Assessing Fuel Treatment Effectiveness Using Satellite Imagery and Spatial Statistics Author(s): Michael C. Wimberly, Mark A. Cochrane, Adam D. Baer and Kari Pabst Reviewed work(s): Source: Ecological Applications, Vol. 19, No. 6 (Sep., 2009), pp. 1377-1384 Published by: Ecological Society of America Stable URL: http://www.jstor.org/stable/40346253 . Accessed: 15/06/2012 09:30 Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org. Ecological Society of America is collaborating with JSTOR to digitize, preserve and extend access to Ecological Applications. http://www.jstor.org Communications EcologicalApplications,19(6), 2009, pp. 1377-1384 © 2009 by the Ecological Societyof America fueltreatment effectiveness Assessing usingsatelliteimagery and spatialstatistics Michael C. Wimberly,1Mark A. Cochrane, Adam D. Baer, and Kari Pabst ScienceCenterof Excellence,SouthDakota State University, Brookings,SouthDakota 57007 USA GeographicInformation Abstract. Understandingthe influencesof forestmanagementpractices on wildfire ecosystemsof the westernUnited States. Newly available severityis criticalin fire-prone vegetation,fuels,topography,and burnseverityoffernew geospatialdata setscharacterizing effectiveness at regionalto nationalscales. In this forstudyingfueltreatment opportunities and sequentialautoregression we used (SAR) (OLS) regression ordinary least-squares study, theCamp 32 firein on burnseverity forthreerecentwildfires: effects to analyzefueltreatment westernMontana,theSchool firein southeastern Washington,and theWarmfirein northern normalizedburnratio(dNBR) maps Arizona. Burnseveritywas measuredusingdifferenced developedby the MonitoringTrendsin Burn Severityproject.Geospatial data setsfromthe LANDFIRE projectwereused to controlforprefirevariabilityin canopy cover,fuels,and thatincorporatedprescribedburningweremore topography.Acrossall threefires,treatments than thinningalone. Treatmenteffectsizes were lower,and standarderrorswere effective higherin the SAR models than in the OLS models. Spatial errortermsin the SAR models indirectlycontrolledfor confoundingvariables not captured in the LANDFIRE data, variabilityin fireweatherand landscape-leveleffectsof reducedfire includingspatiotemporal of carryingout severityoutsidethe treatedareas. This researchdemonstratesthe feasibility assessmentsof fuel treatmenteffectiveness using geospatial data sets and highlightsthe to controlforunmeasuredconfoundingfactors. potentialforusingspatial autoregression normalizedburnratio (dNBR); firebehavior;fireweather;fuel Key words: burnseverity;differenced treatments; spatialautoregression. Introduction Fuels managementhas emergedas thecornerstoneof effortsto mitigatethe impacts of large, destructive in thewesternUnitedStates.Fuels are of parwildfires ticularinterestfroma managementstandpointbecause, unliketopographyand weather,theycan be modified throughmanagementactivities.Various types of fuel are currently treatments beingappliedin westernforests. and subcanopytrees of overstory In particular,thinning has been proposedas a techniqueto reducehorizontal and verticalcontinuityof the forestcanopy and thus decrease the hazard of fast-movingand destructive crownfires(Agee and Skinner2005). Othertreatments such as burning,piling,mastication,and compaction can alter the amount, size distribution,and spatial of surfacefuels,reducingthe spread rate arrangement received9 September 2008; revised5 March Manuscript Editor:D. 2009; accepted25 March2009. Corresponding McKenzie. E-mail: michael.wimberly@sdstate.edu and intensityof surface fires(Kalabokidis and Omi 1998, Jermanet al. 2004, Stephens and Moghaddas have thepotentialto reducetree 2005). Fuel treatments at a local level(Ritchieet and otherfireeffects mortality al. 2007), limitthe rate of firegrowthacross broader landscapes(Finney2001), and increasetheeffectiveness of firesuppressionactivities(Moghaddas and Craggs 2007). However,despitethe widespreadapplicationof fuel treatments,there is only limited information in modifying available about theiractual effectiveness wildfire effects. effectiveness Most empiricalstudiesof fueltreatment have beencase studiesfocusingon one or a fewwildfires. studiesconductedin a varietyof forest Not surprisingly, methodologieshave produced ecosystemsusingdifferent