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

net/publication/282589911

Inferring spatial distribution of oil spill risks from proxies: Case study in the
north of the Persian Gulf

Article in Ocean & Coastal Management · November 2015

DOI: 10.1016/j.ocecoaman.2015.08.017

CITATIONS READS

29 4,981

7 authors, including:

Sahar Mokhtari Seyed Mohsen Hosseini

Tarbiat Modares University Tarbiat Modares University
3 PUBLICATIONS 30 CITATIONS 241 PUBLICATIONS 3,084 CITATIONS

SEE PROFILE SEE PROFILE

Afshin Danehkar Masoud Torabi Azad

University of Tehran Islamic Azad University North Tehran Branch
143 PUBLICATIONS 685 CITATIONS 79 PUBLICATIONS 361 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Babak Naimi on 24 May 2019.

The user has requested enhancement of the downloaded file.

 Ocean & Coastal Management 116 (2015) 504e511

Contents lists available at ScienceDirect

Ocean & Coastal Management

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

Inferring spatial distribution of oil spill risks from proxies: Case study
in the north of the Persian Gulf
Sahar Mokhtari a, Seyed Mohsen Hosseini b, *, Afshin Danehkar c, Masoud Torabi Azad d,
Ji
rí Kadlec e, Ari Jolma f, Babak Naimi g
a
Department of Environment, Faculty of Natural Resources, Tarbiat Modares University, Emam Reza Ave., Noor, Mazandaran, Iran
b
Department of E Forestry, Faculty of Natural Resources, Tarbiat Modares University, Emam Reza Ave., Noor, Mazandaran, Iran
c
Department of Environmental Sciences, Faculty of Natural Resources, College of Agriculture & Natural Resources, University of Tehran, Karaj, Iran
d
Department of Physical Oceanography, IAU, North Tehran Branch, Tehran, Iran
e
Department of Civil and Environmental Engineering, Brigham Young University, Provo, USA
f
Department of Civil and Environmental Engineering, Aalto University, Finland
g
Department of Environment and Energy, Science and Research Branch, Islamic Azad University, Hesarak, Pounak, Tehran, Iran

a r t i c l e i n f o a b s t r a c t

Article history: Maps of oil spill probability are important tools in environmental risk assessment and decision making in
Received 7 May 2015 coastal zone management. This paper describes the development of a spatial predictive model for the
Received in revised form probability of oil spills in the northern part of the Persian Gulf. The model estimates the probability of oil
20 August 2015
spills at a pixel level as function of four proxies, i.e., ship routes, coastlines, oil facilities, oil wells. It uses a
Accepted 29 August 2015
Available online xxx
generalized linear model (GLM) with a polynomial function that is implemented in the R software
environment. For training the model, we used reported oil spill events that represent the location of their
occurrences. We trained and tested the model in 100 iterations, using a different subset of data for each.
Keywords:
Persian Gulf
The model evaluations showed mean accuracy of 0.79% (range 0.68%e0.89%), expressed by the area
Oil spill under the curve (AUC). In the northern part of the Persian Gulf, the largest probability of oil spills was
Probabilistic model predicted in areas where actual oil facilities in combination with high intensity ship traffic are in evi-
GLM dence. The model can predict the probability of oil spills as raster map in a standard R data format. It can
be used in environmental risk assessment as well as an input for more detailed oil spill simulation
models. The advantages of the model include its’ high spatial resolution, accounting for uncertainty in oil
spill locations, and the possibility of sharing as an open-source R script with other users.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction as a combination of an event, its consequences, and the uncertainty

related to both (Aven, 2010; Burgman, 2005). An environmental
Oil spills are a serious environmental problem with critical risk assessment must have the capacity to support decision making
impacts on marine and coastal habitats, wildlife, fisheries and hu- for preventing future oil spill disasters as well as to develop optimal
man activities (Allan et al., 2012; Fattal et al., 2010; Carson et al., planning and management strategies in response to oil spill events.
1992). Coastal areas provide important resources for livelihood This requires identifying and prioritizing areas in the coastal zone
activities, fisheries, transport industries and recreation (Vafai et al., where there is a need for an allocation of resources, while mini-
2013; Singkran, 2013). At the same time, these areas are at risk of oil mizing the risks and costs (Keeney, 1982; Mata and Corchado,
spills caused by vessel navigation, oil transport via cargo ship or 2009). For large marine areas, oil exploration activities are not
submerged pipelines and by oil exploration activities (Jolma et al., uniform in space, therefore, a spatially distributed oil spill risk
2014; Ng and Song, 2010; Kankara and Subramanian, 2007; assessment is required (Burgman, 2005).
Bigano and Sheehan, 2006). Risk of oil spill is commonly defined Many studies develop methods and tools for identifying and
mapping potential oil spill sites and supporting oil spill risk man-
agement (Ko and Chang, 2010; Lehikoinen et al., 2012; Jolma et al.,
2014; Juntunen et al., 2005). A very large number of studies have
* Corresponding author.
been conducted to model and predict the impact of major oil spill
E-mail address: [email protected] (S.M. Hosseini).

