**The Science and Opportunity of Wildfire Risk Assessment**

Matthew P. Thompson, Alan A. Ager, Mark A. Finney, Dave E. Calkin and Nicole M. Vaillant *US Forest Service USA* 

#### **1. Introduction**

98 Novel Approaches and Their Applications in Risk Assessment

Zanobetti, A. and J. Schwartz. (2009). A Novel Approach to Estimate Distributed Lag Model

Zhao Y., J. Staudenmayer, B.A. Coull and M. P. Wand (2006). General design bayesian

generalized linear mixed models. *Statistical Science*. 21(1), pp 35-51.

*Epidemiology*: November - Volume 20 - Issue 6 - p S62

Between Hospital Admissions and Ozone: A Multi-City Time Series Analysis.

Wildfire management within the United States continues to increase in complexity, as the converging drivers of (1) increased development into fire-prone areas, (2) accumulated fuels from historic management practices, and (3) climate change potentially magnify threats to social and ecological values (Bruins et al., 2010; Gude et al., 2008; Littell et al., 2009). The need for wildfire risk assessment tools continues to grow, as land management agencies attempt to map wildfire risk and develop strategies for mitigation. Developing and employing wildfire risk assessment models can aid management decision-making, and can facilitate prioritization of investments in mitigating losses and restoring fire on fire prone landscapes. Further, assessment models can be used for monitoring trends in wildfire risk over space and across time.

The term risk is generally used to measure the chance of loss, as determined from estimates of likelihood and magnitude of particular outcomes. Probabilistic approaches to risk assessment estimate the expected value of the conditional probability of the even occurring and the consequence of the event given that it has occurred. Risk assessments are conducted when predicted outcomes are uncertain, but possible outcomes can be described and their likelihoods can be estimated. Wildfire risk assessment entails projecting wildfire extent and intensity, and the consequences of fires interacting with values-at-risk.

We begin by introducing a conceptual model of wildfire management (Figure 1) that considers the major drivers of wildfire risk and strategic options for mitigation. Ignition processes influence the spatiotemporal pattern of wildfire occurrence (natural and humancaused), and strategic prevention efforts can reduce the number of wildfires and associated damage (Prestemon et al., 2010). Given an ignition that escapes suppression, fuel, weather, and topography jointly drive wildfire behavior. Of these, only fuel conditions (loading, structure, continuity) can be altered to induce desirable changes in fire behavior (Agee & Skinner 2005). Suppression efforts are intended to slow the growth of active wildfires and reduce the chance of loss. Collectively these factors influence wildfire extent and intensity, which in turn determine the consequences (detrimental and beneficial) to social and ecological values. Wildfire losses can also be prevented or reduced by activities that lessen

The Science and Opportunity of Wildfire Risk Assessment 101

geographic locations and ownerships. First we review concepts of hazard and risk in the wildfire management context. Second, we describe newer developments in the application of burn probability modeling for exposure analysis, and illustrate how this modeling approach can inform fuel management and wildfire suppression efforts. Third, we discuss challenges in quantifying risk for the array of non-market values that are the primary management concern on federal lands, and how expert judgment can be used to advance wildfire effects analysis. We use examples from recent and ongoing broad scale risk assessments and describe their use for informing strategic policy. Lastly we conclude by discussing potential benefits to wildfire management and policy from embracing risk

In is important to recognize the difference between wildfire hazard and wildfire risk, since these terms are often used interchangeably in the literature. Wildfire hazard characterizes the potential for wildfire to harm human life and safety or damage highly valued resources and assets (HVRAs) (Keane et al., 2010). Wildfire risk, by contrast, includes quantification of the magnitude of fire outcomes (beneficial and detrimental) as they relate to fire hazard (Finney, 2005). From this perspective, mapping fire hazard can reveal patterns of one component of risk, but offers less complete information to decision-makers faced with

