This content has been downloaded from IOPscience. Please scroll down to see the full text.

Download details:

IP Address: 182.73.193.34

This content was downloaded on 22/07/2016 at 11:33

Please note that terms and conditions apply.

The spatially varying influence of humans on fire probability in North America

View the table of contents for this issue, or go to the journal homepage for more

2016 Environ. Res. Lett. 11 075005

(http://iopscience.iop.org/1748-9326/11/7/075005)

Home Search Collections Journals About Contact us My IOPscience

iopscience.iop.org/page/termshttp://iopscience.iop.org/1748-9326/11/7http://iopscience.iop.org/1748-9326http://iopscience.iop.org/http://iopscience.iop.org/searchhttp://iopscience.iop.org/collectionshttp://iopscience.iop.org/journalshttp://iopscience.iop.org/page/aboutioppublishinghttp://iopscience.iop.org/contacthttp://iopscience.iop.org/myiopscience

Environ. Res. Lett. 11 (2016) 075005 doi:10.1088/1748-9326/11/7/075005

LETTER

The spatially varying influence of humans on fire probability in NorthAmerica

Marc-André Parisien1,4, CarolMiller2, SeanAParks2, EvanRDeLancey1,3, François-Nicolas Robinne3 andMikeDFlannigan3

1 Northern Forestry Centre, Canadian Forest Service, Natural Resources Canada, 5320-122nd Street, Edmonton, ABT6H3S5, Canada2 Aldo LeopoldWilderness Research Institute, RockyMountain Research Station, USDAForest Service, 790 EBeckwithAvenue,Missoula,

MT59801,USA3 Western Partnership forWildland Fire Science, University of Alberta, Department of Renewable Resources, 751General Services

Building, Edmonton, ABT6G2H1, Canada4 Author towhomany correspondence should be addressed.

E-mail:marc-andre.parisien@canada.ca

Keywords:wildfire, NorthAmerica, anthropogenic influence, climate, topography, lightning

AbstractHumans affectfire regimes byproviding ignition sources in some cases, suppressingwildfires in others,and altering natural vegetation inways thatmay either promote or limitfire. InNorthAmerica, severalstudies have evaluated the effects of society onfire activity; however,most studies have been regional orsubcontinental in scope anduseddifferent data andmethods, therebymaking continent-widecomparisons difficult.We circumvent these challenges by investigating the broad-scale impact ofhumans onfire activity using parallel statisticalmodels offire probability from1984 to 2014 as a functionof climate, enduring features (topography andpercent nonfuel), lightning, and three indices of humanactivity (populationdensity, an integratedmetric of humanactivity [HumanFootprint Index], and ameasure of remoteness [roadless volume]) across equally spaced regions of theUnited States andCanada.Through a statistical control approach,wherebywe account for the effect of other explanatoryvariables,we found evidence of non-negligible human–wildfire association across the entirecontinent, even in themost sparsely populated areas. A surprisingly coherent negative relationshipbetweenfire activity andhumanswas observed across theUnited States andCanada:fire probabilitygenerally diminisheswith increasinghuman influence. Intriguing exceptions to this relationship arethe continent’s least disturbed areas,where fewer humans equate to lessfire. These remote areas,however, also often have lower lightning densities, leadingus to believe that theymay be ignitionlimited at the spatiotemporal scale of the study.Our results suggest that there are fewpurely naturalfire regimes inNorthAmerica today.Consequently, projections of futurefire activity shouldconsider human impacts onfire regimes to ensure soundadaptation andmitigationmeasures infire-prone areas.

Introduction

Humans affect fire regimes worldwide, sometimescausing profound changes to ecosystem structure andfunction (McWethy et al 2013). In many parts of theworld, people increase fire activity beyond what isnaturally expected by deliberately or accidentallysetting fire to natural vegetation (Carcaillet et al 2009,Bowman et al 2011). This is the case, for example, inAfrican savannas, where cultural practices of frequent

burning make it the archetype of a ‘human-tended’fire regime (Archibald et al 2012). Conversely, in otherregions or ecosystems of the world, such as in parts ofNorth America, land-use change and a culture ofaggressive fire suppression may limit fire activity(Cumming 2005, Finney et al 2009). Globally, humansare usually responsible for the majority of ignitions;however, many regions where human ignitions areprevalent are also subject to intense fire managementprograms aimed at extinguishing fires before they

OPEN ACCESS

RECEIVED

29 January 2016

REVISED

10 June 2016

ACCEPTED FOR PUBLICATION

6 July 2016

PUBLISHED

20 July 2016

Original content from thisworkmay be used underthe terms of the CreativeCommonsAttribution 3.0licence.

Any further distribution ofthis workmustmaintainattribution to theauthor(s) and the title ofthework, journal citationandDOI.

© 2016 IOPPublishing Ltd

http://dx.doi.org/10.1088/1748-9326/11/7/075005mailto:marc-andre.parisien@canada.cahttp://crossmark.crossref.org/dialog/?doi=10.1088/1748-9326/11/7/075005&domain=pdf&date_stamp=2016-07-20http://crossmark.crossref.org/dialog/?doi=10.1088/1748-9326/11/7/075005&domain=pdf&date_stamp=2016-07-20http://creativecommons.org/licenses/by/3.0http://creativecommons.org/licenses/by/3.0http://creativecommons.org/licenses/by/3.0

become large (Stocks et al 2002, Stephens 2005). Inaddition to these direct effects on fire activity, humansmay exert pervasive and enduring indirect effects onfire regimes by altering, or removing, the naturalvegetation (e.g., change in forest structure, invasiveannuals, logging, conversion to irrigated agriculture)(Cochrane 2003, Pausas and Keeley 2009, Perryet al 2012, Fréjaville andCurt 2015).

Several investigations of wildland fire in NorthAmerica show a dramatic effect on fire activity followingthe arrival of Europeans. An initial surge in area burnedat the time of settlement due to slash-and-burn practicesin some areas (Weir and Johnson 1998)was followed bya large-scale reduction in fire activity that was largely theresult of fire suppression and human modifications tothenatural vegetation cover (Marlon et al2008).Over thepast three decades, however, substantial increases inwildfires have been reported, notably in thewesternUni-ted States (Westerling et al 2006) and in the boreal zone(Kasischke and Turetsky 2006). Although this phenom-enon has been largely attributed to a warmer and drierclimate, it is increasingly recognized that, in some areas,humans may have played a role (Stephens et al 2007,McWethy et al 2013). Paradoxically, some of the recentincreases in fire activity in parts of the continent, notablyin the west, may be due to a legacy of fire managementpolicies that were too successful (Stephens andRuth 2005). Almost a century of active fire suppressionhas led to unintended biomass accumulation that conse-quently increased the average size and intensity, aswell asecological impact, of modern fires (Keane et al 2002).The impact of humans on fire is thus inherently com-plex: some human activities yield more fire, whereas theopposite is true of others (Sturtevant and Cleland 2007,Parks et al 2015), thereby challenging our assessment ofthe net effect of humans on wildland fire activity inNorthAmerica and elsewhere.

In North America, the influence of humans on fireactivity is assumed to be highly variable, ranging fromextreme in areas of high population and altered land useto seemingly negligible in unmanaged wildlands. Areaswith historically active fire regimes (whether natural oranthropogenic), notably in the northeastern UnitedStates and adjacent Canada, now experience virtually nowildland fire (Clark and Royall 1996, Nowacki andAbrams 2008). In contrast, the heavily populated regionsof southeastern North America, where a culture of pre-scribed burning has existed for centuries, have highlyactive fire regimes that are largely human dominated(Waldrop et al 1992, Slocum et al 2007). In the center ofthe continent, in the American Midwest and CanadianPrairies, where vast grasslands were converted to agri-culture, previous high fire activity has been greatlydiminished or even excluded altogether (Collins andWallace 1990, Brown et al 2005). Currently, most fireactivity in North America occurs in the western UnitedStates and in the boreal forests of Canada and Alaska,where, in spite of large pockets of urbanization and agri-culture, fire is promoted by expansive wildlands

combined with fire-conducive climates and vegetation(Swetnam1993,Westerling et al2003).However, thoughquite active, it has been argued that fire regimes in mostforests of western North America are far from natural,having been shaped by a century of fire suppression andforestmanagement (Stephens and Ruth 2005). Themostnatural fire regimes of North America are presumed tooccur in the northern forests of Canada and Alaska(Flannigan et al 2009). Not only are these regions sparselypopulated, but, given their regimes of low-frequency,high-intensity fires, some have suggested that fire sup-pression is less effective here than it is in other foresttypes (Johnson et al2001).

Most evaluations of the influence of human activityon North American fire regimes have been conductedfor specific countries, subregions, or landscapes (e.g.,Cardille et al 2001, Sturtevant and Cleland 2007,Syphard et al 2007, Gralewicz et al 2012, Hawbakeret al 2013, Liu and Wimberly 2015). Moreover, pub-lished studies have been carried out at various temporaland spatial scales and using different data types andmodeling techniques, thereby preventing an unbiasedcomparison among regions (but see Bistinas et al 2013).A continent-wide assessment that uses a consistentapproach would complement previous efforts with theadded benefit of providing a baseline for large-scale pol-icy or management plans. As such, the overarching goalof this studywas to conduct a comprehensive evaluationof the effect of humans on fire activity across NorthAmerica. Our specific objectives were to (1) assess theimportance of human influence on fire activity and (2)evaluate the shape of this relationship (positive, negative,unimodal). Thiswas achievedby partitioning theUnitedStates and Canada into 16 hexagonal polygons(size=1.38×106 km2) and, for each hexagonal poly-gon, creating a statisticalmodel offire probability (basedon area burned) for the 1984–2014 time period as afunction of climate, enduring features (topography andpercent nonfuel), lightning, and indices of anthro-pogenic influence. Thesemodels allowed us to (1) statis-tically control for the effect of other factors whenassessing the influence of humans on fire activity and (2)evaluate the spatial variability in these relationships at acontinental extent.