http://dx.doi.org/10.1016/j.ocecoaman.2015.08.017
0964-5691/© 2015 Elsevier Ltd. All rights reserved.
 S. Mokhtari et al. / Ocean & Coastal Management 116 (2015) 504e511 505

accidents, for example of the Exon Valdez accident (Carson et al., activities and heavy ship traffic are in evidence. Furthermore, the
2003; Wiens and Parker, 1995), BP Gulf of Mexico oil spill International Marine Organization (IMO) has declared the Persian
(Sammarco et al., 2013; Allan et al., 2012), Prestige accident Gulf as Special Area, requiring the adoption of special mandatory
(Bernabeu et al., 2013; Díaz et al., 2008), Braer oil spill (Spaulding methods for the prevention of sea pollution by oil and/or garbage
et al., 1994), and Gulf War (Bejarano and Michel, 2010; Alam, (Mahmoudi, 1997).
1993). More than a thousand research articles and studies can be
found about these catastrophic oil spills. However, the small, but 2. Materials and methods
more frequently occurring events of routine shipping operations
also affect the coastal and marine environment, and should also be 2.1. Study area
included in the environmental risk assessment (Ng and Song,
2010). Several authors attempted to detect the spatial distribution The Persian Gulf, a shallow and semi-enclosed basin in a typi-
of oil spills by remote sensing observations (Fingas and Brown, cally arid climate, is located in the southwest of the Asian continent
2014; O'Hara et al., 2013; Serra-Sogas et al., 2008) or calculate the between 48
and 57
E longitude 24
e30
N latitude (Fig. 1). It is
probability of ship collision accidents (Bigano and Sheehan, 2006; connected to the deep Sea of Oman through the narrow Strait of
van Dorp and Merrick, 2011). Density of the events was also used Hormuz. The Persian Gulf covers an area of approximately
to visualize the oil spill incident frequency as a map (Lu, 2003). 226,000 km2, more than 990 km in length, its width varying be-
Most of these studies relied only on the reported oil spill events, but tween 56 km (Strait of Hormuz) to 338 km. The basin has an
exploring the causal relationship is also important in order to un- average depth of about 35 m, the deepest water approximately at
derstand and quantify what and to what extent different factors 107 m (Ka €mpf and Sadrinasab, 2005). The circulation in the Persian
influence the risk of oil spills. Relying only on the reported events Gulf is driven by wind-stress, surface buoyancy fluxes, fresh water
may not fully represent the potential risk, as it is likely that some runoff, water exchange through the Strait of Hormuz, and tides
events are not reported, or their measurements are subject to error (Hassanzadeh et al., 2011; Pous et al., 2015). In this study, we
and uncertainty (Fingas and Brown, 2014; O'Hara et al., 2013). focused on the northern part of the Persian Gulf defined approxi-
Exploring the association between the factors that may cause an mately by the Exclusive Economic Zone (EEZ) of I.R.Iran, and based
event (causal factors) and the data on oil spills can lead to con- on the availability of data on oil spill events.
struction of a model that can be used to predict the spatial distri- By volume of oil transit, the Strait of Hormuz is one of the
bution of potential risks of an oil spill over the whole study area. If a world's most strategic chokepoints. The volume of crude oil and
significant relationship between the factors and the events is found, petroleum products transported through the Strait was 15.7e17.5
the factors (explanatory variables) can be used as proxies to predict million barrels per day during 2007e2011 (Rahmanpoor et al.,
the spatial distribution of the events. Such methods have been used 2014). A combination of factors, including tanker traffic, acci-
in many disciplines where observations are scarce, for example, in dental spills, industrialization, port areas and refineries may
species distribution (Naimi et al., 2014) or air pollution source directly or indirectly affects increasing of the pollution risk in the
modeling (Hoek et al., 2008). The spatial distribution of the prob- marine environment of the Persian Gulf (Sheppard et al., 2010).
abilities for an oil spill occurrence has been generated based on During the year 2012, most of the oil pollution were from the oil-
simulation of oil spill trajectories and used in an oil spill risk field leakages with tremendous amounts of heavy oil. It has been
analysis (Price et al., 2003). The importance of including causal noticed that most of the oil spills in the northern and southern
factors (e.g., ship route proximity) in mapping operational oil dis- parts of the Persian Gulf is from the oilfields and the shipping
charges has been highlighted by Serra-Sogas et al. (2014). Liu et al. traffics, respectively. Besides, a very serious incident took place in
(2015) modeled the spatial risk of oil spill in China's Bohai Sea using the middle part of the Persian Gulf (MEMAC, 2012). Although the
the ships and platforms as potential oil spill locations. Locations frequency of big oil spill incidents has decreased, the chronic oil
with high probability of oil spill in the Gulf of Finland were iden- spills from ballast water washing and from oil platform activities
tified by Jolma et al. (2014) using the Bayesian network method. remain widespread that make a critical concern in integrated
Moreover, Hong et al. (2010) developed a model for estimating the coastal zone management (Ebrahimi-Sirizi and Riyahi-Bakhtiyari,
probability of oil spill in the Luoyuan Bay in China using the ship 2013; Mirvakili et al., 2013).
traffic intensity as an explanatory variable. Singkran (2013) iden-
tified zones in the coastal waters of Thailand by classifying impacts 2.2. Modeling oil spill risk
of oil spills using explanatory variables such as the number of ports
and boats. 2.2.1. Oil spill events
Although most of the above mentioned studies have focused on The oil spill events in the Persian Gulf were reported by MEMAC
modeling the fate of oil spills either for simulating a big event after (Marine Emergency Mutual Aid Center) in the time period
an accident, or for assessing the oil spill vulnerability in coastal 2010e2013 (near 250 events). Many of the reported events are
areas, there is still a need to predict spatial distribution of such based on satellite observations and some are local reports of the
events and the probability of potential future events. Moreover, northern part of the Persian Gulf. In this research, we considered
exploring and mapping the spatial distribution of potential oil spills the following information: latitude, longitude, and occurrence of
is the first step in locating priority areas at high risk of oil spill (Liu events. To be able to fit the model, our response variable needs to
et al., 2015). In the other words, we need to know the locations follow a binomial distribution (i.e., data on locations where the oil
where oil spill events are most likely to occur. Defining the spill events have occurred as well as those where the events have
vulnerability of the marine environment (Garcia et al., 2013) can not occurred). However, our data include only the occurrence
lead to the locations that are at high risk of oil spills, i.e. those with (presence) of the events. A common approach to overcome this
high probability and vulnerability. The main objective of this study problem is to generate the absence sites, called pseudo-absences
was to investigate the factors that may explain the occurrence of oil (Hengl et al., 2009; Chefaoui and Lobo, 2008). We selected the
spills. By linking the causal factors with the oil spill events, we number of pseudo-absences equal to the number of presence lo-
developed a spatial predictive model to generate a probability map cations, as recommended by Hengl et al. (2009). The pseudo-
of oil spill risks in the Persian Gulf. Our study concerns a high-risk absence locations were generated as random spatial samples of
area for oil spills, due to its location where intense oil exploration the study area. Based on the random choice of the locations, many
506 S. Mokhtari et al. / Ocean & Coastal Management 116 (2015) 504e511

