Forest Ecology and Management 310 (2013) 875–886

Contents lists available at ScienceDirect

Forest Ecology and Management

journal homepage: www.elsevier .com/locate / foreco

Assessing possible shifts in wildfire regimes under a changing climatein mountainous landscapes

0378-1127/$ - see front matter � 2013 Elsevier B.V. All rights reserved.http://dx.doi.org/10.1016/j.foreco.2013.09.036

⇑ Corresponding author at: Graduate School of Environment and InformationSciences, Yokohama National University, 79-7 Tokiwadai, Hodogaya, Yokohama240-8501, Japan. Tel./fax: +81 045 339 4355.

E-mail address: akkym@kb3.so-net.ne.jp (A.S. Mori).1 Tel.: +1 403 220 7635; fax: +1 403 289 9311.

Akira S. Mori a,b,⇑, Edward A. Johnson b,1

a Graduate School of Environment and Information Sciences, Yokohama National University, 79-7 Tokiwadai, Hodogaya, Yokohama 240-8501, Japanb Biogeoscience Institute Dept. of Biological Sciences, University of Calgary, 2500 University Dr. NW, Calgary, Alberta T2N 1N4, Canada

a r t i c l e i n f o a b s t r a c t

Article history:Received 29 July 2013Received in revised form 17 September 2013Accepted 18 September 2013Available online 15 October 2013

Keywords:Ecosystem-based managementExtreme droughtFuture climate and fire scenariosNatural disturbanceWildfire

Climate change may affect the probability of extreme events such as wildfires. Although wildfires aresome of the most important ecological processes in forest ecosystems, large-scale wildfires are often per-ceived as an environmental disaster. Since failure to include the dynamic nature of ecosystems in plan-ning will inevitably lead to unexpected outcomes, we need to enhance our ability to cope with futureextreme events coupled with climate change. This study presents several future scenarios in three differ-ent time periods for Canada’s Columbia Montane Cordillera Ecoprovince, which is prone to wildfires.These scenarios predict the probability of occurrence of widespread wildfires based on the hierarchicalBayesian model. The model was based on the relationships between wildfires and the Monthly DroughtCode (MDC). The MDC is a generalized monthly version of the Daily Drought Code widely used acrossCanada by forest fire management agencies for monitoring of wildfire risk. To calculate future MDC val-ues, we relied on different possible future conditions of climate, given by the Global Circulation Models.We found a regime shift in drought intensity with abrupt decreases in lightning-caused wildfire activityaround 1940, suggesting that future wildfire risks can be inferred primarily from the summer droughtcode. For future periods, we found increasing trends in the probabilities of large-scale fires with timein most areas. It should be notable that, by the 2080s, there is a probability of some areas having morethan 50% of large-scale wildfires under the ‘‘average’’ climatic conditions in the future, indicating that,even without ‘‘extreme’’ weather conditions, some ecosystems will have a fundamental probability ofexperiencing catastrophic fires under the condition of average summer. However, the rate of progressiontoward a fire-prone condition is quite different among the three climate change scenarios and among theregion analyzed. Given such scenario-sensitive, spatially-heterogeneous patterns of wildfire probabilityin response to climate variability, management strategy should be flexible and more localized. By draw-ing on this knowledge, it may be possible to mitigate climate change impacts both before they arise andonce they have occurred. These considerations are critical for maintaining the integrity of systems shapedby large-scale natural disturbances to increase their resilience to the changing climate while protectinghuman society and infrastructures. Working with alternative scenarios will facilitate our adaptation toclimate change in managing fire-prone forest ecosystems.

� 2013 Elsevier B.V. All rights reserved.

1. Introduction

Efforts at management and conservation of ecological systemsare facing challenges of a global climate change and the need tocope with an accompanying uncertainty (Millar et al., 2007; Lawleret al., 2010). One of the critical components of an understanding ofthe responses of ecosystems to a changing climate is alterations ofthe disturbance regimes (Dale et al., 2001; Seidl et al., 2011b).

Natural disturbances are currently treated by ecologists not asevents bringing destruction, but as fundamental sources of diver-sity and heterogeneity (Spugel, 1991; Wallington et al., 2005).When natural disturbances are excluded, ecological variability de-creases, resulting in homogenization and reduction of ecosystemfunctioning, degradation of ecosystem services, and loss of biodi-versity (Wallington et al., 2005; Mori, 2011b; Pakeman, 2011).On the other hand, most infrequent and large-scale natural distur-bances such as wildfires, especially those consuming more than10,000 ha (Keane et al., 2008), are perceived as disasters to humancommunities and infrastructure, bringing social and economic dev-astation (Lindenmayer, 2004; Lindenmayer and Noss, 2006). Inaddition to such difficulties in evaluating the roles of large-scalenatural events, climate instability poses further uncertainty in

876 A.S. Mori, E.A. Johnson / Forest Ecology and Management 310 (2013) 875–886

the management of ecosystems prone to major disturbances (Mori,2011b; Seidl et al., 2011a). Human-driven climate change may alterthe frequency, extent, and severity of major disturbances (Dale et al.,2001); ecosystem conditions in various regions may deviate beyondthe historical ranges of variability that land management plans andmanagers often use as conservation and restoration targets (Millaret al., 2007; Jackson and Hobbs, 2009). Therefore, an urgent reassess-ment of management strategies is critical to the ability to cope withincreased probability of future shifts in disturbance regimes andecosystem states beyond historical reference conditions.

Among natural disturbances, wildfire is one of the most cli-mate-sensitive events. Numerous studies have indicated that in-creases in the frequency and the severity of temperature-drivendroughts are likely, thus creating more wildfire-prone conditionsin many regions (Flannigan et al., 2005, 2009; Gonzalez et al.,2010; Pechony and Shindell, 2010; Mastrandrea et al., 2011; West-erling et al., 2011). In addition to tree mortality, wildfire affectsvarious biological, geochemical, and geophysical characteristicsand processes, including carbon dynamics, nutrient cycles, soilproperties, seed germination, wildlife habitat, landscape heteroge-neity, successional trajectory, community assembly, and biologicaldiversity, among others (Dale et al., 2001; Turner et al., 2003; Nosset al., 2006). Therefore, in the near future, processes and function-ing of ecosystems prone to large-scale wildfires may differ vastlyfrom those found in the past. Although climate change is impactingtoday’s ecosystems, other, possibly more serious impacts willemerge decades into the future (Lawler et al., 2010). Jackson andHobbs (2009) noted that ecological management should aim toconserve and restore historical ecosystems where viable, whilesimultaneously preparing to steer emerging novel ecosystems toensure maintenance of ecological goods and services. In the faceof uncertainty associated with a rapid climate change, attemptsto conserve or restore past conditions will require enormous ef-forts from managers and could potentially create ecosystems moreprone to undesirable changes (Millar et al., 2007; Mori et al., 2013).Therefore, we need to accept that future ecosystems, especiallythose that may become more vulnerable to widespread wildfires,may be different and unique (Millar et al., 2007; Mori et al., 2013).

Insights on future wildfire regimes have been provided mainlyat broader spatial scales such as intercontinental comparisons, na-tion-level trends, and spatially-heterogeneous responses (Flanni-gan et al., 2005; Gonzalez et al., 2010; Pechony and Shindell,2010; Moritz et al., 2012; de Groot et al., 2013). However, muchof the information about climate change impacts is too broad tofully inform the managers of specific ecosystems (Lawler et al.,2010). Furthermore, actual data, projections, and possibilities forspecific local areas are generally not available. Thus, a new ap-proach needs to be incorporated into the future management op-tions, one that addresses local responses of ecosystems toclimate change and simultaneously compensates for limitationsof local data. Here, we aim to develop such a model to evaluatethe future probability of large-scale wildfires and a possible wild-fire regime shift in a mountainous ecoprovince in southwesternCanada. The model is based on the hierarchical Bayesian approachand is applied to show future scenarios of wildfire vulnerability atthe more localized scale within the study ecoprovince. Based onthese scenarios, we discuss flexible approaches to cope with inher-ent variability and uncertainty under the changing climate.

2. Methods

2.1. Study area

The national ecological framework for Canada has hierarchicallevels (ecozone, ecoprovince, ecoregion, and ecodistrict) as

ecological management units (Marshall et al., 1999). These levelswere deemed most suitable for reporting on national issues and re-gional issues of national significance concerning the environmentand sustainability of resources. Ecodistricts are the smallest ofthe management units, and are characterized by distinctive assem-blages of landform, relief, surficial geologic material, soil, waterbodies, vegetation, and land uses. Ecodistrict size is a function ofregional variability of these defining attributes.

