Forest Ecology and Management

Transcription

1 Forest Ecology and Management 258 (2009) Contents lists available at ScienceDirect Forest Ecology and Management journal homepage: Modeling fire susceptibility in west central Alberta, Canada Jennifer L. Beverly *, Emily P.K. Herd, J.C. Ross Conner Canadian Forest Service, Northern Forestry Centre, Street, Edmonton AB T6H 3S5, Canada ARTICLE INFO ABSTRACT Article history: Received 5 February 2009 Received in revised form 29 June 2009 Accepted 30 June 2009 Keywords: Fuels treatment Fuel management Fire behavior Landscape simulation Burn-P3 Prometheus Fire growth model Forest management Fire risk Spatial scale Strategic modification of forest vegetation has become increasingly popular as one of the few preemptive activities that land managers can undertake to reduce the likelihood that an area will be burned by a wildfire. Directed use of prescribed fire or harvest planning can lead to changes in the type and arrangement of forest vegetation across the landscape that, in turn, may reduce fire susceptibility across large areas. While among the few variables that fire managers can influence, fuel conditions are only one of many factors that determine fire susceptibility. Variations in weather and topography, in combination with fuels, determine which areas are more likely to burn under a given fire regime. An understanding of these combined factors is necessary to identify high fire susceptibility areas for prioritizing and evaluating strategic fuel management activities, as well as informing other fire management activities, such as community protection planning and strategic level allocation of fire suppression resources across a management area. We used repeated fire growth simulations, automated in the Burn-P3 landscape-fire simulation model, to assess spatial variations in fire susceptibility across a 2.4 million ha study area in the province of Alberta, Canada. The results were used to develop a Fire Susceptibility Index (FSI). Multivariate statistical analyses were used to identify the key factors that determine variation in FSI across the study area and to describe the spatial scale at which these variables influence fire susceptibility at a given location. A fuel management scenario was used to assess the impact of prescribed fire treatments on FSI. Results indicated that modeled fire susceptibility was strongly influenced by fuel composition, fuel arrangement, and topography. The likelihood of high or extreme FSI values at a given location was strongly associated with the percent of conifer forest within a 2-km radius, and with elevation and ignition patterns within a 5-km radius. Results indicated that prescribed fire treatments can be effective at reducing forest fire susceptibility in community protection zones and that simulation modeling is an effective means of evaluating spatial variation in landscape fire susceptibility. Crown Copyright ß 2009 Published by Elsevier B.V. All rights reserved. 1. Introduction The composition and configuration of landscape vegetation types in northern forest ecosystems can influence wildfire susceptibility. Stands characterized by different species and age classes will have different fuel complexes and structures that will either support or restrict the spread of fire from adjacent stands. Homogenous fuel, weather, and topographic conditions across an area are generally associated with larger, more uniform fires because continuous areas become simultaneously receptive to fire ignition and spread (Turner and Romme, 1994; Delong, 1998; Ryan, 2002). In the Boreal and Foothills natural regions of Alberta, large but infrequent high-intensity fires are characteristic (Tymstra et al., 2005). These fires kill the majority of the trees and over long periods and large spatial scales, produce a diverse mix of landscape patches that differ by age and cover type. Dominant tree * Corresponding author. Tel.: address: jbeverly@nrcan.gc.ca (J.L. Beverly). species in these ecosystems can be divided into two groups the conifer fuels that fires select for, such as white spruce (Picea glauca), black spruce (Picea mariana), and lodgepole pine (Pinus contorta) and the deciduous fuels, such as aspen (Populus tremuloides), that limit or arrest fire ignition and spread (Cumming, 2001; Krawchuk et al., 2006). Because the fires characteristic of these areas can affect large areas over short periods, they can pose a sudden and significant threat to forest resources, forest-based communities, and the industries and economic activity associated with both. Fire suppression has reduced area burned in northern forest ecosystems (Cumming, 2005; DeWilde and Chapin, 2006; Martell and Sun, 2008) and there are concerns that effective suppression over long time periods in these ecosystems may potentially contribute to increasing fuel continuity that could result in some fires reaching sizes that would otherwise have been constrained by the landscape mosaic. The mechanisms of this process are summarized by Chapin et al. (2008) for black spruce forests in boreal Alaska, where the moisture content of deciduous early post-fire successional species are thought to create fuel breaks between adjacent /$ see front matter. Crown Copyright ß 2009 Published by Elsevier B.V. All rights reserved. doi: /j.foreco

2 1466 J.L. Beverly et al. / Forest Ecology and Management 258 (2009) black spruce stands, thus reducing landscape fire susceptibility, or the likelihood that an area will be burned by a fire. Landscape management initiatives such as FireSmart (Hirsch et al., 2001) have been proposed for northern forest ecosystems and aim to reduce fire susceptibility through the spatial allocation of activities such as prescribed burning, harvest cut-blocks and roads. These activities create breaks in continuous flammable areas with non-fuel barriers or the introduction of less flammable vegetation types or ages into the landscape mosaic. The goal of this type of landscape fuel management is to strategically reduce the likelihood of large and costly fires in or near high-value areas, such as communities. While vegetation type and configuration are among the few fire determinants that can be influenced by land managers, many other factors will influence the susceptibility of an area to wildfire. To be effective, strategic fuel management planning must account for the combined variations in fuel, weather and topography that together determine which areas of the landscape are more likely to burn under a given fire regime. Integrated mapping of fire regime and fire environment factors has applications to many fire management activities and has been an important area of fire research for some time. Geographic information systems (GIS) have been used to rate fire susceptibility at individual locations on the basis of spatial data for biophysical, fire regime, and other characteristics, sometimes in combination with