Fig. 1. Map of the Persian Gulf with the Iran EEZ boundary.

of the points happened to fall near the presence data, while there of the traffic density. All pixels with high traffic density were
were some areas with no presence or pseudo-absence. We draw the marked as ship route. Finally, the distances of every pixel to the
pseudo-absences from the areas that were environmentally and edge of the nearest ship route were calculated as:
geographically far from the presence locations to avoid overlapping rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
with the presence sites. The presence and pseudo-absence points  2  2
proximityi ¼ xi  xp þ yi  yp
were combined into one collection for the modeling step.

where xi, yi are the coordinates of the center of each pixel and xp, yp
2.2.2. Explanatory variables the coordinates of the nearest ship route. The raster datasets then
As pre-requisite for the model, we prepared four raster datasets were masked into the boundary of Persian Gulf EEZ area (Fig. 2).
containing the geographical distance to variables including nearest The distance value for each cell varies by a high order of
oil facility, oil well, coastline, and ship route. The distance was magnitude. Therefore, we transformed the data by calculating an
calculated from the center of each raster cell in the study area. We inverse normalized distance, defined as:
chose a raster cell size equal to 500 m in a metric coordinate system
(UTM zone 39) as a compromise between map detail and compu- di
dni ¼ 1 
tation speed. Our goal was to use these raster datasets as proxies for maxðdÞ
predicting the spatial distribution of oil spill events. In the original
data, the oil facilities, the oil wells and the coastlines were repre- where dni is the inverse normalized distance at pixel i, d the dis-
sented as polygons, points and lines, respectively. The ship routes tance at pixel i and max (d) is the maximum distance within the
were in the form of raster dataset that representing the tanker raster.
traffic density, and obtained from the Automatic Identification
System (AIS; www.marinetraffic.com). Because the original map of
traffic density had no legend, we classified the values of tanker 2.2.3. Model creation
traffic density at each cell (pixel) of the raster dataset. A cell clas- To explore the relationship between the occurrence of oil spill
sified as high when the density was equal or greater than two times events and the potential causal factors, the generalized linear
of standard deviations above the mean density, defined as: model (GLM; Nelder and Baker, 1972) was used. Using GLM, a
multiple regression model was fitted and used for predicting the
Di;j > D þ 2sD spatial distribution of the oil spill probabilities in the Persian Gulf.
The GLM method supports a binary response variable (present or
where Di,j is the density at a given pixel and the standard deviation absent) and with one or more independent variables. It has been
 S. Mokhtari et al. / Ocean & Coastal Management 116 (2015) 504e511 507

Fig. 2. Explanatory variables used to explain the oil spill events: distance to oil wells (A); distance to coastlines (B); distance to ship routes (C); and distance to oil facilities (D).

