A New Method to Assist Decision-Making of Water Environmental Emergency in Expressway Region

Water environmental emergency (WEE) in expressway region is a special kind of risk event with several characteristics, such as rarity, unconventionality, and harmfulness. The emergency decision-making (EDM) features, procedures, and methods are considerably different from the general decision-making problems. EDM quality is directly related to the timely implementation of a reasonable emergency plan. Therefore, methods should be developed to respond to emergencies immediately and scientifically and minimize the damage to water environment. This work introduces risk source identification and emergency classification and develops an emergency decision model based on scenario retrieval and case-based reasoning, according to the existing EDM model and characteristics of WEE in expressway region. The proposed method is validated through case analysis of Daguang expressway in China. This method provides an effective solution for EDM of WEEs in expressway region. The emergency measures can be implemented quickly and effectively after the occurrence of water environmental emergencies to control pollution events, provide scientific and feasible action guides for emergency processes, and enrich the case base of decision-making systems.


Introduction
During expressway construction and operation, many pollutants, such as solids, nutrients and heavy metals will be produced [1] and the surface runoff will have a negative impact on the water environment [2]. Also, most expressways are linear, and crossing some important rivers, tourist sites, and water conservation areas is inevitable. Thus, during construction and operation periods, the effect of road hazardous chemical transportation accidents on the water environment in expressway region should be given great attention. These water environmental emergencies (WEE) exhibit several characteristics, such as infrequency, timeliness and dynamics. The major water pollution emergencies in expressway region have severely damaged the water environment and caused serious pollution to drinking water sources, thereby posing a direct threat to people's health and life [3]. Therefore, it is of great practical significance to study how to make decisions and take measures to effectively deal with water environment emergencies to avoid water environment pollution and other losses [4,5].
The pollution sources of water pollution emergencies are complex, it is essential to select an appropriate disposal technology for a specific pollution scenario. Studies on water pollution accidents are mainly carried out from two aspects: simulating the spread of pollutants and assessment. He et al. developed an integrated water pollution management information system to simulate the transfer and diffusion of accidental pollutants in the river and quickly identify the affected area and how it will change over time within a few minutes of an accident occurring [6]. Rui et al. proposed an emergency response system with a hydraulic/water-quality model [7]. Ding et al. combined with two-dimensional water quality model and GIS to simulate accidents and developed an accident trend prediction for managers to deal with them, emergency decision methods need to be studied. WEE in expressway region are typical and special and need an individual study on the event disposal. Therefore, the identification and classification of risk sources is also very important in the treatment of water pollution emergencies. Such water pollution emergencies require a more comprehensive case base, which can quickly and accurately select the best plan according to the actual emergency events and assist the managers in making decisions. Moreover, The EDM of water environmental emergency in expressway region has some similarities between general EDM, but because of the infrequency, timeliness and complex risk resources, there is a need of special research on the WEE. Based on the background of WEE in expressway region, we attempt to introduce risk source identification and emergency classification into the EDM process and propose a EDM model combining scenario analysis and CBR research methods. This study aims to build a model for emergency decision-making and verify the feasibility of the model through specific case analysis, to help with water environment emergency decision making and provide reference for the development of decision support system.

Introduction of Research Methods
Prior research has performed substantial work about simulation of the spread of pollutants, assessment for emergency response plan, or methods and applications of EDM. However, former decision-making methodologies applied to other research topics are not entirely applicable to EDM research on WEE in expressway region. In addition, few studies consider the relationship between emergency classification and EDM. Therefore, this study proposes a model and introduces risk source identification and emergency classification into this model. Also, this study build a case base and conduct case studies to maintain the system learning and update. The proposed method of this study, assisting decision-making of water environmental emergency in expressway region, mainly involve scenario analysis and CBR. This work describes methods in the WEE decision-making process for addressing unconventionality and scenario-dependency of emergencies to bridge this gap.

