Post on 01-Feb-2018
transcript
1
RISK ANALYSIS OF THE TRANSIT VESSEL TRAFFIC IN THE STRAIT
OF ISTANBUL
Özgecan S. Ulusçu (a), Birnur Özbaş (b),
Tayfur Altıok (c), İlhan Or (d)
(a , c) CAIT – DIMACS Laboratory for Port Security, Rutgers, The State University of New Jersey,
Piscataway, NJ 08854, USA, ozgecanu@eden.rutgers.edu
(b , d) Boğaziçi University, Department of Industrial Engineering,
34342, Bebek, Istanbul, TURKEY, birnur@ozbas.com.tr
Appered in J. of Risk Analysis
Volume 29, Number 10, pp. 1454 - 1472
2
ABSTRACT
The Strait of Istanbul, the narrow waterway separating Europe from Asia, holds a strategic importance in
maritime transportation as it links the Black Sea to the Mediterranean. It is considered as one of the
world’s most congested and difficult-to-navigate waterways. Over 55,000 transit vessels pass through the
Strait annually, roughly 20% of which carry dangerous cargo.
In this study, we have analyzed safety risks pertaining to transit vessel maritime traffic in the Strait of
Istanbul and proposed ways to mitigate them. Safety risk analysis was performed by incorporating a
probabilistic accident risk model into the simulation model. A mathematical risk model was developed
based on probabilistic arguments regarding instigators, situations, accidents, consequences and historical
data as well as subject-matter expert opinions. Scenario analysis was carried out to study the behavior of
the accident risks, with respect to changes in the surrounding geographical, meteorological and traffic
conditions.
Our numerical investigations suggested some significant policy indications. Local traffic density and
pilotage turned out to be two main factors affecting the risks at the Strait of Istanbul. Results further
indicate that scheduling changes to allow more vessels into the Strait will increase risks to extreme levels.
Conversely, scheduling policy changes that are opted to reduce risks may cause major increases in average
vessel waiting times. This in turn signifies that the current operations at the Strait of Istanbul have reached
a critical level beyond which both risks and vessel delays are unacceptable.
Keywords
: Risk Analysis, Risk Evaluation, Expert Judgment, Simulation, Maritime Traffic, Strait of
Istanbul
3
1 INTRODUCTION
The Turkish Straits, which consist of the Strait of Istanbul and Çanakkale and the Sea of Marmara, have
for centuries been one of the world’s most strategic waterways since they constitute the Black Sea's sole
maritime link to the Mediterranean and the open sea beyond. As such, they are a vital passageway not just
for trade but for the projection of military and political power, while their extreme narrowness, winding
contours and densely populated shores make navigation quite treacherous in these waterways.
The Strait of Istanbul is approximately 31 km long, with an average width of 1.5 kilometers and a mere
660m at its narrowest point. It features many sharp turns, forcing the ships to alter course at least 12 times,
some necessitating turns of up to 80 degrees. Additionally, frequent adverse meteorological conditions,
such as dense fogs and high currents and winds, contribute to the complexity of navigation in the Strait.
Figure 1. The Strait of Istanbul
There are also some non-natural factors making navigation through the Strait of Istanbul hazardous. One
of them is the dense local traffic, such as intra-city passenger boats, fast ferries, fishing boats, pleasure
boats etc. (1). Another important non-natural factor that negatively effects navigation in the Strait is the
frequency and cargo characteristics of transit vessels. About 56,600 vessels (10,000 being dangerous
material carriers) traveled through the Strait of Istanbul in 2007.
4
The legal framework of the transit traffic through the Strait of Istanbul is governed by the 1936 Montreux
Convention (2). When the Convention was put in effect, less than 5,000 vessels used to pass through the
Strait annually, whereas today, changes in international shipping patterns and volume have led to a ten-
fold increase. Additionally, during the 1930s, transport of hazardous materials posed little concern due to
the infrequent passages and small vessel sizes. However, the increases in traffic and vessel sizes have
raised the likelihood and potential severity of accidents. The unusual characteristics of the Strait of
Istanbul, coupled with the failure to request pilotage in this treacherous waterway, have led to over 200
accidents during the past decade.
Some major accidents have occurred since 1960, when the Greek-flagged M/T World Harmony collided
with the Yugoslavian-flagged M/T Peter Zoranic, leading to the death of 20 crew members, severe oil
pollution and fire that lasted several weeks, suspending the transit traffic. In 1979, Romanian-flagged
Independenta and the Greek freighter M/V Evriyalı collided at the southern entrance of the Strait. 43 crew
members died, 64,000 tons of crude oil spilled into the sea and 30,000 tons burned into the atmosphere. In
yet another catastrophe, the Greek Cypriot vessels M/T Nassia and M/V Shipbroker collided in the Strait.
29 officers and crewmen perished and 20,000 tons of crude oil burned for five days, suspending the traffic
for a week (3).
In order to control and mitigate maritime accident risks in the described dire environment and ensure the
safety of navigation, life, property and environment, the Turkish State unilaterally adopted in 1994 (and
later revised in 1998) the Maritime Traffic Rules and Regulations (R&R) for the Turkish Straits Region (4).
Then, in 2003, the Bureau of Turkish Strait’s Vessel Traffic Services (VTS) has been set up and
empowered to administer the R&R, through a sophisticated Vessel Traffic Control & Monitoring System,
(covering not only the Strait, but also 20 miles into the Black Sea and the Sea of Marmara).
Even though the number of accidents decreased after the adoption of the R&R, the vulnerability of the
Straits was evident once again in a 1999 incident; Voganeft-248, a Russian tanker, ran aground and broke
5
apart at the Southern entrance of the Strait. About 150,000 tons of oil spilled into the sea, and clean-up
efforts lasted several months (5).
The goal of this research is to analyze the risks involved in the transit vessel traffic in the Strait of Istanbul
and provide suggestions to reduce safety risks. We have carried out a model-based mathematical risk
analysis to be used in the risk mitigation process to improve safety in the Strait. In the first step of the risk
analysis, the transit vessel traffic system in the Strait of Istanbul was thoroughly analyzed and a simulation
model was developed to mimic and study maritime operations and surrounding environmental conditions.
In addition to transit vessel traffic through the Strait and geographical conditions, the current vessel
scheduling practices were modeled using a scheduling algorithm (6).
Risk analysis of the Strait was performed by incorporating a probabilistic accident risk model into the
simulation model. Probabilistic arguments utilized historical accident data, as well as subject-matter
expert opinions. Then a scenario analysis was performed in order to study the behavior of accident risks
and arrive at some critical policy suggestions. This analysis allowed us to investigate the impact of various
factors on the risk profile of the Strait. These factors included vessel arrivals, scheduling policies, pilotage,
overtaking, and local traffic density.
2 LITERATURE ON MARITIME RISK ANALYSIS
The existing risk analysis literature in maritime systems mainly focuses on probabilistic risk analysis
arguments, simulation modeling and statistical analysis of data. Below, we present a brief overview.
Amrozowicz (7) and Amrozowicz et al. (8) focus on the determination of the probability of oil tanker
grounding using fault-tree analysis. Dougligeris et al. (9) provides a methodology of quantifying and
assigning risk cost estimates in maritime transportation of petroleum products. Similarly, Iakovou (10)
considers the maritime transportation of crude oil and petroleum products using a multi-objective network
flow model allowing routing strategies for risk mitigation. Slob (11) presents an optimizing study to
minimize spills in the Dutch inland waterways.
