**A Model-Based Engineering Methodology and Architecture for Resilience in Systems-of-Systems: A Case of Water Supply Resilience to Flooding**

#### **Demetrios Joannou 1,\*, Roy Kalawsky <sup>1</sup> , Sara Saravi <sup>1</sup> , Mónica Rivas Casado <sup>2</sup> , Guangtao Fu <sup>3</sup> and Fanlin Meng <sup>3</sup>**


Received: 5 February 2019; Accepted: 3 March 2019; Published: 8 March 2019

**Abstract:** There is a clear and evident requirement for a conscious effort to be made towards a resilient water system-of-systems (SoS) within the UK, in terms of both supply and flooding. The impact of flooding goes beyond the immediately obvious socio-aspects of disruption, cascading and affecting a wide range of connected systems. The issues caused by flooding need to be treated in a fashion which adopts an SoS approach to evaluate the risks associated with interconnected systems and to assess resilience against flooding from various perspectives. Changes in climate result in deviations in frequency and intensity of precipitation; variations in annual patterns make planning and management for resilience more challenging. This article presents a verified model-based system engineering methodology for decision-makers in the water sector to holistically, and systematically implement resilience within the water context, specifically focusing on effects of flooding on water supply. A novel resilience viewpoint has been created which is solely focused on the resilience aspects of architecture that is presented within this paper. Systems architecture modelling forms the basis of the methodology and includes an innovative resilience viewpoint to help evaluate current SoS resilience, and to design for future resilient states. Architecting for resilience, and subsequently simulating designs, is seen as the solution to successfully ensuring system performance does not suffer, and systems continue to function at the desired levels of operability. The case study presented within this paper demonstrates the application of the SoS resilience methodology on water supply networks in times of flooding, highlighting how such a methodology can be used for approaching resilience in the water sector from an SoS perspective. The methodology highlights where resilience improvements are necessary and also provides a process where architecture solutions can be proposed and tested.

**Keywords:** architecture modelling flood resilience; resilience engineering; system-of-systems water systems

#### **1. Introduction**

Water is essential to people, businesses and the environment. The sustainability of water resources is fundamental to life as we know it and access to clean water is essential to society. A generalised depiction (rich picture [1]) of the water "system-of-systems" (SoS) can be seen in Figure 1, which illustrates the complexity and high-level connectivity between stakeholders and

constituent systems. An SoS is a large set of interconnected systems which collaborate to bring about behaviours and capabilities greater than the sum of the individual parts.

**Figure 1.** Rich picture of UK water sector.

Climate projections (UKP09) [2] suggest that the change in climate [3] and the rise in demand for water will increase the strain on UK water resources. General trend projections indicate warmer and drier summers that will result in less water being available for all uses in the coming decades. Additionally, what is expected is that winters will generally be wetter with heavier downpours of rain which will increase the risk of flooding [4]. To this end, the water needs of the UK population, industry and associated dependencies is expected to increase due to forecast population growth [5].

The risks posed by flooding on supply water to demand points ought to be considered just as much as the immediately obvious and devastating impact flooding has on the lives of people, buildings and the environment. This would imply that water is successfully supplied to all demand regions (domestic, industrial and agricultural) whatever the circumstances are of the environment and that the water SoS holistically demonstrates levels of resilience in different operational scenarios. This entails both the aforementioned scenario of less water availability and increased temperatures, and also in crisis scenarios such as natural and man-made disasters, where water may become unavailable or available only at a depleted level, for a certain period of time. Therefore, when attempting to "engineer resilience" [6,7] into such systems, there needs to be an understanding of levels of performance, acceptable levels of performance, and also recovery strategies to mitigate against loss of service.

The objective of this paper is to demonstrate how the resilience methodology created can explore where resilience improvements are necessary and then increase the resilience of water supply systems during times of flooding, by adopting an SoS-based approach and applied model-based processes. A key contribution is the inclusion of a resilience viewpoint within an existing architecture framework to allow engineers to model certain aspects of SoS resilience using different views and model types.

#### *1.1. Resilience in the Water Sector*

'Resilience' is a term that has been used for a long time and in many different domains [6,8,9]. The popularity of the concept is currently on the rise [10] and has attracted great interest from engineers [11] and safety analysts [12] alike. Many variations of the definition exist but the core

connotation remains constant with the Latin origin of the word "resilire", to spring back or rebound [13] i.e., to recover. Traditionally, resilience has been closely linked with safety and risk management strategy planning, taking a hindsight and "lessons-learnt" approach from past incidents to improve existing and future systems. For knowledge purposes, there are many advantages in studying past failures and disasters to gain an insight into why things go wrong, how response systems performed, taking the positives and negatives from a given instance to inform better decision-making in the future. That said, a more effective approach is a proactive one, to test possible future scenarios and see how the system would respond, in order to develop systems with the capability to deal with change prior to an incident and mitigating the severity of the potential impact.

For the purpose of this research, the water system and the occurrence of a flooding event was considered as a system-of-systems (SoS) [14] as a means to appreciate the complexity of water supply systems and associate systems in times of flooding. This permitted the development of model-based methodology and to propose methods to design future phases of water systems that ensure resilience strategies are included early on in their lifecycle phases. In this specific case, the methodology supports increasing the overall resilience of the systems delivering water to consumers in times of flooding.

Literature shows resilience frameworks being developed from an "adaptation" perspective [14] which is interesting from the standpoint of existing infrastructure having to cope with new threats. Nelson et al. [14] provide a definition for systems which are resilient as being able to "undergo change and still retain the same function and structure while maintaining the options to develop." This suggests that systems are able to support a degraded level of performance and provide the core functions required to meet the system's goals. The notion of "surprises" is introduced by Hollangel [15] in his definition of resilience which suggests the idea of unanticipated events, making the challenge for engineering for certain types of events more difficult because they are less predictable. Certain definitions highlight the importance of time and costs whilst a system is in the recovery stages of an incident [16] and the significance of reducing these as optimally as possible. With all definitions, the implications of resilience are all context dependent, therefore defining what is meant by resilient in any given scenario is a critical part of engineering for resilience. This is equally true in our scenario of continuing the supply of water during extreme weather events like flooding. Legacy systems (pumps, for example) may not be able to cope with the changes in flood frequency and flood severity, and evolutions to existing systems should to be made, say by replacing aging systems, ensuring sustainability in the UK's water supply SoS.

The concept of resilience in this instance would attempt to mitigate against this risk of failure and to ensure the availability of water, even in times of severe conditions, i.e., infrastructure resilience. The resilience curve [17,18] is a common representation of the responsiveness of a system to undergo a disturbance and to recover from it in terms of performance; it has four main stages; reliability, unreliability, recovery, and, recovered steady state [8], as seen in Figure 2a.

**Figure 2.** (**a**) Generic resilience curve; (**b**) different impacts on resilience.

The objective when considering the resilience of the water system, and the constituent systems which enable the supply of water is that when the occurrence of a flood for example occurs, then the resilience curve is more like profile (i) in Figure 2b, than profile (ii). Profile (ii) shows complete failure of the system and recovers to a state that is substantially below its original level. Where ideally, there would be no loss in performance, however profile (i) shows an "acceptable loss" in performance, and a full recovery to its normal operational state over a period of time. Of course, this is context dependent, but describing the system resilience in this way enables engineers and architects to apply the resilience methodology and test future designs against variables which are of importance.

Having a succinct methodology which enables engineers, architects and decision makers in the water domain to assess the resilience of current systems, and to then propose and test future architectural designs would be invaluable. The methodology which has been created within this article demonstrates a workflow which defines the scope of interest, and then models the elements within a reference architecture using systems modelling languages and architecture frameworks. Additionally, these static architectures can be examined by migrating them into simulation environments to assess their performance. Favourably, alternative designs and architectures can be modelled, simulated and tested to see which provides the most resilient solutions in terms of water availability and continual supply.

The value of adding the resilience viewpoint within the methodology is to specifically address aspects of resilience like criticality and vulnerability between certain nodes of the network and to identify key risks in certain scenarios. This allows the architect to think about strategies which enable the operational systems to mitigate against certain incidents and remain reliable and resilient at all times. Assigning metrics to individual constituents within a system is seen to be a step in the right direction for developing a set of metrics to measure the resilience in the water sector. This is seen to be another feature of the resilience viewpoint, an ongoing development process and research initiative.

#### *1.2. Systems-of-Systems (SoS) in the Context of Water Supply/Flooding*

Several definitions of what constitutes an SoS have been provided by Jamshidi 2005, 2008 [19], though the most favourable definition from the book is "systems of systems are large-scale integrated systems that are heterogeneous and independently operable on their own, but are networked together for a common goal," [20]. Maier [21,22] claimed that the term 'system-of-systems' did not have a distinct and recognised definition, although he did acknowledge that the SoS idea is widely accepted and generally recognised. He cited a number of examples such as integrated air defence networks, the internet, intelligent transport systems, and enterprise information networks, which are an emergent class of systems comprising large-scale systems in their own right. This has led to the categorisation of an SoS into five important characteristics:


Accordingly, a definition of resilience has been created by the authors in the context of SoS, such as that of the water supply system, which is of topic in this case; "The dynamic ability of the SoS to re-adjust and recover when faced with change and disruption, at both the SoS and constituent system level. To continue to provide operational capacity at a certain level of (degraded) functionality and performance."

There is a clear and evident requirement for a conscious effort to be made towards a resilient water SoS within the UK, in terms of both supply and in terms of flooding. This effort is as stated, an 'SoS-wide' effort which must be supported by a range of stakeholders, including; water companies, emergency response agencies, wastewater companies, regulators, and consumers of water, to name a few. Natural disasters (including flooding) increase the probability of the water infrastructure being damaged which also could jeopardize the supply of water to customers of all types (domestic, industry, agriculture).

#### **2. Materials and Methods**

#### *The SoS Resilience Framework/Methodology*

The SoS resilience methodology is an encompassing set of methods and processes which helps decision makers, engineers and other stakeholders "design" resilience into SoS. The framework has been developed with the goal of evaluating resilience in the 'as is' state of the system and then implementing changes to target resilience in certain areas of the system which could be improved. The water scenario is a very useful case to test the methodology as it is representative of these types of mega-systems. Modelling and simulation (M&S) methods form the basis of the methodology, however it is a tool independent framework and can be replicated in a different set of modelling languages and tools, if required.

The methodology is systematic and takes a holistic approach to engineering future resilient systems. Stakeholder interaction is a fundamental characteristic of the methodology and subject matter experts (SMEs) from the water sector have been consulted throughout the application of the methodology at all phases. An overview of the developed framework is shown in Figure 3.

**Figure 3.** System-of-systems (SoS) resilience methodology.

#### **Phase 1.** CONOPs and Resilience Requirements

Phase 1 commences with deriving a set of requirements specific to the area of concern and understanding the concept of operations (CONOPs) of the operational systems involved. The tools applied here are rich picture diagrams—a soft systems method [1]—and causal loop models [23,24] to understand the dynamic interactions between system elements. Subsequently a reference architecture [1,4,25] is created using systems modelling languages such as SysML and the Department of Defense Architecture Framework (DoDAF) [26] which is common language used to create systems architecture descriptions of systems and enterprises.

#### **Phase 2.** SoS Reference Architecture

The reference architecture forms the backbone to the methodology as it defines the interactions between constituent systems (CS) and sub-elements within the water SoS. A reference architecture, as the name suggests is a source of reference when regarding systems from different architectural perspectives. A set of views that form the reference architecture are shown in Figure 4. The architecture is a resource which is typically shared amongst project stakeholders, and engineering teams in the development phases of software, a product, a service, a system or an SoS. A common set of information is one of the key benefits of model-based systems engineering (MBSE) [27–29] and the purpose of a reference architecture, which is an MBSE practice.

**Figure 4.** Architecture modelling views process.

#### **Phase 3.** Resilience Viewpoint

A novel viewpoint within the framework has been created solely focused on the resilience aspects of the architecture, and proposes three model types (Figure 4) to address; static structural analysis of the nodes within a network of systems and to explore the cascading effects of a failure or disruption; risks and those who are responsible for mitigating against risks, and; resilience attributes which are being designed into the architecture, for example, non-functional features referred to as "ilities" [30,31] such as flexibility, robustness, security, and availability. These non-functional properties are assessed through observation from the subsequent simulation models created in the next phases of the methodology.

#### **Phase 4.** Model-Integration Reference Architecture

The model-integration architecture provides the "big picture" view of the model architecture and puts it in context with the rest of the stakeholder's model portfolio, showing how the project's models and model platforms fit together. The importance of a modelling and the simulation integration framework cannot be overlooked, as it provides details at both high and low levels of granularity, conditional of the MBSE approach employed. This architecture specifies the tools used to develop models, data types, and the exchange of models and data between platforms and engineering teams. This process is highly recommended for when translating architecture to an executable model as it defines exactly what data is required for setting the parameters in the model. The integration framework is independent from the simulation tool; therefore, the engineer can select tools which are apt at a later date from when the architecture was created. The integration framework is beneficial in communicating requirements between stakeholders, and the "chief architect" and has the responsibility of gathering the essential information from individual model owners to generate a complete picture of the integration framework.

#### **Phase 5.** SoS/CS Simulation

Determining the resilience of an architecture is not a trivial task. A predominant reason for this is the nature of the architecture. Architectural designs are essentially static representations of the SoS whereas the real SoS is a dynamic entity whose behaviour is created from the interactions of the constituent elements. For successful analysis of SoS resilience to be conducted the SoS framework must consist of dynamic simulations to explore emergent behaviours which may arise from alternative architectures. The SoS reference architecture strongly supports the depiction of the SoS and its constituent systems, but it is desirable to represent the models that can be simulated in suitable simulation environments. In this case, the platform Simulink (MATLAB\_R2018b, MathWorks, Cambridge, UK) was used—in conjunction with IBM Rhapsody which was used to create the architecture in phases 2, 3 and 4—to run simulations of supply and demand scenarios within a particular region of the UK.

#### **Phase 6.** Resilience Architecture Selection

This process is illustrated by Figure 5, where IBM Rhapsody architectures are manually translated into Simulink for simulation, and subsequent analysis in an additional visual analytics tool, in this case Tableau software (Tableau Desktop 10.5, Seattle, WA, USA). Simulink was selected as it permits the use of sliders and other parameter controls as the simulation runs in real-time. This enables decision makers to explore different parameters and immediately visualize the outputs.

**Figure 5.** Modelling and simulation workflow.

Although an overall SoS resilience metric is desirable, it gives little information of the resilience at the local, CS level (or lower) which could be problematic when resilience is required in a specific area of the SoS. Uncertainty and unknown parameters are likely to make the task of achieving overall SoS resilience extremely difficult for the architect, hence why targeting resilience at the CS level is probably more logical. Metrics associated to executable architectures provide insight to the performance of technical and social systems which assist in the design space exploration process for more resilience solutions.

#### **3. Case Study Results**

#### *3.1. Case Study Overview*

This case study considers a scenario where flooding has resulted in pump failure (directly from flood water or from power failure), where the pump is a constituent system that forms part of the SoS. If we consider this case and assume that the flood has created a redundancy problem as the pump supplies multiple demand points within a network, it becomes evident that this challenge is greater than a redundancy problem and we must consider alternative resilience strategies to overcome the difficulties. Let us consider demand centres which are supplied by a water source of some kind (e.g., primary service reservoir) and utilise a pump to feed water up against gravity, to a set of demand centres via a number of service reservoirs. Of the four demand centres as seen in Figure 6, it is assumed (for demonstration purposes) that two are for domestic use, one for industrial and one for agriculture.

**Figure 6.** Case study scenario.

As previously mentioned, the resilience is the capacity of the water supply system to meet demands during emergency situations like floods, and in this case during the failure of a pump in the network. As it can be seen from Figure 6, this would cease supply to all four service reservoirs, enabling supply to the demand points until those service reservoirs completely depleted. Multiple strategies can be implemented, and numerous architectural variants can be suggested to increase the resilience here, however before getting to these, the upcoming results section shows the application of certain phases of the SoS resilience methodology.

#### *3.2. Case Study Results*

Following the phases outlined previously in Figure 3, a great understanding of the water SoS is achieved through applying CONOPs methods such as rich picture diagrams (Figures 1 and 6) and causal loop models (Figure 7). The advantages of causal loop diagrams are that it helps analyse complex systems and helps identify key dynamic variables for later simulations. Cause and effect diagrams are a crucial aspect of designing resilience systems, particularly for understanding the dynamics between key variables. The loops which are created show an overall effect on the system and this is one of the elements of a causal loop diagram as illustrated in Figure 7, where the overall effect is a decrease in water available due to damaged infrastructure in times of a flood event. Although this seems a basic construct, reducing to this level of complexity is beneficial in exploring potential solutions, and determining the types of data sets required for simulating later in the process. Thus, from the causal loop it can be determined that the solution that needs exploring is the flexibility of the current systems in place to search for strategies and also architectural solutions which regulate the availability of water and hence, the supply of water to those that require it.

**Figure 7.** Causal loop of water problem.

The rich picture and causal loop models helped define the requirements for this case study, and eliciting requirements for resilience in this case was done in collaboration with a range of stakeholders via small workshops, including; water companies, water consultancies and academics interested in water applications.

A subset of architecture models has been provided here to illustrate the type of modelling, but the reality is there would be numerous instances of some models from different stakeholders' standpoints to describe the CSs and to elicit further resilience requirements. The two models here show a general Operational Viewpoint (OV-5b) of the water processing procedure and a Systems Viewpoint (SV-1) of the water companies' systems arrangement and connectivity (Figures 8 and 9, respectively). Figure 8 shows a high-level process model of taking raw water and passing it through a series of stages such as initial water storage, water screening, filtering etc. prior to being supplied for water customers and consumption. Figure 9 shows four high level components of a water system; a water company, distribution network, customers and consumers and waste water distribution network. Again, these are then populated with further subsystems which must be present to make the systems functional. Multiple instances of each model type can be created to model specific system structures, but these are for demonstration purposes.

#### *Water* **2019**, *11*, 496

**Figure 8.** OV-5b—Operational view of high-level water treatment process.

**Figure 9.** SV-1 Water Company systems.

Evolutionary development of the SoS architecture should be addressed by iteratively updating the specific constituent systems with the overall SoS in mind. The reference architecture is only as good as the quality of the information it holds and therefore, it is crucial that quality information is captured in its views/models. These static architectural representations go a long way in the design phases, allowing effective communication with stakeholders who have an influence on shaping the CS and the requirements definition process for future phases of the SoS evolution.

Subsequently, modelling the system using the resilience viewpoint (more specifically, Resilience View 2—RV-2) allows the engineer to highlight which nodes are critically dependent on each other and also to show which nodes are vulnerable if failure was to occur at some point in the network. For instance, modelling the case study example, shown in Figure 6, the viewpoint shows the four demand points being critically dependent on the functioning of the pump, and vice versa, the demand points being vulnerable on the pump failing. Similarly, the water storage component, or the primary service reservoir becomes vulnerable if the pump is no longer working, because that clean water has a set period which it can be stored for before becoming unsafe to distribute. Figure 10 shows the capability of RV-2 to model further details of each SoS component, for instance, the main supply pipeline has certain attributes which can be assigned to that element. These may include important parameters such as flow rates, flow direction, capacity, and others which are important in later simulation phases.

**Figure 10.** Resilience View 2 (RV-2) pump failure.

How would this look on the resilience curve? Referring back to Figure 2, the outage of a pump in a single pump supply system could result in curve 2b (ii), the performance dropping off to zero until that pump was restored. Alternatively, an architectural solution to the problem could be to add a secondary pump (Figure 11), that could be turned on immediately after failure, and performance would again increase back to a lower level of performance, or back to ideal performance P0. On the other hand, a decrease in performance could be avoided altogether if the failure of Pump 1 was anticipated, and Pump 2 switched on prior to failure.

**Figure 11.** RV-2—Two pump architecture solution.

At first glance, this may seem like an obvious problem with some obvious solutions to increase the resilience of supplying water to the consumers. Resilience here is considered as a redundancy issue. However, resilience can be a set of operational strategies implemented during a time of disruption to solve the supply problem. For example, prioritizing which demand points are more critical, decision-makers can set constraints on where to supply water and for which certain periods of time. For instance, it may be acceptable to supply the two domestic areas from say 6.00 am to 10.00 am and then from 3.00 pm to 10.00 pm in attempt to cut back on water consumption and resultant water waste. Another solution would be to decrease the pressure of which water is supplied to these points which would decrease the overall volume of water available to be used. Furthermore, depending on what is prioritised, the system may gracefully degrade its capacity to supply the two domestic points and the industrial point, reducing its supply to agriculture completely for a given period of time. This is one resilience strategy, although there will be many like this which should be explored architecturally and simulated to evaluate the feasibility of each one. These are just a few of many strategies which could be implemented to resolve the problem and, in the short-term at least, solve some issues of increasing resilience against flood scenarios.

From a risk perspective, the resilience viewpoint offers a model (Resilience View 3, RV-3) to capture risks which can be communicated with all stakeholder groups. Additionally, the risk view as illustrated in Figure 12, permits the architect or engineer to allocate the risks to constituent systems or stakeholders within the water SoS. This allocation of risks establishes responsibility of each identified risk and provides accountability to whom or to what is responsible for mitigating against that risk. The advantage from an architectural perspective is that risks can be avoided if mitigation strategies are embedded within the design of constituent systems. An example of an RV-3—Risk View Description—is shown in Figure 12. The example shows the risk of failure (or partial failure) of the water pump from the above example. Although this a simple example for demonstration purposes, the number of risks in a scenario like this are likely to be in the tens or hundreds, and this is a valued way to manage these risks as the complexity can be captured in multiple instances. This view (RV-3) within the resilience view package is a novel inclusion to the architecture framework which historically does not consider risks (in this way) as part of the architecture modelling phases of systems lifecycle modelling or development.

