See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/351344003

An empirical ship domain based on evasive maneuver and perceived collision

risk

Article in Reliability Engineering & System Safety · September 2021

DOI: 10.1016/j.ress.2021.107752

CITATIONS READS

92 457

6 authors, including:

Lei Du Osiris A. Valdez Banda

Aalto University Aalto University
31 PUBLICATIONS 686 CITATIONS 99 PUBLICATIONS 2,503 CITATIONS

SEE PROFILE SEE PROFILE

Yamin Huang Floris Goerlandt

Wuhan University of Technology Dalhousie University
90 PUBLICATIONS 2,748 CITATIONS 204 PUBLICATIONS 8,007 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Lei Du on 05 May 2021.

The user has requested enhancement of the downloaded file.

1 An empirical ship domain based on evasive maneuver and perceived collision risk
2 Lei Du1, Osiris A. Valdez Banda1, Yamin Huang2,3, Floris Goerlandt4, Pentti Kujala1, Weibin Zhang5,*
3 1 Aalto University, School of Engineering, Department of Mechanical Engineering, Espoo, Finland
4 2 National Engineering Research Center for Water Transport Safety (WTS), Wuhan, China
5 3 Wuhan University of Technology, Intelligent Transportation System Research Center, Wuhan, China

6 4 Dalhousie University, Department of Industrial Engineering, Halifax, Nova Scotia, B3H 4R2, Canada

7 5 School of Electronic and Optical Engineering, Nanjing University of Science and Technology, Nanjing, 210094, China

8 Abstract
9 This paper introduced a new ship domain concept and an analytical framework. The ship domain takes the point
10 of the ship’s first evasive maneuver as a basis and correlates it with the navigator-perceived collision risk level.
11 The first evasive maneuver of a ship is detected based on the ship turning point identification and ship intention
12 estimation. The available maneuvering margin (AMM) is utilized as a proxy to measure the perceived collision
13 risk by the navigator. Interpreting the first evasive maneuver in terms of this AMM over a large sample of vessel
14 encounters taken from automatic identification system (AIS) data finally enables an empirical estimation of the
15 size of this ship domain. The method is applied to AIS data in the Northern Baltic Sea, and separate ship domains
16 are constructed for the give-way and stand-on vessels with different maneuverability characteristics. Compared
17 to the existing proximity-based ship domain, this ship domain explicitly incorporates the dynamic nature of the
18 encounter process and the navigator’s evasive maneuvers. Several advantages of this proposed ship domain
19 concept and limitations of the presented modeling approach are discussed. Finally, possible future applications
20 are explained, including waterway safety assessment and navigational decision support systems to reduce ship-
21 ship collision risk.
22 Keywords: Maritime safety; Ship maneuverability; Ship domain; AIS data; Velocity obstacle; Ship-ship collision

23 1. Introduction
24 Ship collision, as one of the most frequently occurring accidents at sea (Kujala et al., 2009; Du et al., 2020b and
25 2021; Zhang et al., 2020), has attracted significant attention in academic research. In the past decades, various
26 concepts and techniques have been proposed for preventing ship collision accidents or analyzing their spatio-
27 temporal occurrence patterns and risks in waterways (Valdez Banda et al., 2015, 2019; Szlapczynski and
28 Szlapczynska, 2016; Zhang et al., 2016; Fan et al., 2020; Kulkarni et al., 2020; Gil et al., 2019 and 2020; Zhu et
29 al., 2020; Rong et al., 2021). However, the occurrence of marine casualties and incidents remains stable at a high
30 level (EMSA, 2020).
31 The timing when a ship takes evasive maneuvers is critical for the success of collision avoidance. Many methods
32 have been proposed to help navigators find the proper timing for collision avoidance actions. For instance, the last
33 time to take action (LTTA) (Zhuo and Tang, 2008), the minimum distance to collision (MDTC) (Montewka et al.,
34 2010, 2014), the last line of defense (LLoD) (Baldauf et al., 2017) have been adopted to inform the navigator of
35 imminent danger. Szlapczynski et al. (2018) combine ship maneuverability and ship domain (SD) to determine
36 the last moment when a particular collision avoidance maneuver can still be successfully performed. By using this
37 approach, the critical condition for a ship to take evasive action is quantified, beyond which a collision cannot be
38 avoided. There are many studies on the construction of a collision alert system (CAS) to alert the ship to act
39 properly (Baldauf et al., 2011; Simsir et al., 2014; Goerlandt et al., 2015). However, their applicability is limited
40 by certain assumptions made to simplify the encounter process, such as the ship sailing in a straight line with
41 constant speed. Other work has focused on the concept of the SD as an area which the navigator would like to
42 keep clear from other ships, for navigational safety reasons, see Szlapczynski and Szlapczynska (2017) for a recent
43 review. Such domains are typically defined based on proximity information, i.e., on the distance between the
44 vessels in the encounter, see e.g., Hansen et al. (2013) and Zhang and Meng (2019) for empirically estimated
45 domains based on AIS data. While many SD formulations have been proposed, these originate from the idea that
46 a certain area around a vessel needs to be clear from other vessels, based on ideas originally presented by Goodwin
47 (1975). However, these existing SDs cannot be utilized as a critical criterion with direct application in collision
48 avoidance (Montewka et al., 2020). Therefore, efforts are still required on how to assist navigators to perform
49 maneuvers for collision avoidance based on these existing SDs.
50 Existing research attests that the timing of ship taking evasive maneuvers is primarily affected by the risk
51 perceived by the navigator (Chauvin and Lardjane 2008; Kim 2020). In this work, the navigator-perceived risk
52 refers to their understanding and tolerance of a collision, which depends on their experience, their understanding
53 of the regulations, and the difficulty in performing evasive maneuvers for collision avoidance. Hence, in this
54 article, a new concept of a ship domain is proposed, based on the timing of evasive maneuvering to avoid a ship-
55 ship collision, which is associated with the perceived risk level at the time of taking such an evasive maneuver.
56 The risks intuitively perceived by the navigator are normally interpreted from Automated Radar Plotting Aid
57 (ARPA) information (Statheros et al., 2008) by setting some critical values based on the navigators’ understanding
58 of ship maneuverability or even company rules, etc. Information, such as the Distance at Closest Point of
59 Approach (DCPA) and Time to Closest Point of Approach (TCPA) plays a key role in this assessment. The
60 perceptions of risk may vary for different navigators (Nicholas, 2006) and many methods have been proposed to
61 analyze factors affecting the perceived risk of collision (Aydogdu et al., 2010; Zhang et al., 2019; Kim, 2020).
62 The pilot-perceived collision risk has been used by Chin and Debnath (2009) as a basis for developing a collision
63 alert system. The stakeholders’ perception of risk has been measured by a generic fuzzy analytical hierarchy
64 process method (Aydogdu, 2014). Nonetheless, there is currently no SD concept which is defined based on the
65 relationship between the collision risk perceived by the navigator and the timing of taking action for collision
66 avoidance.
67 Inspired by the techniques for constructing an empirical SD, such as those used by Pedersen et al. (2013) and
68 Zhang and Meng (2019), statistical features of the collision risk level as perceived by the navigator at the time
69 when the evasive maneuver is initiated can be obtained from AIS data. The concept of available maneuvering
70 margins (AMM) developed in (Du et al., 2020c; Huang and van Gelder, 2020) is adopted as a proxy for the
71 navigator-perceived risk when the ship starts to take evasive maneuvers, for two reasons. First, the perceived risk
72 of a conflict relates strongly to the level of handling difficulty forced on the mariner (Inoue, 2000). When
73 navigators manipulate a ship, its responsiveness is restricted by its maneuverability (Hong and Yang, 2012).
74 Second, risk measures independent of conflict resolution may lead to inaccurate detection of actual danger (Chen
75 et al., 2018). Under the same circumstances, ships with a higher AMM have a greater possibility of eliminating
76 risks, so the risks are relatively lower. However, this is not often taken into consideration (Du et al., 2020c).
77 Therefore, the aims of this article are two-fold. First, to measure the risk perceived by the navigator, a risk
78 perception-based SD is proposed based on the concept of AMM. Second, the boundary of this SD concept is
79 empirically determined based on historical AIS data, utilizing various models to interpret vessel encounters in
80 terms of the timing of evasive maneuvers and the navigator-perceived collision risk. This proposed SD is a
81 practical SD that reveals under what conditions the navigator will start to maneuver for collision avoidance. The
82 conception and construction of this novel SD provides new insights into the behavior of navigators in collision
83 avoidance contexts. The utilization of this risk perception-based SD can support collision detection and conflict
84 resolution. Specifically, this risk perception-based SD can help navigators to understand the collision risk and
85 prompts them to prepare evasive maneuvers for collision avoidance. This also can help the traffic management
86 authorities understand the traffic risks and then some instructions can be made to guide the ship maneuver.
87 Therefore, this contribute to a reduction in ship collisions and the improvement of water traffic management
88 through various possible applications of this domain concept.
89 The remainder of this work is arranged as follows. Section 2 consists of a literature review on the existing research
90 about the SD. Section 3 explains the methodology of building this SD. Case studies are introduced in Section 4 to
91 demonstrate the proposed methodology using AIS data from the Northern Baltic Sea. Discussion and conclusions
92 are addressed in Section 5 and Section 6, respectively.

