterways based on traffic distribution,’’
J. Mar. Sci. Eng., vol. 8, no. 5, p. 331, May 2020.
entities, e.g., fishing activities, resource extraction, military [12] W. Hasbi, ‘‘Tracking capability and detection probability assessment
districts, and marine energy development. If an accurate of space-based automatic identification system (AIS) from equato-
shipping area is selected first, operation safety and marine rial and polar orbiting satellites constellation,’’ IEEE Access, vol. 8,
pp. 184120–184136, 2020.
environment protection can be promoted, and if a waypoint [13] G. Pallotta, M. Vespe, and K. Bryan, ‘‘Vessel pattern knowledge discovery
is created, it will be applied effectively to MASS. As a result, from AIS data: A framework for anomaly detection and route prediction,’’
future maritime routes can be created automatically based on Entropy, vol. 15, no. 6, pp. 2218–2245, 2013.
[14] D. H. Douglas and T. K. Peucker, ‘‘Algorithms for the reduction of
highly reliable waypoints. the number of points required to represent a digitized line or its
caricature,’’ in Classics Cartography: Reflections Influential Articles From
VI. CONCLUSION AND FUTURE WORK Cartographica, vol. 2. Toronto, ON, Canada: Univ. Toronto Press, 1973.
[15] O. Daae Lampe and H. Hauser, ‘‘Interactive visualization of streaming data
In this study, the maritime traffic route framework has been with kernel density estimation,’’ in Proc. IEEE Pacific Visualizat. Symp.,
proposed. The proposed framework represents a new route Mar. 2011, pp. 1–8.
detection method that considers both total maritime traffic [16] E. d’Afflisio, P. Braca, L. M. Millefiori, and P. Willett, ‘‘Maritime anomaly
detection based on mean-reverting stochastic processes applied to a real-
and statistics to calculate ship routes, including route width. world scenario,’’ in Proc. 21st Int. Conf. Inf. Fusion (FUSION), Jul. 2018,
Based on a vessel’s movement pattern, the vertex points are pp. 1171–1177.
extracted to identify the waypoints of a vessel. In addition, [17] J. Montewka, S. Ehlers, F. Goerlandt, T. Hinz, K. Tabri, and P. Kujala,
‘‘A framework for risk assessment for maritime transportation systems—
the route evaluation through the KDE analysis result indicates A case study for open sea collisions involving RoPax vessels,’’ Rel. Eng.
that the created route is considered according to the type and Syst. Saf., vol. 124, pp. 142–157, Apr. 2014.

VOLUME 10, 2022 23365

 J.-S. Lee et al.: Maritime Traffic Route Detection Framework Based on Statistical Density Analysis

[18] F. Goerlandt and J. Montewka, ‘‘Maritime transportation risk analysis: [41] L. Zhao and G. Shi, ‘‘A trajectory clustering method based on Douglas-
Review and analysis in light of some foundational issues,’’ Rel. Eng. Syst. Peucker compression and density for marine traffic pattern recognition,’’
Saf., vol. 138, pp. 115–134, Jun. 2015. Ocean Eng., vol. 172, pp. 167–456, Jan. 2019.
[19] K. D. Schwehr and P. A. McGillivary, ‘‘Marine ship automatic identifica- [42] D. H. Douglas and T. K. Peucker, ‘‘Algorithms for the reduction of the
tion system (AIS) for enhanced coastal security capabilities: An oil spill number of points required to represent a digitized line or its caricature,’’
tracking application,’’ in Proc. OCEANS, Sep. 2007, pp. 1–9. Cartographica, Int. J. Geographic Inf. Geovisualization, vol. 10, no. 2,
[20] A. Serry, The Automatic Identification System (AIS): A Data pp. 112–122, 1973.
Source for Studying Maritime Traffic. Barcelona, Spain: Maritime [43] A. Ram, S. Jalal, A. S. Jalal, and M. Kumar, ‘‘A density based algorithm
Transport, 2018. [Online]. Available: https://hal.archives-ouvertes.fr/hal- for discovering density varied clusters in large spatial databases,’’ Int. J.
01724104/document Comput. Appl., vol. 3, no. 6, pp. 1–4, Jun. 2010.
