Expert Systems With Applications 175 (2021) 114755

Contents lists available at ScienceDirect

Expert Systems With Applications

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

Visual analytic based ship collision probability modeling for ship

navigation safety
Ülkü Öztürk a, *, Hasan Alp Boz b, Selim Balcisoy b
a
Turkish Naval Forces, Ankara, Turkey
b
Sabanci University, Faculty of Engineering and Natural Sciences, Istanbul, Turkey

A R T I C L E I N F O A B S T R A C T

Keywords: This study presents a tangible visual analytic tool to analyse maritime traffic in spatio-temporal basis using AIS
Visual analytic data. This novel approach helps in understanding the macroscopic safety structure of both fairways and indi­
Ship collision vidual ships with evidences in microscopic level. Qualification of our system is demonstrated with 7-days AIS
Probability
trajectory collected from Mexican Gulf. We find out that spatio-temporal position pattern of encountered ships in
Maritime safety
Port Houston varies over time, significantly. In addition, the spatial distribution of ship accidents coincide with
proposed near-miss density areas. Furthermore, proposed tool is capable of capturing real accident cases. Field
experiments with domain experts have demonstrated that our approach helps in making realistic inferences
about navigational safety behaviour of both individual vessel and water area.

1. Introduction present challenges in extracting navigation safety measures. For

example, efforts of employing excessive number of parameters (Bukhari
Maritime transportation is the main artery of the world trade. Any et al., 2013) in the models provoke implicit risk (Chen et al., 2019)
maritime accident poses significant risks not only to human and envi­ identification that hinder the end-users to understand the risky situa­
ronment but also to economy. Proactive measures such as monitoring, tion, in details. In addition, some efforts on extracting navigational risk
controlling and decision-support tools are indispensable parts of such degree based on the cognitive knowledge of expert (Kao et al., 2007;
safety–critical domain. There is much more attention on navigation Wen et al., 2016; Zhang, Meng, & Fwa, 2017) are lacking the perceptual
safety since maritime transportation increases steadily. Therefore, nav­ perspective of navigational situation which in turn can support the
igation safety needs to be examined in every details to provide sus­ decision-making. Furthermore, most of these navigational collision risk
tainable and safe flow of world trade. assessment models focus on absolute risk assessment in current
Navigational collision risk (NCR) assessment studies in maritime encounter situation (Perera, Carvalho, & Guedes Soares, 2011) rather
domain compromise a number of significant output such as collision than a comparative risk assessment suggested by The World Association
avoidance (Lazarowska, 2017), navigation safety (Huang, Nieh, & Kuo, for Waterborne Transport Infrastructure (PIANC) for safe navigation.
2019; Zhang et al., 2015), facilitating Vessel Traffic Services (VTS) Due to the fact that there is no common agreement on NCR assessment
(Bukhari, Tusseyeva, & Kim, 2013), collision probability of waterways methodology, comparative risk assessment (in time or space) may offer
(Montewka et al., 2010). In maritime transportation, conventional distinguished opportunities for navigation safety. In contrast to NCR
collision risk assessment approaches include realist, proceduralist and methodology disagreement, more than half of the NCR assessment
constructivist models (Goerlandt & Montewka, 2015). The models of studies employ Time to Closest Point of Approach (TCPA) and Distance
these approaches employ system parameters (Montewka et al., 2010; to Closest Point of Approach (DCPA) parameters (Ozturk & Cicek,
Hashemi et al., 1995), layout persons (Qu, Meng, & Suyi, 2011) and 2019). Although TCPA and DCPA are the most common navigation
expert judgements, respectively to elicit the severity of ship encounters parameters, they can be insufficient to represent NCR degree in any
by means of collision risk. However, the complex nature of maritime encounter situation.
environment, and including many parameter such as ships variables, Visual analytic is a reasoning approach achieved by an interactive
traffic variables, environment variables and topographic variables, design. Since there is no fully automatic solution (Munzner, 2016) to

* Corresponding author.
E-mail addresses: [email protected] (Ü. Öztürk), [email protected] (H.A. Boz), [email protected] (S. Balcisoy).

https://doi.org/10.1016/j.eswa.2021.114755
Received 9 May 2020; Received in revised form 2 July 2020; Accepted 16 February 2021
Available online 24 February 2021
0957-4174/© 2021 Elsevier Ltd. All rights reserved.
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