93 2. Literature review on ship domain

94 The SD, which has a profound impact on modern navigation technology, was proposed in the 1970s for estimating
95 waterway capacity. Fujii and Tanaka (1971) observed from radar data that most navigators would avoid entering
96 a certain region around a vessel, and this region was defined as the SD. Based on their work, various types of SD
97 have been developed, and some major changes are briefly highlighted below (Figure 1). A comprehensive review
98 of SD and its applications can be found in Szlapczynski and Szlapczynska, 2017.
99 First, the probabilistic boundary of the domain was developed. In the early stage, the boundary of the domain was
100 deterministic, e.g., in (Goodwin, 1975). Thus, when we apply the domain in collision alarms, the collision risk
101 only has two values, 1 or 0 (i.e., violation or not), which differs from the interpretation of navigators. Zhao et al.
102 (1993) then explained the SD by analogy with the personal space from physiology and introduced a fuzzy SD
103 with a probabilistic boundary. This idea was widely accepted, and many researchers developed probabilistic
104 domains, see for example papers by (Gucma and Marcjan, 2012, Zhang and Meng, 2019), etc.
105 Second, the data-based SD has been intensively studied in recent years. The original SD was based on offshore
106 radar, e.g. papers by (Coldwell, 1983, Fujii and Tanaka, 1971). Later on, many researchers modified the original
107 SD by incorporating experts’ knowledge (Zhu et al., 2001). For instance, Goodwin (1975) suggested three-section
108 domains, incorporating the responsibilities of the ship addressed in International Regulations for Preventing
109 Collisions at Sea (COLREGs, 1972). Kijima and Furukawa (2003) adopted a two semi-ellipse domain concerning
110 ship maneuverability, where the two semi-ellipses share the same short axis. Wang (2010) formulated a quaternion
111 SD that consists of four quarter-ellipses (or triangles). In brief, over a period of time, modifications based on the
112 original SD that incorporated experts’ knowledge were popular. However, in recent years, due to the spreading of
113 AIS (Mou et al., 2010), researchers have gained a powerful tool to investigate the SD from the perspective of the
114 ship. Since then, many researchers have proposed a new form of the SD, e.g. (Hansen et al., 2013, Zhang and
115 Meng, 2019).
116 Third, the dynamic SD was developed for collision prevention. The static SD is helpful for capacity estimation,
117 but not for dynamic collision prevention since working conditions are time-varying. To fill the gap, one group of
118 researchers developed a series of static SDs and chose the most appropriate one in for corresponding working
119 conditions (Pietrzykowski, 2008), e.g., open sea or restricted area (Wielgosz, 2017), different courses
120 (Pietrzykowski and Uriasz, 2008), different encounter types (Coldwell, 1983, Fiskin et al., 2020), etc. Another
121 group of researchers employed motion simulations in the construction of the domain (or distance), e.g., MDTC
122 (Montewka et al., 2012), action lines (Szlapczynski et al., 2018), etc. For instance, Gil et al. (2020) simulated two-
123 ship encounters and identified the region surrounding one ship that the other ship needs to take action before
124 violating and this region is also called the “critical area”.

Fujii et al. (1971)

One Ellipse
OS

Ship Domain Goodwin (1975)

Three Sections

Radar-based ColdWell (1983) Davis (1980)

Domain Restricted Waters Circle domain

Cooley (1983)
Irregular three
sections
Zhu (2001)
Kijima (2003) Neutral Network
Two semi-ellipses
Expert-based
Domain Zhao (1993)
Fuzzy Boundary

Zhang & Meng

(2019)
Wang (2010) Probabilistic SD
Quaternion Domain Pietrzykowski
(2008) Ship Fuzzy
Domain
Wang (2013)
Quaternion Domain + Pietrzykowski
more factors (2009) Fuzzy
+Open Sea
AIS-based
Domain Hansen et al., (2013)
One Ellipse

125
126 Figure. 1 Typical ship domains to illustrate the development of its application in collision prevention
127 From the development of the SD, we can clearly see that the purpose of the SD switches from capacity estimation
128 (or waterway safety assessment) to collision avoidance, specifically, e.g., conflict precaution (finding the moment
129 in time to pay attention to approaching dangers), and collision avoidance alarm (finding the moment in time to
130 take evasive actions), etc. To determine the time for taking evasive actions, the performance of the ship avoiding
131 collision should be evaluated, otherwise, the alarm might result in the under-estimation of risk (Huang and Gelder,
132 2020). However, most SDs neglect this aspect. Thus, the AMM is introduced to evaluate the performance of the
133 ship avoiding collision in each encounter, with the aim of offering a new tool for the construction of the SD.

134 3. Methodology
135 The key to the construction of this risk perception-based SD is to link the timing of the ship taking evasive
136 maneuvers with the corresponding perceived risk by the navigator. This involves three aspects. First, ship conflict
137 needs to be detected. Second, it needs to be determined at what point a ship takes evasive maneuvers to avoid a
138 ship-ship collision. Third, the navigator’s perceived risk level at that point in time when evasive action is
139 undertaken needs to be estimated. In this context, the COLREGs need to be considered, as these are the
140 background for navigators’ interpretation of ship encounter situations. Combining the three above steps using a
141 set of data mining and analysis algorithms will enable the empirical estimation of this SD.
142 The detection of ship conflict can be easily obtained as various methods have been proposed (Gil et al., 2020; Du
143 et al., 2019). However, it is difficult to determine when a ship takes evasive maneuvers for conflict elimination.
144 The link of the actions and the ship intention is a challenge. During the process of conflict elimination, the ship
145 can take many actions for multiple purposes. The method of ship intention estimation proposed in Du et al., (2020a)
146 helps find a ship’s evasive maneuvers. The second challenge is the extraction of an indicator that can reflect
147 characteristics of the navigator’s perceived risk when the ship responds to the conflict. The concept of AMM
148 reflects the capacity of resolving a conflict in terms of how many maneuvering options leads to a successful
149 conflict elimination compared to how many options there are available, as developed in (Du et al., 2020c; Huang
150 and van Gelder, 2020). Therefore, the concept of AMM is adopted here.

Step I Step II
Step III
Input Trajectory Collision Candidate Output
Action Analysis
Reconstruction Identification

Data cleaning Ship-pair encounter Turning points

AIS data Risk perception-based
and filtering event detection extraction
ship domain

Ship trajectory Ship evasive

reconstruction Conflict identification maneuvers
identification

ship COLREGs identity

Ship AMM
determination
calculation

Statistical analysis

151
152 Figure 2 The method proposed for constructing risk perception-based ship domain
153 The entire process of constructing a risk perception-based SD is presented in Figure 2.
154 Step I: Reconstructing ship trajectories from AIS data. This step extracts the sailing information of each
155 ship from the raw AIS data. After data cleaning and filtering, some errors are deleted. Then, the trajectory
156 of the ship is reconstructed based on a linear interpolation so that it is updated at each same time step.
157 Step II: Identifying collision candidates, i.e., pairs of ships in conflict. This step identifies all the ships that
158 have a conflict with each other from historical AIS data, which contains (1) the ship-pair encounter event
159 detection, (2) conflict identification, and (3) ship COLREGs identity determination.
160 Step III: Ship action analysis is employed to find the moment when the ship takes evasive maneuvers for
161 conflict elimination. This process includes extraction of turning points by the Douglas–Peucker (DP)
162 algorithm, identification of ship evasive maneuvers by ship intention estimation based on non-linear velocity
163 obstacles (NLVO) algorithm, and the calculation of AMM. For each pair of collision candidates, the AMM
164 value when the ship starts to take evasive maneuvers is recorded and statistically analyzed for the
165 construction of a collision risk perception-based SD.
166 3.1 Step I: Trajectory reconstruction
167 AIS data consists of static and dynamic messages from each ship and the database stores the messages in
168 chronological order of receipt. To obtain the trajectory of each ship, the database needs to be re-sorted by MMSI
169 number and then the messages belonging to the same ship are updated by the time series. Hence, the trajectory of
170 the ship is obtained.
171 Many factors can affect AIS data quality. The raw AIS data contains noise and error information, including ship
172 position errors and abnormal speeds (Zhang et al., 2018). Therefore, these incorrect ship trajectories are deleted
173 after data cleaning and filtering (Zhang et al., 2015).
174 The time interval of AIS broadcasts varies, so the AIS data could be interpolated with a predefined time interval.
175 In that case, a linear interpolation method is employed, and the predefined time interval is set to one minute.
176 3.2 Step II: Collision candidate identification
177 3.2.1 Ship-pair encounter event detection
178 It is beneficial to analyze a conflict from the perspective of regarding a ship-pair encounter as a process (Chen et
179 al., 2018). Hence, the concept of ship-pair encounter event (SPEE) is employed. An SPEE signifies the process
180 of an encounter between a ship pair within a specified time period. Each SPEE is given a number of attributes,
181 relevant for conflict assessment.
182 The SPEE can be detected by the following two steps. The first is the determination of a ship pair encountering
183 each other. The targeted ship pair refers to a ship pair whose minimum relative distance between them is less than
184 the distance limit DisLimit . The second step is the determination of the time period tSPEE . This work focuses on the
185 ship’s first evasive maneuver. Many ships may maneuver for collision avoidance quite early to control the
186 situation (Robert et al., 2003; Chauvin and Lardjane, 2008). To avoid missing the real first evasive maneuver,
187 tSPEE is preliminarily determined as a consecutive one-hour period, which is half an hour before and after the
188 moment of closest point of approach (CPA). Then, the distance limit is employed to shorten this time period to
189 limit the computation time.
190 tmin Dis − 30  tSPEE  tmin Dis + 30 & Dis(tSPEE )  DisLimit , (1)

191 where tmin Dis is the moment of CPA. DisLimit is set as 12 nm as it is the normal radar range setting (Juszkiewicz,
192 2016). The sailing information of one ship in one SPEE is data(tSPEE ) = Lon(tSPEE ), Lat (tSPEE ), v(tSPEE ), c(tSPEE ), h(tSPEE ) .
193 Lon (°) and Lat (°) are the longitude and latitude, respectively. v (kn) is the ship speed, c (°) is the ship course,
194 and h (°) is the ship heading.

195 3.2.2 Conflict identification

196 The NLVO algorithm considers the dynamic nature of ship action throughout the encounter process (Huang et al.,
197 2017), and therefore it is employed to improve the accuracy of conflict identification.

1, if VTS (tr )  S NL _ VO (tr )  , tr  tSPEE ,


198 IC (tr ) =  (2)

0, else
199 where IC represents the conflict index, S NL _ VO is the velocity obstacle (VO) zone in the velocity space of the
200 target ship, see the red marked area in Figure 3. The formula derivation process of S NL _ VO is elaborated in Huang
201 et al., 2017.  is the empty set. If the TS’s velocity VTS falls in the VO zone S NL _ VO , IC = 1 and a conflict exists,
202 see V1 in Figure 3. Otherwise, there is no conflict, see V2 in Figure 3. tr is the time when the conflict exists. The
203 simulation time for conflict detection is limited by tSPEE . More details are provided in Du et al., 2020c. Through
204 this formula, the conflict between each SPEE is identified.
 velocity
v s
obstacle zone
S NL _ VO

V2 V1
a

+
205 0 u
206 Figure 3 Collision risk detection based on NL-VO algorithm
207 3.2.3 Ship COLREGs identity determination
208 The ship COLREGs identity is classified in terms of its action obligation for conflict elimination, which includes
209 the give-way ship (GW) and the stand-on ship (SO). The ship COLREGs identity is determined according to their
210 relative bearing (RB) and relative heading (RH) (COLREGs, 1972). In Figure 4, the RB and RH of the target ship
211 (TS) seen from the own ship (OS) are divided into eight sectors, adopted from Tam and Bucknall, 2010 and
212 Goerlandt et al., 2015. The OS is located in the center of Figure 4.