Scenario Analysis
The description of a possible future and its corresponding path constitutes a scenario [29]. Scenarios, as an important technique of studying the future, have been used by government agencies, corporate industries and military unites as powerful tools to help in the decision-making of uncertainty for a long time [30]. Scenario analysis is an effective way to formulate strategic planning towards future [31,32]. When faced with high levels of uncertainty and complexity, scenarios analysis can help make decisions that are difficult to obtain through traditional predictions [33]. Basing on similar existing scenarios, it becomes more efficient to newly describe target scenarios. Scenario retrieval method using differential information between two scenarios was proposed, which enables the detection of similar scenarios for a given scenario [34].
Scenario analysis was first applied to military and has now been successfully applied in various areas, for example, environment [35][36][37][38][39][40], conomics [41,42], emerging technology [43], infrastructure [44] and construction project [45,46]. In recent years, scenario analysis has also been widely used in disaster risk and emergency management. Li et al. proposed that different assessment models and methods are needed to match with different stages in unconventional emergency disposals [47]. Unconventional emergency management and decision-making frameworks were developed on the basis of "scenarioresponse" [48][49][50]. Lv et al. studied emergency evacuation management and risk analysis under multiple uncertainties with a scenario-based modeling [51]. Davies et al. used a scenario-based approach to co-produce and integrate hazard and risk knowledge towards disaster resilience [52].
The WEE decision-making in expressway region is characterized by uncertainty and information asymmetry. Bayesian network is an effective theoretical model of uncertain knowledge representation and reasoning. This network can deduct with incomplete, imprecise, or indeterminate messages to support modeling and calculating complex decisionmaking problems. The network characteristics determine the applicability to solve such problems. Compared to the entire emergency, a scenario is relatively stable during a certain period of time and is not significantly rare. Therefore, this study uses scenario analysis methods, combined with Bayesian networks, to derive scenarios and obtain feature values to be prepared for CBR.

Case-Based Reasoning
Case-based reasoning (CBR) provides a new way to address the problem in the artificial intelligence domain. This approach refers to the mode of thinking and dealing with problems by referring to human beings. When encountering a problem that needs to be solved, CBR will first look for similar events in the past, then repair solutions of such time, and finally form a solution to the problem. CBR is based on approximate reasoning and knowledge [53,54]. Retrieval is a key phase in CBR because it lays the foundation for the overall effectiveness of CBR systems [55]. Bottlenecks of CBR method application are knowledge acquisition and case representation and need to be handled [56]. CBR can be translated into mature practical applications [57,58]. The research proves the evolution rules of CBR in drilling engineering and data collection of crossed applications. The popularity of CBR applications in CM research [59], such as building safety [60], and bidding risk [61], is increasing.
EDM, the main field of CBR technology, has received extensive attention from scholars, such as: (1) an emergency procurement DSS (EPDSS) on the basis of key technologies, including case representation, case retrieval algorithm, and hybrid reasoning mechanism [62]; (2) a method for emergency resource demand prediction using CBR [63]; (3) a marine oil spoil emergency response decision-making system with CBR technology [64]; (4) analysis of the response to risks connected with urban water supply network in the CBR method [65]; and (5) the use of CBR for contingency decision support for the United States Air Force [66].

Introduction of the Conceptual Model
Zhou et al. [17] explain that decision-making is a process of selecting an optimal scheme in various alternatives. The EDM process includes six stages, namely, problem definition, goal setting, project design, project selection, organization implementation, and feedback modification. The CBR approach is an analogical reasoning method. This method includes the following steps [64]: demonstration of cases to be solved, index search, and updated solutions of similar cases. Yuan and Li [49] research emergency management with scenario analysis and CBR. This study combines SR with CBR in EDM based on the work of Yuan and Li [49], according to the above-mentioned research models. Then, a method of generating an EDM plan is developed for water environmental emergencies.
The particularity of water environmental emergencies in expressway region exhibits the typical characteristics of risk decision-making. Defining risk sources of water environmental emergencies is the premise and basic work for EDM and emergency management of such emergencies. Therefore, this study proposes that risk sources should be first identified based on previous research. Taking appropriate emergency methods and measures for emergencies classification is a common experience of countries in the world when dealing with emergencies. Different types and levels of emergencies directly determine the main body, measures and resources that must be mobilized in response to emergencies. Scientific and accurate classification can avoid insufficient reaction and overreaction. In the case study and storing process, classification can help the management of the case base and ensure the accuracy of the next case reasoning. Therefore, this study introduces classification measures for emergencies.
As shown in Figure 1, the conceptual model proposed in this study includes the following processes: 1. In the risk identification stage, risk sources and their potential impacts are determined through expert interviews and questionnaires. 2. In the emergency classification stage, the dynamic fuzzy comprehensive classification algorithm is used to classify the accident, and the cluster analysis and minimum distance method are further used to classify the emergency. 3. In the scenario representation stage, the feature values of WEE were obtained by combining scene analysis and Bayesian network method. 4. In the CBR stage, the nearest-neighbor method is selected to retrieve similar scenarios from case base, and the emergency plan of the similar case scenario forms the basis of making EDM plan for WEE. world when dealing with emergencies. Different types and levels of emergencies direct-ly determine the main body, measures and resources that must be mobilized in response to emergencies. Scientific and accurate classification can avoid insufficient reaction and overreaction. In the case study and storing process, classification can help the management of the case base and ensure the accuracy of the next case reasoning. Therefore, this study introduces classification measures for emergencies.
As shown in Figure 1, the conceptual model proposed in this study includes the following processes: 1. In the risk identification stage, risk sources and their potential impacts are determined through expert interviews and questionnaires. 2. In the emergency classification stage, the dynamic fuzzy comprehensive classification algorithm is used to classify the accident, and the cluster analysis and minimum distance method are further used to classify the emergency. 3. In the scenario representation stage, the feature values of WEE were obtained by combining scene analysis and Bayesian network method. 4. In the CBR stage, the nearest-neighbor method is selected to retrieve similar scenarios from case base, and the emergency plan of the similar case scenario forms the basis of making EDM plan for WEE.