[21] H. Namgung and J.-S. Kim, ‘‘Collision risk inference system for maritime [44] N. Lu, M. Liang, R. Zheng, and R. W. Liu, ‘‘Historical AIS data-
autonomous surface ships using COLREGs rules compliant collision driven unsupervised automatic extraction of directional maritime traffic
avoidance,’’ IEEE Access, vol. 9, pp. 7823–7835, 2021. networks,’’ in Proc. IEEE 5th Int. Conf. Cloud Comput. Big Data Analytics
[22] L. Zhang, Q. Meng, and T. Fang Fwa, ‘‘Big AIS data based spatial-temporal (ICCCBDA), Apr. 2020, pp. 7–12.
analyses of ship traffic in Singapore port waters,’’ Transp. Res. E, Logistics [45] United Nations Convention on the Law of the Sea, UNCLOS, UN (United
Transp. Rev., vol. 129, pp. 287–304, Sep. 2019. Nations), New York, NY, USA, 1982, p. 45.
[23] H.-T. Lee, J.-S. Lee, H. Yang, and I.-S. Cho, ‘‘An AIS data-driven approach [46] M. Maritime and C. Agency, MGN 543(M+F) Safety of Navigation:
to analyze the pattern of ship trajectories in ports using the DBSCAN Offshore Renewable Energy Installations (ORELs). Guidance on UK Nav-
algorithm,’’ Appl. Sci., vol. 11, no. 2, p. 799, Jan. 2021. igation Practice, Safety and Emergency Response, MCA, Southampton,
[24] H.-T. Lee, J.-S. Lee, W.-J. Son, and I.-S. Cho, ‘‘Development of machine U.K., 2016, p. 18.
learning strategy for predicting the risk range of Ship’s berthing velocity,’’ [47] Marine Spatial Plan for the Belgian Part of the North Sea, Federal
J. Mar. Sci. Eng., vol. 8, no. 5, p. 376, May 2020. Public Service, Brussels, Belgium, 2016, pp. 42–43. [Online]. Available:
[25] J. Canny, ‘‘A computational approach to edge detection,’’ IEEE https://www.health.belgium.be/sites/default/files/uploads/fields/
Trans. Pattern Anal. Mach. Intell., vol. PAMI-8, no. 6, pp. 679–698, fpshealth_theme_file/19094275/Summary%20Marine%20Spatial%20
Nov. 1986. Plan.pdf
[26] N. Otsu, ‘‘A threshold selection method from gray-level histograms,’’ IEEE [48] A. H. Kvanli, R. J. Pavur, and K. B. Keeling, Concise Managerial
Trans. Syst., Man, Cybern., vol. SMC-9, no. 1, pp. 62–66, Jan. 1979. Statistics. Boston, MA, USA: Cengage Learning, 2006, pp. 81–82.
[27] M. Liang, R. W. Liu, S. Li, Z. Xiao, X. Liu, and F. Lu, ‘‘An unsupervised
learning method with convolutional auto-encoder for vessel trajectory JEONG-SEOK LEE received the B.S. degree
similarity computation,’’ Ocean Eng., vol. 225, Apr. 2021, Art. no. 108803. from the Division of Navigation Science, Korea
[28] R. W. Liu, M. Liang, J. Nie, W. Y. B. Lim, Y. Zhang, and M. Guizani,
Maritime and Ocean University, Busan, South
‘‘Deep learning-powered vessel trajectory prediction for improving smart
Korea, in 2014. From 2014 to 2017, he worked
traffic services in maritime Internet of Things,’’ IEEE Trans. Net. Sci. Eng.,
vol. 14, no. 8, pp. 2–3, Aug. 2015. as a Navigator for the third and second officers
[29] L. Wang, P. Chen, L. Chen, and J. Mou, ‘‘Ship AIS trajectory clustering: at SK Shipping Company, with experience in
An HDBSCAN-based approach,’’ J. Mar. Sci. Eng., vol. 9, no. 6, p. 566, bulk carriers, product oil-tanker carriers, and very
May 2021. large crude oil carriers. In 2018, he performed
[30] V. Fernandez Arguedas, G. Pallotta, and M. Vespe, ‘‘Maritime traffic related works as a Maritime Simulator Operator
networks: From historical positioning data to unsupervised maritime traffic and is conducting research using maritime traffic
monitoring,’’ IEEE Trans. Intell. Transp. Syst., vol. 19, no. 3, pp. 722–732, data. Conducted research using maritime data-based computing algorithm.
