See discussions, stats, and author profiles for this publication at: http://www.researchgate.net/publication/248880426 A Framework for Modeling Mass Disasters ARTICLE in NATURAL HAZARDS REVIEW · JANUARY 2010 Impact Factor: 1.14 · DOI: 10.1061/(ASCE)NH.1527-6996.0000033 CITATION READS 1 67 1 AUTHOR: Hamed Assaf 27 PUBLICATIONS 191 CITATIONS SEE PROFILE All in-text references underlined in blue are linked to publications on ResearchGate, letting you access and read them immediately. Available from: Hamed Assaf Retrieved on: 09 December 2015 This is a draft copy of the following paper: Assaf, H. (2011), “A Framework for Modeling Mass Disasters”. ASCE Journal of Natural Hazards Review. 12(2): 47- 61. doi:10.1061/(ASCE)NH.1527-6996.0000033. http://link.aip.org/link/?QNH/12/47. A Framework for Modeling Mass Disasters Hamed Assaf, Ph.D., P.Eng, MASCE1 Abstract: The paper introduces a generic mass Disaster Modeling Framework (DMF) to support integrated socio-physical risk assessment and management of disasters. The DMF draws on the principles that disasters have many common features regardless of their instigating hazards and are the outcome of interaction between a hazard and society. The modular architecture of the DMF weakly couples three key applications: the Community Simulation Model (CSM), the Hazard Simulation Model (HSM) and the Evacuation Simulation Model (ESM). The CSM utilizes census, cadastral, and other data to construct an individual-based representation of the potentially impacted community. The CSM estimates people and built environment at risk at any given time of the day, week, and year. The CSM assessment along with the hazard simulation of the HSM feed into the ESM to simulate the progression of fatalities throughout the event. Dam failures, for which vivid and detailed accounts of people’s interaction with the incoming flood exist, served as case studies. CE Database subject headings: risk assessment; loss-of-life models; evacuation models; emergency planning; agent. 1 Assistant Professor, Department of Civil and Environmental Engineering Faculty of Engineering and Architecture, American University of Beirut AUB POBox 11-0236 Riad El Solh, Beirut 1107 2020 Lebanon Tel. (961) 1-350000 x3460, Fax: (961) 1-744462, E-mail: hamedassaf@hotmail.com 1 Introduction One of the pressing and immortal concerns of people is the sudden advent of a mass disaster that could bring death and wreck havoc and destruction on their communities. This concern is shared by societies of different times and places. Whether hazards are floods, earthquakes, tsunamis, volcanic eruptions, wildfires, hurricanes, terrorist attacks, or landslides, societies have placed high priorities on efforts to mitigate, prepare for, respond to and recover from disasters. Regardless of its nature, a mass disaster episode is the result of an interaction between two highly complex, dynamic and generally hard-to-predict phenomena: a human community and a hazard. A massive earthquake in a desolated area or in prehistoric times is of little significance in comparison to an event involving the death of several thousands of people as a result of a much less powerful earthquake in a densely populated area. Consequently, managing the risk of mass disasters requires not only a good understanding of each phenomenon individually, but also a thorough investigation of the mechanics of their interaction. The presented Disaster Modeling Framework (DMF) provides a virtual platform for creating potential manifestations of a given disaster under different community and hazard conditions to assess its risk and devise effective mitigation policies. The DMF is unique in bringing together the two disparate physical and social sciences disaster research fields into a common ground to study and analyze the socio-hazard aspects of disasters. 2 The DMF incorporates three main components: the Community Simulation Model (CSM), the Hazard Simulation Model (HSM) and the Evacuation Simulation Model (ESM). Utilizing census, cadastral and field collected data, the CSM captures details of the potentially impacted community in terms of the characteristics of individuals, their families, and associated objects of building, vehicles and infrastructure. The CSM is used to produce snapshots representing the state of the community at selected times of the year which serve along with output from the HSM as input into the ESM to produce alternative disaster simulations. The process can be streamlined to carry out disaster risk assessment studies or used in setting up disaster management policies. The DMF is fundamentally anchored on the concepts of individual behavior and interactions implemented based on the agent-based approach. In general this approach represents a given system as a virtual society of individuals or agents who follow a set of rules or heuristics in making decisions in response to changes in their environment or in interacting with other individuals. Their collective behavior and interactions give rise to an overall system behavior or phenomena that cannot be deduced from the simple sum of their individual actions (Bonabeau 2002). The DMF views mass disasters and emergencies as complex, nonlinear and highly dynamic systems that evolve or emerge from interactions at the individual level. From this perspective mass disasters and emergencies cannot be adequately represented by aggregate/empirical functions of average conditions that obscure the internal dynamics of these phenomena. Researchers from a wide array of disciplines including ecology, sociology, urban planning, public health and traffic simulation have successfully applied the agent-based approach to solve problems deemed extremely challenging to the more traditional 3 aggregate methods (see Grimm and Railsback 2005, Macy and Willer 2002, Bonabeau 2002, Verburg et al. 2004, Epstein et al. 2002 and Reynolds 1987). Agent-based modeling provides a more natural and meaningful representation of real systems in contrast to the abstract high-level empirical approach (Bonabeau 2002). This has practical implications in terms of validity, applicability and transparency of models. For example empirical methods used in assessing loss of life from dam failures and flash floods are mainly based on fitting statistical functions to historical records (Zhai et al. 2006, Brown and Graham 1988, DeKay and McClelland 1993) with little insight into the low level processes that could greatly sway the outcome of a given episode in one direction or the other. Consequently these methods do not provide explanation for the high variability in the outcome of flood disasters even when their aggregate characteristics, such as size of population count, percentage of exposure, severance of event are within the same range (Assaf 2007). Modeling at the individual level provides a mechanism for normalizing and sharing information and insight into the disaster phenomena by breaking it into low-level processes common to other disasters or related phenomena, which can be analyzed and investigated individually in field or lab work. For example, the lab and analytical results on the stability of people in moving water (Lind et al. 2004, and Abt et al. 1989) can be readily incorporated into the DMF. A large body of information on related processes such as driver behavior, stability of buildings, behavior of people under stress, etc. can be readily used with minimal reprocessing or interpretation. The paper presents implementation of the Life Safety Model (LSM), a forerunner of the DMF, in dam risk studies. The LSM was used to virtually reconstruct the 1959 Malpasset 4 Dam failure capturing highly detailed and visual simulation of people’s response to incoming flood wave and the impact it had on them. Background White et al. (2001) described how hazard research evolved from two distinct schools. Geographers view hazards as natural phenomena with a wide range of variations and implications including positive (environmental) ones. Sociologists on the other hand considered disaster a social construct and directed their efforts on understanding its human dimension and enhancing community response and recovery capabilities. Worldwide concerns over unabated increases in the frequency of hazards and consequent human and material losses have fueled great interest in hazard research. These renewed research efforts led to the convergence of the two lines in hazard research into the now well established vulnerability assessment approach which considers disaster as a manifestation of a hazard facilitated by intrinsic social and geographic conditions (Adger 2006). The concept of vulnerability has been well received in the hazard research community and was adopted as a core principle in the Disaster Risk Reduction initiative by the International Strategy for Disaster Reduction (ISDR) (International Strategy for Disaster Reduction 2004). The proliferation of research in the multi-disciplinary vulnerability research has brought about a long and confusing array of the definitions and interpretations of the concept (Cutter 1996). Several high level frameworks and conceptual models were introduced to streamline and formalize the process of vulnerability analysis (Turner et al. 2003). Brooks (2003) draws a distinction between biophysical vulnerability and social 5 vulnerability. The latter relates to the inherent characteristics of the human system, e.g. poverty and social status that make it susceptible to hazards. The former refers to the potential loss of life and physical damage that could arise from the interaction of hazard with social vulnerability. In presenting the hazards-of-place conceptual model, Cutter (1996) viewed the place as the natural unit for assessing vulnerability as it logically anchors biophysical vulnerability, defined as the probability of occurrence of hazards at the place, to social vulnerability, which reflects characteristics of the people residing in the place. Cutter et al. (2000) applied this conceptual model to assess the overall vulnerability and its spatial distribution for Georgetown County in South Carolina. They applied Geographic Information Systems (GIS) technology to compile, analyze and present information on common hazards and their rate of occurrence in the area, and main socio-economic indicators of social vulnerability. Their analysis showed significant variability in both vulnerabilities and in the overall vulnerability. Significantly, social vulnerability did not always coincide with biophysical vulnerability, which underscores the risk of relying on biophysical vulnerability as the sole indicator of overall vulnerability. Cutter et al. recognized however that their approach assumes equal weight for biophysical hazards and socio-economic indicators, which does not necessarily reflect the actual significance of one hazard, or socio-economic factor over the other. The logical appeal of the hazardsof-place model and its pragmatic approach based on utilizing readily available information and technology to seamlessly identify areas of high vulnerability makes it an ideal tool for emergency management professionals and urban planners. 