Methods

The study area covers theUnited States (US) andCanada(hereafter North America; 19 437 101 km2) (figure 1(a)),excluding Hawaii and the Caribbean islands. Full cover-age ofNorthAmericawas not possible due to lack of datafor Mexico. The study area was partitioned into 16hexagonal polygons (hereafter hexels) (hex 1 to hex 16)having a width of 1260 km and an area of 1 374 902 km2

(figure 1(b)). Explorations showed that this size andplacement of hexels represent a good tradeoff betweencapturing subcontinental variability in fire activity andproducing robust models of fire activity across the study

2

Environ. Res. Lett. 11 (2016) 075005

area. The response variable is sampled within burnedareas, and consequently themodeled response quantifiesthe probability that any given pixel burned over the1984–2014 time period (hereafter fire probability). Foreach hexel, we built a statistical model of fire probabilityas a function of climatic normals, enduring features(topography and nonfuel), lightning, and measures ofanthropogenic influence.

The fire probability models were fully comparableamong hexels, as they were built using the same set ofexplanatory variables and used the same model settings.We evaluated the importance of indices of anthro-pogenic influence to determine the human impact onfire activity by controlling (statistically) for the effect ofclimate, enduring features, and lightning patterns ineach hexel. We then plotted the relationship of each ofthese metrics to assess whether fire probabilitydecreased, increased, or had a nonlinear (e.g., unimodal)relationship to the anthropogenic variables in eachhexel.

Data

Models of fire probability were built as a function ofseveral explanatory variables likely to influence fireactivity for each hexel. Although a large set ofindependent variables was initially considered formodeling, we selected a subset of variables that were

not highly correlated with one another and also hadgood explanatory power for predicted area burnedacross North America (table 1; figure B1). The descrip-tion that follows focuses on the variables selected forthe modeling. The full list of variables is found intable A1. We processed all model variables using aNorth America Albers equal-area conic projection at a1 kmpixel resolution.

FireThe US fire data were obtained from the MonitoringTrends in Burn Severity (MTBS) project (Eidenshinket al 2007), whereas Canadian data were obtained fromthe Canadian Forest Service National Fire Database(Parisien et al 2006). Because only fires 400 ha areincluded in the western US, this size threshold wasapplied to the entire study area; fires

Fire occurrence, as integrated in the model, consistsof points randomly sampled within fire perimeters ofmapped fires 400 ha from 1984 to 2014. We con-sidered areas as either burned or unburned over the 31year time period and did not distinguish areas that mayhave burned more than once. This approach wasdeemed themost suitable given the lowproportion of re-burned areas during the study time period (∼8% of thetotal area burned). The points were sampled within thepolygon data (rather than a raster dataset); therefore,therewasno loss of spatial resolution in this process.

ClimateA suite of 30 year climate variables describing gradientsof energy (temperature) and moisture (precipitationand humidity) for the 1981–2010 time period wasgenerated. This time period represents a slight mis-match to that of the fire data (1984–2014), but is notexpected to greatly influence the results. The ClimateWNA software (Wang et al 2012) was used to inter-polate the climate data from a digital elevation model.Of the climate variables selected for modeling, twodepicted patterns of moisture: mean annual

precipitation (MeanPrecip) and mean annual relativehumidity (RelHumid). One variable was a measure oftemperature, the degree days under 18 °C (Deg-DaysU18), and another, the climate moisture index(CMI), was an annual integrativemeasure of energy andmoisture (precipitation–potential evapotranspiration).

Enduring featuresEnduring features are those components of the landscapethat vary little, if at all, during the time scale of the study.Elevation-derived metrics describing topography com-prisedmost of this category of variables. Heat load index(HeatLoadIndex), which is a measure of potential sunexposure, and surface area ratio (SurfAreaRatio), whichis a measure of surface roughness, were computed withGeomorphometry and Gradient Metrics Toolbox 2.0(Evans et al2014). Topographic position index (TopoPo-sIndex) describes the relative position along a valley-to-peak gradient; this metric was computed with a 2000mwindow. Percent permanent nonfuel (PctNonfuel),which included openwater, glaciers, barren ground, andurban areas from the 2005 global land cover (Friedlet al 2010), was computed at 100 and 10 000 km2 scales

Table 1.Description of variables used in the statisticalmodeling offire probability in theUnited States andCanada. The time period coveredis 1984–2014 for the area burned and 1981–2010 for climate variables. Themean and range values of pixels are given for each variable. Allvariables were resampled to a resolution of 1 km.

Variable name Description Units Mean (range)

Fire

Area burned USfiresmergedwithCanadian fires. USfires fromMTBS

database 1984–2014. Canadian fires fromCanadian

National Fires database 1984–2014. Fires over

400 ha used.

ha 5311 (400–2 205 060)

Climate (1981–2010 normals)CMI Hargreave’s climaticmoisture index Dimensionless 299.4 (0–1754)DegDaysU18 Degree days under 18 °C Degree days 5043 (0–11 539)MeanPrecip Mean annual precipitation mm 722 (50–10 149)RelHumid Mean annual relative humidity (%) 59.6 (36–85)

Enduring

HeatLoadIndex Heat load index, an index calculating the southwestness of a

slope

Dimensionless 0.64 (0.096–1.03)

SurfAreaRatio Surface area ratio Dimensionless 1.01 (1–1.40)TopoPosIndex Topographic position index calculated at 2000 m scale Dimensionless –0.34 (–489–843)PctNonfuel Percent nonfuel at a 1000 km2movingwindow size. Non-

fuel classified as barren lands, water, snow, and ice. From

2005 land cover

% 4.5 (0–99.7)

Lightning

Lightning Average number of lightning strikes per year from1995

to 2005

Flashes km–2 year–1 5.9 (0–46.4)

Anthropogenic

HumanFoot Human Footprint Index, an index of human influence for

the year 2004. Calculated from a 100 km2moving

window.

Dimensionless 16.5 (0–98.1)

LogPopDens Log of population density for the year 2000 People km–2 –0.51 (–4 to 4.3)RdlsVol Roadless volume calculatedwith a 10 000 km2moving

window

km3 6.5 (0.067–100)

4

Environ. Res. Lett. 11 (2016) 075005

using moving window averages at each of the radii toaccount for possible scale-dependent effects (Parisienet al 2012); PctNonfuel at the 10 000 km2 scale wasretained formodeling.

LightningPatterns of lightning were used to describe naturalignition potential across the study area. The density ofcloud-to-ground lightning strikes for the 1995–2005time period was obtained from the NASA LIS/OTD0.5° high-resolution annual climatology dataset(Christian et al 2003). In spite of the lightning datasethaving only partial overlap with the time period of thestudy and a coarser resolution than the other variables,it was deemed acceptable for depicting lightning strikepatterns at the spatial scale of this study.

Anthropogenic influenceThree anthropogenic variables were considered in ourevaluation of the influence of humans on fire activityin North America. Population density (log10-trans-formed; LogPopDens) was obtained from the Centerfor International Earth Information Network (Balket al 2006). High values of this metric are concentratedin urban areas and, as such, it effectively separatesurban versus rural (including wild) areas. The HumanFootprint Index (HumanFoot) is an integrated indexof human influence derived from roads, urban areas,and nighttime light (Sanderson et al 2002). Because itis based on a range of features, this index reveals moreabout human activities than LogPopDens. We evalu-ated LogPopDens and HumanFoot variables at the1 km2 (resolution of raw data) and 100 km2 scales(moving window average); only the best-performingscale was selected. Roadless volume (RdlsVol; Wattset al 2007) is a metric of isolation from humaninfluence computed from road data (ESRI 2008). Itrepresents the product of the footprint (area of cellscontaining roads) and the mean distance to roads andwas computed at 100 and 10 000 km2 scale. Thoughsimilar to HumanFoot, RdlsVol is simpler and betteremphasizes remoteness. Road density was initiallyconsidered but subsequently dropped because of itsstrong correlationwith RdlsVol.

Variable selectionA subset of 12 variables from the initial pool ofexplanatory variables (table A1)was selected formodelbuilding according to their degree of correlation andability to predict fire activity (table 1; figure B1). First,correlations were computed between each pair ofvariables to identify those correlated at |r|>0.7 acrossthe study area and within each hexel. Althoughcorrelations varied among hexels, explorationsshowed that a hexel-wise selection yielded similarvariable sets than for the entire study area. Thisvariable selection approach thus provides a good

trade-off between model performance and the abilityto compare results among hexels. In each grouping ofcorrelated variables, a single variable was chosenaccording to its ability to explain fire activity inbivariate MaxEnt models of fire according to the areaunder the curve (AUC) metric (see Statistical model-ing). Because we were primarily interested in humaninfluences on fire activity, LogPopDens, HumanFoot,and RdlsVol were retained in spite of correlations|r|>0.7. Of these variables, LogPopDens and RdlsVolwere correlated to the DegDaysU18 climate variable atr=0.80 and r=−0.72, respectively. Anthropogenicvariables were also correlated with one another:LogPopDens–HumanFoot at r=0.86; LogPopDens–RdlsVol at r=−0.78; and HumanFoot–RdlsVolat r=−0.75.