NCR assessment for navigation safety, a visual analytic approach can dynamic (sea state, wind speed, visibility) factors of ship to assess NCR.
resolve these kinds of complexities and ambiguities mentioned above. Because of the some insufficient characteristics of DCPA and TCPA
The visual analytic approach in NCR assessment may has a number of (Goerlandt et al., 2015), Zhang et al. (2015) propose new parameter
contribution such as explicit risk identification, perceptual perspective called vessel conflict ranking operator based on distance, speed and
of encounter severity, filling the spatial information of TCPA/DCPA ships headings. In fact, this insufficiency arises from the fact that low
parameters and comparative risk assessment. Another compelling side DCPA value doesn’t mean a risky encounter when the encounter situa­
of visual analytic approach in NCR assessment is the competence of tion happens behind the port-starboard line. Similarly, (Szlapczynski,
trajectory data to this approach. Because the trajectory data has spatial 2006) proposed the temporary approach factor- the ratio of the distance
heterogeneity, auto-correlation in time and space scale (Andrienko between the two encountering ships to minimum acceptable distance-
et al., 2008), standard statistical techniques requiring independence of substituting the TCPA and DCPA parameters. Although this factor re­
variables may be improper for spatio-temporal data analysis. Although solves the misleading aspect of DCPA, its risk result is implicit for
there are visual analytic approaches analysing maritime trajectory data perceptual understanding of risk. As can be inferred from the detailed
(Scheepens et al., 2015; Scheepens et al., 2011a), domain-knowledge review (Ozturk & Cicek, 2019) of these studies, in addition to implicit
point of view is missing in them. In this study, we attempt to focus on risk identification (Balmat et al., 2009; Smierzchalski, 2000; Szlapc­
visualising the maritime traffic navigational risk by overcoming the zynski, 2006) and lacking spatial information (Chin & Debnath, 2009;
limitations mentioned above. In this, we leverage the huge potential of Ahn et al., 2012), there is neither a widely accepted NCR assessment
discovering knowledge from massive AIS ship movement trajectories. approach nor a NCR reference (Goerlandt & Montewka, 2015) unique to
Our main contribution is twofold. First, we present a novel approach, maritime application area.
interactive visual analytic in navigation safety (See supplementary files On the other hand, the limit values (safe distance or time of last
for introduction video). Second, we introduce a visual abstraction of two minute of avoidance) in determining collision risk base on the as­
spatio-temporal parameters (DCPA/TCPA) to show the users a terse sumptions of authors. For example, Tam and Bucknall (2013) state safe
representation of navigational collision risk structure. By this way, we passing distance as 0.25 nautical miles while Xu, Meng, Wang, and Ning
present comparative NCR assessment of both waterways and individual (2010) assume it as one nautical mile. Smierzchalski and Michalewicz
ships base on ship collision probability. Interviews with domain experts (1998) propose that safe passing distance can be changed by external
and studies with end-users needs lead us to focus on the visualisation of conditions and ship characteristics. Due to these ambiguities, applicable
comparative spatio-temporal NCR that can eliminate the bottlenecks of and realistic NCR assessment in dense traffic environment requires ev­
navigation safety management. idences that users/analysts can view because the validation of these NCR
The general framework of our study is as follows. In the next section, assessment studies is insufficient as Goerlandt and Kujala (2014) have
we review the related literature on NCR and visualisation of maritime stated.
traffic. In the third section, we introduce the design framework of visual
analytic approach, in detail. The results of our approach are presented in 2.2. Visualisation of maritime traffic
the forth section. The conclusions and feedbacks are given in the last
section. Further efforts on NCR assessment include visualisation, which can
provide not only abstraction of spatio-temporal maritime data but also
2. Related works evidence of NCR identification. NCR assessment and visualisation of
NCR can be deemed as nested tasks. While NCR assessment studies focus
Our focus is on the problem of exploring and visualisation of mari­ on extracting navigational risk indexes of individual vessels, visual­
time traffic safety associated with navigational collision risk in a realistic isation of NCR employs these indexes to map trajectory data.
way. Therefore, we have first conducted a literature survey related to Visualisation of maritime traffic behaviour has become a hot topic
navigational collision risk assessment. Then, we have reviewed the since AIS spatio-temporal trajectory data is an accessible repository.
maritime traffic trajectory visualisation literature, elaborately. Andrienko et al. (2008) classify these trajectory visualisations into three
subcategories, direct depiction, summary visualisation and pattern
2.1. Navigational collision risk extraction. While direct depiction shows each record for the analysts,
summary visualisation aggregate, generalise or sample the whole data.
Navigational collision risk assessment is a vital task and has great On the other hand, pattern extraction involves more sophisticated
importance in maritime navigation safety. It encompasses the decision techniques to present patterns to the user. Visualisation studies of AIS
making processes of many stakeholders such as deck officers, ship data have many summary visualisation goals such as collision risk,
management, port management, port authorities etc. While some of the traffic density, port statistics, anomaly detection (Iphar, Ray, & Napoli,
NCR assessment studies leverage the advantage of AIS data (Mou, Van 2020) and situational awareness etc. For example, some studies (Kim &
Der Tak, & Ligteringen, 2010), others employ simulation experiments or Jeong, 2016; Lei, 2019; Liu, Zhaolin, & Zheng, 2019; Lei, 2019) take
surveys. However, each of these studies differs in parameters they used. advantage of TCPA and DCPA parameters to extract ship collision
For example, Smierzchalski (2000) defines hexagon-shaped area in probability or near-miss encounters, and visually represent the obtained
which NCR has been formulated based on TCPA and DCPA, own speed, so-called absolute risk indicators. However, none of these mentioned
relative speed, ship length and ship breadth parameters. The safe dis­ studies take spatial positions of future encounter situation emphasised
tance in front of the bow is assumed to be 2 to 3 miles in this study. Chin by Zhang et al. (2015) into consideration. On the other hand, Szlapc­
and Debnath (2009) adopt TCPA and DCPA parameters with a regres­ zynski and Krata (2018) propose a visualisation approach for safe mo­
sion model fitted by survey results to assess NCR. Although this model is tion parameters mainly focusing on International Regulations for
employed in the Navigational Traffic Conflict Technique (Debnath & Preventing Collisions at Sea (COLREG) and Closest Point of Approach
Chin, 2010) to handle limited collision record problem, the spatial in­ (CPA). While this study takes future encounter situation positions into
formation, which will be discussed later, is lacking. Similarly, Ahn, account, it doesn’t summarise the maritime traffic behaviours from the
Rhee, and You (2012) employ adaptive neural network based on TCPA whole data. Silveira, Teixeira, and Guedes Soares (2013) conduct sta­
and DCPA parameters obtained by an interviewee-independent method. tistical analysis of Portugal Coast and assess risk profile of ship-to-ship
They obtained values of parameters from fuzzy membership functions collision risk. The study characterises the shipping routes by speed
modified with rule-based collision risk table. On the other hand, Balmat and bearings. This ship-to-ship collision risk model base on ship geom­
et al. (2009) adopt a fuzzy approach combining static (year of con­ etry, velocity and distance. In addition, some studies focus on densities
struction, flag state, gross tonnage, the number of companies) and to extract maritime traffic behaviours. Jin, Luo, and Gao (2018) provide

2
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

Fig. 1. Nine-stage design methodology framework (Sedlmair et al., 2012).

visual analysis of vessel trajectory, vessel density and vessel speed dis­ 3. Visual Analytic Framework for Navigational Collision Risk
tribution based on trajectory clustering algorithm. Zhang et al. (2017) Identification
find out that dense traffic and traffic accidents spots coincide by visu­
alising ship trajectories, speed and density. However, this finding con­ Determining the visual abstraction of a system is challenging and
tradicts that of Zhang et al. (2017). Other than NCR based visualisation, hosts many pitfalls. Due to the critical aspects of navigation safety, we
Santos et al. (2012) introduced a visual representation of ships routes consult one of the most important visual analytic design (Sedlmair,
clustered by their positions and directions. The primary goal of this Meyer, & Munzner, 2012) introducing nine-stage framework as shown
approach is to automate identification of traffic routes. Wen et al. (2016) in Fig. 1. Readers can follow our work in each step of this framework,
also extracted shipping routes to facilitate route planning. Cazzanti, including our problem characterisation and visual abstraction approach.
Davoli, and Millefiori (2016) incorporate Centre of Maritime Researches The description of subsections are as follows:
and Experimentation AIS database to provide visual analysis of port
visits statistics for maritime situational awareness. Wang et al. (2017) (i) Learn: This step includes building a solid knowledge for NCR
and Riveiro et al. (2009) apply a visual analytic tool to find anomalous visualisation from literature.
vessel behaviours to support coastal security authorities in maritime (ii) Winnow: This step aims to identify the collaborators with whom
situational awareness. Another maritime traffic visualisation study is the research study can process without any bottleneck.
proposed by Chen, Feng, and Peng (2015) incorporating AIS data to (iii) Cast: This step includes identifying the collaborators roles within
extract principal fairways of a navigation area. our research.
There are also visual analytic based approaches focusing on AIS (iv) Discover: The goal of this stage is to identify problems, needs and
trajectory data. For example, Scheepens et al. (2011a, 2011b) focus on requirements in NCR assessment domain.
trajectory density maps with line kernel density functions. This kind of (v) Design: This step introduces the visual solution to the problem
visualisation approach make maritime trajectory data more meaningful that revealed in discover step.
regarding drifters, ship density and sea lanes. Furthermore, they propose ( vi) Implement: This step produce software prototype to meet scal­
collision risk map based on mass and velocity. Other studies employing ability and other requirements.
line kernel density approach are proposed by Willems, Niels, Huub Van (vii) Deploy: This stage involves deploying visual NCR assessment
De Wetering, and Van Wijk (2011) trying to find stopping/fast-moving tool and gathering feedback about its use.
objects, and Lampe and Hauser (2011) focusing on drilling operations. (viii) Reflect: This step provides informative findings of our research
Wen et al. (2014) also propose a framework of route mining for traffic for visualisation design guidelines.
management use cases. Wang et al. (2019) introduce a brief review of (ix) Write: This includes writing stage of our research.
maritime traffic data visualisation and reveal the challenges in maritime
traffic data visualisation. However, these studies, focusing on visual­
isation technique, lack the identification of real interests of navigation 3.1. Learn: visualisation literature
domain regarding NCR. The main concern of these studies is density of
trajectory data rather than NCR of maritime traffic in spatio-temporal Visualisation can undertakes unique roles in complex problems such
dimension as discussed above, elaborately. as NCR assessment in which there is no common agreement. The
In this section, we have presented the studies related to NCR knowledge that human obtain from visualisation not only depend on the
assessment together with the visualisation of maritime traffic safety. In well designed human–computer interaction (HCI) but also the appro­
that, we have aimed to reveal that both NCR assessment and maritime priate way of visual abstraction that designer build. Our problem-driven
traffic visualisation studies are lacking to represent evidence-based and visualisation research claims to handle and produce validated solutions
comparative spatio-temporal characteristics of navigation safety, as well to a real-world problem. This venture may become a failure because of
as there is no visual analytic design study regarding NCR assessment. some pitfalls, unexpectedly. Therefore, we have adopted a design
methodology (Sedlmair et al., 2012) to refrain from these pitfalls. The
primary objective in analysing movement data is to understand and
characterise the movement behaviour of the entire population of moving
entities over the whole time period the data refer to (Andrienko et al.,

3
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

Fig. 2. Expert contribution.