fire behavior models (e.g., Burgan and Shasby, 1984; Chuvieco and Congalton, 1989; Hawkes and Beck, 1997; Núñez-Regueira et al., 2000; Haight et al., 2004). Statistical models have also been used to map spatial variations in fire susceptibility as a function of variables such as vegetation, weather, topography, and human activity on the basis of historical data for locations where fires occurred (e.g., Cardille et al., 2001; Dickson et al., 2006; Syphard et al., 2008). In regions where fires can have dramatic and sustained impacts on vegetation composition and structure, explanatory variables related to forest fuels must pre-date the historical fire activity under consideration. The resulting statistical model can then be applied to current conditions or projected future conditions to map fire susceptibility. Fire growth simulations can be used to assess fire potential under current landscape conditions. Simulation models are a particularly powerful tool for assessing fire susceptibility because they incorporate the influence of surrounding areas on the potential for a particular location to be burned by a fire. Spatial context is captured when a computer fire growth model is used to simulate the ignition and spread of fires. One approach involves simulating many fires over large landscapes to identify general patterns in fire prone areas (e.g., Mbow et al., 2004). Impacts of proposed fuel management activities, such as prescribed fires or harvest treatments, can be evaluated by altering fuels and conducting subsequent simulations for comparison (e.g., LaCroix et al., 2006; Duguy et al., 2007; Schmidt et al., 2008; Suffling et al., 2008). Simulation modeling has also been used to explore fundamental interactions between fire activity and landscape structural features by comparing simulated fires on hypothetical landscapes that are created to represent specific conditions (e.g., LaCroix et al., 2008). When fire ignitions and burned areas are simulated repeatedly over a large number of model runs or iterations, the frequency of times a location burns can provide a spatially continuous rating of fire susceptibility (e.g., Farris et al., 2000; Carmel et al., 2009). We used repeated fire growth simulations, automated in the Burn-P3 landscape-fire simulation model (Parisien et al., 2005), to assess spatial variation in fire susceptibility across a 2.4 million ha study area in west central Alberta, Canada. We combined simulation results with multivariate statistical analyses to identify the key factors that determine variation in fire susceptibility across the study area and to describe the spatial scale at which these variables influence fire susceptibility at a given location. Finally, we assess the impact of proposed fuel management activities (prescribed fire treatments) on fire susceptibility in our study area. The objectives of our study were threefold: to explore the usefulness of simulation modeling for evaluating the effectiveness of fuel management activities; to provide insight into those factors that determine variation in forest fire susceptibility in northern forest ecosystems; and to characterize the spatial context captured in this type of fire simulation modeling exercise, where the characteristics of surrounding areas influence fire susceptibility at any given point. 2. Methods 2.1. Study area Bounded to the north by Lesser Slave Lake, the 2.4 million ha study area falls within the Boreal and Foothills natural regions in the province of Alberta, Canada (Natural Regions Committee, 2006) (Fig. 1). These regions are characterized by short, wet summers with peak precipitation occurring in July. Mean annual precipitation ranges from 461 mm to 588 mm and mean daily temperatures between May and September are C. Dominant tree species include aspen (P. tremuloides Michx.), lodgepole pine (P. contorta Dougl. ex Loud. var. latifolia Engelm.), white spruce (P. glauca (Moench) Voss), and balsam poplar (Populus balsamifera L.). There is considerable variation in vegetation patterns across the study area with mixedwood and deciduous stands dominating the Lower Foothills subregion and conifer dominating the Upper Foothills Subregion. The Dry Mixedwood Subregion is characterized by aspen-dominated forests. Elevations range from 350 m to 1350 m ASL. The average slope in the study area is less than 10%, with just a few areas in the Upper Foothills Subregion having slopes between 20% and 50%. Between 1976 and 2003 annual fire occurrence in the study area varied substantially from year to year, with an average of 109 fires each year. Fire sizes during this period ranged from less than 0.1 ha to 168,863 ha. Small fires (<4 ha) accounted for 93% of fire events, and only 1% of fires exceeded 200 ha. Only five fires reached a final size greater than 5000 ha, but these fires each burned an average of 50,000 ha, or 2% of the burnable area in the study area. Significant large fires in the study area are shown in Fig. 2, the largest being the 1998 Virginia Hills fire, which burned 167,000 ha. Annual area burned also varied substantially between 1976 and 2003, from 49 ha to 229,995 ha. The majority (93%) of the total area affected by fire during this period was burned during 3 years (1981, 1998, and 2001). On average, 11,191 ha of the study area burned each year; this represents 0.48% of the burnable area contained within the study area, very close to the estimate of 0.42% reported by Stocks et al. (2002) for the Boreal Plains ecozone, which contains our study area. Ninety-seven percent of area burned occurred between late April and early September. About half the fires that occur in the study area are from lightning ignitions and half are from human sources. The majority (77%) of lightning-caused fires occurred between June 1 and August 15. Human-caused fires begin in early April and peak at the beginning of May, then decline through to October Simulation model We assess fire susceptibility for the current state of our study area by repeatedly simulating the ignition and spread of fires. We used Burn-P3 (Parisien et al., 2005) to automate the simulations and to incorporate fire regime characteristics associated with our study area. Burn-P3 is a fire-landscape simulation model that can