A variety of approaches have been adopted to characterize wildfire hazard. Typically hazard is described in relation to factors affecting the fire environment and likely fire behavior, including fuel and vegetation properties, topography, climate and weather variables, and ignition characteristics (Hessburg et al., 2007; Vadrevu et al., 2010). Conceptually, probabilistic, spatially-explicit models of wildfire hazard are most relevant for risk assessment. For instance, hazard can be described with a probability distribution for a given fire characteristic at a given location, such as fire occurrence or behavior. Fire occurrence likelihood is often estimated using logistic regression models (Brillinger et al., 2009; Finney et al., 2011a; Martínez et al., 2009; Prasad et al., 2007; Priesler & Westerling 2007). Some approaches have considered likelihood of wildfire occurrence as a separate component, and characterized hazard instead as the potential to cause harm given a wildfire occurs (i.e., hazard is measure of conditional fire behavior). Here we include wildfire likelihood in our definition of hazard, which incorporates not only the likelihood of ignition for any particular area on the landscape but also the likelihood of burning due to fire spread

Modeling fire behavior given fire occurrence typically entails estimating spread rate, flame length, fireline intensity, and crown fire activity, and involves the integration of multiple sub-models (Ager et al., 2011; Cruz & Alexander 2010). Modeling fire spread allows the computation of fire travel pathways and fire size distributions, and a robust characterization of the spatial process. Simulating fire growth across heterogeneous landscapes can identify emergent behavioral properties that may not be predictable and may not be captured with localized estimates of fire behavior (Carmel et al., 2009; Parisien et al., 2007). Modeling fire

questions of how to understand and mitigate potential impacts to HVRAs.

management principles.

**2.1 Wildfire hazard** 

from remote ignitions.

**2. Wildfire hazard and risk assessment** 

the consequences of an interaction with fire, for instance the use of fire-resistant materials in home construction.

The challenge of wildfire management is to find efficient combinations of investments in mitigation options, recognizing heterogeneity in the environmental and socioeconomic factors contributing to wildfire risk. Assessing wildfire risk and evaluating mitigation options are highly complex tasks that integrate multiple interacting components including fire simulation modeling, mapping valued resources and assets, characterizing first- and second-order fire effects, quantifying social and managerial preferences and priorities, and exploring feasible management opportunities. Wildfire risk analysis is therefore fundamentally interdisciplinary, requiring the pairing of substantive expertise (fire behavior modeling, silviculture, fire ecology, etc.) with methodological expertise (statistics, engineering, decision analysis, etc.). Improved assessment of wildfire risk in turn ideally leads to improved strategic risk management across planning scales, and ultimately to enhanced resource protection and ecosystem resiliency.

Fig. 1. Conceptual Model of Wildfire Management (Modified from Calkin et al., 2011).

The major drivers of fire extent and intensity are represented as ovals, and the major strategic options for mitigating risk are represented as rectangles.

In this chapter we review the state of wildfire hazard and risk analysis, in particular highlighting a risk assessment framework that is geospatial, quantitative, and considers multiple social and ecological values. Contextually our focus is federal wildfire management in the United States, although the framework we present has broader applicability across geographic locations and ownerships. First we review concepts of hazard and risk in the wildfire management context. Second, we describe newer developments in the application of burn probability modeling for exposure analysis, and illustrate how this modeling approach can inform fuel management and wildfire suppression efforts. Third, we discuss challenges in quantifying risk for the array of non-market values that are the primary management concern on federal lands, and how expert judgment can be used to advance wildfire effects analysis. We use examples from recent and ongoing broad scale risk assessments and describe their use for informing strategic policy. Lastly we conclude by discussing potential benefits to wildfire management and policy from embracing risk management principles.

### **2. Wildfire hazard and risk assessment**

In is important to recognize the difference between wildfire hazard and wildfire risk, since these terms are often used interchangeably in the literature. Wildfire hazard characterizes the potential for wildfire to harm human life and safety or damage highly valued resources and assets (HVRAs) (Keane et al., 2010). Wildfire risk, by contrast, includes quantification of the magnitude of fire outcomes (beneficial and detrimental) as they relate to fire hazard (Finney, 2005). From this perspective, mapping fire hazard can reveal patterns of one component of risk, but offers less complete information to decision-makers faced with questions of how to understand and mitigate potential impacts to HVRAs.