results.For example,an analysisof a rangeofconflicting in Montana,Washington,California,and fourwildfires Arizona foundthat thinningalone, prescribedburning alone, or thinningfollowedby prescribedburningall reducedfireseverity comparedto untreatedareas (Pollet and Omi 2002). In contrast,on theCone firein northern 1377 1378 3 I MICHAEL C. WIMBERLY ET AL. California, tree mortality was lower in stands treated with thinning and prescribed fire, but not in stands treatedwith thinningalone, compared to untreated areas (Ritchie et al. 2007). On the Biscuit firein southwestern Oregon, fire severitywas higher in stands subjected to thinning and salvage logging than in untreated areas (Raymond and Peterson 2005, Thompson et al. 2007). To enhance fuels management efforts,an understanding of the conditions under which various types of fuel treatmentsreduce fire severity is needed. In particular, there is a need to analyze large numbers of fuel treatments across multiple fires and forest types to draw conclusions at regional to national levels. The availability of data from several national projects now provides an opportunity for carrying out this type is a multi-agency project of assessment. LANDFIRE that is producing geospatial data sets of vegetation, fuels, and fire regimes across the entire United States (Rollins and Frame 2006). The Monitoring Trends in Burn Severity (MTBS) project is another multi-agency effortusing satellite imagery to map burn severityfor all large wildfiresoccurring in the United States between 1984 and 2010 (Eidenshink et al. 2007). Both of these projects produce spatial data sets at a 30-m resolution that can be used to study burn severitypatterns within individual fires. Although a comprehensive national database of geolocated treatments does not currently exist, GIS data on recent fuel treatments is widely available from various public land management agencies. However, integrating these data to study fuel treatment effectivenessrequires a consistent analytical strategy that can be applied to multiple fires across a variety of ecosystems. There are several methodological challenges associated with using geospatial data sets to assess treatment effectiveness.One issue is that fuel treatments are not randomly located within fires,and are thereforeusually confounded with other environmental variables. A treated stand with fireseveritylower than an untreated stand may result from the treatment itself, or from differencesin topography and vegetation characteristics between the treated and untreated areas. For this reason, simple overlays of treatment polygons onto burn severitymaps can result in misleading conclusions about the effectivenessof treatments. Regression analysis is frequentlyused in observational studies to make inferences about treatment effects conditional on the effectsof one or more confounding variables (Gelman and Hill 2007). However, not all confounding variables are measurable. For example, spatial and temporal variability in temperature, wind speeds, and fuel moisture all influencefirebehavior as the flaming front moves across the landscape. With geospatial data sets, one approach to this missing variable problem is to utilize spatial autoregressive models, in which unmeasured but spatially structured environmental variables can be modeled indirectly using a spatial error term (Haining 1990). Although several studies have used Ecological Applications,Vol. 19, No. 6 spatial autogression and related techniques to analyze fire severity patterns (Finney et al. 2005, Thompson et al. 2007, Wimberly and Reilly 2007), none has explicitly examined the information captured by the spatial error term. The main goals of this research were to develop a methodology for combining data from LANDFIRE and MTBS with spatial data on fuel treatment locations to quantify treatmenteffectson burn severity,and to apply this method in a case study to assess fuel treatmentson three fires occurring in differentecosystems. Specific research objectives were to assess the effectiveness of LANDFIRE data in controlling for the influences of fuels, vegetation, and topography on burn severity patterns; and to determine whether spatial autoregression can be used to capture spatiotemporal variability in fireweather, landscape-level