Statistical modelingStatistical models of fire probability using 12 explana-tory variables (table 1) were produced for each of the16 hexels. Models were built using MaxEnt v.3.3.3k(Phillips et al 2006), a machine-learning techniquedesigned for presence-only data (i.e., when trueabsences are unknown). A presence-only frameworkwas deemed adequate because a lack of fire need not beinterpreted as a true absence over the 31 years of firedata for which we conducted this study (i.e., manyareas that did not burn will likely burn in future years);however, presence-only or presence–absence frame-work have been shown to produce similar models offire probability (Parisien and Moritz 2009). MaxEntevaluates the environmental space of the pointssampled within the fire perimeters (‘fire presence’) incontrast to that of the entire environment of study area(‘background’). It does so by fitting the probabilitydistribution of maximum entropy to the environmen-tal variables at each fire-presence point. Thismodelingtechnique allowed us to model highly complex rela-tionships without overfitting the responses. MaxEntsettings were selected following the advice of Elith et al(2011) and through careful explorations in which theaim was to produce robust models for the ensemble ofhexels. To this end, we opted for a regularization valueof four and did not use the ‘hinge’ feature, which tendsto produce unrealistic (i.e., overfit) responses.

A large number of fire presences and backgroundpoints were sampled to ensure that the appropriatedegree of variability was captured for statistical model-ing. Points were randomly sampled in burned areas at adensity of 1 point per 5 km2 for fire presences and acrossthe hexel’s landmass at a rate of 1 point per 50 km2 forthe background. Points were never sampled within per-manent nonfuels (water, glacier, urban, rock). Because ofvariation in both area burned and hexel landmass, thenumber of sample points varied among hexels. Toreduce spatial autocorrelation in model residuals, a ran-dom subset of total fire-presence points were selected formodel building. One hundred such subsets were used

5

Environ. Res. Lett. 11 (2016) 075005

and their results were subsequently averaged (see Par-isien et al 2012). This strategy was appropriate across thestudy area because it limited overfitting while yieldingrobustmodels of area burned for eachhexel.

Four models of fire probability were produced foreach hexel. The first model (hereafter ‘full model’) wasbuilt using all 12 variables, and its output was used todetermine the importance of the four categories of vari-ables: climate, enduring features, lightning, and anthro-pogenic influence. The other three models (simplynamed after the variable of interest; e.g., LogPopDens)were created to highlight the singular effect of each of theanthropogenic variables. These three models were eachbuilt using a total of 10 explanatory variables, includingclimate, enduring features, lightning, and a singleanthropogenic variable. Thesemodels were used to eval-uate the response of area burned to each of the threeanthropogenic variables; the reason for incorporatingonly one of these variables was to avoid any distortion ofthe response due to the anthropogenic variables’ correla-tion toone another.

Model evaluation was performed on each of the100 subset models and subsequently averaged for eachhexel. As a measure of overall of model performance,the AUC was computed from the true positives andfalse positives (Liu et al 2005). AUC values may rangefrom0.5, where prediction accuracy is no better than ifsamples were randomly selected, to 1, which indicates

perfect classification accuracy. However, in a pre-sence-only framework, as in this study, it is impossibleto achieve an AUC value of 1 because absences (hencefalse positives) are unknown. The maximum achiev-able AUC in a presence-only framework is equal to1−a/2, where a is generally the fraction of the studyarea covered by fire (i.e., the prevalence). For the sakeof adjusting the AUC value, we considered a to be thepercentage of 1 km2 pixels where fire was observed.This provides a fair, yet underestimated, approx-imation of prevalence. Additional evaluation metricsare provided in table C1.

Assessment of anthropogenic influenceThe contribution of each of the four variable categories(climate, enduring features, lightning, and anthropo-genic influence) for each hexel was calculated from thefull model by summing the relative importance of thevariables in each category. Variable importance wascalculated as the model gain associated with eachvariable (Phillips et al 2006). Although it was impos-sible to compute an absolute measure of variableimportance, this could be approximated by comparingthe values to the model performance metrics (e.g.,AUC). The relationship of each anthropogenic vari-able by hexel was then plotted from the results of theLogPopDens, HumanFoot, and RdlsVol models. Weuse partial dependence plots, whichmeasure the effect

Figure 2.The relative contribution of variables for fullmodels of fire probability by hexel in each of the four variable categories:climate (C), enduring features (E), lightning (L), and anthropogenic influence (A). The value (in red) in the top-center of each hexelrepresents the adjusted area under the curve (seemethods). The y-axis is a percentage and is the same for all plots.

6

Environ. Res. Lett. 11 (2016) 075005

of a variable when the value of all other variables areheld constant at their means, to statistically control forthe effect of non-anthropogenic variables on fireprobability. To estimate the effect of this statisticalcontrol, bivariate relationships of fire probability as afunction of individual anthropogenic variables werealso produced.

Results

Overall, models of fire probability as a function of allvariables (i.e., the fullmodel)performedwell (figure 2).The test AUC averaged 0.776 among hexels, andranged from 0.672 (hex13) to 0.891 (hex8). AUCvalues increased, sometimes substantially, whenadjusted for prevalence (mean 0.804).

The relative importance of climate, enduring fea-tures, lightning, and anthropogenic variables variedsubstantially among hexels (figure 2). For all hexels,the two most influential categories of variables wereclimate and anthropogenic. The anthropogenic cate-gory was ranked as the most important in four hexels(hex2, hex7, hex9, hex12) and the second-mostimportant in the remaining 12 hexels. Climate was themost frequent dominant control on fire activity, hav-ing the highest variable importance in all but four hex-els (i.e., the ones dominated by anthropogenicvariables). Enduring features ranked third in terms ofimportance in all hexels except four (hex1, hex5, hex8,hex10), where this category was the least important

and lightning had the third-highest contribution.Lightning was otherwise the least important variable.The importance of each anthropogenic variable, cal-culated both from the full model and models with asingle anthropogenic variable, is reported by hexel intable 2. All three anthropogenic variables were usefulin explaining fire probability models, but their relativeimportance varied substantially among hexels.

The relationship between fire activity and humaninfluence varied as a function of the specific anthro-pogenic variable, but was overall fairly coherent amonghexels. Area burned usually decreased as a function ofincreasing LogPopDens, except in hexels where therewas a strong (hex8) or moderate unimodal response(hex1, hex2, and hex3) (figure 3). The HumanFoot vari-able had a decaying form in all hexels, though the rangeof slopes varied considerably (figure 4). The response tothe RdlsVol was the most diverse of the three anthro-pogenic variables (figure 5). Fire activity increasedmonotonically as a function of RdlsVol in some hexels(hex3, hex4, hex5, hex6, hex7, hex13), decreased insome hexels (hex10, hex12, hex15, hex16), and wasunimodal in others (hex3, hex8, hex9, hex14). Thebivariate relationship between fire activity and anthro-pogenic variable show broadly similar patterns to thepartial dependence plots (figureD1).

Discussion

Because of its extensive wildlands and low humandensities over much of its territory, North America isoften viewed as having fire regimes that are closer tonatural than those in other parts of the world (Lavorelet al 2007). Although this may hold true to some degree,Cardille and Lambois (2010) have shown in the con-terminous US that, even in seemingly ‘intact’ areas, veryfew landscapes are completely free of some humanimprint. The results of this study similarly suggest thatthe human impact on fire regimes is widespread andpervasive. Climate is nevertheless the dominant environ-mental control onfire inmost ofNorthAmerica, therebysupporting the claim that its fire activity today—as it hasbeen for millennia—is mainly regulated by top-downcontrols (Swain 1973, Flannigan et al 2005, Guyetteet al 2012). There are clearly, however, other ‘non-natural’ controls that exert an influence on fire, evenoverriding the effect of climate in some portions of thecontinent, as shown by Hawbaker et al (2013) in theconterminous US. We found a non-negligible impedinginfluence of humans on fire, even in some of the leastdensely populated areas. Though our results show aremarkably coherent negative influence of humans onfire, the specific nature—and hence outcome—of thehuman–fire relationship is far from uniform. A poten-tially interesting exception to the negative effect of peopleon fire is observed in some of North America’s most

Table 2.Relative contributionof the three anthropogenic variables formodels offire probability for eachhexel.Valueswere calculated accord-ing two tomodel types: one that incorporates only the anthropogenicvariable of interest (single) in combinationwith theother explanatoryvariables and theother (full) that incorporates all three anthropogenicvariables. LogPopDens, populationdensity (logged);HumanFoot,HumanFootprint Index;RdlsVol, roadless volume.