widely used in species distribution modeling (Engler et al., 2004; Development Core Team, 2014), an open source programming
Hengl et al., 2009). language for data analysis and visualization (Jolma et al., 2008) was
To fit the model, the occurrence of oil spill events was used as used. We also used the raster package in R (Hijmans and van Etten,
the response variable and the distance to oil facilities, oil wells, 2012) for the spatial analysis and visualization.
coastlines and ship routes were used as covariates (explanatory The GLM model is defined as:
variables). We chose these four variables for the following reasons:
according to the MEMAC 2012 report of the Persian Gulf, the oil EðYÞ ¼ g1 ðXbÞ
spills were associated with oil exploration activities (oil facilities,
oil wells); studies from other areas demonstrated that also the in- where Y is the response variable, E(Y) the expected values of the Y
tensity of ship traffic (Singkran, 2013; Serra-Sogas et al., 2014), assumed to be from binomial distribution, The Xb is the linear
presence of oil platforms (Liu et al., 2015) and coastline (Singkran, predictor which is a linear combination of unknown model co-
2013) are important variables for predicting oil spills. efficients b and g is a link function.
The probability of oil spill occurrence at each raster cell is a For our generalized linear model (GLM), we defined the linear
function of the explanatory variables (Fig. 3). After combining both predictor as a 3rd order polynomial function:
the reported oil spill events and the generated pseudo-absences
into a single dataset, we randomly draw a 75% of records to train
yi ¼ S3 s3 þ S2 s2 þ S1 s þ W3 w3 þ W2 w2 þ W1 w þ F3 f 3 þ F2 f 2
the model. The values of the explanatory variables as well as that of
the response variable (1 for presence and 0 for absence sites) were þ F1 f þ C3 c3 þ C2 c2 þ C1 c þ b
used in model training, and the resulted model was used for pre-
diction of the probability of oil spills at each pixel over the study where yi is the response variable (occurrence of oil spill), and s, w, f,
area. The remaining 25% of the records was used for validation (i.e., c are the normalized explanatory variables, while S, W, F and C are
accuracy assessment). The randomly sampling procedure into the regression coefficients of the linear predictor for the indepen-
training, and validation datasets were repeated 100 times, and for dent variables, both types of variables set in the same order of the
each time, the model was fitted and a probability map was gener- factors, i.e., ship routes, oil well, oil facilities and coastline distance,
ated. The final probability map is the mean of the 100 prediction respectively. Each coefficient has a degree (exponent) between 1
maps. and 3, and b is the intercept.
To conduct the analyses, the R statistical software (R We used a polynomial function (order 3) for flexibility in form of
508 S. Mokhtari et al. / Ocean & Coastal Management 116 (2015) 504e511

Fig. 3. Schematic flowchart representing the modeling procedure to generate the oil spill probability map.

the non-linear relationship between response and predictor vari- predicts a probability (between 1 and 0 range) and the validation
ables. By using the polynomial function, the non-linear relationship sample shows an occurrence (either 1 or 0). The recommended
(in the form of flexible curves, depending on the degree) can be methods for validation of this model are the area under the curve
explored, while a linear combination would explore only a straight (AUC), extracted from the Receiver Operating Characteristic (ROC)
line relationship between the said variables. Adding more flexibility and the percent correctly classified (PCC), (Lobo et al., 2008). The
to the model (by considering higher order) increases the risk of ROC is created by plotting the true positive rate (sensitivity),
over-fitting (Ramsey and Schafer 2012). To avoid this problem, we against the false positive rate (1 e specificity). The AUC is an indi-
chose the degree 3 as recommended by many other studies (e.g., cator of the model performance (Metz, 1978). It has possible values
Naimi et al., 2011). The link function (g) is defined as: between 0 and 1. An AUC value of 1 indicates the best possible
  prediction with 100% sensitivity (no false negatives) and 100%
p specificity (no false positives), and an AUC value of 0.5 implies a
gðpÞ ¼ log
1p random predictive discrimination. The values less than 0.5 indicate
discrimination worse than chance.
where g(p) is the link function and p the predicted probability
(between 0 and 1).
For each coefficient, the p-value was checked in order to
determine whether the corresponding variable significantly
explained the response variable. All the variables that were not
significant in the model were removed. This procedure ended up
with appropriate model with the right features. Then, the model
was used to predict the probability of the oil spill at each pixel over
the study area. The values of the resulting raster range within a
minimum of 0 and a maximum of 1, indicating the probability of
occurrence of oil spill events.
The values of the model coefficients and the p-value of the co-
efficients are highly dependent on the choice of the presence and
pseudo-absence data points in the initial training sample. By using
resampling method to generate 100 samples and consequently 100
models, we tried to account for the uncertainty in the model input
data (Fig. 3). Together with the calculating the mean predicted
probability map at each raster cell, we also calculated the variance
of the probabilities over the 100 runs, that shows the uncertainty of
the model prediction (Fig. 4).

2.2.4. Model validation

For model validation, we compared the predictions of each of Fig. 4. Spatial variability of model prediction over the 100 runs at the cell level; greater
the 100 trained models with the 25% validation sample. The model values of variance indicate higher uncertainty.
 S. Mokhtari et al. / Ocean & Coastal Management 116 (2015) 504e511 509

Percent Correctly Classified (PCC) is calculated as:

true positive þ true negative

PCC ¼
N

where N is the total number of validated events.

To calculate sensitivity, specificity, AUC and PCC, we need to
define a threshold value. The threshold is the occurrence proba-
bility above which we consider the value as “present”. In our vali-
dation we used the default threshold value of 0.5.
In order to assess the overall model performance, we calculated
the mean, minimum, and maximum of the ROC curves and the PCC
values, obtained through the 100 runs of the model validation.

