Identifying Key Nodes and Enhancing Resilience in Grain Supply Chains Under Drought Conditions

Abstract

Grain supply chains remain stable in the face of natural disasters, and the resilience of the grain supply chain plays an important role. In a complex scenario of exposure to shocks, it is significant to identify the critical nodes of the grain supply chain and propose countermeasures accordingly to enhance the resilience of the grain supply chain. In this paper’s study, firstly, a triangular model of contradictory events is used to describe complex scenarios and obtain Bayesian network nodes. Secondly, the fragmentation of the scenario is based on the description of the scene, the scene stream is constructed, the event network is obtained, and the Bayesian network structure is built on the basis. Then, combining expert knowledge and D–S evidence theory, the Bayesian network parameters are determined, and the Bayesian network model is built. Finally, the key nodes of the grain supply chain are identified in the context of the 2022 drought data in the Yangtze River Basin in China, and, accordingly, a strategy for improving the resilience of the grain supply chain is proposed in stages. This study provides a new research perspective on issues related to grain supply-chain resilience and enriches the theoretical foundation of research related to supply-chain resilience.

1. Introduction

In recent years, the frequent occurrence of unexpected events, such as geopolitical conflicts and natural disasters, has had a significant impact on global food security [1]. Grain security has become a significant concern worldwide and is becoming an increasingly important global issue [2].

Due to high temperatures and reduced rainfall, grain production in Italy has been severely reduced [3]. In addition, other extreme events, such as the worst flooding in Pakistan’s history, have caused huge losses [4]. The frequency and intensity of extreme weather events are increasing. [5]. The grain supply chain is increasingly subject to unexpected events. At the same time, over the past few decades, supply chains have become more global and fragmented [6]. In today’s highly interconnected world economy, the failure of one entity can cause the entire supply chain to collapse [7]. Therefore, research on the resilience of the grain supply chain is of great practical significance.

Supply-chain resilience focuses on unexpected events in non-routine emergencies. The concept of resilience is integrated into the supply chain to ensure supply-chain stability through corresponding resilience strategies in the event of unexpected events. In the past few years, supply-chain resilience has evolved from an emerging topic to a rapidly growing research field. The research on supply-chain resilience has mainly focused on the fundamental theoretical aspects of resilience characteristics and mechanisms, such as flexibility, redundancy, robustness, etc. [8,9,10,11]. With in-depth research on the fundamental theories of the connotation and mechanism of supply-chain resilience, the measurement of supply-chain resilience has gradually been resolved. The methods mainly focus on evaluation [12,13], mathematical modeling [14], and simulation [15,16]. The application of new technologies in the context of socioeconomic development will also enhance supply-chain resilience, such as blockchain technology [17], digitalization [18,19], and artificial intelligence [20,21,22]. The research on supply-chain resilience has gradually evolved from theory to practical application, which will be the central theme of future research.

Currently, there is a wide range of supply-chain types, and different types of supply chains face different risks, emergencies, and priorities. This also leads to differences in the application of supply-chain resilience enhancement strategies in different supply chains and the need to conduct resilience research on specific supply chains. Due to the high degree of dependence of food production on the natural environment, the grain supply chain is often hit by extreme weather. Extreme weather has a high degree of uncertainty, and the grain supply chain is also a complex socio-economic system. Together, they form a complex scenario, and resilience is inherent in it. Identifying key nodes is an important starting point for efficiently researching strategies to improve the resilience of the grain supply chain. By identifying the critical nodes of grain supply-chain resilience, resilience-enhancement strategies that are adapted to the supply chain itself can be derived.

Several scholars have already conducted research for key node identification, such as complex networks [23,24,25], data-driven-based [26,27], intelligent algorithms [28,29,30], and model analysis [31]. From the above literature, most of the identification methods for key nodes focus on complex network theory. Few scholars have studied Bayesian networks for the identification of key nodes. It is mentioned in the literature [9] that Bayesian networks can be used for supply-chain resilience-related problems in the future. Secondly, Bayesian networks are powerful tools for uncertainty knowledge representation and inference problems [32], and they have been successfully applied to problems such as risk assessment, fault diagnosis, accident analysis, and prediction [33,34,35,36,37]. The event transmission network composed of grain supply chains and emergencies contains uncertainty and is itself an accident problem, fitting the scope of Bayesian networks.

Therefore, this paper mainly addresses the following key issues:

(1) How to describe complex scenarios to obtain Bayesian network nodes.

(2) How to establish the Bayesian network model from complex scenarios.

(3) How to identify and analyze the supply chain’s key nodes (links) using the Bayesian network model.

(4) How to develop countermeasures to improve the resilience of the grain supply chain to ensure grain security.

In summary, this paper proposes scenario-driven Bayesian network modeling as the solution. Firstly, the triangular model of contradictory events describes complex scenarios, divides them into scenario fragments, and obtains Bayesian network nodes. Secondly, considering the difficulty of constructing Bayesian networks from complex scenarios, the event is first expanded into a single Bayesian network. The scene flows constructed using the scene fragments were then used to derive an events network as a basis for merging the individual Bayesian networks into an overall Bayesian network. The construction of the Bayesian network model was achieved by determining the Bayesian network parameters in combination with expert knowledge and D–S evidence theory. Finally, the Bayesian network model was analyzed using GeNIe(Academic Version 4.1) software to identify and analyze key nodes to provide appropriate decisions.

The remainder of this paper is structured as follows: Section 2 provides an overview of the literature on the resilience of the grain supply chain and methods for node identification and points out current deficiencies. In Section 3, the data and methods used in this paper are described. Section 4 builds on the previous section to validate the theory and build a Bayesian network model. Section 5 experiments with the model from the previous section and produces results. Section 6 provides a relevant discussion based on the results. Section 7 summarizes the research process of this paper and its shortcomings.

2. Literature Review

This section mainly reviews the literature related to the research. These studies provide the relevant theoretical basis for this paper. Some limitations of the existing research are also discussed. Firstly, the literature related to the grain supply chain is discussed. Secondly, it analyzes the issue of grain supply-chain resilience. In view of this subject, there are still some problems to be solved. Then, the research status of Bayesian network theory is discussed. What are the shortcomings of Bayesian network node acquisition, structure construction, and parameter determination for the problems studied in this paper? Finally, the literature is reviewed.

