Optimizing the Recovery Efficiency of Finnish Oil ... - ACS Publications

Jan 17, 2013 - ABSTRACT: Oil transport has greatly increased in the Gulf of Finland over the years, and risks of an oil accident occurring have risen...
0 downloads 7 Views 2MB Size
Policy Analysis pubs.acs.org/est

Optimizing the Recovery Efficiency of Finnish Oil Combating Vessels in the Gulf of Finland Using Bayesian Networks Annukka Lehikoinen,*,† Emilia Luoma,‡ Samu Man̈ tyniemi,‡ and Sakari Kuikka‡ †

Fisheries and Environmental Management Group, Department of Environmental Sciences, University of Helsinki, Kotka Maritime Research Center, Heikinkatu 7, FI-48100 Kotka, Finland ‡ Fisheries and Environmental Management Group, Department of Environmental Sciences, P.O. Box 65, FI-00014 University of Helsinki, Finland S Supporting Information *

ABSTRACT: Oil transport has greatly increased in the Gulf of Finland over the years, and risks of an oil accident occurring have risen. Thus, an effective oil combating strategy is needed. We developed a Bayesian Network (BN) to examine the recovery efficiency and optimal disposition of the Finnish oil combating vessels in the Gulf of Finland (GoF), Eastern Baltic Sea. Four alternative home harbors, five accident points, and ten oil combating vessels were included in the model to find the optimal disposition policy that would maximize the recovery efficiency. With this composition, the placement of the oil combating vessels seems not to have a significant effect on the recovery efficiency. The process seems to be strongly controlled by certain random factors independent of human action, e.g. wave height and stranding time of the oil. Therefore, the success of oil combating is rather uncertain, so it is also important to develop activities that aim for preventing accidents. We found that the model developed is suitable for this type of multidecision optimization. The methodology, results, and practices are further discussed.

1. INTRODUCTION Marine traffic, especially oil transport, has been growing fast in the Gulf of Finland (GoF) and is predicted to grow further also in the near future.1 The opening of new oil ports in Russia and Estonia are mainly responsible for this rapid growth.2 This increase in oil transport exposes the area to heightened probability of oil spills. The GoF is the easternmost part of the Baltic Seadesignated as a Particular Sensitive Sea Area (PSSA) by the International Maritime Organization (IMO). Being a shallow, brackish-, and cold-watered ecosystem with low biodiversity, a major oil accident could have severe consequences in the area.3 Risk is often defined as the product of the probability that some undesirable event materializes and the amount of harm that is caused in that case. Thus, in addition to the sufficient preparedness for oil combating, another important recourse for avoiding ecological disaster in the GoF is to improve maritime safety. We see that there is an obvious need to develop models that can be used to determine the correct allocation of resources between the oil combating activities and the actions that aim to prevent accidents. The model presented in this paper is our first step toward that goal. Oil combating in Finland is based on mechanical recovery only, which is in accordance with the recommendation of the Baltic Marine Environment Protection Commission (HELCOM) stating that mechanical means are preferable options instead of chemical and other nonmechanical means.4 Finland has 16 oil combating vessels which collect oil independently © 2013 American Chemical Society

and operate in the open sea. The vessels are owned by four different organizations: the Finnish Navy, Meritaito PLC, the Finnish Border Guard, and the Åland government. The number, location, preparedness, and combating capacity of the vessels’ influence the amount of oil recovered in the case of an accident. In addition, several environmental factors remarkably affect the oil combatingsometimes preventing it all together, which makes the success of oil combating highly uncertain and difficult to predict.3,5,6 In this paper, we introduce a Bayesian network (BN) that can be used for the probabilistic analysis of the oil recovery efficiency of Finnish oil combating vessels in different conditions and for studying their optimal disposition (Figure 1). The analysis takes into account the combinations of environmental settings and their effect on the individual vessels, and finally on the whole fleet. This model was originally developed to run as a part of a larger risk and decision analysis meta-model.7

2. METHODS 2.1. Bayesian Networks (BNs). BNs are frequently used in environmental studies.10−13 Recently they have also been used Received: Revised: Accepted: Published: 1792

September 6, 2012 January 14, 2013 January 17, 2013 January 17, 2013 dx.doi.org/10.1021/es303634f | Environ. Sci. Technol. 2013, 47, 1792−1799