6
Harrald et al. (12) describes the modeling of human error related accident event sequences in a risk
assessment study of maritime oil transportation in Prince William Sound, Alaska. Merrick et al. (13) and
Merrick et al. (14) present a detailed simulation model of the Prince William Sound oil transportation
system developed primarily for risk mitigation purposes. Merrick et al. (15) discusses the Washington State
Ferries Risk Assessment. A modeling approach was developed combining simulation, expert judgment
and risk analysis. Van Dorp et al. (16) describes a study that has been carried out to assess the sufficiency of
passenger and crew safety in the Washington State ferry system, while focusing on the estimation of the
level of risk present and possible risk reduction measures. As a supplement to Merrick et al. (15) the
potential consequences of collisions are modeled to determine the requirements for onboard and external
emergency response procedures and equipment. Merrick and Van Dorp (17) combines a Bayesian
simulation of the occurrence of situations with accident potential and a Bayesian multivariate regression
analysis. Kuroda et al. (18) proposes a mathematical model for estimating the probability of the collision of
vessels passing through a uniform channel. Kaneko (19) considers probabilistic risk assessment methods to
estimating the probability of collision. Yip (20) studies port traffic risks in Hong Kong by using a negative
binomial regression model and historical data, and provides a statistical evidence that the port of
registration, the vessel type and the accident type are critical to the number of injuries and fatalities.
Literature also includes work based on statistical analysis of data to model accident probabilities and
casualties. Maio et al. (21) developed a regression model to estimate the waterway casualty rate depending
on the geographic conditions. Roeleven et al. (22) describes approaches to obtain a forecasting model for
the probability of accidents as function of waterway attributes and circumstances. Anderson and Talley (23)
studies causal factors of oil cargo spills, and tanker barge vessel accidents, while Talley (24) investigates
the main factors of the risk and the severity of container cargo accidents. Similarly, Psaraftis et al. (25)
presents an analysis on the factors that are important determinants of maritime transportation risk. Le
Blanc and Rucks (26) describes the cluster analysis performed on a sample of over 900 vessel accidents that
occurred on the lower Mississippi River. Kite-Powell et al. (27) developed physical risk model is based on
7
a set of risk factors, including human error and vessel characteristics. Yudhbir and Iakovou (28) presents
the development of an oil spill risk assessment model. Moller et al. (29) reviews the current status of the
government-industry partnerships for dealing with oil spills as the result of maritime transportation. Talley
(30) discusses issues surrounding maritime safety and security in ports and coastal waterways.
The early work in Strait of Istanbul is somewhat limited. Kornhauser and Clark (31) used the regression
model developed by Maio et al. (21) to estimate the vessel casualties resulting from additional oil tanker
traffic through the Strait of Istanbul. Otay and Özkan (32) developed a simulation model to estimate the
probability distribution of vessel casualties using the geographical characteristics of the Strait of Istanbul.
Tan and Otay (33) and later Tan and Otay (34) present a physics-based stochastic model to investigate
casualties resulting from tanker accidents in the Strait of Istanbul. Or and Kahraman (35) investigates
possible factors contributing to accidents in the Strait of Istanbul using Bayesian analysis and simulation
modeling. Akten (36) studies maritime accidents in the Strait of Istanbul and states that current, daytime
and location are the most important factors affecting casualties.
3 A MATHEMATICAL RISK MODEL
As mentioned earlier, numerous instigators have been witnessed resulting in an accident or a closure while
a transit vessel navigates in the Strait of Istanbul. For example, there can be a mechanical failure in the
vessel, or the captain or the pilot can make an error. These causal occurrences that may trigger an accident
are referred to as instigators. They include human error, rudder failure, propulsion failure, communication
and/or navigation equipment failure, and mechanical and/or electrical failure. Clearly, the occurrence of
an instigator depends on the situation, which may be represented by a vector of situational attributes.
Typical accidents that may occur in the Strait include collision, grounding, ramming, sinking and fire
and/or explosion. Accidents may occur in chain in such a way that an accident may cause another
accident. 1st tier accident types include collision, grounding, ramming and fire and/or explosion, while the
2nd tier accident types (that may occur following a 1st tier accident) include grounding, ramming, fire
and/or explosion, and sinking. Potential consequences of the 1st and 2nd tier accidents include human
8
casualty, property and/or infrastructure damage, environmental damage and loss of traffic effectiveness
and throughput. This framework is shown in Figure 1.
INSTIGATORS 1st TIER ACCIDENTS 2nd TIER ACCIDENTS
Human Error
Steering Failure
Propulsion Failure
Communication / Navigation Equipment
Failure
Mechanical / Electrical Failure
Collision
Grounding
Ramming
Fire / Explosion
No 2nd Tier Accident
Grounding
Ramming
Fire / Explosion
Sinking
CONSEQUENCES
Human Casualty
Property-Infrastructure Damage
Environmental Damage
Traffic Effectiveness
Figure 1. The framework of the risk model
Defining situations (factors and their states) that affect the likelihood and/or impact level of instigators and
accidents is critical for the intended risk analysis. Such factors, called Situational Attributes, are divided
into two groups: attributes influencing accident occurrence and attributes influencing consequences. These
two groups of situational attributes are further classified as vessel and environmental/shore attributes as
displayed in Figure 2 and 3. Vessel attributes influencing accident occurrence include class, reliability and
whether the vessel has pilot and/or tugboat. Reliability is determined by vessel’s age and its flag as
suggested by the 2006 Shipping Industry Flag State Performance Table (37). Environmental attributes
include distance between transit vessels, weather and water conditions, geographical difficulty, density of
local traffic and navigating daytime or nighttime. Attributes influencing consequence impact also include
vessel attributes such as cargo type and vessel length and shore attributes such as population, historical
structures, commercial and residential property, and infrastructure.
9
Figure 2. Situational attributes influencing accident occurrence
Figure 3. Situational attributes influencing the consequences
Note that in order to quantify risks, we need to answer the following questions:
• How often do the critical situations occur?
• For a particular situation, how often do instigators occur?
• If an instigator occurs, how likely is an accident?
• If an accident occurs, what would the damage to human life, property, environment and
infrastructure be?
ENVIRONMENTAL ATTRIBUTES
Vessel Class(Type & Length)
Vessel Reliability(Age & Flag)
Pilot Request
Tugboat Request
Vessel Proximity
Visibility
Current
Geographical Difficulty
Local Traffic Density
ATTRIBUTES INFLUENCING ACCIDENT OCCURRENCE
VESSEL ATTRIBUTES
SHORE ATTRIBUTES
Vessel Class(Type & Length)
Population
Historical Buildings
Property
Infrastructure
ATTRIBUTES INFLUENCING CONSEQUENCE IMPACT
VESSEL ATTRIBUTES
10
In this study, risks are quantified based on historical data, expert judgment elicitation and a simulation
model of the transit vessel traffic in the Strait of Istanbul. 2005 and 2006 data acquired from the VTS
include inter-arrival time, speed, pilot request percentage, tugboat request percentage, anchorage
percentage, anchorage duration, stopover percentage, age and flag of different classes of vessels (nearly
for 100.000 vessels). The local traffic schedules for the ferries, motorboats, and tourist boats obtained
from the websites of companies. 1988-2005 visibility data were obtained from the Kandilli Observatory
and Earthquake Research Institute (38). 1991-2005 visibility data obtained from the International Weather
Information Website (39), Surface current data were obtained from VTS and the Department of Navigation,
and Oceanography of the Turkish Navy (40). The simulation model first generates arrivals of vessels,
according to classes shown in Table I along with a number of vessel attributes, using probability
distributions based on these data. It then schedules vessel entrances and administers their pilotage needs,
navigation through the Strait (with such details as speed and overtaking) and their exit with all the relevant
local traffic, weather and current conditions. All these processes are designed and executed in line with the
R&R, while the scheduling algorithm closely mimics the practices and conventions of the VTS (regarding
decisions on sequencing vessel entrances in daytime and nighttime, as well as start time and duration of
the northbound, southbound and bi-directional travel time-windows. The model was validated using vessel
transit times and scheduling decisions made by the operators in some days of 2005. The results were
highly satisfactory (6). Finally accident data obtained from VTS, Turkish Undersecretariat for Maritime
Affairs and Ece (41) for last 60 years were used for the calibration process.