3. Results

To examine which factors are the most important in the oil spill
model, we analyzed each of the 100 model runs, keeping only the
significant coefficients. For each model, there were up to 13 Fig. 5. Map of oil spill probability in the EEZ of Iran, based on average of 100 model
possible coefficients. i.e., three coefficients for each variable and 1 predictions.
intercept for the model, obtained from the summary of each model.
Table 1 shows the total number of models in which at least one
coefficient was found to be significant for oil facility, oil well, ship Table 2
Summary of model validation.
route or coastline. According to Table 1, the oil facility distance is
the most important factor and it was found to be significant in all of Value AUC PCC Sensitivity Specificity
the 100 models. The second and the last most important factor Mean 0.7902968 0.7227619 0.6822869 0.7667660
were the ship route distance and the oil well distance, respectively. Maximum 0.8795136 0.7904762 0.8431373 0.8909091
The resulting map of a probability of an oil spill was obtained as Minimum 0.6844023 0.6285714 0.4893617 0.6250000
the average of all the 100 model predictions (Fig. 5). A large area of a
high oil spill probability was detected in the northern part of the
study area, where there are large facilities in combination with predicted by the model. In our results, the mean of specificity is
heavy ship traffic. In the southeast area of the Persian Gulf, a higher greater than that of sensitivity, probably because the pseudo-
predicted probability lies farther from rather than near to the absence points were chosen with the constraint of not to overlay
coastline, clearly in association with ship routes. the reported (present) oil spill locations.
For each model the mean, maximum, and minimum of valida- The testing events were selected as a 25% random sample of
tion statistics (Table 2) are based on 100 different model runs, each those reported. The comparison between the events in the test
run using a different subset of the presence and pseudo-absence dataset and the map (Fig. 7) shows that the reported events
data points, for both the model training and validation. The sta- generally were occurred in predicted high-probability areas, while
tistics AUC and PCC can be calculated for different threshold levels,
threshold being the predicted probability between 0 and 1 and
considering the binary prediction as “present”, i.e., if threshold is
0.5, all values with occurrence probability >0.5 are considered as
present and all values with occurrence probability <0.5 as absent.
Table 2 shows the results for the default threshold ¼ 0.5.
It can be seen that the mean AUC is 0.79, ranging between 0.68
and 0.88 for the 100 model validation runs. A value of AUC between
0.70 and 0.80 can be considered as a fair fit (Metz, 1978). Fig. 6
shows the ROC curves for all 100 models with the highlighted
mean of ROC curve.
The mean sensitivity of the model is 0.68. Sensitivity is the true
positive rate and measures the proportion of the reported events
where a high probability of oil spill (P > 0.5) was predicted by the
model. The mean specificity of the model is 0.77, specificity being
the true negative rate or the proportion of randomly allocated
pseudo-absence points, with a low probability of oil spill (P < 0.5)

Table 1
Number of models where the regression coefficient for distance of oil facility, oil
well, coastline or ship route were significant.

Explanatory variable Percentage of significant coefficients

Oil facility 100

Oil well 35
Coastline 62
Fig. 6. ROC curves obtained through validation of 100 models; the mean ROC curve is
Ship route 71
highlighted.
510 S. Mokhtari et al. / Ocean & Coastal Management 116 (2015) 504e511