2008). However, this study focus on NCR of maritime traffic. Hence, we 3.4. Discover: problem characterisation and abstraction
clearly review the literature and limitations to justify our work, in above
section. Around 90 percent of world trade is carried by the ships, and the
ports, busy waters, straits and channels are the bottlenecks of this flow.
3.2. Winnow: determining collaborators In these bottlenecks any accident may cause cascaded delay of de­
liveries, as well as environmental disasters. Therefore, authority make
In this early stage, we sticked to collaborate with a VTS chief oper­ navigation risk assessment studies and support navigation safety with
ator having 15-year experience as a pilot. We also observed the working technology, equipment and personnel to behave proactively. During our
area of VTS personnel evaluating maritime traffic safety and taking observation studies, we have noticed that there is no solution to assess
precautions. In our first meeting and following interviews, our interests maritime navigation safety by means of visual abstraction which can
have been clearly matched. The main problems, stated by VTS solve some problems. For example, while TCPA and DCPA parameters
personnel, are the lacking analysis tool of navigation safety, missing are prominent indicators as mentioned by above studies, they are also
historical data on navigational charts and inability to see evidences of incomplete to summarise maritime traffic risk behaviour in a realistic
dangerous passages. Our other collaborator was an official from a Port way. The fact that low DCPA values behind the port-starboard line are
Authority having 12 years of experience. Furthermore, we closely not risky is ignored by most of the risk assessment studies. Furthermore,
worked with 3 captains. In our researches, we effectively and closely the meaning of the risk results are implicit (Chen et al., 2019) in distance
interact these collaborators to get feedback and avoid some other pitfalls based approaches unless there is no visual representation. In addition to
as stated in (Munzner, 2008). From the literature review and early field difficulty of understanding of risk cause raised from including much
studies we conclude that the ability to analysis the navigation risk of any more constraints into the model, meeting all collision risk constraints
water area with evidences may provide cognitive and perceptual un­ may not be feasible (Szlapczynski & Krata, 2018). However, in our in­
derstanding of ship traffic. Our visual analytic approach is a candidate to terviews, we observed that domain experts insist on viewing the cause of
provide these evidences. risky encounters, explicitly. This is due to the fact that relying on poor
evidence may lead to erroneous inference (Goerlandt & Montewka,
3.3. Cast: identify collaborator roles 2015) that can result in dangerous events. Authority relies on the
assessment of operators or officials to avoid such dangerous events. In
Collaborator roles include front-line analyst, gatekeeper, connector, fact, the reliability of this assessment clearly depend on the experience
translator, co-author and fellow tool builder (Sedlmair et al., 2012). In period of VTS or domain expert. Therefore, providing the users with
our study the front-line analyst is a VTS chief operator handling actual visual abstraction of solid knowledge can enlighten them. Otherwise,
data analysis in the NCR visualisation tool. The gatekeeper is our team they can not clearly comprehend the real structure of waterway that
leader. We have identified 2 connectors (maritime academicians) that they are responsible.
meet us with domain experts and 3 translators (captains) that help us to We have conducted a series of interviews and meetings with our
abstract domain problems. For example, one translator inform us about front-line analyst and Port Authority Officer. We recorded our in­
significance of viewing nautical charts when tracking vessel trajectories. terviews, and went back to interviews in every step of our project design.
On the other hand, our fellow tool builder, who is a Ph.D student in Users point out the ability to find risky spots and relation of these risky
computer science, is in our team. spots with parameters such as TCPA and DCPA. In the course of our in­
terviews, needs and requirements were argued, and users suffer from
insufficient type of visual representations. The ability to track vessel
movement on a nautical chart also seen important. Therefore, the NCR

Fig. 3. Data processing steps.

4
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

Table 1 Here, the probability of collision is defined by two elements. NCandidate

Explanatory statistic of preprocessing results. refers the number of collision candidates which is the main focus of our
Length (meter) SOG Wind Speed Wind visual analytic approach. A ship on a collision course (Montewka et al.,
(knots) (knots) Direction 2010) or pairwise vessel encounters resulting in collision if no evasive
(Tanker/Cargo/ action is provided (Goerlandt & Kujala, 2011) are called collision can­
Passenger)
didates. On the other hand, PCausation is the probabilities of failures from
Mean 145/133/57 5.8 4.5 146 multiple perspectives of maritime environment such as human reli­
Std. 75/52/87 2.2 2.2 110 ability, organisational factors, mechanical failures and etc. However,
Dev.
Min. 0/0/0 0 0 0
NCandidate is a geometrical collision probability assuming no avoidance
25% 90/75/20 3.2 2.9 42 measure are taken, and characteristics of striking ships can play a sig­
50% 103/86/42 6.2 4.3 116 nificant role in determination NCandidate (Christian & Kang, 2017). While
75% 156/182/65 10.4 5.9 243 Macduff (1974) model assume uniform traffic flow, Fowler and Sørgård
Max. 340/289/105 41.6 15.6 360
(2000) model can assume non-collision event in the absence of evasion
action. Furthermore, Fujii et al. (1998) adopt ship domain approach to
literature outcomes and the feedbacks from domain experts lead us to derive NCandidate . Therefore, different models can lead different outcomes.
incorporate TCPA and DCPA navigation variables with subtle visual We follow a five step data process, shown below, in order to obtain
abstraction that users can observe each risky encounters. A detailed NCandidate .
chart representing the contribution of expert has been introduced in Our AIS data has been obtained from National Oceanic and Atmo­
Fig. 2. The interviews with experts helped us greatly to define the spheric Administration (NOAA) web service and include 7-days mari­
problem and possible solutions that we can visually represent in the time traffic data of Port Houston with one minute intervals. Tanker,
modelling stage (see Fig. 3). cargo and passenger ships are taken into consideration in this study. We
Current ship trajectory visualisation studies generally interested in handled missing and noisy data in the preprocessing step. Individual
visualisation techniques (Scheepens et al., 2011b; Scheepens et al., ship trajectories have been extracted from the AIS data by unique MMSI
2011a; Lampe & Hauser, 2011; Scheepens et al., 2015) with density number in the second step. However, AIS data contains various types of
maps. One of the most important corollary of these density map is that errors (Norris, 2007) such as data corruption, erroneous MMSI number,
they assume NCR as a feature of simple moving object. However, we target swaps and faulty position reports. To avoid these noisy data, we
learn from our interviews with domain experts that they are interested deleted trajectories having large discontinuities in time, and MMSI
in also the detailed understanding of risky encounters from these density numbers having insufficient trajectory as described in Aarsæther and
maps. Moan (2009). Although NOAA AIS data is one-minute interval, we split
ship trajectories into fifteen second intervals by interpolating each ship
3.5. Design: Data abstraction, visual encoding and interaction trajectories not to miss close encounters. As a result, we have obtained
3.755.761 unique trajectory observation in which there are 291 tanker,
In this section we eloquently describe the navigational collision risk. 152 cargo and 13 passenger ships. The explanatory statistic of the pre­
Our study proposes a probability-based risk analysis of ship collisions. processing results has been presented in Table 1.
The framework of Macduff (1974), shown below, has been widely In the third step, DCPA and TCPA parameters are obtained. One can
applied among risk studies. obtain DCPA and TCPA navigation variables as follows;

PCollision = NCandidate ∗ PCausation

Fig. 4. Ships positions and courses before the encounter situation (a) and in the CPA positions (b).