11
Table I. Vessel types in the Strait of Istanbul for scheduling purposes
Length (m.)
Draft (m.) Tanker LNG-LPG Carrying
Dangerous Cargo Dry Cargo Passenger Vessels
< 50 < 15
50 - 100 < 15
100 - 150 < 15
150 - 200 < 15
200 - 250 < 15
250 - 300 > 15
> 300 > 15
Type
Class EClass D
Class EClass C
Class P
Class T6
Class A
Class B
Class C
In the model, the Strait of Istanbul is divided into 21 slices (each 8 cables long, where 1 cable = 0.92
miles) for risk presentation purposes, as depicted in Figure 4. The risk in slice s, ( )sR S , is computed
based on the snapshot (defined by S ) of the slice every time a vessel enters it (in either direction). The
snapshot contains information about all the situational attributes (vector S ) as observed by the entering
vessel. First, the risk contribution of the vessel entering the slice is computed and then the contribution of
each vessel in the snapshot is computed and all are aggregated into the slice risk in a cumulative manner.
Since there are about 55,000 transits annually, such snapshots of instantaneous risks are taken that many
times in each slice. Accordingly, even though instantaneous risks are not continously tracked and
recorded, the sampling of these risks (accomplished at each vessel entrance) is sufficiently random and
numerous to provide a healthy analysis. Also, note that an alternative approach would be to take snapshots
of every slice at every properly chosen t∆ time units. Both approaches can be successfully applied to
ports or waterways to create meaningful samples of instantaneous risks.
12
Figure 4. Risk slices at the Strait of Istanbul
The risk in slice s, ( )sR S , as observed by a vessel entering slice s, is given by
( ) ( )1 2
( ) , Pr , Pr |s j vs vs vs j vs vs vssv j j
R S E C A S A S E C A S A Sϕ σ
ϕ ϕ ϕ σ σ σϕ σ∈ ∈ ∈ ∈ ∈
= × + ×
∑ ∑ ∑ ∑ ∑
V A A C C
(1)
where
vsAϕ : 1st tier accident type φ in slice s involving vessel v
vsAσ : 2nd tier accident type σ in slice s involving vessel v
1A : Set of 1st tier accident types
2A : Set of 2nd tier accident types that may be caused by 1st tier accident type φ as indicated in Table II
j vsC ϕ : Consequence type j of 1st tier accident type φ in slice s initiated by vessel v
j vsC σ : Consequence type j of 2nd tier accident type σ in slice s initiated by vessel v
ϕC : Set of consequence types of accident type φ ( 1ϕ∈A ) as indicated in Table III
σC : Set of consequence types of accident type σ ( 2σσ ∈A ) as indicated in Table III
13
sV : Set of vessels navigating in slice s, as seen by the observing vessel
sX : Vector of situational attributes affecting accident occurrence
sW : Vector of situational attributes affecting consequence of an accident
( , )s sS X W∈
Note: When there are no 2nd tier accidents, the second term in (1) disappears.
We can also interpret ( )sR S as the instantaneous risk in slice s, conditioned on the particular realization
of S at the time of a vessel entry into slice s. Thus the expected risk for slice s is obtained by averaging
( )sR S over the number of vessels that have entered slice s. Since this process is done for each slice, a
risk profile is obtained for the entire Strait at the end of a simulation run.
Table II. Causal relationship between the 1st and 2nd tier accident types
No 2nd Tier
Accident
2nd Tier Accident Type
Grounding Ramming Fire / Explosion Sinking
1st T
ier
Acc
iden
t Typ
e Collision Grounding Ramming Fire / Expl.
(For example: a grounding may either not lead to a 2nd tier accident or it may lead to fire/explosion or sinking, as a
2nd tier accident)
Table III. Set of accident consequences
Consequences
Property / Infrastructure
Damage
Human Casualty
Environmental Damage
Traffic Effectiveness
Acc
iden
ts
Type
s
Collision Grounding Ramming Fire/Explosion Sinking
14
The probability of a 1st tier accident type φ in slice s involving vessel v based on the situation vector S
observed by an entering vessel is given by
( ) ( ) ( ) ( )I
Pr | Pr | Pr , Prvs vs s vs is s is si
A S A X A I X I Xϕ
ϕ ϕ ϕ∈
= = ×∑ (2)
where,
isI : Instigator type i in slice s
ϕI : Set of instigators that may cause a 1st tier accident type φ as given in Table IV.
Table IV. Set of instigators that may cause an accident
1st Tier Accidents
Collision Grounding Ramming Fire / Explosion
Inst
igat
ors
Human Error Steering Failure Propulsion Failure Comm./Nav. Equipment Failure
Mechanical/Electrical Failure
The probability of a 2nd tier accident type σ in slice s involving vessel v is given by
( ) ( ) ( ) ( ) ( ) ( )1 1
Pr | Pr | Pr , Pr | Pr Pr | .vs vs vs vs vs vs vs vsA S A X A A X A X A A A Xσ σ σ ϕ ϕ σ ϕ ϕϕ ϕ∈ ∈
= = × ≅ ×∑ ∑A A
(3)
Note we assume that 2nd tier accidents depend only on the 1st tier accidents and not the situation vector S .
The expected consequence j in slice s given 1st tier or 2nd tier accident type φ or σ are respectively given by
( ), , Pr ,h hj vs vs j vs vs s j vs j vs vs s
h LE C A S E C A W C C A Wϕ ϕ ϕ ϕ ϕ ϕ ϕ
∈
= = × ∑ (4)
( ), , Pr ,h h
j vs vs j vs vs s j vs j vs vs sh L
E C A S E C A W C C A Wσ σ σ σ σ σ σ∈
= = × ∑
(5)
15
where,
L : Set of impact levels of which h is the index (low, medium, high)
hj vsC ϕ : Consequence type j at level h of 1st tier accident type φ in slice s contributed by vessel v
hj vsC σ : Consequence type j at level h of 2nd tier accident type σ in slice s contributed by vessel v.
Clearly, conditional probabilities, marginal probabilities and conditional expectations are involved in the
expressions (1) - (5). Among them, ( )Pr ,vs is sA I Xϕ and ( )Pr is sI X
(for human error),
( )Pr ,hj vs vs sC A Wϕ ϕ
and ( )Pr ,hj vs vs sC A Wσ σ are all obtained using expert judgment, and ( )Pr is sI X (for
instigators other than human error) and ( )Pr vs vsA Aσ ϕ
are obtained from historical data.
During this study, a total of 53 questionnaires were answered by numerous experts with different
backgrounds and the average response times varied from hours to days depending on their availabilities.
Also, consistencies of the experts’ answers are double checked and maintained.
• A questionnaire for every Accident Type/Instigator Combination:
– A total of 14 questionairres;
– 32 to 40 questions in each questionnaire.