treatmentinteractions, and other confounding factors not accounted for in the LANDFIRE data set. Methods Data sources Our study focused on threerecentwildfiresin the westernUnitedStatesthatburnedthrougha varietyof fuel treatments:the Camp 32 fire in the Rocky Mountains of northwesternMontana (372 ha), the School fire in the Blue Mountains of southeastern Washington(19 871 ha), and the Warm fire on the Kaibab Plateau in northern Arizona (23 490 ha). Additional site descriptionsand fire narrativesare provided in Appendix A. Maps of the differenced normalizedburnratio (dNBR) wereobtainedfromthe MonitoringTrendsin BurnSeverity(MTBS) projectas 30-m rasterGIS layers(Fig. 1). The normalizedburn ratio (NBR) was computed from pre- and postfire Landsat images using bands 4 and 7, and dNBR was betweenthepre-and postfire computedas thedifference images (Key and Benson 2005). The dNBR index has been shownto correspondwell to field-basedmeasurementsof fireseverityin a varietyof ecosystemsand to compare favorablywith other remotelysensed burn severityindices(Van Wagtendonket al. 2004, Cocke et al. 2005, Eptinget al. 2005,Wimberlyand Reilly2007). Canopy cover,fuelmodel,elevation,slope angle,and slope aspect data sets were obtained fromthe LANDFIRE project as 30-m rasterGIS layers. Aspect was via a cosinefunctionintoa heat load index transformed that was higheston southwestaspects and lowest on northeastaspects.Fuel modelswerefromtheexpanded set of 40 standardfirebehaviorfuelmodels (Scott and Burgan 2005), and were coded using treatmentcontrasts.For each fire,thefuelmodelcoveringthe largest area was selectedas thebaselineclass,and theotherfuel as indicatorvariables. modelswererepresented wereobtainedas shape Maps of recentfueltreatments filesfromthe Kootenai National Forest(Camp 32 fire) the Umatilla National Forest (School fire),and the Kaibab National Forest(Warm fire).These maps were September2009 ASSESSING FUEL TREATMENT EFFECTIVENESS 1379 i r-*v o g Fig. 1. Differencednormalizedburn ratio (dNBR) maps of the Camp 32 (westernMontana), School (southeastern Washington),and Warm(northernArizona) fireswithfueltreatment polygons. convertedto 30-mrasterGIS layers.We did notattempt to map all of thehistoricalmanagementactivitieswithin the fire boundaries. Instead, we focused on specific activitiesthat are currentlybeing applied for treating forestfuels, and that were recentenough for us to acquire accurate spatial data on treatmentlocations. Fuel treatments in the areas burnedby the threefires includedthinning followedbyprescribed alone, thinning burning,and prescribedburningalone (Fig. 1,Appendix A). On the Warm fire,the effectsof historicalshelterwood harvests and historical wildfires were also examined.Treatmentswerecoded as indicatorvariables withuntreatedareas as the baselineclass. Fire progressionmaps wereobtainedforSchool and Warmfires(AppendixA). Therewas no fireprogression map fortheCamp 32 firebecause it burnedforless than one day. These maps consistedof polygons,each of whichencompassedthearea burnedduringa particular thearea burned timeperiod.Most polygonsrepresented a of fire spread,althoughtherewas during singleday some variabilityin the temporalresolution.The fire progressionpolygonswerecoded as indicatorvariables with one time period set as the baseline class (5-6 AugustfortheSchool fireand 25-26 JulyfortheWarm fire).The rateof spreadwas computedforeach polygon as the ratio of the polygonarea to the lengthof time by thepolygon. represented 1380 MICHAEL C. WIMBERLY ET AL. The spatialweightsmatrix,W, was based on an inverse distance rule in which the 12 nearestneighborswere where djj was the assigned weights based on 1/</,,, R2 n AIC Fire and model distancefromthe focalcell /to neighbor/ Camp 32 The fourstatisticalmodelsforeach firewererankedin 13641 0.46 termsof theirfit to the dNBR data 1033 1)OLS using the AIC 0.75 1033 13026 3)SAR statistic (Burnham and Anderson 2002). We also School statisticsfor each model as the 718 755 0.43 computed pseudo-/?2 53 631 1)OLS correlation between theobserveddNBR values 714433 0.47 squared 2) OLS withfireprogression 53 631 and the fitted values and Agresti2000). The 679434 0.77 53 631 (Zheng 3)SAR 679 378 0.77 fittedvalues of the SAR models (Eq. 1) were decom4) SAR withfireprogression 