6 By assessing vulnerability based on place of “residence”, the hazards-of-place approach overlooks the mobility of population, a key factor in determining the location of people when a hazard strikes an area. This is particularly important for rapid onset and intense hazards, such as chemical spills, earthquakes, flash floods and tornadoes which constitute most of those considered by the study. Peoples’ pattern of movement follows diurnal, weekly and seasonal patterns. Consequently, the time of day, week and year is an important factor in assessing vulnerability. For example, a tsunami that hits a popular summer resort would result in vastly different outcomes depending on the season it occurs. The influence of patterns of movement on vulnerability could be accounted for by considering the rate of occupancy of a place (Assaf et al. 1997). The temporal dimension of vulnerability is more appropriately represented through capturing the dynamics of population of the given community. An equally significant and related issue is the ability to simulate peoples’ behavior and action during a disaster or mass emergency. These issues were of a prime concern in the efforts at British Columbia Hydro Corporation (BC Hydro), a large dam owner in Canada, to develop more reliable method for assessing potential dam breaches on downstream population. In the first phase of this initiative, the author developed a new conceptual framework for modeling people’s response to a dam breach flood and its impact on them and buildings. He later led efforts to implement this modeling approach as the BC Hydro Loss of Life (LOL) model (Assaf et al. 1997). Although the LOL model determined survivability per individual as a function of water depth and velocity, it did not represent interaction among individuals and depicted the environment as one-dimensional with escape routes perpendicular to channel flow. To overcome these limitations, BC Hydro supported 7 another phase of activities to develop the 2-dimensional Life Safety Model (LSM) (Assaf and Hartford (2002), Assaf (2002) and (Assaf and Hartford (2001)). The LSM integrates output from a flood simulation model with virtual world made up of individual entities representing people, buildings, vehicles and roads. The DMF evolved from the LSM to address a wider range of mass disasters. While the LSM is fundamentally a support tool for dam risk assessment, the DMF provides a generic platform to capture the dynamics of different types of mass disasters in support of risk assessment and emergency management. Despite being an earlier product, the LSM in retrospect represents an implementation of the DMF architecture for dam safety studies. Consequently, examples of how the DMF could be implemented are based on the LSM. The paper presents the architecture and main components of the DMF and illustrates its application in dam safety risk assessment. Review of disaster/emergency models The issue of mass emergencies and disasters has attracted great interest from researchers in a wide range of disciplines such as dam safety, risk analysis, hydrology, seismology, traffic modeling, and building safety. Multitude of models and procedures were developed to address problems ranging from risk assessment, emergency preparedness, and evacuation planning. Table (1) presents an overview of models reported in the literature classified according to their purpose and approach into loss-of-life empirical formula, building evacuation models, traffic evacuation models, population models, flood evacuation models and loss assessment models. 8 Earlier procedures were empirically based and consisted mostly of formulas that estimate overall loss of life and physical damages as functions of global variables such as total exposed population and property, and parameters that relate to warning and intensity of the hazard. For example DeKay and McClelland (1993) used historical events to develop a formula expressing loss of life from floods and dam breaches as a function of total population, average warning time and lethality of flood wave. Zahi et al. (2006) developed a formula for estimating loss of life from floods in Japan based on the number of houses inundated and type of flood, whether a flash flood or a rainy season one. Badal et al. (2005) estimated number of earthquake casualties in Spain as a function of population density and magnitude of earthquake. These empirical formulas are still widely used due to their simplicity. However, they fail to capture the spatial variability of bio-physical and social vulnerabilities, raising concerns about their use in dam risk assessment (Hartford and Kartha 1995). Challenges in planning emergency evacuation from large and complex buildings have motivated the development of several building evacuation simulation models. The SAFER model represents the indoor environment of rooms, stairways and hallways as a network of nodes, with people movement simulated as flow between these nodes (Gupta and Yadav 2004). The model applies networks optimization methods to determine optimal evacuation paths. The EGRESS model represents the building environment as a grid, with people and obstacles as grid points (Santos and Aguirre 2004.) Grid points representing people move with variable speeds towards exits. The more advanced BuildingEXODUS model represents people as agents evacuating and interacting based on a set of rules (Gwynne et al. 2005). By design, building evacuation models are restricted 9 in their application to enclosed environments and are not readily applicable to simulate natural, open-space or spatially extensive disasters. Emergency management could involve mass evacuation of communities in anticipation of natural hazards such as wildfires and hurricanes. Cova and Johnson (2002) have leveraged the sophisticated capabilities of Paramics, a leading traffic simulation commercial software, to simulate evacuation of a small community exposed to the risk of wildfires. Paramics uses agent-technology to simulate drivers’ behavior and traffic flow. Cova and Johnson used a statistical trip generator to produce input for Paramics and develop a set of optimal evacuation scenarios. However they do not offer a clear guidance on carrying out multi-scenario analysis or selecting values for the statistical parameters. Reliable estimation of population is necessary to carry out risk assessment studies and emergency management. In support of these and other activities including security, public health and energy requirement studies, the U.S. Department of Energy’s (USDOE) Oak Ridge National Laboratory (ORNL) has developed LandScan to estimate population worldwide and USA wide (at 3 arc-second resolution) using census data, GIS and remote sensing based on key indicators including land use, elevation, and proximity to transportation, and places of activities. The model provides both nighttime and baseline daytime population estimates. The first estimate represents resident population, while the later compromises student and worker populations (Bhaduri et al. 2007). In response to major concerns over nation-wide impact of earthquakes on buildings and infrastructure, the US Federal Emergency Management Agency (FEMA) sponsored several studies that led to the development of HAZUS in 1997 as an earthquake loss 10 assessment software tool. Over the following few years FEMA significantly restructured and repositioned the software as a multi-hazard integrated assessment system. Renamed HAZUS-MH (MH refers to multi-hazard) the new system contains an extensive GIS database on buildings and census information and supports a consistent set of tools to assess potential damage from three main hazards: earthquake, hurricanes and floods (Schneider and Schauer, 2006). The impact assessment can be conducted at three levels of ascending complexity and detail of analysis. The first level involves using information from a default database to produce a low cost preliminary loss assessment. Highest level requires using more detailed local data and relatively more sophisticated tools. HAZUSMH calculates loss of life only for earthquakes, and not for hurricanes or floods. This is possibly since HAZUS-MH estimates loss of life as function of damage to buildings. Earthquakes occur with no warning and consequently do not trigger a pre-evacuation, making damages to buildings a good indicator of casualties. In contrast hurricanes and floods develop over hours or days giving population the opportunity to evacuate albeit at different timings and speeds. This makes loss of life less or inconsistently associated with the level of damage to buildings occupied prior to the hazard onset. Each of the applications discussed above handles one or two aspects of assessing the consequences of a mass emergency or disaster. The DMF integrates several modules to represent collectively the key processes that give rise to a disaster. A population module represents the society as a set of individuals with characteristics (age, gender and physical mobility), and links to other individuals, places of activities, vehicles and roads that influence their behavior and interaction as they become aware and make decisions regarding evacuation. A hazard module provides the spatio-temporal simulation of the 11 hazard medium. Inputs from the two modules are then integrated into a simulator that tracks movement and fate of each individual throughout the disaster episode. Architecture of the Disaster Modeling Framework The DMF provides a generic, flexible and intuitive virtual environment to examine the social as well as the physical aspects of a mass disaster and estimate loss of life. Drawing upon experience in earlier modeling efforts to assess loss of life from potential dam failures and realizing the great similarities among disasters regardless of their instigating hazards (Scawthorn et al. 2006) the author developed a more holistic and generic approach to modeling mass disasters as articulated in the high-level architecture of the DMF shown in Figure 1. The DMF couples two modules, the Community Simulation Model (CSM) and the Hazard Simulation Model (HSM), representing the potentially affected community and the hazard, respectively, into a geo-referenced Evacuation Simulation Model (ESM) that tracks the unfolding of a potential disaster at the individual levels of persons, families and other groups, buildings, places of activities, vehicles and infrastructure. The simulation produces individual-based information (numbers, locations, socio-demographic characteristics, escape attempt tracks, reception and compliance of warning and physical conditions) on the potentially affected, injured and killed populations. Simulation output also contains information on damages to buildings and infrastructure. A GIS based database stores simulation results to facilitate data quality control, analysis and presentation using common GIS tools. A central Framework Manager Module (FMM) coordinates processing and interactions among different modules. 