2.1. Grain Supply Chain

Some scholars consider the grain supply chain from the perspective of energy, sustainability, and carbon emissions. Belamkar et al. [38] studied the potential impacts on the economic viability and environmental sustainability of agrifood supply-chain networks (AFSCNs) through a multi-objective mixed-integer transport model. Deng et al. [39] explored ways to improve the environmental and economic sustainability of the wheat supply chain. Cao et al. [40] analyzed how water, energy, and carbon footprints are transferred in the food supply chain and the related issues of evolution driven by socio-economic effects. Nourbakhsh et al. [41] studied the reduction of post-harvest losses in the grain supply-chain network with the goal of reducing the total cost of the system. Some scholars also study the grain supply chain from the perspective of new technology applications such as blockchain, digitalization, and artificial intelligence. Saurabh and Dey [42] conducted research on the potential drivers of blockchain technology in the agri-food supply chain. Donaldson [43] studied the impact of the digitalization of the grain infrastructure on supply chain transparency, power distribution, and information control. Das et al. [44] explored the key factors that will enable the rapid adoption of artificial intelligence in the grain supply chain. Maheshwari et al. [45] studied how digital twins can influence the food supply chain. Zhang et al. [46] constructed a grain supply-chain information management solution based on blockchain technology to achieve information security management throughout the grain supply chain. Some scholars have also researched network optimization of the grain supply chain. Mogale et al. [47] used a multi-objective, multi-modal, multi-period mathematical model to select the location of a grain silo and optimize the grain supply-chain network.

2.2. Grain Supply-Chain Resilience

Supply-chain resilience has been a research hotspot in recent years, and many scholars have explored the resilience of the grain supply chain and achieved relevant research results. In terms of measuring supply-chain resilience, Zheng et al. [48] constructed a comprehensive evaluation system for the resilience of the grain supply chain, measured the resilience of China’s grain supply chain from 1996 to 2021, and analyzed the impact of external shocks on resilience. Some scholars also start with the factors affecting resilience. Umar and Wilson [49] studied how vertical and horizontal collaboration affects the resilience of supply chains in rural communities exposed to natural disasters. Zhao et al. [50] analyzed the factors that contributed to the resilience of the agri-food supply chain through cross-country comparative analysis and established a general explanatory structural model of the interrelations between the factors. Singh et al. [51] identified factors that enhance resilience and sustainability in the food supply chain through digital twin technology and then used the GCM method to analyze the causal factors in a ranked order. Mathematical modeling has also made considerable progress in the study of supply-chain resilience. Sharifi et al. [52] developed a multi-objective, robust, fuzzy stochastic programming model to integrate sustainability, resilience, and responsiveness in designing agrifood supply chains under uncertainty. Clavijo-Buritica et al. [53] explored methods for designing AFSCs by simulating disruptive events and using mathematical programming to find resilient designs that address uncertainty. In addition, some scholars have also used simulation modeling tools to study supply-chain resilience. Orengo Serra and Sanchez-Jauregui [54] proposed the FCS resilience model to address the resilience of the food supply chain. The model explains the transport flow of food production, processing, distribution, and retail after natural disasters. Empirical research is also an effective way to study the resilience of the grain supply chain. Burgos and Ivanov [55] studied the impact of the COVID-19 pandemic on the resilience of the food retail supply chain. Finally, there are relevant studies on the supply-chain resilience framework [56], emergency decision-making [57], and the contribution of new technologies to resilience [58].

2.3. Identification of Key Nodes in the Supply Chain

Bayesian networks are made up of nodes and links, with nodes being the important variables and links being the relationships between nodes. Bayesian networks can be reasoned about in conjunction with conditional probability tables. Firstly, the Bayesian network nodes need to be acquired, and the nodes are the basis for constructing the Bayesian network. For instance, there are node determinations using expert knowledge and relevant literature [59,60], and those that are metrics-based [61], fault-tree-based [62,63], model framework-based [64,65], and data-driven [26]. The way of obtaining nodes is richer, but the acquisition of Bayesian network nodes needs to be considered in the context of the research object to choose the appropriate method. This paper needs to construct Bayesian networks from complex scenarios, so how to get nodes from complex scenarios is the first thing that needs to be solved in this paper. Relevant information from the scenario from which Bayesian network nodes can be constructed needs to be obtained, i.e., the scenario needs to be described. At present, the relevant theories for scenario description are knowledge meta [66,67], based on case studies, etc. [68,69], ontology theory [70,71], and based on scenario response or scenario evolution [72,73]. Second, consider how the Bayesian network structure is constructed. Normally, the acquisition of nodes is closely linked to the construction of the Bayesian network structure, such as model mapping [74], metrics systems [75], fault trees [76], data-based algorithmic learning [77], and real-world-based scenarios [78]. It can be seen that most Bayesian network structures are formed by the conversion of other structures or constructed using algorithms, each of which has a certain range of applications. Finally, consider Bayesian network parameter determination. The Bayesian network parameters are determined in several ways. Learning using algorithms and thus determining the parameters [79,80] and determining Bayesian network parameters through expert knowledge [81,82]. For this paper, the above scholars built a Bayesian network model to provide the corresponding ideas. The way to build a Bayesian network is flexible and varied and needs to be combined with the specific situation to choose.

2.4. Research Gap

First, few scholars have considered research on the grain supply chain in the context of emergencies. With frequent natural disasters and ongoing geopolitical conflicts, this is a cause for concern. Second, in the current environment, resilience provides a critical research perspective. Many scholars have focused on supply-chain resilience measurement, influencing factors, digitalization, empirical research, simulation analysis, etc. However, few have paid attention to the identification of critical links in the supply chain, which is a prerequisite for improving resilience and an essential part of the research perspective on supply-chain resilience. Bayesian networks are a well-established tool for identifying critical nodes in supply chains, but they have certain limitations when applied to the object of this study. Specifically, first, the method for obtaining nodes in a Bayesian network from complex scenarios is not mature and has been studied relatively little. Conventional methods based on expert knowledge, indicators, and fault trees cannot meet the requirements. It is necessary to consider the specific object of the study, i.e., the complex scenario needs to be described and the nodes obtained on this basis. However, some current methods for scenario description mainly focus on general scenarios, are often inefficient and incomplete when applied to complex scenarios, and are unable to include important information about complex scenarios. Second, in terms of Bayesian network structure construction, Bayesian networks, as another representation of the research object, derive their nodes and structure from the same place. Therefore, the basis for constructing Bayesian network structures needs to be derived from complex scenarios. Currently, few scholars have conducted research in this area, and there are still deficiencies. Third, the determination of Bayesian network parameters. In terms of parameter learning, the scarcity of data and the shortcomings of unexpected cases undoubtedly pose a considerable obstacle. Therefore, expert knowledge is required to determine Bayesian network parameters combined with relevant methods for correction.

3. Scenario Data and Methodology

3.1. Data

This paper selects the most severe drought in the past 61 years in the Yangtze River Basin in 2022 as the research scenario [83]. The Yangtze River basin is located in southern China (24°30′~35°45′ N, 90°33′~122°25′ E), with a basin area of 1.8 million km². We combed relevant search engines from government platforms, media platforms, relevant researchers, etc., to extract scenario data that can be used to conduct research for this paper. An overview is provided below:

Early June to July saw a decrease in rainfall and rare sustained high temperatures. By early June onwards, precipitation in the Yangtze River Basin was low, with precipitation in late June being 20% less than in the same period and 40% less in July. Mainly, the lower reaches of the Yangtze River and the water system of Poyang Lake are in short supply by 50 to 70 percent. By mid-July, taking into account the safety of flood control, the reservoirs in the Yangtze River Basin had basically controlled their water levels below the flood limit before the flood season. Water had been “put down,” and reservoirs generally hold little water.