In this study, a total of 24 ecodistricts constituting the ColumbiaMontane Cordillera Ecoprovince in western Canada were includedto determine drought-wildfire relationships (Fig. 1). The studiedecoprovince is one of the four subdivisions of the Montane Cordil-lera Ecozone, which is the most diverse of Canada’s 15 terrestrialecozones, exhibiting some of the driest, wettest, coldest, and hot-test conditions anywhere in Canada (Wiken, 1986). The ecosys-tems in this region are heterogeneous, ranging from alpinetundra and dense conifer forests to dry sagebrush and grasslands.Much of the region is rugged and mountainous. Each ecodistrictin the ecoprovince has different land-cover characteristics (Ta-ble 1), indicating that working at the ecodistrict level will yieldmore appropriate perspectives on ecosystem-wildfire relationshipsthan that at the higher hierarchical levels.

2.2. Wildfire data

The Canadian National Fire Database (CNFDB; http://cwfis.cfs.nrcan.gc.ca/) is a collection of wildfire data, which includeyear, date, location, perimeters, cause (human or lighting), anddata source (provided by Canadian fire management agencies ofprovinces, territories, and Parks Canada) of fires of all sizes. In thisstudy, all the data available for lightning-caused fire up to 2009 inthe Columbia Montane Cordillera Ecoprovince were obtained fromthe CNFDB. The CNFDB had partial fire data for 2010 at the time ofdata access (March 2011); these were excluded from analyses asincomplete. The database includes information on past human-caused fires; however, it is difficult to predict their occurrencesin the future, as human-caused fires are expected to largely dependon societal behavior. Although several recent studies have pro-jected future possibilities of human-induced wildfire activity(e.g., Liu et al., 2012), we therefore focused only on natural wildfireregime, which is tightly associated with the changing climate. Theecoprovince straddles British Columbia (BC) and Alberta (AB) andthe time period of the CNFDB data differs between the two prov-inces (from 1917 in BC and from 1931 in Alberta). However, thelarger portion of the study ecoprovince is within BC so we assumedthat the lack of data for AB did not significantly affect the results. Ininterpreting our results from analyses of temporal changes in wild-fire activity, we addressed the lack of early data in some ecodis-tricts located in AB.

2.3. The drought code

Currently, forest fire management agencies in Canada (Lawsonand Armitage, 2008) and other countries (de Groot et al., 2007)use the Fire Weather Index (FWI) System to assess wildfire risks.FWI relies on three indices, i.e., Fine Fuel Moisture Code (FFMC),Duff Moisture Code (DMC), and Drought Code (DC) (Lawson andArmitage, 2008). Among them, the DC is an index of the net changein evapotranspiration and precipitation on cumulative moisturedepletion in organic soils. The DC is practical for estimating thedanger and risk of fire; it has a slower response time (62 days at15 �C and 44 days at 30 �C), less day to day variability than theother two codes, and is more in tune with the blocking high pres-sure systems associated with large wildfires that determine wild-fire regimes (Johnson and Wowchuk, 1993; Macias Fauria andJohnson, 2006). However, the DC relies on daily weather data,

Fig. 1. Map of the Columbia Montane Cordillera Ecoprovince in Canada. All of the 24 ecodistricts are shown with polygons. The number on each ecodistrict is the ID, listed inTable 1.

Table 1Land area, land cover type, and the summer drought intensity in the studied 24 ecodistricts.

Ecodistrict ID Land area (km2) Land cover (%) Forest cover (%) MDC

I II III IV V VI VII Mean Maximum s.d.

Northern Columbia Mountains 985 29400.5 19 9 48 20 4 0 0 76 30.0 148.6 45.2Bowron Valley 986 5928.9 12 2 86 0 0 0 0 100 74.9 189.0 49.9Quensel Highland 987 7019.3 15 2 83 0 0 0 0 100 83.0 200.5 46.8Shuswap Highland 988 13463.7 35 4 60 1 0 0 0 99 92.1 224.4 51.1Eastern Purcell Mountains 989 5847.8 23 3 51 23 0 0 0 77 59.5 164.9 45.2Central Columbia Mountains 990 15480.9 32 11 45 12 0 0 0 88 90.4 211.5 53.3Southern Columbia Mountains 991 6665.5 51 15 34 0 0 0 0 100 143.0 261.8 50.3McGillivary Range 992 2135.3 67 4 29 0 0 0 0 100 173.1 279.4 43.6Northern Park Ranges 993 7192.6 20 6 52 21 1 0 0 78 2.9 131.7 46.6Central Park Ranges 994 5599.6 18 10 27 40 5 0 0 55 48.1 160.4 41.7Southern Park Ranges 995 10864.7 28 13 35 24 0 0 0 76 35.9 155.0 48.2Willmore Foothills 996 5736.7 17 2 77 4 0 0 0 96 42.2 180.0 46.4Jasper Mountains 997 10805.3 14 3 53 29 1 0 0 70 �0.8 132.0 44.3Luscar Foothills 998 2298.1 15 0 78 7 0 0 0 93 21.5 149.7 41.3Banff Mountains 999 15958.8 13 3 36 48 0 0 0 52 �17.5 100.6 42.4Icefield Mountains 1000 2863.1 7 2 30 51 10 0 0 39 19.1 140.2 41.4Selkirk Foothills 1012 7712.7 37 5 58 0 0 0 0 100 158.4 281.5 55.0Upper Fraser Trench 1013 2405.1 35 7 58 0 0 0 0 100 115.3 234.0 44.7Big Bend Trench 1014 326.8 46 17 36 1 0 0 0 99 158.1 253.4 38.7East Kootenay Trench 1015 4285.9 57 9 31 3 0 0 0 97 218.5 310.1 37.9Morley Foothills 1016 1256.2 44 3 14 7 0 0 32 61 100.1 214.4 47.3Crowsnest Mountains 1017 4836.5 43 20 18 14 0 0 5 81 80.3 212.5 52.1Blairmore Foothills 1018 2562.8 49 1 2 3 0 25 20 52 126.8 248.0 57.0Waterton Mountains 1019 1076.1 62 11 7 19 0 1 0 80 45.2 178.3 64.6

Land cover type. I, Mixed Forest: canopy 26–75% coniferous/broadleaf trees; II, Broadleaf Forest: canopy >75% broadleaf trees; III, Coniferous Forest: canopy >75%; IV, SparselyVegetated/Barren Land: plant cover generally <25%coniferous trees; V, Perennial Snow or Ice: perennial snowfields or glaciers; VI, Cropland: cultivated land; VII, Rangelandand Pasture: native vegetation with <10% tree cover. Land cover data source; National Ecological Framework for Canada, http://sis.agr.gc.ca/cansis/nsdb/ecostrat/.

A.S. Mori, E.A. Johnson / Forest Ecology and Management 310 (2013) 875–886 877

and it is impossible to obtain at daily resolution spanning broaderspatial scales and longer time periods, especially with finer spatialresolutions. Recently, Girardin and Wotton (2009) developed amonthly drought code (MDC), and showed that it was well corre-lated with annual fire statistics at broader spatial scales (ecore-gion). Although the MDC was found to have lower predictivecapacity at maritime locations due to high precipitation regimes(Girardin et al., 2009), we expect it to be useful in studyingdrought-wildfire relationships in interior mountain areas that re-ceive little precipitation during summer. The MDC is unitless witha maximum value of 400 (Girardin et al., 2009; Girardin and

Wotton, 2009). The total monthly precipitation and monthlyaverage daily maximum temperature data for the MDC calcula-tions were obtained from Climate Western North America (ClimateWNA) (Mbogga et al., 2009; Wang et al., 2012). Climate WNA relieson the delta adjustment, where historical data and futureprojections are expressed and interpolated as a difference (referredto as ‘‘delta’’ or ‘‘anomaly’’) from a reference period of the 1961–1990. The weather data were generated for 0.1� latitude by 0.1�longitude grids covering the period 1901–2009. Furthermore, theClimate WNA can downscale the climate predictions of the globalcirculation models (GCMs) that were used in the fourth assessment

878 A.S. Mori, E.A. Johnson / Forest Ecology and Management 310 (2013) 875–886

report (AR4) of the Intergovernmental Panel on Climate Change(IPCC 2007) for the periods of the 2020s (2010–2039), 2050s(2040–2069), and 2080s (2070–2099). The list of available GCMsis summarized in Table A1. Wang et al. (2012) demonstrated thatthe delta approach of the Climate WNA can substantially reducethe amount of error associated with the GCM baseline data, gener-ating reasonable downscaled projections for future periods. Weused the MDC values for the month of August, since peak fire sea-son in the study area is from July to August and the August MDCintegrates the influence of the previous months including July.We calculated the MDC values from all models available in the Cli-mate WNA for each of the future climate scenario (A1B, A2 and B1).