Risk Source Identification
The source of hazard is the risk source of environmental pollution, which is a prerequisite for the occurrence of pollution accidents. Understanding the risk source of water environmental emergencies objectively and accurately in expressway region not only contributes to the pre-existing control of risks but also helps to improve the efficiency of emergency rescue and the targeting of emergency response plans. This approach can minimize various effects, such as casualties, traffic disruption, and property damage. It can also narrow the scope of the disaster, and mitigate the negative effects on society.
In this work, the risk source of water environment in expressway region is defined as not only the toxic and harmful substances that can pollute the water quality but also the leakage, diffusion, and migration of the substances and environmental factors that

Risk Source Identification
The source of hazard is the risk source of environmental pollution, which is a prerequisite for the occurrence of pollution accidents. Understanding the risk source of water environmental emergencies objectively and accurately in expressway region not only contributes to the pre-existing control of risks but also helps to improve the efficiency of emergency rescue and the targeting of emergency response plans. This approach can minimize various effects, such as casualties, traffic disruption, and property damage. It can also narrow the scope of the disaster, and mitigate the negative effects on society.
In this work, the risk source of water environment in expressway region is defined as not only the toxic and harmful substances that can pollute the water quality but also the leakage, diffusion, and migration of the substances and environmental factors that can affect an area within 200 m of each side of the centerline of expressway (or 100 m upstream or 100 m downstream of the centerline when crossing the water) during the life of the expressway project. For example, in the Jinxiu event which caused the leakage of dangerous substances, a car carrying arsenic fell into the river, a large amount of arsenic was scattered on the riverside slope, partially scattered into the river, causing the pollution of water, riverside slope and vegetation.

Emergency Classification
Water environmental emergencies exhibit uncertainty and diffusivity. The toxicity and total quantity of dangerous substances are partially determined for the maximum credible accident in expressway. The classification index presents fuzziness and imprecision, and the material leakage is partially determined. Hence, the classification index of the maximum trusted accident demonstrates fuzziness and imprecision. Therefore, the selection of the classification method for WEE should present good elasticity and flexibility.
The introduction of the triangular fuzzy number method to represent the weights among classification indexes can accurately reflect the uncertainty and precision. The optimal scheme of fuzzy comprehensive evaluation is obtained according to the Kerre method. And the pessimistic decision-making principle is adopted to select the highly pessimistic value from the fuzzy value of the optimal scheme as the definite final value, which accords with the principle of minimum repentance value of the psychological behavior. The corresponding classification table determines the final level of the WEE. Then, multiple water environmental emergencies that are at the same level were clustered by the shortest distance method and subdivided.

Scenario Definition and Representation
A specific scenario of a WEE can be described from three dimensions, namely, disaster body, disaster-resistant body, and disaster-bearing body. The relationships between these three dimensions (3D) are shown in Figure 2.

Emergency Classification
Water environmental emergencies exhibit uncertainty and diffusivity. The toxicity and total quantity of dangerous substances are partially determined for the maximum credible accident in expressway. The classification index presents fuzziness and imprecision, and the material leakage is partially determined. Hence, the classification index of the maximum trusted accident demonstrates fuzziness and imprecision. Therefore, the selection of the classification method for WEE should present good elasticity and flexibility.
The introduction of the triangular fuzzy number method to represent the weights among classification indexes can accurately reflect the uncertainty and precision. The optimal scheme of fuzzy comprehensive evaluation is obtained according to the Kerre method. And the pessimistic decision-making principle is adopted to select the highly pessimistic value from the fuzzy value of the optimal scheme as the definite final value, which accords with the principle of minimum repentance value of the psychological behavior. The corresponding classification table determines the final level of the WEE. Then, multiple water environmental emergencies that are at the same level were clustered by the shortest distance method and subdivided.