Environmental Science & Technology

Policy Analysis

2.2. Expert Interviews. Experts are often used when it is expensive or impractical to gain the information in other ways18 or there is no other data available.19 Expert knowledge has been used in risk assessment studies where unrealized future scenarios need to be evaluated.20−22 Elicitation of experts’ probabilities has played a central role in several BN studies as well.3,17 Besides the conditional probabilities, expert elicitation is used for framing and structuring the research problems.23,24 O’Hagan and colleagues18 give a rigorous presentation on the considerations and methods for elicitation of expert knowledge. In this study, a semistructured expert interview19 was used: the questions had been formed beforehand but the interviewees answered informally. The final probabilities were formed with the help of a facilitator. The experts’ views were also utilized when the model structure was decided. The aim of the interviews was to receive as much information as possible to learn the probabilistic conditional dependencies between the variables related to oil combating vessels and the courses of action in the case of an oil accident. People working in the oil combating vessels were chosen as experts because they were seen to have the most authentic information. Five captains, one navigating officer, and an inspector of the Finnish Environmental Administration were interviewed. As our interviewees had no earlier experience on handling probability distributions, we asked them to evaluate maximum, the most probable, and minimum values, which were used to form the CPDs. Some of the interviewees had no experience in real oil combating situations, which made answering a challenging task for them. For a couple of the vessels, some of the estimates were missed out. However, all the estimates for a majority of the vessels were gained and these were utilized for the construction of the prior distributions for the missing vessels, by taking into consideration all the relevant differences and similarities. 2.3. Model Overview and Variables. Hugin Researcher 7.6 software25 was used for constructing and analyzing the model. The entity consists of the main model (Figure 2) and ten vessel-specific submodels (Figure 3). The main model provides information for the submodels on the uncertain, random conditions during the oil accident. The submodels include vessel-specific knowledge that is relevant for the output: the probability distribution of the amount of oil collected. This

Figure 1. Main shipping lanes of the Gulf of Finland (gray dashed line) and the four home harbors of the Finnish oil combating vessels (black boxes). The GoF is divided into five sectors (C1−C5), which are used when analyzing the traffic statistics and accident probabilities.8,9 The hypothetical accident points (hot spots) are shown with stars. Map data from free database http://www.grida.no/ baltic.

in problems related to oil spills and their management.3,6,14 To populate a BN with appropriate conditional probability distributions (CPD), stochastic simulations or separate data analyses,15,16 as well as interviews of one or more experts to elicit their degrees of belief17,18 can be used. The advantage of BNs is that both qualitative and quantitative knowledge can be used and numerically analyzed. This is important especially in cases like oil combating, where the need for information is high due to high stakes and uncertainties, but historical data sets are rare and most of the knowledge must be obtained from models or experts. In this study BNs were used to integrate different types of data and knowledge compiled from several sources. Experts in oil combating were interviewed to gain information on the courses of action in the real situations and to collect experiential data of the real recovery efficiency and the uncertainty related to it. In addition to the expert elicitation, a variety of existing statistics, measurements, equations, and earlier modeling results were utilized when forming the CPDs.

Figure 2. Structure of the main model. Decision variables are denoted by the purple rectangles, random variables are denoted by the yellow ovals, and submodels are shown as angulated white ovals; the green quadrangle is a utility node. 1793

dx.doi.org/10.1021/es303634f | Environ. Sci. Technol. 2013, 47, 1792−1799

Environmental Science & Technology

Policy Analysis

Figure 3. Structure of the submodel of the oil combating vessel Louhi. The nodes with a gray dashed line border are the input variables from the main model; the one with a gray uniform border (variable Recovered amount) provides the output information back to the main model.

knowledge is again input to the main model, where the final calculation for the probability distribution of the recovery percenttaking into account the likely evaporationis performed. The variables are presented in detail in the Supporting Information. The model is completed for the decision analysis purposes by adding the alternative harbors (the locations of vessels), in form of decision variables, for each of the vessels. The location of a vessel is linked to the time that is elapsed before the vessel arrives at the site of the accident through distance and prevailing winds (Figure 3). Finally, the recovery efficiency is calculated by proportioning the potentially recovered amount of oil to the random size of the leak, taking into account the uncertain evaporation-driven reduction. For the decision optimization (i.e., finding the disposal strategy that maximizes the expected recovery efficiency), a utility node that describes the interests of the society must be added to the model. The utility node assigns value to the alternative states of the variable of interest. This enables the calculation of expected utility (EU), which is a weighted average of the utilities of alternative states, weighted by their realization probabilities. In this case we used the recovery percent as such for the valuation. Thus, the utility node in our model will get values between 0 and 100, the maximum utility being gained in a situation where the recovery efficiency is 100% of the collectable oil. In addition, to study the uncertainty behind the EU, a probability distribution of the recovery efficiency can be explored via a separate random node. Environmental Variables. According to the interviewees, prevailing environmental conditions have a remarkable effect on the success and efficiency of the oil combating. We have grouped nodes Season, Wave height, and Wind direction into this category (Table S1 in the Supporting Information). Season is an independent random variable that affects the probabilities of the two other environmental variables, as the wind direction and strength are highly seasonal in the Baltic Sea. The prior probability distribution of Season illustrates how likely an accident happens in the GoF within a certain season in relation to the others.5,6 So far only spring (March−May), summer (June−August), and autumn (September−November) are included as alternative states. Winter months are left outside because oil behavior and oil combating in icy conditions differ a lot from those of ice-free waters. Wintertime oil combating in the extreme conditions of the Gulf of Finland would require a model of its’ own.