**Figure 12.** Resilience View 3 (RV-3)—Risk view description.

The final phase to be applied is simulation of architectures. Due to the nature of an SoS such as the water SoS, it is extremely important to simulate as much of the SoS and with as many of the CSs included as possible. The purpose of simulation in our case is to determine whether the overall resilience goals of the SoS can be satisfied by evaluating the interactions between the constituent systems and to understand where the current system falls short in terms of resilience. Since we cannot always rely on the CSs being modelled with the same tools, special techniques must be used to transform the architecture into a set of system models that can be executed within the same co-simulation [32] environment. Whilst it seems convenient for all CSs to be modelled using the same tool this is not always possible, nor desirable, due to commercial restrictions i.e., different water companies use different modelling tools and languages to design future capabilities. Also, it is highly probable that certain existing CSs have already been created and tested using other tools and provision should be made to use these wherever possible.

System dynamics (SD) [33] is an approach applied to understand the behaviour of complex, dynamic, nonlinear systems. Defining variables is a key step in creating an SD model, as these are assigned to specific model elements to simulate how changes in the system occur over time, and thus provide understanding of the basic structure of a system and the rationale for its behaviours. A positive of employing such a method in the SoS context comes from the need to understand the behaviour of the whole through understanding the behaviour of the interconnected parts. Numerical values can be assigned to the model and the dynamics can be simulated through execution of the variables. This study used the software tool, Simulink to create the model and simulation for our case study that stemmed from the specifications defined within IBM Rhapsody.

The Simulink model as seen in Figure 13 shows the element that reflects the problem scenario outlined in Figure 6, which looks at the use of pumps and valves to carry water from a primary reservoir, against gravity, to four service reservoirs which supply different types of customers (domestic, industry and agricultural). The model enables the user to set desired tank levels and the prescribed levels and to then manipulate the pumps and valves to direct water to specific service reservoirs depending on the resilience strategy in place. This specific strategy focussed on keeping the supply high to the two domestic customer groups i.e., service reservoirs 1 and 2, and to drop the amount of water being supplied to agriculture for the short term in order to maintain water to the domestic and industry customers. Preliminary results for this simulation run can be seen in Figure 14 where the domestic customers and their respective service reservoirs remain constantly at a high water capacity and the agricultural service reservoir (service reservoir 4) remains low. Service reservoir 3, the industry service reservoir fluctuates depending on the demand.

**Figure 13.** Simulink model.

**Figure 14.** Service reservoir capacity from simulation run.

In order to test the model further, real data from the demand of customer groups and the volume and capacity of real reservoirs from specific locations will enable more accurate simulations to be run that are reflective of incidents in times of flooding. A great advantage of running a model like this is Simulink, in that it enables the engineer or decision maker to manipulate the parameters and variables using sliders and other input types to assess the response of the system in real time. Thus, allowing trade offs to be made in real time with respect to the events of a flood and the resilience strategies which are being tested and implemented.

#### **4. Discussion**

The methodology presented provides a means to explore resilience in SoS and to engineer resilience into systems such as the water SoS at multiple points within its lifecycle. Engineering resilience is a systematic endeavour and the model-based engineering methodology shown enables resilience to be evaluated in the 'as is' state and to explore future states to proactively mitigate against certain risks within particular contexts. Overall, the methodology will stipulate

many outputs which are regarded as tools for engineering resilience in SoS. Such artefacts include; information about the water SoS which is stored within a reference architecture; architectural designs which can be stored as architecture patterns [34] for later re-use in design activities; simulation data which reflects the performance of architectures and helps inform decisions about architecture design; and finally; a detailed and shared understanding of key resilience challenges and issues, within the water sector, across a broad range of stakeholders.

In order to avoid commercially sensitive data, generic data was presented in this paper that is based on real water data to test the methodology. A significant result is the creation of a resilience viewpoint within an architecture framework that solely focusses on resilience. The anticipated future results would be to simulate some of the water architectures suggested to see how they perform using a resilience curve and to assess the responsiveness of certain designs. This would be done using the workflow as suggested in Figure 5, and data would be analysed to infer resilience measures for future SoS evolutions.

The intrinsic characteristic of any SoS is its state of constant flux and evolution. This means the reference architecture and subsequent simulations must be updated on a regular basis, especially when simulation results depart from the observed operation of the SoS. Attempting to perform this evaluation on the actual SoS is fraught with danger, because surprise emergent behaviours may be highly detrimental or unsafe. The reference architecture (blueprint) becomes an essential asset when considering the evolution of an SoS or where the SoS begins to exhibit unpredicted behaviour. However, we must always guard against the differences between the real world and the simulated SoS. Nevertheless, having a reference benchmark is extremely important. Similarly, maintaining a library of patterns for the constituent systems is very helpful not only in representing complex characteristics of the constituent systems but also in their reuse.

It was clear from the engagement with water companies and consulting companies that asset resilience is of high priority to water companies, in ensuring an aging infrastructure can meet the demands of present and future trends. To solve the challenges of modernizing infrastructure and maintaining waste water networks, there is an evident need for an interdisciplinary approach involving experts from water supply companies, systems engineers, climate experts, and key customers (domestic, industry and agriculture). By treating the water sector as an SoS, it should be possible to begin to address the more challenging aspects of ensuring resilience in water supply for the future; for example, the application of a reference architecture for mapping the current resources, surveying the current condition, analysing the problem areas, and understanding the nexus within all the stakeholders like geographical and economic implications. Strengthening the links between several elements from academic research and industry could make an important contribution to solving these challenges via the application of an SoS reference architecture.

Future work would include generating architectures that can be assessed quantifiably allowing resilience metrics to be developed for different types of SoS while measuring the performance of architectures can iteratively be done in modelling and simulation environments bespoke to these kinds of investigations. To also link the resilience curve with specific data from water supply failures would provide a way forward for quantifying resilience using some important metrics.

Where a pre-existing model of a constituent system exists, it is not always possible to guarantee that it will be compatible with the simulation environment. Therefore, care must be taken to understand the limitations of the model and its context of use. Such requirements place specific demands on the co-simulation environment, which must be a component-based simulator where system models are incorporated as the composition of a set of hierarchical modules. Consequently, the reference architecture must be structured to support a composition of constituent system models. Whilst this task is straightforward at the reference architectural modelling level, this requirement restricts the choice of simulation environment and depends on what simulations are needed. However, underlying this is a complex process of transforming system architectures into executable models [32], and something which should be looked into in future work.

#### **5. Conclusions**

This paper presents a methodology to explore resilience in the water sector from multiple perspectives, by adopting an SoS approach to problem solving. Water supply to flooding has been considered from an SoS standpoint, and the creation of a methodology allows resilience to be explored in many water scenarios to assess resilience and propose future designs. Resilience is seen as a new paradigm to risk management. A proactive approach is needed to explore future scenarios which may cause problems for systems with high dependencies and raise concerns with regards to system performance. Resilience engineering is seen as an important topic in the domain of systems engineering and certainly in sectors such as water, where the importance of successful operations is critical to human life and the environment.

Model-based systems engineering, and a structured framework or methodology allows decision makers and engineers to; (i) evaluate the resilience of current systems through static and dynamic methods, and; (ii) to discover future system states and resilience strategies through architecture exploration and implement them through a tested methodology. The water supply and flooding example shown within this paper illustrates the practicality of applying a set of MBSE methods to increase the resilience of systems which deliver water to consumers and demands points within a supply network. Water systems may suffer impairments and disruptions during times of severe weather, e.g., floods and droughts, or even failure due to legacy systems aging and failing naturally. The paper has shown how water may still be made available during a flood scenario, through architectural alternatives to deliver water prior and during a disruption. However, there are multiple variations of strategies and architectural alternatives which could provide a solution to this specific issue. Furthermore, water availability data of a specific region would help validate this process via informed simulation models, developed in conjunction with subject matter experts from a water company, however the methodology is a strong step in the right direction to explore resilience in SoS and the water sector during times of flooding.

**Author Contributions:** Conceptualization, D.J., R.K.; Data curation, D.J. and S.S.; Formal analysis, D.J. and S.S.; Funding acquisition, D.J. and R.K.; Investigation, D.J.; Methodology, D.J.; Project administration, D.J., R.K., M.R.C. and G.F.; Resources, D.J.; Software, D.J. and S.S.; Supervision, R.K.; Visualization, D.J.; Writing—original draft, D.J. and S.S.; Writing—review & editing, D.J., R.K., G.F., M.R.C. and F.M.

**Funding:** The authors would like to thank the EPSRC for the funding on BRIM (EP/N010329/1).

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

## *Article* **Automated Floodway Determination Using Particle Swarm Optimization**

**Huidae Cho 1,\*,†, Tien M. Yee 2,† and Joonghyeok Heo <sup>3</sup>**


Received: 4 September 2018; Accepted: 6 October 2018; Published: 10 October 2018

**Abstract:** The floodway plays an important role in flood modeling. In the United States, the Federal Emergency Management Agency requires the floodway to be determined using an approved computer program for developed communities. It is a local government's interest to minimize the floodway area because encroachment areas may be permitted for human activities. However, manual determination of the floodway can be time-consuming and subjective depending on the modeler's knowledge and judgments, and may not necessarily produce a small floodway especially when there are many cross sections because of their correlation. Very little work has been done in terms of floodway optimization. In this study, we propose an optimization method for minimizing the floodway area using the Isolated-Speciation-based Particle Swarm Optimization algorithm and the Hydrologic Engineering Center's River Analysis System (HEC-RAS). This method optimizes the floodway by defining an objective function that considers the floodway area and hydraulic requirements, and automating operations of HEC-RAS. We used a floodway model provided by HEC-RAS and compared the proposed, manual, and default HEC-RAS methods. The proposed method consistently improved the objective function value by 1–40%. We believe that this method can provide an automated tool for optimizing the floodway model using HEC-RAS.

**Keywords:** floodway; optimization; particle swarm optimization; HEC-RAS; flood mitigation; hydraulic modeling

#### **1. Introduction**

The floodplain and floodway are an essential part of hydrologic and hydraulic studies of riverine flooding. The floodplain shows any area that will be covered by water when a flood event occurs. In the United States, the Federal Emergency Management Agency (FEMA) requires the 100-year and 500-year floodplains as part of the National Flood Insurance Program (NFIP) [1]. For developed communities, FEMA also requires a floodway to be determined within the 100-year floodplain using a computer program approved by them [2]. FEMA approved the Hydrologic Engineering Center's River Analysis System (HEC-RAS) [3] because it is widely accepted and used for floodplain mapping and flood risk modeling around the world [4–8]. HEC-RAS was developed by the US Corp of Engineers [3] and is available to the public at no cost. However, since this computer program is not open source, its source code is not available to the research community and implementing improvements within HEC-RAS is very difficult, if at all possible. To address this challenge, HEC-RAS provides the Application Programming Interface (API) called the HECRASController that allows the user to control its user interface non-interactively [9].

The floodway is defined by vertical encroachments on both sides of the main channel area within the 100-year floodplain as shown in Figure 1 and conveys flood water without raising the water surface elevation by a regulated threshold, typically set to be 0.305 m (1 ft) in the United States by FEMA unless a lesser rise criterion is imposed by the state [10]. Figure 1 shows a schematic of the 100-year flood elevation and the floodway elevation before and after encroachment, respectively. Ideally, development within the floodplain should be avoided. However, in dense urban areas, development that encroaches into the floodplain is unavoidable because an increase in human activities often pushes development closer to rivers and streams, and even into floodplains. At the same time, for safety reasons, many regulations do exist to prohibit excessive encroachments into natural streams because such encroachments can cause rise in the 100-year flood elevation and result in severe flooding upstream. The floodway then becomes a boundary to prevent encroachments from excessively impeding the conveyance of flow and hence excessive rise in the flood elevation. Since encroachment areas on both sides are often used for human activities such as business, leisure, parks, etc., it is one of the major interests for landowners and local governments to maximize these areas. To satisfy their interests in a safe manner, the width of the floodway can be reduced during flood modeling as long as the hydraulic requirements are met to minimize flood hazards. By optimizing the floodway, local governments can achieve more sustainable land use planning, better risk and safety assessment, and will be able to mitigate legal issues due to subjective floodway interpretations. However, floodway optimization is a difficult task because the floodway boundary can be established in many different ways [11].

**Figure 1.** Schematic of the 100-year flood elevation and floodway.

Protocols in creating the floodway boundary are not well defined [12,13] and are primarily left to the practitioner [13]. One major problem without well-defined methods of floodway determination is that floodway boundaries produced by different modelers may vary significantly. Depending on the modeler's motivation, experience level, and understanding of the subject matter, different modelers may produce different floodway boundaries even if given the identical hydraulic model. This subjectivity may not necessarily result in optimal floodway boundaries and hence highlights the need and advantage of a computer-aided optimization method.

Selvanathan and Dymond [14] developed an ArcGIS [15] tool that can run HEC-RAS, post-process the results, and visualize and smooth the floodway from within an ArcGIS environment. Their tool also allows the modeler to adjust encroachments visually inside ArcGIS and has a limited optimization routine that tries to satisfy the surcharge requirement. However, the focus of their research was mainly on automation of iterative HEC-RAS runs and post-processing of its outputs. Franz and Melching [11] introduced the Full Equations Utilities (FEQUTL) model that uses an iterative trial and error procedure to determine the left and right encroachment limits. Their model was not developed to perform reach-wide optimization of the floodway, but it does provide the reader with a glimpse of the technique used for floodway determination. Majority of focus in floodway modeling revolves around the criteria and methods used in floodway modeling [12,13,16,17]. Most discussions are primarily focused on modeling techniques, recommendations to modeling standards and procedures, and evaluating the practicality and feasibility of applying one uniform standard for all floodways. Thomas and Golaszewski [17] suggested an improved iterative procedure that involves a consideration of non-steady section-averaged velocity, variation of velocity-depth product, topographic and geomorphic features, control of hydraulic structures, flow conveyance and results of hydraulic models. They also acknowledged that an experienced practitioner is required to delineate and assess floodways using the improved iterative procedure.

Lots of effort in optimization of floodplains and floodways have been devoted towards the system operations of flow control structures within the floodway [18–20] and floodplain management related issues such as flood risk assessment and cost benefit analysis of different floods and structures [21–31]. Szemis et al. [18] introduced an optimization framework for scheduling environmental flow management alternatives using ant colony optimization [32]. Bogardi and Balogh [19] developed a model that calculates the probability of levee failures and optimizes floodway operations. Luke et al. [20] studied the impact on the floodway and levee damages in the New Madrid floodway in Missouri of the detonation control during May of 2011 and concluded that passive control would have greatly reduced the costs of repairing those hydraulic structures. Lund [21] used linear programming to develop an approach that minimizes the expected flood damages and costs. Shafiei et al. [22] examined different genetic algorithms (GA) [33] for optimizing the levee encroachment design and concluded that GAs are acceptable tools for solving the levee design problems while non-GA-based optimization techniques may not be able to find the global optimum of such problems because of the non-linear nature of the objective function surface. Mori and Perrings [23] developed a model for finding optimal floodplain development decisions. Yazdi and Neyshabouri [24] used the non-dominated sorting GA to find the optimal Pareto solutions of two objective functions minimizing flood mitigation costs and potential damages to the floodplain. Lu et al. [25] proposed an inexact sequential response planning approach for optimal management of floodplains. Porse [26] used linear programming to evaluate decisions for urban floodplain development and assess potential flood damages. Woodward et al. [27,28] developed a decision support system that generates effective mitigation measures and optimizes their performance using a multi-objective optimization algorithm. Lopez-Llompart and Kondolf [29] and Kondolf and Lopez-Llompart [30] studied how floodways in the Mississippi River had been affected by land use conflicts and management. Czigáni et al. [31] used the MIKE 21 model [34] for multi-purpose floodway zoning and floodway delineation along the lower Hungarian Drava section. However, very little has been done to produce an optimized floodway boundary for the entire stream [35] and an extensive literature review revealed Froehlich's [36] and Yu's [35] works.

Froehlich [36] suggested that delineation of the floodway boundary be done in a fair way. He pointed out legal issues including violation of the constitution with floodplain regulations and related those issues towards a need for a just and fair way to delineate a floodway. He further hinted that a reach-wide optimal solution would be an impartial way to define regulatory floodway boundaries. Froehlich's proposed approach uses dynamic programming to delineate the floodway boundary of the steady-state energy balance equation using the standard step method. His research is significant in the way that he realized the importance of reach-wide optimization of the floodway and attempted to use dynamic programming as an optimization tool. However, since the work was done using hydraulic code that is not approved by FEMA, his floodway optimization technique cannot be used to generate floodplains and floodways for FEMA.

Yu [35] used a GA to calculate the floodway encroachment limits within HEC-RAS. For the objective function, he used the sum of the absolute difference of simulated and desired water surface elevations within the floodway for all cross sections. He attempted to find floodway encroachment limits that keep the surcharge for all cross sections within an acceptable range, but he did not consider minimizing the floodway area nor maintaining a subcritical flow state reach-wide. His approach was not able to produce better results than the default methods in HEC-RAS, but the idea of using a heuristic algorithm to determine the floodway aligns well with this study.

In one-dimensional riverine modeling, cross sections are extracted from terrain data along the river where there are hydraulically important features. For each cross section, the left and right encroachment limits are assigned, which define the floodway boundaries. In flood modeling, the flood elevations are interrelated between cross sections. When modeling streams that consist of only a few cross sections, it may be feasible to manually find optimal encroachment limits based on engineering judgments, but as the number of cross sections increases, seeking feasible floodway boundaries becomes more complex and time-consuming. This difficulty usually discourages the modeler from repeatedly running the model using different combinations of encroachment limits to determine the best feasible reach-wide solution that would meet the surcharge and hydraulic requirements. Even if the modeler is willing to put forth the trial and error effort, the floodway determined in this manner may not be optimized. Further, there is no one structured and uniform procedure that exists in defining the floodway area [12,13,16], let alone finding the best feasible one. The default methods built in HEC-RAS are generally used as a starting point [13] and should not be considered a method for reach-wide optimization. Optimization of the floodway in a reach-wide manner involves generating the left and right encroachment limits in each cross section that will result in a regulatory rise in the water surface elevation along the entire stream. At the same time, these encroachment limits should satisfy certain requirements set by the modeler. As we mentioned above, our interest is to minimize the footprint area of the floodway to maximize the land use of both encroachments by local governments while ensuring the hydraulic safety from potential flood hazards.

In this study, we developed software using the HECRASController to communicate with HEC-RAS and implemented an optimization method for reach-wide floodway determination using a heuristic algorithm called Isolated-Speciation-based Particle Swarm Optimization (ISPSO) [37] and the HEC-RAS hydraulic model. In Section 2, we defined the floodway optimization problem and formulated engineering judgments as an objective function to remove any subjectivity. We also briefly introduced ISPSO, and combined the objective function and ISPSO to elaborate on the development of the proposed method called the Automated Floodway Optimizer for HEC-RAS (AFORAS). In Section 3, we conducted a case study using a universally available floodway model from HEC-RAS and discussed its results. Finally, in Section 4, we highlighted the main advantages and limitations of AFORAS and concluded this study with a consideration of future work.

#### **2. Materials and Methods**

#### *2.1. Study Area*

A floodway model called FLODENCR included in the HEC-RAS 4.1.0 installation was used for this study. This model was chosen at random to demonstrate the effectiveness of the optimization method. In addition, since input files for the case study are freely available from the HEC-RAS installation, researchers may duplicate the simulation if desired. Although this case study may not generalize all other cases, it serves as a simple but yet applicable case scenario where the floodway can be obtained using the optimization method and manually for cross-validating if the optimization method is indeed providing results that mimic the behavior of a manually obtained floodway. This model represents a 1.59 km of Beaver Creek near Kentwood, Louisiana, and consists of total 12 cross sections with a bridge structure as shown in Figure 2. Figure 3 shows the 9-pier bridge structure on Highway 1049 located at 8.69 km. The upstream area of the bridge is mostly covered by grass while its downstream area is dominated by forests. Two cross sections just upstream and downstream of the bridge define the geometries of the upstream and downstream faces of the structure, respectively. The opening under the low chord of the bridge is 60.05 m wide and its deck is 12.19 m wide. On both sides of the opening, ineffective areas are defined to model the embankment area that effectively blocks the water.

**Figure 2.** River and cross section geometries for Beaver Creek near Kentwood, Louisiana, in the FLODENCR model. The arrow indicates the direction of flow. A bridge structure is located at 8.69 km.

**Figure 3.** Bridge structure on Highway 1049 located at 8.69 km.

#### *2.2. Mathematical Representation of the Problem*

In order to optimize the floodway in HEC-RAS using an optimization algorithm, an objective function needs to be defined that evaluates the fitness of the model parameters objectively because there cannot be the modeler's intervention during an optimization run. The objective function should reflect the modeler's knowledge about the model and their engineering judgments about the model outputs.

Figure 4 depicts a floodway in a river with cross sections, channel banks, and 100-year floodplain extents. The hatched polygon representing the floodway is created by connecting the left and right encroachment limits in each cross section. The area of the hatched polygon indicates the floodway footprint area denoted by *A*fw in this section. In a similar way, the maximum and minimum floodway areas (*A*fw,max and *A*fw,min, respectively) can be calculated by connecting the 100-year floodplain extents (diamonds) and channel banks (dots), respectively. Our interest is to minimize the floodway area *A*fw while satisfying hydraulic requirements.

**Figure 4.** Plan view of a floodway along with a river, cross sections, channel banks, and 100-year floodplain extents.

In floodway determination, the objective function should consider three criteria: (1) Floodway surface area, (2) surcharge, and (3) flow state indicating whether or not the flow is subcritical. The floodway surface area should be minimized while the surcharge is kept within acceptable limits and the subcritical flow state is maintained. There are many different hydraulic parameters in HEC-RAS. However, both models for the 100-year floodplain and floodway are required to share the

same hydraulic parameters including the channel bottom geometry, Manning's roughness coefficient, structural parameters, etc. because hydraulic modeling for the floodway should simulate the same hydraulic conditions as for the 100-year floodplain. Otherwise, the floodway model would not simulate the 100-year flood condition anymore. The only exception is the left and right encroachment limits which defines the floodway area. In other words, the optimization algorithm only adjusts the shape of the hatched polygon in Figure 4 and evaluates the model outputs to calculate the objective function. The model outputs from HEC-RAS provide the surcharge and cross-sectional Froude number [38], which can be used to evaluate violations of criteria (2) and (3) stated above. Ideally, the surcharge should be between 0 and an allowable limit mandated by the FEMA or the state government while the Froude number should be less than 1 for flow to be subcritical.