Scenario Definition and Representation
A specific scenario of a WEE can be described from three dimensions, namely, disaster body, disaster-resistant body, and disaster-bearing body. The relationships between these three dimensions (3D) are shown in Figure 2.

Development Evolution
Decision Disposal

Disaster--resistant body
Disaster--bearing body Figure 2. 3D structure of the elements of water environmental emergency scenario.
The three dimensions can be expressed as shown in Formula (1): S is the WEE scenario, x is disaster body dimension, y is disaster-resistant body dimension, z is disaster-bearing body dimension, and m is the quantity of critical scenarios during the evolution. The variables m and i can be either the same or different.
The relationships between scenario elements are shown in Figure 3, and the arrows indicate influential relationships. Two types of relationships exist between the elements at one time: (1) The relationships between elements in one dimension. For example, emergency resource arrival time presents an effect on the implementation of emergency measures. These relationships belong to disaster-resistant body dimension. (2) The relationships between elements in different dimensions. The pollutant leakage quantity, as a disaster body, affects the degree to which water is contaminated (disaster-bearing body). The three dimensions can be expressed as shown in Formula (1): S is the WEE scenario, x is disaster body dimension, y is disaster-resistant body dimension, z is disaster-bearing body dimension, and m is the quantity of critical scenarios during the evolution. The variables m and i can be either the same or different.
The relationships between scenario elements are shown in Figure 3, and the arrows indicate influential relationships. Two types of relationships exist between the elements at one time: (1) The relationships between elements in one dimension. For example, emergency resource arrival time presents an effect on the implementation of emergency measures. These relationships belong to disaster-resistant body dimension. (2) The relationships between elements in different dimensions. The pollutant leakage quantity, as a disaster body, affects the degree to which water is contaminated (disaster-bearing body).

Bayesian Network Construction and Deduction
The Bayesian network of a scenario is shown in Figure 4.

Bayesian Network Construction and Deduction
The Bayesian network of a scenario is shown in Figure 4.

Bayesian Network Construction and Deduction
The Bayesian network of a scenario is shown in Figure 4.
In Figure 4, nodes E1, E2, and E3 represent different variables of the scenario. The joint probability of the variables in Figure 4 can be expressed as follows: Figure 3 provides a prototype for the Bayesian network structure of a water environmental emergency scenario from the above-mentioned analysis. Then, Bayesian network node variables along with their relationships and the conditional probability of assignment nodes are determined, as studied by Yuan [67].
The joint probability formula of a Bayesian network can be applied to deduce state probabilities of node variables in the scenario output layer after assigning prior or conditional probabilities of network node variables, as shown in Figure 5. In Figure 4, nodes E 1 , E 2 , and E 3 represent different variables of the scenario. The joint probability of the variables in Figure 4 can be expressed as follows: Figure 3 provides a prototype for the Bayesian network structure of a water environmental emergency scenario from the above-mentioned analysis. Then, Bayesian network node variables along with their relationships and the conditional probability of assignment nodes are determined, as studied by Yuan [67].
The joint probability formula of a Bayesian network can be applied to deduce state probabilities of node variables in the scenario output layer after assigning prior or conditional probabilities of network node variables, as shown in Figure 5.  After the state probabilities are obtained, with other scenario information, the scenario output layer can be the next scenario analysis layer and so on. The deduction process can be completed until it enters into the final scenario.

CBR
The WEE CBR is built on SR, and the goal is to retrieve scenarios with similar features. The nearest-neighbor method retrieval algorithm is adopted according to the fea- After the state probabilities are obtained, with other scenario information, the scenario output layer can be the next scenario analysis layer and so on. The deduction process can be completed until it enters into the final scenario.

CBR
The WEE CBR is built on SR, and the goal is to retrieve scenarios with similar features. The nearest-neighbor method retrieval algorithm is adopted according to the features and contents of WEE cases. A threshold value u is required to determine whether the case scenario in the corresponding case base is similar to the scenario to be solved. If multiple cases exhibit similarity greater than this threshold, the case scenario whose similarity value is closest to one is the most valuable case retrieved.
The realization of the retrieval algorithm involves the following four steps: determination of retrieval feature attributes, evaluation of weights of retrieval feature attributes, calculation of attribute similarity, and calculation of global similarity [68].