Variable Wave height describes the level of significant wave height at the time of accident. Conditional probability distributions of wave height, given the season, have been drawn from wave buoy observations26 by Juntunen and coauthors.5 According to the experts, the newest combating vessel, Louhi, is the only one that is able to collect oil in waves higher than 2 m. Wind direction, in turn, affects the speed of the vessels case-specifically (see Oil Combating Vessel Variables). The observed frequencies of the wind direction, given the season, form the CPDs of the variable Wind direction. Statistics were obtained from the observations of the Finnish Meteorological Institute. Data from 1971 to 2000 was used. Oil Spill Variables. Variables related to the oil spill and behavior of the oil in the sea, thus affecting the possibility to collect it, are the accident location (Hot spot), Stranding time, Spill size, Oil type, and Evaporation. There are five areas in the GoF for which the model was applied.8,9 In the first model version, these were given a uniform distribution, meaning they are assumed to be equally accident-prone. The role of this prior distribution is discussed further below. Stranding time is the time (in days) that elapses before most of the leaked oil is drifted too close to the shoreline to be combated by the recovery vessels. In the model presented, it is dependent on the accident location only. In formulating the distributions, the estimation in Hietala and Lampela27 was utilized. According to them, if the accident occurs in the middle of the GoF, oil will strand within the next 24 h at the earliest and after 9 days at the latest. If the accident happens on the sea area of Finland, the average stranding time of the oil is 3 days. For estimating the prior probability distribution of the independent variable Spill size, an observed (year 2008) proportional distribution for the size of tankers navigating in the GoF was obtained from Kuronen et al.1 A probability distribution of spill size, given the size of the tanker and type of the accident (collision or grounding), was obtained from Montewka et al.28 For evaluating the relative probabilities of the accident types collision and grounding within the areas C1−C5 (Figure 1), we used the accident statistics of HELCOM from the years 1989−2008.29,30 Evaporation is the only weathering process that removes oil from the sea. The probability distribution of the variable Evaporation, given the season, wave height, and oil type, was obtained from the work of Juntunen and coauthors.5 The probability distribution for the oil type leaked to water is conditional to the area (C1−C5) in question, the CPDs of the variable resting on the harbor1794

dx.doi.org/10.1021/es303634f | Environ. Sci. Technol. 2013, 47, 1792−1799

Environmental Science & Technology

Policy Analysis

vessels. To make the calculation process computationally faster, the recovery rates of the vessels are first summed up for each of the organizations separately and put together after that in the variable Potential recovered amount meaning the total recovery capacity (in tons) of all the vessels. The variable Possible recovered amount describes the amount of oil to be combated, given the spill size and evaporation rate. As an end result of the model, the recovery efficiency of the vessels (%), given potential and possible recovered amounts of oil, can be evaluated in form of discrete probability distribution (a random variable). The knowledge is also utilized to optimize the disposition of the vessels.

specific transportation statistics from the year 2007, presented by Kuronen and coauthors.1 The variable has three alternative states: heavy, medium, and light oil, based on the viscosity grouping presented in Lindgren and Lindblom.31 Oil Combating Vessel Variables. The vessel-specific random variables (Figure 3) are Time of the day, State of readiness, Departure decision, Departure time, Direction of the wind to the vessel, Speed, Time of arrival, Combating time, Filling time of the tank, Emptying time of the tank, True recovery rate, and Recovered amount. The CPDs of those presented in bold were formed on the basis of the expert interviews. In addition, each vessel has a decision node of its own including four possible home harbors (shown in Figure 1) to be optimized. When the oil accident happens, prevailing time of the day has an impact on the departure time of the combating vessels, whichfor some of the boatsis longer outside the office hours. Probability distribution for the variable Time of the day was formed utilizing the accident database DAMA.32 Data of 1997−1999 and 2001−2006 from the GoF was used. According to the experts interviewed, in the case of large scale accident, more oil recovery equipment is needed, which lengthens the departure time. Thus the variable Departure time is conditional also to Leak size in the model. The third factor in the model affecting the departure time is the preparedness of the vessel. Links between some of the vessel submodels can be observed in Figure 2. This is because the preparedness of certain boats of the same owner is interrelated as part of them may be carrying out some other tasks, such as maintenance, transport, and guarding. For example, in Figure 3, the State of readiness variables of the vessels Halli and Hylje are affecting that one of Louhi correspondingly with the expert comment that usually only one of these vessels is in its’ home harbor, being ready to start in 4 hours. Thus, the model takes into account the main features of the dynamics of the fleet. The times of arrival to the accident places (shown in Figure 1) were calculated given departure time, speed, and distance. The distance was measured from every harbor to every accident spot by using the shortest voyage along the shipping lanes. The likely speed of the vessel was evaluated by the experts, given varying combinations of the wave height and direction of the wind to the vessel. Variable Direction of the wind to the vessel describes the proportions of the voyage the vessel has downwind, headwind, or side-wind, given the actual wind direction, the departure harbor, and the destination site. Combating time meaning the time (in hours) that the vessel has for performing its task before most of the oil is out of its reachwas calculated taking into account the likely time of arrival and stranding time of the oil. The variable True recovery rate describes how much oil a certain vessel can combat per hour. It was calculated taking into account combating time, tank size, and the filling and emptying times of the tank. The variable Recovered amount is a product of combating time and true recovery rate. It describes how much oil a certain vessel can combat before the oil strands. It is the output variable of the submodel, providing this vessel-specific information for the main model for further summation. Total Recovery Variables. After the vessel-specific recovery rate information is received from the submodels, it is summarized in the main model by the variables Recovered amount by the Border Guard, Recovered amount by the Navy, Recovered amount by Meritaito, Potential recovered amount, Possible recovered amount, and Recovery efficiency of the