Most of the streams in the United States flow at a subcritical state except in the mountainous area, which is most likely undeveloped and underpopulated. Since FEMA only requires the floodway analysis for developed communities, the objective function in this study is formulated to handle subcritical flow conditions. However, if the flow state changes to supercritical, the modeler would have to choose the flow condition to model within HEC-RAS. In cases where the flow turns from subcritical to supercritical, HEC-RAS will default to the critical depth [2] to proceed with its calculation even though the water surface elevation may be erroneous. There are options within HEC-RAS for simulation of mixed-type or supercritical flows, but either the trial-and-error method or a priori knowledge of the flow state is required to be able to select these options. While the optimization algorithm itself does not discriminate between the different flow states, encroaching a stream that is flowing at a critical or supercritical flow state will, most often than not, result in a decrease in the water surface elevation and an excessive increase in the flow velocity. Since a negative surcharge is not allowed by FEMA [39], encroachment analysis is not necessary for the supercritical flow state in most of the case.

There are also two constraints: (1) The floodway should not be narrower than the area defined by the left and right channel banks because it should not affect the effective flow conveyance of the channel; and (2) the floodway limits cannot fall outside the 100-year floodplain by definition. The first constraint reserves the channel area between the left and right channel banks to carry the flood water by not impeding this area by floodway encroachment. The second constraint does not allow the floodway encroachment limits on a non-inundated dry area, which effectively leads to no encroachment at all in the river. In Figure 4, the left encroachment limit is constrained between the left channel bank (dots) and the left extent of the 100-year floodplain (diamonds). The same constraints apply to the right encroachment limit. These valid ranges of the left and right encroachment limits are shown in the cross-sectional view in Figure 1. From these two constraints, the minimum and maximum floodway bounds can be defined so that the objective function can compare different floodways by evaluating how close solutions are to the minimum possible floodway area. Given that the above two hydraulic conditions including the surcharge and Froude number are satisfactory, the floodway with the minimum surface area is deemed the optimal floodway.

By formulating the above three criteria and two constraints, the following objective function can be defined:

$$f(\mathbf{M}(\mathbf{g}(\mathbf{x}))) = \begin{cases} \frac{A\_{\text{fo}} - A\_{\text{fo},\text{min}}}{A\_{\text{fo},\text{max}} - A\_{\text{fo},\text{min}}} & \text{if } \mathbf{s} \text{ and } \mathbf{Fr} \text{ are acceptable,} \\ 1 + \sum\_{l=1}^{N} \left\{ \max\left(0 - \mathbf{s}\_{l}/0.305, 0\right) + \max\left(\mathbf{s}\_{l}/0.305 - 1, 0\right) + \max\left(\mathbf{Fr}\_{l} - 1, 0\right) \right\} & \text{otherwise} \end{cases} \tag{1}$$

where **x** is a parameter sample in a normalized hypercube search space [0, 1] *<sup>D</sup>*, *<sup>D</sup>* <sup>=</sup> <sup>2</sup> <sup>×</sup> *<sup>n</sup>* is the problem dimension, *<sup>n</sup>* is the number of cross sections to optimize, **<sup>g</sup>**(·) is a mapping function **<sup>g</sup>**: **<sup>x</sup>** <sup>→</sup> **<sup>e</sup>** that maps a parameter sample **x** to encroachment limits **e**, which the HEC-RAS model takes as its input parameters to compute **<sup>s</sup>** and **Fr**, **<sup>M</sup>**(·) is the HEC-RAS model, *<sup>f</sup>*(·) is the objective function, *A*fw is the floodway surface area, *A*fw,min and *A*fw,max are the minimum and maximum surface areas of the floodway, respectively as shown in Figure 4, **s** and **Fr** are the surcharge and Froude number

vectors, respectively, for all cross sections, **s***<sup>i</sup>* and **Fr***<sup>i</sup>* are the surcharge and Froude number, respectively, for cross section *i*, and *N* is the total number of cross sections in the model including *n* cross sections to optimize and those used as boundary conditions.

If the surcharges and Froude numbers of all cross sections (**s** and **Fr**, respectively) are within acceptable ranges, the objective function evaluates the ratio of the surface area difference between the current and minimum floodways (*A*fw and *A*fw,min, respectively) to the surface area difference between the maximum and minimum floodways (*A*fw,max and *A*fw,min, respectively). Provided that all the hydraulic conditions are satisfied, the values of this ratio are 0 and 1 when the floodway is at its minimum and maximum widths, respectively. If one or more of the hydraulic conditions are violated, the objective function is set to unity and extra penalties are added based on how severely the surcharge and Froude number are deviating from their respective allowable limits. The penalty functions are designed so that the magnitude of deviation from the intended surcharge range or Froude number range is directly correlated to the magnitude of penalty added to the objective function. Any penalty will force the objective function to take on a larger value, which is an undesirable trait in a minimization problem. Now, floodway optimization is defined as a mathematical equation that can be evaluated objectively by computer code. The main goal of the proposed approach is to minimize the objective function *f*(**M**(**g**(**x**))) by optimizing the variable vector **x** that represents encroachment limits **e**, using a heuristic algorithm.

To better explain how the objective function works, Table 1 shows example outputs from a simple hypothetical model and their objective function values in the Equation (1) column as well as Yu's objective function values in the Equation (2) column. This hypothetical model consists of two cross sections (*N* = 2). Its minimum and maximum floodway areas are 50 km2 and 100 km2, respectively. Trial 1 evaluates the minimum floodway area, but the surcharges and Froude numbers for both cross sections violated hydraulic conditions (i.e., **<sup>s</sup>***<sup>i</sup>* <sup>&</sup>gt; 0.305 m and **Fr***<sup>i</sup>* <sup>≥</sup> 1 for *<sup>i</sup>* <sup>=</sup> 1, 2). Since there are hydraulic violations, the penalty case of Equation (1) is used to calculate the objective function value in the Equation (1) column. That is, *<sup>f</sup>*(**M**(**g**(**x**))) <sup>=</sup> <sup>1</sup> <sup>+</sup> max(<sup>0</sup> <sup>−</sup> 0.366/0.305, 0) + max(0.366/0.305 <sup>−</sup> 1, 0) + max(1.1 <sup>−</sup> 1, 0) + max(<sup>0</sup> <sup>−</sup> 0.397/0.305, 0) + max(0.397/0.305 <sup>−</sup> 1, 0) + max(1.2 <sup>−</sup> 1, 0) = 1.8. Similarly, trials 2–3 have some violations and use the same equation to evaluate the objective function. Trial 4 does not violate any hydraulic conditions, so the acceptable case of Equation (1) is used to calculate the objective function value. In this case, *<sup>f</sup>*(**M**(**g**(**x**))) <sup>=</sup> <sup>80</sup>−<sup>50</sup> <sup>100</sup>−<sup>50</sup> <sup>=</sup> 0.6. Similarly, trials 5–6 do not have any violations and use the same floodway area ratio as their objective function values. Assuming that there are no more trial simulations, trial 4 would be the best floodway model because it minimizes the floodway area while satisfying all hydraulic requirements. In actual optimization runs, trials will evolve based on a heuristic algorithm introduced in Section 2.3.

**Table 1.** Example model outputs and their objective function values. For all trials, *N* = 2, *<sup>A</sup>*fw,min = 50 km2, and *<sup>A</sup>*fw,max = 100 km2.


For comparison, Yu's objective function [35] is defined as:

$$f(\mathbf{M}(\mathbf{g}(\mathbf{x}))) = \sum\_{i=1}^{N} \left| \frac{\mathbf{s}\_i}{0.305} - 1 \right|. \tag{2}$$

This lump-sum way of integrating absolute differences can be problematic because bad performance in some cross sections with a high surcharge can be compensated for in other cross sections with a low surcharge. This compensation prohibits Equation (2) from differentiating good trials from bad ones. For example, both cross sections in trial 3 violated the surcharge requirement (i.e., **<sup>s</sup>***<sup>i</sup>* <sup>&</sup>gt; 0.305 m for *<sup>i</sup>* <sup>=</sup> 1, 2) while the two cross sections in trial 4 did not (i.e., **<sup>s</sup>***<sup>i</sup>* <sup>≤</sup> 0.305 m for *<sup>i</sup>* <sup>=</sup> 1, 2). However, Equation (2) evaluates to 0.07 for both trials. As shown above, this objective function cannot adequately rank different trials for better optimization because information about individual cross sections gets lost. At the same time, since this objective function evaluates surcharges closer to the maximum allowed limit (i.e., 0.305 m) favorably, trials with lower surcharges are penalized even if they are actually more desirable. For example, trial 6 with both surcharges at the maximum 0.305 m evaluates to 0.00 while trial 5 with lower surcharges evaluates to 0.50. In the end, populations in a GA will evolve towards the 0.305 m limit and the surcharge requirement can easily be violated when the surcharge is too close to the allowed limit.

#### *2.3. Isolated-Speciation-Based Particle Swarm Optimization*

Isolated-Speciation-based Particle Swarm Optimization (ISPSO) [37] is a multi-modal heuristic optimization algorithm based on collective intelligence of individual particles in a swarm. In ISPSO, parameter samples are referred to as particles, which are collectively called a swarm. They fly around the parameter space and form multiple species based on spatial proximity. Individual particles keep track of their experience and share information with neighbors in the same species. Their velocities and next positions are determined by combining their private experience and neighbors' experience. In this way, particles in one species converge to a local solution. Since there are multiple species in the search space, particles are able to find multiple local solutions, possibly including the global solution as well. More details about how this optimization algorithm works, mathematical examples, and a practical engineering problem can be found in [37].

Unlike gradient-based techniques, heuristic optimization algorithms do not depend on nor require local slope evaluation of the objective function surface. Since the HEC-RAS model transforms model inputs and parameters to model outputs non-linearly, the objective function surface may not be smooth and can be very complicate. Because of this complicate nature of the objective function surface, it becomes important to avoid dependency on the landscape gradient of the objective function surface to prevent solution finding algorithms from falling into inferior local pits. Particles in ISPSO are able to find multiple solutions in different regions of the search space without getting trapped into such inferior pits. For this reason, ISPSO has successfully been applied to stochastic rainfall generation [40–43], storm tracking [44], uncertainty analysis [45,46], and climate change studies [47,48]. The current version of ISPSO is implemented in the R language [49], which was also used to evaluate the objective function and run the HEC-RAS program.

#### *2.4. Automated Floodway Optimizer for HEC-RAS*

The Automated Floodway Optimizer for HEC-RAS (AFORAS) is a tool that automatically optimizes the floodway in a HEC-RAS model using ISPSO. The system is unique in that the modeler need not make manual adjustment trying different combinations of encroachment limits until an acceptable solution is found. For fully automating the optimization procedure, the Application Programming Interface (API) of HEC-RAS was used to interface the HEC-RAS program and ISPSO's R code. A command-line program called the Command-Line Interface for HEC-RAS (CLIRAS) was developed to control the HEC-RAS program from an R environment. CLIRAS can change HEC-RAS plans, read cross section information such as river stations, bank stations, encroachment stations, flood extents, etc., update encroachment stations, and, finally, execute the HEC-RAS program. These features of CLIRAS are essential in controlling the HEC-RAS program without manual user interventions and for the ISPSO R code to be able to execute the HEC-RAS model for a specified

number of times non-interactively. ISPSO executes CLIRAS internally to update and run the HEC-RAS model and extract results from it to evaluate the objective function.

AFORAS integrates the ISPSO R code, CLIRAS, and HEC-RAS as shown in Algorithm 1. The maximum number of iterations itermax tells ISPSO the total number of iterations to perform for optimization. The maximum number of HEC-RAS model runs is defined by itermax times the swarm size *<sup>S</sup>* (itermax <sup>×</sup> *<sup>S</sup>*). The HEC-RAS model is represented by **<sup>M</sup>**(·) and requires that two plans be defined: (1) 100-year floodplain and (2) floodway. The boundary conditions specify how the downstream or upstream end of the floodway should tie into adjacent existing floodways. There are four possible boundary conditions: (1) No existing floodways at the upstream and downstream ends of the study reach (BC = None), (2) floodway only at the downstream end (BC = DS), (3) floodway only at the upstream end (BC = US), and (4) floodways at both ends (BC = Both). The boundary conditions fix the encroachment limits at the most upstream, downstream or both cross sections, and therefore the problem dimension can be determined based on the number of cross sections and the number of boundary conditions. For example, when there are no upstream or downstream floodways to tie into (BC = None), all cross sections should be optimized, whereas the number of cross sections to optimize reduces by either 1 or 2 if one boundary condition (either BC = DS or BC = US) or two boundary conditions (BC = Both) are specified, respectively. The number of cross sections to optimize is indicated by *n* and the problem dimension is the total number of left and right encroachment limits on those *<sup>n</sup>* cross sections, which is *<sup>D</sup>* <sup>=</sup> <sup>2</sup> <sup>×</sup> *<sup>n</sup>*. The recommended swarm size of *<sup>S</sup>* <sup>=</sup> <sup>10</sup> <sup>+</sup> 2 <sup>√</sup>*<sup>D</sup>* [50] was used.

**Algorithm 1** Pseudocode for automated floodway optimization for HEC-RAS.

**Require:** itermax - Maximum number of iterations **Require: M**(·) - HEC-RAS model with 100-year and floodway plans and profiles **Require:** BC ∈ {None, DS, US, Both} - Boundary conditions for the encroachment limits Extract cross section information from **<sup>M</sup>**(·) *N* ← Number of cross sections *n* ← *N*− Number of boundary conditions *D* ← 2 × *n* - Problem dimension *<sup>S</sup>* <sup>←</sup> <sup>10</sup> <sup>+</sup> 2 <sup>√</sup>*<sup>D</sup>* - Swarm size *A*fw,min, *A*fw,max ← Minimum and maximum possible areas of the floodway **<sup>X</sup>** <sup>∈</sup> [0, 1] *<sup>S</sup>*×*<sup>D</sup>* <sup>←</sup> *<sup>S</sup>* number of *<sup>D</sup>*-tuples randomly sampled from [0, 1] *<sup>D</sup>* - Initial population Let **g**: [0, 1] *<sup>D</sup>* <sup>→</sup> <sup>R</sup>*<sup>D</sup>* that maps particles to encroachment limits iter ← 1 **repeat** - ISPSO loop **for** *i* ← 1, . . . , *S* **do X***<sup>i</sup>* ← Row *i* from **X** *i* th trial encroachment limits or particle *i* in ISPSO Simulate **<sup>M</sup>**(**g**(**X***i*)) using CLIRAS - Execute the HEC-RAS program Evaluate *<sup>f</sup>*(**M**(**g**(**X***i*))) - Equation (1) **if** *<sup>i</sup>* = 1 or *<sup>f</sup>*(**M**(**g**(**X***i*))) < *<sup>f</sup>*(**M**(**g**(**X**best))) **then** - If **X***<sup>i</sup>* is better than **X**best **X**best ← **X***<sup>i</sup>* - Store the best encroachment limits from the current iteration **end if end for if** iter = 1 or *<sup>f</sup>*(**M**(**g**(**X**best))) < *<sup>f</sup>*(**M**(**g**(**x**best))) **then** - If **X**best is better than **x**best **x**best ← **X**best - Store the best encroachment limits so far **end if** Evolve **X** using ISPSO - Evolution of the swarm in ISPSO iter <sup>←</sup> iter <sup>+</sup> <sup>1</sup> **until** iter = itermax or other conditions are satisfied Optimized encroachment limits <sup>←</sup> **<sup>g</sup>**(**x**best) -Found the best encroachment limits

The valid ranges of the floodway encroachment limits vary from cross section to cross section as shown in Figure 4 (dot–diamond). As can be seen in Figures 1 and 4, even for the same cross section, the left and right encroachment limits can have different scales. Because our objective is to find the best feasible left and right encroachment limits, the problem space is constructed from *D* encroachment limits. However, different scales in different encroachment limits highly skew the problem space, which can negatively affect the performance of optimization. To address this issue, the problem space is normalized to [0, 1] *<sup>D</sup>* using a mapping function **<sup>g</sup>**(·). The mapping function **<sup>g</sup>**(·) is defined such that it transforms a particle (i.e., a trial set of normalized encroachment limits) in a hypercube search space [0, 1] *<sup>D</sup>* back to a *D*-tuple of encroachment limits **e**, which is a direct input to the HEC-RAS model **<sup>M</sup>**(·). The particle at the lower limits of all dimensions (i.e., *xi* <sup>=</sup> 0 for 1 <sup>≤</sup> *<sup>i</sup>* <sup>≤</sup> *<sup>D</sup>*) defines the minimum possible floodway (dashed polygon in Figure 4) determined by the bank stations (dots in Figure 4) while the particle at the upper limits (i.e., *xi* <sup>=</sup> 1 for 1 <sup>≤</sup> *<sup>i</sup>* <sup>≤</sup> *<sup>D</sup>*) defines the maximum possible floodway (unfilled solid polygon in Figure 4) determined by the 100-year floodplain extents (diamonds in Figure 4). In other words, bank stations are mapped to *xi* = 0 while 100-year floodplain extents are mapped to *xi* = 1. In this way, the floodway is guaranteed to be wider than the main channel such that the conveyance of flow is not highly obstructed. At the same time, the floodway cannot be wider than the 100-year floodplain. Floodway areas including the minimum and maximum areas *A*fw,min and *A*fw,max, respectively, are calculated by straightening the reach and calculating the area of the polygon defined by the bank stations for *A*fw,min or the 100-year floodplain extents for *A*fw,max. These minimum and maximum floodway areas *A*fw,min and *A*fw,max, respectively, are used in the objective function in Equation (1) to assess the fitness of the floodway defined by trial encroachment limits **<sup>g</sup>**(**X***i*). Once the problem is defined, ISPSO initializes the swarm and starts evolving the particles by evaluating trial encroachment limits. The final solution **<sup>e</sup>**best = **<sup>g</sup>**(**x**best) is a set of optimized left and right encroachment limits, which then becomes input for the HEC-RAS floodway model.

#### *2.5. Numerical Experiments*

Three different approaches of producing the floodway were compared: (1) automated optimization by AFORAS, (2) manual determination by the authors, and (3) the default floodway in the model as a reference. For fair comparisons, the model was not modified except that floodway encroachment limits were updated. To update floodway encroachment limits using Algorithm 1, the mapping function **<sup>g</sup>**(·) is defined to normalize encroachment limits in the HEC-RAS model into [0, 1] *<sup>D</sup>*. To define the mapping function **<sup>g</sup>**(·), the bank stations and 100-year flood extents were extracted and normalized to create a unit hypercube search space [0, 1] *<sup>D</sup>* where *D* is two times the number of cross sections to be optimized. After each iteration of evolution in Algorithm 1, particles in the problem space [0, 1] *<sup>D</sup>* are transformed back to actual floodway encroachment limits, which are in turn input into the HEC-RAS model for evaluating the objective function in Equation (1).

The number of cross sections to be optimized varies depending on the boundary condition. Four encroachment boundary conditions as defined in Section 2.4 were tested: (1) BC = None, (2) BC = DS, (3) BC = US, and (4) BC = Both. The upstream-most and downstream-most encroachment limits of the default floodway in the reference model were used as those encroachment boundaries that an optimized floodway is to be tied into. Only in this way, the default model need not be updated at all and can be used as a reference model for the other two approaches.

Numerical experiments were conducted using the floodway model to observe how AFORAS performs in different boundary conditions. A total of 30 independent runs of each boundary condition were performed. Based on Clerc's suggestion [50] and the number of dimensions in this problem, the recommended swarm size of 18 was used for the boundary condition BC = Both while 19 was used for the other three boundary conditions. For each AFORAS run, total 1000 iterations were performed resulting in total 18,000 and 19,000 model runs for the boundary condition BC = Both and the other boundary conditions, respectively.

#### **3. Results and Discussion**

#### *3.1. Comparison of Different Approaches*

The first AROFAS run for each boundary condition was used to compare the floodways between the three approaches. Table 2 shows the problem dimensions and the objective function values of the final floodways for the three approaches and four boundary conditions. The objective function value for the HEC-RAS floodway remains the same because its floodway was used as a reference and was not independently optimized. Since floodway optimization is a minimization problem, a floodway with the least objective function value is preferred. As can be seen in Table 2, among all the AFORAS floodway runs, the case with BC = Both yielded the worst objective function value (0.338), it but still performed better than the best floodway from the other two approaches (0.342). The numbers inside parentheses show that AFORAS reduced the objective function value by 1–29% from the manual method and by 12–40% from the HEC-RAS method. In terms of the floodway area, these numbers indicate the reduction of the floodway area outside the channel (*A*fw <sup>−</sup> *<sup>A</sup>*fw,min), not the total floodway area (*A*fw). On average for all eight cases (i.e., manual and HEC-RAS cases), the reduction of the floodway area outside the channel was 20%, which can significantly increase encroachment areas for development.

**Table 2.** Objective function values for the test cases with different boundary conditions. Since Automated Floodway Optimizer for HEC-RAS (AFORAS) solves a minimization problem, lower values represent better models. The numbers inside parentheses indicate what percent of the objective function value could be improved by running AFORAS.


Despite high problem dimensions from 20 to 24, AFORAS performed reasonably well and even outperformed the manual optimization. An interesting observation is that the performance improved as the problem dimension increased with more cross sections to optimize. When either the upstream-most or downstream-most cross section of the floodway is specified as a boundary condition, the fixed width of the encroachments in those cross sections prevents the floodway area from being further reduced beyond its boundary condition limits.