12 There are two setup levels for the DMF. At the low level, specialized users in hazard and population modeling set up the HSM and CSM, respectively. At the high level, users such as policy makers, risk assessment and emergency management professionals formulate scenarios by defining time of day, week and year and level of hazard. Users can then process these scenarios to create input files into the ESM, which in turn simulates the disaster event. A DMF scenario represents a unique combination of certain community and hazard static and dynamic conditions and disaster mitigation policies. In dam risk studies, for example, analysts formulate several scenarios, each representing a unique combination of water levels in the reservoir, which determine the scale of the flooding event, and the time of day, week and year of failure which dictates the occupancy density and level of alertness. Figure 2 shows the scenario tree configured for a dam risk study showing the outcome of each scenario in terms of the Population at Risk (PAR), estimated Loss of Life (LOL) and corresponding probability of occurrence. The DMF provides flexibility in the choice of the community and hazard simulation models as long as their outputs can be processed to be compatible with the text-based format of the ESM. This equally applies to the selection of the visualization and analysis tools. The main modules of the DMF are discussed in the following sections. The Community Simulation Model (CSM) Despite its richness and long history, the literature on addressing the human aspect of disasters is lacking in providing explicit and quantitative representation of potentially affected communities necessary for objective, transparent and reliable analysis and 13 management of disasters. The few modeling efforts reported in the literature are empirical and anecdotal in nature and generally address single processes only (Blanchard-Boehm 1998; Gillespie and Murty 1991; Gladwin et al. 2001; Kirschenbaum 1992; Perry 1989; Perry and Lindell 1991; Sorensen 1991). The CSM utilizes information from different sources to create a virtual world of individuals each with a set of characteristics that define its behavior, interactions with other individuals, movement and vulnerability to the hazard medium. The CSM defines a set of classes representing different types of people, e.g. a working man with family, male student, working women, single mother, senior, disabled, toddler, etc. These classes act as master copies from which individuals are instantiated. People classes contain statistical parameters from which individual characteristics are sampled. The CSM uses census, property, survey and other data to determine distribution of classes among the potentially affected population. The class attributes and their statistical parameters are derived from information in the literature or could be determined via dedicated research. For example, Assaf and Hartford (2002) used findings from studies on human stability in floodwater (Lind et al. 2004, RESCDAM 2000, and Abt et al. 1989) to calculate statistical means of toppling water depths and velocity for men and women. They in turn used them to statistically generate these parameters for the virtual individuals. Findings from RESCDAM (2000) were used to generate combinations of water depth and velocities that would topple building of wooden structures. Detailed analysis of past mass emergencies or disasters including Malpasset (Snell and Smith 1959, and French Ministry of Sustainable Development 2009) and Teton (McDonald, 2006) dam failures using technical, documentaries, media-based as well as 14 anecdotal information provides valuable insight on circumstances and decisions and actions made by people that have significant impact on the course of events at the individual or global level. For example, the decision to return to the hazard after a warning was the principle reason for most deaths in the Teton dam failure. Assaf et al. (1997) represented this behavior as a parameter in the BC Hydro LOL model that will trigger the return of some individuals to the site to pick up belongings after a long warning without occurrence of the hazard. The approach here emphasizes learning from detailed accounts of individual actions and events within the larger disaster event to develop better estimates of rules of behavior and survivability factors. As presented in a later section, the approach has been successfully validated and applied in dam breach studies, and is proposed here for application for other disasters and mass emergencies. The CSM Design The CSM portrays a community from two perspectives: a persistent Static View and a Snapshot. The former represents attributes and objects that are permanent or relatively stable during a disaster assessment study including people’s age, ethnic background, physical characteristics, family ties, places of residence, work and study, buildings, infrastructure and vehicles. Generated from the former, the latter is analogous to an image captured of the community at a given moment showing where and with who each individual is situated, his/her sate of alertness (awake, asleep), activity (working, driving, walking, etc), and reception and compliance of warnings. In a typical operation of the CSM, the analyst configures a single static view and then uses it to generate several snapshots representing potential settings of the community at the onset of the disaster. In 15 assessing individual extreme events, such as an earthquake during a busy tourist season or a midnight flash flood in a residential area, the analyst runs a snapshot representing the given time of day, week and year along with an HSM-simulated hazard event into the ESM to produce a simulation of the potential disaster event. In disaster risk assessment studies, the analyst may set up a batch run to generate a set of snapshots representing the full spectrum of the year and run them in the ESM in different combinations with a set of HSM-simulated hazard events. Each disaster scenario generated by a pair of a given snapshot and a given hazard event has a probability of occurrence calculated equal to the product of the frequency of the time of the year of the snapshot and the probability of occurrence of the hazard event. The DMF streamline this assessment process as shown in Figure 3. The CSM represents a potentially affected population as a virtual community of individuals, their families, neighbors, work groups and their associated places of activities, buildings, vehicles, road network and open spaces. Figure 4 illustrates the topology of the CSM community. The CSM uniquely identifies each person in the community and assigns him/her attributes that describe his/her family ties, age, gender, occupation, ethnic background, education, preparedness, mobility, and physical characteristics. The CSM passes on this information to the ESM which uses them to simulate a given person’s behavior and determine his/her survivability in encountering a given hazard medium. The CSM places special focus on the family unit since familial behavior highly influences individual behavior and consequently plays a key role in disasters, particularly with respect to preparedness, reception and compliance of warning and evacuation (Hultåker 1983; Kirschenbaum 2006; Carter et al. 1983; Clason 1983). 