4. RESULTS In this article, two alternative vessel disposition policies are compared: one corresponding with the current placement of the vessels and another optimized by the model. We present the central results of the model in its prior state and compare three scenarios (Sc 1−3) having known oil type (in each case locked in state Heavy) and varying wave heights (Sc1: 0−1 m, Sc2: 1−2 m, Sc3: 2−3 m). Policy optimization performed by Hugin is based on the search of decision combination that maximizes the expected utility (EU)in this case the expectation value for the oil recovery efficiency. The result for the optimal disposition of the oil recovery vessels, together with their current home harbors, is shown in Table 1. The model positions all the vessels in the Table 1. Results of the Decision Optimization Showing Vessels’ Current Locations and the Harbors Recommended by the Model

a

vessel name

current harbor

optimization result

Louhi Halli Hylje Merikarhu Tursas Uisko Seili Oili I Oili II Oili III EUa

Kirkkonummi Turku Kirkkonummi Helsinki Turku Turku Helsinki Helsinki Turku Kotka 76.6

Kotka Kotka Kotka Kotka Kotka Kotka Helsinki Helsinki Helsinki Helsinki 77.2

EU values for both of the disposition policies are presented.

analysis to two easternmost harbors: Kotka and Helsinki. EU values produced by these two solutions are very close to each other. Randomizing the locations of the vessels (each having a uniform distribution for the harbors) does not produce a big difference from these (EU 76.7), thus given the current model settings and assumptions, disposition does not seem to be a central factor when organizing the open sea oil combating in the GoF. The value of information (VOI) can be used to study the sensitivity of decision analysis to variables that could be potentially observed before making the decision.33,34 VOI means the difference between the maximum EU values with and without the variable observed before the decision is made. We tested the VOIs for the decision Location Louhi when it is the first vessel to be located. As the disposition of the vessels has only a minor magnitude to the EU, also the VOIs are small. The VOI of variable Hot spot was naturally the highest (0.2) 1795

dx.doi.org/10.1021/es303634f | Environ. Sci. Technol. 2013, 47, 1792−1799

Environmental Science & Technology

Policy Analysis

large spills as more oil should be collected within a time unit. Unfortunately, oil recovery progress is slower when a spill size is large: in addition to the longer departure time of the vessels, the collecting tanks need to be emptied regularly during the work, which lowers the true recovery rates. In theory, evaporation should correlate positively with the recovery efficiency: the higher the evaporation, the lower the amount of collectable oil, thus the higher the recovery efficiency. Still our results show the higher evaporation producing smaller EU values (Figure 4). This is due to the fact that in Bayesian inference, the probabilities update also backward, from the lower hierarchy level to the higher one. Locking the child variable to a certain state will update our belief on the parents as well. Thus, the model infers that if the evaporation is high, it is more likely that the season is autumn, wave heights higher, and the oil type lighter. These are factors that radically worsen the success of oil recovery. Probability distributions of the recovery efficiency and recovery potential, with the oil type and wave height settings of the scenarios 1−3 and the prior distribution (all the random variables being uncertain), are presented in Figure 5. The charts

meaning that knowing the place of accident would help us most in the decision making. Other variables having potential, although very small, value for the policy optimizing were Oil type, Wind direction, Stranding time, Time of the day, and Evaporation. Instead of the disposition policy, five random factors seem to have remarkable effect on the oil recovery: oil type, wave height, stranding time, evaporation, and spill size. Prior joint distributions of these (in a situation where the accident location is unknown and all the random variables uncertain) are shown in Figure 4. EU values produced with locking the variable differ

Figure 4. Prior probability distributions of Wave height, Oil type, Stranding time, Evaporation, and Spill size. Expected oil recovery efficiency (EU-value) produced if the state will be locked (when the other variables are still uncertain) is shown above each bar for the current disposition of the recovery vessels. Note the different scales on the y-axes.

clearly among the states. In the figure, distributions of oil type and stranding time are joint distributions for all the alternative accident locations (being equally likely) and wave height for all the seasons weighted by their probabilities. Evaporation, in turn is a joint distribution of all possible combinations of the seasons and wave heights. Spill size has no parent variables, thus the distribution presented is an independent prior. These results are consistent with the information gained from the experts. First of all, combating light oil with recovery vessels is rather inefficient. Usually light oil evaporates or mixes with water before the combating even starts, thus the recovery efficiency of the vessels is very low. In optimal conditions, medium and heavy oil can be combated quite effectively and the evaporation is not as remarkable with them. Still, when the wave height is over 2 m, the only vessel capable of combating oil is the latest accession Louhi. In the wave height over 3 m, it is impossible to combat oil with any of the vessels. Stranding time naturally has a great effect on the combating success: the faster the oil strands the less time the vessels have for collecting it. The magnitude of stranding time increases in the case of