3 J.L. Beverly et al. / Forest Ecology and Management 258 (2009) Fig. 1. Location of the study area in west central Alberta, Canada. be used to identify the likelihood that a unit of landscape (landscape pixel) will be burned by fire. The model uses a fire growth submodel, Prometheus (Tymstra et al., 2009), to simulate the growth of many individual fires. Prometheus is a deterministic model that uses spatial input data describing topography (slope, aspect, and elevation) and landscape cover type to predict fire spread as a function of fire weather. Wave propagation algorithms are used to model fire growth using equations from the Canadian Fire Behavior Prediction (FBP) System (Forestry Canada Fire Danger Group, 1992) and fire weather components of the Canadian Fire Weather Index (FWI) System (Van Wagner, 1987). The 16 standard fuel types of the FBP system cover most major boreal forest fuel types in Canada and are well suited for representing the vegetation in our study area Modeling process During each Burn-P3 iteration, fire ignition and growth was simulated until a specified number of fires of a predetermined minimum size has been completed. Fires that did not achieve the minimum size were discarded. At the end of each iteration, the burned or unburned state of each location in the landscape was recorded and the process was then repeated for a total of 25,000 iterations. Fire susceptibility was assessed as the number of times a location burned in relation to the number of burning opportunities (total iterations). Only the burned or unburned state of a location was considered, and no attempt was made to account for other fire characteristics, such as the severity of the burn or any resulting impacts. The complete modeling process is illustrated in Fig. 3. A total of 125,000 fires were included in the completed simulation. Model parameters, including the number of fires simulated per run and the minimum fire size, were determined from a combination of sources including calibration and random draws from distributions determined from historical data (see below). A range of inputs were used to incorporate information about fire environment characteristics (fuel, weather, and topography) and fire regime characteristics, such as fire occurrence patterns (see below) Model parameters and settings Landscape features Landscape cover information for the study area was obtained in a raster format with a 25 m 25 m cell size from Alberta Sustainable Resource Development. The data consisted of a vegetation classification according to the 16 standard fuel types of the Canadian Fire Behavior Prediction (FBP) System (Forestry Canada Fire Danger Group, 1992) which represent the major forest cover types in Canada (Fig. 4). Non-fuel areas such as water were also classified. Most of the fuel-type data were derived from the 1:20,000 scale Alberta Vegetation Inventory (AVI), which is in turn derived from aerial photography. The imagery used to generate the AVI dates as far back as 1987, and despite ongoing efforts to verify the information with field sampling, some areas may not be accurately represented. AVI data were converted to FBP fuel types by the provincial government using an automated conversion program that also incorporates expert knowledge. These data are subjected to ongoing field verification and updating. For areas where no AVI data are available, the Alberta Ground Cover Classification, a classification system based on Landsat 7 imagery, is used to generate FBP fuel types. To reduce computation time during fire growth simulations, fuel-type data were resampled to a 100 m 100 m resolution using the majority option in ArcGIS (Environmental Systems Research Institute, 2005). To minimise

4 1468 J.L. Beverly et al. / Forest Ecology and Management 258 (2009) Fig. 2. Areas burned by forest fires that occurred within the study area the impacts of study-area boundaries on fire growth modeled in Burn-P3, we added a 10-km buffer around the study area to each spatial data layer. These layers are shown in Fig Fire processes Several model inputs are used to ensure fire processes in Burn- P3 are modeled according to known fire regime conditions, such as fire season and fire size. These inputs were determined by observing conditions over a historical baseline period. The baseline period was defined to identify the range of variability in weather and fire activity typical of the study area. Accuracy of model results depends on the degree to which the baseline period adequately represents conditions expected over the next fire season. We chose the period , which we considered long enough to capture relatively infrequent events, while still broadly consistent with current regionally specific conditions capable of influencing variability in fire regime characteristics (i.e., climate, land use, vegetation, and fire management). Because fire weather and associated fire activity can exhibit seasonal patterns, Burn-P3 can accept season-specific inputs. We used the historical frequency of both human-caused and lightningcaused ignitions to identify seasonal fire activity patterns. The spring and summer seasons were defined as April 20 May 20 and May 21 September 5, respectively. For efficiency, the Burn-P3 model is designed to simulate only those fires that have significant impacts on the landscape. These are termed escaped fires and are defined using a minimum fire size input that is considered large relative to the population of fires that can be expected to occur in the area. We set the minimum fire size to 10 ha. Between 1976 and 2003, only 5% of the fires in our study area exceeded this minimum size. The location of a fire ignition is determined by several factors, including ignition grids that provide a coarse relative rating of fire occurrence patterns. To assess these patterns, we used ArcGIS (Environmental Systems Research Institute, 2005) to calculate the point densities of fires during the fire season (April 20 September 5) between 1976 and 2003 for each combination of fire cause (human and lightning) and season (spring and summer). Fires within the study area and a 10-km buffer around the study area were included in the analysis. The resulting densities were smoothed at a 10-km radius to identify coarse ignition patterns, which were classified into an ordinal scale (1 or 2 for lightningcaused fires; 1, 2, or 3 for human-caused fires) to indicate spatial variation. The location of fires is also influenced by escaped-fire rates, which are calculated as the percent of all escaped fires ignited by a given cause during a given season (Table 1) and are used to adjust the fire occurrence ignition grids (see Parisien et al., 2005). Once an ignition has been located, fire growth is simulated with Prometheus. Forest fires achieve their final size over the course of a single day or over several days, and it is not uncommon for large fires to burn over a period of several weeks, during which active fire growth is limited to a smaller number of days. Representative estimates of fire size can be obtained by simulating fire growth during the relatively small number of days when fires achieve the majority of their spread. In Burn-P3, days characterized by significant fire growth are referred to as spread-event days. Large forest fires can be expected to achieve significant fire growth during peak burning conditions, which occur in the late afternoon or early evening. By modeling daily fire growth during the 1 3 h when these peak conditions occur, Burn-P3 attempts to simulate the most significant fire behavior. This approach is computation-

