Risk decision-making is an integral part of natural hazards risk assessment. It is a multi-dimensional and multidisciplinary activity embracing physical, socioeconomic, and management-related factors at different spatial and temporal levels. Some risk decision-making tasks for natural hazards are conflict resolution, assessing risk patterns, calculating risk indices, and making rational policy. This short article takes bushfire as an example to illustrate risk decision-making in natural hazards.
Risk decision-making serves as a very useful exercise to understand semi- or ill-structured decision-making tasks by streamlining thinking and devising rational procedures, rather than prescribing a definitive solution to a problem in the complex real world. Formulating a risk decision-making process is not simple. First, to make an informed risk decision requires the understanding of many aspects of risk assessment tasks, including the evaluation of the integrated hazard environment, hazard occurrence probability, hazard process, community vulnerability and hazard consequences. For example, the hazard environment of a particular bushfire-prone area in the urban fringe can be quantified in the following three ways: (1) from a historical perspective, discovering relationships between bushfire occurrence frequencies and a complex set of environmental and socioeconomic data; (2) from a horizontal perspective, evaluating integrated environmental and socioeconomic characteristics of a bushfire-prone area; and (3) from a landscape pattern - ecological process perspective, conducting a landscape ecology study of different bushlands burnt over the years. Historically, bushfire occurrence probability and behaviour models range from semi-empirical descriptive models to spatially explicit models based on GIS databases and spatial analyses. Semi-empirical models were developed with data on temperature, wind, moist-ure, terrain, fuel loads, etc. from a large number of historical and experimental fires. A classic bushfire danger rating and behaviour system is McArthur’s “Forest Fire Danger Meter Mark 5” using a series of inputs to produce a fire danger index ranging from 0 to 100 (McArthur Mark 5 in Figure 1). Five fire danger categories can be classified: low (<5), moderate (5-12), high (12-24), very high (24-50), and extreme (50-100). This system is hazard-oriented, yet risk rating systems based on community vulnerabilities have a great use for risk managers.
Second, risk decision-making is integrated in nature, incorporating multiple, incommensurate and possibly conflicting multivariate factors. A decision-making process should reflect the interests and values of different stakeholders, therefore it is important to devise specific mechanisms of consensus building. Often, risk decision-making is likely to be complicated by political issues in a society.
Third, procedures to carry out decision-making should be open and interpretable, otherwise, decision-making processes may be conceived as guesswork. Often, a “one answer” scenario is not enough, and a conclusion based on “many answers” derived from a series of decision-making models, is more enlightening. In addition to theoretical methods, the development of operational decision support tools to produce solutions effectively and efficiently is essential. Tools can support and evaluate “what-if” scenarios easily by altering parameters used in different decision-making methods.
Fourth, it is desirable that decision results are spatially dis-aggregated, allowing policy to be evaluated at a spatial scale as detailed as possible. The use of GIS in risk assessment results in the increasing development of spatially distributed models replacing simple spatially aggregated or lumped parameter models. With higher spatial resolutions, detailed risk decision-making can provide much-needed practical information and knowledge which would eventually assist risk managers to manage hazards-susceptible assets and properties effectively.
For risk decision-making at a local geographical scale, the following three levels can be conducted: (1) census collection districts, (2) street blocks, and (3) individual dwellings (Figure 2). The census collection districts are the smallest areal units for Australian census collection and each district contains about 250 dwellings on average in an urban area; streets are physical boundaries in the built environment and contain special social significance for hazards (e.g., fires, floods) prevention, evacuation and emergency access; site-specific individual dwellings are the basic units in the built environment as major hazards recipients. Risk assessment at the individual dwelling level represents an ideal case under many circumstances.
As risk decision-making is concerned with spatial preferences and/or patterns, conventional multicriteria evaluation (MCE) with GIS can support such decision-making tasks as prioritisation and selection. The MCE-GIS approach provides a mechanism to assemble, weight, synthesise and analyse a range of spatial data layers, particularly for some tasks for which simple rules can be formulated among factors. We recently applied the MCE-GIS approach to an example of bushfire prescribed burning planning for a local bushfire-prone area at the urban fringe. Relative risk ratings for PerilAUS II were developed using the same methodology.
Recent developments in spatial decision-making show that artificial intelligence (including artificial neural networks, fuzzy sets theory, approximate reasoning, genetic algorithms) offers new opportunities to combine multiple criteria and explore patterns among them. For example, artificial neural networks (ANN) are generally superior to traditional statistical methods when: (1) data exhibit unpredictable non-linearity, (2) patterns important to the decision-making tasks are deeply hidden, and (3) data are fuzzy-oriented involving human opinions or ill-defined categories. During the past decade many advanced paradigms integrating individual components of artificial intelligence have emerged. A prominent research area in developing hybrid systems is integrating ANN, fuzzy sets theory, and genetic algorithms. It is envisaged that these advanced methodologies could examine complex spatial risk decision-making tasks that would be impossible to approach and address when using individual methodologies.
Using the ANN methodology, we recently conducted a preliminary study of assessing the risk patterns or probabilities of house survival under bushfires, with a simulation data set based on an empirical study by Wilson and Ferguson (1986). They considered a number of independent influencing factors to predict the probability of house survival with logistic regressions. Those independent factors included attendance by residents during bushfires (X1), presence of flammable objects (X2), roof material (X3) and pitch (X4), fire intensity (X5, X6, X7), wall material (X8), and presence of plants (X9), and they were treated as binary variables. For example, there is X1=1 if a house is unattended, X1=0 for attended case. Finally, the estimated probability curve of house survival (P) as a function of a composite hazards score (H) was produced by them, P=[EXP(6.3-H)]/[1+EXP(6.3-H)], where H=2.2*X1+0.8*X2+ 2.4*X3+1.4*X4+5.2*X5+3.6*X6+2.3*X7+1.1*X8+0.6*X9 (Figure 3). The coefficient of each factor indicates its relative importance. In Figure 3 the house No 1 (fire intensity 25,000 kW m-1, unattended, wooden roof, non-brick wall, presence of plants) has a hazard score of 11.5, corresponding to a survival probability of 0.005! The aim of our study was to fit, remodel and predict the relationship between house survival probability and those influencing factors with ANN. Various configurations for input and output variables were tested on a property-by-property basis. The input factor of fire intensity was treated as binary variables, linguistic terms or standardised forms; the output factor of risk patterns was treated as either continuous probabilities or discrete risk levels. Encouraging results were achieved and suggest that the ANN approach has a great deal of capability and flexibility for such risk pattern classification, discovering the hidden non-linear relationships between complex input and output factors. As ANN can approximate virtually any linear or non-linear functions, they can be used to supplement traditional statistical approaches, which are based on the assumption of a linear additive nature of many hazard and vulnerability factors.
In summary, theunderstanding of many aspects of risk assessment tasks can provide a solid base for risk decision-making in support of risk management in practice. Risk decision-making requires innovative decision-making theories, tools and applications to meet the challenges of detailed risk analysis for risk managers and decision makers. The artificial intelligence approach introduced above can be used for many risk decision-making tasks relevant to damage assessment (e.g., the estimation of potential damage patterns for flood-prone residential buildings), and is independent of the geographical scales associated. As such methodologies for risk decision-making in natural hazards are in their infancy, wider applications should be pursued in the future.
Wilson, A.A.G., and Ferguson, I.S., 1986. Predicting the probability of house survival during bushfires. Journal of Environmental Management, 23, pp. 259-270.
For further information please
Risk Frontiers - Natural Hazards