16 For example, the CSM assigns parents links to their children, used by the ESM to prevent family members from evacuating independently. The CSM adopts a geo-referenced spatial system to determine locations of individuals via their associations with three classes of spatial objects: buildings, open spaces and road networks. Individuals are not hard-linked to specific locations. For a selected time of the year, week and year, the CSM assigns on the fly individuals to specific locations based on their patterns of activity as represented by the Whereabouts schedule and their associated places of activities. The Whereabouts schedule specifies the likelihoods of the person being engaged in specific activities including working, staying at home, driving, etc. Typically during the setup of the static view the CSM assigns a person a whereabouts schedule from a set of whereabouts schedule templates defined based on identified categories of people including working person, household parent, school children, etc. For example, Table 2 depicts a whereabouts schedule for a working person. The schedule categorizes activities seasonally, weekly and daily in addition to those during holiday times. For a given snapshot a statistical generator determines based on the Whereabouts likelihoods the activity the person is engaged in, which consequently identifies the place of activity that is either associated with a building, an open space, an outside the area, or a section of the road network if the person is driving or commuting. The buildings are important elements in the CSM since they contain the space where people spend most their time in an average year. Not only they provide safe or perceived safe havens for people, but they are also at risk of getting destroyed or failing to protect their inhabitants during disasters. The building object contains several attributes used in determining a building’s stability including type of material, age, height and construction 17 standards. The building act as a container to the place of activity object, which links people to buildings as described above. The place of activity concept provides flexibility in developing representation of complex multi-use built environments such as business complexes and shopping centers. Mobility of population at risk - in terms of presence of drivers and access to vehicles - is critical to the success of evacuation efforts (Cova 2005). This is particularly important for events such as wildfires and hurricanes where a window of few hours could be available to evacuate people. In many dam breaches cases also, several hours to days are available for evacuation as breach develops, flood wave travels toward sites situated far downstream, or if a good detection and warning system is available. This modernly afforded mobility can however act as a double edge sword. French et al. (1983) found that 43% of 190 deaths reported in 16 flash floods in the US are vehicle related. Drobot et al. (2007) indicated that over half of the death in floods involved vehicles. Vehicles provide people with a fast way out of a danger zone, yet unmanaged vehicular evacuation could lead to massive traffic jams that may seriously hamper evacuation efforts. Rohde (2002) indicated that evacuees flow and emergency force during the Laguna Hills Fire in California were locked up in serious traffic jams that have exacerbated this catastrophe. The CSM represents vehicles as objects with attributes describing their capacity and structural stability, and assign them to individuals based on information on car ownership. The CSM models road networks as collections of topologically connected spatial linear elements and nodes. The ESM uses road networks to simulate traffic movement and evacuation. 18 Generation of a Static View In implementation, the static view is a GIS database composed of a set of interrelated tables each maps to a corresponding object as conceptually represented in Figure 4. Relationships among tables map to corresponding relationships among the conceptual objects. The cost and effort required to populate a static view database vary depending on the population density, land use and size of the area under consideration and the scope and objective of the DMF study. For example, a small (2 to 3 people) investigation team may require one to two weeks to develop a static view database for a small community of up to a 100 dwellings, such as a trailer community in a wildfire-prone area, or a small Tibetan village exposed to the risk of snow avalanches. In contrast, large urban areas may require several experts including some with solid GIS experience working for several months to collect compile and corroborate information from several sources that may include census, cadastral and municipal data. In the former case, the study team can manually enter information in the GIS database on the locations and characteristics of buildings, characteristics of residents (age, sex, etc.) and mobility. In the later case, the size, heterogeneity and the varied quality of information may complicate the process of developing the static view. To overcome these problems, Assaf and Hartford (2002) developed an iterative semi-automated routine to compile census and cadastral data in addition to field surveys and other locally available data to produce a static view of a given urban area. They applied this routine in a BC Hydro dam safety study of the Keenleyside dam located few kilometers from the City of Castlegar, B.C. The study involved assessing the seismic risk to the dam and its implications to downstream population in Castlegar through investigating several alternative dam failure scenarios. 19 Assaf and Hartford developed the static view routine to process Canadian census data, however the routine can be modified to handle census data of different format given that it contains similar type of information. Statistics Canada is responsible for collecting, compiling and disseminating census data in Canada. Statistics Canada collects census data per individual household for most of the population that covers a wide range of inquiries. However, to maintain privacy, Statistics Canada publishes census information aggregated as demographic statistics per enumeration areas (EA), each made up of few to several hundreds of households. The Canadian EA statistics include type of residential groups (e.g. Family with children, single mother, etc.), their size, age and gender structure (number of children and age categories, senior citizens, etc) in addition to other information (Statistics Canada 1996) (see the top table in Figure 5). A typical cadastral data in Canada includes use type of buildings, their construction type, age and physical condition as shown in the sample table in the second row from the top in Figure 5. The cadastral data includes an extensive array of items which are not considered here. The static view routine includes several steps. For a given community, the routine generates the static view as described below with references to the example presented in Figure 5. 1. For each EA the routine generates residential groups with size, gender and age structure that match those of the census demographics for the given EA. To do this the routine first uses a statistical random generator to create the same number and type of residential groups in the census data, which readily determine the number and gender of parents and seniors. The routine then calculates the number and age groups of children through several (3 to 5) iterations to maintain consistency between the total number and 20 age group of children and the family structure demographics. The routine then assigns each residential group a unique ID and stores it in a table that also shows it’s associated EA and the number of individuals in each category (man, woman, adult (unknown gender), senior and children of different age groups). As an example, the routine used the demographics of the EA #59011160 shown in the upper table in Figure 5 to generate over 200 residential groups. Five of these groups appear in the third table from the top, which is a partial display of the complete Residential Groups table. The highlighted residential group #3177 is a family-with-children type that has a father, mother, a son or daughter 18 to 24 years of age and a child less than 6 years old. At this stage, the routine has associated each residential group with a specific EA, but left the determination of its exact location to a later step. 2. Using cadastral data that includes information on the type and use of the buildings, the routine identifies the specific locations where different activities take place including residences, schools, markets, hospitals, etc. It then stores this information in corresponding tables representing places of residence, work, school, shopping, recreational. For example, the routine used the information on the apartment building with ID #2855 in the middle table to the right (Figure 5) to generate several apartments as places of residence. Five of these apartments appear in the middle table to the left, which is a partial display of the Places of Residence table. The routine then used spatial processing to identify the EAs where the buildings are located. 3. The routine assigns the residential groups to specific places of residences located within the same EA through an automated matching process that corroborates information from the two independent sources of information of census and cadastral 21 data. For example, the routine only assigns senior residential groups to places of residences classified as senior residences. It also gives family residential groups higher priority for being assigned to detached homes. In the example presented in Figure 5, the residential group #3177 resides in apartment #4532 located in building #2855. 4. An analyst quality checks the generated GIS database for inconsistencies such as placing residential groups disproportionally in certain parts of a given EA or locating large number of children per several families in an area known to have a low count of children per family. The analyst may modify the database manually or rerun one or more of the first three steps few times to obtain a logically consistent representation of the region. The analyst may also manually modify the database to corroborate additional information from field surveys, eyewitness observations and other sources. In particular, this approach facilitates incorporating higher quality data to provide more detailed representation of highly significant areas or hotspots such as those closer to the hazard medium or those that house highly vulnerable segments of the populations such as children or the elderly, e.g. schools, hospitals and senior residences. 5. The routine uses information on residential groups to instantiate individual persons, where each person is described in terms of his/her age, gender, and family links to other members in the group (father, mother, son/daughter). In the example shown in Figure 5 the routine instantiated the following four individuals for the residential group (#3177): a man (#7884), woman (#7885), a son or daughter aged 18 to 24 years (#7886) and a child less than 6 years (#7887). The routine uses the age and gender of a given individual to assign him/her characteristics that describe his/her behavior and survivability in a disaster. These characteristics include those that are generally common 22 for different types of hazard such as interpretation and compliance of warning, risk taking attitude and mobility. Other characteristics are hazard dependant and include for example the toppling depth and velocity of flood waters (Assaf and Harford 2002). In general behavior and survivability in disasters is highly associated with the person’s demographic category. For example young men are generally more risk taking, agile and more capable in surviving contact with a hazard medium in comparison to young children and elderly citizens who are the most exposed and vulnerable groups in disasters. It is estimated, for example, that third of the Indian Ocean Tsunami victims were children (Penrose and Takaki 2006). 