2.4. Drought-wildfire relationships: regime shift

Trends in the August MDC, annual area burned (AAB) and an-nual frequency of large wildfires were evaluated with two methodsfor the detection of a regime shift. The regime shift is defined as asudden shift in a response variable (MDC and AAB) caused byexceeding a given threshold in a time-series. At first, the decadalchanges were tested with the regime shift detection on a sliding10-year window with correlation for serial persistence in data(sequential regime shift detection (Rodionov, 2004)). This analysismakes it possible to verify that changes in the mean value from oneperiod to another were not a manifestation of a red noise (autore-gressive) process. This study used probability of 0.1, outliersweight parameter of 4, and IP4 method for red noise correction.Then, to further validate the reliability of change year(s) in droughtor wildfire regime detected in the former analysis, another methodof threshold detection was conducted with sequential F-tests (Zei-leis et al., 2003). In this change-point analysis, we selected thenumber of threshold points according to that detected by sequen-tial regime shift detection, and calculated 90% confidence inter-val(s). In this study area, correlation between AAB and MDC washighest in August so that we used the ecoprovincial average ofthe August MDC values from a total of 2024 points to evaluatethe ecoprovince-level drought regime shift. Wildfire regime shiftat the same level was evaluated based on the total burned areain the ecoprovince. Annual frequency of large wildfires was definedas the number of wildfires larger than 400 ha in each year. Thethreshold value of 400 ha was selected according to earlier studies(Westerling et al., 2006; Mori, 2011a). Furthermore, the PalmerDrought Severity Index (PDSI) was used to confirm the climate-wildfire relationships. We used the PDSI of Dai et al. (2004) whichcovers global land areas on a 2.5� by 2.5� grids for the period of1870 to 2002. We used the Regime shift detection software Version3.2 (Rodionov, 2004) and Strucchange R package Version 1.4-6(Zeileis et al., 2003).

The summer MDC is a reliable approximation of wildfire-cli-mate correlation over broad areas and over time (Girardin et al.,2009; Girardin and Wotton, 2009). To evaluate a possible relation-ship between the MDC and AAB, we used the Bayesian approach.The Bayesian approach is a robust method to infer uncertainty orvariability of parameters in hierarchical models (McCarthy,2007), such as when the data are auto-correlated and do not satisfythe criterion of randomness. Here, we added a random effect pos-sibly caused by the temporal auto-correlation on the log-linearmodel of drought and burned area in the form yi � Normal(bi) with

logðbiÞ ¼ a0 þ axi þ e

where yi and xi are burned area and the August MDC in the year i,respectively, and a0 and a are the intercept and slope to be deter-mined. The random effect is represented by e in the equation. Allparameters were assumed to follow a normal distribution. The ran-dom effect was set to follow Normal(0, s), where s is a hyper-parameter that was assumed to follow Uniform(100, 1000). We used

WinBUGS Version 1.4.3 (Lunn et al., 2000) and R2WinBUGS R pack-age Version 2.1–16 (Sturtz et al., 2005) to fit the model using Mar-kov chain Monte Carlo (MCMC) simulation. We performed threechains of 51,000 iterations with different initial values, discardedthe first 11,000, and thinned by 10, which resulted in 12,000 itera-tions used for inference.

2.5. Drought-wildfire relationships: a probability model for large-scalewildfires

The probability of a widespread wildfire at the ecodistrict levelwas modeled with the August MDC using logistic regression. In thisstudy, we used several definitions of wildfire years. The first defi-nition of wildfire years included those wildfires that were largerthan 1000 ha (hereafter, defined as large wildfires) (Mori, 2011a).However, the size of ecodistricts is variable; suggesting that a largefire in one ecodistrict may not meet that definition in another eco-district. Therefore, major wildfires in the Rockies have been alsodefined as burned areas covering more than 0.5% of the total areaburned over the study period (Schoennagel et al., 2005) (hereafter,defined as major wildfires). This is useful for the detection of infre-quent wildfire events in each area when the meaning of ‘‘infre-quent’’ differs among areas.

Stand-replacing wildfires are episodic at local scales. Thismeans that data in each ecodistrict are often insufficient to calcu-late the logistic model, potentially causing overdispersion. To over-come this challenge, we developed an interchangeable modelbased on hierarchical Bayesian modeling (Albert, 2009) that al-lowed us to infer the probability of a wildfire at the ecodistrict le-vel in the form of Yij � Bernoulli(pij) with

log itðpijÞ ¼ bj1 þ bj2Xij

where Yij and Xij are the probability of wildfires and the August MDCin the year i of the Ecodistrict j, respectively, and bj are the regres-sion parameters of the Ecodistrict j. We assumed that the parame-ters specific to each ecodistrict are interchangeable; all ecodistrictsare within the same ecoprovince implying that they are somewhatsimilar. Suppose that bj = (bj1, bj2) is the vector matrix of regressionparameters generated from a common multivariate prior, with amean lb and a variance-covariance matrix V expressed as

bjjlb; R � Normalðlb;VÞ

where R is the precision matrix. Then, we assigned the vague priorto the hyperparameters as follows:

lb � c;V � invWishartðS�1;vÞ

where c is the mean of the mean vector lb and invWishart(S�1, v)denotes the inverse-Wishart distribution with scale matrix S anddegrees of freedom v. We used c = 0, S = 0.001, and v = 2 for the cal-culation. That is, the model seeks to combine the individual regres-sion parameters in a way that reflects the somewhat common firebehavior of all ecodistricts within an ecoprovince.

In the above analysis of wildfire probability, we did not considerthe effect of temporal auto-correlation because it did not signifi-cantly improve the model fit. However, all ecodistricts may notbe treated as equally similar, because of spatial dependence. Eco-districts adjacent to a jth ecodistrict may be more similar com-pared with those not adjacent to the jth ecodistrict. Since theabove interchangeable model does not consider spatial arrange-ment of the ecodistricts, we incorporated another random effect(u) using an intrinsic Gaussian conditionally autoregressive(CAR) prior expressed as

u � CARðscÞ

010203040506070

1900 1920 1940 1960 1980 2000

0

50000

100000

150000

200000

250000

3000003

2

1

0

-1

-2

-30

50

100

150

200

Year

PDSI

Annu

al b

urne

dar

ea (h

a)An

nual

freq

uenc

yof

fire

s >

400

haM

DC

AUG

UST

0

Fig. 2. Regime shifts in drought, annual area burned, and annual frequency of largewildfires >400 ha/year in the studied ecoprovince. MDCAUGUST is the mean MDCvalues in August. PDSI is the Palmer Drought Severity Index. Mean values in eachphase that was detected by the regime shift detection are shown with red lines.Break points detected by the change-point analysis are shown with blue line andtheir 90% confidence intervals are shown with light blue boxes. (For interpretationof the references to colour in this figure legend, the reader is referred to the webversion of this article.)

Fig. 3. Linear relationship between drought and annual area burned in the studiedecoprovinces. Annual area burned was log-transformed and MDC_August is themean MDC values in August. Dotted lines indicate the 90% confidence interval.

A.S. Mori, E.A. Johnson / Forest Ecology and Management 310 (2013) 875–886 879

ujuk–m � Normalðum;1=ðscnmÞÞ

where nm is the number of ecodistricts neighboring ecodistrict k,and sc is the precision. We used sc � Gamma(0.5, 0.0005). This spa-tial random effect was added in formula 2 to see whether the modelfitness could be improved by considering spatial dependence.

In the following section, the wildfire probability model with andwithout the spatial random effect is described as CAR and Non-CARmodel, respectively. Note that we assumed that all ecodistrictsshared a common fire regime to some degree (interchangeable),and the CAR model considers neighbors to be similar. Mean valuesof the August MDC in each ecodistrict, calculated from the pastperiod of 1901–2009 and the future periods of the 2020s, 2050sand 2080s, were used for the models. For future periods, we usedthe MDC values calculated for all models available in the ClimateWNA. In both models, we used the same software and calculationprocedures (chains, burn-in, and iterations of MCMC) as the linearregression model described previously. Based on the projected fu-ture probability of large-scale wildfires for each period under thethree scenarios, we calculated coefficient of variation (CV) to inferpossible variability and uncertainty in the future.

2.6. Evaluation of model fit and model selection

To evaluate the reliability of the wildfire probability models,model fit based on the past MDC values was tested by comparingmodel results with the observed frequency of large-scale wildfires.Our objective was to construct probability models of large-scalewildfires for the future periods. Since scenarios based on the GCMsare to predict climate conditions during three decades, we de-signed our models to show wildfire probability in each of the threefuture decades. First, we counted wildfire years (large and majorwildfires) in the past three-decade periods, 1920–1949 (1930s),1950–1979 (1960s), and 1980–2009 (1990s), recorded in theCNFDB, and calculated the proportion of wildfire years in each per-iod in each ecodistrict. The ordinal linear regression was performedbetween these observations and the wildfire probability calculatedfrom the models. Fitness of the models was then evaluated basedon R-square values representing accountability of the variance.