337.5 0
22.5
RH 8 RH 1
RB8 RB1

RH 7 RH 2
RB7 RB 2

270 90

RH 6 RH 3
RB6 RB3

RH 5 RH 4
RB5 RB 4
202.5 157.5
213 180

214 Figure 4 Regions used for the encounter type categorization and ship COLREGs identity determination, adapted
215 from Tam and Bucknall, 2010 and Goerlandt et al., 2015
216 Table1. Ship COLREGs identity of OS in different encounter types, adapted from Tam and Bucknall, 2010 and
217 Goerlandt et al., 2015

Ship
COLREGs RH1 RH2 RH3 RH4 RH5 RH6 RH7 RH8
identity
RB1 OT SF CR/SO CR/SO HO/GW CR/GW CR/GW OT
RB2 OT SF SF SF CR/GW CR/GW CR/GW OT
RB3 OT SF SF SF SF CR/GW CR/GW OT
RB4 OT SF SF SF SF SF CR/GW OT
RB5 OT CR/SO SF SF SF SF SF OT
RB6 OT CR/SO CR/SO SF SF SF SF OT
RB7 OT CR/SO CR/SO CR/SO CR/GW SF SF OT
RB8 OT CR/SO CR/SO HO/GW CR/GW CR/GW SF OT
218 In Table 1, the combinations of these sectors determine the ship COLREGs identity of OS in different types of
219 encounter. Overtaking (OT), Head-On (HO), and Crossing (CR) refer to different encounter types. SF means safe
220 passing. For the OT, ship COLREGs identity is determined by her speed. The OS is SO if the speed of the OS is
221 lower than that of the TS, otherwise, the OS is the GW. For instance, in Table 1, if a TS is located in RB2 with a
222 relative heading of RH6, then the OS is the GW in a crossing encounter scenario.
223 3.3 Step III: Ship action analysis
224 3.3.1 Ship turning points extraction
225 We assume that a ship only alters her course for conflict elimination with the ship speed unchanged, based on
226 statistical analysis (Baldauf et al., 2017). The turning points are where the ship alters her course.
227 The DP algorithm (Douglas and Peucker, 1973) has been widely adopted in the compression of ship trajectory
228 data (Zhao and Shi, 2018), due to its accuracy and efficiency in simplifying the trajectory. Therefore, the DP
229 algorithm with a reasonable compression threshold can be utilized as a simplification method for ship trajectories
230 to identify the ship turning points (Du et al., 2020a). The sailing information of one ship of each SPEE can be
231  
simplified as data(ttp ) = Lon(ttp ), Lat (ttp ), v(ttp ), c(ttp ), h(ttp ) , ttp  tSPEE . ttp is the turning time identified by the DP
232 algorithm. The compression threshold is set as 15 m based on the sensitivity analysis, as described in Du et al.,
233 2020a.
234 3.3.2 First evasive maneuver identification
235 The ship may adopt evasive maneuvers several times for conflict elimination. This work aims to explore the
236 feature of perceived risk that triggers navigators to take evasive maneuvers. Therefore, the focus of this work is
237 the first evasive maneuver.
238 In Du et al., 2020a, the evasive maneuver is that the ship changes her course or/and speed to eliminate a conflict.
239 By utilizing the NLVO algorithm, the evasive maneuver can be identified by checking whether there is a conflict
240 when the ship is at a turning point.

241 VTS (tea )  S NL -VO (tea )  , tea  ttp


, (3)

t1 = min(tea )


242 where tea indicates the time that the ship takes an evasive maneuver. At tea , the ship makes a turn and the conflict
243 exists. t1 is the first time that the ship takes an evasive maneuver.

244 3.3.3 Ship AMM calculation

245 To present the risk perceived by the navigator, the concept of AMM has been adopted as an indicator. The AMM
246 is measured based on the proportion of maneuvers by which the OS can eliminate conflicts, to all its available
247 maneuvers (Du et al., 2020c):

 AMM (t1 ) =
 s (t1 ) , if V (t )  RV ( (t ), t ):V (t )  S (t ) = 
248   a (t1 ) OS 1 OS s 1 ob OS 1 NL _ VO 1
, (4)
t = max(TCPA, 5)
 ob

249 where AMM (t1 ) is the value of AMM when the ship starts to act to eliminate the conflict at t1 . AMM (t1 ) ranges
250 from 0 to 1. A higher AMM (t1 ) means that the ship has more space and time to execute a maneuver, and hence a
251 higher chance of avoiding a dangerous encounter.  s is the value of the adopted rudder angle that can eliminate
252 the existing conflict, see the dotted arc in green color in Figure 3.  a is all the rudder angles available to the OS
253 and −35   a  35 , see the dotted arc in Figure 3. RVOS ( s (t1 ), tob ) is the OS’s reachable velocity after steering
254 with a demanded rudder angle  s within the observation time tob , which is determined by her current velocity
255 VTS (tr ) and her turning ability, see Du et al., 2020a. The turning ability of a ship is modeled based on the Nomoto
256 model (Nomoto et al., 1956). Two key parameters (i.e., K and T) are determined as K = 2K0 VOS / LOS and
257 T = 2K0  LOS / VOS (Hong & Yu, 2000). Here, we have K0 = 1.5 and T0 = 2.5 for the passenger and cargo ship, and
258 K0 = 1.5 and T0 = 6 for the tanker (Hong & Yu, 2000). The Time to Closest Point of Approach (TCPA) is adopted
259 to determine the observation time tob . The minimum value of tob is set as five minutes in order to have sufficient
260 observation time.
261 3.3.4 Statistical analysis
262 For each SPEE at risk of conflict, the calculated AMM value when the ship starts to take an evasive maneuver is
263 recorded. Afterwards, the statistical analysis of AMM is performed for the construction of the risk perception-
264 based SD.
265 Regarding the shape of this SD, it should be analyzed by visualizing the value of AMM in intensity plots. The
266 AMM intensity plots are generated based on the value of AMM and the relative bearing between this ship pair in
267 polar coordinates.
268 For the size of this SD, the curve of its boundary should be interpreted in such a way that the entire area outside
269 this curve has an intensity greater than the value indicated on the level curve. In this work, the value of the
270 boundary of this SD is determined by the analysis of ship behavior characteristics.

271 4. Case study

272 A profile of marine traffic from AIS data is illustrated in Section 4.1, followed by the detection of ship encounters
273 from AIS data in Section 4.2. The analysis of each SPEE is explained in detail by introducing two typical scenarios
274 in Section 4.3, and the risk perception-based SD from the SO and the GW perspectives are presented in Section
275 4.4. Lastly, the comparison of risk perception-based SD from the GW and the SO perspectives is addressed in
276 Section 4.5.
277 4.1 Ship profiles in AIS database
278 In this section, the proposed method for determining the risk perception-based SD is applied to the Northern Baltic
279 Sea area, which is defined as the Baltic Sea with a latitude exceeding 59°N. In our work, we did not start from
280 raw AIS data. The AIS data adopted in this work had already been processed by Zhang et al. (2015 and 2016).
281 The raw AIS data of these studies originally came from HELCOM (2012). After data processing, including
282 cleaning, filtering and interpolation, this AIS data was applied to detect any possible near misses in the Northern
283 Baltic Sea, as discussed in Zhang et al., 2015 and 2016, and Du et al., 2021. The promising results attest that the
284 quality of this AIS data is acceptable. AIS data from the Northern Baltic Sea in July 2011 was used (Figure 5).
285 One-month voyage data from AIS data consisted of 2757 ships, including specific purpose ships, such as tugs,
286 pilot vessels, wing in ground, high-speed craft, and dredgers. These specific purpose ships, including tugs, pilot
287 vessels, wing in ground, high-speed craft, and dredgers were excluded because their working states are not
288 recorded in the AIS data. Their behaviors in working and non-working states are different (Zhou et al., 2019) due
289 to their different responsibilities for taking evasive action, as specified in Rule 18 of COLREGs. Therefore, this
290 work only investigates the following three types of ships: passenger ships, tankers, and cargo ships. As a result,
291 there were 1638 ships in total, of which around 61.8% were cargo ships (1012), 16% passenger ships (262), and
292 22.2% tankers (364). The average length of the passenger ships, tankers, and cargo ships were 95.6 m, 153.2 m,
293 and 123.9 m, respectively.
294 The navigator-perceived risk may vary with the size of the ship as maneuverability diminishes as the vessel size
295 increases (Pérez and Clemente, 2007). Therefore, ships are further divided into three categories. The length of a
296 small-size ship is less than 100 m and the length of a medium-size ship ranges from 100 m to 200 m. The rest are
297 categorized as large-size ships (above 200 m).
298 Additionally, there are two assumptions in this work for the determination of a ship’s COLREGs identity. The
299 first is that all of the ships are considered to be power-driven ships. Therefore, the ship COLREGs identity of each
300 ship can be determined based on their relative position and relative heading. The second assumption is that the
301 visibility in the summertime (July 2011) in the Northern Baltic Sea is good, which is an assumption also made in
302 Kujala et al. (2009) and Asmi et al. (2011). Hence, Rules 11 to 18 in COLREGs are applicable.
303
304 Figure 5 Size of ship in the Northern Baltic Sea from the AIS database
305 4.2 Detection of Ship-pair Encountering Events from AIS data
306 By adopting the methods described in Section 3.2.1, 30344 SPEEs were detected. More than 26% of these
307 encounters (7969 encounters) presented a conflict through the method mentioned in Section 3.2.2. The ship
308 COLREGs identity, i.e., either GW or SO, was also identified using the method in Section 3.2.3. For HO, both
309 ships should turn to starboard for safe passing, according to COLREGs. Hence, both ships were regarded as GW.
310 Table 2 shows the number of GW and SO belonging to different ship types and ship lengths in all SPEEs with
311 conflict. For passenger ships, most of the encountering ship lengths were small. Passenger ships were considered
312 a GW and an SO 2474 times and 2227 times, respectively. For tankers, it was mostly medium size ships that
313 appeared mostly in the studied region. 723 tankers were classified as GW, and 689 tankers as SO. For cargo ships,
314 most of the ship’s lengths were less than 200 m, i.e., small and medium size. The cargo ships were GW and SO
315 1761 times and 1686 times, respectively.
316 Table 2 The counts of GW and SO belonging to different ship types and ship length in all SPEEs with conflict,
317 resulting from the method in Section 3.2.2 and Second 3.2.3

Passengers Tankers Cargo ships

Ship Length
GW SO GW SO GW SO
Small 1201 1145 91 93 493 560
Medium 801 689 533 526 1168 1049
Large 472 396 99 70 100 77