5
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

− ((Lon(t)OS − Lon(t)TS ) ∗ (Vx (t)OS − Vx (t)TS ) + (Lat(t)OS − Lat(t)TS ) ∗ (Vy (t)OS − Vy (t)TS ))
TCPA(t) =
(Vx (t)OS − Vx (t)TS )2 + (Vy (t)OS − Vy (t)TS )2
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
DCPA(t) = [(Lon(t)OS − Lon(t)TS ) + (Vx (t)OS − Vx (t)TS ) ∗ TCPA(t)]2 + [(Lat(t)OS − Lat(t)TS ) + (Vy (t)OS − Vy (t)TS ) ∗ TCPA(t)]2

While Lon(t)OS , Lat(t)OS , Vx (t)OS and Vy (t)OS represent coordinates CPA positions. However, some limit and threshold values have to be
and speed vectors of own ship (OS), Lon(t)TS , Lat(t)TS , Vx (t)TS and Vy (t)TS determined in order to extract these collision candidates. We consult
represent that of target ships (TS) that own ship encounters in time t, domain rules for these values. According to COLREG, all vessels should
respectivelty. These parameters informs about minimum distance and make a full appraisal of the situation and of the risk of collision (Rule 5) and
time that OS and TS will encounter. However, spatial information, very proper use shall be made of radar equipment to obtain early warning of risk of
critical for collision encounters, needs to be detailed in these parameters collision (Rule 7). There is no defined distance limit exist to determine
in order to avoid contradictory results. The detailed representation of the beginning of collision risk in COLREG as can be seen in Rule 5 and 7.
this situation is illustrated in Fig. 4. Nevertheless, IMO states that all ships have to comply with the advance
Own ship has a close encounter situation with two target ship as can and the tactical diameter of 4.5 and 5 ship length, respectively. Due to
be seen in Fig. 4(a) with DCPA value of 0.5 nm after 5 min. Most of the these limits of turning ability, this study assumes 5 ship length as a safe
current collision risk studies may designate same risk degree for these domain limit.
encounter situations. However, as can be seen in Fig. 4(b), one of the There should be critical range (Rc ) and distance (Dc ) values in order
target ship (Target Ship 2) has the close encounter situation in the aft to count an encounter situation as a collision candidate. Since COLREG
(back) side of the own ship and the other (Target Ship 1) has the close assume collision risk in head-on (Rule 14) and crossing situation (Rule
encounter situation forward side of the own ship. While, own ship and 15) clearly, forward of ship port-starboard line (-90 < θRelative < + 90)
target ship 2 encounter situation can be deemed as normal, the other can can be assumed as the critical range (Rc ) of CPA to be a collision
have some concerns. It is essential to evaluate TCPA and DCPA param­ candidate. As for the encounter situations of crossing and overtaking
eters along with spatial positions of ships as can be seen in Fig. 4 that happen between 90 and 112.5 degree on each of the sides of the
Therefore, it is important to sort out these encounters from AIS data to ship, the CPA position mostly happen in front of the port-starboard line,
infer realistic navigation safety results. We propose this visual analytic because CPA position is the future encounter, not the present. It is
tool in order to provide these spatial positions. Forth step deals with CPA important to note that we consult our domain experts regarding the
positions. Assume that there are N target ship in time t for one own ship. assumption of safe CPA position behind port-starboard line and they
The CPA positions for own ship can be obtained as follows; perceive CPA positions as safe behind the port-starboard line. On the
other hand, the Fujii’s approach assumes that the collision happens
Lon(t)TSi = Lon(t)TSi + TCPA(t)TSi ∗ Vx (t)TSi , i = 1, 2, …., N
when two ships are in the ”collision diameter” indicating an actual
physical contact rather than distance. However, ships may not avoid
Lat(t)TSi = Lat(t)TSi + TCPA(t)TSi ∗ Vy (t)TSi , i = 1, 2, …., N
collision when the distance is greater than ”collision diameter”. In fact,
this situation is regulated by PCausation factor. Furthermore, including
Lon(t)OSi = Lon(t)OSi + TCPA(t)OSi ∗ Vx (t)OSi , i = 1, 2, …., N
near-miss and close encounter situations can enhance the assessment of
navigation safety (Inoue & Kawase, 2007). Some studies (Pedersen,
Lat(t)OSi = Lat(t)OSi + TCPA(t)OSi ∗ Vy (t)OSi , i = 1, 2, …., N
2002; Fowler & Sørgård, 2000) adopt constant critical distance rather
Lon(t)TSi /Lat(t)TSi and Lon(t)OSi /Lat(t)OSi are the future positions of i. than ”collision diameter”. Montewka et al. (2010) propose a dynamic
th target ship and own ship at the CPA, respectively. Distances and true minimum distance to collision approach taking critical distance research
bearing of these CPA position from own ship has been calculated in further. Critical distance can be defined also as an ship area (Kaneko,
Python environment with geopy package. After obtaining distances (Di ) 2002) whose violation indicate collision. Therefore, CPA positions
and true bearing of CPA positions (Bearingi ), relative CPA positions can below 0.7 ship length have been assumed as critical distance in this
be obtained with own ship course over ground (COG) as; study according to field experiments and domain experts advice in order
to include near-miss and close encounter situations. This assumption
Bearing(t)Rel
i = Bearing(t)i − COG(t), i = 1, 2, …., N have based on the fact that AIS transponder is generally in the middle-aft
section of the ship, and ship location has been represented as a point
Lon(t)Rel rather than a two dimensional object. In the following sections, these
Rel
TSi = Di ∗ sin(Bearing(t)i ), i = 1, 2, …., N
encounters is called collision candidates.
Lat(t)Rel Rel
TSi = Di ∗ cos(Bearing(t)i ), i = 1, 2, …., N
TCPA is another critical parameter in determining the severity of

Lon(t)Rel Rel
TSi and Lat(t)TSi are relative CPA positions of TSi from own ship
Table 2
by which spatial information of DCPA and TCPA can be obtained for
Some examples for T90 .
realistic navigation safety assessment. A similar approach can be found
in (Marcjan, Gucma, & Voskamp, 2013) that focus on ship domain Ship Type Ship Length T90 (sec.)

construction. These CPA positions are the spatial indicators of naviga­ Tanker 113 90
tion safety that how close maritime traffic flows. It is important to note Tanker 241 159
that we assume linear ship movement in the calculation of CPA positions Tanker 391 203
LNG 285 138
for simplicity. Other approaches considering ship dynamics and envi­
Cargo 185 141
ronmental factors can obtain different CPA positions. However, low Cargo 294 126
current and wind speeds in Port Houston at the relevant time show that Cargo 318 126
this difference may be negligible especially in low TCPA encounters. Passanger 79 68
The fifth step focuses on extracting collision candidates from these Bulk 225 151

6
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

Fig. 5. The graphical representation of CPA positions.