Figure 5. Probability distributions for the variables Recovery ef f iciency of the vessels (upper chart) and Potential recovered amount (lower chart) in three scenarios (Sc 1−3) and when the model is in its prior state. Vessels are located to their current home harbors. In each scenario, variable Oil type is locked to the state Heavy and the wave height varies, being Sc1: 0−1 m, Sc2: 1−2 m, and Sc3: 2−3 m. In prior state, both of the variables are uncertain. EU values for the scenarios were Sc1: 95.7, Sc2: 95.4, Sc3: 30.5.

show clearly how dependent on the prevailing weather conditions the success of the oil recovery is. In the wave heights between 0 and 2 m (Sc 1−2) the probability to gain the recovery efficiency of 80−100% for the heavy oil is over 0.9. Also, it seems very likely that the amount of collected oil could be over 10 000 tons, the probability for that being around 0.9 in scenarios 1−2 as well. Even exceeding the evaluated provision level of 30 000 tons27 seems quite likely (P > 0.4) under those circumstances. In Sc3, the whole combating lies on the shoulders of the vessel Louhi alone. The probability for the 1796

dx.doi.org/10.1021/es303634f | Environ. Sci. Technol. 2013, 47, 1792−1799

Environmental Science & Technology

Policy Analysis

We have shown that the oil recovery efficiency is a kind of either−or issue: it is more likely that the end result represents extremes than the midvalues. In a real situation, the accident parameters can be anything: any of the possible combinations of the alternative states of the variables may materialize. Still, we know that the elements are interlinked and that some of the combinations are more likely than others. By taking into account the relative frequencies and mutual cause−effect relations of these uncertain variables, we can get a more holistic view of the problem, which to our minds provides a better basis for the policy making than single scenario analyses. To make the analysis more realistic, the rest of the Finnish coastal areas should also be included in the model. The length of western continental coastline, bordered by the Gulf of Bothnia (GoB) and Archipelaco Sea (AS) is about 2.7 times the length of the GoF. Including accident points and possible harbors from this area in the analysis would probably change the results. For example, in the current analysis, the harbor of Turku is quite far away from most of the accident locations included and this is why the model did notin any of the cases testedsuggest that the placement of any vessels to Turku would be an optimal solution. Adding the GoB and AS to the analysis would likely make Turku a quite favorable home harbor as it is in the middle of the whole continental coastline. The accident points used in the model are scattered and unevenly distributed along the GoF, which may skew the results. More dense mesh of the possible locations might provide a better analysis of the disposal policy. We noticed that taking into account the relative accident probabilities of the locations did change the optimization result. As certain areas are found to be more accident prone than others,9,36 this is definitely worth acknowledging in the model as well. By adding spatial knowledge concerning the likely oil drifting and local nature valuese.g., the occurrences of threatened species and their conservation values37to the analysis, it is possible to find particularly risk-prone areas at the open sea that need higher provision level. When it comes to the areal protection needs (per length unit of the continental coastline), the huge amount of islands and islets with rich biodiversity would give special weighting to the GoF and AS compared with the GoB. These types of valuation factors, related to the amount of likely harm caused by an oil accident materializing within a certain sea area could be used as criteria for spatial weighting in the analysis of optimal recovery vessel disposition. Linking BNs with spatial data and map presentation seems to be a quite promising method for that type of risk modeling.38 With the improvements outlined above, our model could be used for the holistic planning of the Finnish open sea oil recovery. Despite the complexity of the current model, all the details were impossible to be taken into account. However, its inference seems to be in line with the original logic of the experts. The model shows clearly how coincidental the success of the oil recovery is, thus we propose that the precautionary management actions that potentially decrease the probability of oil accidents should be analyzed together with the oil recovery efficiency. In such an analysis their capabilities to diminish the overall risks should be compared in order to enable costeffective investments. We argue that there is a high need in oil risk analysis to learn more effectively from all historical experiences, from laboratory studies, and from different kinds of modeling techniques.1−3,6 Bayesian methodology, in general, is able to learn effectively from all information sources. For example, the use of

recovery efficiency under 10% is over 0.5 and the probabilities for smaller classes of collected amount of oil high. Still, if the other conditions happen to be optimal (e.g., small spill size and long stranding time), it might be possible to gain the recovery efficiency of 80−100% with the probability of 0.22 or collect over 10 000 tons with P = 0.07. As the accident location was observed to have the highest VOI, we wanted to test how the areal differences in the oil accident probabilities would affect our analysis. For that purpose, we changed the prior distribution of the variable Hot spot from uninformative (uniform) to informative. The distribution we used was a posterior taken from another BNmodel,7 analyzing tanker collisions and the probability they would lead to an oil leakage within the areas C1−C5. The most accident-prone area was C3 (P = 0.28) while the area with the lowest probability was C4 (P = 0.10). Although the EU-value gained with the optimal disposition (77.6) was still very close to the ones presented above, the recommended disposition was different: two of the vessels (Oili I and Seili) in Helsinki and the rest in Kotka. When testing the policy optimization for cases where the accident location was locked, we noticed that the model sometimes gives several, equally good, alternatives for some of the boats’ placement. We think these observations tell about the applicability of our approach to this type of multidecision optimization problems under uncertainty.