318 4.3 Demonstration of analyzing SPEEs for the construction of risk perception-based SD
319 Two typical encounter scenarios were selected from the AIS data to demonstrate the process of analyzing SPEEs.
320 The results of the encounter process analysis from the GW perspective are discussed in Section 4.3.1 and Section
321 4.3.2. The ship attributes are shown in Table 3. The encountering processes are illustrated in Figures 6 and 7, in
322 which the line in black is the trajectory of the SO and that of the GW is colored. Four figures for each encounter
323 scenario are introduced to demonstrate the process in Figures 6 and 7. Figure 6(a) and 7(a) show the layout of the
324 whole encounter process, and some results of the analysis are highlighted. Specifically, the result of collision risk
325 analysis, turning point identification, evasive action extraction, and the AMM value at the moment when the first
326 evasive action is taken are presented. The blue circle on the GW’s trajectory is the GW’s position when the GW
327 turns, and the black circle is the position of the SO when the GW takes action. The red solid dot shows the position
328 of the two ships when the ships reach the Closest Point of Approach (CPA) (the position where the ship is
329 identified as being in danger by traditional methods). The arrows at the ends indicate the ending points of the ship
330 trajectory. Figures 6(b) and 7(b) focus on the conflict development process, where the ship heading and course
331 are shown. Figures 6(c) and 7(c) present the ship course and course change, and the ship heading and heading
332 change as the conflict develops. Figures 6(d) and 7(d) show the relative distance between the ship pair during the
333 conflict development process. In this work, the relative distance between the ship pair is the Euclidean distance
334 between the ship positions. The black line represents the relative distance between the ship pair and the blue circle
335 on this black line indicates when the GW makes a turn.
336 Table 3. Ship attributes in three typical encounter scenarios

Ship COLREGs Length Width

MMSI Type
attributes identity (m) m
Scenario GW 27335xxxx Tanker 81 14
1 SO 27343xxxx Tanker 125 16
Scenario GW 31158xxxx Passenger ship 290 30
2 SO 26552xxxx Passenger ship 38 8
337 The results of the encounter analysis are collected in Table 4. Table 4 shows the time when the conflict exists,
338 when the ship makes a turn, when the ship pair reaches the CPA, and when the ship takes an evasive maneuver.
339 The red marked area means the moment that the conflict exists, and the blue marked area means when the situation
340 became safe. (trs , tre ) is the starting and ending moment of the conflict. (tTP , C ) is the time the ship was turning
341 and the amount of course change. C  0 means the ship turns to starboard, and C  0 refers to a portside turn.
342 tCPA is the time when this ship pair arrives at CPA. tea is the time when the ship takes an evasive maneuver. The
343 unit of time is a minute, and the unit of course change is a degree.
344 Table 4. Result of the encounter process analysis of three encounter events

SPEEs (trs , tre ) (tTP , C ) tCPA tea

Scenario (22, 15, 21, 27, 41,
30 27
1 30) 25.6 -11.9 -15.1 -21.5
Scenario (1, 6, 8, 12, 14, 20, 25, 26, 34, 39, 6, 8, 12,
18
2 21) -14.9 -4.5 -13.8 -7.5 -3.1 13.4 1.8 10.7 -2.7 14, 20
345 Note: The red marked area means that a conflict exists, and the blue marked area means it is safe. A negative
346 course change refers to a portside turn and a positive number refers to a starboard side turn. The unit of time is a
347 minute and that of the course change is a degree.
348 4.3.1 Scenario 1
349 Figure 6 presents the result of Scenario 1 from a GW perspective. The encounter duration tSPEE was 61 min. The
350 conflict was detected from 22 min to 30 min.
351 The GW took actions four times during the whole encounter process (Figure 6(a)). First of all, the GW ship turned
352 to starboard around 25.6° at 15 min (Figure 6(c), Table 4). This turn did not generate a conflict and the GW would
353 safely pass the SO ship's stern if this sailing state remained. The second turn happened at 21 min, when the ship
354 turned to port around 11.8° (Figure 6(c), Table 4). The second turn re-caused the conflict from 22 min. To
355 eliminate the conflict, the GW turned 15.1° to port at 27 min. The ship pair continued to approach each other and
356 the relative distance between them dropped to 0.315 nm at the CPA (Figure 6(d)). Afterwards, the ship pair
357 gradually moved apart from each other and the conflict was over at 30 min. The fourth turn at 41 min was to
358 further extend the relative distance between them for safety.
359 The third turn that occurred at 27 min was determined as an evasive maneuver (Figure 6(b)), when the AMM of
360 GW was 88.6%. Thus, for this SPEE, we infer that the ship could bear the risk until the AMM fell to 88.6%. More
361 details can be seen in Table 4.
362
363 (a)

364
365 (b)

366
367 (c)
368
369 (d)
370 Figure 6 Illustration of the AMM calculation for the SPEE from the GW perspective in Scenario 1
371 4.3.2 Scenario 2
372 Figure 7 shows the results of Scenario 2. The duration of this crossing encounter was 45 min. From the GW
373 perspective, this ship pair experienced a conflict before 21 min.
374 In this case, eight turns from the GW ship were detected (Figure 7(a)). The GW first turned around 14.9° to port
375 at 6 min, with the aim of avoiding the conflict with the SO. This action was not sufficient, so the conflict still
376 existed. To eliminate the conflict, the GW turned left several times (Figure 7(c), Table 4), such as the turn at 8
377 min (-4.5°), 12 min (-13.8°), 14 min (-7.5°) and 20 min (-3.1°). At 18 min, their relative distance fell to the lowest,
378 around 0.11 nm (Figure 7(d)). After 18 min, their relative distance increased and there was no conflict after 21
379 min. The following turns of the GW at 25 min, 26 min, 34 min, and 39 min aimed to increase their relative distance
380 and return to its original track.
381 The five actions of the GW at 6 min, 8 min, 12 min, 14 min, and 20 min were evasive maneuvers (Figure 7(b)).
382 Since the first evasive maneuver was taken at 6 min when the AMM of the GW was 37.1% (Figure 7(a)), we infer
383 that the ship could bear the risk until the AMM fell to 37.1%. More details can be seen in Table 4.

384
385 (a)
386
387 (b)

388
389 (c)

390
391 (d)
392 Figure 7 Illustration of the AMM calculation for the SPEE from the GW perspective in Scenario 3
393 4.4 Risk perception-based SD from different perspectives
394 The boundary of the risk perception-based SD is determined based on the statistical analysis of the AMM of each
395 ship when a ship starts to take evasive maneuver. The impact of a ship COLREG identity on ship behavior is also
396 considered in this work.
397 4.4.1 Give-way ship’s risk perception-based SD
398 Figure 8 visualizes the AMM value of the GW the moment at which the GW starts to take an evasive maneuver,
399 in two different modes. Figure 8(a) presents a histogram and Figure 8(b) shows a polar diagram.
400 Figure 8(a) presents the result in a histogram, where the x-axis shows the AMM value and the y-axis indicates the
401 ratio of the GW taking actions at different AMM levels. From Figure 8(a), we can observe that most GWs take
402 actions when the AMM is still at a high level, which implies that the GW would take actions early, once dangers
403 were detected.
404 Furthermore, the timing for taking actions differs for ships of different lengths. Approximately 95% of GW
405 passenger ships, which are small in size, start to take evasive maneuvers before the AMM drops to 0.92. If we set
406 AMM = 0.95 as a threshold, 93.6% of small-size passenger ships will take action, while this rate drops to 87.1%
407 and 77.3% for medium- and large-size passenger ships. Likewise, we can observe a similar pattern for the tankers
408 and cargo ships. From the data, we can conclude that smaller size ships take actions earlier than larger size ships.
409 Figure 8(b) uses a polar diagram to explain the data. Each gray point represents one record. The length of the
410 point indicates the AMM level when the GW starts to take evasive maneuvers, the angle of which indicates the
411 RB of this ship pair at that moment from the OS perspective. The OS is located in the origin and points to the 0°.
412 By analyzing each dangerous encounter, we constructed scatter plots.
413 Furthermore, we simplified the shape of the risk perception-based SD as a circular shape by visually analyzing
414 the value of the GW’s AMM in intensity plots in Figure 8(b). We added a circle at the origin in the diagram, which
415 cuts the diagram into two parts and the radius of the circle is the AMM level. The points inside the circle mean
416 the ships take actions after the AMM drops to the pre-set AMM level.
417 Let us set 90% as a threshold to determine the boundary of risk perception-based SD, inspired by the work done
418 by Hörteborn et al., 2019. Then the circle that excludes 90% of the ships can be found, which is the circular risk
419 perception-based SD, see the red ring in the figure. 90% of the points beyond the red ring and the point inside the
420 ring imply some unusual cases where action was taken later than in the other 90% of the cases. The AMM values
421 of the boundary of the risk perception-based SD for different ships are listed in Table 5.
422 - When a passenger ship is the GW and of small, medium, or large size, the boundary of the risk
423 perception-based SD is AMM=0.986, 0.914, and 0.814, respectively.
424 - When a tanker is the GW and of small, medium, or large size, the boundary of the risk perception-
425 based SD is AMM=0.843, 0.829, and 0.8, respectively.
426 - When a cargo ship is the GW and of small, medium, or large size, the boundary of risk perception-
427 based SD is AMM=0.9, 0.886, and 0.871, respectively.
428 - These results demonstrate that not only the ship length but also the type of ship affects the navigator-
429 perceived risk. The boundary of the risk perception-based SD of passenger ships is larger than that of the
430 other two ship types. One possible cause is that passenger ships have stricter requirements for safe operations.
431 Therefore, passenger ships usually take evasive maneuvers early to eliminate any potential risk. As a GW,
432 the boundary of risk perception-based SD of a tanker is generally smaller than that of a cargo ship. We
433 discuss possible causes of this phenomenon in Section 5.1.
434
435 (a) Statistical histogram of AMM of GW