6. The routine utilizes Information on schools and colleges records, labor records and additional census data to assign individuals categories as dependent, students, employed, housewives, and unemployed. The routine then assigns each person based on his/her employment status, attendance of school, and dependency, a whereabouts table that reflects his/her pattern of activity. The routine selects the whereabouts table from a set of templates configured for different types of persons including full-time workers, home staying parent, and students. 7. Lastly, the routine uses information on vehicle ownership and traffic intensity patterns to assign vehicles to residential groups and persons. It also assigns vehicles characteristics that determine their vulnerability to the given hazard medium. For example vehicles are reported to be less stable in flood water in comparison to people (Walsh and Benning 1998). The static view database of Castlegar study contained approximately 9,000 people, 3,800 residential groups and 3,500 buildings. BC Hydro team of 5 experts spent about 6 months 23 to develop this database. However, a large share of the effort went into developing and testing the automation routine. Through merging this otherwise disparate information, the static view database provides a new insight into the community’s state of readiness and vulnerability to mass hazards. It could also highlight critical hotspots that would have been missed by following a more traditional approach of data gathering. For example, in the Keenleyside dam safety study, the automation routine highlighted based on the census data the presence of a significant number of senior citizens in a specific EA and linked them to a large senior apartment building residence situated very critically in the 100-year flood zone. A field investigation team confirmed this information and further incorporated additional field information on the exact number of suites and some of the building’s structural characteristics such as its height, construction material, etc. Given that it relies mainly on aggregated data, the above procedure is not expected to produce a virtual world of objects that match their corresponding real entities on a one-toone basis. It does however produce a virtual world with demographics that matches the real one. That also holds true to two static views created via independent runs. In fact an infinite number of different, yet demographically similar, static views can be created through this process. Running these static views through the remaining procedures of the DMF over a wide spectrum of conditions would produce different results. Sensitivity runs should be conducted where several static views are run through the model to assess the expected range of outcome. 24 Generation of Snapshots A configured static view contains the necessary information to generate snapshot views of the community at selected times of the day, week and year. The generation process involves randomly placing each person in the community at a place of activity based on the probabilities specified in the Whereabouts table. Using Table 1 for example, there is 84% chance that the person will be located at his/her place of work during daytime hours between 8 am and 5 pm. In contrast, there is only 5% chance that he/she is on the road during the same period. The snapshot view also captures several attributes that relate to the readiness of a given individual to the hazard including the state of alertness (sleep or awake), access to vehicles, proximity to independent/dependent individuals. The CSM explicitly considers this latter attribute through the group object which ties individuals through family or work links. For example, the CSM model parents to maintain proximity to their children during a crisis, even if their chances of survival were to improve if they act individually. The CSM can also capture group behavior under nonfamily situations such as the tendency of people to help strangers in a mass emergency (Goltz 1984). The snapshot view also contains attributes related to the group or individual warning compliance and evacuation decision making process during crisis. For example, people vary in their responses to the threat of floods in terms of evacuating buildings, delaying evacuation or returning to get belongings, driving through running water (a common cause of death), and evacuating by vehicles or on foot. The CSM assigns these behaviors to individuals and groups based on statistical values obtained from established disaster research (e.g. Kirschenbaum 1992). 25 The Hazard Simulation Model The DMF approach calls for leveraging existing hazard simulation models by projecting their output onto the virtual disaster simulation environment. As such an HSM can be chosen from established hazard-specific simulation models. In the case of flooding disasters, HSM can be TELEMAC or MIKE 11 which are leading flood simulation models. The DMF processes output from the HSM to produce estimates of key variables at the grid nodes of the ESM. Key variables vary by the type of hazard and could include wind speed, water depth and velocity, and chemical spill plume concentration and temperature. The upstream position of the HSM in the DMF modular architecture indicates that the HSM is not required to process feedback from the CSM or ESM. This design arrangement provides flexibility in the choice of the HSM, and reduces the need to develop, usually at very high cost, an in-house HSM or significantly customize an external one. The Evacuation Simulation Model The ESM simulates the response of people to an impending mass hazard and the outcome when they and the main elements of their environment come in contact with the hazard medium. The ESM superimposes the HSM output on a geo-referenced environment it creates based on the given CSM snapshot view. It then determines the outcome of the hazard/community contact at each grid point per given simulation time step based on the values of superimposed hazard variables and characteristics of community objects. The grid points could be spaced irregularly so that more grid points are used in areas where higher details are required and vice versa. 26 The ESM incorporates several modules to capture the dynamics of a disaster. It tracks each person individually or as a member of a group throughout the disaster event. The person receives warning, responds to it by evacuating by foot or in vehicle or staying put, may survive the event without encountering the hazard medium or may come in contact with it, survives with no injury, or gets injured or killed. The ESM simulates these processes via a central modular algorithm as shown in Figure 6, where each process is represented as a block in the algorithm flow chart. The first block in Figure 6 represents reception of warning through formal and informal means including automatic warning systems, police, media, other people and physical observation. The ESM emulates the warning process as a grid of interconnected virtual warning centers distributed throughout the impacted area. The DMF analyst can set up one or more warning center(s) to become active on a given time or get activated once it comes within certain distance from the hazard medium. Once triggered a warning center initiates a warning which it passes on with specified delay and rate of dissemination, reflecting the reliability and efficiency of the warning system, to people and other warning centers. A person can also initiate a warning once he/she becomes aware through own observation or via information by other people or warning centers. This parallel propagation of warning mimics real life warning where people become aware of the hazard through several interdependent sources such police warnings, telephone calls by relatives and friends, radio and TV broadcasts, or own sensing of the hazard or signs of its impending arrival, such as smoke of wild fire or the sight of a hurricane from a distance. 27 The following decision block represents the response to a warning, which the ESM implements as a person/group behavioral and mobility process that involves the decision to evacuate or stay in place and delay evacuation to overcome shock, get ready, assist or rescue other family members, and collect belongings. The ESM determines how individuals behave by assigning them logical tags that relate to their perception of protection provided by the building, personal risk taking attitudes, and confidence of the warning. The ESM assign these tags based on statistical means derived from disaster literature. For example, family links take precedence over decision to evacuate, where parents are modeled to delay evacuation until getting all family members ready to evacuate. This is a particularly important aspect of disaster social behavior that has been strongly supported by disaster research (Kirschenbaum 2006). Depending on the reception of warning and the outcome of the decision to evacuate, a person/group has three alternatives to consider: staying in the building, or evacuating on foot or by vehicle. For those who were already outside buildings either walking or on the road at the onset of the disaster will follow the corresponding logic shown for pedestrians and those in vehicles. The outcome of not evacuating a given building depends on whether the building would come in contact with the hazard medium, and if so whether it can protect its occupants from the hazard effect. This depends on the nature of the hazard and the building’s structural characteristics. Tall and well constructed concrete buildings are considered resilient to floods especially if occupants manage to escape to the upper floors of the building. Light wood buildings are considered more ideal in earthquake disasters. The ESM represents these characteristics as parameters passed on from the CSM. These 28 parameters could include age of the structure, construction material, number of floors, etc. The ESM processes these parameters and the hazard variables at a given