3. Results

3.1. Shifts in drought and wildfire regimes

Shifts in drought and wildfire regimes in the Columbia MontaneCordillera Ecoprovince are shown in Fig. 2. The August MDCshowed regime shifts in 1916 and 1940. These shifts were ob-served with both regime shift detection and change-point analysis.AAB showed a similar tendency with the MDC. The PDSI exhibitedregime shifts in 1916 and 1941, which are consistent with theMDC. The August MDC had finer spatial resolutions and more datathan the PDSI, so this study primarily relied on the MDC for the fol-lowing analyses. Annual frequency of large wildfires and AABshowed a threshold change in 1937 and 1939, respectively. Thesewere within the confidence intervals of changes in the droughtcode. Since there were no data in the CNFDB before 1917, thisstudy could not find the change-point when the wildfire activityincreased. The AAB increased with the August MDC values in theColumbia Montane Cordillera Ecoprovince (Fig. 3), indicating thatshifts in wildfire regime that coincided with those in the droughtregime were primarily caused by a change in the threshold of thedriver. With MDC values less than 100, AAB was variable, indicat-ing an uncertainty in the occurrence of stand-replacing wildfires.However, with MDC values greater than 100, there were moreextensive wildfires, suggesting that the wildfire model based onthe MDC-fire relationship was more reliable under drier situations.

880 A.S. Mori, E.A. Johnson / Forest Ecology and Management 310 (2013) 875–886

Consistent with the results at the ecoprovince-level, the AABalso showed a threshold change at the ecoregion-level (Fig. A1).Four of the six ecoregions in the Columbia Montane Cordillera Eco-province showed shifts in the AAB around 1940 (Fig. A1). Regard-ing frequency of large wildfires, although change-point yearswere variable among the ecoregions from 1928 to 1942, five eco-districts showed a decrease in wildfire activity (Fig. A1). Lack ofshift in the AAB in the other two ecoregions was caused by the lackof data in the period before 1931 (Fig. A1). In spite of the data lim-itation, one of these two ecoregions showed a decrease in largewildfire frequency in 1936 (Fig. A1). However, at the ecodistrict-scale, only eight of 24 ecodistricts showed the shift in AAB around1940 (Fig. 4), indicating that wildfire regime was more heteroge-neous at the ecodistrict scale, compared with the coarser spatialresolution. The ecodistricts that exhibited shifts in the wildfire re-gime were not biased to a specific ecoregion and were found in fiveof the six ecoregions (Fig. 4).

3.2. Wildfire probability models

Diagnostics for the wildfire probability models are summarizedin Table 2. The Non-CAR model explained 32–90% of the variance,while the CAR-model explained 18–81%. In predicting large(>1000 ha yr�1) and major (0.5% of sum of observed wildfires inthe CNFDB) wildfires, the Non-CAR model explained the variance

Fig. 4. Regime shift in annual area burned at the ecodistrict level. Each polygonrepresents ecodistrict area. Cross-shaded ecodistricts showed an abrupt decrease inannual area burned during 1916–1939 (Table 3). Ecoregion groups enclosed bythick black lines belong to the same ecoregion.

Table 2Results of the diagnosis regressions.

Model Large fire category Period

1930s 1960s 1990s

Non-CAR Major wildfires 0.62 0.75 0.69Large wildfires 0.90 0.62 0.32

CAR Major wildfires 0.47 0.64 0.62Large wildfires 0.78 0.81 0.18

R-square values between the observed frequency and the probability of widespreadwildfires in the past three periods are shown for the two models. Visualized rela-tionships are also given in Fig. A2.

somewhat better than the CAR-model Hence, this study uses theNon-CAR model to propose wildfire scenarios for the futureperiods.

Fig. 5 shows that the future probability of major wildfires (>0.5%of total observations in each ecodistrict) generally increased withtime. Modeled future probability of major wildfires was differentdepending on the climate scenarios. In the 2020s, differences inthe probability of major wildfires were not remarkable amongthe three scenarios. In the 2050s, most of the western ecodistrictshad a high probability of major wildfires under the A1B and A2 sce-narios, while progressive increases in the probability of majorwildfires were variable under the B1 scenario. In particular, onlythe A1B scenario exhibited greater than 50% probability of majorwildfires in the Northern Columbia Ecodistrict (ID = 985 in Fig. 1)by the 2050s (Fig. 5). In the 2080s, climate changes caused furtherincreases in the probability of major wildfires throughout the eco-province except for the northeastern ecodistricts. Predictions forthe western ecodistricts were variable. The Central Columbia Eco-district (ID = 990) generally showed a lower probability of majorwildfires than the Northern Columbia Ecodistrict, but the probabil-ities of major wildfires in the 2080s were higher in the Central thanNorthern Columbia Ecodistrict under the A2 scenario (Fig. 5). Nota-ble is that CV values of wildfire probability in each period werelarge for the B1, suggesting relatively-high uncertainty under thisscenario.

Future probability of large (>1000 ha yr�1) wildfires alsoshowed gradual increases with time (Fig. 6); overall trends of in-creases in the probabilities of large wildfires were consistent withthe major wildfires shown in Fig. 5. In the 2020s, wildfire probabil-ity was not different among the scenarios, although southwesternecodistricts showed slightly higher probability of a large fire underthe A1B and A2 (Fig. 6). In the 2050s, the A1B scenario yielded ahigher probability of large wildfires in the western ecodistricts(Fig. 6). In the 2050s under the B1 scenario, only the NorthernColumbia Ecodistrict showed a probability higher than 20% forlarge wildfires (Fig. 6). In the 2080s, progressive increases in theprobabilities of large wildfires were not substantial under B1,while those of the southwestern ecodistricts and the NorthernColumbia Ecodistrict increased under the A1B and A2 (Fig. 6). Inparticular, only the A1B scenario in the 2080s exhibited greaterthan 50% probability of large wildfires in the Northern ColumbiaEcodistrict (Fig. 6). Similar to the result for major wildfires, rela-tively high uncertainty (i.e., larger CV values) was found for theB1 scenario (Fig. 6).

3.3. Future changes in drought

Future changes in the August MDC are illustrated in Fig. 7.Although the MDC values increased with time under all of the sce-narios, responses were spatially heterogeneous. Under the A1B sce-nario, some ecodistricts reached more than 50% of the thresholdlevel of 1916–1939 by the 2050s and most of them exceeded thesame level by the 2080s (Table 3). However, no ecodistrictsreached the same level by the 2050s under the A2 and B1 (Table 3).By the 2080s, the August MDC in some ecodistricts reached morethan 50% of the threshold level of 1916–1939 under the A2 sce-nario (9 of 24 ecodistricts), while this was not observed underthe B1 scenario even in the 2080s (Table 3).

4. Discussion

4.1. Shifts in wildfire and drought regimes

In the Columbia Montane Cordillera Ecoprovince of westernCanada, regime shift in drought intensity was responsible for

Fig. 5. Future probability of major wildfires (annual area burned >0.5% of total burned areas in the fire database in each ecodistrict). For each scenario in each period,probabilities gained from different GCM models were averaged and then visualized. Coefficient of variation (CV) for each each scenario in each period was shown under eachresult.

Fig. 6. Future probability of large wildfires (annual area burned >1000 ha/year). For each scenario in each period, probabilities gained from different GCM models wereaveraged and then visualized. Coefficient of variation (CV) for each each scenario in each period was shown under each result.

A.S. Mori, E.A. Johnson / Forest Ecology and Management 310 (2013) 875–886 881

Fig. 7. Future changes in summer drought intensity in the studied ecoprovince. Colors in each grid cells are according to the MDC values in August. The value of 0 indicatesthe mean drought level in the dry phase in 1916–1939. Note that, differing from Table 3, values in each grid cell are changes in actual values compared with the drought levelin the early 20th century.

Table 3Percentage of grid cells that exceeded the drought level in 1916–1939.