• Regarding Human Error Instigator:
– 1 questionnaire;
– 40 questions.
• A questionnaire for every Accident Type/Consequence Level Combination:
– A total of 38 questionnaires;
– 8 to 20 questions in each questionnaire.
16
• Interviews with 23 Experts:
– 5 Academicians (ITU School of Maritime Affairs)
– 10 Pilot Captains
– 3 VTS Marine Traffic Experts
– 5 Open Sea Captains (from Private Sector)
3.1 Probability of a First Tier Accident
In order to compute the probability of a 1st tier accident in a slice based on the situation vector sX , one
needs to evaluate (2) involving,
i) probability of a 1st tier accident given an instigator and a situation,
ii) probability of an instigator given sX .
In addition to these evaluations, a calibration process needs to be carried out to make sure that the long-
run accident probabilities are legitimate.
3.1.1 Probability of a 1st Tier Accident Given an Instigator and a Situation
Due to lack of data to determine the contribution of various situational attributes to accident risks,
estimation of the conditional probability of an accident given an instigator requires elicitation of expert
judgment.
In this research, “a paired-comparison elicitation method” (42) is deployed in the extraction of expert
judgment. This is because experts are regarded to be more comfortable in making paired comparisons,
rather than directly assessing a probability value for a given situation. This approach was also used by
Merrick et al. (15), Merrick and Van Dorp (17) and Szwed et al. (43).
17
The paired-comparison approach focuses on the functional relationship between situational attributes
1 2( , , , )TpX X X Xϕ
= where pφ is the number of situational attributes influencing occurrence of 1st tier
accident φ the given accident occurrence) and an accident probability rather than a value function. The
probability of a 1st tier accident given an instigator and the situation vector sX can be written as,
( ) ( )Pr , expTi iA I X P Xϕ ϕ ϕβ= (6)
where iϕβ is the vector of parameters and iPϕ is the calibration constant for the 1st tier accident type φ
and instigator type i. The accident probability model (6) is also used by Roeleven et al. (22), Merrick et al.
(13) and Van Dorp et al. (16). It is based on the proportional hazards model originally proposed by Cox (44),
which assumes that accident probability behaves exponentially with changes in covariate values.
The probability of an accident defined by (6), where [0,1]pX ∈ , i pRϕβ ∈ and [0,1]iPϕ ∈ , is assumed to
depend on the situational attributes listed in Table V. A cardinal ranking is obtained through expert
opinion to create a consistent scale for each situational attribute. This is accomplished by offering each
expert a series of pairwise comparisons of different levels of each attribute iX , such that the aggregation
of his/her binary responses to these comparisons provide a subjective cardinal ranking of this expert
regarding the described levels of a specific attribute. Finally, the averages of such weights over all experts
constitute the cardinal ranking of the described levels of each attribute. The values of iX are normalized
so that 1iX = describes the “worst” case scenario, while 0iX = describes the “best” case scenario. For
example, regarding the 11th attribute, (time of the day), 11 1X = represents the nighttime, while 11 0X =
represents the daytime. The possible settings of the situational attributes influencing accident occurrence
are listed in Table V.
18
Table V. Possible settings of situational attributes influencing accident occurrence
Variable Attribute Name Number of
Possible Settings
Description
1X 1st Interacting Vessel Class 9 Vessel types in Table I rearranged
from 1 to 9
2X 2nd Interacting Vessel Class 11 Similar to 1X , including 2 types of
local vessels
3X 1st Vessel Tugboat Request 2 Yes, No
4X 1st Vessel Pilot Request 2 Yes, No
5X
Nearest Transit Vessel Proximity
(from a given transit vessel)
9
same direction 0-4 cables, same direction 4-8 cables; same direction >8
cables, 1 knot/hr speed difference overtaking lane, 2 knots/hr speed
difference overtaking lane, 3 knots/hr speed difference overtaking lane, 4
knots/hr speed difference overtaking lane, opposite direction normal lane, opposite direction overtaking lane
6X Visibility 3 <0.5 mile, 0.5-1 mile , >1 mile
7X Current 8
0-2 knots/hr same direction with 1st vessel, 2-4 knots/hr same direction with 1st vessel, 4-6 knots/hr same
direction with 1st vessel, > 6 knots/hr same direction with 1st vessel, 0-2 knots/hr opposite to 1st vessel, 2-4 knots/hr opposite to 1st vessel, 4-6 knots/hr opposite to 1st vessel, > 6
knots/hr opposite to 1st vessel
8X Local Traffic Density 3 1-2, 3-5, >5
9X Zone 12
Anadolu Feneri-Sarıyer SB, Anadolu Feneri-Sarıyer NB, Sarıyer-Beykoz SB, Sarıyer-Beykoz NB, Beykoz-Kanlıca SB, Beykoz-Kanlıca NB,
Kanlıca-Vaniköy SB, Kanlıca-Vaniköy NB, Vaniköy-Üsküdar SB,
Vaniköy-Üsküdar NB, Üsküdar-Kadıköy SB, Üsküdar-Kadıköy NB
10X Vessel Reliability 9 Age (New, Middle Age, Old) x Flag Category (Low Risk, Medium Risk,
High Risk)
11X Time of the Day 2 Daytime, Night time
19
In addition to the individual situational attributes listed in Table V, attributes describing interaction effects
are included in the model. For example, 12 1 9X X X= ⋅ represents the interaction between the 1st
interacting vessel class and the zone. The aim is to capture the combined impact of certain key attributes
on the accident probability. There are 12 interaction attributes as seen in Table VI.
Table VI. Interaction attributes
Interaction Description 12X 1 9X X⋅ 1st Interacting Vessel Class x Zone
13X 4 7X X⋅ 1st Vessel Pilot Request x Current
14X 4 9X X⋅ 1st Vessel Pilot Request x Zone
15X 3 9X X⋅ 1st Vessel Tugboat Request x Zone
16X 3 7X X⋅ 1st Vessel Tugboat Request x Current
17X 5 6X X⋅ Nearest Transit Vessel Proximity x Visibility
18X 5 7X X⋅ Nearest Transit Vessel Proximity x Current
19X 7 9X X⋅ Current x Zone
20X 6 8X X⋅ Visibility x Local Traffic Density
21X 6 9X X⋅ Visibility x Zone
22X 9 8X X⋅ Zone x Local Traffic Density
23X 10 4X X⋅ Time of the Day x 1st Vessel Pilot Request
To assess the accident probability given an instigator, subject matter experts are asked to compare a series
of situation pairs ( 1 2,X X ) differing only in one attribute (regarding their ratio of probabilities to cause an
accident). If the expert thinks that the likelihood of an accident is the same in both situations, then he/she
circles “1”. Otherwise, the expert circles a value towards the likelier event (the higher ratio of the
probability, the higher the circled value). Figure 5 provides a sample question from one of the related
questionnaires.
For each 1st tier accident, a separate questionnaire was prepared containing 4 questions per attribute (one
question representing the worst case scenario, one representing the best case, and two others
corresponding to ordinary cases).