Mar. 2018. In particular, uses the GIS program in maritime. His research interest
[31] N. Forti, L. M. Millefiori, and P. Braca, ‘‘Unsupervised extraction of includes fusion research using artificial intelligence techniques.
maritime patterns of life from automatic identification system data,’’ in
Proc. OCEANS Marseille, Jun. 2019, pp. 1–5. HYEONG-TAK LEE received the B.S. degree in
[32] L. M. Millefiori, P. Braca, K. Bryan, and P. Willett, ‘‘Modeling nautical science course from the Department of
vessel kinematics using a stochastic mean-reverting process for long- Maritime Transportation Sciences, Korea Mar-
term prediction,’’ IEEE Trans. Aerosp. Electron. Syst., vol. 52, no. 5,
itime & Ocean University (KMOU), Busan, South
pp. 2313–2330, Oct. 2016.
[33] L. M. Millefiori, P. Braca, and G. Arcieri, ‘‘Scalable distributed change
Korea, in 2013, the M.E. degree from the Graduate
detection and its application to maritime traffic,’’ in Proc. IEEE Int. Conf. School, KMOU, in 2019, and the Ph.D. degree
Big Data (Big Data), Dec. 2017, pp. 1650–1657. from the Ocean Science and Technology (OST)
[34] L. M. Millefiori, P. Braca, and P. Willett, ‘‘Consistent estimation of School, KMOU, in 2022. He worked at Hanjin
randomly sampled Ornstein–Uhlenbeck process long-run mean for long- Shipping Company Ltd., as a Navigating Officer
term target state prediction,’’ IEEE Signal Process. Lett., vol. 23, no. 11, and worked for four years on container vessels.
pp. 1562–1566, Nov. 2016. He also worked at Training Ship HANBADA as an Instructor and a Chief
[35] M. Ester, H.-P. Kriegel, J. Sander, and X. Xu, ‘‘A density-based al- gorithm Officer for two and a half years. His research interests include maritime
for discovering clusters in large spatial databases with noise,’’ in Int. Conf. artificial intelligence and data science.
Data Mining Knowl. Discovery, vol. 4, pp. 226–231, 1996.
[36] R. Xu and D. C. Wunsch II, Sequential Data Clustering. Hoboken, NJ, IK-SOON CHO received the B.Sc. degree in
USA: Wiley, 2008, ch. 7, pp. 179–212. nautical science course, in 1996, the M.E. degree
[37] G. Wang, J. Meng, and Y. Han, ‘‘Extraction of maritime road networks in maritime transportation science from Korea
from large-scale AIS data,’’ IEEE Access, vol. 7, pp. 123035–123048, Maritime & Ocean University (KMOU), Busan,
2019.
South Korea, in 2000, and the Ph.D. degree from
[38] Z. Yan, Y. Xiao, L. Cheng, S. Chen, X. Zhou, X. Ruan, M. Li,
R. He, and B. Ran, ‘‘Analysis of global marine oil trade based on the Graduate School of Science and Technology,
automatic identification system (AIS) data,’’ J. Transp. Geography, vol. 83, Kobe University, Kobe, Japan, in 2005. He has
Feb. 2020, Art. no. 102637. worked with the Traffic Safety Audit Center,
[39] Technical Characteristics for an Automatic Identification System Using KMOU based on the Korea Maritime Safety Act,
Time Division Multiple Access in the VHF Maritime Mobile Frequency where he is currently working as a Professor with
Band, document ITU-R M. 1371-5, Radiocommunication Sector of the Division of Navigation Convergences Studies, College of Maritime
International Telecommunication Union (ITU-R), 2014, p. 114. Sciences. His research interests include maritime safety evaluation, ship
[40] D. P. Huttenlocher, G. A. Klanderman, and W. J. Rucklidge, ‘‘Comparing operation, and mooring safety analysis based on big-data using the machine
images using the Hausdorff distance,’’ IEEE Trans. Pattern Anal. Mach. learning and deep learning algorithm.
Intell., vol. 15, no. 9, pp. 850–863, Sep. 1993.

23366 VOLUME 10, 2022