LogPopDens HumanFoot RdlsVol

Single Full Single Full Single Full

Hex1 30.9 11.8 4.43 5.25 29.5 6.05

Hex2 44.4 50.3 23.9 0.31 24.3 0.276

Hex3 4.65 7.97 2.54 3.23 4.39 8.37

Hex4 44.5 34.4 44.4 15.3 75.7 4.58

Hex5 16.6 15.0 21.6 13.2 8.46 6.42

Hex6 13.9 10.2 2.12 0.764 18.8 1.84

Hex7 14.3 12.4 16.3 11.3 15.2 4.05

Hex8 33.3 32.9 9.13 6.18 20.0 7.24

Hex9 51.5 24.2 54.3 21.3 61.3 5.19

Hex10 6.32 3.12 9.09 8.25 10.6 5.30

Hex11 10.6 4.70 2.81 9.74 10.1 10.3

Hex12 9.91 3.52 51.6 52.7 21.5 3.83

Hex13 14.0 5.18 7.49 6.54 29.4 12.3

Hex14 6.37 0.924 29.4 32.4 8.59 4.06

Hex15 9.07 4.86 16.0 18.0 8.77 4.55

Hex16 2.00 5.32 25.4 20.0 4.23 3.10

Mean 19.5 14.2 20.0 14.0 21.9 5.46

7

Environ. Res. Lett. 11 (2016) 075005

remote areas where, despite the lack of humans, we sawlittlefire activity.

One of the most surprising aspects of this study’sresults is the relatively strong influence of humans onfire activity across North America. Anthropogeniceffects were usually substantially greater than endur-ing features and lightning, and even greater than cli-mate in some areas. Our expectations were that moresparsely populated parts of the continent would havea relatively lower anthropogenic effect on fire. Asshown in all of the boreal hexels (hex10, hex12,hex13, hex14, hex15, hex16), this expectation did notbear out. More aligned with our expectations was thehigh anthropogenic impact of the Midwest (hex4,hex8), Great Plains (hex5, hex9), and Gulf of Mexico(hex2, hex4), areas that have undergone widespreadland conversions and that are at a more advancedstage of anthropogenic fire-regime transformation(Guyette et al 2002, Bowman et al 2011). By contrast,some heavily human-altered hexels of the con-terminous US show an unexpectedly moderate effectof humans on fire activity, perhaps because climate issimply much more dominant than other environ-mental controls. This may be the case in coastal Cali-fornia (hex3), where a clear anthropogenic effect(Syphard et al 2008)may bemasked by strong climateforcing (Moritz et al 2010).

Any large-scale effect of population density on fireactivity can be confounded by the type of human activ-ities on the landscape (Hawbaker et al 2013). For exam-ple, the culture of prescribed burning in the southeasternUS and, more recently the tolerance of lightning-ignitedfire in some of the protected areas in western NorthAmerica, may be idiosyncratic and not found elsewhereat similar population densities. Overall, our results agreewith those of Knorr et al (2014) andHantson et al (2015),who found a globally negative relationship between fireand population density, and are coherent with the con-clusions of Marlon et al (2008) and Mouillot and Field(2005), who reported a decrease in area burned in rela-tion to increased population trends in boreal and tempe-rate biomes. As observed in this study, Bistinas et al(2013) found this relationship to be spatially variable, butthere are discrepancies in the direction of relationships(positive and negative) between the two studies. In someparts of the continent (southwesternUS, Florida, and theGreat Lakes area), we observed the unimodal relation-ship reported by Syphard et al (2007), whereby the peakin fire ignitions occurred at intermediate populationdensities, suggesting that both intense human activityand the near absence of humans is associated with fewerfires. We also observed that the per capita effect ofhumans onfire activity varieswidely acrossNorthAmer-ica. Notably, in the boreal zone we found a non-

Figure 3.Partial dependenceplots of themodeled responseoffire probability as a function of log10 of populationdensity. The red lineindicates themean response, whereas the blue areas represent the standard deviation calculated from the 100model subsets.Note that thex-axes are variable amonghexels and its values havebeen scaled from0 to100 for easeof visualization; the y-axis is the same for all hexels.

8

Environ. Res. Lett. 11 (2016) 075005

negligible human impact on fire activity in spite of thatregion’s low population densities. This may be due inpart to the spatially expansive nature of human featuresand policies of aggressive fire suppression (Cum-ming 2005, Martell and Sun 2008). In this region, theeffects of humans on fire are spatially structured, in thatmost fires are ignited by humans close to roads, but it isthe remote lightning-ignited fires that are responsible forthemajority of the area burned (Gralewicz et al2012).

While many large-scale wildland fire studies mayinclude a single variable describing anthropogenicimpacts on fire (e.g., Krawchuk et al 2009, Bistinaset al 2013), we included three conceptually differentvariables to capture varying aspects of the fire–human relationship. We were expecting the HumanFootprint Index, which accounts for land use andinfrastructure in addition to population density, tohave a different relationship with area burned com-pared to population density. Results, however, showthat population density and the Human FootprintIndex have a remarkably similar (negative) relation-ship with fire. This provides further support to theidea that, at the spatiotemporal frame of this study,the anthropogenic factors that impede fire activity(e.g., suppression, landscape fragmentation) out-weigh those that promote it (e.g., ignitions) in NorthAmerica. Our third index, the roadless volume, tells a

somewhat different story, and therefore we cautionagainst using a single index to measure human influ-ence on fire regimes. Even our use ofmultiple indexescan probably not describe all important humanimpacts. For instance, the invasion of exotic cheat-grass species (Bromus spp.) have profoundly mod-ified fire regimes in parts of the western US (Balchet al 2013, Parks et al 2015), but are not fully capturedby our three indexes. The same could be said for theeffect of past fire management policies on current fireactivity, given that these have likely altered the natureof the fire−environment relationship in some areas(Higuera et al 2015, Ruffault andMouillot 2015).

The roadless volume represents a more accuratedepiction of the degree of isolation from humanactivities than the population density and the HumanFootprint Index. That is, the ‘stretch’ of values in thismetric is more heavily weighted towards the distancefrom roads (i.e., low-impact areas), translating into amore diverse set of human−fire relationships. Muchof the northern and western parts of North America,for instance, show a unimodal response of fire toroadless volume, which suggests less fire in the mostisolated areas (mainly protected areas) compared toareas of moderate-to-high remoteness. These resultsare consistent with what Parisien et al (2012) reportedin the western US; however, our results show that a

Figure 4.Partial dependence plots of themodeled response offire probability as a function ofHuman Footprint Index. The red lineindicates themean response, while the blue areas represent the standard deviation calculated from the 100model subsets. Note thatx-axes are variable among hexels; the y-axis is the same for all hexels.

9

Environ. Res. Lett. 11 (2016) 075005

single response for the western US masks the varia-bility in the fire responses across such a large area.Notwithstanding, there is a relative decrease in fireprobability in some of the most remote areas of somehexels. Reasons why some remote areas may haveexperienced relatively less fire than adjacent anthro-pogenized areas may be due to differences in fireenvironments, past fire-management policies, orsimply that there have experienced comparativelyfewer ignitions (Cardille et al 2001, Miller et al 2012,Haire et al 2013).

We suspect that the unimodal response of fire tohuman influence (roadless volume) in some areas ofNorth America is the result of ignition limitation, asobserved in some studies in California (Syphardet al 2007, Keeley et al 2011) and other parts of theglobe (Bradstock 2009, Kraaij et al 2013). Flanniganet al (2005) hinted that this may be the case in the lar-gely uninhabited parts of the lightning-poor borealzone. Some of the most remote areas of the westernUS may also be ignition limited simply because theyencompass some of the highest mountains andbecause lightning ignitions are deficient at very highelevations (Dissing and Verbyla 2003) (figure E1). Itmust be noted that some remote areas may haveexperienced more fire in the past if they were sub-jected to frequent burning by humans, but the extent

to which these past fire regimes were ‘natural’ or notis debated and likely varies greatly among areas(Vale 2002, White et al 2011). Nevertheless, in an eraof rapid and sometimes drastic change, it is impor-tant to evaluate the potential effect of changes in igni-tion loads, especially given the sensitivity of someecosystems to changes in fire activity (McWethyet al 2013).

LimitationsAs is the case with all modeling studies, this study’sresults are subject to limitations both in terms of thedata and the modeling approach. In spite of ourdataset being largely comprehensive, the missing firedata of hex12 (northern Ontario and Manitoba)mayhave significantly affected the results; therefore,results for this hexel should be interpreted withcaution. Although our systematic partitioning of thecontinent into hexels allows us to see broad patternsin human–fire relationships, comparison amonghexels is complicated by the widely varying range ofvalues in the three anthropogenic variables. Forexample, the maximum value for roadless volume inhex4 (southernMidwest) is lower than the minimumvalue for in hex16 (Alaska). The dichotomization ofthe response variable into burn and unburnedprecluded consideration of areas that burned more

Figure 5.Partial dependence plots of themodeled response offire probability as a function of roadless volume. The red line indicatesthemean response, while the blue areas represent the standard deviation calculated from the 100model subsets. Note that x-axes arevariable among hexels and its values have been scaled from0 to 100 for ease of visualization; the y-axis is the same for all hexels.

10

Environ. Res. Lett. 11 (2016) 075005

than once. Although the fraction of these areaswas low for the study time period it may be importantto explicitly consider them as fire history atlasesare updated with new data in future years. Interpreta-tion may also be complicated by the modelingtechnique. For instance, even though an effort wasmade to statistically control for the effects ofother variables, the influence of anthropogenic vari-ables may still conflate with non-anthropogenicvariables in some parts of North America. Thismay be the case at the forest/tundra transitioncharacterizing the northernmost band of hexels, forexample, where the presumed human influence maybe partly masked by other gradients in climate andvegetation.