risks. It should be noted that the even the less important variable in
our study (i.e., oil wells) was significant in their association to the
oil spill events. Thus, the importance of each variable is a relative
measure comparing to those for the other variables.
In our study, we generated pseudo-absence records to be able to
use logistic regression. The pseudo absences, however, may intro-
duce false information to the model and be considered as a source
of uncertainty. When the collected data on oil spill events are
systematically collected or the data are not subjected to
geographical bias (i.e., all parts of study area were monitored
through data collection), the real absence data may be inferred and
allocated in the area where no oil spill events were reported. If
these conditions are not met, we cannot be sure whether the
inferred absence is a real or a false one. Under these circumstances,
generating a random pseudo-absence (as used in this study) may be
a preferable method, as the same and balanced level of false/true
absences were introduced over the whole study area.
Fig. 7. Predicted probability of oil spills in northern part of the Persian Gulf. The yellow A limitation in our study was that the model was trained by
points show the 25% random sample of the reported events used for model validation. using events only from the northern part of the Persian Gulf. This
(For interpretation of the references to colour in this figure legend, the reader is may be a reason why the distance to the oil wells was the less
referred to the web version of this article.) important variable in our study as many of the active oil wells are
located in the southern part of the Persian Gulf. Moreover, the ac-
curacy of the map in the other parts of the Persian Gulf could not be
localities with no reports of events are distributed in the low
validated due to the restricted data availability. As a future study, it
probability areas. A few errors in the model predictions lie in the
would be interesting to evaluate the performance of the model to
eastern region of the Persian Gulf, where the events considered for
predict within those areas (the procedure is also called extrapola-
validation are located while the probability are low.
tion) by comparing the results with the reported events from those
areas. The second limitation is the possible inaccuracies in the input
4. Discussion data. For example, we used the tanker traffic density category map
from the AIS, but the real marine traffic intensity may be slightly
In this study, we developed a predictive model that generates different from our approximation. Also, the exact locations of the
the spatial distribution of the probability of oil spill events in the reported oil spill events may be inaccurate or imprecise.
northern part of the Persian Gulf. Moreover, a set of factors that An advantage of our model is that it was built using free and
may cause a risk in occurrence of oil spill events was scrutinized open source software. Therefore, it can be reproduced in other
and investigated whether and to what extent each factor actually marine regions, and the output file of the model is interoperable
influences the oil spill risk. The generated probabilities provide with other oil spill simulation software.
some insights into the “where and why” of oil spill risks. The key
importance of this study is that the model can be used to estimate 5. Conclusions
the level of risk in the areas with little or no surveillance. The
probability map can be used to identify the locations that have a In this study, we explored whether and to what extent the
high potential for oil spills. This model provides a useful informa- causal factor can be used as proxies to estimate the risk of oil spill
tive tool and may be used in studies that need such estimation on events in the north of the Persian Gulf. We employed a predictive
the risk level such as scenario-based planning as well as in decision modeling approach to fit the model between reported oil spill
support systems. events and four explanatory variables including oil facilities, oil
Our results show a successful use of the modeling approach in wells, ship routes and coastlines. Our results showed that the
identifying areas with high probability of oil spill risks and are in explanatory variables significantly explained the spatial distribu-
agreement with those concluded by Serra-Sogas et al. (2014). The tion of oil spill events. Moreover, oil facilities and ship routes were
initial motivation of their study was, however, to address the the most important variables to estimate the probability of the oil
challenges of surveillance programs in how to effectively allocate spill events. Of key importance of this study is the development of a
its resources to improve the accuracy of detecting oil discharges model that can be used to estimate the level of risk in the areas with
through such surveillance. Despite these differences and also the little or no surveillance. We conclude that the indirect information
differences in the data types and consequently model settings, or proxies can potentially be used to infer the risk of oil spills.
these two studies deliver the same conclusion that underlying
factors associated with the spatial patterns of oil spill events can be Author's contribution
identified through the modeling procedures and then used to
predict the spatial distribution and risk level of oil spills. The work presented here was carried out in collaboration be-
There is a similarity between our study and one by Liu et al. tween all authors. S.M., S.M.H., B.N., J.K and A.J conceived the ideas
(2015) as they also used ship routes and oil platforms to predict and reviewed the results. S.M and S.M.H implemented the study,
the spatial distribution of oil spills. Apart from the difference in analyzed the results and drafted the manuscript. S.M.H., A.D and
input data formats, our study expands the work by Liu et al. (2015), M.T.A supervised the research and reviewed the manuscript. All
as we used actual reported oil spill events to train and validate the authors read and approved the final manuscript.
spatial predictions of the models. This enabled us to find out which
proxies have significant association with oil spill events. In addi- Acknowledgments
tion, our study provided a ranking of the factors according to their
level of importance to explain the spatial distribution of oil spill Thanks to the Aalto University and Louisiana state university for
 S. Mokhtari et al. / Ocean & Coastal Management 116 (2015) 504e511 511