The graphical representation of safe and collision candidate area and

Table 3
CPA positions are presented in Fig. 5 to explain the CPA position
PCausation from literature (Friis-Hansen et al., 2008).
approach, further.
Location PCausation (x10− 4 ) Reference The collision risk area and near-miss area are semi-circles with five
Dover Strait 5.18 Macduff (1974) and one ship length radius, respectively. Small circles are future relative
Dover Strait 3.15 Macduff (1974) positions of other ships. While CPA1 is the CPA position of TS1 that will
Oresund,Denmark 0.27 Karlsson et al. (1998) be in that relative position in t1 +10 min, CPA2 is the CPA position of TS2
Japanese Straits 0.49 Fujii et al. (1998) that will be in that position in t1 +1 min. Although TCPA of TS2 below
Japanese Straits 1.23 Fujii et al. (1998)
Dover Strait 1.11 Macduff (1974)
TC , it is not a collision candidate because its DCPA is more than Dc . CPA3
Dover Strait 0.95 Macduff (1974) and CPA4 are in the collision candidate area. TS3 will be in CPA position
Japanese Straits 1.10 Fujii et al. (1998) in t1 +10 min and it is not a collision candidate. However, TS4 will be in
Great Belt, Denmark 1.30 Pedersen et al. (1997) CPA position in t1 +1 min which can result in a collision candidate sit­
uation. Although TS5 has similar DCPA and TCPA values with TS4 , its
geographical position makes that encounter safe. In CPA positions point
encounter situation. There are also many studies proposing some critical
of view, a simple rule (TCPA<TC and DCPA<Dc ) or supervised learning
limits about TCPA. In our study, we adopt the Hilgert and Baldauf
model based on these parameters (Chin & Debnath, 2009; Smierzchal­
(1997)’s approach assuming development of danger of collision below 6
ski, 2000; Ahn et al., 2012) may not identify a collision candidate
min. They also define a critical time (T90 ) below which there is a danger
without a spatial information. The collision candidate extraction can be
of collision. In fact, T90 is the time required to accomplish 90 degree
represented as follows.
change of heading, corresponding to exact advance position that we
{
adopt in safe domain limit. T90 of some ship types according to their 1 if TCPAi < TCi andDistancei < Dci and − 90 < θRelativei < +90
pilot cards have been presented in Table 2. Candidatei =
0 otherwise
As can be seen in Table 2 T90 time is generally above 100 s, and 203 s
is the time required for a 391 meter length tanker to change its heading Second element of collision probability is PCausation examples of which
exactly 90 degree, which in turn provide to avoid a collision. A fitted is shown in Table 3.
polynomial function (R-square root value is 0.99) can be used to obtain It is possible to find out various PCausation values, acting as a refining
Tci of i.th individual ship. probability, as can be seen in Table 3. It is clear that possible variations

7
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

Fig. 6. The general flow of navigation safety visualisation.

Fig. 7. Density of ship collision candidates in Port Houston by geographic location.

8
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

Table 4 have been presented in the result and discussion section.

Number of collision candidates and expected weekly number of accidents.
Encounter type Collision candidates Causation factor Number of accidents 3.6. Implement: Prototypes, tool & usability
Overtaking 5 1.1 x 10 − 4
5.5 x 10− 4
It is greatly important to meet scalability requirement and choosing
Crossing 236 1.3 x 10− 4
306.8 x 10− 4
the right algorithm. First, we have built a quick prototype software in
Head-on 255 0.5 x 10− 4
127.5 x 10− 4
Python Dash environment to check drawbacks about visual abstraction
Total 496 - 439.8 x 10− 4 and HCI-orientation with front-line analysts. Then, the final software
has been written in JavaScript environment. The focus of tool usability is
performing a scalable interactive process of navigation safety to elicit
implicit risk identification. In our case, it takes approximately 192 h to
of PCausation may affect the final collision probability. Furthermore, calculate collision candidates in Intel Core i7, 8 GB RAM and 256 GB SSD
including all factors can be insurmountable. For brevity, we adopt computer. There is no doubt that choosing large time interval (more
PCausation factor of Waterway Risk Assessment Program (IWRAP) of In­ than 15 s.) may improve computing time and allow to process more data
ternational Association of Marine Aids to Navigation and Lighthouse (more than one week). However, employing large time interval (Goer­
Authorities (IALA), 1.1 x 10− 4 , 1.3 x 10− 4 , 0.5 x 10− 4 for over-taking, landt & Kujala, 2011) can also lead to ignoring the dangerous encoun­
crossing and head-on encounters, respectively (Friis-Hansen, Peter, ters. A map showing yearly maritime traffic density of Port Houston
Ravn, & Engberg, 2008). obtained with one hour interval data can provide new information to a
Multiplication of causation factor and the obtained total number of person from a different port management or industry. But, only more
collision candidates gives the collision probability of Port Houston. The frequently divided data set may provide new information to Port
general flow of our approach can be seen in Fig. 6. Further outcomes Houston management, having already much knowledge.

Fig. 8. Positions of ship collisions in Port Houston during 2001–2014.

9
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

Table 5 tool. The results are concerning Port Houston entrance, fairways and
Comparison of yearly accident number with expected number of collision. inner basins in March 2015. Fairway navigation safety and individual
Vessel type Encounter Number of Expected number of ship navigation safety perspectives have been introduced to provide a
type collision collision better understanding of the applicability of CPA positions approach and
Tanker Ship All 1.66 1.28 visual analytic tool, further.
Head-on 0.41 (% 25) 0.34 (% 28)
Crossing 0.25 (% 15) 0.92 (% 71)
Overtaking 1.0 (% 60) 0.02 (% 1) 4.1. Fairway Navigation Safety
Cargo Ship All 0.92 0.91
Head-on 0.33 (% 36) 0.27 (% 20) The Port Houston is ranked first in the United States in foreign
Crossing 0.41 (% 44) 0.64 (% 70) waterborne tonnage. 8,200 vessels and 215,000 barges move through
Overtaking 0.18 (% 20) 0 (% 0) 52-mile-long Houston Ship Channel each year. Obtaining navigation
Passenger All 0.33 0.09 safety indicators of that dense maritime traffic is of vital importance.
Ship The ship collision candidates density by geographic locations obtained
Head-on 0 (% 0) 0.06 (% 66) by proposed model is shown in Fig. 7.
Crossing 0.33 (% 100) 0.03 (% 34)
The results show that Galvestone entrance, Point Bolivar, upper and
Overtaking 0 (% 0) 0 (% 0)
middle Galvestone Bay and inner Houston Ship Channel geographical
locations are more prone to ship collisions. Number of collision candi­
3.7. Deploy, reflect and write dates and expected weekly number of accidents based on 2015 year AIS
data have been shown in Table 4.
The final interactive visual analytic tool, conforming graphical The total number of collision candidates is 496, of which density map
integrity, has been deployed for evaluating navigation safety. Findings has been shown in Fig. 7. % 47 and % 51 of the collision candidates are
show that front-line analysts are indeed assisted by the contribution of crossing and head-on encounters, respectively. In order to validate the
novel visual analytic tool approach for navigation safety. In fact, this conflict water areas that observed by proposed approach, we consult
assistance is especially visible in exploring the risk causes and spatial real ship collision events provided by IMO Global Integrated Shipping
encounter positions of collision candidates. Feedback and findings are Information System platform. It turns out that there are 35 ship collision
explained in conclusion section. accidents in Port Houston during the period 2001–2014. While our
Since our work is concerning navigation safety, the reflection of our model indicates 2.28 ship collision accidents per year, there are 2.5 in
design study relates to the other transportation modes such as air and Port Houston. This finding is also supported by a clear agreement be­
land transportation. Similar proximity based approaches in air and land tween the accident locations and the density of collision candidates as
transportation can enhance safety evaluation and reveal risky situation can be seen in Fig. 8. Although making inferences about model validity
encountered in analyses. from accident cases can be obscure because of the low number of such
Beside introducing a clear domain background, we have also events (Goerlandt & Kujala, 2011), this inference can be justified by
employed a visual analytic design methodology for a better under­ analysing concerning encounter types in detail.
standing of proposed ship collision probability model. Therefore, the Comparison of yearly accident number with expected number of
writing phases of this study include both visualisation and ship collision collision, by vessel and encounter types has shown in Table 5. There is a
model structures. clear discrepancy in collision numbers of all encounter type but the
trend of collision number by vessel type is resemble. However, the same
4. Results and discussion situation doesn’t prevail in encounter type distribution. Although the
number of expected number of collision of tanker ships in overtaking
This section presents the results of CPA positions based ship collision encounters is very rare, % 60 of tanker accidents took place in over­
probability model integrated with a novel interactive visual analytic taking encounters. An important reason of this can be shortness of time
interval (one week) that has been studied. Besides, it should be taken

Fig. 9. Daily CPA positions of Point Bolivar.