5 J.L. Beverly et al. / Forest Ecology and Management 258 (2009) Fig. 3. Burn-P3 modeling process.

6 1470 J.L. Beverly et al. / Forest Ecology and Management 258 (2009) Fig. 4. Composition of the study area by Canadian Forest Fire Behavior Prediction (FBP) System fuel types. ally efficient and responds to fire growth modeling limitations that make it impracticable to model the entire lifetime of a fire; however, it necessitates user-defined inputs on the duration of fire growth to ensure that simulated fire sizes are realistic. The total duration of fire growth for each fire consists of the number of spread-event days multiplied by the daily burning period (hours). These inputs are determined from a calibration process to ensure that Burn-P3 produces realistic fires, given the actual escaped fire size distribution of the study area. By varying the daily hours of burning and the distribution of spread-event days, we produced a series of fire susceptibility maps for the study area, each consisting of 10,000 iterations. Inputs that produced a distribution of simulated fire sizes consistent with the historical escaped fire distribution over the period were selected. The daily burning period was defined as 2 h and is always the same, for every fire and every spread-event day. The number of spreadevent days associated with each fire is input into the model as a distribution that designated 85% of fires with a single spread-event day and the remaining fires with 2 (5%), 10 (5%), or 11 (5%) spreadevent days. The number of escaped fires simulated per iteration is defined from historical data. Historical fire records were used to produce a distribution of the number of escaped fires that can be expected in a given year. In our study area, there was an average of five escaped fires per year over the period. A spatial coverage consisting of coarse fire zones is input into the model to define the boundaries between regions where fire processes are known to differ substantially, for example, between ecological regions or between major forest ecosystem types. Comparison of fire regime characteristics across our study area by natural region indicated that fire processes did not differ sufficiently to warrant the use of multiple fire zones. For example, over the period, the mean annual proportion burned was in the Boreal Natural Region and in the Foothills Natural Region. Other model inputs were used to capture known spatial variation in factors that influence fire processes, such as ignition patterns, weather zones, and escaped-fire rates defined by season and cause Weather conditions Daily weather data recorded at stations within our study area or outside but near the study area boundary were used to characterise fire weather conditions over the period. Weather station records obtained from Alberta Sustainable Resource Development contained noon observations of fire weather variables for each day during the fire season (April 20 September 5). We selected 12 primary weather stations on the basis of the breadth and completeness of their records, and classified the 21 remaining stations as secondary. On average, primary weather stations were missing data for 3.4% of daily records, which we approximated from data recorded at the nearest primary or secondary station that had a complete record for the day in question. In Burn-P3, fire weather zones are used to define the boundaries between regions where fire weather conditions vary. To define fire weather zones for the study area, we calculated 90th percentile maps for key weather station variables (temperature, wind speed, relative humidity, and precipitation) and for components of the Canadian Fire Weather Index (FWI) System (Van Wagner, 1987) using the Spatial Fire Management System (SFMS) Fire Climatology extension (Englefield et al., 2000). The SFMS uses an inverse distance weighting interpolation method found to adequately fit

7 J.L. Beverly et al. / Forest Ecology and Management 258 (2009) Fig. 5. Spatial layers used in Burn-P3 modeling: (A) landscape cover type, (B) topography, (C) fire zones, (D) weather zones, (E) ignition pattern (summer, human-caused), (F) ignition pattern (spring, lightning-caused), (G) ignition pattern (spring, human-caused), and (H) ignition pattern (summer, lightning-caused). observed low and high FWI values in northern boreal ecosystems (Flannigan and Wotton, 1989). Visual analysis of 90th percentile maps indicated a band of relatively lower fire weather values from the central north to the central southwest portion of the study area. Wind speed and Initial Spread Index (ISI) maps showed a slightly different pattern, with a concentration of high values in the central, high-elevation section of the study area. Weather zone boundaries were defined from three FWI components (Fine Fuel Moisture Code (FFMC), Duff Moisture Code (DMC), and ISI) as follows: 90th percentile values were standardised and multiplied, and the results were smoothed at a 2.5-km radius. To simulate the growth of each fire with the Prometheus fire growth submodel, Burn-P3 randomly selects a fire weather record from a list of available records defined for the specific weather zone and fire season where the ignition occurred. Each record in the fire weather list contains fire weather variables and FWI System components that represent peak daily burning conditions. If a simulated fire is assigned multiple spread-event days, a unique weather record is drawn for each day. For each spread-event day,

8 1472 J.L. Beverly et al. / Forest Ecology and Management 258 (2009) Table 1 Summary of model parameters. Parameters Input format Description Static parameters Landscape cover type ASCII grid Fire Behavior Prediction System fuel-type grid for the study area, at a m resolution. Integer values are used to identify fuel types and other landscape cover types, such as water and non-fuel areas. Topography ASCII grids Slope, aspect, and elevation grids for the study area at a 100 m 100 m resolution, entered as integer or floating point values. Ignition pattern ASCII grids Coarse spatial pattern of historical fires in the study area. Individual grids for the study area are input for each combination of fire cause and season: spring lightning, spring human, summer lightning, and summer human. Fire zone ASCII grid Grid used to define broad areas that differ in terms of fire processes. Because the study area has just one fire zone, the grid value is the same across the entire area. Weather zone ASCII grid Grid used to define areas that differ in terms of fire weather conditions. Two weather zones are represented in the study area. Fire season NA Dates defining seasonal changes in fire processes. Two fire seasons are defined for the study area: spring (April 20 to May 20) and summer (May 21 to September 5). Used primarily in the preparation of other parameters (escaped-fire rates, ignition patterns, weather lists). Minimum fire size GUI keystroke Minimum fire size (ha) used to define escaped fires, which are considered large relative to the population of fires that have occurred historically in the study area (10 ha in this study). Simulated fires that do not achieve this size are discarded. Escaped-fire rate delimited text file The percentage of escaped fires attributed to a given season and cause, according to historical data: in this study, 55.4% spring human, 4.0% spring lightning, 9.9% summer human, and 30.7% summer lightning. Used in combination with ignition grids to create an ignition table, which is then used by Burn-P3 to locate ignitions across the study area and assign them a specific cause and season. Ignition rules Delimited text file Rules used to ensure that ignition processes are realistic. For example, lightning-caused ignitions were not permitted in the aspen fuel type. Variable parameters Fire