Case Learning and Storing
After implementing the obtained emergency response plan, the scenario and emergency response plan can be supplemented with the necessary information. And relevant feedback information can be recorded. The learning rule is used to determine whether to add the case to case base for implementing the self-learning function of the CBR system. No in-depth research of the case study process was conducted in this work. Below is a brief introduction: Machine learning is a branch of the artificial intelligence field. Any research on learning algorithms through data training belongs to machine learning, including several technologies that have been developed for many years, such as KNN, SVM, linear regression, K-means, random forest, PCA, and ANN. This work proposes the cluster analysis method to classify emergencies, which also means that this study is applicable to unsupervised learning algorithms. Hidden information in data sets can be extracted using clustering results to classify and predict future data. However, the actual system contains a considerable amount of case data, and the traditional machine learning algorithm seems to be unable to deal with big data.
Hinton made a breakthrough in the artificial neural network [68], thereby giving the neural network a "depth", set off a wave of deep learning, and provided an effective solution for big data. The main differences between deep learning and traditional machine learning lies are as follows: Traditional machine learning needs to be characterized by industry experts and then manually coded. The model rules are relatively simple. Meanwhile, a deep learning algorithm attempts to learn features from the data itself, thereby making the internal rules of deep learning model difficult to understand. In addition, deep learning heavily relies on high-end hardware facilities and long training time due to the substantial amount of computation. The top-level deep learning algorithm ResNet takes two weeks to train. Adopting traditional machine learning method at the beginning is feasible during the expressway WEE since ordinary machine learning exhibits no particularly high hardware requirement, and the training usually takes just a few seconds to hours. However, the application of deep learning algorithm is inevitable with the increase of the number of samples and data.

Description of Case Study Scenario
Concentrating on one case allows us to collect substantial data, which is important to analyze a phenomenon that is yet to be fully understood [69]. The "4.25 event" of Daguang(DG) expressway is chosen as a case study scenario. A tanker carrying 16.13 tons of petroleum rolled over near the Liuxi River water source on DG expressway. The petroleum leaked into the river, and the tanker caught fire. During handling of the accident, a rear-end collision happened to two tank trucks carrying hazardous chemical behind the former accident. One of the trucks carrying white oil and solvent naphtha burned and exploded, and the other one carrying styrene got fired and burned. Moreover, petroleum and other toxic and harmful substances leaked, then spread through the water, thereby causing serious river pollution. The entire WEE was called the "4.25 event". This incident is a typical WEE with serious hazard, huge impact, and sufficient data. Therefore, the incident is selected as a research case.

Data Collection and Analysis of Indicators
This study got the identification of potential risk events and the most credible accident grading indicators through questionnaires, interviews and expert scoring, providing a basis for case studies. After the questionnaire was collected, SPSS 17.0 statistical software was used and the Mean and standard deviation (S.D) was obtained for data analysis. When Mean > 3, the result was chosen, which means that the respondents think it's important. Reliability is an important indicator to measure the quality of a set of data. It can test the stability or reliability of the questionnaire and the inherent consistency of the survey results. Generally, coefficient value of Cranbach α is between 0 and 1. When the coefficient does not exceed 0.6, the internal consistency is insufficient; when it reaches 0.7-0.8, the scale has considerable reliability; when it reaches 0.8-0.9, the reliability of the scale is great. According to the investigation results, events that may pollute the water environment (potential risk events) are presented in Table 1.  Table 1, we can find that the potential risk events of water environment in expressway region covers design problems of the infrastructures, improper construction materials and methods, leakage from machine, equipment and vehicle, improper storage and the lack of safety awareness.
From Table 2, the Cranbach α value of this part is 0.889, and the standardized Cranbach α value of the evaluated items is 0.879. According to the Cranbach α, the reliability of the survey scale is great. Before the questionnaire, unimportant risk events had been eliminated through experts' interviews. According to the questionnaire, the Mean of each risk event is above 4 (except case 9 with Mean being 3.99), which indicates that these risk events are more likely to pollute the water environment. The results of the study on factors affecting the consequences of the most credible accidents and the grading indicators are shown in Table 3. From Table 3, after further expert scoring, it's found that the risk events that have a greater impact on the water environment include leakage substances, treatment of leakage substances, total amount of leakage substances, emergency time, emergency duration, emergency form, leakage location, water flow velocity, wind speed.
From Table 4, the Cronbach α value of this survey is 0.747, and the standardized Cronbach α value of the evaluated item is 0.747. High reliability of the survey scale is proved in terms of the Cronbach α value. The nine grading indicators were scored and evaluated by experts. Except for the "wind speed" scoring 3.88 which meant the experts had minor differences, the Means of other eight grading indicators were all above "4" points, which meant they were crucial.