Meanwhile, Duchang County turned on the flood control work. On 21st June, the water level of Xingzi Station of Poyang Lake exceeded the 19 m warning line, and the water level had been rising slowly since then in preparation for flood control. After a few days, the water level began to fall rapidly, and after mid-July, it was receding at a rate of more than a meter a day, and the drought was progressing rapidly. By the time August rolled around, there was a sustained heatwave, and the water gap widened further during the critical water-demanding period of rice uprooting and pregnancy, as well as spike and flower production. These had a severe impact on the growth of crops, and the drought broke out under a combination of unfavorable factors. By September, the drought was gradually affecting the grain supply, and the market was volatile. In June and July, the Hunan Provincial Water Resources Department began preparing for the drought by increasing the amount of water stored in reservoirs. Grain production bases adjusted their stocks and production capacity to ensure the supply of grain, among other things. By August, the Ministry of Water Resources had launched the “Special Operation on Joint Dispatch of Reservoir Groups in the Yangtze River Basin for Drought Relief and Water Supply Protection,” with a variety of measures to protect irrigation and, at the same time, safeguard the supply of grain. By September, the grain supply chain was coordinated to stabilize the grain market. The area of failing crops was counted, subsidies were increased, severely damaged agricultural facilities were repaired, etc.

The drought has been long-lasting, widespread, and far-reaching and has had a significant impact on the grain supply chain. On this basis, the drought is also evolving, with different evolutionary aspects that are difficult for decision-makers to control. In between, this paper employs relevant theories and methods such as complex scenario description, scenario evolution, Bayesian networks, and D–S evidence theory to investigate so that relevant actors in the grain supply chain can intervene effectively at the right points to reduce the impacts of the drought.

3.2. Methodology

3.2.1. Triangular Model of Contradictory Events

The emergence, evolution, etc., of droughts cannot be entirely designed or prescribed by man. Conversely, droughts have a wide range of scenario spaces and unknown pathways of scenario formation, which often contain complexity ([84] pp. 82–83) belonging to complex scenarios. Specific scenario description methods are necessary to describe it. In emergencies such as droughts, many scholars have classified scenarios into initial, developmental, evolutionary, and end scenarios based on time [85,86]. For complex scenarios, there are limits to this division, because in the process of scenario description, the state of the scenario needs to be identified. In complex scenarios, the state of each scenario is difficult to determine, and therefore, there is a certain degree of error in the analysis process [87]. Some scholars have also classified scenarios according to their content into disaster-causing factors, disaster-bearing bodies, disaster-conceiving environments, emergency-response activities [88], etc. This division reflects the attributes of the scenario and allows for an effective description of the scenario. However, the process of scenario description should not be overly detailed, and describing scenarios in terms of pre-specified concepts poses two problems. First, a particular conceptual framework imposes certain limitations. Secondly, a complex scenario description involves heavy elements, and this method of description is not efficient [70,89]. A scenario is a collection of events, and the description of the scenario should be determined by the events contained within the scenario. When describing complex scenarios, identifying the most antagonistic and conflicting events in the scenario is a crucial step, that is, the contradictory events. Therefore, this paper proposes a ‘triangular model of contradictory events’ based on the ‘triangular model of urban public safety’ [89] to describe complex scenarios (as shown in Figure 1). Contradictions promote the development of things, and the evolution of the scenario itself is also the result of contradictions, which represent opposite attributes. For ease of distinction, contradictory events are divided into positive events and negative events. In general, events that are in line with human beings and the common good are positive events, while events that are detrimental to human development and disasters are negative events. Examples are shown in

Table 1. Positive and negative events.

Scenario	Positive Events	Negative Events
Droughts hit the grain supply chain	Water resources scheduling	Reduced rainfall, persistent high temperatures
COVID-19	vaccine development	Virus mutation, epidemic spreads
Heavy rain caused urban flooding	Flood prevention and drainage measures are activated	Heavy, persistent rain
…	…	…

.

Due to the contradictory relationship between positive events and negative events, when judging the attributes of an event, it can be determined by directly identifying whether the event is a positive event or whether it is a negative event. Scenarios contain events, so the categories of scenarios can be divided into three types according to the attributes of the contained events: disaster prevention scenarios, disaster evolution scenarios, and disaster decline scenarios, which corresponds to Figure 1.

Disaster prevention scenarios (DPS) only include positive events, such as emergency measures and safeguard policies; disaster evolution scenarios (DES) only include negative events, such as the development of hazard factors; and disaster decline scenarios (DDS) include both positive and negative events, such as the scenario of emergency response and the simultaneous effect of the disaster. The formula is described below:

In this paper, based on the formula for emergencies given in [90], combined with the triangular model of contradictory events, emergencies are described as shown in Equation (1):

$$E=(I,S,O,m)$$

(1)

Among them, $E$ is the event, $I$ is the set of input attributes of the event, $S$ is the set of state attributes of the event, $O$ is the set of output attributes of the event, and $m$ is the Positive events or negative events.

Scenarios contain events, and the description of a scenario is shown in Equation (2):

$$MS=(T,N,E,M)$$

(2)

In this formula, $MS$ denotes the scenario, $T$ is the timeframe in which the scenario occurs, $N$ is the scope of the scenario’s effects, $E$ is the events included in the scenario, and $M$ is the set of scenario attributes (DPS, DES, DDS).

3.2.2. Evolutionary Analysis of Scenarios

Overall scenarios are complex, and individual scenarios are simpler. Analyzing them by simply breaking them down into individual scenarios loses their wholeness. Therefore, scenario evolution needs to be analyzed to link scenarios organically and enable the analysis of complex scenarios. The evolution of scenarios is portrayed through event chains and scene flows.

(1) Event chain

Scenarios contain more than one event, and the connections between events form an event chain where inter-event relationships are categorized into causal and coupling relationships, as shown in Equation (3):

$$\left\{\begin{array}{c}{EC}_{ij}=(E_i,E_j,f_{ij})\\ {EC}_{ijk}=(E_i,E_j,E_k,f_{ijk})\end{array}\right.$$

(3)

In the formula, ${EC}_{ij}$ is the event chain formed by event $i$ and event $j$ having a causal relationship; ${EC}_{ijk}$ is the event chain formed by event $i$, event $j$, and event $k$ having a coupling relationship; $E_i$ is event $i$; $E_j$ is event $j$; $E_k$ is event $k$; $f_{ij}$ is the causal relationship formed by event $i$ and event $j$; and $f_{ijk}$ is the coupling relationship formed by event $i$, event $j$, and event $k$.