436
437 (b) The polar coordinate diagram of GW’s AMM value and its risk perception-based SD
438 Figure 8 Visualization of the AMM value of GW when the GW starts to take evasive maneuver
439 4.4.2 Stand-on ship’s risk perception-based SD
440 Figure 9 presents the result of the SO’s AMM when the SO starts to act for conflict elimination. Figure 9(a) is a
441 statistical histogram and Figure 9(b) is a polar coordinate diagram.
442 Similarly to Figure 8(a), the x-axis in Figure 9(a) is the AMM value when the SO starts to take evasive maneuvers
443 and its y-axis indicates the ratio of the SOs that take actions at different AMM levels. In Figure 9(a), most of the
444 SOs also prefer to take collision avoidance actions early, when their AMM is still at a high level, so as to master
445 the situation, which is consistent with the findings in Chauvin and Lardjane, 2008. This can be supported by the
446 fact that approximately 95.8% of the small-size passenger ships start to take evasive maneuvers before the AMM
447 decreases to 0.8. Before the AMM drops to 0.8, almost 90% of the small-size tanker ships and 88% of the small-
448 size cargo ships have started to take actions to avoid conflict.
449 Furthermore, the timing for ships of different lengths to take evasive maneuvers is also different. Specifically, the
450 AMM when the SO starts to take evasive maneuvers decreases as the ship size increases. About 95% of SO
451 passenger ships of small, medium, and large size begin to take evasive maneuvers before AMM drops to 0.829,
452 0.5 and 0.486 respectively.
453 AMM = 0.95 is set as the threshold, below which the rate of a small-size oil tanker that is the SO starts to take
454 evasive maneuvers is 77.4%, while that decreases to 72.2% for medium size and 38.5% for large size. This similar
455 trend holds true for cargo ships. For a cargo ship that is the SO, the percentage of ships starting to take evasive
456 maneuver is 78.8% for small-size, 72.7% for medium-size, and 71.8% for the large-size group.
457 Figure 9(b) shows a polar coordinate diagram that presents the SO AMM value when the evasive maneuver starts.
458 The angular coordinate and the radius of the polar coordinate diagram have the same meaning as in Figure 8(b).
459 We also simplified the shape of the risk perception-based SD to make it circular. Its boundary is the critical AMM
460 value before which 90% of the ships will act, which is indicated by the red circle in each polar coordinate diagram
461 in Figure 9(b). The size of the risk perception-based SD is affected by the ship size. The AMM values of the
462 boundary of the risk perception-based SD for different ships are listed in Table 5.
463 - When a passenger ship is the SO and of small, medium or large size, the boundary of risk perception-
464 based SD is AMM=0.943, 0.786, and 0.729, respectively.
465 - When a tanker is the SO and of small, medium or large size, the boundary of risk perception-based
466 SD is AMM=0.857, 0.629, and 0.486, respectively.
467 - When a cargo ship is the SO in small, medium or large size, the boundary of risk perception-based
468 SD is AMM=0.729, 0.5, and 0.486, respectively.
469 Further, we observed that the size of the risk perception-based SD of the SO was different for different types of
470 ship. Generally, when being as the SO, the size of the risk perception-based SD of a passenger ship is larger than
471 that of a tanker, which is larger than that of a cargo ship. We discuss the possible causes in Section 5.1.
472
473 (a) Statistical histogram of AMM of SO

474
475 (b) The polar coordinate diagram of SO’s AMM value and its risk perception-based SD
476 Figure 9 Visualization of the AMM value of SO when the SO starts to take evasive maneuver
477 4.5 Comparison between the risk perception-based SD of GW and OS
478 The size of risk perception-based SD of a ship whose identity is GW is generally larger than that of a ship whose
479 identity is SO (see Table 5). When being as the SO, ships are more inclined to let the GW respond, and therefore
480 their time to act is relatively late. For instance, for a large-size passenger ship, the AMM value at the boundary of
481 risk perception-based SD is 0.814 for being as a GW and 0.729 for being as an SO. For a large-size tanker, whose
482 identity is SO, the AMM value at the boundary of risk perception-based SD is 0.486, while the value is 0.8 for a
483 tanker whose identity is GW. For a large-size cargo ship that is the SO, the AMM value at the boundary of the
484 risk perception-based SD is 0.486, while the value is 0.871 when it is the GW. More details can be seen in Table
485 5.
486 Table 5. The AMM value at the boundary of risk perception-based SD for different ships
Ship type Small Medium Large
GW 0.986 0.914 0.814
Passenger Ship
SO 0.943 0.786 0.729
GW 0.871 0.829 0.8
Tanker
SO 0.857 0.629 0.486
GW 0.9 0.886 0.871
Cargo Ship
SO 0.729 0.5 0.486

487 5. Discussion
488 5.1 The feature of risk perception-based SD
489 Based on the present analysis, we found that the proposed risk perception-based SD has the following features:
490 First, the ship’s COLREGs identity affects the size of the risk perception-based SD, see Figures 8 and 9. The size
491 of the risk perception-based SD of a GW ship is larger than that of a SO ship (Table 5). The rules as specified in
492 COLREGs provide a possible explanation. The GW and SO have different action responsibilities during different
493 encounter stages, see Rules 16, 17, and 18 in COLREGs, 1972. When a conflict exists, the SO is not allowed to
494 act to avoid conflict at the onset of the encounter. The SO can, however, take an evasive maneuver when the
495 conflict becomes serious due to the GW’s improper strategy of conflict elimination. Therefore, the SO seems to
496 be more likely to let the GW respond first, waiting to initiate collision avoidance actions until the navigator of the
497 SO vessel believes the risk levels become too high and action is required. Nevertheless, a common feature of the
498 risk perception-based SD of both GW and SO vessels is that navigators prefer to act early. This is in accordance
499 with the provisions of the COLREGs and earlier research findings (Robert et al., 2003; Chauvin and Lardjane,
500 2008). This widely adopted strategy of conflict elimination (Olsson and Jansson, 2006) aims at mastering the
501 interaction situations, leading to fewer very close near misses (Belcher, 2003).
502 Second, the size of the risk perception-based SD decreases with increasing ship sizes, see Figures 8 and 9. In the
503 present analysis, ships are divided into three groups based on their length. From the results shown in Section 4.4,
504 smaller ships take actions earlier than larger ships. For instance, for a passenger ship being as an SO, the AMM
505 at the boundary of its risk perception-based SD is 0.943 for small-size vessels, which drops to 0.786 for medium-
506 size and 0.729 for large-size, see Table 5. This shows that navigators of larger ships prefer to act at higher levels
507 of perceived risk. A similar trend can be observed for tankers and cargo ships. To interpret the results, we need to
508 clarify that the AMM is an indicator for judging whether the ship takes evasive action early or late. A ship acting
509 early basically implies that the ship starts evasive action with a higher AMM. This does not mean that the action
510 time is earlier. Compared with smaller vessel categories, the course changes of larger ships require more effort
511 and time due to their relatively limited maneuverability. Let us take two ships as an example, one a small-size
512 ship with better maneuverability and the other is a large-size ship with a relative lower maneuverability. Even
513 though the starting action time of the smaller ship is slightly later than that of the larger ship in the same encounter
514 scenario, if the AMM of the smaller ship is bigger than that of the larger ship, the smaller ship is regarded as
515 acting earlier than the larger ship. Zhou et al. (2019) show that the behavioral characteristics of navigators of ships
516 of different sizes are different. This may be a plausible explanation for the observation that larger vessels respond
517 with a relative lower AMM to avoid a collision than smaller ships.
518 Third, the size of the risk perception-based SD varies by ship type, see Figures 8 and 9. The size of the risk
519 perception-based SD of passenger ships is generally larger than that of the other two ship types. The safety of life
520 has been the top priority for the passenger ship industry for decades, due to the huge threat of accidents and loss
521 of human lives (Iqbal et al., 2008; Lu & Yang, 2011). This may be a reason why navigators on passenger ships
522 have a low tolerance for conflict and take collision avoidance actions at lower levels of risk perception. Therefore,
523 this may explain why the timing for a passenger ship to take an evasive maneuver is earlier. Further, when being
524 as an SO, the risk perception-based SD’s size for a tanker is larger than that for a cargo ship (Table 5). This
525 indicates that the risk tolerance of a tanker that is the SO is lower than that of a cargo ship, as the tanker takes
526 evasive maneuvers earlier than a cargo ship. We could explain this by taking into account ship maneuverability.
527 The load conditions significantly influence a ship’s maneuverability. The maneuverability of a fully loaded tanker
528 is relatively poorer than that of a cargo ship, which directly reduces the AMM value of the tanker when an evasive
529 maneuver starts.
530 5.2 Advantages of using AMM in the construction of risk perception-based SD
531 There are many concepts that can represent temporal closeness between a pair of vessels. TCPA is one of the most
532 commonly used parameters because both spatial proximity and speed are integrated (Szlapczynski and
533 Szlapczynska, 2017). TCPA represents the remaining time for two ships to reach their closest points if the course
534 and speed remain the same.
535 The proposed risk perception-based SD based on the concept of AMM is more realistic than TCPA as the
536 formulation of conflict based on the proposed risk perception-based SD is linked to conflict resolution. The risk
537 solution presents the difficulty of performing evasive maneuvers to successfully avoid collisions. Determining the
538 risk levels of conflict in ship-ship encounter situations independent of their potential for conflict resolution may
539 lead to inaccurate detection of actual danger (Chen et al., 2018), considering that ship maneuverability helps to
540 measure the level of risk more accurately (Baldauf et al., 2015; Huang and van Gelder, 2020). However, most of
541 the existing research in terms of conflict based on TCPA ignore the impact of ship maneuverability.
542 Let us take the encounter scenarios in Figure 10 as an illustration. The only difference between them is the loaded
543 condition of the GW. The half-loaded GW and fully loaded GW are marked in blue and red, respectively. The SD
544 of the GW will be violated, so the conflict exists in both Figure 10(a) and Figure 10(b). Ship maneuverability
545 affects this resolution solution (Hong and Yang, 2012). The load conditions significantly affect the ship
546 maneuverability. Although the TCPA is the same for these two ship-pair encounter scenarios, their collision risk
547 is different because the GW has different maneuverability. Therefore, the conflict analysis based on risk
548 perception-based SD is more precise.