#### **2.1 Wildfire hazard**

100 Novel Approaches and Their Applications in Risk Assessment

the consequences of an interaction with fire, for instance the use of fire-resistant materials in

The challenge of wildfire management is to find efficient combinations of investments in mitigation options, recognizing heterogeneity in the environmental and socioeconomic factors contributing to wildfire risk. Assessing wildfire risk and evaluating mitigation options are highly complex tasks that integrate multiple interacting components including fire simulation modeling, mapping valued resources and assets, characterizing first- and second-order fire effects, quantifying social and managerial preferences and priorities, and exploring feasible management opportunities. Wildfire risk analysis is therefore fundamentally interdisciplinary, requiring the pairing of substantive expertise (fire behavior modeling, silviculture, fire ecology, etc.) with methodological expertise (statistics, engineering, decision analysis, etc.). Improved assessment of wildfire risk in turn ideally leads to improved strategic risk management across planning scales, and ultimately to

Fig. 1. Conceptual Model of Wildfire Management (Modified from Calkin et al., 2011).

strategic options for mitigating risk are represented as rectangles.

The major drivers of fire extent and intensity are represented as ovals, and the major

In this chapter we review the state of wildfire hazard and risk analysis, in particular highlighting a risk assessment framework that is geospatial, quantitative, and considers multiple social and ecological values. Contextually our focus is federal wildfire management in the United States, although the framework we present has broader applicability across

enhanced resource protection and ecosystem resiliency.

home construction.

A variety of approaches have been adopted to characterize wildfire hazard. Typically hazard is described in relation to factors affecting the fire environment and likely fire behavior, including fuel and vegetation properties, topography, climate and weather variables, and ignition characteristics (Hessburg et al., 2007; Vadrevu et al., 2010). Conceptually, probabilistic, spatially-explicit models of wildfire hazard are most relevant for risk assessment. For instance, hazard can be described with a probability distribution for a given fire characteristic at a given location, such as fire occurrence or behavior. Fire occurrence likelihood is often estimated using logistic regression models (Brillinger et al., 2009; Finney et al., 2011a; Martínez et al., 2009; Prasad et al., 2007; Priesler & Westerling 2007). Some approaches have considered likelihood of wildfire occurrence as a separate component, and characterized hazard instead as the potential to cause harm given a wildfire occurs (i.e., hazard is measure of conditional fire behavior). Here we include wildfire likelihood in our definition of hazard, which incorporates not only the likelihood of ignition for any particular area on the landscape but also the likelihood of burning due to fire spread from remote ignitions.

Modeling fire behavior given fire occurrence typically entails estimating spread rate, flame length, fireline intensity, and crown fire activity, and involves the integration of multiple sub-models (Ager et al., 2011; Cruz & Alexander 2010). Modeling fire spread allows the computation of fire travel pathways and fire size distributions, and a robust characterization of the spatial process. Simulating fire growth across heterogeneous landscapes can identify emergent behavioral properties that may not be predictable and may not be captured with localized estimates of fire behavior (Carmel et al., 2009; Parisien et al., 2007). Modeling fire

The Science and Opportunity of Wildfire Risk Assessment 103

resources from fire (Finney, 2005). In the above formulation, the components required to estimate wildfire risk are wildfire hazard maps generated from wildfire simulation models, HVRA maps, and characterization of fire effects to HVRAs. Exposure analysis intersects mapped HVRAs with spatially-explicit measures of wildfire hazard (burn probability and conditional fire intensity). Effects analysis quantitatively defines the response of the HVRA to wildfire hazard, in this case using response functions. Collectively exposure and effects analysis characterize risk to the HVRAs in question, which can be analyzed separately or aggregated using valuation techniques and/or multi-criteria decision analysis (Thompson &

Fig. 2. Burn probability (a) and conditional flame length (b) for National Forests in the states of Oregon and Washington, in the Pacific Northwest of the United States. Figure from (Ager