For $f_{ij}$, there are arbitrary $E_i$ and $E_j$, $\left({E_i\ne E}_j\right)$, satisfying the following conditions, $v_x\subset O_i,v_y\subset I_j$, and $v_x\ne \varnothing ,v_y\ne \varnothing$, with $v_x\cap v_y\ne \varnothing$, or $\exists f(v_x,v_y):v_x|\to v_y$ then $E_i$ and $E_j$ have some causal relationship. $f(v_x,v_y)$ is a mapping relationship in which $v_x$ has an influence on $v_y$ in some way, as shown in Figure 2.

For $f_{ijk}$, there is an arbitrary $E_i$ and $E_j$, $(E_i\ne E_j)$, and the following condition is satisfied, $\exists f\left(\right(v_x,v_y),v_z):(v_x,v_y)|\to v_z$, then $E_i$ has a coupling relationship with $E_j$. $f\left(\right(v_x,v_y),v_z)$ is a mapping relationship between $(v_x,v_y)$ and $v_z$ that affects in some way, as shown in Figure 3.

(2) Scene flow

Scene fragments are part of the Overall scenario. In complex scenarios, there may be different time periods with different scenarios occurring in different regions, with certain boundaries between these scenarios. The evolution of the scenarios can be understood more clearly by splitting the overall scenario and forming a scene flow. Scene fragments have the same description formula as the scenario. The constraint on scene fragments is that a scene fragment can only contain a whole, i.e., a unique event or event chain. No unrelated events or event chains can be included. When the scene fragments do not satisfy the constraints, then the unrelated $E_i$ or $E{C}_{ij}$ and $E{C}_{ijk}$ need to be split out until the constraints are satisfied.

The description of the scene flow formula is shown in Equation (4):

$$S{F}_{ij}=\left(M{S}_i,M{S}_j,s{f}_{ij}\right)$$

(4)

Among them, $S{F}_{ij}$ is the scene flow formed by scene fragments $i$ and scene fragments $j$, $M{S}_i$ is scene fragment $i$, $M{S}_j$ is scene fragment $j$, and $s{f}_{ij}$ is the relationship between scene fragments; the scene flow is shown in Figure 4.

3.2.3. Bayesian Network Structure Construction

Bayesian networks can be used as a tool for scenario extrapolation, and their structure reflects the evolution of the scenario to some extent. On the contrary, the evolutionary process of the scenario can also have a certain degree of orientation towards the construction of the Bayesian network structure, and the two are complementary to each other. In addition, the nodes of the Bayesian network are derived from the set of attributes of the events. When confronted with complex scenarios, the number of events included in the Overall scenario is high, and it is more challenging to construct Bayesian networks directly. Therefore, this paper draws on scenario orientation to guide the construction of Bayesian networks. An event is first viewed as a single Bayesian network structure, represented as follows:

Adopting the representation in the literature of [91], the input subset, state subset, and output subset of events are used as nodes in the Bayesian network, and the interrelationships between the subsets are used as directed edges to construct the Bayesian network as shown in Equation (5):

$$BN=\left(V,R,P\right)$$

(5)

Among them, $V=\left\{v_i|v_i\in I\cup S\cup O\right\}$, $R=\left\{r\left(v_x,x_y\right)|v_x,v_y\in V\right\}$, and $P=\left\{p\left(v_x|v_y\right)|v_x,v_y\in V\right\}$.

Secondly, according to the events network formed by the scene flow, the Bayesian networks of different events are merged, and finally, the complete Bayesian network structure is formed. The scenario-driven Bayesian network modeling is shown in Figure 5:

Construct an events network guided by the scene flow. The events network is then used to merge the Bayesian network. The steps are described below:

(1) Acquisition of scene fragments (As shown in

Table 2. Acquisition of scene fragments.

Step	Content	Graphical Display
1	Based on the complex scenario description, identify all the events from the Overall scenario to get the Event Overall;
2	The Overall scenario is bounded according to the time horizon and scope of action to obtain Initial scene fragments;
3	Construct possible event relationships within scene fragments, $EC1$;
4	Determine whether the events within the scene fragments satisfy the constraints;
5	Split the scene fragments that do not satisfy the constraints until the constraints are satisfied;
6	Get the scene fragments and end.

)

(2) Scene flow and events network construction (As shown in

Table 3. Scene flow and events network construction.

Step	Content	Graphical Display
1	Define the scene fragments association threshold;
2	The experts judge that the scene fragments are related to each other by scoring them according to the time and scope of action between the scene fragments and getting the relationship matrix between the scene fragments;
3	Deletion of scene fragments relationships below the scene fragments association threshold to get the initial scene flow;
4	The remaining scene fragments relationships are analyzed to determine whether the events contained between the scene fragments have a relationship, and if they do, they are retained, and if they do not, the original relationship is deleted;
5	At this point, the event relationship $EC2$ between the scene flow SF and the scene fragments is obtained;
6	Add $EC2$ to $EC1$ to obtain $EC$, the preliminary events network;
7	Judge whether it matches the reality based on the scene flow and events network; If it does not match, return to Step 1 to adjust the scene fragments association threshold. If it matches, then get the event’s network and end.

)

The Overall Bayesian network structure is more difficult to construct, and the individual Bayesian network structures that make up the totality are simpler. The events themselves have some orientation to the Bayesian network structure, such as the input, state, and output attributes of the events. Therefore, the Bayesian network of individual events can be constructed by combining expert knowledge and attribute interrelationship table. Then, Bayesian network merging is performed based on the event’s network with the following steps:

(3) Bayesian network merger (As shown in

Table 4. Bayesian network merger.

Step	Content	Graphical Display
1	All events are represented as $B{N}_i$ according to the Bayesian network representation;
2	If $EC\ne 0$, go to Step 3, otherwise go to Step 6;
3	Read $EC$, and judge the specific connecting relationship. If it is causal, then go to Step 4, coupling relationship; go to Step 5, and delete the relationship in $EC$;
4	Perform causal connection for $B{N}_m$ and $B{N}_n$: for $v_x\in O_i$ and $v_y\in I_j$, merge $v_x$ with $v_y$ when $v_x=v_y$; when $v_x\ne v_y$, then construct the directed edge $r(v_x,v_y)$ based on the causal connection. Move to Step 2;
5	Perform coupling connection for $B{N}_m$, $B{N}_n$, and $B{N}_k$: for $v_x\in O_i$, $v_y\in I_j$, and $v_z\in I_k$, when $v_x\ne v_y$, construct $r(v_x,v_z)$ and $r(v_y,v_z)$ according to the coupling connection; when $v_x=v_y$, merge $v_x$ with $v_y$ and then construct the directed edge $r(v_x,v_z)$. Go to Step 2;
6	The events network Bayesian correlation is complete, end.