The difficulty in pinpointing specific mechan-isms by which humans affect fire ignition and spreadfrom statistical relationships is further exacerbatedby the potential scale-dependence of these relation-ships. The fire–human relationships reported in thisstudy are not necessarily expected to hold over muchsmaller or greater spatial extents. For instance,whereas the negative association of wildfire withhumans over large areas, such as our hexels, is partlydue to greater area burned in some protected areas,this relationship could in fact be positive over areaswhere humans are responsible for most (or all) of thefire ignitions. Parisien et al (2011) observed a con-sistently negative response of area burned to theHuman Footprint Index at four spatial scales in theboreal forest of Canada. The scale-dependence of thefire–human relationship, however, requires furtherinvestigation and, as such, the results of this studyshould thus not be extrapolated to greatly varyingspatio-temporal scales. Humans influence aspects ofthe fire regime other than fire frequency or areaburned. For instance, human-induced changes to theseasonality of fire occurrence, though difficult todetect, may lead to fairly drastic ecological changes insome biomes (Le Page et al 2010). It would be inter-esting and important to consider other componentsto paint a more complete picture of human influenceon this disturbance process (Whitman et al 2015).Liu and Wimberly (2015), for example, foundthat humans had a greater impact on fire size thanfire occurrence of high-severity burns in the wes-ternUS.

Except for the percentage of permanent nonfuel,the fire probability models of this study did not expli-citly consider vegetation. There is compelling evidencethat biota affects spatial and temporal patterns of fireactivity (Girardin et al 2013, Terrier et al 2013), andvegetation-related variables have been often incorpo-rated into to large-scale biophysical models of fire(Sturtevant and Cleland 2007, Hawbaker et al 2013,Liu andWimberly 2015). Although these variables canbe very useful in explaining fire activity, they are also

highly correlated with climatic gradients at the spatio-temporal scale of study. In fact, Parisien and Moritz(2009) showed there was virtually no loss of modelperformance when vegetation class was removed frommodels of fire activity built for the conterminous USthat included several climate variables. Informationon vegetation is certainly useful in explaining potentialfeedback mechanisms that can regulate fire occur-rence via postfire succession, but this level of detail isbeyond the scope of this large-scale study. Anotherreason for focusing on climate and not incorporatingvegetation composition into this study is that it isusually impossible to know the state of the vegetationat the time of burning. This caveat convinced us that,to meet the goal of our study, it was more appropriateto use climatic gradients; however, it must beacknowledged that detailed vegetation information atthe time of burn would have likely substantiallyimproved ourmodels.

Conclusion

The results of this study suggest that there may be fewtruly natural fire regimes in North America today.While this study points to a general impeding effect ofpeople on fire across North America, it also paints acomplex picture of anthropogenic effects on fire acrossthe continent—one where fire is as variable as thebiophysical environment that defines it. Because of thiscomplexity, the specific mechanisms by which humansalter fire ignition and spread may be difficult, or evenimpossible, to identify. Further, humans can indeeddistort or mask fire–climate relationships, therebyundermining our ability to predict current and futurefire activity in a changing climate (Parks et al 2014).Consequently, this study’s results should be viewed asanother building block towards more in-depth investi-gations of the spatial variability of human impacts onNorth American fire regimes. The statistical relation-ships reported here could, for example, be incorporatedinto a process-based model to enhance our under-standing of changing human–fire relationships (e.g.,Thonicke et al 2010). As adaptation and mitigationmeasures are developed for fire-prone areas, it willbecome imperative that wildland fire scientists and landmanagers account for human influences in projectionsof futurefire.

Acknowledgments

This study was partially funded by the AdaptationProgramofNatural ResourcesCanada.We are gratefulto Nicolas Mansuy and two anonymous reviewers forconstructive comments.

11

Environ. Res. Lett. 11 (2016) 075005

AppendixA

TableA1. List of all variables initially considered for themodeling offire probability. A subset of variables was selected to build thefireprobabilitymodels in this study.

Variable name Description Units Mean (range)

Climate (1981–2010 normals)StartFrostFree Julian date onwhich the frost-free period begins Day of year 135 (0–208)CMI Hargreave’s climaticmoisture index Dimensionless 299 (0–1754)DegDaysA18 Degree days above 18 °C Degree days 345 (0–3044)GrowDegDays Degree days above 5 °C (growing degree days) Degree days 2081 (0–7643)ChilDegDays Degree days under 0 °C (chilling degree days) Degree days 1358 (0–5339)DegDaysU18 Degree days under 18 °C Degree days 5043 (0–11 539)EndForstFree Julian date onwhich the frost-free period ends Day of year 271 (0–365)MinTemp Extrememinimum temperature over 30 years °C –35 (–63 to 10)Eref Hargreave’s reference evaporation Dimensionless 749 (0–1924)MaxTemp Extrememaximum temperature over 30 years °C 36 (19–53)FrostFree frost-free period Days 136 (0–365)MeanPrecip Mean annual precipitation mm 722 (50–10 149)MeanTemp Mean annual temperature °C 4.9 (–13 to 26)MeanTempCold Mean temperature of the coldestmonth °C –9.1 (–32 to 22)SumPrecip Mean summer (May–September) precipitation mm 347 (6–3175)MeanTempWarm Mean temperature of thewarmestmonth °C 18.7 (2–38)NumFrostFree Number of frost-free days 179.9 (0–365)SnowPrecip Precipitation as snow mm 154.2 (0–6608)SumPrecip Summer (June–August) precipitation mm 217.5 (0–1678)WintPrecip Winter (December–February)precipitation mm 150.5 (8–3013)RelHumid Mean annual relative humidity % 59.6 (36–85)SumHeatMoist Summer heatmoisture index, calculated asMWMT/

(MSP/1000)°C mm–1 89.9 (1–4085)

SumTemp Summer (June–August)mean temperature °C 17.3 (1–36)WintTemp Winter (December–February)mean temperature °C –7.7 (–31 to 22)

Anthroprogenic

HumanFoota Human Footprint Index, an index of human influence for

the year 2004

Dimensionless 16.5 (0–100)

LogPopDensa Population density for the year 2000 People km–2 –0.51 (–4 to 4.3)RdsDensb Road density using 2014 roads datawith only primary and

secondary roads used

Roads area–1 0.11 (0–4.3)

RdlsVolc Roadless volume km3 6.5 (0–100)

Enduring

HeatLoadIndex Heat load index, an index calculating the southwestness of a

slope

Dimensionless 0.64 (0.096–1.03)

SurfAreaRatio Surface area ratio Dimensionless 8178 (8100–42 944)TopoPosIndex Topographic position index calculated at 2000 m scale Dimensionless –0.34 (–489 to 843)PctNonfuelc Percent nonfuel. Non-fuel classified as barren lands, water,

snow, and ice. From2005 land cover.

% 4.5 (0–99.7)

GPP Gross primary productivity from2014 kg_carbonm–2 7144.6 (0–35 508)

Lightning

Lightning Average number of lightning strikes per year from1995

to 2005

Flashes km–2 year–1 5.9 (0–46.4)

Fire

Area burned USfiresmergedwithCanadian fires. US fires fromMTBS

database 1984–2014. Canadian fires fromCanadian

National Fires database 1984–2014. Fires>400 ha used.

ha 5311 (400–2 205 060)

a Variables calculated at 1 and 100 km2 scale.b Variables calculated at 10, 50, and 100 km2 scale.c Variables calculated at 100 and 10 000 km2 scale.

12

Environ. Res. Lett. 11 (2016) 075005

Appendix B

AppendixC

Model evaluation followed the method of Parisien et al(2012) and was performed on each of the 100 subsetmodels and subsequently averaged for each hexel. As ameasure of overall of model performance, the AUC wascomputed from the true positives and false positives.AUC values may range from 0.5, where predictionaccuracy is no better than if samples were randomlyselected, to 1, which indicates perfect classificationaccuracy. However, in a presence-only framework, as inthis study, it is impossible to achieve an AUC value of 1

because absences (hence false positives) are unknown.The maximum achievable AUC in a presence-onlyframework is equal to 1–a/2, where a is generally thefraction of the study area covered by fire (i.e., theprevalence). For the sake of adjusting the AUC value, weconsidered a to be the percentage of pixels where firewasobserved. This provides a fair, yet underestimated,approximation of prevalence. Also calculated were theestimated fraction of the suitable (i.e., burnable) area,which represents an approximation of the false positiverate, and the omission, which is the false negative rate.These two metrics were measured at the fire probability

Figure B1.Mapof the explanatory variables incorporated in the statisticalmodels offire probability for each hexel.

13

Environ. Res. Lett. 11 (2016) 075005

threshold that minimizes the sum of these errors of thetwo metrics (Liu et al 2005). As such, they can beinterpreted as the expected rate of false negatives for agivenpredicted suitable area.

AppendixD

Table C1.Performancemetrics of thefire probabilitymodel for each hexel (seefigure 1).

Test AUC

Adjusted

AUC

Suitable

area (%)Omission

error (%)

Hex1 0.752 0.776 0.370 0.247

Hex2 0.831 0.858 0.293 0.172

Hex3 0.797 0.831 0.333 0.191

Hex4 0.827 0.836 0.231 0.273

Hex5 0.700 0.711 0.295 0.411

Hex6 0.881 0.877 0.249 0.118

Hex7 0.701 0.749 0.424 0.281

Hex8 0.891 0.894 0.211 0.115

Hex9 0.839 0.860 0.298 0.167

Hex10 0.837 0.875 0.308 0.155

Hex11 0.700 0.706 0.394 0.315

Hex12 0.704 0.765 0.539 0.154

Hex13 0.672 0.684 0.507 0.192

Hex14 0.746 0.827 0.501 0.094

Hex15 0.715 0.764 0.487 0.184

Hex16 0.818 0.849 0.343 0.137

Mean 0.776 0.804 0.361 0.200

FigureD1.Modeled fire probability as a function of the three anthropogenic variables obtained frombivariatemodels (i.e., areaburned as a single anthropogenic variable, without controlling for the effects of climate, enduring features, and lightning). The red lineindicates themean response, whereas the blue areas represent the standard deviation, calculated from the 100model subsets.