53 631 Warm posed intomaps of thespatialtrend,spatial signal,and 0.22 noise components(Haining 1990). These maps were 58 726 756377 1)OLS 0.40 derived from model 740952 2) OLS withfireprogression 58 726 3, the SAR model withoutfire 0.75 58 726 699278 3) SAR of the signal termswere examined progression. Maps 1 0.75 726 699 34 4) SAR withfireprogression 58 visuallyto searchforpatternsof any importantdriving Note: Abbreviations are: OLS-, ordinary least-squares variablesthat were not capturedin the trendcomporegression;SAR, sequential autoregression;AIC, Akaike nent.For the School and Warm fires,the signal maps information criterion;w,samplesize foreach model. wereoverlainon thefireprogressionmaps and Pearson correlationcoefficients betweenmean signalvalues and Analysismethods therateof burningforeach fireprogressionperiodwere The analyseswerecarriedout on a 25% subsampleof calculatedto test the hypothesisthat the spatial error on the30-mpixels,withthesampleof pixelsdistributed capturesspatiotemporalvariabilityin fireweatherand a 60 X 60 m square lattice.All analyseswerecarriedout the resultingratesof firespread. in the R statisticalanalysisenvironment (R DevelopResults mentCore Team 2008). Four typesof statisticalmodels wereexamined,withdNBR as thecontinuousdependent For all three fires,the OLS models without fire variable for each model: (1) ordinary least-squares progressionhad the weakestfitas quantifiedby high (OLS) regressionwithindependentvariablescharacter- AIC valuesand low R2 values(Table 1). Addingthefire fuels,vegetation,and fueltreatments; progressionvariablesto theOLS modelsforthe School izingtopography, with date of burningadded to theindependent and Warm firesimprovedfitslightly.The SAR models OLS (2) variablesfrommodel 1; (3) simultaneousautoregression withoutfireprogressionimprovedmodel fitconsidervariablescharacterizing topog- ably compared to the OLS models. The SAR models (SAR) withindependent (4) SAR with fire progression improved model fit slightly raphy,fuels,vegetation,and fuel treatments; variables comparedto the SAR modelswithoutfireprogression. withdate of burningadded to theindependent models were The model coefficients, frommodel 3. The spatial autoregressive standarderrors,and the results fittedusingthespautolmfunctionfromthespdepspatial of statisticaltestson thecoefficients weresimilarforthe analysispackagein R (Bivand 2002). Sample R code for SAR models fittedwith and withoutfireprogression the models is variables(AppendixB). These resultsdemonstrated that computingthe spatial weightsand fitting (1) the spatialerrortermin the SAR modelsaccounted providedin the Supplement. The SAR models explicitlyaccounted for spatial fornearlyall of the variabilityin fireseveritythatwas a spatialtermintothe captured by the fire progressionmaps, and (2) the autocorrelation by incorporating decision to include or exclude the fire progression standardregressionmodel: variablesin the SAR models had a minimalinfluence = Y Xp + A.W(Y XP) + e effectson burnseverity. about treatment on inferences whereY is a vectorofdependentvariables,X is a matrix in theSAR models effects In mostcases,thetreatment Xis werelowerthan in the comparableOLS models (Table ofindependent variables,P is a vectorofparameters, theautoregressive parameterthatcapturesthedegreeof 2). On the School fire,the positiveeffectof prescribed autocorrelation among neighboringobserva- burningwas higherin the SAR model thanin the OLS spatial a W is tions, spatial weightsmatrix,and £ is a vector model, althoughthe effectwas not statistically signifion all three of uncorrelatederrors.The spatial trend,Xp, reflected cant in eithermodel. For all fueltreatments thatwas predicted fires,the standarderrorsof the treatmenteffectswere in burnseverity thespatialvariability effects. higherin the SAR models than in the OLS models. by LANDFIRE variablesas well as treatment based on effectiveness about treatment The spatial signal, XW(Y - XP), captured spatially Thus, inferences autocorrelateddeviations from the trend that were OLS overestimatedboth the effectsize and statistical We therefore based our of fueltreatments. functionof deviationsin significance modeledas an autoregressive without the models effects on SAR deviations of treatment The noise