)

3.2.4. D–S Evidence Theory

Often, experts can only give more abstract quantifiers that describe the strength of the causal relationship, etc., in the context of their own experience, for example, “high impact” and “may be relevant”. No definite probability can be given. However, there is a certain correspondence between this strength and the actual probability, which can be illustrated by a probability marker (as shown in Figure 6) [92]. But this is a rougher probability, and the accuracy needs to be corrected. D–S evidence theory allows for the fusion of expert empirical information, which in turn diminishes the influence of subjective factors when experts make judgments [93].

Assume that the two independent trust functions in D–S evidence theory are $p_1$ and $p_2$, and that the mess functions of this trust function are expressed as $m_1$ and $m_2$. The focal elements $A$ and $B$ of the trust function are, respectively, $A_1,A_2,A_3,\dots ,A_i,B_1,B_2,B_3,\dots B_j$. The fusion formula for D–S evidence theory is shown in Equation (6):

$$m_{DS}(A)=\left\{\begin{array}{c}{\displaystyle \frac{{\sum }_{A_i\cap B_j=A}{m}_1\left(A_i\right)m_2\left(B_i\right)}{1-{\sum }_{A_i\cap B_j=\varnothing }{m}_1\left(A_i\right)m_2\left(B_i\right)}},A\ne \varnothing \hfill \\ 0,A=\varnothing \hfill \end{array}\right.$$

(6)

If there exist three independent states $A,B,C$ for a node variable, there are i existing experts who assign values to each of the above three state probabilities according to the probability scale, $\left\{m_1\right(A),m_1(B),m_1(C\left)\right\}$, $\left\{m_2\right(A),m_2(B),m_2(C\left)\right\}$, $\left\{m_i\right(A),m_i(B),m_i(C\left)\right\}$. The value of the fusion probability of the three states $A,B,C$ can be found by using the above equation when there is more than one piece of evidence, because the Dempster synthesis rule has the properties of exchange and union laws. Thus, the synthesis of multiple pieces of evidence can be transformed into a two-by-two synthesis of multiple pieces of evidence, and the final result is independent of the order in which each piece of evidence is involved in the synthesis.

4. Calculation

4.1. Bayesian Network Node

Drought emergencies last a long time and have many scenarios. It is necessary to find contradictory events and identify positive events and negative events. At the same time, in order to improve the accuracy of modeling, by combining the drought data of the Yangtze River Basin in recent years as a sample, key variables are selected analytically to describe the events, see Supplementary Material File S1: Event Description.

4.2. Bayesian Network Structure

Expert knowledge is required when constructing a Bayesian network. This paper invited seven experienced experts in the fields of logistics and supply-chain management (PhD), emergency decision-making (PhD), grain supply chains (PhD), government personnel (with extensive experience in drought relief), and large grain growers (with many years of experience in grain farming). This paper finally determined 30 detailed and valid drought reports as sample data on droughts by statistically screening drought reports in recent years. These include the large-scale spring and summer drought in the north in 2000, the spring and summer drought in the north in 2001, the large-scale and severe autumn drought in southern China and the middle and lower reaches of the Yangtze River in 2004, the Yunnan drought in 2010, the drought in the middle and lower reaches of the Yangtze River in 2011, the drought in Guizhou in 2013, the drought in the Yangtze River basin in 2022, and other typical droughts (as shown in

Table 5. Historic drought data.

{Title}-{Source}	{Date Type}-{Time}-{Location}	General Description
{2001 Northern Spring and Summer Drought}-{China Meteorological Administration, China News Network}	{Text}-{2001-February-June }-{most of the area north of the Yangtze River}	From February to early June, most of the area north of the Yangtze River had abnormally low precipitation, generally high temperatures, and high evaporation. Farmland lost moisture quickly, and a large-scale and persistent drought occurred. Due to years of drought and lack of rain, water conservancy projects did not have enough water storage, and groundwater was not significantly replenished, leading to a worsening water shortage.
…	…	…

), as well as some other reports that provide descriptions of droughts. Based on the sample drought data, a statistical analysis of the interactions (causal relationships and coupling relationships) between the attributes of drought events was carried out and corrected using expert knowledge. A correlation coefficient table was obtained (see Supplementary Material File S2: Attribute Correlation Table for details).

In this paper, only attributes with correlation coefficients above 0.5 were selected for correlation. At the same time, the Overall scenario is divided into scene fragments to obtain initial scene fragments and identify causal and coupling relationships to analyze event relationships within a scenario. The final scene fragments are shown in

Table 6. Scene fragments.

Num	Scene Fragments	T	N	E	Attributes
NO.1	Sharp decrease in rainfall	June	Yangtze River Valley	E1	DES
NO.2	Preventive work	July	Yangtze River Valley	E6	DPS
NO.3	The main flood season is anti-desiccation; the situation is not optimistic	Late July	Reservoirs in the Yangtze River Basin	E2 E3	DES
NO.4	Drought relief operations	August	Grain supply chain	E4 E5 E7	DDS
NO.5	Grain supply-chain recovery	August–September	Grain supply chain	E8	DPS

:

The expert scored the correlation assessment between the scene fragments by T and N to obtain a correlation matrix, as shown in Equation (7):

$$\left[\begin{array}{ccccc}0& 0.74& 0.7& 0.72& 0.1\\ 0.16& 0& 0.08& 0.82& 0.1\\ 0.12& 0.72& 0& 0.76& 0.24\\ 0.02& 0.1& 0.04& 0& 0.8\\ 0& 0& 0& 0& 0\end{array}\right]$$

(7)

From left to right (top to bottom) are NO.1, NO.2, NO.3, NO.4, and NO.5. The scenario association threshold is 0.5; i.e., above 0.5 it is considered to be associated. From this, a preliminary scene flow is constructed, followed by judging the event relationships between the scene fragments, thus determining the scene flow, as shown in Figure 7:

After the scene flow was acquired, the Bayesian networks were merged from the event network using the method provided in Table 4. Due to space constraints, the merging of the E₂ and E₃ Bayesian networks is shown in Figure 8.

The process of merging all relationships is shown in Supplementary Material File S3: Merger process, and the final result is shown in Figure 9.

The same output and input variables may appear during the merge. When the same input variable occurs in different networks, the same nodes should be merged first, and the impact of the variable on different subsequent nodes under different events (scenarios) should be considered separately when the probabilities are determined. For identical output variables, it is necessary to combine identical variables in order to consider the total amount of impact caused.