Figure 5 shows the final floodway for all cases with different boundary conditions. Vertical lines at encroachment station 0 represent the straightened river while negative and positive stations represent the left and right floodway encroachment limits on both sides of the river. For the case with BC = None, at the downstream-most cross section at river station 9.66 km, the floodway is narrower than those of the cases that have a restriction to the same cross section as a boundary condition (i.e., BC = DS and BC = Both). Similarly, for the case with BC = None, at the upstream-most cross section at river station 8.05 km, the floodway is narrower than those of the cases with BC = DS and BC = Both. Since these boundary conditions affect the objective function, lower objective function values in the BC = None case do not necessarily mean that AFORAS performs better when there are no boundary conditions specified. However, this result shows that AFORAS was able to take full advantage of cross sections that are not constrained to help reduce the footprint of the floodway and performed consistently better than the other two approaches.

**Figure 5.** Graphical views of the left and right encroachment limits for the four different boundary conditions. Vertical lines at encroachment station 0 represents the river line. Negative and positive stations represent the left and right encroachment limits, respectively. Straightened river, Automated Floodway Optimizer for HEC-RAS (AFORAS) floodway, Manual floodway, and Hydrologic Engineering Center's River Analysis System (HEC-RAS) floodway.

#### *3.2. Sensitivity of Encroachment Limits to the Boundary Condition*

In optimization, it is beneficial to visualize the landscape of the objective function surface to be able to understand a problem. However, since most of floodway optimization problems deal with more than one cross section, their problem dimensions are usually higher than two, which makes it difficult to visualize the objective function surface without projecting it onto a much smaller-dimensional space. Figures 6 and 7 show two-dimensional projections of 10,000 particles with grayscaled objective function values satisfying hydraulic requirements for BC = None and BC = Both, respectively. These plots

show how sensitive particles were to the performance of optimization. The more spread out particles are along one encroachment limit, that encroachment limit was less sensitive to the performance because particles did not converge to a narrower band of the dimension. For example, in Figure 6e,f, particles tend to spread out almost randomly across the entire projected space. This widespread presence of particles means that they did not have any preference over specific encroachment limits in these cross sections at 8.67 km and 8.71 km during optimization and they were very insensitive to the performance compared to those in the other cross sections. Ideally, AFORAS should be able to take advantage of this insensitivity of particles to the performance of floodway encroachments in these two cross sections. The result of BC = None in Figure 5a (the very first run for this boundary condition) shows this behavior where AFORAS decided to take very narrow floodway encroachments as the final solution at these two cross sections. Since a bridge is located between these cross sections, ineffective flow areas were established around the bridge structure to model embankment areas. The areas immediately before and after the structure prevent the water from flowing freely downstream and result in ineffective flow areas. As a result, floodway encroachments within these ineffective flow areas could not affect the flood elevation significantly because those areas blocked by the floodway encroachments were already ineffective.

As shown in Figure 6a–d,g–k, in other cross sections except the upstream-most cross section, particles spread out more along the left encroachment axis than along the right encroachment axis. The narrow distribution of particles in the right encroachment axis means that the right encroachment limit was more sensitive to the performance and played a bigger role than the left encroachment limit during optimization. This result can be explained by the relative width of possible encroachment areas on the left overbank versus those on the right overbank. As shown in Figure 2, the portions of cross sections on the left side of the river line (i.e., left overbanks) are much narrower than those on the right side (i.e., right overbanks), especially at the downstream area. Because the encroachment widths on both overbanks are normalized to create a unit hypercube search space, wide-spread particles across the left encroachment do not necessarily mean that the left overbank area is wider. Also, for the same reason, significant movements of particles along the left encroachment dimension only slightly affect actual changes in the encroachment width, effectively making the left encroachment insensitive to the performance. However, since the left overbank gradually expands as it traverses upstream, this effect started slowly diminishing from 8.84 km.

The upstream-most cross section in Figure 6l showed a similar pattern to those around the bridge structure at 8.67 km and 8.71 km. Its insensitivity to the performance is due to the fact that HEC-RAS computes the flood elevation from downstream to upstream. Since the flood elevation within the floodway encroachment at the upstream-most cross section does not affect any downstream flood elevations, encroachment limits at this cross section may be placed anywhere as long as the hydraulic criteria are satisfied. For AFORAS, this insensitivity means that it should pick the narrowest possible floodway, which can be seen in the result in Figure 5. What it also means is that the problem with BC = None can be simplified by constraining the encroachment limits in the upstream-most cross section to the bank stations and solving the problem as if it had the boundary condition BC = US. In this way, the problem dimension of BC = None can be reduced by two and AFORAS should be able to achieve better performance and a faster convergence rate.

Two-dimensional plots for the boundary conditions BC = DS and BC = US are not presented in this paper because of space limitations and a prohibitively large number of data points. However, cross sections from 8.26 km to 9.27 km in these boundary conditions showed similar patterns to Figure 6c–j or Figure 7c–j. Cross sections at 8.05 km and 8.15 km in the boundary condition BC = US behaved similarly to Figure 6a,b, respectively (i.e., both cases do not constrain the downstream-most cross section). Also, cross sections at 9.45 km and 9.64 km in the boundary condition BC = DS behaved similarly to Figure 6k,l, respectively (i.e., both cases do not constrain the upstream-most cross section).

**Figure 6.** Two-dimensional projections of all particles **<sup>X</sup>***<sup>i</sup>* <sup>∈</sup> [0, 1] *<sup>D</sup>* for 1 <sup>≤</sup> *<sup>i</sup>* <sup>≤</sup> *<sup>S</sup>* satisfying the three hydraulic criteria from all 30 AFORAS runs for BC = None. Since the total number of particles meeting hydraulic requirements was excessively large—130,395 out of 570,000 (30 AFORAS runs × 19, 000 model runs/AFORAS run)—10,000 particles were sampled to construct each subplot, which represents one cross section (XS). Particles that perform better are plotted darker in front of those that perform worse and are in a lighter gray.

When the downstream-most cross section is constrained (i.e., BC = DS or BC = Both), the right encroachment became even more sensitive to the performance at 8.15 km, just upstream of the downstream boundary cross section, and its encroachment plot for BC = DS is very similar to Figure 7b. This neighbor cross section was highly affected by the downstream boundary condition because of the backward calculation of HEC-RAS. When the upstream-most cross section is constrained (i.e., BC = US or BC = Both), AFORAS was able to achieve faster convergence by avoiding unnecessary trial-and-error sampling in the upstream-most cross section, which does not affect any other cross sections at all. The encroachment plot for 9.45 km, just downstream of the upstream boundary condition, became much narrower than Figure 6k with most particles clustering around the diagonal line from top-left to bottom-right as shown in Figure 7k. Table 3 summarizes these observations of two-dimensional projection subplots of all particles from 30 AFORAS runs.

**Figure 7.** Two-dimensional projections of all particles **<sup>X</sup>***<sup>i</sup>* <sup>∈</sup> [0, 1] *<sup>D</sup>* for 1 <sup>≤</sup> *<sup>i</sup>* <sup>≤</sup> *<sup>S</sup>* satisfying the three hydraulic criteria from all 30 AFORAS runs for BC = Both. Since the total number of particles meeting hydraulic requirements was excessively large—142,937 out of 540,000 (30 AFORAS runs × 18, 000 model runs/AFORAS run)—10,000 particles were sampled to construct each subplot, which represents one cross section (XS). Particles that perform better are plotted darker in front of those that perform worse and are in a lighter gray.

**Table 3.** Summary of two-dimensional projection subplots of all particles by Figures 6 and 7. A ∼ symbol indicates a similar pattern to the subplots given on the right. Figures 6c–j and 7c–j, respectively, have a similar pattern.


#### *3.3. Optimization Performance*

The convergence lines of all 30 AFORAS runs for different boundary conditions are presented in Figure 8 as 30 gray lines with the average performance as a black line. All the runs converged exponentially to the final values of the objective function. The minimum and maximum of the final values are 0.265 and 0.280, respectively, and their mean and standard deviation are 0.272 and 0.004, respectively. The standard deviation is approximately 1.5% of the mean, which indicates that the performance of AFORAS is very robust and reliable. The black line shows the mean of the 30 cumulative minimum values of the objective function. The average performance converges very quickly until 5000 model runs and experiences gradual improvement until about 15,000 model runs for BC = None and BC = DS, and 10,000 model runs for BC = US and BC = Both, after which the convergence rate slowed down significantly. Figure 8 shows that AFORAS achieved a faster convergence rate for the cases where the upstream-most cross section is constrained as a boundary condition. However, their objective function values are higher or worse than those for the other two cases. The higher objective function values for BC = US and BC = Both is because the upstream-most cross section used as a boundary condition is fairly wide compared to the bank stations, which act as the lower limits in that cross section for the other cases. Particles in BC = None and BC = DS where the upstream-most cross section was also optimized, found a very narrow floodway width at this cross section as the final solution. On the other hand, this cross section was fixed at a much larger width as a boundary condition for BC = US and BC = Both and produced worse objective function values. Since the two sets of boundary conditions have different search spaces, the higher objective function values do not necessarily mean poor performances.

**Figure 8.** Cumulative minimum values of the objective function vs. the number of model runs. The gray lines and black line represent 30 runs of AFORAS and the mean of those runs, respectively.

#### *3.4. AFORAS as a Tool for Floodway Optimization*

To the authors' knowledge, Yu's study [35] was the first attempt to determine the floodway encroachment limits using HEC-RAS and a GA, but his approach did not consider the floodway area and subcritical flow state. Also, as we compared our objective function to his in Section 2.2, the lump-sum way of integrating information across all cross sections leads to information loss and a lack of differentiation between good and bad models. AFORAS takes a different approach for floodway determination by taking into account the flow state and minimizing the floodway area. At the same time, its objective function is formulated to separate out favorable models from those with hydraulic violations.

AFORAS integrates ISPSO and HEC-RAS for reach-wide optimization of the floodway area. The derivative- free search of ISPSO and a mathematical representation of floodway area optimality were integrated as AFORAS in a way that HEC-RAS can be executed automatically by computer code without user interventions. AFORAS was successful in seeking and improving the floodway encroachment limits for the case study. As discussed above, AFORAS was able to identify those cross sections that are insensitive to its performance and almost fully encroached the floodway at these locations so that the floodway area was kept to the minimum. Overall, AFORAS performed consistently better than the manual approach and the reference floodway, and showed reliable and consistent performance across different boundary conditions. It is also interesting to see that ISPSO, as a heuristic algorithm, is able to reliably solve high dimensional problems in this study. As summarized in Table 2, the problem dimension for this study varied from 20 to 24. Different problem dimensions result in different landscapes of the objective function, but, as shown in Figures 6 and 7, two-dimensional projections onto individual cross sections of the objective function surface exhibit similar patterns depending on which cross section was used as a boundary condition. Despite the differences in the surface of the objective function and its final values, convergence was achieved for all cases and the solutions found by AFORAS minimized the floodway area while keeping the surcharge and flow state within the allowable limits. These observations suggest that AFORAS can be a suitable tool for reach-wide floodway optimization.

While AFORAS can be a good candidate for reach-wide floodway optimization, it is only limited to handle reaches with subcritical flows. The proposed objective function within AFORAS has to be modified to accommodate cases with mixed or supercritical flows, which are not as common as subcritical flows already handled by AFORAS. More tests are needed to observe the performance of ISPSO given a different objective function that can handle varying flow states. Also, when calculating the floodway area, AFORAS currently assumes that the river line is straight and the floodway encroachment limits vary linearly between consecutive cross sections. While this simplification of geometries may be a reasonable assumption for optimization, a more realistic evaluation of the floodway area can be attained by incorporating the natural curvature of the river line. Finally, a sensitivity analysis needs to be performed to see how the numbers of particles and iterations for an ISPSO run affect the performance of AFORAS and find the right balance between computing time and the performance. Better a priori estimation of the total number of model runs may help reduce computing time. Future work and research on AFORAS will include addressing these limitations and running more HEC-RAS floodway models with different geometries and structures.

#### **4. Conclusions**

Since the floodway is an essential part of hydrologic and hydraulic studies of riverine flooding, in the United States, FEMA requires one to be determined for developed communities using their approved computer programs, one of which is HEC-RAS. HEC-RAS has widely been used for flood risk and floodway regulation studies by many researchers. Because the floodway encroachment area is often used for human activities, it is a local government's interest to expand this area by minimizing the floodway footprint area that conveys the flood water without affecting the water surface elevation too much. Our objective is to minimize the floodway area while maintaining the

surcharge and subcritical flow state reach-wide, both of which are required by FEMA. The authors' literature review has revealed that very little work has been done in terms of floodway optimization. A recent attempt to determine the floodway encroachment limits using HEC-RAS and a GA did not consider the floodway area and subcritical flow state, which most of the streams in the United States exhibit. The proposed objective function takes into account the floodway area, surcharge, and subcritical flow state to make sure that the final optimized floodway not only meets FEMA's hydraulic requirements but also maximizes floodway encroachment areas for human activities in a reach-wide manner. By integrating the objective function and a heuristic algorithm called ISPSO, we proposed a floodway area optimization tool named AFORAS for reach-wide optimization of the floodway using HEC-RAS. We used a readily and freely available floodway model from the HEC-RAS 4.1.0 installation for a case study so that other researchers can replicate our results if they decide. Comparisons of the AFORAS, manual, and HEC-RAS approaches showed 1–40% improvements in the objective function value by AFORAS. AFORAS consistently provided superior results for all the boundary conditions. We also conducted a sensitivity analysis of encroachment limits to the boundary condition and a convergence test by running AFORAS 30 times for four different boundary conditions. Both left and right encroachment limits were insensitive to the performance in cross sections adjacent to a bridge structure while these encroachment limits exhibited different level of sensitivity to the performance in other cross sections. Because of the bridge opening and ineffective areas, encroaching these cross sections could not affect the flood elevation much and did not help improve the objective function compared to the other cross sections. The surface of the objective function may vary significantly for different HEC-RAS models or even for different combinations of the boundary conditions in the same floodway model. In this regard, it is advantageous for AFORAS to employ ISPSO over gradient-based optimization techniques because of the capability of ISPSO to solve high dimensional problems without requiring the derivative of the objective function. Limitations in the current AFORAS method include the lack of support for mixed and supercritical flows in the objective function, and the linear approximation of the river geometry and floodway. Also, since the total number of HEC-RAS runs has to be specified a prior, a quantitative analysis would be beneficial to reduce computing time by estimating how many model runs are required in advance. Addressing these limitations and recommending the required number of model runs a prior will be left for research in the near future.

**Author Contributions:** Conceptualization, T.M.Y.; Methodology, H.C. and T.M.Y.; Software, H.C.; Validation, T.M.Y., H.C. and J.H.; Formal Analysis, T.M.Y. and H.C.; Investigation, T.M.Y.; Resources, J.H.; Data Curation, H.C.; Writing—Original Draft Preparation, H.C., T.M.Y. and J.H.; Writing—Review & Editing, H.C., T.M.Y. and J.H.; Visualization, H.C.; Supervision, T.M.Y.; Project Administration, T.M.Y.; Funding Acquisition, N/A.

**Funding:** This research received no external funding.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


c 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

## *Article* **Intelligent Storage Location Allocation with Multiple Objectives for Flood Control Materials**

#### **Wei Wang 1,2, Jing Yang 3, Li Huang 4,\*, David Proverbs <sup>5</sup> and Jianbin Wei <sup>6</sup>**


Received: 16 June 2019; Accepted: 18 July 2019; Published: 25 July 2019

**Abstract:** Intelligent storage is an important element of intelligent logistics and a key development trend in modern warehousing and logistics. Based on the characteristics of flood control materials and their intelligent storage, this study established a flood control material storage location allocation model reflecting the multiple objectives of retrieval efficiency and shelf stability and used a weighting method to transform a multi-objective optimization problem into a single-objective optimization problem. We then used the facilities and equipment planning and storage location allocation in the intelligent storage area for provincial flood control materials at the Zhenjiang warehouse of the Jiangsu water conservancy and flood control material reserve center as a case study. Empirical analysis was performed and used the genetic algorithm and Matrix Laboratory (MATLAB) software to optimize the storage location allocation of provincial flood prevention supplies at this warehouse, and it achieved effective results.

**Keywords:** flood control materials; intelligent warehousing; location allocation; multi-objective optimization

#### **1. Introduction**

While water resources play an important role in human society, the world suffers from all types of natural disasters [1]. In many countries, floods are the most likely natural disaster, and compared with other natural disasters, they are easily predicted and prevented [2]. Therefore, it is very important for flood disaster management departments to formulate scientific and comprehensive plans for flood control. Consequently, government agencies spend massive amounts of money and manpower on flood control and rescue.

Floods are easily predicted, which means that preparations can be made to prevent flooding from actually happening. Developed countries such as England, the United States, Holland, Denmark, Germany, France, Belgium, and Austria have invested significant time and effort into flood control and disaster mitigation research since the 1960s and have achieved significant results in meteorology, hydrology, water conservancy, water quality, topography, the influence of social and economic activities on the prevention of flood damages, the runoff-conflux model of floods and its prediction and analysis, and other fields. During the process, the Wallingford National Institute of England, the Delft Hydraulics Institute, the Danish Hydraulic Institute, (DHI, Hesholm, Denmark), the United States Environmental Protection Agency, (EPA, Washington, DC, USA) and Army Corps of Engineers, (ACE, Washington, DC, USA), and other research institutes have stood out [3,4]. In the 1970s, the United States proposed

flood control that used non-structural measures, which realized flood management through legislation, flood forecasting, flood dispatching, flood detention, flood insurance, floodplain management, and soil and water conservation, developed flood contingency plans and other methods, standardized people's precautions and preparations against heavy rain and flooding, and provided guidance according to actual circumstances so that the losses from floods could be reduced and better social and economic benefits could be achieved. Compared with structural measures, non-structural measures were no longer the focus of flood control; instead, they emphasized the timely and scientific implementation of flood control command and dispatch through the collection, analysis, and processing of relevant flood information. This was done in order to improve the potential capability of the existing flood-fighting structural measures and standardize people's actions against floods and the developmental activities within the floodplain, thus realizing flood control and disaster mitigation [5]. Non-structural measures were important supplements to structural measures.

Among these non-structural measures, flood control material and emergency logistics were important for fighting floods. Flood control material is one of the three major elements of flood control and rescue, including material for preparation, response, and mitigation phases, and the reserve management of flood control materials is the key link in carrying out flood control work and plays an irreplaceable role in national flood control and rescue work. However, flood emergency preparedness lacks logistical insights [6,7], and relevant scholars have already conducted some primary studies. Garrido, Lamas, and Pino (2015) put forward a flood logistics model. The model attempts to optimize emergency supply inventories and vehicle availability [8]. Leeuw, Vis, and Jonkman (2012) developed an emergency logistics framework that supports preventing catastrophic breaches of flood defenses during extreme situations [9]. Alem, Clark, and Moreno (2016) developed a new two-stage stochastic network flow model to determine how to rapidly supply humanitarian aid to victims of floods [10].

Intelligent storage is an important element of intelligent logistics and a key development trend in modern warehousing and logistics. The effectiveness of intelligent storage is chiefly influenced by the storage location allocation, and optimal storage location allocation can enhance the storage space utilization rate, shorten the storage and retrieval distances and times, accelerate the turnover of goods, ensure inventory stability, and increase the operating performance of an intelligent storage system.

The key to effective storage is optimizing the storage location allocation strategies, and scholars have performed in-depth, systematic studies of storage location allocation. Roodbergen et al. analyzed storage location strategies [11] and found that the storage location allocation methods of commercial intelligent storage systems chiefly consist of the following types: fixed storage location assignment [12–15], random storage location assignment [16–18], class-based storage location assignment [19,20], random class-based storage location assignment [21], and shared storage location assignment. Most of these studies assume that a warehouse is initially empty, which is clearly not in accordance with the needs of real projects. Furthermore, even when the studies in the literature consider a warehouse with a non-empty initial state, they perform simulation experiments only involving a single batch of goods entering the warehouse. However, available empty storage locations will be abundant when a single batch of goods is put into storage, and there will thus be considerable freedom when selecting an optimal storage location. Consequently, it is possible that only sub-optimal locations will be left for the next batches of goods entering storage, and the arrangement of the storage locations as a whole will be irrational. Lee et al. proposed similarity coefficients to cluster goods and then assigned goods to storage locations in accordance with the clustering results [15]. As for algorithms, most studies have adopted intelligent optimization algorithms, such as the genetic algorithm [22], the simulated annealing algorithm [23], the tabu search method [24], the data mining-based algorithm [25], and other algorithms [26,27], which can greatly shorten the computing time and enable optimal solutions to be found after relatively few iterations.

The foregoing review reveals that most scholars have constructed intelligent storage systems via a systems engineering approach and synergistically applied technologies, including message identification, communications technology, automatic control, and intelligent algorithms, in their intelligent storage systems. In particular, many scholars have used mathematical modelling to construct mathematical models of intelligent storage systems and then used simulation experiments to validate the models [28–30]. Furthermore, operation optimization methods are most commonly used in intelligent storage analysis and decision-making [31], and the development of commercial intelligent storage location allocation methods has provided a foundation for research and development (R&D) related to intelligent storage for emergency logistics.

Because emergency supplies tend to be infrequently used, intelligent storage systems are seldom used for emergency supplies, and therefore, there has been little research on storage location allocation in intelligent emergency logistics [32–36]. The emergency supply reserve warehouses of some power agencies have used the Internet of Things to construct intelligent storage systems for their emergency power supplies [37], which has enabled substantial increases in storage automation and storage and retrieval performance, reduced injuries and supply losses, and enhanced the emergency response capabilities.

There has been relatively little research on the application of intelligent storage in emergency logistics, and no research on key intelligent storage technologies for emergency disaster relief supplies has yet been published. Furthermore, because intelligent storage systems have not been used for flood control, there have been few studies concerning this aspect. However, the research on the key intelligent storage technologies that are used in commercial logistics and particularly the research on the storage strategies that are used in intelligent storage systems are already quite mature. However, because of the characteristics of the storage and management of flood control materials, including limited types of supplies, large quantities, low batch numbers, large quantities in a batch, the need for quick retrieval in the event of an emergency, and strong constraints on response times, the findings of the research on the application of intelligent storage systems to commercial logistics are not fully applicable to flood control materials. Therefore, there is an urgent need for research on the practical application of technologies for the intelligent storage of emergency flood control materials.