5. DISCUSSION Our results demonstrate that the current placement of the oil combating vessels in Finland is not optimal. However, the expected utility of the current placement is very close to that of the optimal placement, which means that there is no obvious need to change the home harbors. This is due to the fact that the disposition has a minor effect on the estimated recovery efficiency of the oil combating fleet, while the environmental and accident conditions have the biggest impact. The results still show that the applied BN-structure is suitable tool for this type of multidecision optimization. We collected and integrated into the BN a large amount of knowledge concerning the oil combating vessels and their courses of action in a case of an oil spill. In addition to the policy optimization, the model also enables us to create and analyze scenarios by locking certain variables to the state under our interest and considering the rest as uncertain. Still, one of its main added values, compared with other modeling techniques, is that the analysis can (and should) be made in the light of full uncertainty related to each of the variables. Then the result is a synthesis of all the possible scenarios, weighting to the most likely but still taking into account the possibility of the unlikely ones. The model in its prior state (without any settings in the random variables) depicts our current knowledge and prevailing uncertainty related to it, whichin our opinionis a good basis for the decision ranking and policy analysis. The novelty of this work is that it provides us probabilistic estimates of the oil recovery fleet efficiency by taking into account combinations of environmental conditions, which has not been available earlier. Earlier BN applications have focused on estimating the likely biological and ecosystem impacts of an oil spill as well as the pros and cons of alternative oil combating methods.3,5,6,35 We think that in order to get even better estimates for these issues, the efficiency of oil recovery under uncertain environmental conditions must be first properly estimated. 1797

dx.doi.org/10.1021/es303634f | Environ. Sci. Technol. 2013, 47, 1792−1799

Environmental Science & Technology

Policy Analysis

Bayesian network for the Gulf of Finland; the Baltic Sea. J. Hazard. Mater. 2011, 185, 182−192. (7) Lehikoinen, A.; Luoma, E.; Hänninen, M.; Storgård, J.; Kuikka, S. Probabilistic Risk Assessment and Decision Support Tools for the Evaluation of Oil Transport in the Gulf of Finland, North-Eastern Baltic Sea. International Environmental Modelling and Software Society (iEMSs) 2012 International Congress on Environmental Modelling and Software Managing Resources of a Limited Planet, Sixth Biennial Meeting, Leipzig, Germany, 2012; http://www.helsinki.fi/science/fem/articles/ 1365paper_long.pdf. (8) Ylitalo, J.; Häninen, M.; Kujala, P. Acccident probabilities in selected areas of the Gulf of Finland; TKK-AM 6; Helsinki University of Technology, 2008; pp 1−54; http://www.merikotka.fi/safgof/Ylitalo_ 2008_Accident.pdf. (9) Kujala, P.; Hanninen, M.; Arola, T.; Ylitalo, J. Analysis of the marine traffic safety in the Gulf of Finland. Reliab. Eng. Syst. Saf. 2009, 94, 1349−1357. (10) Varis, O.; Kuikka, S. Joint use of multiple environmental assessment models by a Bayesian meta-model: the Baltic salmon case. Ecol. Model. 1997, 102, 341−351. (11) Bacon, P. J.; Cain, J. D.; Howard, D. C. Belief network models of land manager decisions and land use change. J. Environ. Manage. 2002, 65, 1. (12) Borsuk, M. E.; Stow, C. A.; Reckhow, K. H. A Bayesian network of eutrophication models for synthesis, prediction, and uncertainty analysis. Ecol. Model. 2004, 173, 219. (13) Bromley, J.; Jackson, N. A.; Clymer, O. J.; Giacomello, A. M.; Jensen, F. V. The use of Hugin to develop Bayesian networks as an aid to integrated water resource planning. Environ. Modell. Softw. 2005, 20, 231−242. (14) Aps, R.; Herkül, K.; Kotta, J.; Kotta, I.; Kopti, M.; Leiger, R.; Mander, Ü .; Suursaar, Ü . Bayesian inference for oil spill related Net Environmental Benefit Analysis. In Coastal Processes; Brebbia, C., Benassai, G.; Rodriguez, G., Eds.; WIT Press: Boston, 2009; Vol 126, pp 235−246. (15) Mäntyniemi, S. Bayesian fisheries stock assessment: integrating and updating knowledge. PhD. dissertation, Rolf Nevanlinna Institute, Department of mathematics and statistics, University of Helsinki, 2006. (16) Fernandes, J. A.; Kauppila, P.; Uusitalo, L.; Fleming-Lehtinen, V.; Kuikka, S.; Pitkänen, H. Evaluation of Reaching the Targets of the Water Framework Directive in the Gulf of Finland. Environ. Sci. Technol. 2012, 46, 8220−8228. (17) Uusitalo, L.; Kuikka, S.; Romakkaniemi, A. Estimation of Atlantic salmon smolt carrying capacity of rivers using expert knowledge. ICES J. Mar. Sci. 2005, 62, 708−722. (18) O’Hagan, A.; Buck, C. E.; Daneshkhah, A.; Eiser, J. R.; Garthwaite, P. H.; Jenkinson, D. J.; Oakley, J. E.; Rakow, T. Uncertain Judgements: Eliciting Experts’ Probabilities; Wiley: Chichester, 2006; Vol. 2. (19) Leal, J.; Wordsworth, S.; Legood, R.; Blair, E. Eliciting Expert Opinion for Economic Models: An Applied Example. Value in Health (Wiley-Blackwell) 2007, 10, 195−203. (20) van der Gaag, L. C.; Renooij, S.; Witteman, C. L. M.; Aleman, B. M. P.; Taal, B. G. Probabilities for a probabilistic network: A case study in oesophageal cancer. Artif. Intell. Med. 2002, 25, 123−148. (21) Gevaert, O.; De Smet, F.; Kirk, E.; Van Calster, B.; Bourne, T.; Van Huffel, S.; Moreau, Y.; Timmerman, D.; De Moor, B.; Condous, G. Predicting the outcome of pregnancies of unknown location: Bayesian networks with expert prior information compared to logistic regression. Hum. Reprod. 2006, 21, 1824−1831. (22) Mallor, F.; García-Olaverri, C.; Gómez-Elvira, S.; MateoCollazas, P. Expert judgment-based risk assessment using statistical scenario analysis: A case study - Running the bulls in Pamplona (Spain). Risk Anal. 2008, 28, 1003. (23) Haapasaari, P.; Mäntyniemi, S.; Kuikka, S. Baltic Herring Fisheries Management: Stakeholder Views to Frame the Problem. Ecol. Soc. 2012, 17, 36.