5. Results

A node may simultaneously belong to the attributes of different emergencies. For ease of expression, it is uniquely represented by $S\left(i\right),i=1,2,3\dots$. The accuracy of the parameters in a Bayesian network determines whether Bayesian inference is realistic. In this paper, the initial conditional probabilities are first determined using experts combined with probability benchmarks (see Supplementary Material File S4: Initial Condition Probability Table for details). The probabilities were subsequently corrected using evidence theory, using Equation (4) to obtain the D–S evidence theory corrected probabilities (see Supplementary Material File S5: Modified Conditional Probability Table for details). The specific steps are as follows:

A total of seven experts made judgments. Based on the (as shown in Figure 10) node relationships, it is now necessary to determine the conditional probability of S3. Then, the initial conditional probability is obtained by the expert’s assessment of the node state based on his knowledge combined with the probability benchmark. It is shown in

Table 7. Initial probability table.

Status Node	Conditional Node	Ex1	Ex2	Ex3	Ex4	Ex5	Ex6	Ex7
Climate S3-T	S1-seriousness, S2-seriousness,	0.35	0.35	0.5	0.75	0.25	0.5	0.5
	S1-seriousness, S2-not_serious,	0.35	0.5	0.5	0.75	0.35	0.35	0.5
	S1-not_serious, S2-seriousness,	0.35	0.5	0.5	0.5	0.35	0.25	0.75
	S1-not_serious, S2-not_serious,	0.35	0.35	0.75	0.5	0.75	0.35	0.35

.

Ex1 to Ex7 are seven experts, respectively. After obtaining the initial probability table, we synthesized sequentially from Ex1 onwards. The post-fusion probabilities were calculated according to Equation (6) and are shown in

Table 8. D–S evidence theory synthesis process.

Expert	T	F
Ex1	0.35	0.65
Ex2	0.35	0.65
Ex1&Ex2	0.224771	0.775229358
Ex3	0.5	0.5
Ex1&Ex2&Ex3	0.224771	0.775229358
Ex4	0.75	0.25
Ex1&Ex2&Ex3&Ex4	0.46519	0.534810127
Ex5	0.25	0.75
Ex1&Ex2&Ex3&Ex4&Ex5	0.224771	0.775229358
Ex6	0.5	0.5
Ex1&Ex2&Ex3&Ex4&Ex5&Ex6	0.224771	0.775229358
Ex7	0.5	0.5
the corrected value	0.224771	0.775229358

.

The conditional probabilities for the final determination of S3 are shown in Figure 11:

In this paper, probabilistic inference was performed using GeNIe software, and the results are shown in Figure 12.

The probability that the market is stable can be derived from forward reasoning, which is 76%. The drought in the Yangtze River Basin has lasted for a long time and had a great impact on the economy and society to a certain extent. The figure of 76% is between what can be expected and what is possible, which is consistent with reality and indicates that the constructed Bayesian network is accurate. The evolution of an emergency can be determined from the scene flow. Disaster scenarios begin with a decrease in rainfall, followed immediately by a new disaster scenario as the situation begins to deteriorate. At the same time, the supply-chain-related entities carried out certain preventive measures. Then, the drought broke out, and the supply-chain subjects fought against it. When the drought subsided, recovery of the grain supply chain was initiated to ensure its stability. The evolution of drought shows that it first appears, then accumulates, builds to an outbreak, and finally disappears. The behavior of the grain supply-chain subjects then goes through a process of prevention, confrontation, and recovery.

6. Discussion

Based on the analysis of the above results, the process of complex scenario evolution is discussed in three phases: the prevention phase, the antagonistic phase, and the recovery phase. The prevention phase mainly involves preventing an impending drought, thereby reducing the initial impact of the drought. The confrontation phase indicates that the drought has had a significant impact on the grain supply chain, with grain being seriously affected and crop failure further affecting the grain supply process. The grain supply chain confronts the drought through a variety of measures. The recovery phase is when the stability of all entities in the grain supply chain decreases after the impact of the drought, which is not conducive to the sustainable development of the grain supply chain and needs to be restored. These three stages correspond to the process of drought impacting the grain supply chain. Each stage has a corresponding objective, which is analyzed below.

6.1. Prevention Phase

The first stage is the prevention phase, which accurately identifies the disaster. Appropriate preventive measures correspond to the target nodes, including S20 availability of groundwater, S21 reservoir capacity, S22 river available water, S23 grain stocks, and production capacity adjustment and advance. GeNIe performed a sensitivity analysis, and fragments were captured. Results are shown in Figure 13.

A redder color means that the node is more sensitive. Change the target node to 100% and record the difference of the parent node. A larger difference indicates a greater impact on the target node. The results are shown in Figure 14.

It can be found that the node of implementation of preventive measures significantly influences the prevention phase objectives. The key node provided the maximum difference. This was followed by the decrease in reservoir water storage capacity and the decline in river level decline. On the contrary, it is found that the decrease in reservoir capacity is smaller than other changes because the state of the decrease in the reservoir capacity is easy to grasp. Its primary function is to influence the degree of implementation of preventive measures. The drought situation can be evaluated based on changes in the amount of water in the reservoir. Since implementing preventive measures affects subsequent nodes along the network, the reservoir’s storage capacity alone has little influence. Besides, the average rainfall is not the only factor that decreases the reservoir water storage capacity in the network. After tracing the source, it was found that opening the gate and releasing the water were the main factors because the drought forecast is a significant difficulty. Additionally, the abnormal situation of the main flood season in 2022 is reversed. As a result, the water in the reservoir has been artificially drained, and the situation has reversed, thus reducing the water storage. The river level decline dropping node also has a significant effect because farmers’ direct irrigation water mainly comes from rivers, which increases its sensitivity to river levels. When an exception occurs, appropriate measures, such as drainage channels, are taken to rectify the fault. The reservoir capacity status in the prevention phase indicates the drought development. Specific preventive measures are mainly taken according to this sign. Besides, the forecast level of disaster will profoundly affect the water storage state. Since droughts and floods are related to this quantity, coordination is difficult.

6.2. Antagonistic Phase

The second stage is the antagonistic phase, where its target is the S29 grain supply. This node is set as the target node for sensitivity analysis. The clip was intercepted, and the results are shown in Figure 15.

Sensitivity analysis indicates that the reserve grain supply is the key node. The grain supply has three main components: the salvage area, the reserve grain supply, and the grain production. The salvage area mainly occurs in the producer and is affected by the degree of implementation of the overall rescue plan. The better the rescue plan implementation, the more the salvage area and supply. The grain supply is also related to the reserve grain supply. Stockpiling grain is mainly a government matter. The government should dispatch supplies immediately to ensure a stable grain supply. Finally, grain production can be scheduled. Node enterprises in the grain supply chain can schedule inventory to ensure grain supply stability in disaster-stricken areas. This requires planning and adjustment ahead of time. The reserve grain supply is darker and has a more significant impact. The government plays an essential role in the grain-supply process during the antagonistic phase. At this stage, attention should be paid to the causes of the drought outbreak. The node of the crop disaster area was changed to 100%, and the parent node variations were observed, as shown in Figure 16. In the first place, the decrease in irrigation water is a significant reason for the outbreak of drought, followed by the high temperature and the increase in the grain water requirement. The reduction in water volume was the main output of the previous phase. The causes of high temperatures are divided into two aspects. One is the climate, and the other is that the decrease in water volume decreases the humidness of the air, exacerbating the occurrence of high temperatures. The increase in the grain water requirement is also an essential cause of the crop drought. Since crops are in a growing cycle that depends on adequate water, this drought is more severe, indicating eruption signs, judged by the cause of the outbreak. The internal cause is the increase in water demand, while the external cause is the decrease in water resources and high temperatures. Most of the external causes come from the previous stage. The direct and indirect conduction effects are water resources and high temperatures, respectively. It is a by-product of the first stage.