10
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

Fig. 10. Daily CPA positions of Point Bolivar by TCPA values. Each CPA position plots can be filtered by TCPA feature in order to extract immediate encounters.

Fig. 11. Comparison of three different waterway. Zone A, B and C can be assessed both individually and comparatively.

into consideration that the traffic density can be changed significantly (Fig. 10) with details on demand approach. It can be inferred from the
during the different period of year and navigational safety measures can Fig. 9 that there are many close port-to-port and starboard-to-starboard
be improved year by year. encounter situations and some of these encounters are immediate.
Since above mentioned explanations offer implicit risk identifica­ However, a detailed look as in Fig. 10 reveals that there are no imminent
tions, the evidences of these identifications can be obtained step by step encounter (TCPA below one minute) in head-on region. A close look can
with the proposed visual analytic tool. For example, navigation safety also reveals that there are more imminent port-to-port encounters that of
behaviour of a water area in Point Bolivar, one of the bottleneck of the starboard-to-starboard because of the traffic lane direction. It is also
port area, has been shown in Fig. 9 with CPA position approach. These interesting to find out that CPA positions in head-on region are more in
plots represent daily CPA positions belonging to all ships in that selected the weekend (8 and 14 of March 2015). The CPA position plots of
area (Colored Square). Furthermore, low TCPA values can be filtered selected water area indicate that in spite of the existence of close

11
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

Fig. 12. Macroscopic-microscopic relation of navigation safety. “Near-Miss Density Map” can be observed by CPA Position Plots, comparatively. Further individual
ship evaluation can be mapped in ”CPA Position Plots of Selected İndividual Ship Encounter” section. Encounter situations of each CPA points can also be viewed in
”Encounter Situation of Selected Ship” section.

there are no close CPA position in head-on region. However, there are
Table 6
more close encounters in March 13 and 14. Zone B is in the middle of the
Feedbacks of experts.
straight channel line as can be seen in Fig. A.14. As a result of the
Expert Assessment about the tool’s contribution location of Zone B one could expect that there are close encounter sit­
No.
uations since the total width of navigable space is about 160 meter
Expert 1 CPA plots are helpful in discovering approaching pattern of other ships, throughout the fairway. It can be seen in Fig. A.15 that there are close
and capturing the DCPA pattern of navigating ship.
encounters in port and port bow side below half ship length. In contrast
Expert 2 Revealing the evidences/details of dangerous situations, observing
waterway safety structure in time and distance dimension is the most
to Zone C and B, Zone A has many imminent encounters as can be seen in
critical and beneficary part of tool. Fig. 11. At first look one can say that Zone A is the most risky waterway
Expert 3 CPA plot pattern can be used as a benchmark for safe passage margin. among these Zones. While this difference can not be obtained in near-
Dense CPA positions below TCPA value of one minutes (especially in miss density plot (Fig. 6), CPA position plots allow to capture real
hean-on region) may indicate more risky positions. Circle patterns of
navigational conflict situations by our proposed visual abstraction
critical CPA positions can be used as a ship domain.
technique. In fact, Zone A has more accident event than Zone B and C as
can be seen in Fig. 8. A close look to Zone A can be seen in Fig. A.16.
encounter situations in time and distance dimension, these encounters
are not in ships course direction and most of DCPA values are above half 4.2. Individual ship navigation safety
ship length. Zoomed CPA position plots of selected area with TCPA
values below one minute have been presented in Fig. A.13. This Our visual analytical tool not only provide comparative risk assess­
approach can help us to understand the navigation risk structure of ment in selected waterways but also allow users to map selected en­
Fig. 6 in a more realistic way. counters of individual ships. Chen et al. (2016) also state that vessel
One of the most important outcome of our proposed visual analytical encounters are vital patterns noticed by ship management. In this step, it
tool is the opportunity to make comparative risk assessment rather than is critically important to explicit risky encounters, not provided by
absolute since there is no common agreement on navigational collision former studies. Such approach may provide evidences of macroscopic
risk assessment. CPA position plots (TCPA value of below one minutes is risk situation and enhance the validity and reliability of proposed nav­
highlighted) of three waterway region have been presented in Fig. 11. igation collision risk model. In Zone A there are close and imminent
This plots allow users to both evaluate and compare waterways navi­ encounters in head-on region in March 8 and 9 (see Fig. A.17). This
gation conflict situation. Although Zone C is in the port entrance area, means that there are ship encounters below one minute at the distance of
less than half ship length in head-on region. There are two possible

12
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

explanation to these close encounters, one is berthing/unberthing since change the granularity and abstraction level and as a result better pre­
there are berthing areas in Zone A, the other is the accident/near-miss. serve their mental map.
The graphical representation of macroscopic to microscopic point of
view in the case of March 9 with one of the close encounters in head-on 5. Conclusion
region has been shown in Fig. 12. At first waterways are selected as
described above. Then a ship (636092096) having a close encounter in In this study, we propose a first tangible visual analytic tool
March 9 has been selected (Selection 1) in Zone A. The CPA positions providing explicit risk identification regarding navigational safety. Real
that this ship has encountered with all other ships have been presented AIS data records of Port Houston were used. Our tool has two perspec­
as ”CPA Position Plots of Selected Individual Ship Encounters”. In fact, this tive, fairway navigation safety and individual ship navigation safety.
plot gives the navigation characteristic of that ship regarding TCPA and Fairway navigation safety perspective unveils the realistic situation (by
DCPA. However, the main difference from the former studies is that it CPA positions) in the macroscopic view of near-miss density map. On the
has geographical dimension providing the allocation of risky encoun­ other hand, individual navigation safety perspective, analyse the navi­
ters. The other selection (Selection 2) of a close encounter in head-on gation safety and critical encounters of a ship by CPA position plot.
region of our ship (636092096) in March 9 gives the ”Encounter Situa­ Furthermore, analyst can capture any two ship encounter situation on a
tion of Selected Ship”. This map show the current and next (after 0.89 map which allow evidences. Also the feedbacks from domain experts
min) position of individual ship (636092096) and selected ship (SS). It support our efforts in dealing with navigation safety structure.
turns out that next ship positions (after 53 s.) result in collision. The blue Our visual analytic approach fulfils three main gaps in navigation
circle in ”Near Miss Density Map of Port Houston” indicates the location of safety. At first, the shortcoming of spatial information in TCPA and
encounter situation. Considering the location, time and distance this DCPA has been complemented by CPA position approach. Then, implicit
should be a collision that happened in Port Houston. After a short navigational risk identification has been revealed by visual analytic
research, we have found that this collision happened between ”Carla approach. Another important contribution of visual analytic approach is
Maersk” and ”Conti Peridot” in March 9, 2015 at the same location. This to allow users to conduct comparative navigational risk assessment since
finding shows us that close encounters in head-on region are vital there is no common agreement on this subject. Furthermore, our field
regarding navigational safety. Furthermore, selected ship (636092096) experiments with domain experts have shown that this proposed CPA
CPA position plots by days introduce a unique representation that how position based visual analytic tool is realistic, reliable and serviceable.
close it interacts with other ships. For example, there are no imminent By applying our visual analytic approach, we find out that spatio-
ship encounters in March 8 and 11. In addition, there is an obscure inner temporal distribution pattern of ship encounters varies over time,
circle that covers close encounters about two ship length radius. Below significantly.
CPA positions plot, a line plot indicates average, minimum and Since we concentrate only upon limited information regarding
maximum CPA position distance for selected ship by day. This step-by- maritime environmental context, representing additional information
step evaluation of navigational collision risk is a typical example of how such as depth, current and ship characteristic can improve the inference
a macroscopic risk mapping (density maps etc.) turns into a microscopic capability. Nonlinear CPA position calculation, considering ship dy­
and more a realistic navigation safety measure. namics, wind speed and current speed, with different AIS data splitting
In order to validate our proposed visual analytic tool, we consult our can be researched in the future. Furthermore, some predictive tools can
domain experts and allow them to use our visual analytic tool (See be implemented as future works.
supplementary files). Experiments have been conducted by remote
screen sharing method. Main results and feedbacks of our experts is
Declaration of Competing Interest
presented in Table 6. Detained transcript of videos can be found in
supplementary files.
The authors declare that they have no known competing financial
Existing NCR visualisation studies provide a static platform for the
interests or personal relationships that could have appeared to influence
users. However, such systems require analysts to generate visualisations
the work reported in this paper.
per various configurations to perform comparisons on different granu­
larity levels. As a result, a significant cognitive load is burdened on the
Acknowledgements
analysts. In this study, the proposed visual analytic tool enables its users
to analyse the maritime traffic in the desired granularity and abstraction
Thanks to anonymous domain experts for supporting and being a
levels. Based on Shneiderman’s information-seeking mantra Shneider­
part of our visual analytic tool study. This research did not receive any
man (1996), analysts are able to observe the overall distribution of near-
specific grant from funding agencies in the public, commercial, or not-
miss encounters and then choose a group of waterways to compare the
for-profit sectors.
spatio-temporal CPA positions with a filtering mechanism to highlight
the desired TCPA range. Moreover, users can focus on a particular ship’s
CPA position and observe the encounter situation in detail. To conclude, Appendix A. Zoomed CPA plots
the tool presents a visual environment in which users can interactively