Figure 3 presents our conceptual model for assessing wildfire risk combining exposure and effects analysis. Here an integrated assessment is illustrated, using a representative set of HVRAs (air quality, wildlife habitat, municipal watersheds, and human communities) for which federal agencies manage. Equation 2 presents the mathematical formulation for calculating risk (Finney, 2005), where E(NVCi) is the expected net value change to resource j, and RFi and is a "response function" for resource j as a function of fire intensity i and a

E(NVCj) = ΣBPiRFj(i,Xj) (2)

vector of geospatial variables Xi that influence fire effects to resource j.

Calkin, 2011).

et al., submitted).

spread also allows for estimates of fireline intensity as a function of fire spread direction (flanking, heading, or backing).

Rapid advancements in geospatial data management, fire behavior modeling, and computing power have vastly improved the spatial assessment of fire impacts on HVRAs. In particular, estimation of burn probability (BP), an estimate of the likelihood of a point burning under a predefined set of assumptions about ignition and fire behavior, is now feasible for large landscapes. Explicit consideration of fire spread from remote ignitions is particularly important in parts of the western United States, where large lightning-caused fires typically spread over large distances. In other locations and in different planning environments ignition likelihood may be much more of a driver.

Simulation modeling can further produce burn probabilities for fire intensity (BPi) as a function of the number of times a pixel burned at a given intensity level. The intensity with which a fire burns is an important variable for predicting fire effects. Fire intensity (KW/m) is typically converted to flame length to measure fire effects. Fire intensity is relative to the spread direction, and thus quantifying intensity for a particular point needs to consider all possible arrival directions and their probabilities. The conditional flame length (CFL), or the probability weighted flame length given a fire occurs (Scott, 2006; Equation 1) is used for this purpose, and is a statistical expectation, summing over burn intensity probabilities multiplied by the midpoints of the corresponding flame length category (Ager et al., 2010).

$$\text{CFL} = \Sigma \text{BP}\_i \text{FL}\_i \tag{1}$$

Figure 2 displays burn probability maps (a) and conditional flame lengths (b) for National Forests in the states of Oregon and Washington, in the Pacific Northwest of the United States. These estimates were derived from the large fire simulation model FSim (Finney et al., 2011a). Maps of BP and CFL differentiate regions and forests with higher relative wildfire hazard, for instance the eastern-most National Forests. Hazard is lower in the western portion of the region, where forests are generally moister and where annual rainfall is much higher.

#### **2.2 Wildfire risk**

A widely accepted ecological risk assessment framework was developed by the U.S. Environmental Protection Agency that entails four primary steps: (1) problem formulation, (2) exposure analysis, (3) effects analysis, and (4) risk characterization (U.S. Environmental Protection Agency, 1998). The two primary analytical components are exposure analysis, which explores the predicted scale and spatiotemporal relationships of causative risk factors, and effects analysis, which explores the response of HVRAs to varying levels of the risk factors (Fairbrother & Turnley, 2005). Risk characterization integrates information from exposure analysis and effects analysis to formulate a conclusion about risk. The ability to characterize risk in a common metric facilitates the integration of multiple HVRAs and allows for economic analysis of management alternatives on the basis of cost-effectiveness, although challenges exist especially for non-market resources (Chuvieco et al., 2010; Venn & Calkin, 2011).

Assessing wildfire risk requires an understanding of the likelihood of wildfire interacting with valued resources, and the magnitude of potential beneficial and negative effects to

spread also allows for estimates of fireline intensity as a function of fire spread direction

Rapid advancements in geospatial data management, fire behavior modeling, and computing power have vastly improved the spatial assessment of fire impacts on HVRAs. In particular, estimation of burn probability (BP), an estimate of the likelihood of a point burning under a predefined set of assumptions about ignition and fire behavior, is now feasible for large landscapes. Explicit consideration of fire spread from remote ignitions is particularly important in parts of the western United States, where large lightning-caused fires typically spread over large distances. In other locations and in different planning