SO SO

AMM 1 AMM 2
−35 35 −35 35

GW(Half loaded) GW(Full loaded)

(a) (b)
549
550 Figure 10 Ship conflict analysis based on risk perception-based SD
551 5.3 Application of risk perception-based Domain
552 The risk perception-based SD provides statistical information about the timing of a ship starting an evasive
553 maneuver in terms of a perceived collision risk level. It shows how the navigators choose the timing to avoid the
554 conflict in a specific area. The proposed risk perception-based SD based on the concept of AMM can be applied
555 for the following purposes, although more work is needed to validate this application before practical
556 implementation.
557 First, it can be used for providing information about waterway safety based on historic AIS data, such as near miss
558 detection. Various methods have been proposed to analyze near misses from historical AIS data. One typical
559 method measures the collision risk by obtaining insights into ship behavior characteristics during the process of
560 collision avoidance. Any abnormal ship behavior during the process of collision avoidance normally leads to a
561 serious encounter. Frenetic rudder actions may occur in the last moment before ship collision to prevent it
562 happening (Mestl et al., 2016). The use of statistical techniques for detecting abnormal ship behavior has attracted
563 increasing attention in maritime transportation research (Pallotta et al., 2013). The statistical characteristics of the
564 risk perception-based SD in a given sea area can be obtained from a given AIS database. Then, abnormal behavior
565 can be observed by comparing the AMM levels of encounters found in a new data time series to the normal risk
566 perception-based SD determined earlier. When two ships approach each other and the give-way ship violates the
567 risk perception-based SD, this means that the GW acts beyond normal operational conditions and probably
568 approaches the boundaries of acceptable safety levels, since most ships (around 90% of the ships navigating in
569 the area) would take evasive maneuvers before this moment. Combining the information derived from vessel
570 encounters detected in AIS data in terms of risk perception-based SD violations, evasive maneuvers with other
571 information about navigational safety can provide a comprehensive picture of the navigational safety levels in a
572 given sea area. This can be done, for instance, by combining the risk perception-based SD with the ship’s
573 obligation for collision avoidance at different stages of the encounter as specified in the COLREGs, e.g., through
574 a delineation of four safety levels, as specified in Du et al. (2020c, 2021).
575 Second, the risk perception-based SD could be further developed as a basis for a collision alert system onboard
576 ships or in remote control centers. Real-time alerts for ship-ship encounter situations could significantly contribute
577 to the reduction of collision accidents (Lehikoinen et al., 2015), and considerable work has been dedicated to
578 proposing ship collision alert systems, see Gil et al. (2020). The proposed risk perception-based SD can be used
579 as a basis for defining intelligent collision alerts, incorporating conflict resolution and comparing the
580 characteristics of an ongoing encounter with historic patterns of normal operation in the area. When appropriate
581 AMM threshold levels are used for raising alarms, such an intelligent collision alert system can reduce the number
582 of unnecessary alarms, which is known to be a problem with existing collision alert systems (Baldauf et al. 2011).
583 Third, the risk perception-based SD could be used as a benchmark for testing the CAS in maritime autonomous
584 surface ships (MASS), or for assessing the safety levels of introducing MASS in mixed traffic environments. The
585 risk perception-based SD reflects the statistical features of the historically observed traffic flow. If the CAS in
586 MASS can handle collision avoidance equally as well as human navigators, traffic flow with a mixed composition
587 of MASS and conventional vessels would have similar statistical features to historically observed traffic flows.
588 On the other hand, if the statistical features of the risk perception-based SD of the mixed traffic (MASS and
589 conventional traffic) are different from the historic traffic characteristics, this would imply that the CAS in MASS
590 might not perform as well as human navigators, and that mixed traffic would not be as safe as conventional vessel
591 traffic. Thus, the risk perception-based SD could be used as a basis for simulation models to analyze safety levels
592 under various traffic conditions, or as a benchmark when implementing MASS in real-world environments.
593 5.4 Limitations and future improvements
594 This work has introduced a risk perception-based SD and proposed a methodology for determining its shape and
595 size. This was done by linking the timing of conflict elimination with the AMM as a proxy for collision risk levels
596 perceived by a navigator. The maneuverability of the ship has been taken into account, measuring the capability
597 of a vessel to eliminate the conflict. Although the results of the case studies and empirical findings are promising,
598 several factors could further improve the proposed method and strengthen the findings and future applicability.
599 5.4.1 Ship maneuverability
600 The accuracy of modeling ship maneuverability also affects the accuracy of calculating a ship’s AMM. The
601 Nomoto model was employed in this work to measure ship maneuverability as it only requires limited input
602 parameters. Although the Nomoto model is widely used as it is effective and comparatively simple, it may not be
603 appropriate in some situations, e.g., for vessels with non-conventional steering arrangements. Ship
604 maneuverability is improving significantly due to the fast development of the ship industry.
605 To measure the reliability of AMM computation based on the Nomoto model, a sensitivity analysis was conducted.
606 100 SPEEs were randomly selected to be the database for this sensitivity analysis. Figure 11 illustrates the impact
607 of ship maneuverability on the value of AMM. There are two key parameters (i.e., K0 and T0 ) in the Nomoto
608 model. For passenger and cargo ships, K0 = 1.5 and T0 = 2.5 , and for tankers K0 = 1.5 and T0 = 6 (Hong & Yu,
609 2000). We varied K0 and T0 and checked how much the AMM varied relative to the baseline K0 and T0 values.
610 Figure 11(a1-a8) shows the change in AMM when K is the only variable. Figure 11(b1-b8) shows the change in
611 AMM when T is the only variable. The change rate of AMM ( AMM ) is the magnitude of change in AMM
612 divided by the reference value of AMM. The AMM with ( K0 , T0 ) input was set as the reference. For instance,
613 if the K0 decreases by 0.1 and T remains the same, AMM = ( AMM ( K0 − 0.1,T0 ) − AMM ( K0 ,T0 )) / AMM ( K0 ,T0 ) . In
614 Figure 11, the x-axis is AMM and the y-axis is the probability of AMM ( PAMM ).

615 The sensitivity analysis revealed the following three findings. First, AMM increases with an increase of K, but
616 decreases with an increase of T. This is consistent with the fact that the combination of a large K and small T is a
617 characteristic of excellent steering performance (Nomoto et al., 1956). A larger K means a large turning moment
618 and a smaller T means a quick rudder response. The distribution of the ratio of AMM is positively skewed
619 when K increases or T decreases, however, it is negatively skewed when K decreases or T increases. M AMM is
620 the mean value of AMM . For example, when K increases by 0.1 and 0.4, M AMM increases by 0.012 and 0.035
621 respectively, see Figure 11. When T increases by 0.1 and 0.4, M AMM decreases by 0.003 and 0.011 respectively.
622 Second, the impact of K on the AMM is higher than that of T. For instance, when K increases by 0.1, the increment
623 of AMM is 0.012. When T increases by 0.1, the decrement of AMM is 0.003. Third, when the ship maneuvers
624 with a higher AMM, the change in K and T has a minor impact on AMM. When the ship maneuvers late so that
625 the AMM is low, the impact of K and T on AMM is more serious.
626 In brief, as many ships prefer to act early according to the statistical analysis in Section 4.4, the calculation of
627 AMM is less affected by K and T. The M AMM is less than 6%, and therefore the calculation of AMM based on
628 the Nomoto model is acceptable. Nonetheless, better information about the maneuverability characteristics of
629 vessels could improve the definition of the risk perception-based SD. In future work, the MMG model (Tao et al.,
630 2019) or Abkowitz’s model (Zhang and Zou, 2011), which are more accurate than the Nomoto model, could be
631 used to define a more accurate risk perception-based SD.

632
633 Figure 11. The impact of ship maneuverability on the value of AMM
634 5.4.2 Ship intention estimation
635 The intention estimation can be further improved, as it is a fundamental step for identifying a ship’s evasive
636 maneuvers. In this work, the intention is estimated from ship movement based on historical AIS data. However,
637 ship movement is not only determined by the intention of the navigators but can also be influenced by several
638 internal and external factors, such as wind and currents in ports and inland waterways (Zhou et al., 2020). Further
639 addressing these influencing factors may be important for more accurate intention estimation. Additionally, when
640 a ship encounters multiple targets, it is difficult to judge the intention of each action using AIS data alone. In this
641 study, multi-vessel encounters were divided into multiple ship-pair encounters, and conflicts between multiple
642 ships were not explicitly considered. This issue may lead to an inaccurate understanding of the action intention of
643 the target ship (Du et al., 2020b).
644 5.4.3 Violation of COLREGs
645 Whether or not the ship action violates the COLREGs is not considered in this work. The ship domain proposed
646 in this work is a data-driven model and the parameters were determined by historical data. The AIS data records
647 what really happens during the collision avoidance process. From this study and other studies (Chauvin and
648 Lardjane, 2008), the violation of COLREGs by the navigator does happen. For instance, the give-way ship may
649 turn to portside for easy operation. Some safe navigators on board the SO prefer to act early to control the situation,
650 whereas some risky navigators may choose to act late (Huang et al., 2020). However, this work only considered
651 the conditions under which the ship took the first evasive action. We just reflect what really happened during the
652 encounter process. Whether the behavior of a navigator complies with the COLREGs or not requires future work.
653 5.4.4 Size and shape of the proposed domain
654 The determination of the size and shape of the risk perception-based SD was simplified and can be further
655 investigated. In this work, we simplified the shape of the risk perception-based SD to be a circular shape and its
656 size was determined by choosing 90% as a threshold value for AMM. The circular shape of the SD is acceptable
657 as a first analysis using this ship domain concept, but it is acknowledged that other researchers have proposed
658 various irregular-shaped SDs such as ellipses or polygons (Pietrzykowski and Uriasz, 2009; Wang et al., 2010).
659 5.4.5 Other limitations
660 This work aims to make a conceptual proposal, rather than firm statements about the exact shape and size of the
661 AMM-based ship domain. The algorithm designed to compute the AMM in this work is complex and in the current
662 implementation takes a very long time to run. Therefore, only one-month AIS data is utilized to demonstrate the
663 results. To propose the new concept, we believe one month of data is sufficient, but indeed for estimating the
664 limits of the AMM-based domain accurately for a range of vessels, we acknowledge that one-month data is limited .
665 More analyses with more data are needed in the future work, but as a prerequisite this work the current code of
666 the algorithm needs to be optimized, or a meta-model needs to be developed to achieve a faster computation speed.
667 This can help to increase the accuracy and reliability of the values obtained. As the current focus is to propose the
668 concept of an AMM-based ship domain rather than defining its exact size and shape for a range of vessels, we
669 leave this further development and more extensive data analysis for future work.