informative prior probabilities, based on earlier experiences, and the provision of posterior probabilities for future impact analysis and modeling studies is required in holistic oil spill risk analysis. Bayesian models are viable solutions for such a need, and we aim for developing models that cover the accident probabilities, biological impacts,3,15 and the costs and benefits in probabilistic format.39



ASSOCIATED CONTENT

S Supporting Information *

Additional information on the Bayesian Networks, a link for the model file, and a table presenting all the model variables and their alternative state classes. This information is available free of charge via the Internet at http://pubs.acs.org/.



AUTHOR INFORMATION

Corresponding Author

*Tel.:+358 50 4150607; e-mail: annukka.lehikoinen@helsinki. fi. Notes

The authors declare no competing financial interest.



ACKNOWLEDGMENTS We thank all the interviewees and experts for their help, especially Jouko Pirttijärvi (Finnish Environment Institute), Inari Helle (University of Helsinki), Jutta Ylitalo (Aalto University), Maria Hänninen (Aalto University), and Eveliina Klemola (University of Helsinki). The study was part of the projects SAFGOFa and MIMICb financed by European Regional Development Funda,b; The Central Baltic INTERREG IV A Programme 2007-2013b; Regional Council of Kymenlaaksoa; City of Kotkaa,b; Kotka-Hamina Regional Development Company Cursora,b; Centre for Economic Development, Transport and the Environment of Southwest Finland (VARELY)b; Port of Haminaa; Finstashipa; Machine Technology Center Turku Ltda; Kotka Maritime Research Centrea; Kymenlaakso University of Applied Sciencesb; Swedish Meteorological and Hydrological Instituteb; Finnish Environment Instituteb; Tallinn University of Technologyb and University of Tartub.



REFERENCES

(1) Kuronen, J.; Lehikoinen, A.; Tapaninen, U. Maritime Transportation in the Gulf of Finland in 2007 and three Alternative Scenarios for 2015. 2009 Annual Conference of the International Association of Maritime Economists (IAME 2009) 2009; http://www. merikotka.fi/julkaisut/Kuronen_2009_Maritime.pdf. (2) Hänninen, S.; Rytkönen, J. Oil transportation and terminal development in the Gulf of Finland. VTT Publ. 2004, 547; http://vtt. fi/inf/pdf/publications/2004/P547.pdf. (3) Lecklin, T.; Ryoma, R.; Kuikka, S. A Bayesian network for analyzing biological acute and long-term impacts of an oil spill in the Gulf of Finland. Mar. Pollut. Bull. 2011, 62, 2822−2835. (4) Helsinki Commission. Restricted Use of Chemical Agents and Other Non-Mechanical Means in Oil Combating Operations in the Baltic Sea; HELCOM Recommendation 22/2; 2001. (5) Juntunen, T.; Rosqvist, T.; Rytkonen, J.; Kuikka, S. How to Model the Oil Combating Technologies and Their Impacts on Ecosystem: A Bayesian Networks Application in the Baltic Sea; ICES Council Meeting documents; International Council for the Exploration of the Sea, Palaegade 2-4 DK 1261 Copenhagen K Denmark; 2005; http://www. ices.dk/products/cmdocs/2005/s/S0205.pdf. (6) Helle, I.; Lecklin, T.; Jolma, A.; Kuikka, S. Modeling the effectiveness of oil combating from an ecological perspective - A 1798