sites. term, £, analysis captured neighboring fromthe trendthat were not spatiallyautocorrelated. fireprogressionvariablesforthe Camp 32 fire,and the normalizedburn Table 1. Regressionmodels of differenced ratio(dNBR) forthreewildfires. I 5 Ecological Applications,Vol. 19, No. 6 September2009 ASSESSING FUEL TREATMENT EFFECTIVENESS 1381 modelsof dNBR forthreewildfires. Table 2. Treatmenteffectsfromordinaryleast-squaresand sequentialautoregression Ordinaryleast squares Sequentialautoregression P SE P P SE Camp 32J Thinning burn Thinning/prescribed 96.12 -178.48 13.26 22.13 O.001 <0.001 57.41 -77.50 25.53 35.04 0.025 0.027 School§ Prescribedburn Thinning burn Thinning/prescribed 1.48 -110.39 -145.10 5.15 10.40 5.42 0.774 <0.001 <0.001 11.46 -26.81 -27.43 12.08 12.93 6.76 0.343 0.038 <0.001 -20.70 -40.54 204.47 -39.25 2.75 7.05 22.82 16.95 <0.001 <0.001 <0.001 0.021 -14.11 -43.14 69.23 -26.53 3.35 15.56 24.87 23.02 <0.001 0.006 0.005 0.249 Fire and treatmentf Warm§ Shelterwood Prescribedburn Thinning Historicalwildfire P t Coded as indicatorvariableswithuntreatedas the baselineclass. variables(models 1 and 3 in Methods:Analysis %The OLS and SAR modelsfortheCamp 32 firedid notincludefireprogression methodsand Table 1). § The OLS and SAR modelsforthe School and Warm firesincludedfireprogressionvariables(models 2 and 4 in Methods: Analysismethodsand Table 1). evidentin theWarmfire(Fig. 2c), in which SAR models with fire progressionvariables for the particularly School and Warmfires. high spatial signal values were concentratedin the For the Camp 32 fire (Table 2), thinningalone easternand southernportionsof the burn.This is the firebehaviordrove rapid increasedburn severity,whereas the combinationof area whereplume-dominated from 25 Juneto 26 June. the southwest toward severburn decreased and spread prescribedburning thinning zone of low a distinctive fire exhibited School The and of the combination the School For fire, thinning ity. of the in the southeastern values than burn in lower resulted portion signal spatial severity burning prescribed alone also decreasedburn burnedarea (Fig. 2b). The boundariesof thiszone did untreatedareas,and thinning The effectof prescribedburningalone was not not correspondwiththefireprogressionmap, but were severity. For the Warm fire(Table 2), insteadassociatedwithan area wheretherewas a high statisticallysignificant. and shelterwoodharvesting density of fuel treatments.This pattern provided alone prescribedburning effect,in which thanuntreatedareas. In evidenceof a landscape-leveltreatment resultedin lowerburnseverities contrast,thinningalone increasedburn severityover influencesof fuel treatmentson firebehaviorreduced boundaries. untreatedareas. The small negativeeffectof previous burnseverityoutsideof the fueltreatment not was wildfires Canopy cover, statistically significant. Discussion one or moretopographicvariables,and one or morefuel Results of these analyses supportthe assertionthat models were statisticallysignificantin all three fires (Appendix B). Most of the fireprogressionvariables treatmentof surfacefuels is criticalfor fireseverity were also statisticallysignificantin the School and reduction (Agee and Skinner 2005). In particular, Warm fires,reflecting higherfireseverityduringdays prescribedburninghas been argued to be the most withextremefireweatherat thebeginningof theSchool effectivetreatmentfor reducingthe fine surfacefuel of firebehaviorand loadingsthathave a majorinfluence fire,and at theend of theWarmfire(AppendixB). In the SAR model forthe Camp 32 fire,the spatial fire effects.In our study, combined thinningand signalwas generallyhighestin thecentralportionof the prescribedburningreducedburnseverityon the Camp burn and lowest at the edges (Fig. 2a). These signal 32 and School fires,and prescribedburning alone patterns likely representedtemporal variability in reduced burn severityon the Warm fire.In contrast, weatherand firebehaviorthroughoutthe course of the thinningwithout treatmentof the resulting slash heterogene- actuallyincreasedburn severityon the Camp 32 and day,and may have also capturedtreatment ity withinthe thinnedand burned areas. The mean Warm fires.This result corroboratesstudies of the Oregonwhichfoundhigher spatial signal fromthe SAR model withoutfirepro- Biscuitfirein