Based on the characteristics of flood control materials and their intelligent storage, this study established a flood control material storage location allocation model with the multiple objectives of retrieval efficiency and shelf stability and used a weighting method to transform a multi-objective optimization problem into a single-objective optimization problem. The study then optimized the storage location allocation of provincial flood prevention supplies using MATLAB. The optimized allocation can comprehensively improve the support ability, utilization efficiency, technical level, and management level of the flood control material reserve system, reduce the losses and hazards caused by emergencies, and achieve remarkable social and economic benefits.

#### **2. Analysis of the Storage Strategies for the Intelligent Storage of Flood Control Materials**

After summarizing the commonly used storage strategies (including fixed location storage, random location storage, class-based storage, class-based random storage, shared storage, and item-location coupling storage) for location assignment, and considering the attributes of flood control materials, we decided to adopt the class-based fixed location storage [1].

Storage strategies constitute the major principles of flood control material storage area planning and must be naturally combined with storage location allocation principles in order to determine a storage operating model. To scientifically and rationally implement the storage of flood control materials, the storage location allocation for such materials must adhere to the following principles:


#### **3. Intelligent Storage Location Allocation Model for Flood Control Materials**

#### *3.1. Construction of a Multi-Objective Optimization Model*

The types of the flood control materials include motor oil, life jackets, tents, boats, portable lights, etc., and the material specifications cover the lengths, widths, heights, and importance of the materials. Here, in order to simplify the problem, a certain number of single-class materials will be grouped as a single standard pallet group based on their size. The combination is the standard pallet group of the material and the length, width, height, and weight of the standard pallet group of the various materials satisfy the shelf restrictions.

The paper takes single-row shelves as the research target, and we establish a multi-layer shelf coordinate system in the optimization module with the starting position of the work vehicle at the exit table as the origin *O*(0,0), the *X-*axis is the longitudinal axis direction of the shelf, the *Y*-axis is the vertical direction of the shelf, each column and each layer of the shelf is one unit in length in the *X*-axis and *Y*-axis directions, and the position of the materials is determined as *P*(*i*,*j*).

The main variables and parameters are defined as follows.

We assume that a certain set of shelves in a warehouse has *n* levels and *m* rows, and the position of the materials is determined as *P*(*i*,*j*) (0 ≤ *i* ≤ *m*, 0 ≤ *j* ≤ *n*).

The flood control materials include *r* total categories; *Wk* is the material weight (single standard pallet group weight) of the type-*k* material (0 ≤ *k* ≤ *r*); *tij* is the time that is needed for a forklift to transport the goods at the storage location at the *i*th row and *j*th level; *vx* and *vy* are, respectively, the horizontal and vertical operating speeds of the shuttle vehicles or forklifts; *L* and *H* are respectively, the length and height of a storage location; *pk* is the calling frequency of the materials (0 ≤ *k* ≤ *r*), the total number of times the materials are used within a certain period of time, which is equivalent to the number of times that the materials are retrieved divided by the time.

We define the decision variable as *xijk*. When the type-*k* material (0 ≤ *k* ≤ *r*) is stored in *P*(*i*,*j*), *xijk* = 1; otherwise, *xijk* <sup>=</sup> 0, where 0 <sup>≤</sup> *<sup>i</sup>* <sup>≤</sup> *<sup>m</sup>*, 0 <sup>≤</sup> *<sup>j</sup>* <sup>≤</sup> *<sup>n</sup>*, 0 <sup>≤</sup> *<sup>k</sup>* <sup>≤</sup> *<sup>r</sup>*.

Based on the classification of materials, different materials should be stored in different warehouse areas, and storage location allocation must be performed in different warehouse areas. Based on storage location allocation principles, this study constructed the following model.

(1) In accordance with the principle of lighter materials on top and heavier materials on the bottom, assuming that a certain set of shelves in a warehouse has *n* levels and *m* rows, where the level closest to the floor is the first level and the row closest to the exit is the first row, the goal of storage location allocation optimization is to minimize the sum *S* of the products of the weights of the materials on pallets and the levels on which the materials are located. The first objective function of the shortest optimization objective function is as follows:

$$
tau S = \sum\_{i=0}^{m} \sum\_{j=0}^{n} \sum\_{k=1}^{r} \mathcal{W}\_k \mathbf{x}\_{ijk} (i - 1) \tag{1}
$$

(2) In accordance with the principles of close access and quick turnover, minimizing the sum *T* of the transport times of the materials in each storage location and minimizing the sum of the products of the usage frequency of each material and the forklift operating time when retrieving the material. The second objective function that optimizes the shelf stability is as follows:

$$
\min T = \sum\_{i=0}^{m} \sum\_{j=0}^{n} \sum\_{k=1}^{r} t\_{ij} x\_{ijk} p\_k \tag{2}
$$

Based on the weighting method, the weights α and β (where 0 ≤ α ≤ 1, 0 ≤ β ≤ 1, and α + β = 1) of the fusion model are determined according to the importance of the shortest delivery time and shelf stability. The two objective functions for the shortest delivery time and shelf stability are merged, and the final warehouse optimization multi-objective model is established as follows:

$$\text{minim} = \alpha \sum\_{i=0}^{m} \sum\_{j=0}^{n} \sum\_{k=1}^{r} \mathcal{W}\_{k} \mathbf{x}\_{ijk} (i - 1) + \beta \sum\_{i=0}^{m} \sum\_{j=0}^{n} \sum\_{k=1}^{r} t\_{ij} \mathbf{x}\_{ijk} p\_{k\star} \text{ s.t.} \begin{cases} \sum\_{k} \mathbf{x}\_{ijk} = 1\\ \mathbf{x}\_{ijk} = 0 \text{ or } 1 \end{cases} \tag{3}$$

#### *3.2. Simplifying the Storage Location Allocation Model*

Storage location allocation is a composite multi-objective optimization problem, and multi-objective optimization problems are typically solved in two ways. One way uses a weighting method, a maximum method, a constraint method, or goal programming to quantitatively address the multiple objectives and obtain a unique feasible solution. The second way uses a multi-objective optimization algorithm based on artificial intelligence, such as a multi-objective genetic algorithm, an ant colony optimization algorithm, or a simulated annealing algorithm, to perform the optimization.

In this study, considering the characteristics of a flood control materials warehouse, a weighting method was used to transform the multi-objective problem into a single-objective problem. In view of the equal importance of the two objectives of the shortest delivery time and shelf stability, in this study, the two objectives were both assigned weights of 0.5, which resulted in the following objective function:

$$\min \mathbf{h} = 0.5 \sum\_{i=0}^{m} \sum\_{j=0}^{n} \sum\_{k=1}^{r} \mathcal{W}\_{k} \mathbf{x}\_{ijk} (i - 1) + 0.5 \sum\_{i=0}^{m} \sum\_{j=0}^{n} \sum\_{k=1}^{r} t\_{ij} \mathbf{x}\_{ijk} p\_{k} \tag{4}$$

Each side is multiplied by 2 to yield the final optimization model for location allocation as follows:

$$\min H = \sum\_{i=0}^{m} \sum\_{j=0}^{n} \sum\_{k=1}^{r} \mathcal{W}\_{k} \mathbf{x}\_{ijk} (i - 1) + \sum\_{i=0}^{m} \sum\_{j=0}^{n} \sum\_{k=1}^{r} t\_{ij} \mathbf{x}\_{ijk} p\_{k\prime} \text{ s.t.} \begin{cases} \sum\_{k} \mathbf{x}\_{ijk} = 1\\ \mathbf{x}\_{ijk} = 0 \text{ or } 1 \end{cases} \tag{5}$$

#### *3.3. Determining the Parameters*

#### 3.3.1. Facility and Equipment Status and Their Parameters

In accordance with the distances of shelves from the exit and their lifting heights, an optimal storage location is a location in an intelligent storage system that is at a height within 20% of the storage area that is closest to the floor and within the 20% of the storage area that is closest to the exit. These storage areas have the characteristics of easy storage, short pathways, and low mechanical operating losses. An intelligent storage system is chiefly composed of a material storage and transport system and a warehouse management system. Here, the material storage and transport system comprises shelves, the storage and retrieval entrance/exit, and warehousing equipment.


set of shelves. Because vehicles can typically drive directly into flood control material warehouses for loading or unloading, this study considered only the transport of materials from shelves to a storage and retrieval entrance/exit.

(3) Warehousing transport equipment: Only one shuttle vehicle was used for single pick-up actions and was responsible for serving one set of shelves. The shuttle vehicle was located at a fixed initial position at the storage and retrieval entrance/exit in the beginning, and the time that is needed for the shuttle vehicle to leave its initial position, complete the placement (retrieval) of goods and return to its initial position was defined as the operation time. The warehousing equipment includes a shuttle car and a forklift. The shuttle car is responsible for the horizontal work. The forklift is responsible for the vertical work and other forklift operations. The working speed of the vehicle includes the maximum idle speed of the shuttle *vx1* and the maximum speed of the shuttle load *vx2*. The vehicle's horizontal acceleration is denoted as a*x*, the forklift's vertical speed is denoted as *v*y, and the forklift's speed is denoted as *vf*. Here, we can refer to the forklift's basic operating parameters to calculate the value or use our experience to set the value.

The shuttle's movement consists of its horizontal movement, the shelves, and the use of a forklift to move vertically up and down the shelves. The speed of a loaded shuttle vehicle is different from that of an empty shuttle vehicle. Its linear acceleration rate when starting is identical to the linear deceleration rate when stopping. After the shuttle vehicle reaches its maximum speed, it maintains that speed during its operations. See Table 1 for the warehouse shelving parameters.


#### 3.3.2. Determining the Storage Location Operation Time

In the coordinate system encompassing multiple levels of shelves, the *X*-axis represents the length of the shelves, and the *Y*-axis represents the height of the shelves. Goods were stored within storage locations. The time that is needed for the transport system consisting of shuttles and forklifts to move to individual storage locations on the fixed shelves was calculated using kinematics. We constructed a time-minimization model to obtain the amount of time that is needed to access goods at each storage location.

Storage locations were designated using two-dimensional coordinates, where *x* indicated the row coordinate and *y* indicated the level coordinate. The origin *O*(0, 0) was set as the shuttle's initial location on the entrance/exit platform, and the shuttle's movement from *O* to the storage location *P*(*i*, *j*) and back again completed an operating cycle (see Figure 1).

The horizontal movement distance is calculated as follows:

$$L\mathbf{x} = \mathbf{i} \times \mathbf{l} \tag{6}$$

The vertical operating distance is calculated as follows:

$$Hy = (j-1) \times h \tag{7}$$

**Figure 1.** Operating pathway coordinate system.

The horizontal operating times were designated as *tx*<sup>1</sup> and *tx*2, where *tx*<sup>1</sup> is the time for the loaded shuttle vehicle to reach the storage location, and *tx*<sup>2</sup> is the return time of the loaded shuttle vehicle. The vertical operating time was designated as *ty*, and the time-minimization model assumed that the shuttle vehicle accelerated uniformly until reaching the maximum velocity *vy*<sup>1</sup> and then continued to move at its maximum velocity. The shuttle vehicle decelerated at a uniform rate after approaching its target storage location. Here, it was also necessary to consider the situations where the shuttle vehicle did not reach its maximum velocity when travelling a short route and where it travelled a sufficiently long route to reach its maximum velocity.

When the horizontal movement distance was too short, the shuttle vehicle could not reach the maximum velocity *vx*<sup>1</sup> before it had to decelerate uniformly from its original velocity. When the movement distance was sufficiently long, the shuttle vehicle accelerated until reaching the maximum velocity *vx*1, then moved at a uniform velocity and finally decelerated at a uniform rate. While the shuttle vehicle moved in the same manner when returning, it was loaded and therefore could not reach the maximum velocity *vx*<sup>1</sup> of the unloaded condition. At this time, the maximum velocity that it reached was *vx*2.

In the vertical direction, the forklift lifted the shuttle vehicle from the entrance/exit platform at a constant velocity of *vy*, and the operation time was calculated using the data parameters in Table 1 as follows:

$$t\_{x1} = \begin{cases} 2 \times \sqrt{\frac{l\_X}{a\_x}}, & L \ge \frac{v\_{x1}^2}{a\_x} \\ \frac{v\_{x1}}{a\_x} + \frac{l\_X}{v\_{x1}}, & L\_X > \frac{v\_{x1}^2}{a\_x} \end{cases} (x = 1, 2, \dots, m) \tag{8}$$

$$t\_{\mathbf{x}2} = \begin{cases} 2 \times \sqrt{\frac{L\_{\chi}}{a\_{\chi}}} \cdot L\_{\chi} \le \frac{v\_{\mathbf{x}2}^2}{a\_{\mathbf{x}}}\\ \frac{v\_{\mathbf{x}2}}{a\_{\mathbf{x}}} + \frac{L\_{\chi}}{v\_{\mathbf{x}2}} \cdot L\_{\chi} > \frac{v\_{\mathbf{x}2}^2}{a\_{\mathbf{x}}} \end{cases} (\mathbf{x} = 1, 2, \dots, m) \tag{9}$$

$$t\_{\mathcal{Y}} = \frac{H\_{\mathcal{Y}}}{v\_{\mathcal{Y}}}, \ (y = 1, 2, n) \tag{10}$$

*;*

When operating, the shuttle vehicle first moved vertically to the level of the target storage location and then moved horizontally to the appropriate location. The total time needed by the shuttle vehicle for a single operation was therefore calculated as follows:

$$t\_{i\bar{j}} = t\_{x1} + t\_{x2} + t\_{\bar{y}} + t\_{\bar{f}} \tag{11}$$

Taking shelves with three levels and 10 rows as an example, the operation time for the shuttle vehicle to reach each storage location is as shown in Table 2.


**Table 2.** Storage location operating time.

#### **4. Empirical Analysis**

This study empirically analyzed the storage location allocation in the intelligent storage area for flood control materials at the Zhenjiang warehouse of the Jiangsu water conservancy and flood control material reserve center [1].

#### *4.1. Case Warehouse*

To emphasize the issues in intelligent storage location allocation with multiple objectives in Chinese flood control material reserve management, a case study was conducted.

The Jiangsu Provincial Hydraulic and Flood Control Material Reserve Centre (HFCMRC) Zhenjiang Warehouse is the only central- and provincial-level flood-fighting material warehouse in Jiangsu Province. It is one of the most representative flood-fighting material reserve warehouses in China. For this reason, this study chose the HFCMRC Zhenjiang Warehouse as the case example.

Interviews, document analyses, and observations were used for the data collection in this case study. A series of face-to-face semi-structured interviews with managers and staff members from the government sector (flood control and food control materials) and the business sector (food control materials) were conducted from September 2017 to August 2018. The inventory, invocation, and warehousing data of Jiangsu provincial flood control materials were provided by HFCMRC and the HFCMRC Zhenjiang Warehouse.

#### *4.2. Retrieval of Materials in Storage*

In accordance with the types and quantities of materials that were used at the Jiangsu provincial water conservancy and flood control material reserve center and in line with the warehouse's size and shelving arrangement, we selected five types of supplies: motor oil, life jackets, tents, outboard motors, and powerful handheld flashlights. These supplies are most frequently used, have regular shapes, and can be suitably placed on multi-level shelves. See Table 3 for the specific quantities of these items in storage.



#### *4.3. Determining of the Intelligent Storage Area Location, Size, and Dimensions*

According to the general construction plan of the Zhenjiang warehouse of the Jiangsu water conservancy and flood control material reserve center, the warehouse's storage room has a minimum length and width of 30 m and 18 m, respectively. Because the location of the warehouse has not yet been determined, the warehouse's dimensions were set as 30 m × 18 m, which provided a total of 540 m2. This ensured that the design would be applicable to any storage area in the warehouse.

#### *4.4. Facility Layout and Equipment Types of Intelligent Storage Area*

#### (1) Selection of shelves

There are limited types of stored flood control materials. They have large batch quantities, are often heavy and bulky, are not easily to manually carry, are not frequently used, and must be accessed quickly when needed. In view of these characteristics, we considered the use of pallet racks, drive-in racks, shuttle racks, and cantilever racks.

The specifications of the intelligent shelf storage locations in this study were preliminarily set as 1.5 m × 1.1 m × 1.5 m. Table 4 lists the advantages and disadvantage of the various types of shelves and the other equipment required.


**Table 4.** Shelf types.

From the above types of shelves, pallet racks and drive-in racks must be equipped with forklift shuttle racks, and these warehouses must possess shuttles and forklifts. Intelligent access warehouse shelves also must be equipped with forklifts. Among these types of shelves, the order of space utilization of warehouses from large to small is as follows: intelligent access warehouse shelves, shuttle racks, drive-in racks, and pallet racks. The total valuation of shelves in descending order is as follows: intelligent access warehouse shelves, shuttle racks, drive-in racks, and pallet racks. The price of a shuttle is 100,000 yuan, and a 500 square meter warehouse needs to be equipped with two to three sets. The shelves can be customized. Here, they are specified to be 1.5 m × 1.1 m × 1.5 m. The prices of each type of shelf are different, and the price of pallet racks is the lowest. A 500 square meter warehouse needs three to four shuttles. Considering the low utilization rate of the flood control warehouse, we can rent forklifts, which can save costs and avoid idle assets. The price of a forklift is approximately 500,000 yuan. If the decision is made to purchase a forklift, a Linde forklift or Zhejiang Nori forklift is recommended. Furthermore, electric forklifts are economical and environmentally friendly and require narrower lanes compared with diesel forklifts, thereby making electric forklifts more preferable. Their price is between 60,000 yuan and 80,000 yuan.

#### (2) Selection of pallets

Different countries have different pallet specifications. The most common pallet specification in China is 1200 mm × 1000 mm, which is also one of the most common pallet specifications in Europe. These pallets are low-price, flat, wooden pallets with good durability, which makes them well suited to flood control materials, and they have a price of approximately Renminbi (RMB) 35–60 each.

When pallets are used to store materials, attention should be paid to the reasonableness of the materials that are stored on the pallets. The materials should cover at least 80% of the pallet area, the height of the center of gravity of the stored materials should not exceed two-thirds of the pallet width, and the height of the materials above the pallet should not exceed 1200 mm. In this study, the various types of materials, their quantities, and their stacking arrangements are given in Table 5. The materials in this table were placed on shuttle racks, and generators and towed water pumps were stored on the floor. If more materials are added in the future, they can be placed on shuttle racks and placed at an upper level. Figures 2–6 present the schematic diagrams of the arrangements of the materials.


**Table 5.** Material types, quantities, and stacking arrangement.

\* HP: horsepower.

**Figure 2.** Motor oil stacking model.

**Figure 3.** Life jacket stacking model.

**Figure 4.** Flashlight stacking model.

**Figure 5.** Tent stacking model.

**Figure 6.** Outboard motor stacking model.

#### *4.5. Storage Location Allocation in the Intelligent Storage Area*

We first arranged the warehouse's internal layout (see Figure 7) in accordance with the existing intelligent warehouse area and material storage needs.

**Figure 7.** Plan layout.

Compartment A and compartment B are both intelligent shelf areas, and compartment C is for storing special materials, such as tent poles and very small amounts of materials. It is equipped with ordinary shelves and cantilevered shelves. Compartment D is the pending area, which can be used as the storage area for supplementary materials, such as generators, pumps, and other supplies. Compartment A is approximately 11 × 18 = 198 square meters, compartment B is approximately 12.1 × 7.5 = 84 square meters, compartment C is approximately 12 × 1 = 12 square meters, and compartment D is approximately 12 × 6.5 = 78 square meters. The blank areas are the lanes. The middle lane is 6 meters wide, and the other lane is approximately 3 meters wide. The upper part is the entrance, and the bottom is the exit. Temporary sorting areas can be established on both sides of the entrance and exit. If a one-time delivery is sufficient, the upper entrance can be used as a temporary exit to improve the distributional efficiency of flood control materials. In addition, if the warehouse is subject to realistic conditions, the entrances and exits can also be combined together, and the exit can be used as the entrance.

Based on our storage location allocation model and the flood control material warehouse's material use records, we assigned outboard motors, life jackets, tents, motor oil, and flashlights to one category, and generators and towed water pumps to another category. The materials that are suitable for storage in an intelligent warehouse were roughly divided into three areas: Area I contained outboard motors, life jackets, and tents; Area II contained motor oil and flashlights; and Area III contained generators, towed water pumps, and space for other materials that might be stored in the intelligent warehouse in the future. The materials in Area I were the most frequently used and had high inventory levels, the materials in Area II were frequently used but had low inventory levels, and the materials in Area III were moderately used, had low inventory levels, and were very heavy, bulky, and difficult to move.

Therefore, we arranged the storage area in accordance with the material types and quantities and the number of storage locations on the shelves. Type I materials were placed in areas A and B, with life jackets placed in Area A and outboard motors and tents placed in Area B. Type II materials consisted of tents, tent poles, and flashlights. Because of the close relationship between the tents and tent poles, they were placed in Area C. Type III materials consisting of generators and towed water pumps were placed sequentially on the floor in Area D.

After the materials had been placed in these sub-areas in accordance with the storage location allocation principles, the materials could be quickly and precisely located and retrieved from the warehouse, which increased the efficiency and facilitated their inspection, inventory, and maintenance.

Depending on their form, the shelves were classified into two main types. Type 1 consisted of pallet racks, which required many lanes and had a relatively low overall spatial utilization rate. Type 2 consisted of close-packed shelves, including drive-in racks, shuttle racks, and an intelligent 3-D storage area. These shelves required fewer lanes and used relatively little space.

The storage locations on the pallet racks are shown in Figure 8. A total of 90 storage locations were situated on each level in areas A and B, and five levels could be used for the storage of existing materials.

**Figure 8.** Pallet-type shelf storage locations.

The type 2 close-packed shelves, including drive-in racks, shuttle racks, and intelligent 3-D storage area shelves, could be arranged in the following manner. Objects could be placed horizontally on the shelves, and the main aisles could be used to ensure that the life jackets could be stored and removed via separate pathways. There were 176 storage locations on each level and four levels of shelves, which resulted in a total of 704 storage locations. See Figure 9 for a plan diagram of the storage locations.