dx.doi.org/10.1021/es303634f | Environ. Sci. Technol. 2013, 47, 1792−1799

Environmental Science & Technology

Policy Analysis

(24) Mäntyniemi, S.; Haapasaari, P.; Kuikka, S.; Parmanne, R.; Lehtiniemi, M.; Kaitaranta, J. Incorporating stakeholders’ knowledge to stock assessment: Central Baltic Herring. Can. J. Fish. Aquat. Sci. 2013, DOI: 10.1139/cjfas-2012-0316. (25) Madsen, A. L.; Jensen, F.; KjaeRulff, U. B.; Lang, M. The Hugin tool for probabilistic graphical models. Int. J. Artif. Intell. Tools 2005, 14, 507−543. (26) Kahma, K.; Pettersson, H. Wave Statistics from the Gulf of Finland (Aaltohavaintoja Suomenlahdelta 1982−1985); Finnish Institute of Marine Research; Internal Report 1/1993: Helsinki, 1993 [In Finnish with English summary]. (27) Hietala, M.; Lampela, K. Pollution Preparedness on the Open Sea − Final Report of the Working Group (in Finnish). The Finnish Environment 2007, 41. (28) Montewka, J.; Ståhlberg, K.; Seppala, T.; Kujala, P. Elements of risk analysis for collision of oil tankers. In Reliability, Risk and Safety: Back to the Future; Ale, B. J. M., Papazoglou, I. A., Zio, E., Eds.; Taylor & Francis Group: London, 2010; pp 1005−1013. (29) HELCOM. Compilation on ship accidents in the Baltic Sea area 1989−1999; 2001; http://www.helcom.fi/stc/files/Maps/accidents/ accidents1989to1999.pdf. (30) HELCOM Report on shipping accidents in the Baltic Sea area for the year 2008; 2009; http://www.helcom.fi/stc/files/shipping/ shipping_accidents_2008.pdf. (31) Lindgren, C.; Lindblom, E. Short-Term Effects of Accidental Oil Pollution in Waters of the Nordic Countries. IVL Swedish Environmental Research Institute Ltd; U964; 2004; http://www.norden.org/da/ publikationer/publikationer/2004-429/at_download/publicationfile. (32) Kallberg, V. P. Accident statistics for different modes of transport - preliminary survey (in Finnish); 1/2011; Publications of the Finnish Transport Safety Agency, 2011. (33) McDonald, A. D.; Smith, A. D. M. A tutorial on evaluating expected returns from research for fishery management using Bayes’ theorem. Nat. Resour. Model. 1997, 10, 185−216. (34) Mäntyniemi, S.; Kuikka, S.; Rahikainen, M.; Kell, L. T.; Kaitala, V. The value of information in fisheries management: North Sea herring as an example. ICES J. Mar. Sci. 2009, 66, 2278−2283. (35) Carriger, J. F.; Barron, M. G. Minimizing Risks from Spilled Oil to Ecosystem Services Using Influence Diagrams: The Deepwater Horizon Spill Response. Environ. Sci. Technol. 2011, 45, 7631−7639. (36) Goerlandt, F.; Kujala, P. Traffic simulation based ship collision probability modeling. Reliab. Eng. Syst. Saf. 2011, 96, 91−107. (37) Ihaksi, T.; Kokkonen, T.; Helle, I.; Jolma, A.; Lecklin, T.; Kuikka, S. Combining conservation value, vulnerability, and effectiveness of mitigation actions in spatial conservation decisions: An application to coastal oil spill combating. Environ. Manage. 2011, 47 (5), 802−813. (38) Jolma, A.; Lehikoinen, A.; Helle, I. In In Coupling Bayesian networks and geospatial software for environmental risk assessment; Chan, F., Marinova, D. and Anderssen, R. S., Eds.; MODSIM2011, 19th International Congress on Modelling and Simulation; Modelling and Simulation Society of Australia and New Zealand; 2011; 1881−1887; http://www.mssanz.org.au/modsim2011/E3/jolma.pdf. (39) Klemola, E.; Kuronen, J.; Kalli, J.; Arola, T.; Hänninen, M.; Lehikoinen, A.; Kuikka, S.; Kujala, P.; Tapaninen, U. A crossdisciplinary approach to minimising the risks of maritime transport in the Gulf of Finland. World Rev. Intermodal Transp. Res. 2009, 2, 343.

1799

dx.doi.org/10.1021/es303634f | Environ. Sci. Technol. 2013, 47, 1792−1799