building’s location into a hazard-specific module to determine the state of the building throughout the simulation run. For example, the ESM uses the height of the building and construction material along with the flood depth and velocity to estimate the damage to the building during a flood disaster simulation run. Individuals who choose to evacuate a building are faced with the decision to select a method of evacuation and a destination. Assaf and Hartford (2001) have introduced the concept of a Perceived Safe Haven (PSH) to represent potential evacuation destinations that are deemed, and may not be necessarily, safe by a given person/group. Depending on the nature of the hazard, preparedness level and type of transportation, the ESM assigns each person/group a set of PSHs, which could be shared with other persons/groups. Spatially, the PSH can be a point, a line or an area. For example, the DMF analyst can designate a 200-year flood contour line as a PSH. The PSH concept offers the analyst flexibility in representing unwarranted evacuation schemes or situations where the disaster outcome got worse as people evacuated to unexpectedly unsafe locations. Several factors influence a person’s choice of the method of evacuation including the nature of the hazard, the location of the PSH, availability of vehicle and condition of the traffic network. For example, an effective method to avoid being swamped by a flash flood is to climb the river valley wall by foot rather than driving a car along a river road. The ESM includes a pedestrian and a vehicular transportation networks to simulate both evacuation by foot and vehicle. A person/group evacuating by foot will move with a speed calculated initially based on physical characteristics tied to gender and age 29 information obtained via the CSM. The ESM may modify the speed of a certain group to reflect the presence of physically less capable members such as children. The model assesses the status of the person as he/she comes in contact with the hazard medium based on the hazard’s variables and the person’s characteristics. In the case of floods, the ESM uses water depth and velocity along with age and sex of the person to estimate his/her chances of being drowned (Assaf and Hartford 2002). Evacuation by vehicle differs in important aspects from that on foot. Vehicles are much faster and may provide additional protection, yet their movement is dependent on the functionality of the transportation system, which could be greatly impacted by traffic jams and potential disruption, e.g. due to flooding, or destruction, e.g. due to earthquakes, of the road network. These are important issues in emergency and disaster management, where uncoordinated efforts could lead to unwarranted all-out evacuation creating traffic jams that impedes the evacuation of truly impacted areas. The ESM includes a fully integrated individual-based traffic simulation model, which incorporates basic driver behavior including car following, collision avoidance and gap acceptance. The ESM models vehicles as individual entity with characteristics that determine their capacity and resilience to hazards. It models the traffic infrastructure as networks of road segments and nodes - created from GIS road network data or prepared or edited manually - that determine road connectivity, capacity and flow direction. It is important to emphasize the importance of separating the functionalities and the weak coupling among the three DMF core models. The ESM acts as a simulation engine that expects information on the state of both the community and hazard from the CSM and HSM. Consequently the ESM is not restricted to a particular CSM or HSM. In fact input 30 to ESM can be edited and prepared manually giving great flexibility to use in a variety of case studies. For example, the DMF analyst can choose to manually prepare the ESM for small scale disasters involving remote communities or villages, where the data can be efficiently field collected and processed in a timely fashion. Demonstration of the Potential of the DMF as a virtual disaster simulation environment As a demonstration of the potential of DMF in simulating disaster events particularly at the individual level, this section presents a brief description of an extensive forensic analysis of the Malpasset Dam failure disaster to validate the LSM. An investigation team from BC Hydro and a consulting firm conducted this validation study over several months where it examined a large set of information sources including official French national census, cadastral and mapping data, published documents, newspapers and magazines, video documentaries, photos, interviews and exchange with experts and eyewitnesses (Johnstone et al. 2003). The team developed a GIS database to store and link these pieces of information. The database provided input data to the LSM and reference information to assess its simulation of actual events. At the time of its completion in 1955, the Malpasset concrete dam was a symbol of France’s technical edge in developing aesthetically appealing and structurally sound dams. The exceptionally thin double curvature dam was dubbed as “a beautiful, shapely lady - very daring and very delicate” (Snell and Smith 1959). The dam reservoir provided water supply and irrigation to the downstream Reyran river valley and city of Frejus on the French Riviera. It took 4 years till the dam was first filled at the end of 1959. Due to an oversight in geotechnical analysis of the bedrock, the dam foundation was not 31 adequately designed and gave way under the immense momentum and torque of the 66.5 m-high water mass. A 40 m high wave moving initially at 70 km/hr rushed through the steep narrow valley washing everything in its way including debris, buildings, trees, vehicles and human bodies (Snell and Smith 1959). Large several-ton concrete pieces torn from the dam core were found strewn few hundred meters downstream of the dam. It took 25 minutes for the wave to travel 12 kms to reach Frejus and finally emptying into the Mediterranean 16 kms downstream of the dam leaving several hundred (estimates range from 423 to over 550) people dead and destroying over 150 buildings in addition to roads, railways and crops. Close to 7,000 people were believed to be affected by the disaster (Johnstone et al. 2003). To represent the flood wave characteristics the validation study relied on a well calibrated hydrodynamic simulation of the Malpasset Dam breach produced by the Laboratoire National d’Hydraulique (LNH) of the Electricité de France (EDF) using the TELEMAC2D model (Hervouet 2000a and Hervouet 2000b). The LSM estimated loss of life from the Malpasset dam failure to range from 424 to 514 which is well within the reported values of 423 and 550 (Johnstone et al. 2005). The model was also able to virtually replicate detailed individual events including the escape of the dam operator with his wife to a high ground nearby after hearing loud cracking sounds very shortly before the collapse of the dam. It also captured the sudden death of 30 highway construction workers camped few hundred meters downstream of the dam (Snell and Smith 1959). Four computer screen snapshots of the Malpasset event simulation are presented in Figure 7 to illustrate the dynamic capabilities of the LSM environment. The figure shows the progression of the Malpasset dam breach disaster in an area located 1 kilometer 32 downstream of the dam. The upper left image shows the flood wave approaching the area at 1:50 minutes after the dam breach, with some residents groups (labeled 1, 2, 3, 4 and 5) already evacuating the area to a higher area via a bridge spanning the river. The upper right image shows the situation 10 seconds later, with the flood wave only few meters away from some residences. The lengths and directions of flow vectors indicate the magnitudes and the directions of corresponding flows, respectively. Resident groups #1, #3, #4 and #5 have moved northwest towards the safe area, with resident #2, who happened to be pedestrian, lagging behind. 10 seconds later, the flood wave has overcome resident groups #10, #11, #14 and #15 as shown in the lower left image. Residents #12 and #13 have moved south towards the bridge. Another resident (#18) is shown moving in from the south towards the bridge. The last image shows the flood wave moving swiftly to the south, sweeping away additional residents (#6, #7, #8 and #9) and closing in on residents #2, #12, #13 and #18. These simulation results constitute a portion of 1.5 hour ½ second-step simulation run that covers the whole area from the dam site to the river outlet into the Mediterranean. Summary and Conclusions The Disaster Modeling Framework (DMF) offers a novel approach to capture the highly complex, dynamic and spatially variable human and physical aspects of disasters. Its flexible architecture integrates the individual-based Community Simulation Model (CSM) with the physical Hazard Simulation Model (HSM) to simulate a combined human/hazard scenario through the visual environment of the Evacuation Simulation Model (ESM). The CSM untangles the highly intertwined dynamic and static aspects of a human society into the snapshot-static views pattern to facilitate creating logically 33 coherent and consistent scenarios. The ESM utilizes Agent based approach to simulate people’s behavior, interaction and movement as they become aware of an impending hazard or come in contact with its medium. It is important to emphasize that the objective of the DMF is not to predict but rather to provide a range of potential outcomes of mass disasters under given conditions. The aftermath of a historic disaster represents one realization of the ensemble of potential outcomes that the disaster could have turned into. Calibrating the DMF should not lean heavily on matching the outcomes of real disasters but rather to gain insight into the processes that led to these outcomes. Although the DMF was designed to provide a one-stop and comprehensive environment to assess and plan for disasters and emergencies, its architecture is flexible to integrate with other existing systems. In particular, HAZUS-MH with its vast database and standardized toolset can provide valuable input to populate and calibrate the DMF components. It can also be used as a prescreening tool to identify and prioritize high risk areas. The DMF constitutes an important step towards a more integrated modeling approach to analyze and manage mass disasters. It also has the potential of providing a virtual environment for analyzing the impact of risk perception and social factors on loss of life. It should be noted however that since the DMF has not been verified for applications other than those of dam breach studies, the framework is still considered in the experimental stage with respect to modeling other hazards. Acknowledgement 34 The author thanks the editor and the three anonymous reviewers for their valued comments and suggestions which have considerably improved the quality of the paper. It should be noted that the first version of this manuscript was submitted for publication on November, 2006. 