**Figure 9.** Close-packed shelves.

Area C, which was used for the storage of tents and poles, had an overall volume of 5 m × 1.2 m × 4 mm and a floor area of approximately 12 m2. The tent poles were placed on the shelves in bundles of 10 poles. The storage locations were at a vertical distance of 0.5 m from each other, their total height was 3 m, and there were 12 storage locations. The size of the storage locations in the ordinary shelves located on the right was 1.4 m × 1.0 m × 1.5 m. Each level had five storage locations, and there were four levels. These shelves were used for the placement of small items, such as motor oil and flashlights that required convenient access. Area D contained 78 m2 of empty space that could be used for the storage of additional materials or to meet other needs in the future.

See Table 6 for the material parameters at the Zhenjiang warehouse.



Our storage location optimization model was programmed using the MATLAB software and yielded an *H*-function variation curve when run (see Figure 10).

The following conclusions can be derived from the curve in Figure 10. As the number of iterations increases, the objective function value *H* steadily decreases and reaches a relatively constant value after the number of iterations reaches 170. This indicates that the objective function has basically achieved convergence. The final value was 1.44 <sup>×</sup> 105.

**Figure 10.** Optimizing process.

The solution that was obtained above optimized the model using the genetic algorithm, and MATLAB yielded the storage location allocation results for the close-packed shelves in different areas at the Zhenjiang intelligent warehouse (see Table 7).


**Table 7.** Simplified storage location allocation results.

#### **5. Conclusions**

This paper introduces the intelligent storage of flood control materials in detail, including the storage principles and allocation strategy, and establishes a multi-objective optimization model and a simplified method for the allocation of flood control materials. Based on the optimization strategy obtained from MATLAB, this paper empirically analyses the distribution of storage positions for the provincial flood control materials in the Zhenjiang branch warehouse of Jiangsu Province and obtains some good results. It can provide referential value and recommendations for effectively improving the storage efficiency and management of flood control materials and reducing losses from floods.

Finally, we make the following related contributions.

(1) We scientifically assess the management of flood control materials. A flood control material classification management system is established. It is based on the attributes, the quantity, the occupied capital, and the frequency of the use of the flood control materials, and manages each material following the principles of scientific classification and a rational layout. It can be used to continuously improve the reserves and management of flood control materials and emergency material supplies.


Of course, different types of intelligent warehousing technology and equipment have different characteristics. When applied to other intelligent warehousing technology and equipment, the research in this paper must adjust the model accordingly. In future research, we will further enhance the generality of the model to solve these problems. In addition, we also hope that more emergency logistics, such as drought-proof materials and wind-proof materials, will be included to optimize the location allocation of intelligent storage of integrated emergency materials.

**Author Contributions:** All authors contributed equally to this work. In particular, W.W. developed the original idea for the study and designed the methodology. W.W. and L.H. revised the manuscript. J.Y. drafted the manuscript. J.Y. and L.H. performed the case study and collected the data. J.W. performed the investigation. D.P. helped analyze the results and further reviewed the literature. All authors have read and approved the final manuscript.

**Funding:** This work was funded by the Humanities and Social Sciences of Ministry of Education Planning Fund (No. 18YJAZH092); the Jiangsu Water Conservancy Science and Technology Project (Nos. 2017059 and 2018071); Key Laboratory of Coastal Disaster and Defence of Ministry of Education, Hohai University (Nos. 201913); the Fundamental Research Funds for the Central Universities (Nos. 2019B20414); and Sichuan University (No. SKSYL201819).

**Acknowledgments:** The authors would like to thank all the HFCMRC staff and especially Jianbin Wei, Yi Han and Qingli Hua for their help with the data collection.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

## *Article* **Guidelines for the Use of Unmanned Aerial Systems in Flood Emergency Response**

**Gloria Salmoral <sup>1</sup> , Mónica Rivas Casado 1,\*, Manoranjan Muthusamy <sup>1</sup> , David Butler <sup>2</sup> , Prathyush P. Menon <sup>3</sup> and Paul Leinster <sup>1</sup>**


Received: 15 December 2019; Accepted: 8 February 2020; Published: 13 February 2020

**Abstract:** There is increasing interest in using Unmanned Aircraft Systems (UAS) in flood risk management activities including in response to flood events. However, there is little evidence that they are used in a structured and strategic manner to best effect. An effective response to flooding is essential if lives are to be saved and suffering alleviated. This study evaluates how UAS can be used in the preparation for and response to flood emergencies and develops guidelines for their deployment before, during and after a flood event. A comprehensive literature review and interviews, with people with practical experience of flood risk management, compared the current organizational and operational structures for flood emergency response in both England and India, and developed a deployment analysis matrix of existing UAS applications. An online survey was carried out in England to assess how the technology could be further developed to meet flood emergency response needs. The deployment analysis matrix has the potential to be translated into an Indian context and other countries. Those organizations responsible for overseeing flood risk management activities including the response to flooding events will have to keep abreast of the rapid technological advances in UAS if they are to be used to best effect.

**Keywords:** drone applications; deployment time; monitoring; flood modelling; evacuation; rescue; resilience

#### **1. Introduction**

In recent decades, significant flood events have affected many countries around the world including those caused by Hurricane Katrina (Florida and Louisiana, August 2005), Hurricane Leslie (France, October 2018) and the 2018 monsoon season in India (Kerala, August 2018). The impacts of these flood events on people and communities are wide and varied and can be catastrophic. The Katrina floods resulted in over 1100 deaths in Louisiana [1], with estimated economic losses of \$149 billion [2]; the 2018 floods in France resulted in 14 deaths and over \$500 million of damage [3]; whilst the 2018 Kerala in India floods resulted in more than 400 deaths, displaced 1.8 million people and caused an estimated \$3 billion worth of damage [4]. The last major flood event in England took place in winter 2015/2016, with December 2015 being the wettest month ever recorded in England. A total of 17,000 properties across the north of England were affected with named storms Desmond, Eva and Frank. The total economic damages were estimated to be £1.6 billion [5].

Effective and efficient flood emergency response has a key role in reducing the adverse impacts of flooding. Coordinating the response—including warning and informing prior to and during events, evacuation prior to the flooding, the rescue of people and the organization of volunteers [6]—has been a priority for governments in recent decades. Although there are clearly lessons to be learned from experiences in other countries, often the detailed arrangements need to be context specific.

In England, revised flood emergency operational arrangements were put in place building on the experience to date. These are outlined in the National Flood Emergency Framework [7]. The primary aims of the emergency response include the protection of human life, the alleviation of suffering and the restoration of disrupted services (e.g., water, electricity, transport). Within this framework and based on documented command and control protocols, decisions are taken at the lowest appropriate level with coordination at the highest necessary level.

In India, the Central Government has established the National Disaster Management Division within the Ministry of Home Affairs. This Division has introduced various initiatives and set up several organizations to deal with disasters, including floods. In 2016, a National Disaster Management Plan was published to provide the overall direction and national goals. Under the plan, the various ministries and departments at state and district levels have to develop their own specific management and response plans, and related operating procedures [8].

Various emergency response activities rely on the information provided by monitoring, models and multiple data sources. For example, in India hydrological and flood models are used by the Central Water Commission for modelling and forecasting purposes [8] to provide water level and river flow information to the authorities [9] with an online dissemination portal [10]. In England, there has been a continuing interest in developing flood models for fluvial, pluvial and sea flooding [11], with different data needs and outcomes [11,12], to help reduce the impacts on people, property and critical infrastructure [13]. From the response side, emergency responders request and collate a varied range of information, from aerial imagery to individual eyewitness reports, to support decision-making. Different information is required pre-, during- and post-events. For instance, during a flood event real-time or near real-time local information on how many people, buildings, and other infrastructure are at risk is required [14]; post-event aerial imagery can provide vital and detailed information about the extent of flooding and damage to properties [15].

In recent years, emergency responders have used Unmanned Aircraft Systems (UAS) to acquire core information pre-, during- and post-events [16–18]. UAS are small and light (less than 20 kg) remotely piloted aircraft generally equipped with a range of sensors for the collection of information. There are two main types of UAS platforms: fixed wing and vertical take-off and landing (VTOL). The former relies on wings that generate lift to fly, whereas VTLO rely on rotors. UAS can be equipped with different sensors [19] from cameras to warning systems. RGB cameras are able to provide high resolution imagery of up to 2 cm. Emergency responders in various countries have identified the added benefits of UAS in humanitarian responses in terms of the rapid assessment of damage, such as collapsed buildings or blocked roads and search and rescue operations [20,21]. In England for example, during winter 2015/2016, UAS were used by the Environment Agency (EA) to assess smaller scale flooding incidents in high detail; in particular, UAS were used to provide an up close and detailed live stream of an inaccessible river breach. This enabled an effective and efficient assessment of the area [22]. However, the rapid uptake and continuous development of the technology have resulted in the ad-hoc and opportunistic use of UAS over a strategic appraisal of how best to use them and for what purpose pre-, during- and post-events. Various types of UAS missions have been identified as being used in flood emergency responses (including strategic situation awareness, inspections, ground search, water search, debris/flood/damage estimation and tactical situation awareness) with an indication of the data products (e.g., images, videos, and orthomosaics) generated in the flights [18]). However, there is not yet a purpose-driven approach defining which UAS products would be of benefit at each stage of the preparation for and response to a flood event. The need of such logic-based decision support approach has been identified by multiple research and governmental organizations within the

context of catastrophe response in India and England through two knowledge exchange workshops organized in Delhi (30 September 2018) and Bangalore (18 September 2019) within the engagement activities organized by the EPSRC research project 'Emergency flood planning and management using unmanned aerial systems' (www.efloodplan.net). To the authors' knowledge, a purpose-driven approach detailing how and when UAS with specific embedded sensors should be used to collect data to assist in flood event responses is not yet documented. It is envisaged that such an approach will be context-specific and influenced by the nature of the flood events that occur within a particular area, region or country, the data available from other sources, as well as the airspace regulatory framework for UAS use.

Based on these premises, the aim of this study is to develop purpose led guidelines for the efficient and effective deployment of UAS for flood risk management activities including emergency response pre-, during- and post-event phases. We demonstrate how a deployment analysis matrix can be designed and used to assist flood emergency response requirements in the context of catchment response and the nature of flood events for England and explore its potential to be translated into an Indian context. This will be achieved through the following four overarching objectives: (1) to map out the current role of existing organizations involved in emergency response in England and India; (2) to identify existing UAS applications within the components of a flood risk management system; (3) to determine context-specific requirements for UAS products to assist in flood risk management activities; and, (4) to develop an adaptive and transferable matrix analysis framework that can then be used as the basis for guidelines for the effective deployment of UAS for flood risk management activities leading to more resilient urban environments and including emergency response before, during and after a flood event.

#### **2. Materials and Methods**

#### *2.1. Flood Emergency Response in England and India and the Potential Use of UAS*

The institutional arrangements for flood emergency response and the current and potential applications of UAS technology were determined through a literature review and face-to-face interviews with key personnel with detailed understanding of the flood response arrangements in England and in India. England and India were selected for this study based on recent flood events occurring in these areas (Cockermouth, England, 2015 and Kerala, India, 2018). A total of 14 interviews were held in India (7) and England (7), including participants with experience in flood emergency response from the national and regional authorities, private sector and Non-governmental organizations (NGOs). A semi-structured questionnaire based on a set of twenty open questions was used (Supplementary Materials). This format enabled the interviews to be focused on the research objectives, but with the flexibility to evaluate responses and explore issues that emerged during interviews. The raw data products (i.e., those are provided without extra processing or internet connection) and derived products (i.e., produced by post-processing of the raw data products) produced by UAS applications, as well as the key factors affecting UAS deployment and flight plan configuration were also identified in the literature review. Additional information on the main applications and potential use of UAS was also obtained from a knowledge exchange workshop organized in India within the engagement process of the EPSRC research project "Emergency flood planning and management using unmanned aerial systems" (Delhi, 30 May 2018). The UAS applications were grouped into a set of five flood management components, which included flood warning, flood monitoring and flood risk assessment, evacuation route identification, damage assessment and rescue.

#### *2.2. Development of an UAS Deployment Analysis Matrix*

This study uses a 3 × 3 matrix to identify potential uses of UAS in flood risk management activities and as the basis of guidelines on the deployment of UAS for flood risk management activities. A number of factors were considered for the x and y axes of the matrix when considering UAS

deployment and flight plan configurations. The three main factors identified were related to catchment size, flood source type, and phase of a flood event (Table 1). Catchment size influences the amount of data gathered [23] and the type of UAS that is required to provide the spatial coverage [24].

Catchment flood response was chosen as one of the key factors because this gives an indication of the time available to deploy an UAS and the use of particular applications and technologies in a given situation. The catchment flood response was determined based on the time between the start of a rainfall event and the potential for the flooding of properties. Based on climatic and catchment conditions in England, the flood response was considered to be 'slow' when flooding occurs more than 8 h after the rain event, 'medium' when flooding occurs between 3 and 8 h and 'fast' when the onset of flooding takes place in less than3h[25,26]. We also considered in the deployment analysis matrix the phase in which UAS will be deployed in the overall approach to flood risk management activities [27]: 'pre-event', 'during-event' and 'post-event' (Figure 1). Pre-event refers to activities such as flood modelling activities, the construction of flood risk reduction assets and the planning that will be needed to respond effectively to a defined flood magnitude (i.e., based on a return period). During-event starts as soon as the first flood warning is issued, whilst post-event refers to the recovery and clean-up phase when the water has receded and for example is no longer within people's houses or blocking transport routes.


**Table 1.** Key factors identified as relevant for the development of the Unmanned Aircraft Systems (UAS) deployment analysis matrix.

**Figure 1.** Format of the UAS deployment analysis matrix.

Each of the nine cells within the matrix were populated with the UAS applications identified from the literature review and via a second set of ten one-to-one interviews in England. The interviews targeted specialists in the use of UAS for monitoring, surveying and incident response within the EA, Cranfield University, University of Exeter and an independent expert in flood risk management who had extensive experience of emergency responses at a senior level. During each interview, responders were presented with a set of cards defining the UAS applications, the processing time required to obtain the UAS products and the accuracy or resolution of such products. Processing time, accuracy and resolution were defined based on values reported in the literature. Responders were able to allocate the UAS applications with a given processing time and accuracy within the context of a recent flood event in which they were involved. The data gathered in the matrix was analyzed to identify the consistency between participants in allocating an application to a particular matrix cell. Consistency between participants was assessed through direct comparison of the number of responses per application and cell.

A workshop organized in India (Bangalore, 18 September 2019) helped provide insights about the transferability of the designed deployment analysis to an Indian context. Discussions with experts on flood risk management activities including emergency response in India were held to determine the potential transferability of the matrix to other countries.

#### *2.3. Technological Needs for the Use of UAS in Flood Emergency Response*

To assess how the technology should be further developed to meet flood emergency response needs, an extended online survey was also carried out (See Supplementary Materials for details of the survey). The UAS applications—flood extent, flood depth and flow velocity—were selected as they are considered to be key for making decisions during a flood event [33,34]. For the three UAS applications, participants were asked about the current and desired time to process these specific geomatics products and the associated accuracy requirements. Accuracy refers to the expected error range in flood extent (m), flood depth (cm) and flow velocity (m s−1) in the generated geomatics product. The survey was built in Qualtrics software and distributed to relevant experts and at two flood risk management related events: Oasis Conference (London, 18 June 2019) and Flood and Coast (Telford, UK, 20 June 2019). A total of 25 participants completed the survey. Data collected were compared using descriptive statistics to assess the current and desired accuracies and time to process each application. The information gathered enabled an assessment to be made as to whether there are any knowledge and technology gaps that need to be addressed to achieve a desired time and accuracy for a given application.

#### **3. Results**

#### *3.1. Organizations Involved in Flood Emergency Response: England and India*

Results from the literature review and one-to-one interviews highlighted that in England there are over 17 organizations involved in flood emergency response. This is similar to the number in India, where 16 key organizations were identified (see Supplementary Materials).

In England, the response to localized flooding is led by the local emergency responders without any significant involvement from central government [35]. For some flood events, local responders are supported by central government via Department for Environment, Food and Rural Affairs (Defra) as the designated Lead Government Department for responding to floods. Defra normally co-ordinates the cross-government response to lower level national flooding events (level 1) and manage it within the department. As the extent and impact of the flooding increases, it is likely that there will be increasing involvement by others in central government with the activation of the Cabinet Office Briefing Room (COBR), which brings together ministers, seniors government officials, representatives from national response agencies and organizations impacted by the flood event. Level 2 events (serious impact) are still coordinated by Defra but through COBR. More serious events (level 3—catastrophic) are fully escalated to central co-ordination by the Civil Contingencies Secretariat within COBR [7] (Figure 2). The Army and other military forces may be requested to help in a flood response [36].

**Figure 2.** Schematic diagram of flood emergency response in England showing the main agencies and groups involved, the levels of emergency responses at the local level (operational, tactical, strategic), the categories of responding organizations (Category 1 and Category 2) and the likely government arrangements (from local response to central direction from COBR) which depends upon flood extent (local, regional, cross-region and national). Category 1 comprises the organizations that are at the core of the response to most emergencies, whereas Category 2 responders are co-operating bodies involved in incidents that affect their sector. The color schemes in the government arrangements reflect the increasing levels of emergency response (from green to red). COBR: Cabinet Office Briefing Rooms, Defra: Department for Environment, Food and Rural Affairs, MHCLG: Ministry of Housing, Communities and Local Government, RED: Resilience and Emergency Division, and LGD: Lead Government Department.

At present, there are a number of organizations that may use UAS during a flood event. These include the EA, the Fire and Rescue Service, the Police and insurance companies as well as private individuals. As an example, in England, the local or national incident responders may request the deployment of UAS to the specialist Geomatics Team of the EA to provide information related to flood damage and impacts [37]. The Geomatics Team will evaluate if UAS are the most appropriate means of obtaining the information. One of the interviewees informed us that arrangements are in place that allow the EA's Geomatics Team to deploy UAS in any part of England within six hours. The UAS images can be sent via a live feed to an EA incident room. The decision as to who will fly UAS in a particular situation is agreed locally event by event. To date, there is not an established approach to decide which organization will fly UAS for what purpose during and after an event. This can result in a duplication of effort or in important information not being gathered during a particular event.

India also has a tiered approach to flood emergency response (Figure 3). The national government develops policies and provides advice and assistance when there are major events, whilst the States are the responsible for carrying out risk assessments and planning and implementing mitigation measures [8]. At the district level, flood events are categorized into three levels of impact [38]: Level 1—there are sufficient resources and capacity to respond at the district level; Level 2—the impact is beyond existing capacities and support from State agencies is needed; and Level 3—the impact is beyond the existing capacities of district and state resources and support from national agencies is needed (Figure 3). If an emergency escalates beyond their capabilities, the local administration must seek assistance from the district administration or the State Government. If the State Government considers it necessary, it can seek central assistance [8]. The Ministry of Water Resources (river flooding) and Ministry of Urban Development (pluvial flooding) function under the overall guidance of the Ministry of Home Affairs [8,39] when responding to flood events.

In India the police, navy and army have permission to fly UAS for security and rescue reason. However, as in England, there is not an established system to deploy UAS in flood emergency response activities. For example, in the 2013 Uttarakhand floods, the National Disaster Response Force (NDRF) deployed UAS with technical support from research institutions [40,41]. The Indian security forces and the Indo-Tibetan Border Police also deployed UAS to assist in the relief efforts of the National Disaster Response Force by helping find survivors in remote locations [42] and in areas cut off by landslides [43]. During the interviews performed in Delhi, Indian participants highlighted the opportunities to use UAS at the district level by the District Collector, who is responsible for district-level responses to a flooding event.

**Figure 3.** Schematic diagram of flood emergency response in India showing the main agencies and groups involved and the levels of emergency responses (level 1, level 2 and level 3) which depends upon the flood extent (district, state, cross-states, national).

#### *3.2. The Potential Use of UAS in Flood Emergency Response*

From the existing scientific literature, sixteen UAS applications that could be used in flood risk management activities were identified. The UAS applications can be assessed in terms of their use before an event in flood risk assessments, determining terrain elevations, flood extent modelling, identifying evacuation routes and flood warning. During an event to inform responders about actual flood extents, flood sources and routes, whether evacuation routes remain clear, identifying people in need of rescue and provision of emergency relief supply. Post events as part of damage and impact assessments. UAS raw products included high definition (HD) video, infrared imaging, Red-Green-Blue (RGB) imagery, RGB video, RGB video streaming and thermal imaging. A total of fourteen post-processing outcomes were identified: those derived from models (e.g., flood models, evacuation models), bespoke algorithms (e.g., image feature recognition), UAS-specific software (e.g., terrain elevation measurements) (Table 2).


 **2.** Detailed list of flood risk management components, with identified applications, UAS raw products and post-processed outcomes from the literature

**Table**

 within


study shows values of resolution and/or accuracy, it is indicated in the table the details of the lowest values for resolution and/or accuracy found. Assumed the value of 0.2 m based the real-time optical flood extent detection. 5 It is assumed that above 3 m/s, there are damages in the infrastructure [68], which is multiplied by the median of two coe fficients of variation in flow velocity [52] to estimate the medium accuracy in m/s.

1

#### *3.3. The UAS Deployment Analysis Matrix for England: Pre-, During- and Post-Event*

Results (Figure 4) showed that pre-event for all catchment responses, the UAS applications were primarily concerned with digital elevation models for use in flood models, the condition of flood risk management assets, identification of safe shelter points and evacuation routes and providing warnings.

During the event applications providing information in real-time were prioritized. A combination of rapid visualization with high resolution of flood extent and flood depth were chosen. This can be provided with the current UAS technical capabilities. In fast response catchments, the participants' preference was for flow velocity with medium accuracy (instead of high accuracy). Additional time and costs are needed to achieve higher accuracies. Real-time applications relating to rescue activities (i.e., identification of safe shelter points, detection of stranded people and delivery of ad-hoc supplies) and damage assessment (i.e., visual detection of affected properties) were also identified as priorities. Participants stated that applications requiring more than 4 h of processing time to generate products are unlikely to be of use to responders in many flood events.