Simulation modeling can further produce burn probabilities for fire intensity (BPi) as a function of the number of times a pixel burned at a given intensity level. The intensity with which a fire burns is an important variable for predicting fire effects. Fire intensity (KW/m) is typically converted to flame length to measure fire effects. Fire intensity is relative to the spread direction, and thus quantifying intensity for a particular point needs to consider all possible arrival directions and their probabilities. The conditional flame length (CFL), or the probability weighted flame length given a fire occurs (Scott, 2006; Equation 1) is used for this purpose, and is a statistical expectation, summing over burn intensity probabilities multiplied by the midpoints of the corresponding flame length category (Ager et al., 2010).

Figure 2 displays burn probability maps (a) and conditional flame lengths (b) for National Forests in the states of Oregon and Washington, in the Pacific Northwest of the United States. These estimates were derived from the large fire simulation model FSim (Finney et al., 2011a). Maps of BP and CFL differentiate regions and forests with higher relative wildfire hazard, for instance the eastern-most National Forests. Hazard is lower in the western portion of the region, where forests are generally moister and where annual rainfall

A widely accepted ecological risk assessment framework was developed by the U.S. Environmental Protection Agency that entails four primary steps: (1) problem formulation, (2) exposure analysis, (3) effects analysis, and (4) risk characterization (U.S. Environmental Protection Agency, 1998). The two primary analytical components are exposure analysis, which explores the predicted scale and spatiotemporal relationships of causative risk factors, and effects analysis, which explores the response of HVRAs to varying levels of the risk factors (Fairbrother & Turnley, 2005). Risk characterization integrates information from exposure analysis and effects analysis to formulate a conclusion about risk. The ability to characterize risk in a common metric facilitates the integration of multiple HVRAs and allows for economic analysis of management alternatives on the basis of cost-effectiveness, although challenges exist especially for non-market resources (Chuvieco et al., 2010; Venn &

Assessing wildfire risk requires an understanding of the likelihood of wildfire interacting with valued resources, and the magnitude of potential beneficial and negative effects to

CFL = ΣBPiFLi (1)

environments ignition likelihood may be much more of a driver.

(flanking, heading, or backing).

is much higher.

**2.2 Wildfire risk** 

Calkin, 2011).

resources from fire (Finney, 2005). In the above formulation, the components required to estimate wildfire risk are wildfire hazard maps generated from wildfire simulation models, HVRA maps, and characterization of fire effects to HVRAs. Exposure analysis intersects mapped HVRAs with spatially-explicit measures of wildfire hazard (burn probability and conditional fire intensity). Effects analysis quantitatively defines the response of the HVRA to wildfire hazard, in this case using response functions. Collectively exposure and effects analysis characterize risk to the HVRAs in question, which can be analyzed separately or aggregated using valuation techniques and/or multi-criteria decision analysis (Thompson & Calkin, 2011).

Fig. 2. Burn probability (a) and conditional flame length (b) for National Forests in the states of Oregon and Washington, in the Pacific Northwest of the United States. Figure from (Ager et al., submitted).

Figure 3 presents our conceptual model for assessing wildfire risk combining exposure and effects analysis. Here an integrated assessment is illustrated, using a representative set of HVRAs (air quality, wildlife habitat, municipal watersheds, and human communities) for which federal agencies manage. Equation 2 presents the mathematical formulation for calculating risk (Finney, 2005), where E(NVCi) is the expected net value change to resource j, and RFi and is a "response function" for resource j as a function of fire intensity i and a vector of geospatial variables Xi that influence fire effects to resource j.

$$\mathbb{E}(\text{NVC}\_{i}) = \Sigma \text{BP}\_{i} \text{RF}\_{j}(\text{i}, \text{\textdegree{X}}\_{i}) \tag{2}$$

The Science and Opportunity of Wildfire Risk Assessment 105

reliance on local knowledge by resource managers is a common substitute for formal effects