14

Environ. Res. Lett. 11 (2016) 075005

FigureD1. (Continued.)

15

Environ. Res. Lett. 11 (2016) 075005

Appendix E

References

Archibald S, Staver AC and Levin SA 2012 Evolution of human-drivenfire regimes in AfricaProc. Natl Acad. Sci. USA 109847–52

Balch J K, Bradley BA,D’AntonioC andGomez-Dans J 2013Introduced annual grass increases regional fire activity acrossthe aridwesternUSA (1980–2009)Glob. Change Biol. 19173–83

BalkD L,DeichmannU, YetmanG, Pozzi F,Hay S I andNelsonA2006Determining global population distribution:methods,applications and dataAdv. Parasitol. 62 119–56

Bistinas I, OomD, SáAC,Harrison S P, Prentice I C and Pereira JM2013Relationships between human population density andburned area at continental and global scalesPloSOne 8e81188

BowmanDM et al 2011The human dimension of fire regimes onEarth J. Biogeogr. 38 2223–36

BradstockRA 2009 Effects of largefires on biodiversity in south-easternAustralia: disaster or template for diversity? Int. J.Wildland Fire 17 809–22

BrownK J, Clark J S, GrimmEC,Donovan J J,Mueller PG,Hansen BC S and Stefanova I 2005 Fire cycles inNorthAmerican interior grasslands and their relation to prairiedroughtProc. Natl Acad. Sci. USA 102 8865–70

Carcaillet C, Ali AA, BlarquezO,Genries A,Mourier B andBremond L 2009 Spatial variability offire history insubalpine forests: fromnatural to cultural regimes Ecoscience16 1–12

Cardille J A and LamboisM2010 From the redwood forest totheGulf Streamwaters: human signature nearly ubiquitousin representativeUS landscapes Front. Ecol. Environ. 8130–4

Cardille J A, Ventura S J andTurnerMG2001 Environmental andsocial factors influencing wildfires in theUpperMidwest,United States Ecol. Appl. 11 111–27

ChristianH J et al 2003Global frequency and distribution oflightning as observed from space by theOptical TransientDetector J. Geophys. Res. 108 4005

Clark J S andRoyall PD1996 Local and regional sediment charcoalevidence for fire regimes in presettlement north-easternNorthAmerica J. Ecol. 365 382

CochraneMA2003 Fire science for rainforestsNature 421913–9

Collins S L andWallace L L (ed) 1990 Fire inNorth AmericanTallgrass Prairies (Norman:University ofOklahoma Press)

Cumming SG2005 Effective fire suppression in boreal forestsCan.J. For. Res. 35 772–86

DissingD andVerbylaD L 2003 Spatial patterns of lightning strikesin interior Alaska and their relations to elevation andvegetationCan. J. For. Res. 33 770–82

Eidenshink J, Schwind B, BrewerK, ZhuZ-L,Quayle B andHoward S 2007A project formonitoring trends in burnseverity Fire Ecology 3 3–21

Elith J, Phillips S J, Hastie T, DudíkM,Chee Y E andYates C J 2011Astatistical explanation ofMaxEnt for ecologistsDivers.Distrib. 17 43–57

ESRI 2008Tele Atlas North America: StreetMap 2008North America(Redlands, California: ESRI)

Evans J S, Oakleaf J, Cushman S andTheobaldD 2014AnArcGISToolbox for SurfaceGradient andGeomorphometricModeling, version 2.0-0 (http://evansmurphywixcom/evansspatial)

FinneyM,Grenfell I C andMcHughCW2009Modelingcontainment of large wildfires using generalized linearmixed-model analysis Forest Sci. 55 249–55

Figure E1.The relationshipbetween roadless volume (x-axis) and lightningdensity (y-axis), as used in the statisticalmodels offireprobability. The red line indicates thefit between the twovariables using a self-fitting generalized additivemodel.Note that in someof thehexels themost remote areas (highvalues of roadless volume)have lower ignitiondensities than in the areas of higher human influence.

16

Environ. Res. Lett. 11 (2016) 075005

http://dx.doi.org/10.1073/pnas.1118648109http://dx.doi.org/10.1073/pnas.1118648109http://dx.doi.org/10.1073/pnas.1118648109http://dx.doi.org/10.1073/pnas.1118648109http://dx.doi.org/10.1111/gcb.12046http://dx.doi.org/10.1111/gcb.12046http://dx.doi.org/10.1111/gcb.12046http://dx.doi.org/10.1111/gcb.12046http://dx.doi.org/10.1016/S0065-308X(05)62004-0http://dx.doi.org/10.1016/S0065-308X(05)62004-0http://dx.doi.org/10.1016/S0065-308X(05)62004-0http://dx.doi.org/10.1371/journal.pone.0081188http://dx.doi.org/10.1371/journal.pone.0081188http://dx.doi.org/10.1111/j.1365-2699.2011.02595.xhttp://dx.doi.org/10.1111/j.1365-2699.2011.02595.xhttp://dx.doi.org/10.1111/j.1365-2699.2011.02595.xhttp://dx.doi.org/10.1071/WF07153http://dx.doi.org/10.1071/WF07153http://dx.doi.org/10.1071/WF07153http://dx.doi.org/10.1073/pnas.0503621102http://dx.doi.org/10.1073/pnas.0503621102http://dx.doi.org/10.1073/pnas.0503621102http://dx.doi.org/10.2980/16-1-3189http://dx.doi.org/10.2980/16-1-3189http://dx.doi.org/10.2980/16-1-3189http://dx.doi.org/10.1890/080132http://dx.doi.org/10.1890/080132http://dx.doi.org/10.1890/080132http://dx.doi.org/10.1890/080132http://dx.doi.org/10.1890/1051-0761(2001)011[0111:EASFIW]2.0.CO;2http://dx.doi.org/10.1890/1051-0761(2001)011[0111:EASFIW]2.0.CO;2http://dx.doi.org/10.1890/1051-0761(2001)011[0111:EASFIW]2.0.CO;2http://dx.doi.org/10.1029/2002JD002347http://dx.doi.org/10.2307/2261199http://dx.doi.org/10.1038/nature01437http://dx.doi.org/10.1038/nature01437http://dx.doi.org/10.1038/nature01437http://dx.doi.org/10.1038/nature01437http://dx.doi.org/10.1139/x04-174http://dx.doi.org/10.1139/x04-174http://dx.doi.org/10.1139/x04-174http://dx.doi.org/10.1139/x02-214http://dx.doi.org/10.1139/x02-214http://dx.doi.org/10.1139/x02-214http://dx.doi.org/10.4996/fireecology.0301003http://dx.doi.org/10.4996/fireecology.0301003http://dx.doi.org/10.4996/fireecology.0301003http://dx.doi.org/10.1111/j.1472-4642.2010.00725.xhttp://dx.doi.org/10.1111/j.1472-4642.2010.00725.xhttp://dx.doi.org/10.1111/j.1472-4642.2010.00725.xhttp://evansmurphywixcom/evansspatialhttp://evansmurphywixcom/evansspatial

FlanniganMD,KrawchukMA, deGrootW J,Wotton BMandGowmanLM2009 Implications of changing climate forglobal wildlandfire Int. J.Wildland Fire 18 483–507

FlanniganMD, LoganKA,Amiro BD, SkinnerWR and Stocks B J2005 Future area burned inCanadaClim. Change 72 1–16

FriedlMA, Sulla-MenasheD, Tan B, Schneider A, RamankuttyN,Sibley A andHuangX 2010MODISCollection 5 global landcover: algorithm refinements and characterization of newdatasetsRemote Sens. Environ. 114 168–82

Fréjaville T andCurt T 2015 Spatiotemporal patterns of changes infire regime and climate: defining the pyroclimates of south-eastern France (Mediterranean Basin)Clim. Change 129239–51

GirardinMP, Ali AA,Carcaillet C, BlarquezO,Hély C, Terrier A,Genries A andBergeron Y 2013Vegetation limits the impactof a warm climate on boreal wildfiresNewPhytol. 1991001–11

GralewiczN J, NelsonTA andWulderMA2012 Factorsinfluencing national scalewildfire susceptibility inCanadaForest Ecol.Manage. 265 20–9

Guyette R P,Muzika RMandDeyDC2002Dynamics of ananthropogenicfire regime Ecosystems 5 472–86

Guyette R P, StambaughMC,DeyDCandMuzika RM2012Predictingfire frequencywith chemistry and climateEcosystems 15 322–35

Haire S L,McGarigal K andMiller C 2013Wilderness shapescontemporaryfire size distributions across landscapes of thewesternUnited States Ecosphere 4 15

Hantson S, LasslopG,Kloster S andChuvieco E 2015Anthropogenic effects on globalmean fire size Int. J.WildlandFire 24 589–96

HawbakerT J,RadeloffVC, Stewart S I,HammerRB,KeulerNS andClaytonMK2013Humanandbiophysical influences onfireoccurrence in theUnited StatesEcol. Appl.23 565–82