With the limitations of current UAS applications, the updating of evacuation routes was identified as being important only in catchments with a flood response longer than 12 h or where the duration of flooding in a faster responding catchment persists for more than 12 h. Applications that need more than 48 h of processing time, such as modelling flood extent and the identification of resilience and resistance measures, were still identified as being important as the data collected can be used subsequently to improve flood models and the response for future flood events.

During an event time is the priority, whereas after the event accuracy was most relevant. The focus in post event data collection was on the provision of more precise information for flood extent, flood depth and flood source so that flood impact can be estimated more accurately. After an event, there is also a continuing need for information that will assist with the rescue and recovery activities and the estimation of property level flood impacts.

#### *3.4. Preferences in England for UAS Applications in Flood Emergency Response*

The online survey evaluated the existing and desired processing time and accuracy in UAS applications for flood emergency response in England, as a way to determine whether technological development is needed to better inform emergency response. Results from the online survey indicated that 44% of participants currently have access to flood extent data within 12 h, with accuracies from 2 cm to 50 m. Only 3 of the 25 participants indicated they have access to flood extent data in less than 1 h with accuracies between 1 and 50 m. The preference of 52% of the participants was to have access to flood extent data within 0.5 h (i.e., near real time) with an accuracy of <10 m. There was also another significant group of participants (28%), who would seek to have access to flood extent data within 12 to 24 h with an accuracy of <10 m. Similarly, for flood depth and flow velocity respondents considered that having data more quickly with improved accuracy compared with the current products would be of benefit, with a desire for data to be available in <0.5 h with accuracies of 1 to 5 cm in flood depth and 0.1–0.5 m s−<sup>1</sup> in flow velocity (Figure 5). There are current UAS technologies that are able to meet these requirements (Table 2).


**Figure 4.** Respondents' preferences in England for each UAS application in relation to catchment response and UAS deployment at different emergency response phases (before, during, post). The numbers indicate the number of participants. There were 10 participants, and therefore the maximum score possible is 10. Only UASapplicationswithascore≥4areshownasameansofindicatingthemajorityopinion.

**Figure 5.** (**a**) Actual and preferred accuracy values and time needed to process flood extent, flood depth and flow velocity. The marker indicates the average of flood extent, flood depth and flow velocity, whereas the vertical lines shows the minimum and maximum values. The number of participants with a preference for a given combination are indicated against each measure. (**b**) Type of organizations that completed the survey. (**c**) Participants' experience level in flood emergency response, remote sensing and UAS. Results obtained from 25 participants.

For flood extent 13 participants considered time more important than accuracy when generating a product that will assist flood emergency response, whereas 12 participants thought accuracy was more important than time. Some participants stated that the most important factor for them was the trustworthiness of the data source. For flood depth (16) and flow velocity (17) participants were more interested in improved accuracy than in the time taken to obtain the data.

#### **4. Discussion**

#### *4.1. A Purpose-Driven Approach to UAS Deployment in the Context of Flood Emergency Response*

The operational use of UAS in flood emergency response is still limited [69]. A more systematic analysis of their application and capabilities in relation to their use in flood risk management including as part of an effective response to an event is, therefore, required if a purpose-driven approach to their deployment is to be realized.

During Storm Desmond in Cockermouth (Cumbria, England) in December 2015, more than 300 mm of rain fell over a 24 h period with an estimated <1% annual exceedance probability for both rainfall and river flows [70]. The Ministry for Housing, Communities and Local Government—Resilience and Emergencies Division (DCLG-RED), the Department for Environment, Food and Rural Affairs (Defra) and the Cabinet Office Briefing Room (COBR) were involved in coordinating the emergency response and supporting the local Cumbrian Strategic coordinating Group (SCG) [71]. UAS were used in Cockermouth during and after the storm to estimate the flood extent and identify impacted properties [72]. However, the use of UAS in Cockermouth could also have facilitated the identification of different types of flood sources (e.g., pluvial versus fluvial), as highlighted by [15]. In 2015 the range of applications UAS could be used for and how best to deploy them during flooding events had not been studied in a systematic way and, therefore, they were used in a reactive rather than strategically planned way.

Kerala is one of the Indian States that experiences the highest monsoon rainfall every year [73] and was affected by flooding in August 2018. The rain caused thousands of landslides in mountainous regions. Nearly 500 people died in the event [74]. Parts of the city of Cochin—the commercial capital of Kerala—were flooded, with a 90% increase in water cover and a water level rise of up to 5 m to 10 m [75]. As a result, major infrastructure assets including the airport, roads and railways had to be closed for safety reasons. The government issued evacuation orders and deployed the National Disaster Response Force teams within the area. During the emergency response over 223,000 people were evacuated to emergency relief camps [76]. UAS were used in Kerala to support rescue operations [77] and deliver aid [78]. There were also examples of people using UAS independently of the official response.

Although in both examples UAS were used in the response, the full capabilities were not necessarily exploited and the deployment of UAS was largely uncoordinated within the emergency response. The deployment analysis matrix developed here will enable those involved in flood risk management, including incident response, to take a structured approach to determining how best to use and deploy UAS within their specific context. The matrix-based approach will enable guidelines to be produced for the purpose-driven deployment of UAS within flood risk management activities including emergency response, as we discuss in the next section. This will help reduce duplication of effort and ensure the timely capture of important information that can be used to inform the current and future responses.

#### *4.2. Guidelines for the Deployment of UAS within Flood Risk Management Activities Including Emergency Response*

There are many benefits that can be derived from the use of UAS, to help reduce flood risk and the impacts on people, properties and the economy, if they are deployed in a structured and considered way that are currently not being fully utilized or exploited. The use of UAS has to be considered within the strategic planning for flood risk management activities including the response to flood events. This can build on experiences from the development of integrated flood forecasting, warning and response systems [79,80] and the use of real-time modelling to assist flood emergency response [81]. Our deployment matrix approach can be used as the basis for developing guidelines for the use of UAS within flood risk management before, during and post events. These guidelines are summarized in the following paragraphs.

Before a flood event:


During a flood event:


After an event:


#### *4.3. Selecting the Correct UAS Platform*

Multiple UAS platforms are available for use. The selection of the most appropriate platform for a particular application is a complex task. Factors that need to be considered include the capability of the gimbal to integrate the payload, weather conditions, the extent of the area to be flown, and the availability of pilots with the rights skills and regulatory permissions. UAS can be classified into vertical take-off-and-landing (VTOL), fixed wing and hybrids [84]. In Rivas et al. [85], the authors highlight that VTOL UAS are able to hover over a point and provide high resolution still imagery whereas fixed wing platforms enable wide area surveying [85,86].

The flooding of large areas, which will most likely occur in catchment areas with a slow response to floods, will require the use of fixed wing rather than VTOL platforms, as the former have longer endurance, although they are more difficult to fly and require specific training [87]. However, some fixed wing platforms do not have the capability of slowing down to speeds that enable them to collect high resolution imagery. Rivas Casado et al. [85] report 8 h to map the river channel of a 1.4 km reach, when using a rotor platform (Falcon 8 octocopter, ASCTEC, Krailling, Germany) whereas the same author reported a coverage of 142 ha within four hours in two flights undertaken with a Sirius Pro fixed wing [15]. Fixed wing platforms such as the Sirius Pro enable flights of up to 50 min at a cruising speed of 18 m s−<sup>1</sup> [88]. In certain locations, VTOL UAS will also be required to overcome the limitations of terrain in terms of take-off and landing [19]. VTLOs can hover on site with high location accuracy and, therefore, take more detailed photographs at locations of interest. Hybrid models are able to combine the advantages of both fixed wing and VTOL platforms. The WingtraOne PPK VTO is a hybrid rotor and fixed wing platform able to provide Ground Sampling Distances of 0.7 cm/pixel and map 400 ha in a single flight [89]. The battery endurance is 55 min. The platform is able to fly under wind conditions of up to 45 km h−<sup>1</sup> in cruise and up to 30 km h−<sup>1</sup> for landing.

Battery endurance can also compromise performance. Overall, fixed wing platforms provide better battery endurance than VTOL platforms. Figure 6 and Table 3 show an alternative classification for UAS based on battery endurance and work range. Low-cost close range UAS include platforms with a range of generally up to 5 km and an endurance time of less than 45 min. Examples of such platforms include standard small VTOL platform such as the DJI Phantom 4 Pro (DJI Technology Co, Shenzhen, China) which can fly continuously for over 30 min [90] at a maximum speed of 20 m s<sup>−</sup>1. Such platforms will be suitable to cover small catchments or specific areas of medium to large catchments. More expensive close-range platforms offer a working range of up to 50 km with battery endurance of up to 6 h. The mdMapper4-3000DμoG VHR VTLO (microdrones) is an example of such a platform able to capture RGB imagery at a ground sampling distance (GSD) of up to 0.6 cm/pixel [91]. The platform has an endurance of approximately 40 min when flying at an altitude of 120 m. The UAS can cover between 64 ha (80 mm lens, GSD = 0.6 cm/pixel) and 150 ha (35 mm lens, GSD = 1.3 cm/pixel) at a constant speed of5ms−<sup>1</sup> within a single flight. A fixed wing example of a close-range platform is the Sirius Pro (Topcon Positioning System Inc., Livermore, CA, USA) which has a 50 min flight endurance and is able to operate under windy conditions (50 km h<sup>−</sup>1) with gusts of up to 65 km h−<sup>1</sup> [92]. Another example is the eBeeX which is able to gather RGB imagery at 1 cm/pixel and cover 220 ha in a single flight when flying at an altitude of 120 m. Short-range, medium-range and high-endurance platforms all require runways for their deployment and, although they provide a larger working range and endurance, are difficult to deploy in urban settings, especially during flood events.

In large flooded areas, there is likely to be a need to coordinate the deployment of multiple UAS within an affected area. The surveying of large areas will result in larger data sets and this will have a consequential impact on the time taken to generate products. More stable and perhaps heavier platforms, such as the microdrone md4-1000 [93], are needed for more extreme wind and rainfall conditions. Additional advances will enable the miniaturization of sensors, enhance the level of autonomy, increase battery life and the capability of flying in more extreme weather conditions.

**Figure 6.** Simplified classification of UAS platforms based on their work range (km) and battery endurance time (h). The different classes include low-cost close-range (CR), close-range (CR), short-range (SR), medium-range (MR) and high-endurance (Endurance) platforms. A full description of these classes is provided in Table 3.


**Table 3.** Description of the simplified classification of UAS.

#### *4.4. The Deployment Decision Approach to an Indian Context*

Our deployment analysis matrix approach for the use of UAS in flood risk management activities in England has the potential to be transferred to other countries (e.g., India) with different climatic, topographic and socioeconomic contexts. In India, environmental conditions can be extreme in terms of the intensity and extent of the rainfall. Transferability of the matrix will need to take into account the catchment response. In India, some catchments are of a larger scale than those in England and can be flooded for weeks with a need for recurrent monitoring of large areas. One of the challenges faced is the need to evacuate large numbers of people over extended areas in a short period [82]. The size of the areas affected will require the deployment of certain types of UAS (Section 4.3) for particular applications. In India, rural areas can also present access, travel time and maintenance challenges [82,94] with landslides limiting access to remote areas [95,96].

#### **5. Conclusions**

UAS are currently used in a largely ad hoc manner in flood risk management activities with practice differing significantly even within countries. Even so, their use has proved to be beneficial. However, if they were to be used in a purpose-driven and strategically coordinated way, they can provide more coherent and targeted information that will have added value for flood risk management activities, including during the response to events. The data and information produced by UAS can be used to improve flood risk management activities, structures, tools and approaches helping to reduce flooding and its associated impacts on people, properties, infrastructure and the economy. We have identified a range of products that can be delivered by UAS and have developed an analysis matrix approach to help target their deployment. The UAS deployment matrix forms the basis for developing guidelines to assist those involved in flood risk management activities, including emergency responders, in developing a more strategic and targeted approach to the use of UAS before, during and after flood events. The approaches developed will need to be context specific including who will use what type of UAS and for what purpose before, during and after an event. The deployment matrix we have developed for England has the potential to be translated into an Indian context, and in other countries.

Further research should focus on exploring future technological developments of UAS platforms and sensors, their potential applications within a flood emergency response context and how these will feed into the existing deployment guidelines. Technological developments would be particularly helpful in the miniaturization of sensors, their integration on more stable UAS platforms and increased flight (i.e., battery) endurance. The fast pace of technological advances within the field of UAS requires a flexible and adaptive approach, which facilitates operational uptake as soon as advances are commercially available. The various organizations involved in the use of UAS in flood risk management will have to keep the deployment guidelines under review if they are to make the best use of the available and developing technologies to achieve flood management and resilience targets.

**Supplementary Materials:** The following are available online at http://www.mdpi.com/2073-4441/12/2/521/s1, Table S1: Organisations involved in the flood emergency response in England and India and their associated role.

**Author Contributions:** Conceptualization, G.S., P.L., M.R.C., M.M., D.B. and P.P.M.; methodology, G.S., M.R.C., P.L. and M.M.; validation, G.S.; formal analysis, G.S. and M.M.; investigation, G.S., M.R.C., P.L. and M.M.; resources, M.R.C., P.L., D.B. and P.P.M.; data curation, G.S. and M.M.; writing—original draft preparation, G.S., M.R.C. and P.L.; writing—review and editing, G.S., P.L., M.R.C., M.M., D.B. and P.P.M.; supervision, M.R.C. and P.L.; project administration, M.R.C.; funding acquisition, D.B., P.L., P.P.M. and M.R.C. All authors have read and agree to the published version of the manuscript.

**Funding:** This research was funded by EPSRC, EP/P02839X/1 "Emergency flood planning and management using unmanned aerial systems" and EPSRC EP/N010329/1 "BRIM: Building Resilience Into risk Management". Due to the ethically sensitivity nature of the research (interviews), supporting data cannot be made available.

**Acknowledgments:** The authors acknowledge the support from the following institutions and experts from England and India: Environment Agency, Central Water Commission, National Institute Disaster Management, International Federation of Red Cross, Goonj, Seeds, R.K. Dave, and Ruchi Saxena.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

## *Article* **Analysis of the Public Flood Risk Perception in a Flood-Prone City: The Case of Jingdezhen City in China**

#### **Zhiqiang Wang 1,2,3,\*, Huimin Wang 1,3,\*, Jing Huang 1,3, Jinle Kang 1,3 and Dawei Han <sup>2</sup>**


Received: 13 October 2018; Accepted: 1 November 2018; Published: 4 November 2018

**Abstract:** Understanding and improving public flood risk perception is conducive to the implementation of effective flood risk management and disaster reduction policies. In the flood-prone city of Jingdezhen, flood disaster is one of the most destructive natural hazards to impact the society and economy. However, few studies have been attempted to focus on public flood risk perception in the small and medium-size city in China, like Jingdezhen. Therefore, the purpose of this study was to investigate the public flood risk perception in four districts of Jingdezhen and examine the related influencing factors. A questionnaire survey of 719 randomly sampled respondents was conducted in 16 subdistricts of Jingdezhen. Analysis of variance was conducted to identify the correlations between the impact factors and public flood risk perception. Then, the flood risk perception differences between different groups under the same impact factor were compared. The results indicated that the socio-demographic characteristics of the respondents (except occupation), flood experience, flood knowledge education, flood protection responsibility, and trust in government were strongly correlated with flood risk perception. The findings will help decision makers to develop effective flood risk communication strategies and flood risk reduction policies.

**Keywords:** flood risk perception; natural flood management; disaster mitigation; flood-prone city; questionnaire survey

#### **1. Introduction**

Natural disasters are a major threat to the social and economic structure and they can easily wipe out the wealth accumulated in the past. In the future, flood risk is projected to increase in many regions due to effects of climate change and an increased concentration of people and economic properties [1,2]. Besides, the natural disaster frequency appears to be increasing in recent years [3], and the threat to development and economic losses from flood disasters are increasing too. Although many efforts have been done to reduce the risk and damage from natural disasters, floods remain the most devastation natural hazard in the world (World Bank, 2012). In 2017, Emergency Events Database (EM-DAT) data showed 318 natural disasters in the world, affecting 122 countries. These disasters resulted in 9503 deaths, 96 million people affected, and \$314 billion in economic losses. Among them, nearly 60% of the population affected by the disaster in 2017 were affected by the flood. Similar to previous years, China was the most disaster-affected country, with 25 events (c): 15 floods/landslides and 6 storms. During 2006–2015, flood disasters in China killed 6641 people, affected about one half billion people, and caused more than US\$87.5 billion damage (https://www.cred.be).

In order to study and reduce the negative impact of flood disasters on society and economy, researchers began to pay attention to flood risk assessment and flood risk management. Many researchers studied the objective flood risk, such as flood occurrence probability, flood inundation, and economic loss based on risk perspective. Other researchers believe that the subjective factors of the individuals can influence the judgment of the objective flood disaster risk. One of the important factors is the individuals' flood risk perception, and it has become an important topic to policy makers that are concerned with flood risk management [4]. Generally, risk perception refers to people's beliefs, attitudes, judgements and feelings towards events, and researchers believe that flood risk perception is the direct cause of flood risk prevention awareness and response behaviors [5–7]. Studying people's risk perception level is conducive to the implementation of effective flood risk management and disaster reduction policies, which has very important practical significance [8,9]. This is because:


Actually, research on risk perception began in the 1940s, when Gilbert White published the human adjustments to floods in the United States [10]. White found that people's behaviors could be directly affected by their previous flood experience, which created a precedent for study on human dimensions of risk in a multi-hazard environment [11,12]. The key early paper about public risk perception was written in 1960s by Chauncey Starr. Starr explored the correlation between the social acceptance of technological risks and the perception of social benefits and justified social costs from these technologies [13,14]. This method of revealed preference influenced the subsequent research. In 1978, Fischhoff and Slovic first proposed the use of psychometrics to assess risk. They used a scaled questionnaire of expressed preferences to directly capture people's different perceptions of risk and benefits, responding to limitation of Starr's revealed preference [15]. In the flood disasters area, relevant studies are based on their research, using psychological experiments and social surveys to assess people's flood risk perceptions. These studies on flood risk perception are mainly:


affective (feelings, perceived control, etc.) aspects. Most studies use a different set of items to measure the flood risk perception. However, Gotham [41] and Horney [42] measured the risk perception with only one item or question.

Although China is a country with a long history of flood disasters, few studies have assessed the flood risk perception and analyzed the influence factors of flood risk perception. Kellens reviewed 57 studies of flood risks perception in leading journals and found the study of flood risk perception is mainly in the Western world (more than 40 studies) [43]. Related research in China is still in its infancy, and the published literature is very limited. Besides, the current study areas were usually rural area or the area near the river, less concerned about the differences of flood risk perception in different regions of the same small and medium-sized and flood prone city. Given the lack of research of flood risk perception in China and the significance and important role of risk perception research in flood risk management, the aims of this study are


#### **2. Materials and Methods**

#### *2.1. Study Area*

The study was undertaken in the Jingdezhen City as it is seriously affected by floods almost every year, causing huge economic losses and wide impacts. At the same time, Jingdezhen City is a typical small and medium-sized city in China, with rapid social and economic development. The local economic structure is more fragile than the bigger cities, and the structure and function of the river network can be more easily damaged.

In addition, the Jingdezhen government is also actively developing Integrated Flood Risk Management Plan to reduce the flood risk and it is expected to provide reference for other small and medium-sized cities in China. The concept of the plan is fully realized by the transition from flood control to flood risk management, combining engineering and non-engineering measures to form a comprehensive flood control and disaster reduction system for cities and meeting the needs of the whole society for water security. Therefore, it is very practical to choose Jingdezhen City as a research area. This study is one part of the proposed Integrated Flood Risk Management in Jingdezhen City.

Jingdezhen is located in the northeast of Jiangxi Province, China, and it belongs to the transition zone between the extension of Huangshan Mountain, Huaiyu Mountain and Poyang Lake Plain (as shown in Figure 1). It lies between 116◦57 –117◦42 E longitude and 28◦44 –29◦56 N latitude. Jingdezhen City covers a total area of 5256.23 km<sup>2</sup> and governs two counties and two districts, namely Leping County (1982.76 km2), Fuliang County (2580.84 km2), the Changjiang District (391.83 km2), and Zhushan District (30.80 km2). The agricultural land, construction land, and unused land in Jingdezhen City are 4754.47 km2, 355.9 km2 and 147.8 km2, respectively, accounting for 90.4%, 6.8%, and 2.8% of the total area of the city. The highest and lowest elevations in Jingdezhen are 1618 m and 20 m, respectively, with plains on the southern part having an average altitude of 200 m. This region is characterized by a subtropical monsoon climate, with abundant sunshine and rainfall. The annual average temperature is 17 ◦C and the annual average sunshine time is 2009.8 h. The annual precipitation is 1763.5 mm, and the distributions of precipitation are quite uneven, with about 46% of precipitation occurring in the rainy season (from April to June). In the past 10 years, the economy of Jingdezhen has grown at an annual rate of more than 8%. In 2017, the regional GDP reached 87.825 billion Yuan. At the end

of 2017, the total resident population was 1.665 million, of which the urban resident population was 1.098 million.

**Figure 1.** Study area: Jingdezhen City, Jiangxi Province, China.

Jingdezhen City is prone to floods. On the one hand, this is due to the extreme rainfall and flow during the flood season. Taking the flood in 2016 as an example, the city's 24-h average rainfall exceeded 200 mm with a return period of 20 years and the maximum flow of the Changjiang River was 7090 m3/s. The flood return period of the main stream of the Changjiang River was 20 years, and some river sections were 50 years. On the other hand, Jingdezhen City lacks an effective flood control and drainage project. The existing flood control project can only defend floods with 5–8 years return period. The drainage site design standards are low as well, and some sites stopped working due to flooding. Besides, the length of the urban drainage pipe network is only 26.4 km and only 19% meets the drainage standard of 1-year return period flooding event. Meanwhile, Jingdezhen City has problems with its drainage pipe network, such as many bottleneck pipe sections, mild pipe network slopes, and insufficient pipe network outlet, resulting in regular flooding in Jingdezhen City.