35 References Abt, S. R., Wittler, R. J., Taylor, A., and Love, D. J. (1989). “Human Stability in a High Hazard Flood Zone.” JAWRA Journal of the American Water Resources Association, 25(4), 881-890. Adger, W.N. (2006). “Vulnerability.” Global Environmental Change, 16(3), 268-281. Assaf, H. (2007). “Discussion –"An Empirical Model of Fatalities and Injuries Due to Floods in Japan" by Guofang Zhai, Teruki Fukuzono, and Saburo Ikeda.” JAWRA Journal of the American Water Resources Association, 43(5), 1344-1346. Assaf, H. (2002). “Understanding dam emergencies and formulating plans through virtual reality modeling.” EAP 2002 International Workshop for Emergency Preparedness at Dams, sponsored by FERC and ASDSO, Niagara Falls, USA. Assaf, H. and Hartford, D. N. D. (2002). “A virtual reality approach to public protection and emergency preparedness planning in dam safety analysis.” Proc., Canadian Dam Association (CDA) 2002 Annual Conference, CDA, Victoria, Canada. Assaf, H., and Hartford, D.N.D. (2001). “Physically-based Modeling of Life Safety Considerations in Water Resource Decision-Making.” Proc., The ASCE World Water and Environmental Resources Congress, Orlando, Florida. Assaf, H., Hartford, D. N. D, and Cattanach, J. D. (1997). “Estimating Dam Breach Flood Survival Probabilities”. Proceedings of the Australian Committee on Large Dams (ANCOLD) Conference on Dams, Perth & Kununurra, Australia. 36 Badal, J., Vazquez-Prada, M., and Gonzalez A., (2005). “Preliminary quantitative assessment of earthquake casualties and damages.” Natural Hazards, 34(3), 353-474. Bhaduri, B., Bright, E., Coleman, P. and Urban, M. (2007). “LandScan USA: A High Resolution Geospatial and Temporal Modeling Approach for Population Distribution and Dynamics.” Geo Journal, 69(1-2), 103-117. Blanchard-Boehm, R. (1998). “Understanding public response to increased risk from natural hazards: application of the hazards risk communication framework.” International Journal of Mass Emergencies and Disasters, 16(3), 247-278. Bonabeau, E. (2002). “Agent-based modeling: Methods and techniques for simulating human systems.” Proceedings of the National Academy of Sciences of the United States of America, 99(2), 7280-7287. Brooks, N. (2003). “Vulnerability, risk and adaptation: A conceptual framework.” Working Paper 38, Tyndall Centre for Climate Change Research, University of East Anglia, Norwich. Brown, C. A. and Graham, W. J., (1988). “Assessing the Threat to Life From Dam Failure.” Water Resources Bulletin, 24(6):1303-1309. Carter, T. M., Kendall, S., and Clark, J. P. (1983). “Household response to warnings.” International Journal of Mass Emergencies and Disasters, 1(1), 95-104. Clason, C. (1983). “The family as a life-saver in disaster?” International Journal of Mass Emergencies and Disasters, 1(1), 43-62. 37 Cova, T.J. (2005). “Public safety in the urban-wildland interface: should fire-prone communities have a maximum occupancy?” Natural Hazard Review, 6(3), 99-108. Cova, T. J., and Johnson, J. P. (2002). “Microsimulation of neighborhood evacuations in the urban – wildland interface.” Environment and Planning A, 34(12), 2211 – 2229. Cutter, S. (1996). “Vulnerability to environmental hazards.” Progress In Human Geography 20(4), 529–589. Cutter, S. L., Mitchell, J. T., and Scott, M. S. (2000). “Revealing the vulnerability of people and places: a case study of Georgetown County, South Carolina.” Annals of the Association of American Geographers, 90(4), 713–737. DeKay, M. L. and McClelland, G. H., (1993). Predicting Loss of Life in Cases of Dam Failure and Flash Flood. Risk Analysis, 13(2), 193-205. Drobot, S.D., Benight, C., and Gruntfest, E., (2007). “Risk factors for driving into flooded roads.” Environmental Hazards, 7(3), 227-234. Epstein, J. M., Cummings, A. T., Chakravarty, S., Singa, R. M. and Burke, D. S. (2002) “Toward a containment strategy for smallpox bioterror: an individual-based computational approach”, Center on Social and Economic Dynamics, Working Paper No. 31. French, J., Ing, R., von Allmen, S., and Wood, R. (1983). “Mortality from flash floods: a review of the national weather service reports, 1969 – 1981.” Public Health Reports, 98 (6), 584 – 588. 38 French Ministry of Sustainable Development (2009) “Burst of a dam: 2 December 1959 Malpasset [Var].” French Ministry of Sustainable Development, France. Gillespie, D. F. and Murty, S. A. (1991). “Setting boundaries for research on organizational capacity to evacuate.” International Journal of Mass Emergencies and Disasters, 9(2), 201-218. Gladwin, C. H., Gladwin, H., and Peacock W. G. (2001). “Modeling hurricane evacuation decisions with ethnographic method.” International Journal of Mass Emergencies and Disasters, 19(2), 117-143. Goltz, J. D. (1984). “Are the news media responsible for the disaster myths?: a content analysis of emergency response imagery.” International Journal of Mass Emergencies and Disasters, 2(3), 345-368. Grimm, V., and Railsback, S. F. (2005). “Individual-based Modeling and Ecology”. Princeton, NJ, Princeton University Press. Gupta, A. and Yadav, P. (2004). “SAFE-R: a new model to study the evacuation profile of a building.” Fire Safety, 39(7), 539 - 56. Gwynne, S., Galea, E., Owen, M., Lawrence, P.J., and Filippidis, L. (2005). “A systematic comparison of buildingEXODUS predictions with experimental data from the Stapelfeldt trials and the Milburn House evacuation.” Applied mathematical modeling, 29(9), 818–851. 39 Hartford, D. N. D., and Kartha, C. V. (1995). “Dam breach inundation and consequence evaluation. How safe is your dam? Is it safe enough?” British Columbia Hydro Cooperation, Canada, Report no. MEP11-5. 264 p. Hervouet, J-M. (2000a). “TELEMAC modelling system: an overview.” Hydrological Processes, 14(13), 2209–2210. Hervouet, J-M. (2000b). “A high resolution 2-D dam-break model using parallelization.” Hydrological Processes, 14(13), 2211-2230. Hultåker, Ö. (1983). “Introduction: family and disaster (special issue).” International Journal of Mass Emergencies and Disasters, 1(1), 7-18. International Strategy for Disaster Reduction (2004). Living with Risk: A global review of disaster reduction initiatives. UN Publications, Geneva. Johnstone, W., Assaf, H., Sakamoto, D., and Hartford, D. (2003). “Analysis of the Malpasset dam failure using GIS and engineering models” Proceedings of GeoTec 2003. Vancouver. Johnstone, W.M., Sakamoto, D., Assaf, H., and Bourban, S. (2005). “Architecture, Modelling framework and validation of BC Hydro’s Virtual Reality Life Safety Model.” Proceedings of the International Symposium on Stochastic Hydraulics, Nijmegen, the Netherlands. Kirschenbaum, A. (1992). “Warning and evacuation during a mass disaster: a multivariate decision making model.” International Journal of Mass Emergencies and Disasters, 10(1), 91-114. 40 Kirschenbaum, A. (2006). “Families and disaster behavior: a reassessment of family preparedness.” International Journal of Mass Emergencies and Disasters, 24(1), 111– 143. Lind, N., Hartford, D., and Assaf, H. (2004). “Hydrodynamic Models of Human Stability in a Flood.” JAWRA Journal of the American Water Resources Association, 40(1) 89-96. Penrose, A. and Takaki, M. (2006). “Children’s rights in emergencies and disasters.” The Lancet, 367(9511), 698-699. Macy, M. W. and Willer, R. (2002). "From factors to actors: Computational sociology and agent-based modeling", Annual Review of Sociology, 28(1), 143-166. McDonald, D. J, (2006). “Images of America: The Teton Dam Disaster”. San Francisco, CA, Arcadia Publishing. Perry, R. W. (1989). “Taxonomy and model building for emergency warning response.” International Journal of Mass Emergencies and Disasters, 7(3), 305-327. Perry, R. W. and Lindell, M. K. (1991). “The effects of ethnicity on evacuation decisionmaking.” International Journal of Mass Emergencies and Disasters, 9(1), 47-68. RESCDAM (2000). “The use of physical models in dam-break flood analysis.” Helsinki University of Technology, Helsinki, Finland. Reynolds, C. W. (1987). “Flocks, herds, and schools: a distributed behavioral model”. Comput. Graphics 21(4), 25–34. 41 Rohde, M.S. (2002). Command decisions during catastrophic urban-interface wildfire: A case study of the 1993 Orange County Laguna Fire. M.A. thesis, California State University, Long Beach, CA. Santos, G. and Aguirre, B. E. (2004). “A critical review of emergency evacuation simulation models.” Disaster Research Center, University of Delaware, Newark, DE, USA. Scawthorn, C., Schneider P. J., and Schauer, B. A. (2006). “Natural hazards – the multihazard approach.” Natural Hazards Review, 7(2), 39. Schneider P. and Schauer B. (2006). “HAZUS- its development and its future”, Natural Hazard Review, 7(2), 40-44. Sorensen, J. H. (1991). “When shall we leave?: factors affecting the timing of evacuation departures.” International Journal of Mass Emergencies and Disasters, 9(2), 153-165. Snell, D. and Smith, G. (1959). “Valley is washed in horror when a dam gives way”, Life Magazine, December 14. Statistics Canada (1996). GeoSuite, 1996 Census, Catalogue 92F0085XCB. Turner II., B.L., Kasperson, R. E., Matson, P., McCarthy, J. J., Corell, R. W., Lindsey, C., Eckley, N., Kasperson, J. X., Luers, A., Mertello, M. L., Polsky, C., Pulsipher, A., and Schiller, A. (2003). “Framework for vulnerability analysis in sustainability science.” Proceedings of the National Academy of Sciences of the United States of America, 100(14), 8074–8079. 