Ecodistrict Regime shift in annual burned area Grid cells with increased MDC (%)

A1B A2 B1

Direction Shift year 2020s 2050s 2080s 2020s 2050s 2080s 2020s 2050s 2080s

Northern Columbia Mountains – 0.0 39.1 62.2 14.1 13.1 37.1 0.4 14.4 25.6Bowron Valley – 0.0 42.1 59.3 16.1 12.5 33.5 0.5 8.5 22.3Quensel Highland – 0.0 41.5 53.0 16.7 11.4 33.3 0.1 9.9 22.8Shuswap Highland Decrease 1941 0.0 36.0 50.0 15.7 11.3 36.3 0.4 8.5 24.7Eastern Purcell Mountains Decrease 1935 0.3 43.9 66.4 13.7 11.3 39.2 0.6 8.3 25.1Central Columbia Mountains Decrease 1941 6.0 59.5 75.0 11.5 12.8 44.7 2.7 7.6 22.8Southern Columbia Mountains – 6.7 57.5 72.3 10.5 12.5 47.8 5.1 6.1 20.0McGillivary Range Decrease 1941 0.0 46.2 54.2 7.5 11.8 47.8 7.4 6.5 17.1Northern Park Ranges – 1.0 35.9 64.6 10.7 15.6 48.9 5.3 5.6 18.2Central Park Ranges – 0.3 38.7 49.7 10.3 15.6 49.1 5.7 5.9 20.1Southern Park Ranges Decrease 1941 0.0 35.1 49.4 9.0 11.5 47.7 4.9 4.5 15.4Willmore Foothills – 0.0 36.4 49.8 9.5 6.5 47.2 4.5 2.7 16.6Jasper Mountains – 1.7 35.7 65.5 8.0 4.6 49.7 2.3 0.5 13.0Luscar Foothills – 0.1 43.9 59.1 7.1 3.0 52.5 0.4 0.0 12.6Banff Mountains – 0.5 41.1 56.1 12.9 12.2 61.8 0.3 1.5 22.8Icefield Mountains – 0.0 39.9 54.2 16.7 33.3 74.1 0.0 4.8 36.5Selkirk Foothills Decrease 1941 7.0 62.3 71.9 18.3 32.1 76.4 0.0 3.6 37.1Upper Fraser Trench – 14.9 51.7 71.1 20.0 26.7 76.1 3.3 2.9 31.4Big Bend Trench – 0.0 38.9 53.5 25.0 41.7 83.3 8.3 2.4 38.1East Kootenay Trench Decrease 1933 1.6 57.5 75.4 19.7 35.2 78.5 3.1 4.9 35.8Morley Foothills – 2.3 56.6 74.2 21.6 31.4 77.5 5.0 0.0 36.1Crowsnest Mountains Decrease 1937 4.8 60.0 69.5 18.4 24.2 69.0 2.5 0.7 38.5Blairmore Foothills – 5.0 58.0 71.4 13.2 0.0 45.6 0.0 0.0 21.0Waterton Mountains – 0.0 51.3 56.5 15.2 5.3 48.5 0.0 0.0 24.7

Ecodistricts that showed a aburupt decrease in annual area burned within the confidence interval of the change point 1916–1939 for the annual area burned of the wholeecoprovince are indicated. The percentage of grid cells that exceeded the mean drought level during the dry period of 1916–1939 within each ecodistrict is shown for the allthree scenarios in the three future periods. Values indicate means from all GCM models used. Values larger than 50.0 are shown in bold.

882 A.S. Mori, E.A. Johnson / Forest Ecology and Management 310 (2013) 875–886

A.S. Mori, E.A. Johnson / Forest Ecology and Management 310 (2013) 875–886 883

abrupt decreases in lightning-caused wildfire activity around 1940(Fig. 2). This regime shift is consistent with the smooth pressure-status relationship explained by Andersen et al. (2009) becausethe fire regime shift in our study resulted from the linear responseto the pressure (i.e., drought) (Fig. 3). This indicates that futurewildfire risks can be inferred primarily from the summer droughtcode. Because large-scale fires have significant ecological effectsand social consequences (Stocks et al., 2002; Kitzberger et al.,2007; Keane et al., 2008; Williams and Bradstock, 2008; Flanniganet al., 2009), the possibility of a fire-prone future is relevant for re-gional management and policymaking. Although this simplifica-tion may miss some important feedbacks between ecosystemsand physical environments, wildfire projection using regional val-ues of the summer drought code demonstrates a reasonable andpractical approach to thinking of a critical threshold of future cli-mate conditions. Fire-risk assessment strictly founded on trendsor phases is not sufficient, as illustrated by several extremedroughts after 1940 (Fig. 2); therefore recognition of such pulses(Bouchard et al., 2008) is also important even before the mean cli-mate conditions exceed certain critical levels. A shift to a phasethat is more prone to an increased frequency and intensity of nat-ural disturbances can substantially change ecosystems throughalterations of post-disturbance successional pathways (Turneret al., 1998), and substantially raise costs of fire management(Flannigan et al., 2009). Therefore, attempts to couple the re-sponses of drought and wildfires to the ongoing climate trendsare worthwhile.

Another important issue is scale-dependency of disturbance re-gimes (Peters et al., 2004; Jackson, 2006; Falk et al., 2007; Mori,2011b; Grigulis et al., 2013). In this study, abrupt decreases inthe frequency and the extent of wildfires by 1940 were commonlyobserved at both the ecoregion and the ecoprovince scales (Fig. 2;Fig. A1). One may consider that these results suggest homogeneouspatterns of drought-induced wildfires throughout the region. How-ever, based on the past regime shift (Fig. 4) and the future proba-bility of wildfires (Fig. 5 and 6), we found more heterogeneousresponses to climate variability at the more local scale of ecodis-tricts. The relatively poor performance of the wildfire probabilitymodel that accounted for the similarity between adjacent ecodis-tricts (Table 2) also suggested that each ecosystem defined as anecodistrict had specific characteristics of the wildfire regime. Theecodistrict is designed to represent relatively uniform biophysicaland climatic characteristics (Strong and Zoltai, 1989). Therefore,although it is difficult to determine the optimal spatial scale, eco-district level is meaningful for regional managers in facing cli-mate-driven wildfire risks. Furthermore, an evaluation of apossible future shift in wildfire regime at finer scales may not beadequate. Kootenay National Park (approximately 1400 km2),which is in the Southern Park Ranges Ecodistrict (ID = 995), expe-rienced the largest AAB since the 18th century in 1926 (Masters,1990; Mori and Lertzman, 2011). Although this may be perceivedas an exceptional large-scale fire event within this park, it com-prised only 17.7% of the AAB in the ecodistrict in the same year.Wildfires that burn more than 10,000 ha yr�1 are often observedat a coarse management unit such as the Montane Cordillera Ecoz-one, but are infrequent and episodic at the finer geographical scalesuch as that of Kootenay National Park. Thus, careful considerationof spatial scales is needed to define disturbance regime for effec-tive ecological management.

In analyzing regime shifts, we treated wildfires as a responsevariable governed by a driver (weather). However, in the theoreti-cal and empirical frameworks of ecological thresholds, disturbanceis often the driver that causes abrupt changes in the state of an eco-system (Scheffer and Carpenter, 2003; Folke et al., 2004; Sudinget al., 2004). Increased frequency of large-scale fires is known to al-ter key characteristics and processes of ecosystems, such as the

composition of the dominant species (Hallett and Hills, 2006; Carc-aillet et al., 2007) and the carbon cycle (Bond-Lamberty et al.,2007; Metsaranta et al., 2010). An ecological regime shift drivenby natural disturbances is a complex phenomenon mediated byvarious processes and feedbacks (Scheffer and Carpenter, 2003;Suding et al., 2004); thus, responses of ecosystems often show non-linear dynamics. Although predicting future behaviors and re-sponses of an ecosystem state following a shift in wildfire regimeis beyond our scope, we provide some insights. First, frequentoccurrences of severe fires can devastate forest vegetation and re-sult in the dominance of barrens, inducing subsequent changessuch as soil erosion (Dale et al., 2001; Colombaroli and Gavin,2010). Loss of a dominant species may result in a reduction of bio-tic associations, possibly preventing communities from recovery(Mueller et al., 2005). Therefore, compared with substantial effortsrequired for ecological restoration after cumulative changes areobserved, preventive management actions, for example, reductionof fire risks by fuel treatment as means of addressing a potentialshift of a wildfire regime, would be relatively easier to implement.Adaptive strategies might include constructing of complete fuelbreaks around areas of highest risk or highest value, such as wild-land–urban interfaces and forests with high ecosystem services.Based on the smooth and relatively predictable pressure-statusrelationship between climate and wildfires, which is relatively pre-dictable, a precautionary approach is useful in confronting a chang-ing climate.

4.2. Drought as the major driver of wildfires

In addition to drought, various factors including other fire-weather conditions (such as wind speed and solar radiation),topography, vegetation type, fuel accumulation, and human factors(such as land-use and ignition factors) also determine wildfireactivity (Schoennagel et al., 2004; Zumbrunnen et al., 2008; Bow-man et al., 2009; Littell et al., 2009). However, climate fundamen-tally governs wildfire activity throughout many regions (Johnsonand Wowchuk, 1993; Kitzberger et al., 2007; Macias Fauria andJohnson, 2008; Miller et al., 2008; Girardin et al., 2009) includingthose under fire suppression and exclusion policies such as thosewidely adopted in North America (Littell et al., 2009; Mori andLertzman, 2011). Therefore, focus on weather variables such astemperature and precipitation may provide a uniform method toevaluate future risks of wildfires in forest ecosystems independentof other secondary factors (Nitschke and Innes, 2008). This ap-proach facilitates comparisons of wildfire sensitivity to a changingclimate among different regions (Girardin et al., 2009; Meyn et al.,2010). Such comparisons play a critical part in mitigating climatechange through their possible contributions to evaluating changesin positive/negative feedbacks between terrestrial ecosystems andclimate, mediated by forest fires (Bowman et al., 2009). Further-more, comparisons help in land management planning at variousgeographical scales (Nitschke and Innes, 2008; Flannigan et al.,2009).

