757
1 INTRODUCTION
Accidents to navigating vessels have the potential to
result in loss of life, environmental pollution and
economic losses. To better understand where and why
these accidents occur, a significant body of literature
that might be described as maritime risk analysis has
developed, particularly within the last decade [23].
These methods seek to apply quantitative methods to
measure the likelihood or consequence of hazards
such as collisions, groundings and allisions from
occurring. These risks have an inherent spatial
component and mapping where these risks are
highest can help decision makers in both allocation of
risk control measures and marine spatial planning.
The multitude and variety of these methods is
significant, with reviews undertaken by several
authors [4, 21–23, 28]. A key challenge when
comparing these methods is that each have their own
assumptions and limitations that could introduce
biases and therefore it would be unreasonable to
assume that any one method works better than others
in all situations. Whilst this makes evaluation between
established models difficult, it also makes judgements
on the suitability and validity of novel methods
similarly challenging.
One such novel approach is that of the use of
machine learning (ML) techniques for maritime risk
assessment. ML might be described as a subset of
artificial intelligence whereby algorithms improve
through experience rather than being explicitly
programmed. These models can be supervised,
whereby the model is constructed on data containing
both input and outputs, or unsupervised, whereby
structure is sought on unlabelled data. Few have
sought to apply ML to ship navigation [9, 36] and
even fewer have attempted to use these methods to
assess navigation safety [5, 19]. Whilst some have
From Conventional to Machine Learning Methods for
Maritime Risk Assessment
A. Rawson
1
, M. Brito
1
, Z. Sabeur
2
& L. Tran-Thanh
3
1
University of Southampton, Southampton, UK
2
Bournemouth University, Bournemouth, UK
3
University of Warwick, Warwick, UK
ABSTRACT: Within the last thirty years, the range and complexity of methodologies proposed to assess
maritime risk have increased significantly. Techniques such as expert judgement, incident analysis, geometric
models, domain analysis and Bayesian Networks amongst many others have become dominant within both the
literature and industry. On top of this, advances in machine learning algorithms and big data have opened
opportunities for new methods which might overcome some limitations of conventional approaches. Yet,
determining the suitability or validity of one technique over another is challenging as it requires a systematic
multicriteria approach to compare the inputs, assumptions, methodologies and results of each method. Within
this paper, such an approach is proposed and tested within an isolated waterway in order to justify the
proposed advantages of a machine learning approach to maritime risk assessment and should serve as
inspiration for future work.
http://www.transnav.eu
the International Journal
on Marine Navigation
and Safety of Sea Transportation
Volume 15
Number 4
December 2021
DOI: 10.12716/1001.15.04.06
758
argued that such methods have numerous advantages
over traditional risk assessment techniques [15], such
benefits have not been demonstrated within the
maritime domain.
A possible solution to this challenge is through a
systematic evaluation of different methods within a
common framework, using set criteria against which
each method can be judged. In [28], a comparison is
made between 20 models against standardised
criteria, but without implementing these techniques,
evaluation is more challenging, particularly where the
models are proprietary.
This paper sets out to develop such a framework
by comparing 10 different maritime risk analysis
methods and identifying suitable criteria that could be
used for an evaluation. We make a number of
contributions; firstly, we conduct a systematic and
applied evaluation of a selection of the most widely
researched maritime risk models, in order to highlight
their methodological strengths and weaknesses.
Secondly, we introduce four ML techniques and how
they could be utilised for predicting the likelihood of
accidents, with some high-level implementations and
a discussion of opportunities for greater application of
these techniques. Thirdly, we propose a list of criteria
through which these methods can be directly
compared, proposing further work for a multi-criteria
evaluation of maritime risk models. Whilst the
evaluation requires further work, we make a number
of observations on the different techniques to provide
initial feedback on the capability of ML for maritime
risk assessment.