6.3. Recovery Phase

The recovery phase aims to realize the S39 (market stability) and S41 (the future of the grain supply chain is stable). Sensitivity analysis was performed similarly, as shown in Figure 17. There is a progressive relationship between the target nodes, and the target nodes of the lower level are focused first. It can be found that the stabilization level of the affected people is the key node. People’s consumption interacts with the market, and grain is a daily necessity. When the grain supply fluctuates, the affected people become dissatisfied, thus affecting the market stability. At the same time, network analysis indicates that the market stability mainly comes from three aspects: the producer’s stability level, the stability of the affected people, and the supply-chain intermediate node enterprise stability level. The area of failing crops and the stability measures of the producers, such as agricultural insurance and grant subsidies, determine the stability factors of the producers. The influence of the area of the failing crop and stabilization measures determine the degree of stability. When it is unstable, it will cause many growers to withdraw from the producer field, thus affecting the market stability of the grain supply chain. Intermediate enterprises in this emergency mainly play the role of adjusting inventory. The emergency adjustment will have a specific cost requiring some compensation measures. In contrast, more and more people will quit the industry when there is instability.

Now, the future stability of the S41 grain supply chain should be verified. The future stability of the grain supply chain refers to its absorptive capacity and resilience after an emergency. Market stability is the absorption performance. The stability of the grain supply chain will depend on the future stability of the supply chain, which is a key node affecting future stability. The degree of disaster repair will also be affected, as indicated by the ability to recover. Losses are bound to follow events. Repair work, such as some infrastructure damage, water resources retention facilities repair, and reconstruction, will affect the future stability of the grain supply chain.

6.4. Overall Discussion

In summary, the following steps are necessary to analyze the influence of different organizations at different stages. First, the nodes that map the relevant measures should be found. Second, the main organizations of node mapping should be distinguished. Finally, the directly connected sub-nodes below this node should be set as target nodes for sensitivity analysis. The degree of role of different organizations in the grain supply chain at different stages can be obtained using the sensitivity value as the data source.

Table 9. Statistics on the role degree of different organizations in the grain supply chain at different stages.

Phase	Mapping Node	Designation	Sensitivity	Sensitivity Ratio	Major Organization
Prevention phase	Availability of groundwater	S20	0.244	24.40%	Water resources management organization
	Reservoir capacity	S21	0.272	27.22%	Water resources management organization
	River availability	S22	0.261	26.07%	Producer
	Grain stocks, production capacity adjustment, and advance	S23	0.223	22.31%	Intermediate node enterprise
Antagonistic phase	Salvage area	S28	0.625	62.50%	Water resources management organization
	Reserve grain supply	S27	0.250	25.00%	Government department
	Grain production can be scheduled	S26	0.125	12.50%	Intermediate node enterprise
Recovery phase	Market stability	S39	0.667	66.67%	Government department
	Degree of disaster recovery	S40	0.333	33.33%	Government department

describes the cumulative proportion of the sensitivity of different organizations at each stage, as plotted in Figure 18.

Figure 18 shows that the water resources management organization plays a significant role in the prevention phase. The producer and the intermediate node enterprise play a secondary role. The water resources management organization and antagonistic phase play the main roles. The intermediate node enterprise plays a secondary role, while the production side has a negligible effect on the sudden drought event. After the recovery phase, the government mainly performs the recovery.

Further analysis indicates that the prevention, antagonistic, and recovery phases are closely related and progressive. The events of the previous stage will directly or indirectly affect the next stage. Indirect effects are easy to ignore and more critical. For example, the network’s S15 (humidness) and S23 (grain stocks, production capacity adjustment, and advance) are indirectly affected nodes. The prevention phase potentially affects subsequent-stage events.

6.5. Insights and Practical Implications

Multiple factors often cause instability in the grain supply chain during emergencies. The grain supply chain is a complex system in which different factors influence and interact with each other in space and time. At the same time, measures and factors related to the resilience improvement of the grain supply chain are consistent in pace. This paper divides the complex scenarios of the grain supply-chain resilience into prevention, antagonistic, and recovery stages, which were analyzed, respectively. Combined with relevant studies, the following toughness improvement strategies can be obtained:

(1) Prevention phase. It should be performed in the early days of drought in the grain supply chain. The water resources management organization plays an essential role in improving the grain supply-chain resilience, disaster monitoring capacity, and drought forecasting system. Then, producers should improve agricultural facilities, such as irrigation, to improve water utilization. At the same time, relevant instruments should be employed to monitor the growth of crops. The intermediate node organization shall pay timely attention to the production and inventory situation, thus making relevant predictions and realizing resource scheduling.

(2) Antagonistic phase. This phase should be performed during a drought in the grain supply chain. Water resources management organizations and government departments dominate the expression of the resilience of the grain supply chain. An intermediate node enterprise occupies a secondary position. Improving grain supply-chain resilience first targets water resources management organizations. Regional water resources should be reasonably evaluated, and an improved water resources dispatching scheme should be established using digital twin technology. Coordinating the use of domestic, agricultural, and industrial waters is necessary. Government departments should improve the information exchange of granaries and break the information barriers. Combined with cloud storage technology, the reserve grain supply should be ensured. Second, optimizing the grain inventory model of the intermediate node enterprise is necessary. Advanced scheduling reduces costs and ensures the sustainability of the grain supply. Although the producer cannot play a role at this stage, related problems have emerged. Producers should consider local resources, such as water resources and climate. At the same time, drought-resistant seeds should be employed, and planting methods should be improved to avoid widespread water shortage.

(3) Recovery phase. This is the final stage of a drought in the grain supply chain. The government department dominates the expression of the grain supply-chain resilience to improve the disaster emergency resource allocation system and ensure the fairness of material distribution. The grain supply-chain risk-transfer channels, such as third-party grain insurance, are improved. The land water retention capacity is strengthened using wild ponds, waste wells, and other water storage. Timely repair of damaged agricultural facilities, improving post-disaster policies for intermediate node enterprises, and enhancing the resilience of the grain supply chain are some tasks in this phase.

In short, improving the resilience of the grain supply chain should not be limited to a specific time, space, or organization. This paper verifies the whole process of the grain supply chain, finds out the key “nodes”, and employs system thinking to solve the problem.