The Columbia Montane Cordillera Ecoprovince is one of themost diverse regions in Canada. The diversity results from largevariations in biological, meteorological, and geophysical character-istics, and may be expected to make wildfire prediction uncertaindue to various interactions among factors. Nevertheless, our modelshowed a high predictability of large-scale wildfires (Table 2). Thedrought code used in the model does not fully account for someoverwintering effects of hydroclimatic conditions such as snowaccumulation, which affect the length of the following fire seasonand resultant wildfire activity (Westerling et al., 2006; Morganet al., 2008). However, a high predictability suggests that, it is use-ful to estimate wildfire risks even in a region characterized by acomplex mountainous terrain.

884 A.S. Mori, E.A. Johnson / Forest Ecology and Management 310 (2013) 875–886

4.3. Future wildfire probabilities

There is no consensus on the probability of fire in a warmerworld. Girardin et al. (2009) demonstrated that, despite a warmingsince the mid-19th century, trends of decreased occurrences of ex-treme droughts were noted for the Canadian boreal forests in thepresent times. This was consistent with the trends toward a moist-er regional climate shown in tree-ring studies, and indicated thattemperature increases are not always accompanied by an in-creased wildfire activity. In contrast, Flannigan et al. (2009) notedthat temperature was the most important predictor of area burnedin Canada and Alaska, with future warmer temperatures associatedwith further increases in the area burned. Parisien et al. (2011) alsoshowed that spatial pattern in temperature extremes was the mostimportant determinant of patterns of area burned in boreal Can-ada. According to the simulation model recently developed byPechony and Shindell (2010), while the global wildfire activitywas governed by precipitation in the past, a temperature-drivenglobal wildfire regime is projected in the near future (�2050)resulting from unprecedented increases in temperature. Thus, fu-ture probability of drought-induced wildfires is expected to be lar-gely influenced by the degree of anthropogenic warming (Terrieret al., 2013).

Regardless of the definition of the extent of fire, we foundincreasing trends in the probabilities of large-scale fires with time(Figs. 5 and 6) in all areas, except for the northeastern ecodistricts.The northeastern ecodistricts are at high elevations in the RockyMountains with relatively larger percentages of glaciated and al-pine terrains and lower drought levels (Table 1). It should be nota-ble that, by the 2080s, there is a probability of some areas havingmore than 50% of large-scale wildfires under the ‘‘average’’ climaticconditions in the future (Figs. 5 and 6). This means, even without‘‘extreme’’ weather conditions, some ecosystems will have a funda-mental probability of experiencing catastrophic fires under thecondition of average summer in a given period (especially in the2080s). However, the rate of progression toward a fire-prone con-dition is quite different among the three scenarios (Figs. 5–7); bythe 2050s, climate change effects became profound under A1B,but not under A2 and B1 (Table 3). Also, variations of wildfire prob-ability were often large, especially under the B1 scenario (Figs. 5and 6). Our result that wildfire susceptibility is scenario-dependentreminds us that a future wildfire activity will be largely influencedby human activities which will determine the degree of changes infuture climate. In the Columbia Mountains Ecoregion, the NorthernEcodistrict was not always more prone to widespread wildfiresthan the Central Ecodistrict (Fig. 5); this highlights the fact that firemanagement based on an ecoregion-scale may still miss importantecosystem-specific properties. Given such scenario-sensitive, spa-tially-heterogeneous patterns of wildfire probability in responseto climate variability, management strategy should be flexibleand more localized. Considering that wildfire probabilities in thenear future will increase, associated costs of fire management,physical/economic limitations, and protection efforts may needto be focused on high-value areas and resources (Flannigan et al.,2009). In this regard, projection of wildfire probability specific toeach ecodistrict will help in assessing localized risks posed by in-creased wildfires, and in choosing priorities in order to mitigatethe impacts of the changing wildfire regime.

The three scenarios provide different rates of increase indrought intensity (Fig. 7). Although the drought level observed inthe 1910–30s cannot be regarded as an absolute threshold, our re-sults of more severe droughts by the 2050s in large areas (Table 3)suggest a high probability of broad-scale shifts in wildfire regimesunder the A1B scenario. Compared with the gradual drying pro-jected by B1 (Fig. 7), this rapid change makes it more difficult forboth humans and the biota to mitigate for a disturbance regime

characterized by a higher frequency, probably more severity, anda larger spatial extent than currently. In this scenario (A1B), thepriorities of fire management should be urgently placed on thewestern ecodistricts which are more prone to severe droughts(Fig. 7), especially on the southwestern ecodistricts such as theCentral Columbia Mountains and the Selkirk Foothills that experi-enced a wildfire regime shift in the last century (Fig. 4; Table 3). Inthese areas, given high probability of wildfires under the averageclimate conditions expected from the GCMs (Figs. 5 and 6), ex-treme summer conditions would be no longer a prerequisite forinducing catastrophic fires. Under such future conditions, proactivefire management such as a fuel treatment would be far more effec-tive than reactive approach such as fire suppression (also see deGroot et al., 2013), if land management is primarily oriented to-ward the reduction of risks of catastrophic fires. By taking into con-sideration human communities, infrastructures, vegetationproperties, and the historical disturbance regime, an urgent anddifficult decision is needed on whether to permit natural fires thatmay spread at unprecedented scales in the drier future.

In the wildfire modeling of Pechony and Shindell (2010), the A2scenario yielded the warmest future with more wildfires when fo-cused on the western United States, while the model results for theA1B and B1 scenarios led to greater biomass burning when ana-lyzed at the global scale. Such inconsistency between differentscales implies the need of multi-scaled mitigation for the fire-prone future. Furthermore, we found the large discrepancy be-tween the A2 and B1 scenarios in the future drought levels in eachecodistrict (Table 3), indicating the high uncertainty of future re-sponses at the local scale; for instance, Luscar Foothills, BanffMountains, Icefield Mountains. Selkirk Foothills, Upper FraserTrench, Big Bend Trench, East Kootenay Trench, Morley Foothills,and Crowsnest Mountains would be prone to wildfires and mightexperience a possible wildfire regime shift by the 2080s underA2, while such probability in these ecodistricts is very low underthe B1 scenario (Table 3). These areas need further consideration,because of the high variability depending on the scenario.

4.4. Limitations and implications

Despite the fact that human influences on climate and wide-spread wildfires are mainly recent in North America (Williamsand Bradstock, 2008), they have caused significant variability anduncertainty in the climate–wildfire relationship (Pechony andShindell, 2010). Note that, although our study dealt with humaninfluences in terms of anthropogenic warming, other factors suchas fire suppression and fuel management were not directly consid-ered in this study, likely producing the future uncertainty. Never-theless, this study highlights that future wildfire responses toclimate are spatially heterogeneous at the finer geographical-scale,regardless of the climate scenarios. Therefore, in coping withuncertainty associated with a changing climate, it is at the local-ized level that future scenarios are meaningful to decision-makersand the practice of ecological management.

Landscape management can be often informed from the evalu-ation of historical range of variability (HRV) (e.g., Wimberly, 2002;Mori and Lertzman, 2011; Grigulis et al., 2013). Here, it is notablethat the variability expected from the last century is not necessar-ily a useful reference. This is especially the case for areas with a rel-atively longer fire interval of stand-replacing fires. In other words,the possible shift in wildfire regimes projected for this centurydoes not necessarily indicate the deviation from the HRV. However,we found that widespread wildfires could become more commonphenomena (especially in the southwestern ecodistricts underthe A1B scenario). Given a high potential to see novel ecosystemsunder the changing climate (Williams and Jackson, 2007; Jacksonand Hobbs, 2009), management should be more flexible to face

A.S. Mori, E.A. Johnson / Forest Ecology and Management 310 (2013) 875–886 885

the uncertainties and surprises. Such management (so-called, resil-ience-based management, Chapin et al., 2009; Pakeman, 2011;Mori et al., 2013) helps managers and stakeholders to choose pri-orities in conserving ecosystems by considering the expected socialand ecological consequences of mitigating potential shifts in dis-turbance regimes. At this point, projective models such as shownhere could be useful, making it possible to prepare for the futurethat may be different from the present.

Acknowledgments

This study was funded by the Japan Society for the Promotion ofScience (JSPS) and by the Mitsui & Co., Ltd. Environment Fund.ASM’s implication for fire risk analyses is from the project at theYokohama National University under the support of the JSPS GlobalCOE Program (E03). This paper was written during ASM’s stay atthe University of Calgary under the researcher exchange programbetween the Natural Sciences and Engineering Research Councilof Canada (NSERC) and the JSPS. We acknowledge Simon Fukadaand Naoko Mori in helping with this research. We thank editorsand anonymous reviewers for their valuable comments in improv-ing our manuscript.