Higuera P E, Abatzoglou J T, Littell J S andMorgan P 2015Thechanging strength and nature offire–climate relationships intheNorthern RockyMountains, USA, 1902–2008PloSOne10 e0127563

Johnson EA,Miyanishi K andBridge S R J 2001Wildfire regime inthe boreal forest and the idea of suppression and fuel buildupConserv. Biol. 15 1554–7

Kasischke E S andTuretskyMR2006Recent changes in thefireregime across theNorthAmerican boreal region—spatial andtemporal patterns of burning across Canada andAlaskaGeophys. Res. Lett. 33 L09703

Keane RE, RyanKC,Veblen TT, AllenCD, Logan J A andHawkes B 2002The cascading effects offire exclusion inRockyMountain ecosystemsRockyMountain Futures: anEcological Perspective ed J Baron (Washington, DC: Island)pp 133–52

Keeley J E, BondW J, BradstockRA, Pausas J G andRundel PW2011 Fire inMediterranean Ecosystems: Ecology, Evolution andManagement (Cambridge: CambridgeUniversity Press)

KnorrW,Kaminski T, ArnethA andWeberU2014 Impact ofhumanpopulation density onfire frequency at the globalscaleBiogeosciences 11 1085–102

Kraaij T, CowlingRMand vanWilgenBW2013 Lightning and fireweather in eastern coastal fynbos shrublands: seasonality andlong-term trends Int. J.Wildland Fire 22 288–95

KrawchukMA,MoritzMA, ParisienMA,VanDorn J andHayhoeK2009Global pyrogeography: the current and futuredistribution ofwildfirePloSOne 4 pe5102

Lavorel S, FlanniganMD, Lambin E F and ScholesMC2007Vulnerability of land systems to fire: interactions amonghumans, climate, the atmosphere, and ecosystemsMitig.Adapt. Strat. Glob. Change 12 33–53

Le Page Y,OomD, Silva J, Jönsson P andPereira J 2010 Seasonalityof vegetation fires asmodified by human action: observingthe deviation from eco‐climatic fire regimesGlob. Ecol.Biogeogr. 19 575–88

LiuC, Berry PM,DawsonTP and PearsonRG2005 Selectingthresholds of occurrence in the prediction of speciesdistributionsEcography 28 385−393

Liu Z andWimberlyMC2015Climatic and landscape influences onfire regimes from1984 to 2010 in thewesternUnited StatesPLoSOne 10 pe0140839

Marlon J R, Bartlein P J, Carcaillet C,GavinDG,Harrison S P,Higuera P E, Joos F, PowerM J and Prentice I C 2008Climateand human influences on global biomass burning over thepast twomillenniaNat. Geosci. 1 697−702

Martell D L and SunH2008The impact of forestfire suppression,vegetation andweather on burned area inOntarioCan. J. For.Res. 38 1547–63

McWethyDB et al 2013A conceptual framework for predictingtemperate ecosystem sensitivity to human impacts onfireregimesGlob. Ecol. Biogeogr. 22 900–12

Miller J D, Collins BM, Lutz J A, Stephens S L,vanWagtendonk JWandYasudaDA2012Differencesinwildfires among ecoregions and landmanagementagencies in the SierraNevada region, California, USAEcosphere 3 80

MoritzMA,Moody T J, KrawchukMA,HughesMandHall A 2010Spatial variation in extremewinds predicts large wildfirelocations in chaparral ecosystemsGeophys. Res. Lett. 37L04801

Mouillot F and Field CB 2005 Fire history and the global carbonbudget: a 1× 1fire history reconstruction for the 20thcenturyGlob. Change Biol. 11 398–420

NowackiG J andAbramsMD2008The demise offire and‘mesophication’ of forests in the easternUnited StatesBioScience 58 123–38

ParisienM-A andMoritzMA2009 Environmental controls on thedistribution ofwildfire atmultiple spatial scales Ecol.Monog.79 127–54

ParisienM-A, Parks S A, KrawchukMA, FlanniganMD,BowmanLMandMoritzMA2011 Scale-dependent factorscontrolling area burned in boreal CanadaEcol. Appl. 21789−805

ParisienM-A, Peters V S,Wang Y, Little JM, Bosch EMandStocks B J 2006 Spatial patterns of forestfires inCanada1980–1999 Int. J.Wildland Fire 15 361–74

ParisienM-A, Snetsinger S, Greenberg J A,NelsonCR,Schoennagel T,Dobrowski S Z andMoritzMA2012 Spatialvariability inwildfire probability across the westernUnitedStates Int. J.Wildland Fire 21 313–27

Parks SA,Miller C, ParisienM-A,Holsinger LM,Dobrowski S Z andAbatzoglou J 2015Wildland fire deficitand surplus in thewesternUnited States 1984–2012Ecosphere6 13

Parks SA, ParisienM-A,Miller C andDobrowski S Z 2014 Fireactivity and severity in thewesternUS vary along proxygradients representing fuel amount and fuelmoisture PLoSOne 9 e99699

Pausas J G andKeeley J E 2009 A burning story: the role of fire in thehistory of lifeBioScience 59 593–601

PerryG L,Wilmshurst JM,McGloneMS andNapier A 2012Reconstructing spatial vulnerability to forest loss by fire inpre‐historicNewZealandGlob. Ecol. Biogeogr. 21 1029–41

Phillips S J, AndersonRP and Schapire R E 2006Maximumentropymodeling of species geographic distributions Ecol.Model. 190231–59

Ruffault J andMouillot F 2015How anewfire‐suppression policycan abruptly reshape thefire–weather relationship Ecosphere6 19

Sanderson EW, JaitehM, LevyMA,RedfordKH,WanneboAV andWoolmerG2002The human footprintand the last of thewildBioscience 52 891−904

SlocumMG, PlattW J, Beckage B, PankoB and Lushine J B 2007Decoupling natural and anthropogenic fire regimes: a casestudy in EvergladesNational Park, FloridaNat. Area J. 2741–55

Stephens S L 2005 Forestfire causes and extent onUnited StatesForest Service lands Int. J.Wildland Fire 14 213–22

Stephens S L,Martin R E andClintonNE2007 Prehistoric fire areaand emissions fromCalifornia’s forests, woodlands,shrublands, and grasslands For. Ecol.Manage. 251 205–16