6.6. Comparative Analysis

Identifying the key nodes that affect the resilience of the grain supply chain from complex scenarios is a key step in improving the resilience of the grain supply chain. Due to the complexity of resilience issues, constructing a Bayesian network from complex scenarios and identifying key nodes present certain difficulties. Therefore, this paper mainly makes the following improvements to the methodology compared with existing research. First, the method of describing complex scenarios. This paper proposes a triangular model of contradictory events to describe complex scenarios. Different scholars describe scenarios in different ways, and the main difference lies in the different research objectives and objects, which lead to different description methods, as shown in

Table 10. Summary of the scenario method.

Emergency	Description	Year	Authors
Construction safety accident	Scenario State (S); Human Behaviors (H); Hazard-formative Environments (E); Emergency Management (M).	2024	She et al. [94]
Collision accidents	Hazard-forming environments (E); Hazard-affected bodies (B); Hazard-formative factors (F); Accident states (S).	2024	Zhang et al. [95]
Food quality and safety emergencies	Situation state (S); Emergency objectives (T); Emergency decisions (D); External environments (E).	2023	Li et al. [88]
Torrential rain	Factor (F); Bearing (B); Incident (I); Response (R); Environment (E); Casualties (C).	2021	Jiang et al. [96]
Urban flood and waterlog disaster	Disaster-pregnant environment (E); Disaster-causing factors (H); Disaster-bearing carrier (C); Emergency management (M).	2021	Wang et al. [97]

.

Contradictions are the driving force behind the development of events. The research objectives and goals are different due to different unexpected events. If specific descriptions are set according to these, it will limit the scope of application and reduce scalability. Since contradictions accompany the development of events, the description method in this paper is more applicable than other scholars’ description methods and can describe different unexpected events. Secondly, in the process of constructing Bayesian networks, this paper fully considers the guiding role of scenarios in constructing the structure of Bayesian networks. The proposed scenario-driven Bayesian network construction method has a broader application range than general construction methods such as index systems [98] and accident trees [99]. Finally, in the Bayesian network analysis, this paper identifies and analyzes the key nodes affecting the resilience of the grain supply chain in stages based on the impact process of drought on the grain supply chain. Since complex scenarios have multiple evolution processes, this approach can effectively mine information from complex scenarios. However, Bayesian network analysis in stages is still relatively rare in current research.

The relevant results of the Bayesian network analysis are divided into stages. For example, in the prevention phase, this paper emphasizes the need to improve drought-monitoring capabilities in a timely manner and take early preventive measures to reduce the impact of drought. This is consistent with Wu et al. [100], who believe that avoiding risks is more effective than disaster relief in reducing the impact of disasters. In the confrontation phase, we found that producers play a relatively minor role. Mishra et al. [101] believe that producers can adapt to drought conditions, such as adopting water-saving technologies and alternative water sources, but in the case of severe droughts, they will suffer significant losses. This is somewhat different from the conclusions of this paper, but there are also connections. Drought resistance is a long-term process, and producers have a role to play in the early stages when droughts are developing and have not yet reached an explosive state. However, as the drought progresses and becomes more severe, the capacity of producers gradually becomes exhausted. At this point, the confrontation phase has begun, and producers need to rely on external capabilities to fight back. This paper provides a more detailed analysis than the results of existing literature. It shows the changing role of producers in the drought-resistance process from the perspective of the development process of drought. In the recovery phase, this paper not only emphasizes the recovery of producers, intermediate enterprises, and affected people but also emphasizes the need to enhance local water-retention capacity and maintain and update damaged agricultural facilities. This will help to improve the sustainability of the grain supply chain and increase its future resilience. Holman et al. [102] argue that there is a need to move from a response (short-term and reactive) to an adaptation (long-term and anticipatory) strategy to improve the resilience of agriculture to future droughts. They also suggest that drought experiences should be used to develop drought-coping strategies to improve resilience and environmental adaptability. This is similar to the conclusions of this paper but from a different perspective. This paper focuses on the recovery of the grain supply chain and emphasizes improving the stability of each entity to cope with future risks. In addition, measures such as increasing local water-storage capacity reflect the process of learning from drought experience and adapting to the environment.

Overall, Vicario et al. [103] point out the lack of a true supply-chain perspective (from production, processing, distribution, and marketing to consumption) in understanding drought resilience and the need to urgently focus on the drought resilience of the entire chain. This paper, as a whole, includes all perspectives of the supply chain, such as production organizations (production), intermediate node enterprises (processing, distribution, and marketing), and local people (consumption). Third-party perspectives, such as those of water-resource scheduling organizations and government departments, have also been added. A comprehensive analysis of the role played by these organizations in coping with drought from a systems perspective also reflects the integrity of the analysis process.

7. Conclusions

This paper proposes a triangular model of contradictory events to describe complex scenarios and obtain Bayesian network nodes. The Bayesian network structure is then constructed by scene flow. The Bayesian network parameters are determined based on expert knowledge and evidence theory. Finally, the key nodes are identified based on the sensitivity analysis of the GeNIe software. The relevant results are analyzed, and the promotion strategy is obtained. In addition, methods are provided for constructing Bayesian network models from complex scenarios. This is scenario-driven Bayesian network modeling. The proposed triangular model of contradictory events is based on the contradictory events. Since contradiction is the source of the event development, the complex scenarios can include all kinds of emergencies, complex systems, and major accidents, which can be employed for relevant applications. With the continuous development of society, people should deal with more complex phenomena. A complex system has a complex integrity. In complex scenarios, consider contradictory events to reduce complexity. When analyzing models, consider different stages and different organizations to restore reality. The Bayesian network mode is first constructed locally and restored according to the scenario. It embodies the complexity degradation principle ([84] pp. 135–138). Holism should be considered when using reductionism to degrade complexity. After achieving the goal, holism is utilized to restore, which is the application of the complexity degradation principle. The application process of this paper can be extended to analyze various complex problems.

Theoretically, although this paper has achieved the description and analysis of complex scenarios, the relevant theory used (the urban public safety triangle model) is only an improvement on the existing framework. As society continues to develop and the complexity of scenarios gradually increases, there is a need to deepen the existing theory in order to gradually adapt it to the application of complex scenarios. Methodologically, firstly, although this paper adopts a subjective approach to describe complex scenarios when analyzing them, which improves efficiency, it still relies on manual judgment. The amount of data contained in complex scenarios is huge; secondly, when constructing the Bayesian network model in this paper, the model is static and based on unexpected events or scenarios that have already occurred. In reality, real-time analysis of unexpected events will be of greater practical significance. Therefore, the future research direction of this paper will focus on considering the use of methods such as neural networks and deep learning to describe complex scenarios and considering the use of dynamic Bayesian network models to be applied to the analysis process of complex scenarios to achieve dynamic analysis and deduction of complex scenarios.