Appendix A. Supplementary material

Supplementary data associated with this article can be found, inthe online version, at http://dx.doi.org/10.1016/j.foreco.2013.09.036.

References

Albert, J., 2009. Bayesian Computation with R, Second Edition. Springer, New York,US.

Andersen, T., Carstensen, J., Hernandez-Garcia, E., Duarte, C.M., 2009. Ecologicalthresholds and regime shifts: approaches to identification. Trends Ecol. Evol. 24,49–57.

Bond-Lamberty, B., Peckham, S.D., Ahl, D.E., Gower, S.T., 2007. Fire as the dominantdriver of central Canadian boreal forest carbon balance. Nature 450, 89–92.

Bouchard, M., Kneeshaw, D., Bergeron, Y., 2008. Ecosystem management based onlarge-scale disturbance pulses: a case study from sub-boreal forests of westernQuebec (Canada). For. Ecol. Manage. 256, 1734–1742.

Bowman, D.M., Balch, J.K., Artaxo, P., Bond, W.J., Carlson, J.M., Cochrane, M.A.,D’Antonio, C.M., Defries, R.S., Doyle, J.C., Harrison, S.P., Johnston, F.H., Keeley,J.E., Krawchuk, M.A., Kull, C.A., Marston, J.B., Moritz, M.A., Prentice, I.C., Roos,C.I., Scott, A.C., Swetnam, T.W., van der Werf, G.R., Pyne, S.J., 2009. Fire in theearth system. Science 324, 481–484.

Carcaillet, C., Bergman, I., Delorme, S., Hornberg, G., Zackrisson, O., 2007. Long-termfire frequency not linked to prehistoric occupations in northern Swedish borealforest. Ecology 88, 465–477.

Chapin, F.S.I., Kofinas, G.P., Folke, C., 2009. Principles of Ecosystem Stewardship.Resilience-based Natural Resource Management in a Changing World. Springer,New York, US.

Colombaroli, D., Gavin, D.G., 2010. Highly episodic fire and erosion regime over thepast 2,000 y in the Siskiyou Mountains, Oregon. Proc. Natl. Acad. Sci. USA 107,18909–18914.

Dai, A., Trenberth, K.E., Qian, T., 2004. A global dataset of palmer drought severityindex for 1870–2002: relationship with soil moisture and effects of surfacewarming. J. Hydrometeorol. 5, 11171130.

Dale, V.H., Joyce, L.A., McNulty, S., Neilson, R.P., Ayres, M.P., Flannigan, M.D., Hanson,P.J., Irland, L.C., Lugo, A.E., Peterson, C.J., Simberloff, D., Swanson, F.J., Stocks, B.J.,Wotton, B.M., 2001. Climate change and forest disturbances. BioScience 51,723–734.

de Groot, W., Field, R., Brady, M., Roswintiarti, O., Mohamad, M., 2007. Developmentof the Indonesian and Malaysian Fire Danger Rating Systems. MitigationAdaptation Strat. Global Change 12, 168–180.

de Groot, W.J., Flannigan, M.D., Cantin, A.S., 2013. Climate change impacts on futureboreal fire regimes. For. Ecol. Manage. 294, 35–44.

Falk, D.A., Miller, C., McKenzie, D., Black, A.E., 2007. Cross-scale analysis of fireregimes. Ecosystems 10, 809–823.

Flannigan, M.D., Logan, K.A., Amiro, B.D., Skinner, W.R., Stocks, B.J., 2005. Futurearea burned in Canada. Climatic Change 72, 1–16.

Flannigan, M., Stocks, B., Turetsky, M., Wotton, M., 2009. Impacts of climate changeon fire activity and fire management in the circumboreal forest. Global ChangeBiol. 15, 549–560.

Folke, C., Carpenter, S., Walker, B., Scheffer, M., Elmqvist, T., Gunderson, L., Holling,C.S., 2004. Regime shifts, resilience, and biodiversity in ecosystem management.Ann. Rev. Ecol. Evol. Syst. 35, 557–581.

Girardin, M.P., Wotton, B.M., 2009. Summer moisture and wildfire risks acrossCanada. J. Appl. Meteorol. Climatol. 48, 517–533.

Girardin, M.P., Ali, A.A., Carcaillet, C., Mudelsee, M., Drobyshev, I., HÉLy, C.,Bergeron, Y., 2009. Heterogeneous response of circumboreal wildfire risk toclimate change since the early 1900s. Global Change Biol. 15, 2751–2769.

Gonzalez, P., Neilson, R.P., Lenihan, J.M., Drapek, R.J., 2010. Global patterns in thevulnerability of ecosystems to vegetation shifts due to climate change. GlobalEcol. Biogeogr. 19, 755–768.

Grigulis, K., Lavorel, S., Krainer, U., Legay, N., Baxendale, C., Dumont, M., Kastl, E.,Arnoldi, C., Bardgett, R.D., Poly, F., Pommier, T., Schloter, M., Tappeiner, U., Bahn,M., Clément, J.-C., Hutchings, M., 2013. Relative contributions of plant traits andsoil microbial properties to mountain grassland ecosystem services. J. Ecol. 101,47–57.

Hallett, D.J., Hills, L.V., 2006. Holocene vegetation dynamics, fire history, lake leveland climate change in the Kootenay Valley, Southeastern British Columbia,Canada. J. Paleolimnol. 35, 351–371.

Jackson, S.T., 2006. Vegetation, environment, and time: the origination andtermination of ecosystems. J. Veg. Sci. 17, 549–557.

Jackson, S.T., Hobbs, R.J., 2009. Ecological restoration in the light of ecologicalhistory. Science 325, 567–569.

Johnson, E.A., Wowchuk, D.R., 1993. Wildfires in the southern Canadian RockyMountains and their relationship to mid-tropospheric anomalies. Can. J. For.Res. 23, 1213–1222.

Keane, R.E., Agee, J.K., Fule, P., Keeley, J.E., Key, C., Kitchen, S.G., Miller, R., Schulte,L.A., 2008. Ecological effects of large wildfires on US landscapes: benefit orcatastrophe? Int. J. Wildland Fire 17, 696–712.

Kitzberger, T., Brown, P.M., Heyerdahl, E.K., Swetnam, T.W., Veblen, T.T., 2007.Contingent Pacific-Atlantic Ocean influence on multicentury wildfire synchronyover western North America. Proc. Natl. Acad. Sci. 104, 543–548.

Lawler, J.J., Tear, T.H., Pyke, C., Shaw, M.R., Gonzalez, P., Kareiva, P., Hansen, L.,Hannah, L., Klausmeyer, K., Aldous, A., Bienz, C., Pearsall, S., 2010. Resourcemanagement in a changing and uncertain climate. Front. Ecol. Environ. 8, 35–43.

Lawson, B.D., Armitage, O.B., 2008. Weather guide for the Canadian forest firedanger rating system. Natural resources Canada, Canadian Forest Service,Northern Forestry Centre. Edmonton, Alberta, Canada.

Lindenmayer, D., 2004. Salvage harvesting policies after natural disturbance.Science 303, 1303.

Lindenmayer, D.B., Noss, R.F., 2006. Salvage logging, ecosystem processes, andbiodiversity conservation. Conserv. Biol. 20, 949–958.

Littell, J.S., McKenzie, D., Peterson, D.L., Westerling, A.L., 2009. Climate and wildfirearea burned in western U.S. ecoprovinces, 1916–2003. Ecol. Appl. 19, 1003–1021.

Liu, Z., Yang, J., Chang, Y., Weisberg, P.J., He, H.S., 2012. Spatial patterns and driversof fire occurrence and its future trend under climate change in a boreal forest ofNortheast China. Global Change Biol. 18, 2041–2056.

Lunn, D.J., Thomas, A., Best, N., Spiegelhalter, D., 2000. WinBUGS – a Bayesianmodelling framework: concepts, structure, and extensibility. Stat. Comput. 10,325–337.

Fauria, Macias, M., Johnson, E.A., 2006. Large-scale climatic patterns control largelightning fire occurrence in Canada and Alaska forest regions. J. Geophys. Res.111.

Macias Fauria, M., Johnson, E.A., 2008. Climate and wildfires in the North Americanboreal forest. Philos. Trans. R. Soc. London B 363, 2317–2329.

Marshall, I.B., Schut, P.H., Ballard, M., 1999. A national ecological framework forCanada: Attribute data. In. Agriculture and Agri-Food Canada, Research Branch,Centre for Land and Biological Resources Research, and Environment Canada,Ottawa, Canada.

Masters, P.I.P., 1990. Changes in forest fire frequency in Kootenay National Park,Canadian Rockies. Can. J. Bot. 68, 1763–1767.

Mastrandrea, M.D., Tebaldi, C., Snyder, C.W., Schneider, S.H., 2011. Current andfuture impacts of extreme events in California. Climatic Change 109, 43–70.