670 6. Conclusions
671 Existing ship domains are typically constructed based on the idea that navigators intend to keep an area around
672 their vessel clear from other ships. With the realization that such proximity-based ship domains to assess
673 navigational safety are limited, this article proposes a new ship domain concept and describes the empirical
674 investigation of it. This risk perception-based ship domain (SD) was created on the timing of ships taking evasive
675 maneuvers and the associated perceived risk levels. The boundary of this risk perception-based SD was
676 empirically determined based on a large dataset of ship encounter situations detected in AIS data. Ship turning
677 points were extracted through the Douglas–Peucker (DP) algorithm, after which the ship’s evasive maneuver was
678 estimated using the Non-Linear Velocity Obstacles (NLVO) algorithm. The concept of Available Maneuvering
679 Margin (AMM) was employed to reflect the vessel’s capability of conflict resolution, which is used as a proxy
680 for the navigator’s perceived collision risk at the timing of evasive maneuvering. This work helps explain and
681 predict the behavior of ships in encounter situations with a potential ship-ship collision. The timing of the evasive
682 maneuvers and the associated perceived risk levels were investigated, providing novel insights into maritime
683 transport safety.
684 The results of the case study demonstrate that the proposed risk perception-based SD can give an understanding
685 of the ship-ship collision avoidance process, indicating that the concept can be further developed in various
686 applications to improve navigational safety. First, most navigators prefer to take evasive maneuvers early once
687 collision dangers are detected. Second, the size of the risk perception-based SD is affected by many factors,
688 indicating that navigators interpret and perceive collision risks differently in different encounter situations.
689 Important factors in this interpretation include the ship length, ship type, and the ship COLREGs identity. The
690 findings show that the timing of collision avoidance actions is delayed as ship lengths increases. The size of the
691 risk perception-based SD of passenger ships is generally larger than that of tankers and cargo vessels. In terms of
692 the ship COLREGs identity, it was found that the risk tolerance of a navigator of a tanker in ‘stand-on’ status is
693 lower than those of navigators operating a cargo vessel with the same ship COLREGs identity. Third, as the
694 dynamic nature of ship action and ship capability for conflict resolution are explicitly considered in this ship
695 domain concept, we argue that this risk perception-based SD based on the concept of AMM has some advantages
696 over Time to Closest Point of Approach (TCPA).
697 Based on our findings, we conclude that this SD concept can be further applied in various future developments,
698 for instance, to provide information about maritime transportation safety based on historic AIS data (i.e., near
699 miss detection), to develop intelligent collision alert systems, and to analyze the safety performance of maritime
700 autonomous surface ships in future mixed traffic environments. However, more work is needed to validate this
701 application before it is implemented. In light of the promising results of this SD concept, the proposed
702 measurement procedure, and the obtained empirical findings, we would like to highlight several avenues for
703 further development. These include the improvement of evasive maneuver estimation, a more advanced method
704 to more accurately reflect vessels’ maneuverability characteristics, and further integration and development of the
705 proposed risk perception-based SD for waterway safety analysis and collision alert system development.

706 Acknowledgements
707 The first author of this work is supported by the China Scholarship Council (Grant Number: 201606950009) and
708 Marine Technology research group in Aalto University (9170094). This work is also supported by the National
709 Science Foundation of China (NSFC) through Grant No. 52001237. The work has further received financial
710 support from the Canada First Research Excellence Fund, through the Ocean Frontier Institute. This financial
711 support is gratefully acknowledged. Besides, we thank the two anonymous reviewers for their very insightful
712 comments, which have been very instrumental to improve an earlier version of this work.