Risk Source Identification
The DG expressway project passes through headwater protection areas containing the Liuxi River reservoir, Huanglong Belt reservoir, and Liuxi River, which are surrounded by rich water system, considerable mountains with slope steep, and abundant rainfall. The effect of the project on headwater protection areas can be divided into two aspects. The first one is the potential threat to the water quality of vehicles transporting hazardous goods during the operation period. The second one is the potential risk that remained due to unmet environmental protection design requirements and standards in the design period. Risk source identification is mainly carried out by expert interviews and questionnaire surveys according to the actual situation of DG. Based on the work in 3.6, the risk source identification and their potential effects on the water environment in DG expressway region combined with the actual situation of the construction environment are shown in Table 5. Dangerous goods enter the water body, pollute the river and affect water using safety Historical information, relevant engineering, video surveillance, road administration, maintenance personnel daily inspection, expressway monitoring system, alarm and other means are helpful to timely discover emergencies on the spot.

Emergency Classification
The maximum credible accident of water environment in expressway region refers to emergencies, such as leakage of dangerous substances (toxic, hazardous, and dangerous chemicals and oil) and pollution to surrounding water quality caused by a chain reaction of traffic accidents. The "4.25 event" accords with the characteristics of the maximum credible accident. The dynamic fuzzy comprehensive classification algorithm is selected to classify the "4.25" WEE. The events at the same level are further divided by cluster analysis and minimum distance method to improve the scientific value and rationality of EDM.
In this work, the classification index and weight of WEE are applied by a questionnaire survey with nine grading indexes. The various indexes include nature of the leakage substances, treatment of leakage substances, total amount of leakage substances, emergency time, emergency duration, emergency form, leakage location, water flow velocity, and wind speed. These indexes are scored by five-point scale criteria. The Euclidean distance is calculated by comparing the classification standard level in the WEE classification database.
Each grading index is scored, and finally the comprehensive score is 3.366 by the fivepoint scale method. The cluster analysis result with the case information in the database indicated that the "4.25 event" belongs to the first level of WEE, and the corresponding emergency response should be implemented.
The next case analysis and deduction only need to retrieve scenarios and disposal plans corresponding to the first type of water environmental emergencies. Then, the case with high similarity in the case base of this level is searched to generate an emergency disposal plan.

Scenario Analysis
Scenario attribute of the "4.25 event" EDM based on scenario SR is determined before calculating the similarity between the "4.25 event" real scenario and the case scenario.

Node Variable Determination and Network Construction
The scenario network of the "4.25 event" is developed from disaster, disaster-resistant, and disaster-bearing bodies based on scenario 3D representation of WEE. In summary, the scenario network node variables and values of the "4.25 event" are shown in Table 6. Table 6. Scenario network node variables and values of "4.25 event".

Index Value Range and Quantization Score
Disaster--resistant body The similarity satisfies the retrieval requirement by choosing reliability threshold u = 0.85. The process of determining the Bayesian network structure relationship is demonstrated by the determination of network causality between the "pollutant leakage quantity Y 3 " node and the remaining nodes. The basic probability assignment function m i (T, F) of the causality between Y 3 and other variables is scored by five experts. In the function, T means that causality exists between the two variables, while F indicates that no direct causality is present between the two variables. Only the probability assignment function, which the experts regard as a direct causality between the two variables, is listed in the table. The combination m (T) of the causality between Y3 and other variables in the disaster-bearing body dimension is calculated. The causality is established if m (T) is ≥0.85. Calculation results are shown in Table 7. Accordingly, m (T) = 0.21 ≤ 0.85 in Z 3 , so the causality of Z 3 is unestablished, and the variable is rejected. Therefore, disaster-bearing body dimension only exhibits two variables, namely, contaminated area Z 1 and concentration of pollutants Z 2 . The relationships among the node variables of WEE can be calculated through the above-mentioned method. Then, the Bayesian network structure of WEE scenario analysis can be obtained, as shown in Figure 6.
The values of the known variables, namely, X 1 , X 2 , Y 1 , Y 2 , Y 3 , and Y 4 , can be calculated by D-S evidence theory. Accordingly, the Bayesian network structure is obtained. As shown in Figure 6, emergency resource completeness X 1 will affect containment-diversion X 3 and pollutant diluted concentration X 4 . Then, the effect of X 3 will further influence contaminated area Z 1 . This occurrence is consistent with the actual situation wherein the contaminated area is greatly diminished when the effect of containment-diversion is satisfied. Moreover, the contaminated area is expanded when the effect is non-ideal.