Mbogga, M., Hamann, A., Wang, T., 2009. Historical and projected climate data fornatural resource management in western Canada. Agri. Forest Meteorol. 149,881–890.

McCarthy, M.A., 2007. Bayesian Methods for Ecology. Cambridge University Press,New York, US.

Metsaranta, J.M., Kurz, W.A., Neilson, E.T., Stinson, G., 2010. Implications of futuredisturbance regimes on the carbon balance of Canada’s managed forest (2010–2100). Tellus B 62, 719–728.

Meyn, A., Taylor, S.W., Flannigan, M.D., Thonicke, K., Cramer, W., 2010. Relationshipbetween fire, climate oscillations, and drought in British Columbia, Canada,1920–2000. Global Change Biol. 16, 977–989.

Millar, C.I., Stephenson, N.L., Stephens, S., 2007. Climate change and forests of thefuture: managing in the face of uncertainty. Ecol. Appl. 17, 2145–2151.

Miller, J.D., Safford, H.D., Crimmins, M., Thode, A.E., 2008. Quantitative evidence forincreasing forest fire severity in the Sierra Nevada and Southern CascadeMountains, California and Nevada, USA. Ecosystems 12, 16–32.

Morgan, P., Heyerdahl, E.K., Gibson, C.E., 2008. Multi- season climate synchronizedforest fires throughout the 20th century, Northern Rockies, USA. Ecology 89,717–728.

Mori, A.S., 2011a. Climatic variability regulates the occurrence and extent of largefires in the subalpine forests of the Canadian Rockies. Ecosphere 2. Art 7.

886 A.S. Mori, E.A. Johnson / Forest Ecology and Management 310 (2013) 875–886

Mori, A.S., 2011b. Ecosystem management based on natural disturbances:hierarchical context and non-equilibrium paradigm. J. Appl. Ecol. 48, 280–292.

Mori, A.S., Lertzman, K.P., 2011. Historic variability in fire-generated landscapeheterogeneity of subalpine forests in the Canadian Rockies. J. Veg. Sci. 22, 45–58.

Mori, A.S., Spies, T.A., Sudmeier-Rieux, K., Andrade, A., 2013. Reframing ecosystemmanagement in the era of climate change: issues and knowledge from forests.Biol. Conserv. 165, 115–127.

Moritz, M.A., Parisien, M.-A., Batllori, E., Krawchuk, M.A., Van Dorn, J., Ganz, D.J.,Hayhoe, K., 2012. Climate change and disruptions to global fire activity.Ecosphere 3. Art49.

Mueller, R.C., Scudder, C.M., Porter, M.E., Talbot Trotter, R., Gehring, C.A., Whitham,T.G., 2005. Differential tree mortality in response to severe drought: evidencefor long-term vegetation shifts. J. Ecol. 93, 1085–1093.

Nitschke, C.R., Innes, J.L., 2008. Climatic change and fire potential in South-CentralBritish Columbia, Canada. Global Change Biol. 14, 841–855.

Noss, R.F., Franklin, J.F., Baker, W.L., Schoennagel, T., Moyle, P.B., 2006. Managingfire-prone forests in the western United States. Front. Ecol. Envion. 4, 481–487.

Pakeman, R.J., 2011. Multivariate identification of plant functional response andeffect traits in an agricultural landscape. Ecology 92, 1353–1365.

Parisien, M.-A., Parks, S.A., Krawchuk, M.A., M.D.Flannigan, Bowman, L.M., Moritz,M.A., 2011. Scale-dependent controls on the area burned in the boreal forest ofCanada, 1980–2005. Ecol. Appl. 21, 789–805.

Pechony, O., Shindell, D.T., 2010. Driving forces of global wildfires over the pastmillennium and the forthcoming century. Proc. Natl. Acad. Sci. USA 107, 19167–19170.

Peters, D.P., Pielke Sr., R.A., Bestelmeyer, B.T., Allen, C.D., Munson-McGee, S.,Havstad, K.M., 2004. Cross-scale interactions, nonlinearities, and forecastingcatastrophic events. Proc. Natl. Acad. Sci. 101, 15130–15135.

Rodionov, S.N., 2004. A sequential algorithm for testing climate regime shifts.Geophys. Res. Lett. 31, L09204.

Scheffer, M., Carpenter, S.R., 2003. Catastrophic regime shifts in ecosystems: linkingtheory to observation. Trends Ecol. Evol. 18, 648–656.

Schoennagel, T., Veblen, T.T., Romme, W.H., 2004. The interaction of fire, fuels, andclimate across rocky mountain forests. BioScience 54, 661–676.

Schoennagel, T., Waller, D.M., Turner, M.G., Romme, W.H., 2005. ENSO and PDOvariability affect drought-induced fire occurrence in rocky mountain subalpineforests. Ecol. Appl. 15, 2000–2014.

Seidl, R., Fernandes, P.M., Fonseca, T.F., Gillet, F., Jönsson, A.M., Merganicová, K.,Netherer, S., Arpaci, A., Bontemps, J.-D., Bugmann, H., González-Olabarria, J.R.,Lasch, P., Meredieu, C., Moreira, F., Schelhaas, M.-J., Mohren, F., 2011a.Modelling natural disturbances in forest ecosystems: a review. Ecol. Model222, 903–924.

Seidl, R., Schelhaas, M.-J., Lexer, M.J., 2011b. Unraveling the drivers of intensifyingforest disturbance regimes in Europe. Global Change Biol. 17, 2842–2852.

Spugel, D.G., 1991. Disturbance, equilibrium, and environmental variability:what is ‘natural’ vegetation in a changing environment?. Biol. Conserv. 51,1–18.

Stocks, B.J., Mason, J.A., Todd, J.B., Bosch, E.M., Wotton, B.M., Amiro, B.D., Flannigan,M.D., Hirsch, K.G., Logan, K.A., Martell, D.L., Skinner, W.R., 2002. Large forestfires in Canada, 1959–1997. J. Geophys. Res. 108.

Strong, W., Zoltai, S.C., 1989. Ecoclimatic regions of Canada, First approximation. In,Ecological land classification series, Ottawa, Canada..

Sturtz, S., Ligges, U., Gelman, A., 2005. R2WinBUGS: a package for running WinBUGSfrom R. J. Stat. Software 12, 1–16.

Suding, K.N., Gross, K.L., Houseman, G.R., 2004. Alternative states and positivefeedbacks in restoration ecology. Trends Ecol. Evol. 19, 46–53.

Terrier, A., Girardin, M.P., Perie, C., Legendre, P., Bergeron, Y., 2013. Potentialchanges in forest composition could reduce impacts of climate change on borealwildfires. Ecol. Appl. 23, 21–35.

Turner, M.G., Baker, W.L., Peterson, C.J., Peet, R.K., 1998. Factors influencingsuccession: lessons from large, infrequent natural disturbances. Ecosystems 1,511–532.

Turner, M.G., Rome, W.H., Tinker, D.B., 2003. Surprises and lessons from the 1988Yellowstone fires. Front. Ecol. Environ. 1, 351–358.

Wallington, T.J., Hobbs, R.J., Moore, S.A., 2005. Implications of current ecologicalthinking for biodiversity conservation: a review of the salient issues. Ecol. Soc.10. Art15.

Wang, T., Hamann, A., Spittlehouse, D.L., Murdock, T.Q., 2012. ClimateWNA—High-resolution spatial climate data for western North America. J. Appl. Meteorol.Climatol. 61, 16–29.

Westerling, A.L., Hidalgo, H.G., Cayan, D.R., Swetnam, T.W., 2006. Warming andearlier spring increase western U.S. forest wildfire activity. Science 313, 940–943.

Westerling, A.L., Bryant, B.P., Preisler, H.K., Holmes, T.P., Hidalgo, H.G., Das, T.,Shrestha, S.R., 2011. Climate change and growth scenarios for Californiawildfire. Climatic Change 109, 445–463.

Wiken, E.B., 1986. Terrestrial Ecozones of Canada. In, Ecological land classificationseries. Environment Canada, Hull, Quebec..

Williams, R.J., Bradstock, R.A., 2008. Large fires and their ecological consequences:introduction to the special issue. Int. J. Wildland Fire 17, 685–687.

Williams, J.W., Jackson, S.T., 2007. Novel climates, no-analog communities, andecological surprises. Front. Ecol. Environ. 5, 475–482.

Wimberly, M.C., 2002. Spatial simulation of historical landscape patterns in coastalforests of the Pacific Northwest. Can. J. For. Res. 32, 1316–1328.

Zeileis, A., Kleiber, C., Krämer, W., Hornik, K., 2003. Testing and dating of structuralchanges in practice. Comput. Stat. Data Anal. 44, 109–123.

Zumbrunnen, T., Bugmann, H., Conedera, M., Bürgi, M., 2008. Linking forest fireregimes and climate—A historical analysis in a dry Inner alpine valley.Ecosystems 12, 73–86.