weather record Delimited text file A list of daily records of noon fire weather observations and Fire Weather Index components used to model fire growth during peak daily burning conditions. Records are consistent with high or extreme burning conditions known to be associated with large fire spread in the study area and are separated by season. A unique fire weather record is allocated to each fire spread-event day. For a given spread-event day, the weather record is repeated for all hours of burning specified during the model calibration process. Number of escaped fires Delimited text file Frequency distribution of the number of escaped fires per year; based on historical data. Used to determine the number of fires grown by Prometheus during a given Burn-P3 iteration. Varies by iteration. Calibrated parameters Hours of burning GUI keystroke In combination with the distribution of spread-event days, this parameter is calibrated to ensure that the size of simulated fires is consistent with the historical distribution of fire sizes for the study area. Determined to be 2 h for our study area. This is a static parameter that does not vary by spread-event day or fire. Spread-event days Delimited text file Frequency distribution of the number of days of significant spread associated with a fire. In combination with hours of burning, this parameter is calibrated to ensure that the size distribution of simulated fires is consistent with that for historical fires. This parameter is variable and is re-assigned each time a simulated fire is grown by the model. NA = not applicable, GUI = graphical user interface. the same weather record is used (i.e., repeated) for the hours of burning specified during the model calibration process, which in our study was 2 h. Because Burn-P3 is only concerned with modeling large, escaped fires, the weather records in the fire weather list must be suitable for modeling fires that achieve significant growth. To select a pool of weather records from the historical database, data for each station were divided into spring and summer seasons and sorted by FFMC. The top 10% of records at each station for each season were selected and compiled to produce spring and summer weather lists for each weather zone Prometheus settings Several Prometheus fire growth model settings are input. These relate to FBP System settings and calculations in Prometheus and include enabling fire acceleration from a point ignition; using the Buildup Index (BUI) effect on rate of spread; using a terrain effect in which topography inputs are used in FBP calculations; enabling seasonal fuel changes associated with the timing of leaf flush (green-up); and setting the frequency with which the fire front is calculated, every 5 min in our study Fire Susceptibility Index calculations Fire susceptibility, assessed as the number of times a location burned in relation to the number of burning opportunities (total iterations), is output by Burn-P3 at the same cell size as for landscape input data. We used a 100 m 100 m cell size as the input data to ensure accurate fire growth simulations with the Prometheus fire growth submodel. This cell size is considered a relatively fine scale for both landscape simulation modeling and the strategic planning exercises that will make practical use of the results. The scale of Burn-P3 output combined with the large number of stochastic elements in the model necessitates extremely large and computationally impractical numbers of iterations to replicate results. To produce results that could be replicated with a reasonable number of iterations, we developed a Fire Susceptibility Index (FSI) by smoothing the 100 m 100 m Burn-P3 output grid with a 500-m radius circular neighbourhood and reclassifying the resulting data using the natural breaks (Jenks) method to create five fire susceptibility classes (very low, low, moderate, high, and extreme). These five qualitative classes represent a relative rating of the likelihood that a location in our study area will be burned by a fire over the next fire season. The number of iterations required to replicate FSI values reflects the unique combination of fire regime and fire environment characteristics of the area under investigation, as well as the size of this area. To determine an acceptable number of iterations for our study area we produced two replicate FSI maps using increasing numbers of iterations and compared agreement between replicates using difference maps in ArcGIS Spatial Analyst (Environ-

9 J.L. Beverly et al. / Forest Ecology and Management 258 (2009) mental Systems Research Institute, 2005). Replicated FSI maps that were each produced with 25,000 iterations demonstrated 80% agreement, and increasing the iterations beyond 25,000 resulted in minimal improvement. Disagreement in FSI values was limited to a single class difference, with 10% of cases associated with a lower class and 10% with a higher class. We considered this degree of replication accuracy acceptable for the purposes of our analysis. Our assessment of the replication of Burn-P3 results is fundamentally different from the assessment of the stability of individual Burn-P3 maps explored by Parisien et al. (2005). They compared Burn-P3 results produced for a single map, calculating average change as progressive iterations were conducted. While a Burn-P3 map may become stable at a relatively small number of iterations (2000), our results indicate that considerably more iterations (>20,000) are necessary to produce replicate maps, even when using our approach of smoothing and classifying raw Burn- P3 results Statistical analysis To explore the impact of fire environment and fire regime variables on modeled fire susceptibility, we used logistic regression analysis. This method is commonly applied in studies that model the presence or absence of a feature, such as wildlife (see review by Keating and Cherry, 2004) or land use (e.g., Huang et al., 2007) as a function of human or biophysical variables. We applied the same approach to predict the probability of high or extreme Fire Susceptibility Index (FSI) values at a given location in our study area. FSI values were used for statistical modeling due to our lack of confidence in the raw values output by Burn-P3. We combined data on the presence or absence of high or extreme FSI values with a set of explanatory variables representing fire environment and fire regime characteristics that we expected to influence FSI values (Table 2). To explore the influence of surrounding landscape areas on FSI values, each independent variable was derived at four spatial extents using a moving window routine with a radius of 500 m, 1 km, 2 km, or 5 km. Independent variables were calculated for each 100 m 100 m pixel in the study area. Due to the presence of spatial autocorrelation in the data, we randomly selected 500 locations using SAS SQL (SAS Institute Inc., 2004). An equal number of 0 and 1 observations of the dependent variable were selected. For each independent variable, and spatial extent, we initially used Wilcoxon rank sum tests to examine differences in conditions between locations with and without high or extreme FSI values. We modeled the probability of high or extreme FSI values as follows: 1 PðheÞ ¼ 1 þ e ðb 0 þb (1) 1 x 1þ þb n x nþ where P(he) is the probability of high or extreme FSI values, x 1 n are the independent variables, and b 0 n are regression coefficients. Maximum likelihood estimates of model parameters were computed with SAS LOGISTIC (SAS Institute Inc., 2004). Model selection was based on Akaike s Information Criterion. The model s predictive ability and goodness of fit were assessed by the likelihood ratio x 2 test, the Wald x 2 test for individual parameters, and the C statistic, which measures the association between predicted