20
Figure 5. A sample question for impact of instigators on collision probability
Let ,e qz be the response of expert e ( 1, ,e r= ) to question q ( 1, , 4q pϕ= ), where r is the total
number of experts and 4pφ is the total number of questions in each questionnaire. Then, we can write
( )( )
( )( ) ( )( )11
, 1 22 2
expPr ,exp
Pr , exp
T
T
T
i iqq i
e q q qi iq q
P XA I Xz X X
A I X P X
ϕ ϕϕϕ
ϕ ϕ ϕ
ββ
β= = = − (7)
where ( )1 2q qX X− denotes the difference vector of the two situations evaluated by the differences of
cardinality ratings assigned to attribute settings. Therefore, given an instigator, the ratio of accident
probabilities in two situations depends only on the difference vector ( )1 2q qX X− and the parameter
vector iϕβ . Since the experts are asked to assess the ratio in (7), the parameter vector i
ϕβ can be estimated
using these difference vectors and without determining the accident probability itself.
21
The aggregate expert responses are taken as their geometric mean:
1
,1
r r
q e qe
z z=
= ∏ . (8)
Then, linear regression can be deployed to determine the relative effect of the situational attributes on the
accident probability. In other words,
( )1 2
Tiq q q qy X Xϕβ ε= − + (9)
where ln( )q qy z= and qε is the residual error.
Under the assumption that qε is normally distributed (εq ∼ σ 2. . (0, )i i d N ), (9) can be regarded as a set
of linear regression equations, where aggregate expert responses are the dependent variables,
1 2( )q qX X− are the independent variables and }{ ,0 ,1 ,2 ,: , , , ......,i i i i ipϕϕ ϕ ϕ ϕ ϕβ β β β β are the regression
parameters, to be estimated for each (accident type - instigator) pair. The R values associated with 9 are
between 0.955 and 0.985 for all accident type-instigator pairs. As an example normal probability plot of
residuals and residuals vs. predicted variable (collision-human error) pair are given in Figure 6. Also
Figure 7 shows contribution of each attribute and their interactions for the collision probability given a
human error as an instigator as a result of this regression.
Normal Prob. Plot
-1.2-1.0-0.8-0.6-0.4-0.20.00.20.40.60.8
-1.0 -0.5 0.0 0.5 1.0
Normal Distribution
Res
idua
ls
Residuals vs Pred Y
-1.2-1.0-0.8-0.6-0.4-0.20.00.20.40.60.8
-2 -1 0 1 2 3
Predicted Y
Res
idua
ls
22
Figure 6. Normal probability plot of residuals and residuals vs. predicted variable of (collision-human
error) pair
-5
-4
-3
-2
-1
0
1
2
3
4
5
X1 X2 X3 X4 X5 X6 X7 X8 X10 X15 X16 X19 X20 X21 X22
β Coefficients
Figure 7. β coefficients for impact of human error on collision probability
3.1.2 Probability of an Instigator Given a Situation
Due to the lack of data regarding the occurrence of human error in accidents in the Strait, expert judgment
is deployed to derive conditional human error probabilities, while historical accident data is used to derive
occurrence probabilities of the remaining instigators. Clearly, occurrence of human error is affected by the
particular vessel attributes, as well as environmental attributes. So, its conditional probability is estimated
using the paired-comparison approach described earlier. Thus, the probability of human error is written as,
( ) ( )Pr Human Error exp TiX P Xα= (10)
where iP is the calibration constant and α is the parameter vector for the human error probability. To
assess the conditional human error probability, experts were asked to compare a series of situation pairs
( 1 2,X X ). Figure 8 displays a sample question from the human error questionnaire, which consists of 40
questions.
23
Figure 8. A sample question representing the impact of situational attributes on human error
The regression equation describing the relative effect of situational attributes on human error probability is
given by,
( )1 2T
q q q qy X Xα ε= − + (11)
where qε is the residual error. Again, under the assumption that qε is normally distributed, (11) can be
regarded as a set of linear regression equations, and the parameter vector α can be obtained by solving it.
Figure 9 shows contribution of each attribute and their interactions for the human error probability.
24
Figure 9. α coefficients for human error probability
Let us now go back to the probability of a 1st tier accident based on the situation vector sX as observed by
a vessel entering a slice. Recall that the accident probabilities are used to compute slice risks at points of
vessel entry into slices. As an example, consider the probability of collision:
( ) ( ) ( )Pr Collision,Human Error | Pr Collision Human Error, Pr Human Errors s sX X X= × (12)
Combining (2), (6), (10) and (12) and all the instigators, we can write,
( )
( )
( )
1
1
1 1
2 2 21 1, 1,0
1
3 3 31 1, 1,0
1
1 1 11 1, 1,0 1 1, 1,0
1 1
4 41 1,
Pr Collision in slice |
exp Pr Steering Fail
exp Pr Propulsion Fail
exp exp
exp
s
p
i i si
p
i i si
p p
i i i ii i
i i
s X
P x X
P x X
P x P x
P x
β β
β β
β β α α
β
=
=
= =
=
+ + ×
+ + ×
+
+ × +
∑
∑
∑ ∑
( )1
41,0
1Pr Comm/Nav Fail
p
si
Xβ=
+ ×
∑
(13)
25
where each term of the summation represents the joint probability of collision and an instigator, given sX .
We can use (13) in evaluating (1) provided all the calibration constants and the conditional probabilities
are known.
3.1.3 Calibration
Since coefficient iϕβ , for accident ϕ and instigator i, of the conditional probability function (6) are
obtained through the pair-wise comparison process, a further calibration of the constants iP and iPϕ is
necessary to make sure that values generated by (6) will correspond to their counterparts obtained from
historical data. For calibration purposes, we first let 1 41 1 1 1P P P= = = =
and
( ) ( ) ( )Pr Steering Fail Pr Propulsion Fail Pr Comm/Nav Fail 1s s sX X X= = =
in (13) and make a
preliminary run of the simulation model. Averaging vessel observations of each component of (13) in the
long run provides us with the joint probabilities of accident types and instigators conditioned on the
aforementioned parameters being 1. Note that this process removes the condition on the situation
vector sX . Then, these values are compared with their counterparts (e.g. ( )Pr Collision, Human Error ,
etc.) obtained from the historical data. Additionally, we have,
( ) ( ) ( )Pr Collision,Human Error Pr Human Error Collision Pr Collision= × (14)
Using historical data, we can estimate ( )Pr Human Error Collision through,
( ) Number of collisions due to human errorPr Human Error Collision
Total number of collisions,= (15)
and approximate the probability of a vessel being involved in a collision by
( ) Number of vessels involved in collisionsPr A vessel involved in a collision
Total number of vessels.= (16)
26
Then, ( )Pr Collision,Human Error can be estimated using (14) – (16).
Let the long run unconditional P(Collision,Human Error) that
is
P11 exp β
1, i1 x
ii=1
p1
∑ + β1,01
× P
1exp α
1, ix
ii=1
p1
∑ +α1,0
× Pr X
ls( )l∈X∑ in (13) be 1
1 1 1P P C , which reduces to C1
since 11P and 1P are set to unity in the preliminary run of the model. Thus, the comparison of
( )Pr Collision,Human Error with its counterpart obtained from the historical data, will provide an estimate
for the product of calibration constants 11 1P P through
( )1
1 11
Pr Collision,Human Error.P P
C= (17)
Similarly, for the remaining joint accident and instigator probabilities, we will use
( ) ( )21
2
Pr Collision,Steering FailPr Steering Fail ,P
C× = (18)
( ) ( )31
3
Pr Collision,Propulsion FailPr Propulsion Fail ,P
C× = (19)
and
( ) ( )41
4
Pr Collision,Comm/Nav FailPr Comm/Nav Fail .P
C× = (20)
Note we made an approximation here by assuming that instigators do not depend on the situational
attribute vector sX . This was necessary since data did not provide joint probabilities of instigators and
situations. One is now ready to compute conditional as well as marginal accident and instigator
probabilities (used in (2)), and obtain the probability of a first tier accident.