713 References
714 Asmi, E., Kivekäs, N., Kerminen, V. M., Komppula, M., Hyvärinen, A. P., Hatakka, J., ... & Lihavainen, H.
715 (2011). Secondary new particle formation in Northern Finland Pallas site between the years 2000 and 2010.
716 Atmospheric Chemistry & Physics Discussions, 11(9).
717 Aydogdu, Y. V., Yurtoren, C., Kum, S., Park, J. S., & Park, Y. S., 2010. Questionnaire survey on the risk
718 perception in the Istanbul Strait. Journal of Navigation and Port Research, 34(7), 517-523.
719 Aydogdu, Y. V., 2014. A comparison of maritime risk perception and accident statistics in the Istanbul Straight.
720 The Journal of Navigation, 67(1), 129.
721 Baldauf, M., Benedict, K., Fischer, S., Motz, F., & Schröder-Hinrichs, J. U., 2011. Collision avoidance systems
722 in air and maritime traffic. Proceedings of the Institution of Mechanical Engineers, Part O: Journal of Risk
723 and Reliability, 225(3), 333-343.
724 Baldauf, M., Mehdi, R., Deeb, H., Schröder-Hinrichs, J. U., Benedict, K., Krüger, C., ... & Gluch, M., 2015.
725 Manoeuvring areas to adapt ACAS for the maritime domain. Zeszyty Naukowe Akademii Morskiej w
726 Szczecinie.
727 Baldauf, M., Mehdi, R., Fischer, S., Gluch, M., 2017. A Perfect Warning to avoid collisions at sea? Sci. J.
728 Maritime Univ. Szczecin 49 (121), 53–64.
729 Belcher, P., 2003. A day in the life of the Dover Strait. Safety at Sea International, 57(408), 15–16.
730 Chauvin, C., & Lardjane, S., 2008. Decision making and strategies in an interaction situation: Collision avoidance
731 at sea. Transportation Research Part F: Traffic Psychology and Behaviour, 11(4), 259-269.
732 Chen, P., Huang, Y., Mou, J., & van Gelder, P. H. A. J. M., 2018. Ship collision candidate detection method: A
733 velocity obstacle approach. Ocean Engineering, 170, 186-198.
734 Chin, H. C., & Debnath, A. K., 2009. Modeling perceived collision risk in port water navigation. Safety Science,
735 47(10), 1410-1416.
736 Coldwell, T. G., 1983. Marine traffic behaviour in restricted waters. The Journal of Navigation, 36(3), 430-444.
737 Conventions on the International Regulations for Preventing Collision at Sea (COLREGs). 1972. The
738 International Maritime Organization (IMO).
739 Douglas, D.H., Peucker, T.K., 1973. Algorithms for the reduction of the number of points required to represent a
740 digitized line or its caricature. Cartographica: The International Journal for Geographic Information and
741 Geovisualization 10 (2), 112–122.
742 Du, L., Valdez Banda, O. A, & Kujala, P., 2019. An intelligent method for real-time ship collision risk assessment
743 and visualization. In Developments in the Collision and Grounding of Ships and Offshore Structures:
744 Proceedings of the 8th International Conference on Collision and Grounding of Ships and Offshore
745 Structures (ICCGS 2019), 21-23 October, 2019, Lisbon, Portugal (p. 293). CRC Press.
746 Du, L., Goerlandt, F., Valdez Banda, O. A., Huang, Y., Wen Y., & Kujala, P., 2020a. Improving stand-on ship’s
747 situational awareness by estimating the intention of the give-way ship. Ocean Engineering, 201, 107110.
748 Du, L., Goerlandt, F., & Kujala, P., 2020b. Review and analysis of methods for assessing maritime waterway risk
749 based on non-accident critical events detected from AIS data. Reliability Engineering & System Safety,
750 106933.
751 Du, L., Valdez Banda, O. A., Goerlandt, F., Huang, Y., & Kujala, P., 2020c. A COLREG-compliant ship collision
752 alert system for stand-on vessels (In press).
753 Du, L., Valdez Banda, O. A., Goerlandt, F., Kujala, P., & Zhang, W., 2021. Improving Near Miss Detection in
754 Maritime Traffic in the Northern Baltic Sea from AIS Data. Journal of Marine Science and Engineering,
755 9(2), 180.
756 EMSA, 2020. Annual overview of marine casualties and incidents 2020. European Maritime Safety Agency.
757 Retrieved from: http://www.emsa.europa.eu/we-do/safety/accident-investigation/item/4266-annual-
758 overview-of-marine-casualties-and-incidents-2020.html
759 Fan, C., Wróbel, K., Montewka, J., Gil, M., Wan, C., & Zhang, D., 2020. A framework to identify factors
760 influencing navigational risk for Maritime Autonomous Surface Ships. Ocean Engineering, 202, 107188.
761 Fiskin, R., Nasiboglu, E., & Yardimci, M. O., 2020. A knowledge-based framework for two-dimensional (2D)
762 asymmetrical polygonal ship domain. Ocean Engineering, 202, 107187.
763 Fujii, Y., & Tanaka, K., 1971. Traffic capacity. The Journal of navigation, 24(4), 543-552.
764 Gil, M., Wróbel, K., & Montewka, J., 2019. Toward a method evaluating control actions in STPA-based model
765 of ship-ship collision avoidance process. Journal of Offshore Mechanics and Arctic Engineering, 141(5).
766 Gil, M., Montewka, J., Krata, P., Hinz, T., & Hirdaris, S., 2020. Determination of the dynamic critical
767 maneuvering area in an encounter between two vessels: Operation with negligible environmental disruption.
768 Ocean Engineering, 213, 107709.
769 Gil, M., Wróbel, K., Montewka, J., & Goerlandt, F., 2020. A bibliometric analysis and systematic review of
770 shipboard Decision Support Systems for accident prevention. Safety science, 128, 104717.
771 Goodwin, E. M., 1975. A statistical study of ship domains. The Journal of navigation, 28(3), 328-344.
772 Goerlandt, F., Montewka, J., Kuzmin, V., & Kujala, P., 2015. A risk-informed ship collision alert system:
773 framework and application. Safety Science, 77, 182-204.
774 Gucma, L., & Marcjan, K., 2012. Probabilistic model of minimal passing distances of vessels navigating in Polish
775 coastal waters. 11th International Probabilistic Safety Assessment and Management Conference and the
776 Annual European Safety and Reliability Conference 2012, PSAM11 ESREL 2012, 2012. 5536-5543.
777 Hansen, M. G., Jensen, T. K., Lehn-Schiøler, T., Melchild, K., Rasmussen, F. M., & Ennemark, F., 2013.
778 Empirical ship domain based on AIS data. The Journal of Navigation, 66(6), 931-940.
779 Hörteborn, A., Ringsberg, J. W., Svanberg, M., & Holm, H., 2019. A Revisit of the Definition of the Ship Domain
780 based on AIS Analysis. The Journal of Navigation, 72(3), 777-794.
781 Hong, B. & Yu Y., 2000. Ship’s K, T indices statistics analysis. Journal of Dalian Maritime University, 26, 29-
782 33.
783 Hong, B., & Yang, L., 2012. Ship handling. Dalian Maritime University Press, Dalian.
784 Huang, Y., & Gelder, P. H. A. J. M. V., 2017. Non-linear velocity obstacles with applications to the maritime
785 domain. In 7th International Congress on Maritime Transportation and Harvesting of Sea Resources, Lisbon,
786 Portugal.
787 Huang, Y., & van Gelder, P. H. A. J. M., 2020. Collision risk measure for triggering evasive actions of maritime
788 autonomous surface ships. Safety science, 127, 104708.
789 Inoue, K., 2000. Evaluation method of ship-handling difficulty for navigation in restricted and congested
790 waterways. The journal of navigation, 53(1), 167-180.
791 Iqbal, K. S., Bulian, G., Hasegawa, K., Karim, M. M., & Awal, Z. I., 2008. A rational analysis of intact stability
792 hazards involving small inland passenger ferries in Bangladesh. Journal of marine science and technology,
793 13(3), 270-281.
794 Juszkiewicz, W., 2016. Verification of the accuracy requirements for relative course and closest point of approach.
795 Zeszyty Naukowe/Akademia Morska w Szczecinie, (45 (117)), 108-113.
796 Kijima, K., & Furukawa, Y., 2003. Automatic collision avoidance system using the concept of blocking area.
797 IFAC Proceedings Volumes, 36(21), 223-228.
798 Kim, D. H., 2020. Human factors influencing the ship operator's perceived risk in the last moment of collision
799 encounter. Reliability Engineering & System Safety, 203, 107078.
800 Kujala, P., Hänninen, M., Arola, T., & Ylitalo, J., 2009. Analysis of the marine traffic safety in the Gulf of Finland.
801 Reliability Engineering & System Safety, 94(8), 1349-1357.
802 Kulkarni, K., Goerlandt, F., Li, J., Banda, O. V., & Kujala, P., 2020. Preventing shipping accidents: Past, present,
803 and future of waterway risk management with Baltic Sea focus. Safety Science, 129, 104798.
804 Lehikoinen, A., Hänninen, M., Storgård, J., Luoma, E., Mäntyniemi, S., & Kuikka, S., 2015. A Bayesian network
805 for assessing the collision induced risk of an oil accident in the Gulf of Finland. Environmental science &
806 technology, 49(9), 5301-5309.
807 Li, S., Liu, J., & Negenborn, R. R., 2019. Distributed coordination for collision avoidance of multiple ships
808 considering ship maneuverability. Ocean Engineering, 181, 212-226.
809 Lu, C. S., & Yang, C. S., 2011. Safety climate and safety behavior in the passenger ferry context. Accident
810 Analysis & Prevention, 43(1), 329-341.
811 Montewka, J., Hinz, T., Kujala, P., & Matusiak, J., 2010. Probability modelling of vessel collisions. Reliability
812 Engineering & System Safety, 95(5), 573-589.
813 Montewka, J., & Przemyslak, K., 2014. Towards the assessment of a critical distance between two encountering
814 ships in open waters. European Journal of Navigation, 12(3), 7-14.
815 Montewka, J., Gil, M., Wróbel, K., 2020. Discussion on the article by Zhang & Meng entitled "Probabilistic ship
816 domain with applications to ship collision risk assessmentˮ [Ocean Eng. 186 (2019) 106130]. Ocean
817 Engineering 209, 107527. https://doi.org/10.1016/j.oceaneng.2020.107527
818 Montewka, J., Goerlandt, F., & Kujala, P., 2012. Determination of collision criteria and causation factors
819 appropriate to a model for estimating the probability of maritime accidents. Ocean engineering, 40, 50-61.
820 Mou, J. M., Van der Tak, C., & Ligteringen, H., 2010. Study on collision avoidance in busy waterways by using
821 AIS data. Ocean Engineering, 37(5-6), 483-490.
822 Nicholas, B., 2006. Risk perception and safety management systems in the global maritime industry. Policy and
823 Practice in Health and Safety, 4(2), 59-75.
824 Nomoto, K., Taguchi, K., Honda, K., & Hirano, S., 1956. On the steering qualities of ships. Journal of Zosen
825 Kiokai, 1956(99), 75-82.
826 Olsson, E., & Jansson, A., 2006. Work on the bridge–studies of officers on high-speed ferries. Behaviour &
827 Information Technology, 25(1), 37-64.
828 Pallotta, G., Vespe, M., & Bryan, K., 2013. Vessel pattern knowledge discovery from AIS data: A framework for
829 anomaly detection and route prediction. Entropy, 15(6), 2218-2245.
830 Pérez, F. L., & Clemente, J. A., 2007. The influence of some ship parameters on manoeuvrability studied at the
831 design stage. Ocean Engineering, 34(3-4), 518-525.
832 Pietrzykowski, Z., 2008. Ship's fuzzy domain-a criterion for navigational safety in narrow fairways. The Journal
833 of Navigation, 61(3), 499.
834 Pietrzykowski, Z., & Uriasz, J., 2009. The ship domain-a criterion of navigational safety assessment in an open
835 sea area. The Journal of Navigation, 62(1), 93.
836 Robert, G., Hockey, J., Healey, A., Crawshaw, M., Wastell, D. G., & Sauer, J., 2003. Cognitive demands of
837 collision avoidance in simulated ship control. Human factors, 45(2), 252-265.
838 Rong, H., Teixeira, A. P., & Soares, C. G., 2021. Spatial correlation analysis of near ship collision hotspots with
839 local maritime traffic characteristics. Reliability Engineering & System Safety, 209, 107463.
840 Simsir, U., Amasyalı, M. F., Bal, M., Çelebi, U. B., & Ertugrul, S., 2014. Decision support system for collision
841 avoidance of vessels. Applied Soft Computing, 25, 369-378.
842 Statheros, T., Howells, G., & Maier, K. M., 2008. Autonomous ship collision avoidance navigation concepts,
843 technologies and techniques. The Journal of Navigation, 61(1), 129-142.
844 Szlapczynski, R., & Szlapczynska, J., 2016. An analysis of domain-based ship collision risk parameters. Ocean
845 Engineering, 126, 47-56.
846 Szlapczynski, R., & Szlapczynska, J., 2017. Review of ship safety domains: Models and applications. Ocean
847 Engineering, 145, 277-289.
848 Szlapczynski, R., Krata, P., & Szlapczynska, J., 2018. Ship domain applied to determining distances for collision
849 avoidance manoeuvres in give-way situations. Ocean Engineering, 165, 43-54.
850 Tam, C., & Bucknall, R., 2010. Collision risk assessment for ships. Journal of marine science and technology,
851 15(3), 257-270.
852 Tao, J., Du, L., Dehmer, M., Wen, Y., Xie, G., & Zhou, Q., 2019. Path following control for towing system of
853 cylindrical drilling platform in presence of disturbances and uncertainties. ISA transactions.
854 Valdez Banda, O. A., Goerlandt, F., Montewka, J., & Kujala, P., 2015. A risk analysis of winter navigation in
855 Finnish sea areas. Accident Analysis & Prevention, 79, 100-116.
856 Valdez Banda, O. A., Kannos, S., Goerlandt, F., van Gelder, P. H., Bergström, M., & Kujala, P., 2019. A systemic
857 hazard analysis and management process for the concept design phase of an autonomous vessel. Reliability
858 Engineering & System Safety, 191, 106584.
859 Wang, N. 2010. An Intelligent Spatial Collision Risk Based on the Quaternion Ship Domain. Journal of
860 Navigation, 63, 733-749.
861 Wielgosz, M., 2017. Ship domain in open sea areas and restricted waters: an analysis of influence of the available
862 maneuvering area. TransNav: International Journal on Marine Navigation and Safety of Sea Transportation,
863 11.
864 Zhang, L., Meng, Q., Xiao, Z., & Fu, X., 2018. A novel ship trajectory reconstruction approach using AIS data.
865 Ocean Engineering, 159, 165-174.
866 Zhang, L., & Meng, Q., 2019. Probabilistic ship domain with applications to ship collision risk assessment. Ocean
867 Engineering, 186, 106130.
868 Zhang, M., Zhang, D., Goerlandt, F., Yan, X., & Kujala, P., 2019. Use of HFACS and fault tree model for collision
869 risk factors analysis of icebreaker assistance in ice-covered waters. Safety science, 111, 128-143.
870 Zhang, W., Goerlandt, F., Kujala, P., & Wang, Y., 2016. An advanced method for detecting possible near miss
871 ship collisions from AIS data. Ocean Engineering, 124, 141-156.
872 Zhang, W., Zou, Z., Wang, J., & Du, L., 2020. Multi-ship following operation in ice-covered waters with
873 consideration of inter-ship communication. Ocean Engineering, 210, 107545.
874 Zhang, X. G., & Zou, Z. J., 2011. Identification of Abkowitz model for ship manoeuvring motion using ε-support
875 vector regression. Journal of Hydrodynamics, 23(3), 353-360.
876 Zhao, J. S., Wu, Z. L. & Wang, F. C. 1993. Comments on Ship Domains. Journal of Navigation, 46, 422-436.
877 Zhao, L., & Shi, G., 2018. A method for simplifying ship trajectory based on improved Douglas–Peucker
878 algorithm. Ocean Engineering, 166, 37-46.
879 Zhou, Y., Daamen, W., Vellinga, T., & Hoogendoorn, S. P., 2019. Ship classification based on ship behavior
880 clustering from AIS data. Ocean Engineering, 175, 176-187.
881 Zhou, Y., Daamen, W., Vellinga, T., & Hoogendoorn, S. P., 2020. Impacts of wind and current on ship behavior
882 in ports and waterways: A quantitative analysis based on AIS data. Ocean Engineering, 213, 107774.
883 Zhu, M., Sun, W., Hahn, A., Wen, Y., Xiao, C., & Tao, W., 2020. Adaptive modeling of maritime autonomous
884 surface ships with uncertainty using a weighted LS-SVR robust to outliers. Ocean Engineering, 200, 107053.
885 Zhu, X., Xu, H., & Lin, J., 2001. Domain and its model based on neural networks. The Journal of Navigation,
886 54(1), 97-103..
887 Zhuo, Y., & Tang, T., 2008. An intelligent decision support system to ship anti-collision in multi-ship encounter.
888 In 2008 7th World Congress on Intelligent Control and Automation (pp. 1066-1071). IEEE.

889 Appendix
890 Table 1. List of Abbreviations
AIS Automatic Identification System MDTC Minimum Distance to Collision
AMM Available Maneuvering Margin NL-VO Non-Linear VO
ARPA Automated Radar Plotting Aid OS Own Ship
CAS Collision Alert Systems OT Overtaking encounter
COLREGs Convention on the International RB Relative Bearing
Regulations for Preventing
Collisions at Sea
CPA closest point of approach RH Relative Heading
CR Crossing encounter SD Ship Domain
DCPA Distance at Closest Point of SF Safe passing
Approach
DP algorithm Douglas–Peucker algorithm SO Stand-on ship
SPEE Ship-pair Encounter Event TS Target Ship
GW Give-way ship TCPA Time to Closest Point of Approach
HO Head-on encounter TTC Time to Collision
MASS Maritime Autonomous Surface VO Velocity Obstacles
Ships
891
892 Table 2. List of notations

AMM the value of calculated AMM t1 the first time that ship take evasive
maneuver
c ship course tSPEE time period of SPEE
h ship heading tea the time of ship taking evasive maneuver
 IC conflict index tob observation time
K turning ability index tr the period that collision risk exists
Lon longitude trs the starting moment of tr
Lat latitude tre the ending moment of tr
M AMM mean value of AMM ttp turning time
VTS TS’s velocity s the adopted rudder angle that can
eliminate the existing collision risk
v ship speed a all available rudder angle
RV reachable ship velocity C the amount of course change
S NL _ VO velocity obstacle zone at in TS’s AMM the change of AMM
velocity space
T turning lag index
893

View publication stats