southwestern recentforestmanagement with areas in fire correlated was variables severity positively (model 3) gression with the mean rate of fire spread for each fire activitiesthan in unmanaged areas (Raymond and progressionpolygonforboth the School fire(r = 0.70, Peterson2005, Thompsonet al. 2007). The situationin P = 0.05) and theWarmfire(r = 0.75, P = 0.005). These the thinnedareas on the Camp 32 and Warm fires relationshipsindicated that the spatial signal was treatmentsrepresenta worst-casescenario in which a associatedwithspatiotemporalpatternsof fireweather wildfireoccurred before the thinningtreatmentwas and the resultingfirebehavior.This relationshipwas completedand loadingsof surfacefuelswere therefore o 1382 MICHAEL C. WIMBERLY ET AL. Ecological Applications,Vol. 19, No. 6 o Fig. 2. Spatial signalmaps of the Camp 32, School, and Warm fireswithfueltreatment polygons. particularlyhigh. In contrast,four-year-oldthinning (Gelman and Hill 2007). For all threefiresthat we treatmentson the School fire and 14-18-year-old analyzed,canopy cover and at least one topographic shelterwoodharvestson the Warm firereduced burn variable and fuel model had a statisticallysignificant severity comparedto untreatedareas. emphasizingtheimporrelationshipwithburnseverity, effectiveness tanceof includingthesevariablesin assessmentsof fuel Empiricalassessmentsof fueltreatment are by necessitybased on retrospective, observational treatments effects.However,theseanalysescan stillbe studiesof wildfires thathave burnedover one or more confoundedby omittedor "lurking"variables,particufueltreatments. Therefore,it is not possibleto random- larlythoserelatedto fireweather.Althoughtherehave ize treatmentlocations with respect to vegetation, been attemptsto incorporatefireprogressionmaps and topography,weather,and otherfactorsthataffectburn weatherdata into spatial analysesof fireseverity(e.g., severity;and comparisonsof the average burn severity Collins et al. 2007), theyhave only been partiallysucamong treatmentsare typicallyconfoundedby these cessfulbecause of the coarse scale of fireprogression betweenconditionsat otherfactors.Regressionanalysiscan be used to control polygonsand thelargedifferences fortheseeffects ifall confounding variablesare included the flamingfrontand weatherdata collected at the in the model, and if the model is correctlyspecified neareststations. 2009 September ASSESSING FUEL TREATMENTEFFECTIVENESS Spatial regressionanalysisprovidesan approach for information about fireweatherand other incorporating omittedvariables.The phenomenonof spatial autocorrelationcan be framedas a missingvariablesproblem,in which the patternof model errorsrepresentsone or variablesthatare morespatiallystructured independent missingfrom the regressionmodel (Ver Hoef et al. 2001). In thecase of theSAR model,theseunmeasured variablesare modeledindirectly by a spatialsignalterm thatcan be computedand mapped foreach pixel. We foundthatsignalvalues werehighestin theportionsof thefiresthatburnedmostrapidly,indicatingthatburn was highestduringperiodsof rapid firespread severity and extremefirebehavior.Even when indicatorvariables forthe fireprogressionpolygonsare includedas independentvariables,SAR models substantiallyimprovedthefitcomparedto theOLS models.This finding reflectsthe abilityof the SAR models to capture the influencesof weathervariabilityat finerspatial and temporalscales than fireprogressionmaps, in which each polygontypically encompassesone to threedays of firegrowth. Spatial regressionmodelingcan also provideinsights into other missingvariables and processes affecting patternsof burn severity.On the School fire,a large clusterof low signalvalues was associatedwithan area We interthat had a high densityof fuel treatments. of a as evidence this landscape-level pattern preted offire in whichthecombinedinfluences treatment effect, withthetreatedareas reducedburnseverity interactions evenin nearbyareas thatwereoutsidethetreatedunits. These types of landscape-level effects have been demonstratedin simulationmodelingstudies (Finney 2001), and have been observed in the burn severity patternsresultingfromthe 2002 Rodeo and Chediski fires (Finney et al. 2005). In the present analysis, mappingthe spatial signal allows spatiallystructured