1.1 Case Study
To achieve these aims, we utilise a case study of the
waterway between Washington State (United States),
Vancouver Island (Canada) and British Columbia
(Canada). This waterway is known as the Puget
Sound or Salish Sea, and extends from the Pacific
Ocean, through the Strait of Juan de Fuca, before
heading north through the San Juan Islands towards
Vancouver, or south through Admiralty Inlet towards
Seattle (Figure 1). This area is notable for several
reasons. Firstly, it has a significant volume of traffic,
of all types, including cargo and tanker traffic bound
for various ports and terminals, significant
recreational and fishing fleets, and major ferry routes.
Secondly, traffic within the area is managed by Traffic
Separation Schemes (TSS), pilotage districts, escort
towage and a cooperative VTS between the United
States and Canada. Thirdly, the area has been
extensively studied in other maritime risk studies,
most notably Vessel Traffic Risk Assessment (VTRA)
[39].
Vessel traffic data from the Automatic
Identification System (AIS) was obtained from the
MarineCadastre for June 2018 covering the waterway.
AIS is an automatic ship reporting system that
transmits dynamic (positional, speed and course) and
static (ship type and size) information that can be
collected to produce high spatial-temporal resolution
datasets. Furthermore, incident data was available
from the US and Canadian Coastguards for the years
2002-2014.
Figure 1. Study Area with TSS overlaid.
2 CONVENTIONAL METHODS
Six broad conventional maritime risk analysis
methods were identified from the literature and are
discussed below.
2.1 Risk Matrices and Expert Judgement
At an operational level, most decisions on maritime
safety are made using risk matrices. Such an approach
is also recommended for the screening stage of the
Formal Safety Assessment [18]. A list of hazards are
identified and a group of experts or stakeholders
score the likelihood and consequence against set
criteria to produce a risk score. Within the study area,
we might score three hazards as Table 1, noting that
the navigational complexity of the waterway is such
that groundings are more likely, but would have
lower consequences than collisions.
Table 1. Simple hazard table with 5x5 Matrix
_______________________________________________
ID Hazard Likelihood Consequence Risk
(1-5) (1-5) (1-25)
_______________________________________________
1 Collision 3 3 9
2 Grounding 4 2 8
3 Allision 3 2 6
_______________________________________________
Such a method enables the inclusion of non-
modelled issues [2] and may be suitable in situations
where there is little quantitative data. However, such
an approach has received significant criticism
regarding the limitations and bias of expert prediction
[37, 38] or the inherent properties of the matrices [17].
Further, only a single score is provided per hazard
and therefore does not reflect the distribution of risk
across the study area. As such, it is a highly simplistic
method of risk assessment, but a useful means of risk
evaluation [28].
759
2.2 Incident Rates and Analysis
By comparing the number of accidents against some
unit of exposure (such as time or distance), the
relative risk of incidents between locations and
situations can be compared [3]. Figure 2 compares the
number of accidents against the number of hours of
exposure at 1nm resolution in the study area. A key
challenge is the relative sparsity of accident data that
prevents high-resolution output with most locations
having zero incidents, which might be incorrectly
interpreted as zero risk. Other criticisms include the
under-reporting of accidents [14] and the assumed
static relationship between accidents and traffic [30].
One method to overcome this is to calculate an
accident rate and use the statistical relationship
between accidents and traffic to estimate the accident
rate (Figure 2). Whilst this increases coverage of the
risk map, it loses the influence of spatial factors that
might elevate risk in certain locations, becoming
highly sensitive to traffic volume.
Figure 2. Calculated Incident Rates (top) and estimated
incident rate (bottom).
2.3 Weighted Overlay Analysis
A further method to include the influence of other
spatial factors is through a weighted overlay model.
This approach can be summarised that risk is the
product of the scores of likelihood factors (L) and
their weightings (w) with the scores of consequence
factors (C) and their respective weightings.
NL NC
LC
i i i i
ii
Risk w L w C=

(1)
In Figure 3, grounding risk is estimated using a set
of scoring criteria and weightings for traffic volume,
proximity to shore, proximity to traffic lanes and
ports. Higher risk areas are shown to the east with
more complex navigation around the San Juan islands
than in the middle of the Strait. Whilst such an
approach enables inclusion of other risk factors and
the production of high spatial resolution risk maps,
the choice of weightings and factors are to some
extent arbitrary, subjective and lacking in treatment of
uncertainties.