Z3
0.69 0.50 0.58 0.64 0.75 0.21 Accordingly, m (T) = 0.21 ≤ 0.85 in Z3, so the causality of Z3 is unestablished, and the variable is rejected. Therefore, disaster-bearing body dimension only exhibits two variables, namely, contaminated area Z1 and concentration of pollutants Z2. The relationships among the node variables of WEE can be calculated through the above-mentioned method. Then, the Bayesian network structure of WEE scenario analysis can be obtained, as shown in Figure 6.  The values of the known variables, namely, X1, X2, Y1, Y2, Y3, and Y4, can be calculated by D-S evidence theory. Accordingly, the Bayesian network structure is obtained. As shown in Figure 6, emergency resource completeness X1 will affect containmentdiversion X3 and pollutant diluted concentration X4. Then, the effect of X3 will further influence contaminated area Z1. This occurrence is consistent with the actual situation wherein the contaminated area is greatly diminished when the effect of containmentdiversion is satisfied. Moreover, the contaminated area is expanded when the effect is non-ideal.

Scenario Network Deduction
The Bayesian network structure and conditional probability of all nodes are inputted to calculation software Hugin Lite 8.1 (HUGIN EXPERT Company; Aalborg, Denmark), and then the probability value of each node variable are obtained by running the software program.
When obtaining evidence information, emergency time is in day, emergency location is far, pollutant leakage quantity is between 325 kg and 975 kg, water velocity is >1.5 m/s, and emergency resource completeness is adequate. The 3D node variable probability can be calculated, as shown in Table 8.

Scenario Network Deduction
The Bayesian network structure and conditional probability of all nodes are inputted to calculation software Hugin Lite 8.1 (HUGIN EXPERT Company; Aalborg, Denmark), and then the probability value of each node variable are obtained by running the software program.
When obtaining evidence information, emergency time is in day, emergency location is far, pollutant leakage quantity is between 325 kg and 975 kg, water velocity is >1.5 m/s, and emergency resource completeness is adequate. The 3D node variable probability can be calculated, as shown in Table 8. The analysis result of the actual scenario indicated that the deduction result is consistent with the "4.25 event". This finding indicates that the Bayesian network-based WEE scenario analysis method is feasible and effective.

CBR
Scenario representation and case reasoning of the "4.25 event" are carried out according to scenario analysis-based CBR on the basis of scenario analysis. The principle of emergency plan selection is proven.

XML-Based WEE Case Scenario Representation
The main feature attributes of the "4.25 event" scenarios are obtained in accordance with the collection of WEE cases and above-mentioned content of scenario knowledge representation.

CBR of WEE Based on SR
The retrieved feature attributes of the "4.25 event" and case scenarios are shown in Table 9. The similarity calculation is performed according to the above-mentioned similarity algorithm between "4.25 event" and case scenarios S 1 and S 2 . The calculation example of sim(s 0 i , s 1 i ) and sim(s 0 i , s 2 i ) is followed.
Emergency location: The similarity of each retrieved feature attribute is multiplied by the corresponding weight, which is the global similarity between the emergency and the case scenarios. The weights are excluded in the scope of this study. Therefore, the weights are considered to be the same value as 1/9, and the results are: sim(s 0 i , s 1 i ) = 0.73, sim(s 0 i , s 2 i ) = 0.62. Therefore, case scenario S 1 is more relevant to "4.25 event" scenario than case scenario S 2 . Emergency plan (F s1 ) and emergency response (J s1 ) of case scenario S 1 is regarded as the basis for real-time EDM plan. From seven industry experts, the method is well suited to practical cases.