27
3.2 Probability of a Second Tier Accident
The conditional probability of a 2nd tier accident given the occurrence of a 1st tier accident is estimated
using the historical accident data as shown below:
( )st nd
st
Number of type 1 tier accidents that lead to a type 2 tier accidentPr
Total number of type 1 tier accidents.A Aσ ϕ
ϕ σϕ
= (21)
Values of ( )Pr A Aσ ϕ for the Strait are given in Table VII.
Table VII. Values for Pr (2nd tier Accident|1st tier Accident)
2nd tier Accident
No 2nd Tier
Accident Grounding Ramming Fire /
Explosion Sinking
1st ti
er A
ccid
ent Collision 0.8737 0.0289 0.0000 0.0158 0.0816
Grounding 0.9794 0.0041 0.0165
Ramming 0.8325 0.1218 0.0102 0.0355 Fire /
Explosion 0.9355 0.0081 0.0000 0.0565
3.3 Expected Consequence Given an Accident
For slice risk computation, we next need to obtain the expected consequence of accidents that may be
caused by the situations observed by an entering vessel, that is, j vs vsE C Aϕ ϕ and j vs vsE C Aσ σ ) as
presented below.
3.3.1 Probability of a Consequence Given an Accident
Due to the lack of any consequence data, expert judgment is relied upon in the estimation regarding the
probability of specific impact level realizations. It is assumed that the probability of impact level depends
on accident type and situational attributes. The list of considered situational attributes influencing impact
level and their possible settings are given in Table VIII.
28
As seen in Table VIII, the 1st interacting vessel type has three different settings for different consequence-
accident type pairs. For example, regarding (environmental damage, collision) pair, this attribute has five
possible settings in terms of cargo type and amount; while regarding (human casualty, collision) pair, it
has three settings based on the number of people in the vessel.
Table VIII. Possible settings of situational attributes influencing consequence impact
Variable Attribute Name
Number of
Possible Values
Description
1W 1st Interacting Vessel Type
6 LNG-LPG, Tanker, Empty LNG-LPG, Empty Tanker; Passenger,
other vessel 2 Passenger vessel, other vessel
3 Loaded LNG-LPG and Tanker, Passenger, other vessel
2W 2nd Interacting Vessel Type
6 LNG-LPG, Tanker, Empty LNG-LPG, Empty Tanker; Passenger,
other vessel 2 Passenger vessel, other vessel
3 Loaded LNG-LPG and Tanker, Passenger, other vessel
3W 1st Interacting Vessel Length 2 0-150m., 150-300m.
4W 2nd Interacting Vessel Length 2 0-150m., 150-300m.
5W Zone 6
Anadolu Feneri-Sarıyer, Sarıyer-Beykoz, Beykoz-Kanlıca,
Kanlıca-Vaniköy, Vaniköy-Üsküdar, Üsküdar-Kadıköy
6W 1 2W W⋅ 1st Interacting Vessel Type x 2nd Interacting Vessel Type
7W 3 4W W⋅ 1st Interacting Vessel Length x 2nd Interacting Vessel Length
8W 1 5W W⋅ 1st Interacting Vessel Type x Zone
9W 3 5W W⋅ 1st Interacting Vessel Length x Zone
The conditional probability of consequence impact level is estimated using the paired comparison
approach described earlier. Thus, the conditional probabilities of impact level of consequence type j given
an 1st tier or 2nd tier accidents are given by φ or σ and the situation vector W,
29
( ) ( ) ( ) ( )Pr , exp Pr , expT Th h h h h h
j j j j j jC A W G W o r C A W G Wϕ ϕ ϕ σ σ σγ γ= = (22)
where hjGϕ and h
jGσ are calibration constants and hjϕγ and h
jσγ are the parameter vectors. To assess the
probability of impact level given an accident, experts are asked a series of questions, each comparing two
situations 1W and 2W (varying from one another in a single attribute), for each impact level. Table IX
displays the general descriptions provided to the experts regarding impact levels, so that experts’
perceptions of the terms deployed (such as low/medium/high severity) would not vary greatly. Figure 10
displays a sample question from the questionnaire for the (fire/explosion, human casualty) pair. A separate
questionnaire is prepared for each (consequence, accident type) pair, where four questions are asked per
situational attribute (one question representing the worst case scenario, one representing the best case
scenario, and two others corresponding to intermediate scenarios).
Figure 10. A sample consequence question
Similar to (9), the resulting set of regression equations used to determine the relative effect of the
situational attributes on the probability of impact levels given an accident are,
30
( ) ( )1, 2, 1, 2,
T Th hq j q q q q j q q qy W W o ry W Wϕ σγ ε γ ε= − + = − + (23)
where qε is the residual error.
Since the hjϕγ
and hjσγ parameters of the conditional probability functions (22) are derived through the
aforementioned relative comparison process, a further calibration of the constants h
jGϕ and hjGσ is
necessary. We have used the historical accident data in this calibration process. In this regard, the number
of past accidents which received negligible coverage, some coverage, headline coverage in the media
(regarding their consequences) have been deployed to calibrate the conditional probability functions
associated with accident consequences (in terms of low, medium, high impact levels).
Conditional probabilities of consequences can be computed in a manner similar to the ones used in
computing accident probabilities. As an example, consider the probability of low casualty given collision,
that is,
( )( ) ( )( )11
1 1 111 11, 11,0
1
Pr Casualty Low Collision , Pr Casualty Low Collision ,
exp .
s ls
p
i ii
W W
G xγ γ=
=
= +
∑
(24)
Finally, the conditional probabilities of consequences are normalized to make
sure ( ) ( )Pr 1 Pr 1L L
h hj j
h handC A C Aϕ ϕ σ σ
∈ ∈= =∑ ∑ .
3.3.2 Consequence Impact Level
It is assumed that the quantitative values of impact levels of a consequence type j, ( hj vsC ϕ or h
j vsC σ ) are
uniformly distributed within pre-specified ranges, as displayed in Table IX. These values/ranges
(representing the user’s perception of a low, medium and high consequence) do not represent the actual
impact of an accident in a specific unit (e.g. dollars or number of casualties) and can be altered in a
parametric fashion in a series of scenarios, in order to study the effects of different quantification
31
conventions. As such, the computed risk values are meaningful only when compared to each other in a
given context. For example, comparing risk at different slices helps to determine high and low risk zones.
Table IX. Consequence impact levels
Impact Level Value
Property / Infrastructure
Damage
Human Casualty
Environmental Damage
Traffic Effectiveness
Low Uniform (0-1,000) Small boats, tugs, fishing nets, shore No death No spills No closures
Medium Uniform (4,000-6,000)
1-2 vessels , ferries,
breakwater, cables
1 death Less than 1000 tone
Less than 4 hours
High Uniform (8,000-10,000)
Ports, historical buildings, waterside
residences, valuable goods,
More than 1 death
More than 1000 tone
More than 4 hours
Then, the conditional expectation of consequences are obtained using (4) - (5), which in turn allows us to
compute the expected total consequence that is nothing but the risk we have been striving to compute from
the very beginning.