probabilities and observed outcomes. To avoid modeling spurious associations, we used a two-step process to reduce our variable set. First, for each variable that exhibited marginal univariate association at a given scale (P < 0.05), we evaluated independence among scales using Spearman rank correlation. Variables that were highly correlated (rs > 0.7) across scales were deleted at those scales where the association with high or extreme FSI values was weakest. The remaining variables were grouped into three broad subsets that describe classes of fire environment and fire regime characteristics: fuels (composition, arrangement, and breaks), topography, and fire ignition. As above, we evaluated relationships among variables first within and then among groups. Redundant variables were eliminated if they showed a weaker association with high or extreme FSI values. We used a correlogram of Moran s I to test for spatial autocorrelation in Pearson residuals of the final logistic regression model. Only pairs of points separated by less than half the maximum distance observed were considered for the analysis (Rossi et al., 1992). Moran s I statistic was calculated for lag intervals of 2 km up to a distance of 130 km using the PASSAGE program (Rosenberg, 2001). Each value of Moran s I was tested for significant deviations from the expected value under the null hypothesis of random spatial distribution (Cliff and Ord, 1981), using a Bonferroni correction for multiple tests (Oden, 1984). Model validation was conducted with cross-validated predicted probabilities computed in SAS (SAS Institute Inc., 2004). The procedure involved withholding single observations from the data set, building the model without it, and then testing the observations in the model. The process was repeated until all observations had been tested, and cross-validation accuracy was estimated from the frequency of times an observation was not reclassified during the cross-validation procedure. The coefficients derived for the final logistic regression model were used to produce a predicted probability surface for the entire study area. A cut-off probability was selected to assess correspondence between predicted and actual high or extreme FSI values Prescribed fire treatment scenario We investigated the effects of prescribed fire treatments on fire susceptibility for a case study area surrounding the community of Swan Hills, Alberta. In consultation with the provincial forest fire management agency, a set of candidate prescribed fire treatment areas were defined to support community protection objectives. The candidate areas were identified on the basis of expert knowledge of fire processes and fuel types in the area and some consultation with forest industry stakeholders; areas considered to have marginal value for timber production were given priority for such treatments. Table 2 Explanatory variables analysed in the logistic regression model to explain the presence of high or extreme Fire Susceptibility Index (FSI) values at a location. For each landscape pixel, variables were derived at four spatial extents using a moving window routine with a radius of 500 m, 1 km, 2 km, or 5 km. Variable Description Conifer Percent of area classified as having a conifer landscape cover type Contiguity Mean contiguity of conifer fuels; computed in Fragstats software (McGarigal et al., 2002); values between 0 and 1, with large contiguous patches resulting in larger values Fuel break Percent of area classified as having a fuel break landscape cover type Elevation Percent of area with elevation greater than the median elevation of the study area Ignition Average Burn-P3 ignition grid value

10 1474 J.L. Beverly et al. / Forest Ecology and Management 258 (2009) Spatial layers of the proposed treatments were produced in ArcGIS, by heads-up digitising of scanned hard-copy maps generated by hand in collaborative consultations. Thirty-five treatment areas were defined, ranging in size from 16 ha to 156 ha, with a mean of 52 ha (Fig. 6). The complete set of prescribed fire treatments covered an area of 1819 ha or 4.7% of the community assessment area, which we defined as the area extending from the community outward to include the prescribed fire treatments and a 4-km buffer beyond their farthest boundaries. Over 80% of the treatment areas were characterized by the Boreal Spruce fuel type of the FBP System. Treatment areas were merged with the original fuel map of the study area and were reclassified as non-fuel on the assumption that recently burned areas would not be able to support fire growth. The Burn-P3 simulations were then repeated, with all other model inputs held constant, to assess the impact of the prescribed fire treatments on FSI values immediately surrounding the community of Swan Hills. 3. Results 3.1. Fire Susceptibility Index (FSI) map FSI values varied across the study area (Fig. 7). The majority of the study area had very low (31%), low (24%), or moderate (21%) FSI values. Approximately one-quarter of the study area was associated with high (16%) or extreme (8%) FSI values. As would be expected, high and extreme FSI values generally corresponded with contiguous areas of the Boreal Spruce fuel type, whereas very low and low FSI values generally corresponded with deciduous fuels (see fuel types, Fig. 4). High and extreme FSI values were also concentrated in areas of relatively high elevation (see topography, Fig. 5B). Areas of extreme fire susceptibility were concentrated in the central portion of the study area even though this area was modeled as a separate weather zone, with somewhat less extreme fire weather conditions than the edges of the study area (see weather zones, Fig. 5D). The locations of recent large fires (see Fig. 2), which are characterized by less flammable fuel types, corresponded with large contiguous areas of low FSI values Statistical analysis The probability of a high or extreme FSI value in the study area, as determined from logistic regression analysis, was strongly associated with the percent of surrounding areas classified as having a conifer landscape cover type. The association was strongest at the 2-km spatial extent, and conifer variables derived at the other three extents were considered redundant. There was also a strong positive association between mean contiguity of conifer fuels and high or extreme FSI values; however, at 500 m and 1-km extents, this variable was highly correlated with the proportion of conifer fuel within 2-km. Contiguity of conifer fuels at the 2-km spatial extent was not significant in the final model. FSI values were strongly associated with the percent of surrounding areas having an elevation that exceeded the median elevation for the study area as a whole, and this association was strongest for surrounding areas within a 5-km radius. There was a significant but weaker association between FSI values and the fuel break variable derived at 500 m and 5-km extents; however these variables were highly correlated with contiguity of conifer fuels derived at a 2-km extent and were considered redundant. There was also a significant but weaker association between FSI values and the ignition variable derived at a 5-km extent. The final logistic regression model was highly significant (x 2 = , 3 df, P < ), with 96% concordance between predicted probabilities and observed outcomes: 1 PðheÞ ¼ (2) 1 þ e ð 9:9306þ0:1194x 1þ0:0136x 2 þ1:5479x 3 Þ where x 1 is the percent of conifer fuel within a 2-km radius (P < ), x 2 is the percent of area within a 5-km radius with Fig. 6. Location of proposed prescribed fire treatments around the community of Swan Hills, Alberta. Fuel types of the Canadian Forest Fire Behavior Prediction System (Forestry Canada Fire Danger Group, 1992) are shown for a 4-km buffer around the treatment area.