Discussion
Water environmental emergencies in expressway region exhibit typical characteristics, such as unconventionality, low frequency, evident specificity, and great harm. The emergency plan generation method based on conventional plan management demonstrates difficulties in meeting the need of EDM for such emergencies. This paper is devoted to solving the problem of EDM of water environmental emergencies in expressway region, and an emergency decision-making model is constructed by combining CBR and SR methods. This work improves the former model based on previous studies, proposes risk source identification and dynamic fuzzy comprehensive clustering classification of emergencies, and introduces it to the field of WEE in expressway region. Risk source identification, dynamic classification, EDM plan generated with scenario-based case reasoning, and EDM support system are implemented in this work to timely generate effective EDM plan.
The following results are discussed based on theoretical research and case analysis: (1) Based on literature analysis, there is a lack of EDM research on water environment in expressway region and a lack of risk source identification and emergency classification in EDM. Compared with previous studies, this research introduces risk source identification and emergency classification mechanism, which can identify the categories of emergencies more quickly and provide more accurate data for subsequent calculations. The research also has a case base and case study processes, in order to ensure the system is sustainable and the efficiency of scenario retrieval is improved. But water environmental emergencies have a certain degree of rarity and lack plenty of historical data to be verified. The prior probability needs to be assigned by experts and the error can only be reduced but not eliminated. This will affect the causality and influence the degree of scenario elements and cause errors in scenario deduction. There may be mismatches between S 0 and case scenarios in CBR, resulting in wrong results. However, due to the consistency of Bayesian posterior distribution, the reasoning will rely on the prior distribution less and less and the results will be more and more accurate with the increasing sample size.
(2) Through questionnaires, interviews and expert scoring, this study got identification of potential risk events and the most credible accident grading indicators of water environment emergency in expressway region. When an emergency occurs, emergency categories and levels can be immediately identified in the light of risk source identification list, event list, and dynamic classification of large credible incidents to subsequently provide basic information for EDM plan generation. Risk sources may have influences on each other and need further researches to complete the model. Treatment of different toxicities of hazardous chemicals should be serious in case of secondary accidents. Classifying hazardous chemicals, formulating detailed measures to deal with actual situations, and improving emergency disposal plans are recommended further.
(3) The conceptual model and method were established based on scenario analysis and CBR. For different expressway water environmental emergencies, if the risk source type changes, the method in this research can still apply to it. However, this study is based on the assumption that the risk sources are independent of each other. The research model needs to be adjusted when the mutual influence relationship between the risk sources is considered. In addition, if the scenario elements are not disaster body, disaster-resistant body, and disaster-bearing body, but disaster factors, hazard-affected body, disaster-gestating environment or others, the three-dimensional structure of water environmental emergency scenario elements may change.
(4) In the validation of method with a case study, we listed five main potential impacts on the water environment in "4.25" event and classify it, finally concluded that it belongs to the first level of WEE. We can calculate the results of node variables and generate Bayesian network structure to get the influence relationship between variables and obtain the 3D node variable probability. At last, by calculating similarity between "4.25 events" S 0 and case S 1 , S 2 , we can make a conclusion that emergency plan (F s1 ) and emergency response (J s1 ) of case scenario S 1 is regarded as the basis for real-time EDM plan. Eventually, the emergency disposal plan of "4.25 event" scenario S 0 contains information reports, on-site disposal, established field command team, equipment security, personnel protection, hazardous chemical disposal, vehicle towing, on-site cleaning, and water intake emergency treatment. This study only uses one case to verify the applicability of the model, which still needs to be further improved and validated by more cases of different sizes. And although the EDM support system plays a certain role in the case study, its function is still to be fully developed and to be applied by more case. For instance, (1) the case base still needs further classification and supplementation, and (2) the case learning algorithm needs to be developed. This task is difficult and requires a substantial amount of energy and time, but it also has profound significance. Scientific and applicable learning rules can improve the efficiency and accuracy of EDM, shorten emergency response time, achieve rational allocation of resources, and minimize the harm of WEE.

Conclusions
This study focused on the emergency decision-making for the water environmental emergencies in the expressway region, which aims to provide help for timely and effective generation of emergency plans to avoid serious environmental pollution and minimize various economic losses. The method proposed in this paper combines emergency classification with emergency decision-making and constructs an EDM model based on SR and CBR. Compared with the previous emergency decision-making models, our model is more targeted and practical in solving the problem of water environment emergency in expressway region. On the one hand, this study introduced risk source identification and emergency classification into the existing model, and obtained the indicators of risk source and event classification through questionnaires, interviews and expert scoring. On the other hand, this study used scenario analysis and CBR to establish a conceptual model. The model and methods were applied to a water environment emergency "4.25 event" in the DG expressway project to verify its validity. The study contributes to EDM in expressway water environment during the implementation stage, and provide a deeper insight to the development of a decision support system for emergency management.
Nevertheless, the study was still limited in terms of methods and model. These aspects in the EDM field can be further studied in the future: firstly, implicit and intrinsic emotional information requires an in-depth study in the scenario analysis and deduction to solve specific events. Because for specific emergencies, the scenario information includes not only explicit, external things or environment, but also some hidden and intrinsic emotional information, such as the psychological state of the affected population. Secondly, the EDM support system need be further developed through emergency management practices, such as supplementing the case base, studying the case learning algorithm and improving the case learning mechanism.