4 OBSERVATIONS AND CONCLUSIONS
We have experimented with the aggregate simulation/risk model described earlier with the parameters
(such as, vessel arrival rates, overtake and pursuit distances, vessel entrance schedules and local traffic
density, among others.) reflecting the current situation in the Strait, based on year 2005-2006 data. The
risk profiles of this “base scenario” (in terms of average slice risks and average maximum instantaneous
risks), obtained using 25 replications - each of one year length, are displayed in Figure 11. The average
slice risk profile exhibits a steady behavior from the north entrance all the way down to the Bogazici
Bridge, where the local traffic congestion starts showing itself in this highly populated and busy region of
the Strait. Interaction of the transit and local traffic patterns generates a large spike in the average risk in
Slice 19 (i.e. the area of the Strait corresponding to downtown Istanbul and including the main harbor
region) and tapers off around the south entrance. The average maximum instantaneous risk profile also
32
exhibits a similar behavior varying from 80 to 380 fold differences between average risks and average
maximum instantaneous risks observed at various points along the Strait. This remarkable observation
indicates how risky the maritime traffic in the Strait of Istanbul can get at specific instances. That is,
depending on random realizations of accident causing factors, ordinary and safe appearance of the Strait
maritime activity could swiftly change into a very risky environment. For example, a rare realization
observed in Slice 1 involved an excessive level of fog during nighttime and two D-class vessels that just
entered the slice before the Strait is closed. Another rare realization, observed in Slice 19, involved an A-
vessel that was about to leave the Strait just after the night schedule started a D-vessel and an E-vessel
along with 10 local vessels. Such potentially highly dangerous situations may be rare, but a rare disaster is
a disaster too many. So, high risks indicated by the maximum risks should be taken seriously.
Figure 11. Current risk profiles of the Strait of Istanbul
33
It can be seen in Figure 12 how the southbound traffic which is mostly of the laden cargo, varies over time
and generates significantly higher instantaneous risks in slices 19 and 20.
Figure 12. 3D Diagram for all southbound vessel types through 24 days of the year
Using the aggregate model, we have performed an extensive scenario analysis to investigate the
characteristics of accident risks in the Strait under different settings and conditions. This analysis has
provided us with the ability to observe and predict how changes in various policies and practices impact
the risk profile of the Strait. Below we summarize our observations.
The accident risks in the Strait and the average vessel waiting times exhibit a tight balance. For instance,
a small increase in arrival rates may result in rather high waiting times at the entrances (an increase of
60% for some vessel classes). Furthermore, scheduling changes made to reduce vessel waiting times
increase risks in the Strait substantially. Conversely, one has to be very careful in revising the scheduling
Observation 1
Slice Risk
Slice
Time of Day
34
mechanism for the purpose of risk mitigation since the waiting times are highly sensitive to entrance rules.
The benefits obtained in risks may not justify the resultant waiting times. In the future, scheduling changes
may be justified, if significant reductions occur in the transit vessel traffic, perhaps due to alternative oil
transport modes such as pipelines and other routes. Thus, scheduling decisions to balance out delays vs.
risks should be made based on extensive experimentation with the model developed in this study.
The model indicates that pilots are of utmost importance for safe passage, and lack of sufficient pilotage
service significantly increases the risks in the Strait. Currently, vessels longer than 250 m. are mandated to
take a pilot, and it is voluntary for the rest. As a result of our experimentation, we have recommended
mandatory pilotage for vessels longer than 150 m. This will reduce the average risk by 7%, the average of
maximum risk by 11% in Slice 19 and the observed maximum instantaneous risk is 11114 observed in
Slice 3 (6763-fold of its average). Had pilotage been obligatory for vessels longer than 100 m., this would
reduce the average risks by 46 % and the average of maximum risks by 33 % at Slice 19.
Observation 2
Even though current regulations discourage overtaking anywhere in the Strait, results indicate that
overtaking a vessel is less riskier as opposed to requiring a pursuing faster vessel to slow down behind a
slower vessel, where the average slice risk and the average of maximum risk are increased by 28 % and 21
% in Slice 19, respectively. In the latter case, the maximum observed instantaneous risk is 23030 (12633-
fold of its average) observed in Slice 1. Therefore, in the regions where the geography of the Strait
tolerates it, overtaking seems to be a safe practice (as also suggested by the expert opinion).
Observation 3
The most significant contributor to risk appears to be the juxtaposition of the transit vessel traffic and the
local traffic. When the local traffic density in the Strait is decreased by 50% during daytime, it results in
an 83% decrease in the average risk and 31% decrease in the average maximum instantaneous risk of Slice
Observation 4
35
19. Accordingly, for potential risk mitigation, the scheduling procedure maybe revised to enable more
effective night-time traffic where there is almost no local traffic. The difference between risks involved in
day time and night time traffic is depicted in Figure 13 that is demonstrating risks of all southbound cargo
for 24 hours from midnight, averaged over a year. Clearly, this issue requires further research regarding
the kind of modifications that can be done to the scheduling practice to accommodate a larger volume of
night-time traffic, hopefully without increasing overall vessel delays or other risks.
Figure 13. Risks in Slice 19 of all southbound cargo over a year
The nature of the global economy and international politics dictate that the maritime transit traffic in the
Strait of Istanbul cannot be greatly reduced, nor eliminated. Nonetheless, the economic and political
realities versus environmental awareness and risk management need not to be mutually exclusive goals in
the Strait. The risks regarding the transit traffic can be mitigated by operational policies and rules that
adequately regulate and guide the transit traffic, while maintaining the freedom of passage. Until then, the
environment, the priceless historical/cultural heritage and the health and safety of the city’s residents will
be at jeopardy.
Slice Risk
Time of Day
36
In fact, the balance between economic and political realities and concerns for risks (leading to various
rules, regulations and requirements) are common in all constrained maritime environments (and especially
in narrow waterways). Furthermore, it would neither be unrealistic, nor difficult to project the main
observations and policy suggestions made for the Strait of Istanbul to other maritime environments in a
general qualitative sense. However, in order to make any quantitative assessment of the risks involved, a
similar extensive risk study should be performed in each case, for which this study can only act as a guide.
ACKNOWLEDGEMENT: Throughout this study, we have received sincere collaboration from the
Turkish Straits Vessel Traffic Services (VTS), Turkish Ministry of Transportation Directorate General of
Coastal Safety, Turkish Undersecretariat for Maritime Affairs, Turkish Maritime Pilots’ Association,
private industry, Istanbul Technical University Faculty of Maritime, Bogazici University Kandilli
Observatory and Earthquake Research Institute, Turkish Navy Office of Navigation, Hydrography and
Oceanography for which we are utmost thankful. We are also thankful to Prof. Johan René van Dorp of
George Washington University for his valuable suggestions in this study. This work is in part funded by
the Laboratory for Port Security at Rutgers University, NSF Grant Number INT-0423262, and TUBITAK,
The Scientific and Technological Research Council of Turkey through the Research Project 104Y207 and
BAP (Scientific Research Projects Fund of Bogazici University) through the Research Project 09A301D.
REFERENCES
1. VTS Users Guide. Turkish Straits Vessel Traffic Service General Management of Coastal Safety
and Salvage Administrations, 3rd edition, May 2004, Istanbul.
2. Convention Regarding the Regime of Straits. The American Journal of International Law. 1937;
31 (1): 1-18.
3. Turkish Maritime Pilots' Association. Some of the major casualties in the Strait of Istanbul during
the past years, 2004. Available at: http://www.turkishpilots.org.tr, Accessed on July 8, 2008.
4. Traffic Regulations for the Turkish Straits and the Marmara Region. Official Gazette, October 8,
1998, Issue 23515, Regulation No: 98/11860.