Fig. A.13. Zoomed CPA positions in which below one minute TCPA values are highlighted.

13
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

Fig. A.14. Zoomed plot of Zone B.

Fig. A.15. Zoomed CPA positions of Zone B in which below one minute TCPA values are highlighted.

14
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

Fig. A.16. Zoomed plot of Zone A.

Fig. A.17. Zoomed CPA positions plot of Zone A.

15
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

Appendix B. Supplementary data Karlsson, Max et al. (1998). ”Verification of ship collision frequency model”. In: Ship
collision analysis, pp. 117–121.
Kim, Kwang-Il, & Jeong, Jung Sik (2016). Visualization of ship collision risk based on
Supplementary data associated with this article can be found, in the near-miss accidents. In 2016 Joint 8th International Conference on Soft Computing and
online version, athttps://doi.org/10.1016/j.eswa.2021.114755. Intelligent Systems (SCIS) and 17th International Symposium on Advanced Intelligent
Systems (ISIS) (pp. 323–327). IEEE.
Lampe, Ove Daae, & Hauser, Helwig (2011). Interactive visualization of streaming data
References with kernel density estimation. In 2011 IEEE pacific visualization symposium (pp.
171–178). IEEE.
Aarsæther, Karl Gunnar, & Moan, Torgeir (2009). Estimating navigation patterns from Lazarowska, Agnieszka (2017). A new deterministic approach in a decision support
AIS. The Journal of Navigation, 62(4), 587–607. system for ship’s trajectory planning. Expert Systems with Applications, 71, 469–478.
Ahn, Jin-Hyeong, Rhee, Key-Pyo, & You, Young-Jun (2012). A study on the collision Lei, Po-Ruey (2019). ”Mining maritime traffic conflict trajectories from a massive AIS
avoidance of a ship using neural networks and fuzzy logic. Applied Ocean Research, data”. In: Knowledge and Information Systems, pp. 1–27.
37, 162–173. Liu, Zihao, Zhaolin, Wu., & Zheng, Zhongyi (2019). A novel framework for regional
Andrienko, Gennady et al. (2008). Geovisualization of dynamics, movement and change: collision risk identification based on AIS data. Applied Ocean Research, 89, 261–272.
key issues and developing approaches in visualization research. Macduff, T. (1974). The probability of vessel collisions. Ocean Industry, 9(9).
Andrienko, Gennady et al. (2008). ”Visual analytics methods for movement data”. In: Marcjan, K., Gucma, L., & Voskamp, A. (2013). ”Vessel Traffic Stream Analysis in
Mobility, data mining and privacy. Springer, pp. 375–410. Vicinity of The Great Belt Bridge”. In: Marine Navigation and Safety of Sea
Balmat, Jean-Francois, et al. (2009). MAritime RISk Assessment (MARISA), a fuzzy Transportation: Navigational Problems, p. 109.
approach to define an individual ship risk factor. Ocean Engineering, 36(15-16), Montewka, Jakub, et al. (2010). Probability modelling of vessel collisions. Reliability
1278–1286. Engineering & System Safety, 95(5), 573–589.
Bukhari, Ahmad C., Tusseyeva, Inara, Kim, Yong-Gi, et al. (2013). An intelligent real- Mou, Jun Min, Van Der Tak, Cees, & Ligteringen, Han (2010). Study on collision
time multivessel collision risk assessment system from VTS view point based on avoidance in busy waterways by using AIS data. Ocean Engineering, 37(5-6),
fuzzy inference system. Expert Systems with Applications, 40(4), 1220–1230. 483–490.
Cazzanti, Luca, Davoli, Antonio, & Millefiori, Leonardo M (2016). Automated port traffic Munzner, Tamara (2008). ”Process and pitfalls in writing information visualization
statistics: From raw data to visualisation. In 2016 IEEE International Conference on Big research papers”. In: Information visualization. Springer, pp. 134–153.
Data (Big Data) (pp. 1569–1573). IEEE. Munzner, Tamara (2016). Visualization analysis and design: keynote address. Journal of
Chen, Chen, et al. (2016). Information visualization of AIS data. In In: 2016 International Computing Sciences in Colleges, 32(1), 106–107.
Conference on Logistics, Informatics and Service Sciences (LISS) (pp. 1–8). IEEE. Norris, Andy (2007). AIS implementation–success or failure? The Journal of Navigation,
Chen, Jinhai, Feng, Lu., & Peng, Guojun (2015). A quantitative approach for delineating 60(1), 1–10.
principal fairways of ship passages through a strait. Ocean Engineering, 103, 188–197. Ozturk, Ulku, & Cicek, Kadir (2019). Individual collision risk assessment in ship
Chen, Pengfei, et al. (2019). Probabilistic risk analysis for ship-ship collision: State-of- navigation: A systematic literature review. Ocean Engineering, 180, 130–143.
the-art. Safety Science, 117, 108–122. Pedersen, Preben Terndrup (2002). Collision risk for fixed offshore structures close to
Chin, Hoong Chor, & Debnath, Ashim Kumar (2009). Modeling perceived collision risk in highdensity shipping lanes. Proceedings of the Institution of Mechanical Engineers, Part
port water navigation. Safety Science, 47(10), 1410–1416. M: Journal of Engineering for the Maritime Environment, 216(1), 29–44.
Christian, Robby, & Kang, Hyun Gook (2017). Probabilistic risk assessment on maritime Pedersen, Preben Terndrup, Hansen, Peter Friis, & Nielsen, Lars Peter (1997).
spent nuclear fuel transportation (Part II: Ship collision probability). Reliability Probabilistic Analysis of Collision Damages with Application to ro-Ro Passenger
Engineering & System Safety, 164, 136–149. Vessels: Identification of Collision Hazards. In.
Debnath, Ashim Kumar, & Chin, Hoong Chor (2010). Navigational traffic conflict Perera, L. P., Carvalho, J. P., & Guedes Soares, C. (2011). Fuzzy logic based decision
technique: a proactive approach to quantitative measurement of collision risks in making system for collision avoidance of ocean navigation under critical collision
port waters. The Journal of Navigation, 63(1), 137–152. conditions. Journal of Marine Science and Technology, 16(1), 84–99.
Fowler, Timothy G, & Sørgård, Eirik (2000). Modeling ship transportation risk. Risk Qu, Xiaobo, Meng, Qiang, & Suyi, Li (2011). Ship collision risk assessment for the
Analysis, 20(2), 225–244. Singapore Strait. Accident Analysis & Prevention, 43(6), 2030–2036.
Friis-Hansen, Peter, Ravn, E.S. & Engberg, P.C. (2008). ”Basic modelling principles for Riveiro, Maria, & Falkman, Göran (2009). Interactive visualization of normal behavioral
prediction of collision and grounding frequencies”. In: IWRAP Mark II Working models and expert rules for maritime anomaly detection. In 2009 Sixth International