17

Environ. Res. Lett. 11 (2016) 075005

http://dx.doi.org/10.1071/WF08187http://dx.doi.org/10.1071/WF08187http://dx.doi.org/10.1071/WF08187http://dx.doi.org/10.1007/s10584-005-5935-yhttp://dx.doi.org/10.1007/s10584-005-5935-yhttp://dx.doi.org/10.1007/s10584-005-5935-yhttp://dx.doi.org/10.1016/j.rse.2009.08.016http://dx.doi.org/10.1016/j.rse.2009.08.016http://dx.doi.org/10.1016/j.rse.2009.08.016http://dx.doi.org/10.1007/s10584-015-1332-3http://dx.doi.org/10.1007/s10584-015-1332-3http://dx.doi.org/10.1007/s10584-015-1332-3http://dx.doi.org/10.1007/s10584-015-1332-3http://dx.doi.org/10.1111/nph.12322http://dx.doi.org/10.1111/nph.12322http://dx.doi.org/10.1111/nph.12322http://dx.doi.org/10.1111/nph.12322http://dx.doi.org/10.1016/j.foreco.2011.10.031http://dx.doi.org/10.1016/j.foreco.2011.10.031http://dx.doi.org/10.1016/j.foreco.2011.10.031http://dx.doi.org/10.1007/s10021-011-9512-0http://dx.doi.org/10.1007/s10021-011-9512-0http://dx.doi.org/10.1007/s10021-011-9512-0http://dx.doi.org/10.1890/ES12-00257.1http://dx.doi.org/10.1071/WF14208http://dx.doi.org/10.1071/WF14208http://dx.doi.org/10.1071/WF14208http://dx.doi.org/10.1890/12-1816.1http://dx.doi.org/10.1890/12-1816.1http://dx.doi.org/10.1890/12-1816.1http://dx.doi.org/10.1371/journal.pone.0127563http://dx.doi.org/10.1046/j.1523-1739.2001.01005.xhttp://dx.doi.org/10.1046/j.1523-1739.2001.01005.xhttp://dx.doi.org/10.1046/j.1523-1739.2001.01005.xhttp://dx.doi.org/10.1029/2006GL025677http://dx.doi.org/10.5194/bg-11-1085-2014http://dx.doi.org/10.5194/bg-11-1085-2014http://dx.doi.org/10.5194/bg-11-1085-2014http://dx.doi.org/10.1071/WF11167http://dx.doi.org/10.1071/WF11167http://dx.doi.org/10.1071/WF11167http://dx.doi.org/10.1371/journal.pone.0005102http://dx.doi.org/10.1007/s11027-006-9046-5http://dx.doi.org/10.1007/s11027-006-9046-5http://dx.doi.org/10.1007/s11027-006-9046-5http://dx.doi.org/10.1111/j.1466-8238.2010.00525.xhttp://dx.doi.org/10.1111/j.1466-8238.2010.00525.xhttp://dx.doi.org/10.1111/j.1466-8238.2010.00525.xhttp://dx.doi.org/10.1111/j.0906-7590.2005.03957.xhttp://dx.doi.org/10.1371/journal.pone.0140839http://dx.doi.org/10.1038/ngeo313http://dx.doi.org/10.1139/X07-210http://dx.doi.org/10.1139/X07-210http://dx.doi.org/10.1139/X07-210http://dx.doi.org/10.1111/geb.12038http://dx.doi.org/10.1111/geb.12038http://dx.doi.org/10.1111/geb.12038http://dx.doi.org/10.1890/ES12-00158.1http://dx.doi.org/10.1029/2009GL041735http://dx.doi.org/10.1029/2009GL041735http://dx.doi.org/10.1111/j.1365-2486.2005.00920.xhttp://dx.doi.org/10.1111/j.1365-2486.2005.00920.xhttp://dx.doi.org/10.1111/j.1365-2486.2005.00920.xhttp://dx.doi.org/10.1641/B580207http://dx.doi.org/10.1641/B580207http://dx.doi.org/10.1641/B580207http://dx.doi.org/10.1890/07-1289.1http://dx.doi.org/10.1890/07-1289.1http://dx.doi.org/10.1890/07-1289.1http://dx.doi.org/10.1890/10-0326.1http://dx.doi.org/10.1890/10-0326.1http://dx.doi.org/10.1071/WF06009http://dx.doi.org/10.1071/WF06009http://dx.doi.org/10.1071/WF06009http://dx.doi.org/10.1071/WF11044http://dx.doi.org/10.1071/WF11044http://dx.doi.org/10.1071/WF11044http://dx.doi.org/10.1890/ES15-00294.1http://dx.doi.org/10.1371/journal.pone.0099699http://dx.doi.org/10.1525/bio.2009.59.7.10http://dx.doi.org/10.1525/bio.2009.59.7.10http://dx.doi.org/10.1525/bio.2009.59.7.10http://dx.doi.org/10.1111/j.1466-8238.2011.00745.xhttp://dx.doi.org/10.1111/j.1466-8238.2011.00745.xhttp://dx.doi.org/10.1111/j.1466-8238.2011.00745.xhttp://dx.doi.org/10.1016/j.ecolmodel.2005.03.026http://dx.doi.org/10.1016/j.ecolmodel.2005.03.026http://dx.doi.org/10.1016/j.ecolmodel.2005.03.026http://dx.doi.org/10.1016/j.ecolmodel.2005.03.026http://dx.doi.org/10.1890/ES15-00182.1http://dx.doi.org/10.1641/0006-3568(2002)052[0891:THFATL]2.0.CO;2http://dx.doi.org/10.3375/0885-8608(2007)27[41:DNAAFR]2.0.CO;2http://dx.doi.org/10.3375/0885-8608(2007)27[41:DNAAFR]2.0.CO;2http://dx.doi.org/10.3375/0885-8608(2007)27[41:DNAAFR]2.0.CO;2http://dx.doi.org/10.3375/0885-8608(2007)27[41:DNAAFR]2.0.CO;2http://dx.doi.org/10.1071/WF04006http://dx.doi.org/10.1071/WF04006http://dx.doi.org/10.1071/WF04006http://dx.doi.org/10.1016/j.foreco.2007.06.005http://dx.doi.org/10.1016/j.foreco.2007.06.005http://dx.doi.org/10.1016/j.foreco.2007.06.005

Stephens S L andRuth LW2005 Federal forest fire policy in theUnited States Ecol. Appl. 15 532–42

Stocks B J et al 2002 Large forestfires in Canada 1959–1997J. Geophys. Res.—Atmos. 108 FFR5-1–FFR5-12

Sturtevant BR andClelandDT2007Human and biophysicalfactors influencingmodern fire disturbance in northernWisconsin Int. J.Wildland Fire 16 398–413

SwainAM1973Ahistory offire and vegetation in northeasternMinnesota as recorded in lake sedimentsQuat. Res. 3383IN1391–90IN2396

SwetnamTW1993 Fire history and climate change in giant sequoiagroves Science 262 885–885

SyphardAD, Radeloff VC,Keeley J E,Hawbaker T J, ClaytonMK,Stewart S I andHammerRB 2007Human influence onCaliforniafire regimes Ecol. Appl. 17 1388–402

SyphardAD, Radeloff VC,KeulerN S, Taylor R S,Hawbaker T J,Stewart S I andClaytonMK2008 Predicting spatial patternsoffire on a southernCalifornia landscape Int. J.Wildland Fire17 602–13

Terrier A,GirardinMP, Périé C, Legendre P andBergeron Y 2013Potential changes in forest composition could reduce impactsof climate change on boreal wildfiresEcol. Appl. 23 21–35

ThonickeK, Spessa A, Prentice I C,Harrison S P,Dong L andCarmona-MorenoC2010The influence of vegetation, firespread and fire behaviour on biomass burning and trace gasemissions: results from a process-basedmodelBiogeosciences7 1991–2011

Vale T 2002 Fire, Native Peoples, and theNatural Landscape(Washington, DC: Island)

WaldropTA,WhiteD L and Jones SM1992 Fire regimes for pine-grassland communities in the southeasternUnited States For.Ecol.Manage. 47 195–210

WangT,HamannA, SpittlehouseD andMurdockTN2012ClimateWNA—High-resolution spatial climate data forwesternNorth America J. Appl.Meteorol. Clim. 61 16–29

Watts RD,ComptonRW,McCammon JH, RichCL,Wright SM,Owens T andOurenD S 2007Roadless space of theconterminousUnited States Science 316 736–8

Weir JMHand Johnson EA 1998 Effects of escaped settlementfiresand logging on forest composition in themixedwood borealforestCan. J. For. Res. 28 459–67

Westerling AL, GershunovA, BrownT J, CayanDRandDettingerMD2003Climate andwildfire in thewesternUnited StatesBr. Am.Meteorol. Soc. 84 595–604

Westerling AL,HidalgoHG, CayanDRand SwetnamTW2006Warming and earlier spring increase westernUS forestwildfire activity Science 313 940–3

WhiteCA,PerrakisDDB,KafkaVGandEnnisT2011Burning at theedge: integrating biophysical and eco-culturalfire processes inCanada’s parks andprotected areasFire Ecol.7 74–106

Whitman E, Batllori E, ParisienM-A,Miller C, Coop JD,KrawchukMA,ChongGWandHaire S L 2015The climatespace of fire regimes in north‐westernNorthAmericaJ. Biogeogr. 42 1736–49

18

Environ. Res. Lett. 11 (2016) 075005

http://dx.doi.org/10.1890/04-0545http://dx.doi.org/10.1890/04-0545http://dx.doi.org/10.1890/04-0545http://dx.doi.org/10.1029/2001JD000484http://dx.doi.org/10.1071/WF06023http://dx.doi.org/10.1071/WF06023http://dx.doi.org/10.1071/WF06023http://dx.doi.org/10.1016/0033-5894(73)90004-5http://dx.doi.org/10.1016/0033-5894(73)90004-5http://dx.doi.org/10.1016/0033-5894(73)90004-5http://dx.doi.org/10.1016/0033-5894(73)90004-5http://dx.doi.org/10.1126/science.262.5135.885http://dx.doi.org/10.1126/science.262.5135.885http://dx.doi.org/10.1126/science.262.5135.885http://dx.doi.org/10.1890/06-1128.1http://dx.doi.org/10.1890/06-1128.1http://dx.doi.org/10.1890/06-1128.1http://dx.doi.org/10.1071/WF07087http://dx.doi.org/10.1071/WF07087http://dx.doi.org/10.1071/WF07087http://dx.doi.org/10.1890/12-0425.1http://dx.doi.org/10.1890/12-0425.1http://dx.doi.org/10.1890/12-0425.1http://dx.doi.org/10.5194/bg-7-1991-2010http://dx.doi.org/10.5194/bg-7-1991-2010http://dx.doi.org/10.5194/bg-7-1991-2010http://dx.doi.org/10.1016/0378-1127(92)90274-Dhttp://dx.doi.org/10.1016/0378-1127(92)90274-Dhttp://dx.doi.org/10.1016/0378-1127(92)90274-Dhttp://dx.doi.org/10.1175/JAMC-D-11-043.1http://dx.doi.org/10.1175/JAMC-D-11-043.1http://dx.doi.org/10.1175/JAMC-D-11-043.1http://dx.doi.org/10.1126/science.1138141http://dx.doi.org/10.1126/science.1138141http://dx.doi.org/10.1126/science.1138141http://dx.doi.org/10.1139/x98-001http://dx.doi.org/10.1139/x98-001http://dx.doi.org/10.1139/x98-001http://dx.doi.org/10.1175/BAMS-84-5-595http://dx.doi.org/10.1175/BAMS-84-5-595http://dx.doi.org/10.1175/BAMS-84-5-595http://dx.doi.org/10.1126/science.1128834http://dx.doi.org/10.1126/science.1128834http://dx.doi.org/10.1126/science.1128834http://dx.doi.org/10.4996/fireecology.0701074http://dx.doi.org/10.4996/fireecology.0701074http://dx.doi.org/10.4996/fireecology.0701074http://dx.doi.org/10.1111/jbi.12533http://dx.doi.org/10.1111/jbi.12533http://dx.doi.org/10.1111/jbi.12533

IntroductionMethodsDataFireClimateEnduring featuresLightningAnthropogenic influenceVariable selectionStatistical modeling

Assessment of anthropogenic influence

ResultsDiscussionLimitations

ConclusionAcknowledgmentsAppendix AAppendix BAppendix CAppendix DAppendix EReferences