37
5. Birpınar ME, Talu GF, Su G, Gulbey M, Istanbul T. The Effect of Dense Maritime Traffic on the
Bosphorus Strait and Marmara Sea Pollution, 2005. Available at:
http://balwois.mpl.ird.fr/balwois/administration/full_paper/ffp-746.pdf, Accessed on July 8, 2008,
6. Ulusçu SO, Özbaş B, Altıok T Or I, Yılmaz T. Transit Vessel Scheduling in the Strait of Istanbul.
Journal of Navigation. 2009; 62 (1): 59-77.
7. Amrozowicz MD. Quantitative Risk of Oil Tanker Groundings. 1996, Master’s Thesis,
Department of Ocean Engineering. Massachusetts Institute of Technology, Cambridge, MA.
8. Amrozowicz MD, Brown A, Golay M. Probabilistic analysis of tanker groundings. 1997,
Proceedings of the 7th International Offshore and Polar Engineering Conference, Honolulu, HI:
313-320.
9. Dougligeris C, Iakovou E, Yudhbir L. Maritime route risk analysis for hazardous materials
transportation, 1997, Proceedings of the 8th IFAC/IFIP/IFORS, Transportation Systems, Chania,
Greece; 2: 563-568.
10. Iakovou ET. An interactive multi-objective model for the strategic maritime transportation of
petroleum products: Risk analysis and routing. Safety Science. 2001; 39 (1-2): 19-29.
11. Slob W. Determination of risks on inland waterway. Journal of Hazardous Materials. 1998; 61 (1-
3): 363-370.
12. Harrald J, Mazzuchi T, Merrick J, van Dorp R, Spahn J. Using system simulation to model the
impact of human error in a maritime system. Safety Sci. 1998; 30 (1-2): 235-247.
13. Merrick JRW, Van Dorp JR, Harrald J, Mazzuchi T, Spahn JE, Grabowski M. A systems
approach to managing oil transportation risk in Prince William Sound. Systems Engineering.
2000; 3 (3): 128-142.
14. Merrick JRW, Van Dorp JR, Mazzuchi T, Harrald JR, Spahn JE, Grabowski M. The Prince
William Sound risk assessment. Interfaces. 2002; 32 (6): 25-40.
38
15. Merrick JRW, Van Dorp JR, Mazzuchi TA, Harrald JR. Modeling risk in the dynamic
environment of maritime transportation. 2001, Proceedings of the 2001 Winter Simulation
Conference, Arlington, VA; 2: 1090-1098.
16. Van Dorp JR, Merrick JRW, Harrald JR, Mazzuchi TA, Grabowski M. A risk management
procedure for the Washington State ferries. Risk Analysis. 2001; 21 (1): 127-142.
17. Merrick JRW, Van Dorp R. Speaking the truth in maritime risk assessment. Risk Analysis. 2006;
26 (1): 223-237.
18. Kuroda K, Kita H, Kono S. Mathematical Model of Ship Collision Probability. Memoirs of the
Faculty of Engineering, Kyoto University. 1982; 44 (1): 135-157.
19. Kaneko F. Methods for probabilistic safety assessments of ships. Journal of Marine Science and
Technology. 2002; 7 (1): 1-16.
20. Yip TL. Port traffic risks--A study of accidents in Hong Kong waters. Transportation Research
Part E, 2008; 44 (5): 921—931.
21. Maio D, Ricci R, Rossetti M, Schwenk J, Liu T. Port Needs Study. Vol. 1, Report No. DOT-CG-
N-01-91-1.2, USDOT/RSPA/VolpeTSC, U.S. Department of Transportation, Washington, DC,
1991.
22. Roeleven D, Kok M, Stipdonk HL, De Vries WA. Inland waterway transport: modelling the
probability of accidents. Safety Science. 1995; 19 (2-3): 191-202.
23. Anderson E, Talley WK. The oil spill size of tanker and barge accidents: Determinants and policy
implications. Land Economy. 1995; 71: 216-228.
24. Talley WK. Determinants of cargo damage risk and severity: The case of containership accidents.
Logistics and Transportation Review. 1996; 32: 377-388.
25. Psaraftis HN, Panagakos G, Desypris N, Ventikos N. Analysis of maritime transportation risk
factors. 1998, Proceedings of the 8th International Offshore and Polar Engineering Conference,
Montreal, Canada; 4: 477-483.
39
26. Le Blanc LA, Rucks CT. Multiple discriminant analysis of vessel accidents. Accident Analysis
and Prevention. 1996; 28 (4): 501-510.
27. Kite-Powell HL, Patrikalakis NM, Jin D, Abrams SL, Jebsen J, Papakonstantinou V, Lin SC.
Formulation of a Model for Ship Transit Risk. Final Project Report, Massachusetts Institute of
Technology, Sea Grant Coll. Program, Cambridge, MA, 1998.
28. Yudhbir L, Iakovou E. A maritime oil spill risk assessment model. 2005, Proceedings of
International Oil Spill Conference, Miami Beach, FL: 4608-4613.
29. Moller TH, Molloy FC, Thomas HM. Oil spill risks and the state of preparedness in the Regional
Seas. 2005, Proceedings of International Oil Spill Conference Miami Beach, FL: 9266-9269.
30. Talley WK. Maritime Safety, Security And Piracy. Informa Law, 2008.
31. Kornhauser AL, Clark WA. Quantitative forecast of vessel casualties resulting from additional oil
tanker traffic through the Bosphorus. Report; ALK Associates, Princeton, NJ, 1995.
32. Otay EN, Özkan S. Stochastic Prediction of Maritime Accidents in the Strait of Istanbul. 2003,
Proceedings of the 3rd International Conference on Oil Spills in the Mediterranean and Black Sea
Regions: 92-104.
33. Tan B, Otay EN. Modeling and analysis of vessel casualties resulting from tanker traffic through
narrow waterways. Naval Research Logistics. 1999; 46 (8): 871-892.
34. Tan B, Otay E. A stochastic model of vessel casualties resulting from oil tanker traffic through
narrow waterways. 1998, Proceedings of 12th European Simulation Multiconference, Manchester,
UK: 881-885.
35. Or I, Kahraman I. A simulation study of the accident risk in the Istanbul Channel. International
Journal of Emergency Management. 2003; 1 (2): 110-124.
36. Akten N. Analysis of shipping casualties in the Bosphorus. The Journal of Navigation. 2004;
57(3): 345—356.
40
37. Maritime International Secretariat Services Ltd, Shipping Industry Flag State Performance Table,
2006. Available at: http://www.marisec.org/flag-performance/ FlagStatePerformanceTable06.pdf,
Accessed on May 3, 2008.
38. Kandilli Observatory and Earthquake Research Institute, 2005. Official website:
http://www.koeri.boun.edu.tr.
39. Weather Underground, International Weather Information Website, 2005. Official website:
www.weatherunderground.com.
40. The Department of Navigation, and Oceanography of the Turkish Navy, 2005. Official website:
http://www.shodb.gov.tr/english.html
41. Ece J N. İstanbul Boğazı Deniz Kazaları Analizi. DEKAŞ Kültür Yayınları, 2007.
42. Cooke RM. Experts in Uncertainty: Opinion and Subjective Probability in Science. Oxford
University Press, Oxford, U.K., 1991.
43. Szwed P, Van Dorp JR, Merrick JRW, Mazzuchi TA, Singh A. A Bayesian Paired Comparison
Approach for Relative Accident Probability Assessment with Covariate Information. European
Journal of Operations Research. 2006; 169 (1): 157-177.
44. Cox DR. Regression Models and Life Tables (with Discussion). Journal of the Royal Statistical
Society, Series B, 1972; 34:187-220.