Document, pp. 1–59. Conference on Computer Graphics, Imaging and Visualization (pp. 459–466). IEEE.
Fujii, Y. & Mizuki, N. (1998). “Design of VTS systems for water with bridges”. In: Santos, Maribel Yasmina et al. (2012). ”Automated traffic route identification through
Proceedings of the International Symposium on Advances in Ship Collision Analysis. the shared nearest neighbour algorithm”. In: Bridging the Geographic Information
Copenhagen, Denmark, pp. 10–13. Sciences. Springer, pp. 231–248.
Goerlandt, Floris, & Kujala, Pentti (2011). Traffic simulation based ship collision Scheepens, Roeland, et al. (2011a). Composite density maps for multivariate trajectories.
probability modeling. Reliability Engineering & System Safety, 96(1), 91–107. IEEE Transactions on Visualization and Computer Graphics, 17(12), 2518–2527.
Goerlandt, Floris, & Kujala, Pentti (2014). On the reliability and validity of ship–ship Scheepens, Roeland et al. (2011b). ”Interactive visualization of multivariate trajectory
collision risk analysis in light of different perspectives on risk. Safety Science, 62, data with density maps”. In: 2011 IEEE Pacific Visualization Symposium. IEEE, pp.
348–365. 147–154.
Goerlandt, Floris, & Montewka, Jakub (2015). A framework for risk analysis of maritime Scheepens, Roeland, et al. (2015). Visualization, selection, and analysis of traffic flows.
transportation systems: a case study for oil spill from tankers in a ship–ship collision. IEEE Transactions on Visualization and Computer Graphics, 22(1), 379–388.
Safety Science, 76, 42–66. Sedlmair, Michael, Meyer, Miriah, & Munzner, Tamara (2012). Design study
Goerlandt, Floris, & Montewka, Jakub (2015). Maritime transportation risk analysis: methodology: Reflections from the trenches and the stacks. IEEE Transactions on
review and analysis in light of some foundational issues. Reliability Engineering & Visualization and Computer Graphics, 18(12), 2431–2440.
System Safety, 138, 115–134. Shneiderman, B. (1996). The eyes have it: a task by data type taxonomy for information
Goerlandt, Floris, et al. (2015). A risk-informed ship collision alert system: framework visualizations. In Proceedings 1996 IEEE Symposium on Visual Languages (pp.
and application. Safety Science, 77, 182–204. 336–343).
Hashemi, Ray R, et al. (1995). A neural network for transportation safety modeling. Silveira, P. A. M., Teixeira, A. P., & Guedes Soares, C. (2013). Use of AIS data to
Expert Systems with Applications, 9(3), 247–256. characterise marine traffic patterns and ship collision risk off the coast of Portugal.
Hilgert, Helmut, & Baldauf, Michael (1997). A common risk model for the assessment of The Journal of Navigation, 66(6), 879–898.
encounter situations on board ships. Deutsche Hydrografische Zeitschrift, 49(4), Smierzchalski, R. (2000). Ships’ domains as a collision risk at sea in the evolutionary
531–542. trajectory planning. In WIT Transactions on Ecology and the Environment 45.
Huang, Juan-Chen, Nieh, Chung-Yuan, & Kuo, Hsin-Chuan (2019). Risk assessment of Smierzchalski, Roman & Michalewicz, Zbigniew (1998). “Adaptive modeling of a ship
ships maneuvering in an approaching channel based on AIS data. Ocean Engineering, trajectory in collision situations at sea”. In: 1998 IEEE International Conference on
173, 399–414. Evolutionary Computation Proceedings. IEEE World Congress on Computational
Inoue, Kinzo & Kawase, M. (2007). Innovative probabilistic prediction of accident Intelligence (Cat. No. 98TH8360). IEEE, pp. 342–347.
occurrence. In: TransNav, International Journal on Marine Navigation and Safety od Szlapczynski, Rafal (2006). A unified measure of collision risk derived from the concept
Sea Transportation 1.1. of a ship domain. The Journal of Navigation, 59(3), 477–490.
Iphar, Clément, Ray, Cyril, & Napoli, Aldo (2020). Data integrity assessment for maritime Szlapczynski, Rafal, & Krata, Przemyslaw (2018). Determining and visualizing safe
anomaly detection. Expert Systems with Applications, 147, Article 113219. motion parameters of a ship navigating in severe weather conditions. Ocean
Jin, Liang, Luo, Zhengyi, & Gao, Sh.u. (2018). Visual Analytics Approach to Vessel Engineering, 158, 263–274.
Behaviour Analysis. The Journal of Navigation, 71(5), 1195–1209. Tam, CheeKuang, & Bucknall, Richard (2013). Cooperative path planning algorithm for
Kaneko, Fujio (2002). Methods for probabilistic safety assessments of ships. Journal of marine surface vessels. Ocean Engineering, 57, 25–33.
Marine Science and Technology, 7(1), 1–16. Wang, Guizhen, et al. (2017). TraSeer: A visual analytics tool for vessel movements in the
Kao, Sheng-Long, et al. (2007). A fuzzy logic method for collision avoidance in vessel coastal areas. In 2017 IEEE International Symposium on Technologies for Homeland
traffic service. The Journal of Navigation, 60(1), 17–31. Security (HST) (pp. 1–6). IEEE.

16
Ü. Öztürk et al. Expert Systems With Applications 175 (2021) 114755

Wang, Kai, et al. (2019). Maritime Traffic Data Visualization: A Brief Review. In 2019 Willems, Niels, Huub Van De Wetering, & Van Wijk, Jarke J. (2011). ”Evaluation of the
IEEE 4th International Conference on Big Data Analytics (ICBDA) (pp. 67–72). IEEE. visibility of vessel movement features in trajectory visualizations”. In: Computer
Wen, Rong et al. (2016). “Spatio-temporal route mining and visualization for busy Graphics Forum. Vol. 30. 3. Wiley Online Library, pp. 801–810.
waterways”. In: 2016 IEEE International Conference on Systems, Man, and Xu, Q.Y., Meng, X.Y., & Wang, Ning, (2010). ”Intelligent evaluation system of ship
Cybernetics (SMC). IEEE, pp. 000849–000854. management”. In: TransNav, International Journal on Marine Navigation and Safety
Wen, Xu et al. (2016). ”Ship Automatic Collision Avoidance by Altering Course Based on od Sea Transportation 4.4.
Ship Dynamic domain”. In: 2016 IEEE Trustcom/BigDataSE/ISPA. IEEE, pp. Zhang, Liye, Meng, Qiang, & Fwa, Tien Fang (2017). Big AIS data based spatial-temporal
2024–2030. analyses of ship traffic in Singapore port waters. In Transportation Research Part E.
Wen, Yu Ting et al. (2014). “Routeminer: Mining ship routes from a massive maritime Logistics and Transportation Review.
trajectories”. In: 2014 IEEE 15th International Conference on Mobile Data Zhang, Weibin, et al. (2015). A method for detecting possible near miss ship collisions
Management. Vol. 1. IEEE, pp. 353–356. from AIS data. Ocean Engineering, 107, 60–69.

17