patternsof burnseverityto be separatedfrompatterns relatedto vegetation,fuels,and topography,and from errorterm.Thus, it is possibleto verify theuncorrelated thatthe observedclusterof low burnseverityis not an artifactof one or more confoundingvariables, but instead likelyrepresentsa true landscape-leveleffect of firespreadand thehigh fromtheinteraction resulting densityof fueltreatments. Although the present study addressed only three wildfiresoccurringin ponderosa pine forestsof the westernUnited States,the approach could be adapted and applied in a varietyof ecosystems.We focusedon dNBR as a burn severitymetricbecause of its widespread use and acceptancein the firemanagementand forestrycommunities.However, the same modeling approachcould be used withrelativeversionsof dNBR (Millerand Thode 2007) or withany othercontinuous burn severitymetric. Differentsets of confounding variableswill likelybe importantin different regions, and various types of model-selectionmethods (e.g., Burnhamand Anderson2002) could be appliedto select 1383 a parsimoniousset of covariates.Differentspatial regressionmodelingapproaches,such as generalizedleast squares (Thompson et al. 2007) could also be applied. AlthoughGLS and related methods allow for more precisemodelingof spatial autocorrelationvia a semiforanalyzing variogram,we chose spatialautoregression fueltreatment effectiveness because of the potentialfor spatial analysisof largedata sets (>50000 samplesfor the School and Warm fires)using the sparse matrix algorithmsavailable in R. Conditional autoregression (CAR) could also be used (Finneyet al. 2005), although we used the SAR model in thisapplicationbecause its allows fora more straightforward decomspecification positionof the responseinto spatial trendand spatial signalcomponents(Haining 1990). In conclusion,LANDFIRE and similardata products can be used to controlfor at least some of the confoundingeffectsof vegetation,topography,and other environmentalvariables in assessmentsof treatment effectiveness.However, obtaining accurate measurements of other potential confounderssuch as fire weatherstillpresentsa challenge.Spatial autoregression can be used to indirectly controlfortheseunmeasured variablesby modelingthemvia the spatial signalterm. The spatial signal can also be used to visualize other unmeasured effects,such as landscape level effects resultingfrom the interactionof fire spread with multipletreatments. Acknowledgments Fundingforthisresearchwas providedbytheUSDA/USDI JointFire Science Program.Chris Barber, Don Ohlen, and Narayana Ganapathyprovidedassistanceduringfieldvisitsto the Camp 32 and Warm fires.JeffEidenshink,Mark Finney, Matt Reeves, and Zhi Liang Zhu provided feedback and assistanceat variousstagesof thisresearch. Literature Cited Agee,J. K., and C. N. Skinner.2005. Basic principlesof forest Forest Ecology and Management fuelreductiontreatments. 211:83-96. Bivand, R. 2002. Spatial econometricsfunctionsin R: classes and methods.Journalof GeographicSystems4:405-421. Burnham,K. P., and D. R. Anderson.2002. Model selection and multimodelinference:a practicalinformation-theoretic Verlag,New York, New approach.Second edition.SpringerYork, USA. Cocke, A. E., P. Z. Fule, and J.E. Crouse.2005. Comparisonof normalizedburn burnseverityassessmentsusingdifferenced ratio and ground data. InternationalJournalof Wildland Fire 14:189-198. Collins, B. M., M. Kelly, J. W. van Wagtendonk,and S. L. Stephens. 2007. Spatial patternsof large natural firesin SierraNevada wildernessareas. Landscape Ecology22:545557. Eidenshink,J., B. Schwind,K. Brewer,Z. L. Zhu, B. Quayle, and S. Howard. 2007. A projectfor monitoringtrendsin burnseverity.Fire Ecology 3:3-21. Epting,J., D. Verbyla,and B. Sorbel. 2005. Evaluation of in interior sensedindicesforassessingburnseverity remotely Alaska using Landsat TM and ETM+. Remote Sensingof Environment 96:328-339. Finney,M. A. 2001. Designof regularlandscapefueltreatment patternsfor modifyingfiregrowthand behavior. Forest Science47:219-228. 