Figure 3. Weighted Overlay Analysis.
2.4 Geometric Method
A geometric method is one that aggregates vessel
traffic into routes, with known distributions and
frequencies, before performing mathematical
functions to calculate accident candidates. Whilst
variations exist [22, 27] the work of [29] has been
particularly influential and has been adopted by
IALA’s IWRAP Risk Modelling Tool [8]. An IWRAP
model was developed for the study area with the
traffic legs representing the major routes and the
shoreline inputted as a grounding hazard. From this,
the risk of collision and grounding can be calculated
by vessel type and location (Figure 4).
Figure 4. IWRAP Model Results.
Geometric methods have been widely discussed in
the literature and have attracted numerous criticisms.
Firstly, aggregation undermines the individual
760
behaviours of vessels, particularly where vessel
transits are non-linear, such as in a pilot boarding
area. Secondly, movements are averaged over 24
hours and don’t reflect variations in risk throughout
the day, such as where tidal heights dictate channel
access, or recreational and fishing activities are
diurnal. Furthermore, where the data collection is
limited to short periods of time or conducted in
quieter places, establishing representative traffic
distributions may not be possible [24]. Fourthly, the
method omits some hazard types such as drifting
vessel collisions [1]. Fifthly, the choice of legs and
other input parameters is subjective and depending
on the expertise of the analyst [28]. Sixthly, the results
are highly sensitive to the causation probability [27]
which might be chosen with little evidential basis.
Finally, some have questioned the underlying
assumption that risk is directly related to traffic flow
[25, 30].
2.5 Domain Analysis
Ship domain models construct a region of safe water
that a master wants to keep clear of other vessels or
fixed objects [12], and by measuring the frequency
and types of encounters between those vessel
domains, an indicative measure of collision risk is
provided [6]. Whilst a multitude of domain designs
have been proposed [35], we implement the model
proposed by [40] that is dynamic given vessel size and
speed. Figure 5 shows the frequency of domain
encounters across the study area, with the majority
clustered in the key ports and harbours of Victoria,
Port Angeles and Anacortes, rather than in the Straits
to the west.
Figure 5. Domain Analysis Encounters.
Whilst the inclusion of a temporal dimension
within domain models overcome some limitations of
static methods, they have also received criticism.
Some have questioned both the reliability and validity
of such methods [11] and also the statistical
relationship between encounters and collision
frequency [31]. Furthermore, the interactions between
vessels have been shown to vary by many other
factors such as waterway geometry [16] that are not
captured in existing domain models. Finally, domain
models are often limited to one-vs-one collision
situations, omitting the influence of other nearby
vessels on developing collision situations [6].
2.6 Bayesian Networks
Bayesian Networks are a technique for graphically
representing a joint probability distribution of a
selected set of variables and many have argued that
they are well suited to maritime risk analysis [13, 41].
In particular, they enable the inclusion of expert
judgment, a far greater number of conditions that
cannot be easily quantified and can be employed in
situations where there is little historical data, such as
autonomous vessels or Arctic shipping. Furthermore,
the impact of risk controls can be tested by interfering
with specific elements within the model [26]. Finally,
uncertainties can be reflected within the model. Figure
6 shows a highly simplified Bayesian Network for
predicting grounding risk, compared to others
proposed in the literature [26].
Figure 6. Simplified Bayesian Network.
Bayesian Networks are not without their criticisms
[16, 41]. Firstly, the lack of data makes determining
priors challenging. Secondly, the reliance on expert
judgment might introduce biases [37, 38] and limits
the frequency and scope of such assessments, given
the time commitments necessary on experts and
stakeholders.