42 Verburg, P. H., Schot, P. P., Dijst, M.J., and Veldkamp, A. (2004). “Land Use Change Modelling: Current Practice and Research Priorities,” GeoJournal , 61(4), 309-324. Walsh, M., and Benning, N. (1998) “Defining Flood Hazard in Urban Environments.”, Proceedings of the Second SIA Regional Conference on Stormwater, Stormwater Industry Association, Batemans Bay, Australia, April 27-28. White, G. F., Kates, R.W., and Burton, I. (2001) “Knowing better and losing even more: the use ofknowledge in hazards management.” Environmental Hazards, 3(3-4), 81–92. Zhai, G., Fukuzono T., and Ikeda S. (2006). “An Empirical Model of Fatalities and Injuries Due to Floods in Japan.” JAWRA Journal of the American Water Resources Association, 42(4), 863–875. 43 Tables Table 1. Overview of models used in disaster risk assessment and emergency planning Name/Reference Purpose Type Description US Bureau of Risk Loss-of-life Input data: total population Reclamation assessment empirical at risk (PAR), warning time (USBR) method of dam formula and average water depth; (Brown and breach and Graham 1988) flash floods Methodology: a nonlinear equation developed based on 22 cases of dam breach failures and flash floods; Extreme events were excluded; Pros: simple to use; Cons: cannot be used to estimate loss of life based on sub divisions of PAR; does not represent extreme events. Dekay and Risk Loss-of-life Input data: total PAR, McCelland (1993) assessment empirical warning time, and flood of dam formula lethality (assigned 1 if 20% 44 breach and or more of the buildings are flash floods destroyed, zero otherwise); Methodology: based on a large set of flood events. Pros: considered the most rigorous of all empirical formula as it was developed on a much larger database; Cons: cannot be used to estimate loss of life at more than two subdivisions of PAR. Zahi et al. (2006) Risk Loss-of-life Input data: number of flood assessment empirical affected residential buildings of flash formula and whether the event is a floods and flash flood or rainy season low level one. floods. Methodology: developed based on a large set of Japanese flood events. Pros: Easy to use. 45 Cons: can produce great variability in estimates since it relies on logarithmic transformation and does not differentiate between severely or lightly impacted buildings (Assaf 2007). Badal et al. (2005) Risk Loss-of-life Input data: population assessment empirical density, earthquake’s of formula magnitude and duration; earthquakes Methodology: the formula was derived based on logtransformed regression analysis of earthquakes events in the 20th century. Pros: easy to use in risk assessment studies; Cons: logarithmic transformation can results in great underestimation of large events. 46 SAFE-R (Gupta and Yadav 2004) Emergency Building Input data: representation of management evacuation evacuation environment of buildings (stairs, rooms, hallways, etc.) model as a network of nodes, the number of people in each node, travel time between nodes, and designation of destination nodes; Methodology: the model simulates evacuation of people from buildings; network optimization techniques are used to determine optimal evacuation paths; Pros: captures physical characteristics of evacuation environment; Cons: No representation of social interaction. EGRESS Emergency Building Input Data: evacuation 47 (Santos management evacuation space is represented as a grid, of buildings people and obstacles are and Aguirre 2004) model represented as grid cells, hazard (chemical spill, etc.) can be also represented as grid points; Methodology: based on the cellular Automata approach; Pros: individuals can be assigned a limited number of characteristics (e.g. speed); computationally less intensive; Cons: does not represent social interaction. BuildingEXODUS Emergency (Gwynne et al. 2005) Building Input data: selection of management evacuation nodes and arcs representing of buildings spaces and distances between model them, respectively; range of values representing characteristics of population (e.g., age, weight and agility); 48 Methodology: simulates evacuation of people from a building; the evacuation environment is represented as an interconnected set of nodes and arcs; evacuee are represented individually with a set of characteristics and rules determining their agility and behavior; Pros: provides more realistic representation of the evacuation environment; ability to simulate toxicity; Cons: data intensive. Paramics (Cova and Johnson 2002) Emergency Traffic Input data: road network management evacuation and parcel data, destinations, of natural and statistical information model hazards (e.g. used in estimating number of outdoor departing vehicles and fires). departure timing; 49 Methodology: Cova and Johnson (2002) used a commercial traffic simulation software, Paramics, to simulate evacuation scenarios based on trips created by a custom scenario generator; Paramics simulates individual vehicles movements in a transportation network; Pros: Leveraging of a well established commercial microsimulator; Cons: does not facilitate multi-scenario analysis; no clear guidance on selection of statistical parameters. LandScan Estimation Population Input data: Census data, (Bhaduri et al. of model land use, transportation, 2007) population imagery, elevation, 50 educational institutions and jails Methodology: provides estimates of nighttime and daytime population. Nighttime population is estimated by distributing the population for each census block over 1 arc-second grid based on weights representing land use, proximity to places of activities (schools, parks etc…) and transportation elements. The daytime population is estimated based on information related to work and school activities. Pros: provides a global coverage and comprehensive coverage for the USA at 3 arc-second (about 90 meters) 51 resolution. Cons: Temporal distribution only covers nighttime and baseline daytime populations. BC Hydro Loss of Dam risk flood Input data: census data, Life (LOL) model assessment evacuation cadastral and building use model data, elevation, hydraulic (Assaf et al. 1997) modeling, and dam operation policies. Methodology: Census data is distributed based on building use data. Daily, weekly and yearly occupancy factors determine the temporal distribution of population. The model simulates the escape of population away from flood path. It estimates survivability based on water depth and velocity and individual characteristics. Pros: Provided a novel 52 approach to estimate spatial and temporal distribution of population under the risk of a dam failure. Cons: A one-dimensional model with no explicit representation of individual interaction. BC Hydro Life Dam risk flood Input data: census data, Safety Model assessment evacuation cadastral and building use model data, topography, information (LSM) (Assaf and Hartford 2002) on workers student and hospitable, GIS in addition to field surveys. Methodology: utilizes agents technology to model peoples’ behaviors and interactions; simulates vehicular and pedestrian movement; and simulate fate of each individual as function of physical robustness and 53 exposure to flood water. Pros: a significant enhancement of the BC LOL by simulating explicit individual interaction and movement in a more realistic two dimensional virtual environment. Cons: strictly applicable to flood disasters. The Hazards U.S. Multi- Loss Input data: GIS based Multi-Hazard hazard loss assessment database on buildings and (HAZUS-MH) assessment model census data, buildings (Schneider and damage functions for three Schauer 2006) main hazards: earthquakes, hurricanes and floods. Methodology: Provides impact assessment of main hazards at three levels of ascending complexity and accuracy. At the lowest level, the model provides 54 preliminary assessment utilizing information from a default database. At higher levels analysts can utilize more detailed information to develop more reliable assessments. Pros: provides a USA nationwide tool for hazard and risk assessment. Cons: calculate loss of life only for earthquake hazards. Does not simulate evacuation. Table 2 Example of a whereabouts table for a full-time worker. Time of Year September 1 to April 30 May 1 to August 31 Time of Week Weekdays (Mon-Fri) Weekends & Statutory Weekdays (Mon-Fri) Holidays Time of Day 8am- 5pm- 10pm- 10am-8pm 8pm-10am 8am5pm 10pm 8am 5pm 5pm10pm 10pm8am Weekends & Statutory Holidays 10am-8pm 8pm-10am 0.0435 0.0609 Frequency of 0.1694 0.0941 0.1883 period per Year 0.0881 0.1233 At Home 5 55 80 37 80 5 42 70 25 60 At Work 84 5 1 3 1 60 3 1 1 1 At School 0 0 0 0 0 0 0 0 0 0 0.0871 0.0484 0.0968 55 Shopping 3 20 5 25 1 5 20 5 25 2 Recreational 1 10 7 25 10 12 12 15 30 25 On Road 5 7 5 7 5 8 8 5 10 7 Out of Area 2 3 2 3 3 10 15 4 9 5 56 57 Fig. 1. Architecture of the Disaster Modeling Framework Time of Year Reservoir’s Level (m) Time of Week Day 8am - 5pm JAN1 - APR31 425 Weekdays 5pm - 10pm 10pm - 8am Weekends 10am - 8pm 8pm - 10am 8am - 5pm MAY1 - JUN30 430 Weekdays 5pm - 10pm 10pm - 8am Weekends 10am - 8pm 8pm - 10am 8am - 5pm JUL1 - AUG31 440 Weekdays 5pm - 10pm 10pm - 8am Weekends 10am - 8pm 8pm - 10am 8am - 5pm SEP1 - DEC31 PAR = Population at Risk LOL = Loss of Life 435 Weekdays 5pm - 10pm 10pm - 8am Weekends 10am - 8pm 8pm - 10am PAR LOL Prob. # # # # # # # # # # ##E-10 ##E-10 ##E-10 ##E-10 ##E-10 ### ### ### ### ### ## ## ## ## ## ##E-10 ##E-10 ##E-10 ##E-10 ##E-10 #,### #,### #,### #,### #,### ### ### ### ### ### ##E-10 ##E-10 ##E-10 ##E-10 ##E-10 #,### #,### #,### #,### #,### ## ### ### ### ### ##E-10 ##E-10 ##E-10 ##E-10 ##E-10 Fig. 2. Output from a dam safety DMF application 58 Fig. 3. Streamlining CSM output in the DMF Fig. 4. The Topology of the CSM Objects 59 Enumeration Areas (EAs) Places of Residence-(PoR) Buildings (Bldgs) Residential Groups (RGs) Individuals Fig. 5. Example showing the process of populating the static view’s GIS database 60 P/G : Person or Group Bldg: Building Vhcl: Vehicle P/G Receive Warning Stay in Bldg Decide to Use Vhcl Walk P/G Stay in Bldg Hazard Ended? P/G Walk YES Hazard Ended? NO NO Contact w Hazard? NO NO YES YES NO Bldg Destroyed? YES Contact w Hazard? NO NO Reached Safety? YES P/G in Vhcl YES YES Hazard Ended? NO Reached Safety? YES NO Contact w Hazard? NO YES NO Vhcl Destroyed? Injured or Killed YES YES P/G Survived P/G Injured or Killed Fig. 6. The ESM central algorithm 61 Fig. 7. Snapshots of a dam breach/community simulation scenario 62