3 o g 1384 so MICHAEL C. WIMBERLY ET AL. Finney,M. A., C. W. McHugh,and I. C. Grcnfell.2005. Standand landscape-leveleffectsof prescribedburningon two Arizonawildfires. Canadian Journalof Forest Research35: 1714-1722. Gelman,A., and J. Hill. 2007. Data analysisusing regression models. Cambridge University and multilevel/hierarchical Press,New York, New York, USA. Haining, R. 1990. Spatial data analysis in the social and sciences.CambridgeUniversityPress,Camenvironmental bridge,UK. Jerman,J. L., P. J. Gould, and P. Z. Fule. 2004. Slash compressiontreatmentreduced tree mortalityfrom prescribedfirein southwestern ponderosapine.WesternJournal of AppliedForestry19:149-153. Kalabokidis, K. D., and P. N. Omi. 1998. Reductionof fire hazard through thinningresidue disposal in the urban Journalof WildlandFire 8:29-35. interface. International Key, C. H., and N. C. Benson. 2005. Landscape assessment: the compositeburnindex,and groundmeasureof severity, the normalizedburnratio.Pages remotesensingof severity, LA-l-LA-51 in D. C. Lutes, R. E. Keane, J. F. Carattei, C. H. Key, N. C. Benson,S. Sutherland,and L. J. Gangi, editors.FIREMON: fireeffectsmonitoringand inventory system. CD-ROM. RMRS-GTR-146CD. USDA Forest Service,Rocky Mountain Research Station,Ogden, Utah, USA. burnseverity Miller,J.D., and A. E. Thode. 2007. Quantifying in a heterogeneous landscapewitha relativeversionof the delta normalizedburn ratio (dNBR). Remote Sensing of 109:66-80. Environment Moghaddas, J. J., and L. Craggs. 2007. A tuel treatment in a and increasessuppressionefficiency reducesfireseverity Journalof WildlandFire mixedconiferforest.International 16:673-678. Pollet, J., and P. N. Omi. 2002. Etlect ot thinningand prescribedburningon crownfireseverityin ponderosapine Journalof WildlandFire 11:1-10. forests.International R Development Core Team. 2008. R: a language and environmentfor statisticalcomputing.R Foundation for StatisticalComputing,Vienna,Austria. Raymond,C. L., and D. L. Peterson.2005. Fuel treatments alter the effectsof wildfirein a mixed-evergreen forest, Ecological Applications,Vol. 19, No. 6 Oregon, USA. Canadian Journalof Forest Research 35: 2981-2995. Ritchie,M. W., C. N. Skinner,and T. A. Hamilton. 2007. Probabilityof treesurvivalafterwildfirein an interiorpine forest of northernCalifornia: effectsof thinningand prescribedfire.Forest Ecology and Management247:200208. Rollins, M. G., and C. K. Frame. 2006. The LANDFIRE prototypeproject:nationallyconsistentand locallyrelevant geospatial data for wildland fire management.General technicalreport RMRS-GTR-175. USDA Forest Service Rocky Mountain ResearchStation,Fort Collins,Colorado, USA. Scott,J.H., and R. E. Burgan.2005. Standardfirebehaviorfuel models:a comprehensive setforuse withRothermal'ssurface firespread model. RMRS-GTR-153. USDA ForestService, Rocky Mountain ResearchStation,Fort Collins,Colorado, USA. Stephens,S. L., and J. J. Moghaddas. 2005. Experimentalfuel treatment impactson foreststructure, potentialfirebehavior, and predictedtree mortalityin a Californiamixedconifer forest.ForestEcologyand Management215:21-36. Thompson,J. R., T. A. Spies,and L. M. Ganio. 2007. Reburn severityin managed and unmanagedvegetationin a large wildfire.Proceedingsof the National Academyof Sciences (USA) 104:10743-10748. Van Wagtendonk,J. W., R. R. Root, and C. H. Key. 2004. Comparison of AVIRIS and Landsat ETM+ detection capabilitiesfor burn severity.Remote Sensingof Environment92:397-408. Ver Hoef,J.M., N. Cressie,R. N. Fisher,and T. J.Case. 2001. Uncertaintyand spatial linear models for ecological data. Pages 214-237 in C. T. Hunsaker,M. F. Goodchild,M. A. in ecology: Friedl,and T. J.Case, editors.Spatial uncertainty implications for remote sensing and GIS applications. Springer,New York, New York, USA. Wimberly,M. C, and M. J. Reilly.2007. Assessmentof fire severityand species diversityin the southernAppalachians usingLandsat TM and ETM+ imagery.RemoteSensingof Environment108:189-197. thepredictive Zheng,B. Y., and A. Agresti.2000. Summarizing power of a generalizedlinearmodel. Statisticsin Medicine 19:1771-1781. APPENDIX A narrativesforthe Camp 32, School, and Warmfires(EcologicalArchivesA019-057-A1). and wildfire Site descriptions APPENDIX B P values for the ordinaryleast-squaresand sequentialautoregressionmodels and standarderrors, Tables of coefficients, (EcologicalArchivesA019-057-A2). SUPPLEMENT effectson burn severityusing ordinaryleast-squaresregressionand fuel treatment for data and R analyzing Sample script (EcologicalArchivesA019-057-S1). sequentialautoregression