11 J.L. Beverly et al. / Forest Ecology and Management 258 (2009) Fig. 7. Spatial variation in fire susceptibility across the study area. Fire Susceptibility Index (FSI) values represent a classification of raw Burn-P3 output using the natural breaks (Jenks) method. elevation exceeding the median elevation of the study area (P = ), and x 3 is the average ignition grid value within a 5-km radius (P = ). Correspondence between model accuracy and the accuracy computed from cross-validated predicted probabilities indicated that the model is robust. The correlogram of Moran s I indicated significant spatial autocorrelation is associated with only one of the 65 distance classes (16 18 km). This indicates that the model fails to account for the spatial dependence of some autocorrelated environmental factor. The predicted probability surface derived from the final model resulted in an accurate classification of 89.4% of the 2.4 million ha study area when a cut-off probability of 75% was used to classify each 100 m 100 m landscape pixel as having a high or extreme FSI value (Fig. 8). Of the misclassified area, 5.3% was incorrectly predicted to be a 1 and 5.3% was incorrectly predicted to be a 0. The model performed poorly around the margins of high or extreme FSI patches, which is not surprising given the process used to classify raw Burn-P3 output into FSI values Prescribed fire treatment scenario Prescribed fire treatments resulted in a marked decrease in fire susceptibility within the community assessment area, which included the area extending from the community of Swan Hills outward to include the prescribed fire treatments and a 4-km buffer beyond their farthest boundaries. Thirty-nine percent of this 38,000 ha area exhibited a decrease in FSI values following the prescribed fire treatments (Fig. 9). For every 1 ha within the community assessment area that was treated with prescribed fire, 8 ha exhibited a decrease of at least one FSI class. Decreases in fire susceptibility were concentrated in the area between the prescribed fire treatments and the community of Swan Hills. For a 1-km buffer zone around the community, 83% of the 1122-ha area exhibited a decrease of at least one FSI class, and 44% of the area exhibited a decrease of 2 FSI classes. 4. Discussion Our analysis indicates that modeled fire susceptibility in west central Alberta was strongly influenced by fuel composition, fuel arrangement, topography, and ignition patterns. We used logistic regression modeling to identify the key determinants of fire susceptibility in our study area and to characterize the spatial scale at which surrounding areas influence fire susceptibility at a given location. The presence of high or extreme FSI values at a given location was strongly associated with the percent of conifer fuels within a 2-km radius, the percent of surrounding area within 5-km that had elevation values exceeding median values for the study area as a whole, and the average ignition grid value within a 5-km radius. We suspect that slope impacts on fire spread equations contributed to the significance of the elevation variable. Site

12 1476 J.L. Beverly et al. / Forest Ecology and Management 258 (2009) Fig. 8. Validation of the logistic regression model. (A) actual FSI values derived from Burn-P3 output and (B) predicted FSI values derived from the logistic regression model corresponding to high or extreme areas (1) and very low, low, or moderate areas (0). (C) difference between actual and predicted FSI maps. Fig. 9. Fire Susceptibility Index (FSI) values around the community of Swan Hills (A) under baseline conditions and (B) following fuel changes associated with prescribed fire treatments, with (C) the difference in FSI values between baseline conditions and the prescribed fire treatment scenario within a 1-km buffer around the community (indicated by dashed line).