sharing the knowledge and experience about environmental Kankara, R., Subramanian, B., 2007. Oil spill sensitivity analysis and risk assessment
for Gulf of Kachchh, India, using integrated modeling. J. Coast. Res. 1251e1258.
modeling. We would like to acknowledge the Ports and Maritime
Keeney, R.L., 1982. Decision analysis: an overview. Oper. Res. 30 (5), 803e838.
Organization of I.R.Iran for their help in gathering the data, espe- Ko, T.-T., Chang, Y.-C., 2010. Integrated marine pollution management: a new model
cially Mr. Asghar Fallahi and Ms. Akram Beheshti.Thanks to Javadi of marine pollution prevention and control in Kaohsiung, Taiwan. Ocean Coast.
Khorsaghani, Ali Hadadpour, Arash Karimpour in Louisiana state Manag. 53, 624e635.
Lehikoinen, A., Luoma, E., Ha €nninen, M., Storgård, J., Kuikka, S., 2012. Probabilistic
University for providing comments and suggestions. Thanks to risk assessment and decision support tools for the evaluation of oil transport in
Daniel P. Ames, Dustin Woodbury and Ellen Vuosalo Tavakoli for the Gulf of Finland, North-Eastern Baltic Sea. In: International Environmental
help with the English editing. Modelling and Software Society (IEMSs) 2012 International Congress on Envi-
ronmental Modelling and Software Managing Resources of a Limited Planet,
Sixth Biennial Meeting, Leipzig, Germany.
References Liu, X., Meng, R., Xing, Q., Lou, M., Chao, H., Bing, L., 2015. Assessing oil spill risk in
the Chinese Bohai Sea: a case study for both ship and platform related oil spills.
Allan, S.E., Smith, B.W., Anderson, K.A., 2012. Impact of the deepwater horizon oil Ocean Coast. Manag. 108, 140e146.
spill on bioavailable polycyclic aromatic hydrocarbons in Gulf of Mexico coastal Lobo, J.M., Jime nez-Valverde, A., Real, R., 2008. AUC: a misleading measure of the
waters. Environ. Sci. Technol. 46 (4), 2033e2039. performance of predictive distribution models. Glob. Ecol. Biogeogr. 17,
Alam, I.A., 1993. The 1991 Gulf War oil spilldlessons from the past and a warning 145e151.
for the future. Mar. Pollut. Bull. 27, 357e360. Lu, J., 2003. Marine oil spill detection, statistics and mapping with ERS SAR imagery
Aven, T., 2010. On how to define, understand and describe risk. Reliab. Eng. Syst. Saf. in south-east Asia. Int. J. Remote Sens. 24 (15), 3013e3032.
95 (6), 623e631. Mata, A., Corchado, J.M., 2009. Forecasting the probability of finding oil slicks using
Bejarano, A.C., Michel, J., 2010. Large-scale risk assessment of polycyclic aromatic a CBR system. Expert Syst. Appl. 36, 8239e8246.
hydrocarbons in shoreline sediments from Saudi Arabia: environmental legacy Mahmoudi, S., 1997. Legal protection of the Persian Gulf's marine environment. Mar.
after twelve years of the Gulf war oil spill. Environ. Pollut. 158, 1561e1569. Policy 21, 53e62.
Bernabeu, A., Fern ndez, S., Bouchette, F., Rey, D., Arcos, A., Bayona, J.,
andez-Ferna MEMAC, 2012. Pollution Report 2011e2012.
Albaiges, J., 2013. Recurrent arrival of oil to Galician coast: the final step of the Mirvakili, H., Zaker, N.H., Imani, F., 2013. Evaluation of oil pollution and origin in
prestige deep oil spill. J. Hazard. Mater. 250, 82e90. surface coastal sediments of Kharg island in the persian gulf. Coast. Res. 93e98.
Bigano, A., Sheehan, P., 2006. Assessing the Risk of Oil Spills in the Mediterranean: Special Issue.
the Case of the Route from the Black Sea to Italy. Metz, C.E., 1978. Basic principles of ROC analysis. In: Seminars in Nuclear Medicine,
Burgman, M., 2005. Risks and Decisions for Conservation and Environmental vol. 8. Elsevier, p, pp. 283e298.
Management. Cambridge University Press. Naimi, B., Skidmore, A.K., Groen, T.A., Hamm, N.A.S., 2011. Spatial autocorrelation in
Carson, R.T., Mitchell, R.C., Hanemann, W.M., Kopp, R.J., Presser, S., Ruud, P.A., 1992. predictors reduces the impact of positional uncertainty in occurrence data on
A contingent valuation study of lost passive use values resulting from the Exxon species distribution modelling. J. Biogeogr. 38, 1497e2203.
Valdez oil spill. University Library of Munich, Germany. Naimi, B., Hamm, N.A.S., Groen, T.A., Skidmore, A.K., Toxopeus, A.G., 2014. Where is
Carson, R.T., Mitchell, R.C., Hanemann, M., Kopp, R.J., Presser, S., Ruud, P.A., 2003. positional uncertainty a problem for species distribution modeling? Ecography
Contingent valuation and lost passive use: damages from the Exxon Valdez oil 37, 191e203.
spill. Environ. Resour. Econ. 25, 257e286. Nelder, J.A., Baker, R., 1972. Generalized linear models. Encyclopedia of Statistical
Chefaoui, R.M., Lobo, J.M., 2008. Assessing the effects of pseudo-absences on pre- Sciences. John Wiley & Sons.
dictive distribution model performance. Ecol. Model. 210, 478e486. Ng, A.K., Song, S., 2010. The environmental impacts of pollutants generated by
Díaz, B., Pavo n, A., Go
mez-Gesteira, M., 2008. Use of a probabilistic particle tracking routine shipping operations on ports. Ocean Coast. Manag. 53 (5), 301e311.
model to simulate the prestige oil spill. J. Mar. Syst. 72, 159e166. O'Hara, P., Serra-Sogas, N., Canessa, R., Keller, P., Pelot, R., 2013. Estimating discharge
Ebrahimi-Sirizi, Z., Riyahi-Bakhtiyari, A., 2013. Petroleum pollution in mangrove rates of oily wastes and deterrence based on aerial surveillance data collected in