#### Floods in Jingdezhen City

The climate of the Changjiang River Basin in Jingdezhen is significantly changeable, and the precipitation distribution is extremely uneven. The central city of Jingdezhen is located on both sides of the Changjiang River and its tributaries, the Nanhe River and the Xihe River. The terrain along the rivers are low, with an elevation of 24–31 m, which is threatened by floods. In 1955, 1996, 1998, 1999, 2010 to 2012, 2016, and 2017, the floods occurred in the Changjiang River, which caused serious flooding in Jingdezhen City, resulting in large economic losses and social impact.

According to historical records, in 1998, Jingdezhen City suffered a serious flood. The flooding time in the urban area reached 94 h, and the water depth in the urban low-lying area reached 10 m. The urban flooded area reached 31.4 square kilometers, accounting for 94.6% of the total urban area. Besides, the flood affected 354 thousand people and caused 3.23 billion Yuan damage. The flood in 2010 was also very serious, affected 32.52 thousand hectares of crops, with 666.8 thousand people

affected and 1284 collapsed houses. The direct economic loss was about 2.96 billion Yuan. In 2016, from 4 pm on June 18 to 4 pm on the 19th, the rainfall in the Changjiang river basin reached 200.9 mm. The flood level of Dufengkeng Hydrological Station increased by 9.35 m in one day. The city has relocated a total of 119.7 thousand people, which was the largest relocation number of people since the past 10 years. And the direct economic losses amounted to 1.90 billion Yuan.

#### *2.2. Sample Selection*

All four districts were chosen as survey sampling sites because the floods affected all of them and the comparison was need latter. In the heavily flood-hit areas and those in city center, more subdistricts were selected, while fewer subdistricts were chosen in less affected areas. Changjiang District and Zhushan District are the urban centers of Jingdezhen City. The population is large and concentrated. The Changjiang River passes through these two areas and has many low-lying areas. There are more affected populations, so more questionnaires have been set up. Fuliang and Leping are far away from the city center, and their populations are relatively scattered. It is difficult to carry out all investigations in these two areas, thus fewer questionnaire surveys have been set up. As a result, five subdistricts were selected in Changjiang District, eight subdistricts were chosen in Zhushan District, one subdistrict was selected in Leping County, and two townships were selected in Fuliang County.

#### *2.3. Questionnaire Design*

This survey was part of the World Bank Project JIANGXI WUXIKOU INTEGRATED FLOOD MANAGEMENT PROJECT and it was approved by the Jingdezhen City Government. A semi-structured questionnaire was designed to investigate the public flood risk perception. In order to eliminate misunderstanding of the questionnaire, before the formal survey, some respondents with different educational levels were chosen to complete the questionnaire, and their feedbacks on the content of the questionnaire were collected. Then, based on these feedbacks, the project team changed all obscure and professional vocabulary into simple and easy to understand vocabulary. At the same time, the number of questions in the questionnaire was reduced too.

In the introduction part of the questionnaire, the purpose of the investigation and the relevant confidentiality principles were highlighted. This was used to inform respondents that the survey is anonymous and what data were collected. The main content of the questionnaire was divided into three main sections (Table 1). Each section had several items to measure. The first section included six items determining the most important sociodemographic factors of the respondents such as place of residence, gender, age, education level, occupation, and income per month [43]. The second section was other four important factors that could influence the public flood risk perception. The four impact factors were flood experience, flood knowledge education, flood protection responsibility, and trust in government. One impact factor comprised one item. The third section was the measurement of public flood risk perception. The impact and likelihood are most often employed variables to measure the flood risk perception as the flood risk is usually defined by the product of the likelihood of flood disaster with its consequences (impact) [43]. In this study, Jingdezhen City suffers from the flood almost every two years in history. Thus, the impact of the flood was more important than likelihood and only the flood impact was defined as the measurement of flood risk perception.

As mention above, some researchers include many other impact factors such as distance from the river, residence history, etc. But in this study, these factors were not included. This is because in this study, Jingdezhen City is a small and medium-sized city with a small migrant population and local people have lived here for a long time. In addition, not only people near the river are affected by flood, but people living in the low-lying city center are often affected. Likert scale technique was employed for the impact factors in Table 1.


**Table 1.** Definition of measurement and impact factors of public flood risk perception.

#### *2.4. Data Collection*

The data used in this study came from a face-to-face questionnaire survey that was conducted from 8 July 2016 to 14 July 2016. Investigators who attended to the survey were Ph.D. students and MSc students. All of them had basic knowledge and background of natural disaster management. The project team invited four experts in the field of flood risk management to conducted four standardized training sessions for the investigators to make sure that the survey can be carried out smoothly. These trainings mainly focused on the introduction of project objectives and the basic skills for the investigation.

According to the sample selection, in total, 16 subdistricts were selected. Based on the distribution of subdistricts, investigators were divided into four groups of at least four investigators. In each group, there was a senior researcher who monitored the survey process, coordinated the questionnaires collection and checked the completeness and validity of collected questionnaires. Before each interview began, the purpose of the investigation and confidentiality principles were verbally explained by the investigators again. The participation of the respondents in this study was voluntary and consented, and enough time was given. Any confusions of the questions during the survey were explained by the investigators and the respondents had the rights to refuse to participate or withdraw from the survey at any time. In order to incentivize the public to participate the survey, every respondent was given a gift after they completed the questionnaire. At last, 900 questionnaires were distributed, 852 were collected, and the number of valid responses after the exclusion of incomplete questionnaires was 719 (the response rate was 84.4%).

#### *2.5. Statistical Analysis*

Firstly, descriptive statistics were applied to quantitatively describe and summarize the features of the socio-demographic characteristics of the respondents and the other four impact factors (flood experience, flood knowledge education, flood protection responsibility, trust in government). Secondly, analysis of variance (ANOVA) was conducted to examine the mean ranks of two or more independent variables with the null hypothesis of equality. This was to identify whether there existed correlations between the impact factors and public flood risk perception or not. Then, Post Hoc Tests was employed to find and compare the flood risk perception differences between different groups under the same impact factor among all respondents. Thus, the correlations were confirmed to be positive or negative. Besides, with regard to the gender variable, the mean difference comparison among gender using independent T test because only two groups exist, male and female.

All statistical analysis was carried out under a significance test value of 0.05 to confirm whether these impact factors affected the public flood risk perception. Impact factors with the significance value of less than 0.05 were considered to be significantly influential to public flood risk perception, as proposed in similar questionnaire-based studies [2,17,44]. The data was analyzed using the IBM SPSS Statistics software (Version 24.0.0.2, SPSS Inc., Chicago, IL, USA).

#### **3. Results and Discussion**

#### *3.1. Preliminary Analysis*

In this study, 719 valid survey questionnaires were analyzed. Table 2 summarized the response number for each area and the Socio-Demographic Variables of the respondents. The number of respondents in Changjiang District and Zhushan District were higher than in the other areas. Because the more subdistricts were selected, the more responses were collected. Table 3 summarized the distribution of other four important impact factors (flood experience, flood knowledge education, flood protection responsibility, and trust in government).


**Table 2.** The Socio-Demographic variables of the respondents (N = 719).


**Table 3.** The distribution of the respondents regarding flood experience and trust in government (N = 719).

#### 3.1.1. Socio-Demographic Characteristics of Respondents

Of the 719 respondents, 244 (33.9%) were from Changjiang District, 362 (50.3%) from Zhushan District, 52 (7.2%) from Leping County, and 61 (8.5%) from Fuliang County (Table 2). Among the 719 respondents, the percentage of male respondents and female respondents were close, accounting for 46.5% and 53.5%, respectively. Respondents aged from 21 to 50 years old were the majority, with the percentage of 21–35 years old and 35–50 years old were 37.8% and 35.5%, respectively. With regard to the education level, most of the respondents were primary school and below (38.0%) and middle school (42.7%), whereas the percentage of high school, bachelor's degree, and master's degree were only 13.4%, 5.7%, and 0.3%, respectively. Six occupation types were classified, 42.4% and 19.6% of the respondents were self-employed person and worked in company, respectively. Most of respondents earned less than 5000 Chinese Yuan (CNY) every month, income less than 2000 CNY per month accounted for 45.1% and monthly income between 2001 and 5000 CNY were majority, accounting for 49.1%. Only 5.8% earned more than 5000 CNY per month. The specific ratios of the four districts/counties were similar to those of the entire city of Jingdezhen.

#### 3.1.2. Other Important Impact Factors

In this study, when considering the characteristics of Jingdezhen City and the objectives of the study, the flood experience, flood knowledge education, flood protection responsibility, and trust in government were considered as other important factors influencing the public flood risk perception (as shown in Table 3).

Most studies revealed that the flood experience can increase flood risk perception and people with recent flood experience would acquire good knowledge of flood and do well in flood mitigation [9,28]. Most of the respondents in Jingdezhen City were influenced by the flood every year, with 47.5% of the respondents experienced at least one flood every year and only 37.3% of respondents experienced less than one flood every two years. Meanwhile, about 25.4% of respondents in Changjiang District and

15.7% of respondents in Zhushan District were seriously affected by the flood, experiencing more than 2 times flood every year.

It has been found the flood knowledge had a positive correlation with the flood risk perception [45]. Medium level was the largest percentage (28.0%) and only 19.6% of respondents had received many flood knowledge education sessions in Jingdezhen. Besides, 26.3% of respondents never received flood knowledge education. Among the four areas, the respondents in Zhushan District received more knowledge education with 31.5% chosen the answer of "Medium" and 27.9% selected the answer of "Many".

Flood protection responsibility, in many studies, was found to reflect the degree of responsibility to which a person took protection behaviors [46]. In this study, most of the respondents thought the government and flood management experts should take the responsibilities to protect the public from the flood disaster, accounting for 48.4% and 24.6%. In contrast, only 13.9% of respondents believed that they themselves should also be responsible for flood protection and disaster mitigation.

Many findings had showed that people with higher degrees of trust in government perceive lower consequences of disaster and tend to prepare less [47]. When asked about their trust in the local government for protecting them from floods, respondents showed low trust levels in Jingdezhen City, with "low trust" answers accounting for 54.1%, followed by not trust (21.0%), medium trust (18.5%), and high trust (6.4%). In addition, Leping County had the largest percentage of high trust in the government (13.5%), followed by Changjiang District (6.6%), Zhushan District (6.4%).

#### *3.2. Public Flood Risk Perception in Four Districts in Jingdezhen*

The public flood risk perception was compared among four districts. Here, the *p* value was less than 0.05 (Table 4), which showed that the public flood risk perceptions among four districts were statistically significant. Therefore, the comparison of scores between the four districts afterwards was credible and reliable.


**Table 4.** The descriptive statistic of the flood risk perception of case areas.

Note: \* with statistical difference (*p* < 0.05).

When asked about the flood impact, the respondents in Jingdezhen City showed different flood risk perception level. As shown in Figure 2, more than half the respondents selected "medium impact", accounting for 50.3%, and they thought the flood affected their daily life and work but not serious. And 27.1% of the respondents thought that they were largely affected by the flood, causing inconvenience to life and some loss of property, sometimes unable to work or shutdown. Besides, 4.3% of the respondents felt strongly affected by the flood, and they thought their lives and work were greatly affected and sometimes life-threatening property-damaging. On the contrary, "somewhat impact" and "no impact" recorded a small percentage. In specific, 13.5% of respondents thought they were affected by the flood, but this situation only lasted a very short time and 4.7% of the respondents never felt affected by the flood.

**Figure 2.** Distribution of the responses regarding the flood impact, the measurement of the flood risk perception.

The specific difference of flood risk perception was compared among four districts in Jingdezhen City. As shown in Table 4, the mean score of flood risk perception for the respondents in the Jingdezhen City was 3.13. This represented the level of flood risk perception among the Jingdezhen City was not high, close to medium level. Among the four districts, Changjiang District had the highest score of flood risk perception, which was 3.24, 0.11 higher than the whole city average level. In the actual investigation, this may be because the area is the city center, where the wealth and population density are large, and the damage that is caused by the disaster was serious in the past. Meanwhile, some streets and plazas got flooded every year, resulting in a high flood risk perception among the respondents in Changjiang District. The lowest score of flood risk perception was 2.65 in Leping County. This may be due to the respondents in Leping County thought the flood was not a serious disaster. The possible reason is 71.2% of the respondents in Leping County experienced less than one flood every two years (Table 3). With less likelihood of future flood as well as the less flood experience, people in Leping County might believe that flood risk had lower importance. In addition, the flood risk perception in Zhushan District and Fuliang County were similar, being 3.12 and 3.11, respectively.

Overall, these results showed that the respondents in the whole Jingdezhen city had a medium level of flood risk perception, and there were large differences among four districts with the respondents in Changjiang District having the highest flood risk perception while the respondents in Leping County had the lowest.

#### *3.3. Impact Factors of Public Flood Risk Perception*

ANOVA was conducted to examine and analyze the links between the impact factors and public flood risk perception in Jingdezhen City (as shown in Table 5). In this study, the impact factors were related to:


The ANOVA results showed that the impact factors, such as district, gender, age, education level, income per month, flood experience, flood knowledge education, flood protection responsibility and trust in the government have significant relationships with the flood risk perception. Only the occupation factor has insignificant relationship with *p* > 0.05. Then, Post Hoc Tests was conducted to find the flood risk perception differences between different groups under the same impact factor among all respondents (Table 6).


**Table 5.** Analysis of variance (ANOVA) results of impact factors and public flood risk perception.

Note: \* with statistical difference (*p* < 0.05).


Note: \* with statistical difference (*p* < 0.05). The mean difference among gender using independent T test because of only two group exist, male and female.

As far as the district was concerned, the above analysis has found that there was a statistically significant difference between the flood risk perceptions in four districts in Jingdezhen City (see Tables 4 and 5). Compared with other three districts, the respondents in Changjiang District perceived the highest flood risk. Besides, the respondents in Leping County showed the lowest flood risk perception, and this may be because they did not think the flood risk was very important and they believed that the likelihood of future flood in this area was very low.

With regard to gender, the female respondents had higher flood risk perception than the male respondents (*p* < 0.05). This is because women had lower socioeconomic status than men and were more vulnerable when facing the floods, which caused women to be more willing to seek flood information, pay more attention to property losses, and more likely to take self-protection measures in advance. Therefore, female perceived higher flood risk than male. This result was in line with the findings in most published studies [4,48].

In this study, there was a significant positive correlation between age and flood risk perception. In general, the older the respondent, the higher the flood risk perception level. Among the age groups, the respondents aged 51–70 years old had the highest flood risk perception comparing with other age groups younger than 51 years old. This may be because the respondents aged 50–71 years old experienced many historical serious floods and more likely to take the responsibilities in family safety. So, they perceived higher flood risk than other age groups. These results were similar to the findings in most published studies [4,48], although some studies thought that age negatively influenced the flood risk perception [43].

It can be seen from Table 6 that the respondents with higher education level had lower flood risk perception. The respondents with bachelor's degree had lower flood risk perception than those with primary school degree, middle school degree, or high school degree. Meanwhile, there was no big difference between bachelor and master or above. Relevant studies also confirmed the significance of education for risk perception with negative correlation [16]. Ho et al. thought that people with higher education level had lower risk perception because highly educated people were more likely to better understand the flood information and government flood mitigation actions, and thus might feel a higher degree of controllability over a disaster [7].

With regard to the occupation factor, in this study, it had no statistically significant relation to the flood risk perception of the respondents. But, from the empirical data, the self-employed respondents had the highest flood risk perception. This may be because, for the self-employed respondents, the damage caused by the flood needs to be borne by themselves. In the relevant studies, Arnaud et al. revealed that, when discussed about the flood risk reduction, the factor of occupation was never significant [49].

With regard to the monthly income of the respondents, in general, higher educated people had higher income and thus their relation to the flood risk perception was similar. This study found that the respondents with lower income per month showed higher level of risk perception for flood. Compared with the respondents whose average monthly income was less than ¥2000, other three groups had poorer flood risk perception (*p* < 0.05 for all). This result was similar to the findings in some previous studies [43]. For example, Kellens et al. reviewed many relevant studies and found that there was a negative correlation between income and risk perception [50,51], though statistical significance was often absent [7,16,52].

The flood experience of the respondents had been found to be positively correlated with the flood risk perception and it was highly significant in this study, i.e., those with more flood experiences had higher flood risk perception than those with less flood experiences. The respondents experienced more than 2 times flood every year had the highest flood risk perception compared to other three groups. This was because people with more flood experience had more knowledge and better understanding of historical floods, and they were more likely to seek flood information and take measure to protect themselves. In fact, in our field research, it was discovered that the residents who were often affected by disasters made small flood control facilities in front of their homes, stored food during the flood

season, and established mutual assistance agreements between neighbors. These results were also confirmed by the studies of Pagneux et al. [53] and Kellens W et al. [4].

With regard to the flood knowledge education, the flood knowledge is generally found strongly related to the feeling of security. Individuals with little knowledge of the causes of floods had lower flood risk perception [9]. In this study, it has a similar result that the individuals with more flood knowledge education showed higher flood risk perception. Compared with the respondents who received more flood knowledge education, the other three group showed lower flood risk perception. But the difference between the three groups ("Never", "Few" and "Medium") were small, while the gaps between the group of "Many" and other three groups were large. This means that the level of public flood risk perception would be improved only after a certain amount of flood knowledge education. Thus, the government need to adhere to more flood knowledge education.

The view of flood protection responsibilities has been found that can influence the public flood risk perception [33]. In this study, the respondents who believed themselves should be responsible for the flood protection showed higher flood risk perception than other groups. The reason may be that people who feel responsible for taking protective actions usually doubt the effectiveness of 'public' protective measures [34], thus, they perceived higher flood risk perception and preferred to take self-protection measures. Besides, the difference between other groups were not big. These results also reflect the fact that raising public responsibility for flood protection is very helpful for flood mitigation and risk management.

With regard to the trust in the government, it was found to be negatively related to the flood risk perception in this study (see Table 6). The respondents held the "Very low" trust in the government perceived highest flood risk than other three groups (*p* < 0.05 for all). The reason may be that trust in the government is represented by the trust of the government, experts, and the mass media [26]. The high level of trust showed that the respondents believe that the government can cope with flood hazards and do not need to do too much preparedness themselves. People who have lower trust in the government do not believe that the government can issue early flood warning and timely rescue. Instead, they choose to actively understand flood knowledge, seek flood information, and take measures to protect themselves. These results also were confirmed in most published studies [27,54].

#### **4. Conclusions**

This study analyzed the public flood risk perception in Jingdezhen City and explored the impact factors on flood risk perception. Results of ANOVA showed that many impact factors were strongly correlated with the flood risk perception. These include the socio-demographic characteristics of the respondents, previous flood experience, flood knowledge education, flood protection responsibility, and trust in government. The key findings are:


But, the limitation of the study should also be taken into consideration. First, due to the financial and time limitation, only one subdistrict in Leping County and two downtowns in Fuliang District were chosen to carry out the survey. This may influence the results in these two districts. Second, we used a single measure of risk perception that some researchers might not consider to be optimal. Other research tends to use a set of items (e.g., impact, awareness, likelihood and fear) to measure the flood risk perception. Third, although we told respondents that the survey was anonymous, some respondents were still cautious and avoided choosing extreme answers about the government, which could cause some uncertainty in the results. Despite this, we believe that our analysis will help decision makers to develop effective flood risk communication strategies and flood risk reduction policies. First, it can help decision-makers to grasp the difference between the flood risk perceived by residents and the real flood risk, make reasonable risk regulation according to the characteristics of the people, and reduce the irrational behavior caused by this risk perception deviation. For example, too high-risk perception could lead to social panic while too low risk perceptions could lead to negative flood mitigation. Second, this study can be used to understand which groups have a higher risk perception, and the government can specifically encourage and promote flood prevention products such as flood insurance to reduce flood losses. Third, these results about the "flood protection responsibility" and "trust in government" can provide references for local government to strengthen the relationship between the government and the public because the responsibility and trust are very important for policy promotion and implementation.

Future work will focus on how flood risk perception affects people's flood mitigation behavior; conduct a more detailed survey of flood risk perception for specific vulnerable populations, such as students, females, and farmers; highlight the dominant role of the government in flood risk mitigation in small and medium-sized cities, and refine and select more variables related to government; and, analyze and compare the changes in flood risk perception before and after flood disasters. Finally, our findings suggest that the local government should actively promote the government's credibility and enhance the trust between residents and the government. Although this may partially reduce the flood risk perception of residents, higher trust is conducive to the implementation of the government's disaster reduction policy, and also enables residents to cooperate with the government's disaster reduction work. At the same time, it is important to strengthen flood knowledge education and training for residents based on their socio-demographic characteristics to improve the public flood risk perception and the ability of self-protect. Moreover, the government should reconsider their disaster emergency drills, making it simple and understandable, and more operational.

**Author Contributions:** Zhiqiang Wang was responsible for literature search, survey design, data collection, data analysis and he also wrote the initial draft of the manuscript. Huimin Wang, Jing Huang principally conceived the idea for the study and provided financial support. Jinle Kang was responsible for field interviews and data collection. Dawei Han contributed to the figures and the revision of English and style.

**Funding:** This research was funded by Jiangxi Wuxikou Integrated Flood Management Project (Grant number JDZ-WXK-ZX-9), National Natural Science Foundation of China (Grant number 71601070), Postgraduate Research & Practice Innovation Program of Jiangsu Province of China (Grant number KYCX17\_0517) and China Scholarship Council (Grant number 201706710094).

**Acknowledgments:** We would like to express our gratitude to the Jingdezhen district office and street office for the authorization and support to carry out the study. Finally, we thank all the investigators and respondents who participated in the survey.

**Conflicts of Interest:** The authors declare no conflicts of interest.

#### **References**


© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

MDPI St. Alban-Anlage 66 4052 Basel Switzerland Tel. +41 61 683 77 34 Fax +41 61 302 89 18 www.mdpi.com

*Water* Editorial Office E-mail: water@mdpi.com www.mdpi.com/journal/water

MDPI St. Alban-Anlage 66 4052 Basel Switzerland

Tel: +41 61 683 77 34 Fax: +41 61 302 89 18