The design and functionality of simulation-based approaches span a range of intended applications, from modeling a specific fire event given an ignition to projecting wildfire likelihood and intensity at landscape scales across multiple fire seasons. Advances in burn probability modeling have enabled increasing sophistication and analytical rigor across a variety of wildfire management applications. Researchers and practitioners are able to, for instance, project near-term fire behavior using real-time weather information (Andrews et al., 2007) or to project wildfire behavioral changes in response to fuel treatments (Kim et al., 2009). In this section we focus on application of burn probability modeling and exposure analysis to support management of wildfire incidents and to support proactive hazardous

Development of suppression strategies for escaped wildland fires is subject to considerable uncertainty and complexity. Factors to balance include likely weather and fire behavior, topography, firefighter safety, and the availability and productivity of firefighting resources (ground crews, fire engines, air tankers, etc.). Of particular importance is the ability to project where and under what conditions fire is likely to interact with HVRAs. This information can help fire managers decide where aggressive fire suppression may be effectives to protect HVRAs, and where fires may have a positive impact in fire-prone

In the United States, all wildfires occurring on federal lands are cataloged within the Wildland Fire Decision Support System (WFDSS). WFDSS provides decision documentation and analysis functionality to describe the fire incident, create objectives and requirements, develop a course of action, validate key dependencies and evaluate risks (Noonan-Wright et al., 2011). The system combines a suite of fire behavior predictions with identification and quantification of values at risk to inform incident management considering safety,

The two primary risk-based analytical components within WFDSS are the Fire Spread Probability model (FSPro) and the Rapid Assessment of Values at Risk (RAVAR). FSPro calculates the probability of fire spread from a current fire perimeter or ignition point, for a specified time period. Burn probability maps are derived from simulating fire growth for thousands of statistically generated weather scenarios (Finney et al., 2011b). As implemented in WFDSS burn probabilities are mapped as probability zones, or contours, of similar burn probability; exterior contours have lower probability of fire occurrence than

Figure 4 displays an FSPro analysis for the SQF Canyon Fire, a human-caused fire that ignited on September 20, 2010 in California in the Sequoia National Forest. The figure provides a 7-day projection of fire growth as of September 14, 2010. The fire spread probability contours, moving outward from the red center, correspond to intervals of >80%,

**3. Applications of burn probability modeling & exposure analysis** 

analyses.

fuels reduction treatments.

**3.1 Incident management** 

complexity, economics, and risk (Calkin et al., 2011).

ecosystems.

interior contours.

This framework quantifies risk in terms of relative net value change (NVC), or the percentage change in initial value resulting from interaction with fire. That is, response functions address relative rather than absolute change in resource or asset value. Response functions translate fire effects into NVC to the described HVRA. In response functions illustrated in Figure 3 NVC is based on fire intensity, which is a robust fire characteristic that integrates fuel consumption and spread rate, and is often used to estimate fire effects (Thompson et al. 2011a; Ager et al. 2007). In Figure 3 the response function varies according to categorical fire intensity levels, although the framework is perfectly amenable to definition of multivariate response functions incorporating additional geospatial information that influence response to fire (see Section 4).

Fig. 3. Conceptual model for calculating wildfire risk (Modified from Calkin et al., 2010).

Characterizing fire effects has presented a major challenge to risk assessment, due to limited understanding of the spatiotemporal dynamics of ecological changes wrought by wildfire. Many past analyses focused on wildfire risk to commercial values, such as commercial timber (Konoshima et al. 2008), with a much more limited set focusing on broader nonmarket resource values and public infrastructure (Venn & Calkin 2011). There exist a variety of models that can estimate first-order fire effects such as tree mortality, soil heating, fuel consumption, and smoke production, although managers are generally more concerned with second- and third-order effects such as air quality, water quality, and habitat degradation (Reinhardt & Dickinson 2010). The management context, availability of appropriate models, and quality of spatial data will inform selection of the appropriate fire effects modeling approach (Reinhardt et al. 2001). In the absence of fire effects models, a reliance on local knowledge by resource managers is a common substitute for formal effects analyses.