2.7 Summary of Conventional Methods
The six methods described above are broad but
encapsulate a significant portion of the academic and
industry techniques used for maritime risk
assessment. In general, the different methods
consistently identify higher risk to the east of the
study area in the constrained and busy waters of the
San Juan Islands. Whilst we have provided a high-
level introduction to each method and identified some
key criticisms in each case, conventional methods of
maritime risk assessment have more fundamental
challenges. At a workshop of EMSA [7], it was noted
that existing methods used by risk assessment projects
had a high cost, used proprietary models of
consultants, were time-consuming and often failed to
communicate their uncertainties. Given these
criticisms, it may be that alternative methods may be
more suitable for maritime risk assessment.
3 MACHINE LEARNING METHODS
ML consists of numerous approaches and
applications, and here we discuss how three broad
761
categories could be applied to maritime risk
assessment: namely supervised ML, unsupervised ML
and deep learning techniques.
3.1 Supervised Machine Learning
Predicting whether an accident will occur in a certain
location or time extent can be framed as a form of
supervised ML. We seek the function y=f(x) that
describes the relationship between a response variable
y, in this case accident occurrence, with a set of
explanatory factors x, such as traffic volume, depth or
ship size as vectors in the form (x1,x2,…,xn). In order to
learn this relationship, a dataset is split into a training
dataset, through which the model is developed, and
then tested on a set-aside dataset. The accuracy of
such models can be gauged through numerous
metrics, such as classification accuracy (correct
predictions to incorrect predictions) or prediction
value error in the case of regression.
Within the literature, there are relatively few
applications of ML for maritime risk assessment [5,
19]. Where these methods have been used, it is more
common for the training dataset to consist of a static
list of vessels [19] or port state control inspection
outcomes [33]. In such a context, a model is developed
that seeks to predict whether a vessel has an accident
given the characteristics of the vessel, such as age, flag
state or type. However, there is much greater scope to
advance these methods for spatial or real-time
maritime risk assessment using vessel traffic data and
spatial-temporal risk factors.
Figure 7. Maritime risk ML framework.
Figure 7 shows a framework through which such
methods could be applied to predict vessel traffic risk.
In stage 1, AIS, incident and other exploratory
datasets are combined as input features and labels. In
stage 2, the dataset is either labelled as positive or
negative (accident occurred or didn’t occur) as a
classification problem, or the accident frequency is
calculated as a regression problem by aggregating the
datasets. In stage 3, the data is split into a training and
testing dataset, with the ML model developed on the
former and tested on the latter, before the trained
model is deployed as a risk analysis tool.
In this example, the Random Forest ML algorithm
is used which is an ensemble of decision trees using
subsets of the training data, with the final prediction
the average prediction of the individual trees. In the
first case, the study area was subdivided into grid
cells and the traffic volume and depth of water in each
cell used as input features and the number of
groundings used as the label. The dataset is split into
a training and testing dataset with the ratio of 80% to
20%, which once trained, achieved an R2 of 0.58 and a
Mean Squared Error of 0.004 on the testing set. Figure
8 shows the results, highlighting the waters around
the San Juan Islands to the east, where traffic is
concentrated into small and shallow channels,
locations where most grounding have historically
occurred.
Secondly, the same framework is applied but the
individual vessel traffic positions and historical
groundings are used as the positive and negative
classes respectively. By using features such as vessel
size (length and draught), depth of water, distance
from shore and vessel traffic density, the probability
of ship accident can be predicted. In this case, an
accuracy of 98.8% is achieved on the test dataset, with
15 of the 26 groundings and 427,054 of 432,154 non-
groundings correctly predicted. This indicates a high
recall (0.58) but a low precision (0.003), suggesting the
model can differentiate groundings, but at the
expense of a number of false positives. Figure 9 shows
the predicted transit risk for an Aframax tanker
approaching Anacortes, with the risk of grounding
being predicted to increase as it passes through the
Guemes Channel.
Figure 8: Area Grounding probability using Random Forest.
In this work, only a limited number of features are
utilised but there is scope to improve the predictive
capability through inclusion of weather, waterway
geometry, ship characteristics and a plethora of other
risk factors. However, the results show good
potential; firstly, the results exceed the predictive
power that depth or traffic volume alone could have