forests sediments from Qeshm island and Khamir PortdPersian gulf, Iran. western Canadian marine waters. Mar. Pollut. Bull. 69 (1), 157e164.
Environ. Monit. Assess. 185, 4019e4032. Pous, S., Lazure, P., Carton, X., 2015. A model of the general circulation in the Persian
Engler, R., Guisan, A., Rechsteiner, L., 2004. An improved approach for predicting the Gulf and in the Strait of Hormuz: intraseasonal to interannual variability. Cont.
distribution of rare and endangered species from occurrence and pseudo- Shelf Res. 94, 55e70.
absence data. J. Appl. Ecol. 41 (2), 263e274. Price, J.M., Johnson, W.R., Marshall, C.F., Ji, Z.-G., Rainey, G.B., 2003. Overview of the
Fattal, P., Maanan, M., Tillier, I., Rollo, N., Robin, M., Pottier, P., 2010. Coastal oil spill risk analysis (OSRA) model for environmental impact assessment. Spill
vulnerability to oil spill pollution: the case of Noirmoutier island (France). Sci. Technol. Bull. 8 (5), 529e533.
J. Coast. Res. 879e887. R Development Core Team, 2014. R: A Language and Environment for Statistical
Fingas, M., Brown, C., 2014. Review of oil spill remote sensing. Mar. Pollut. Bull. 83 Computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN: 3-
(1), 9e23. 900051-07-0. Available online at. http://www.R-project.org/.
Garcia, D.A., Bruschi, D., Cumo, F., Gugliermetti, F., 2013. The Oil Spill Hazard Index Rahmanpoor, S., Ghafourian, H., Hashtroudi, S.M., Bastami, K.D., 2014. Distribution
(OSHI) elaboration. An oil spill hazard assessment concerning Italian hydro- and sources of polycyclic aromatic hydrocarbons in surface sediments of the
carbons maritime traffic. Ocean Coast. Manag. 80, 1e11. Hormuz strait, Persian Gulf. Mar. Pollut. Bull. 78 (1), 224e229.
Hassanzadeh, S., Hosseinibalam, F., Rezaei-Latifi, A., 2011. Numerical modelling of Sammarco, P.W., Kolian, S.R., Warby, R.A., Bouldin, J.L., Subra, W.A., Porter, S.A., 2013.
salinity variations due to wind and thermohaline forcing in the Persian Gulf. Distribution and concentrations of petroleum hydrocarbons associated with the
Appl. Math. Model. 35 (3), 1512e1537. BP/Deepwater horizon oil spill, Gulf of Mexico. Mar. Pollut. Bull. 73, 129e143.
Hengl, T., Sierdsema, H., Radovi c, A., Dilo, A., 2009. Spatial prediction of species' Sheppard, C., Al-Husiani, M., Al-Jamali, F., Al-Yamani, F., Baldwin, R., Bishop, J.,
distributions from occurrence-only records: combining point pattern analysis, Benzoni, F., Dutrieux, E., Dulvy, N.K., Durvasula, S.R.V., 2010. The Gulf: a young
ENFA and regression-kriging. Ecol. Model. 220 (24), 3499e3511. sea in decline. Mar. Pollut. Bull. 60, 13e38.
Hijmans, R.J., van Etten, J., 2012. Raster: Geographic Analysis and Modeling with Serra-Sogas, N., O'Hara, P.D., Canessa, R., 2014. Modeling the distribution of illicit
Raster Data. R Package Version 1, pp. 9e92. oily discharges detected by aerial surveillance in western Canadian marine
Hoek, G., Beelen, R., de Hoogh, K., Vienneau, D., Gulliver, J., Fischer, P., Briggs, D., waters. Mar. Pollut. Bull. 87, 76e87.
2008. A review of land-use regression models to assess spatial variation of Serra-Sogas, N., O'Hara, P., Canessa, R., Bertazzon, S., Gavrilova, M., 2008. Explor-
outdoor air pollution. Atmos. Environ. 42 (33), 7561e7578. atory Spatial Analysis of Illegal Oil Discharges Detected off Canada's Pacific
Hong, X., Chen, W., Zhang, L., 2010. A probabilistic risk forecast of accidental oil Coast Computational Science and its Applicationseiccsa 2008. Springer,
spills from vessels in Luoyuan Bay, Fujian Province, PRC. Procedia Environ. Sci. 2, pp. 81e95.
49e56. Singkran, N., 2013. Classifying risk zones by the impacts of oil spills in the coastal
Jolma, A., Ames, D., Horning, N., Mitasova, H., Neteler, M., Racicot, A., Sutton, T., waters of Thailand. Mar. Pollut. Bull. 70 (1), 34e43.
2008. Chapter ten free and open source geospatial tools for environmental Spaulding, M., Kolluru, V., Anderson, E., Howlett, E., 1994. Application of three-
modelling and management. Dev. Integr. Environ. Assess. 3, 163e180. dimensional oil spill model (WOSM/OILMAP) to hindcast the Braer spill. Spill
Jolma, A., Lehikoinen, A., Helle, I., Venesj€ arvi, R., 2014. A software system for Sci. Technol. Bull. 1, 23e35.
assessing the spatially distributed ecological risk posed by oil shipping. Environ. Vafai, F., Hadipour, V., Hadipour, A., 2013. Determination of shoreline sensitivity to
Model. Softw. 61, 1e11. oil spills by use of GIS and fuzzy model. Case studyeThe coastal areas of Caspian
Juntunen, T., Rosqvist, T., Rytko €nen, J., Kuikka, S., 2005. How to Model the Oil Sea in north of Iran. Ocean Coast. Manag. 71, 123e130.
Combating Technologies and Their Impacts on Ecosystem: a Bayesian Networks Van Dorp, J.R., Merrick, J.R., 2011. On a risk management analysis of oil spill risk
Application in the Baltic Sea. ICES CM. using maritime transportation system simulation. Ann. Oper. Res. 187, 249e277.
Ka€mpf, J., Sadrinasab, M., 2005. The circulation of the Persian Gulf: a numerical Wiens, J.A., Parker, K.R., 1995. Analyzing the effects of accidental environmental
study. Ocean Sci. Discuss. 2 (3), 129e164. impacts: approaches and assumptions. Ecol. Appl. 5, 1069e1083.

View publication stats