13 J.L. Beverly et al. / Forest Ecology and Management 258 (2009) conditions associated with elevation differences, while potentially influential on fire behavior, are not accounted for in the standard fuel types of the Fire Behavior Prediction (FBP) System used to model fire growth in our simulations. High levels of contiguity or connectedness of conifer fuels was also strongly associated with high or extreme FSI values. Fuel breaks had a weaker influence on FSI values. These results are consistent with a recent study by Carmel et al. (2009). In that study, fire susceptibility in the Mediterranean basin was mapped by simulating the growth of 500 random ignitions using FARSITE (Finney, 1998), which is the American counterpart of the Prometheus fire growth model. Spatial variation in fire susceptibility was associated with fuel patterns, microclimate, topography, land use, and the distribution of ignition locations. The results of our prescribed fire scenario were consistent with logistic regression analysis and highlighted the influence of fuel type on FSI values. Simulated fuel modification, through prescribed fire treatments, was effective at reducing forest fire susceptibility in a community protection zone in our study area. Model results provided a formal means for characterizing the efficiency of prescribed fire activities. In our case there was a 1:8 ratio between the area treated with prescribed fire and the area that experienced a subsequent reduction in fire susceptibility. Effectiveness of prescribed fire treatments for achieving fire management objectives was also demonstrated by Suffling et al. (2008) using the Prometheus fire growth model to evaluate prescribed fire alternatives recommended by experts. The spatial scale at which surrounding areas influenced fire susceptibility values varied from 2 km to 5 km and provides some insight into the spatial context that is captured by landscape-fire simulation models like Burn-P3. Instead of actually growing fires, our statistical model was based on explanatory variables that were defined to reference spatial context to each individual location. This approach was highly effective at reproducing Burn-P3 results, correctly classifying roughly 90% of the area, and may represent an alternative for evaluating fire susceptibility when time, resources or computational issues make simulation modeling impractical. For example, evaluations of large numbers of candidate fuel management designs for an area could be expedited by producing a baseline fire susceptibility map using simulation modeling and then deriving a statistical model that could be used to quickly calculate predicted changes associated with each candidate design. This would eliminate the need for the repeated simulations that were required in our evaluation of the prescribed fire scenario. Another application may be in the area of optimization modeling for locating fuel treatments, which is an important fuel management research area (e.g., Finney et al., 2007; Finney, 2007; Wei et al., 2008). By combining simulation modeling with our statistical modeling approach, the computational manageability of identifying optimal fuel treatments could be improved. Fire activity in our study area can be expected to increase under a warming climate. For northern Alberta, Tymstra et al. (2007) used Prometheus fire growth simulations modeled with current conditions and conditions representative of 2 CO 2 and 3 CO 2 scenarios simulated by the Canadian Regional Climate Model (CRCM) to estimate potential future increases in area burned of 12.9% and 29.4%. Recent predictions for the boreal mixedwood forests of Alberta (Krawchuk et al., 2009) suggest a 1.9-fold and a 2.6-fold increase in area burned could occur for the periods and , respectively, based on 2 CO 2 and 3 CO 2 conditions simulated by the CRCM. Given the prospect of significant future increases in area burned, the use of landscape fuel management activities to strategically reduce fire susceptibility in high-valued areas becomes an ever more important preemptive action that fire and land managers can implement in support of traditional fire response and suppression activities. Simulation modeling approaches, like the one we have used, provide a means of integrating information about fuels, weather, topography and fire regime characteristics in an area in order to map fire susceptibility and prioritize costly fuel management options. These fire susceptibility maps are not limited to fuel management applications and may also be useful for prioritizing community protection planning activities and for assessing strategic level allocation of fire suppression resources across a management area. 5. Conclusions Use of the Burn-P3 simulation model, in combination with our Fire Susceptibility Index (FSI) approach, proved to be an effective means of evaluating the impact of fuel management activities on landscape fire susceptibility and may have widespread practical applications for forest industry and fire management practitioners. Vegetation, topography and ignition patterns were the key determinants of fire susceptibility. The spatial scale at which these variables influenced fire susceptibility at a given location consisted of a circular neighbourhood with a radius of 2 5 km. Our Fire Susceptibility Index (FSI) map derived from Burn-P3 results provides a relative rating of the likelihood that an area will be burned by a fire. These results are only considered applicable over the very short term (i.e., the next fire season) and would have to be repeated using updated inputs in the event that fire regime and/or fire environment conditions change substantially from the baseline period used to parameterize and calibrate the model. Comparisons of our results with future assessments of FSI values in other northern forest ecosystems could be used to further explore how fire environment and fire regime characteristics influence FSI values; and to further characterise the spatial context captured in this type of fire simulation modeling exercise where the characteristics of surrounding areas can be expected to influence fire susceptibility at any given point. Acknowledgements This project was funded by the Forest Resource Improvement Association of Alberta (Project OF-04-P036) and was completed in partnership with Millar Western Forest Products Ltd. and Alberta Sustainable Resource Development. A. Espinoza assisted with early data acquisition, preparation, and analysis. J. Wilkes prepared several model inputs and contributed to model calibration and testing. We thank K. MacDonald (Alberta Sustainable Resource Development) for providing valuable information and expertise regarding the development of the prescribed fire treatment scenario; R. Hilts (Millar Western Forest Products Ltd.); and C. Tymstra (Alberta Sustainable Resource Development) for providing helpful comments on an earlier version of the manuscript. References Alberta Sustainable Resource Development, Understanding fire weather [online]. < (accessed June 5, 2007). Burgan, R.E., Shasby, M.B., Mapping broad-area fire potential from digital fuel, terrain, and weather data. J. Forest. 82, Cardille, J.A., Ventura, S.J., Turner, M.G., Environmental and social factors influencing wildfires in the upper Midwest, United States. Ecol. Appl. 11, Carmel, Y., Paz, S., Johashan, F., Shoshany, M., Assessing fire risk using Monte Carlo simulations of fire spread. Forest Ecol. Manage. 257, Chapin, F.S., Trainor, S.F., Huntington, O., Lovecraft, A.L., Zavaleta, E., Natcher, D.C., McGuire, A.D., Nelson, J.L., Ray, L., Calef, M., Fresco, N., Huntington, H., Rupp, T.S., DeWilde, L., Naylor, R.L., Increasing wildfire in Alaska s boreal forest: pathways to potential solutions of a wicked problem. BioScience 58, Chuvieco, E., Congalton, R.G., Application of remote sensing and geographic information systems to forest fire hazard mapping. Remote Sens. Environ. 29,