Ecological Applications, 15(4), 2005, pp. 1317±1330 ᭧ 2005 by the Ecological Society of America

IMPACTS OF LARGE-SCALE ATMOSPHERIC±OCEAN VARIABILITY ON ALASKAN FIRE SEASON SEVERITY

PAUL A. DUFFY,1,5 JOHN E. WALSH,2 JONATHAN M. GRAHAM,3 DANIEL H. MANN,4 AND T. S COTT RUPP1 1Ecological Dynamics Modeling Group, Department of Forest Sciences, University of Alaska, Fairbanks, Alaska 99775 USA 2International Arctic Research Center, Fairbanks, Alaska 99775 USA 3Department of Mathematical Sciences, University of Montana, Missoula, Montana 59812 USA 4Institute of Arctic Biology, University of Alaska, Fairbanks, Alaska 99775 USA

Abstract. Fire is the keystone disturbance in the Alaskan boreal forest and is highly in¯uenced by summer weather patterns. Records from the last 53 years reveal high vari- ability in the annual area burned in Alaska and corresponding high variability in weather occurring at multiple spatial and temporal scales. Here we use multiple linear regression (MLR) to systematically explore the relationships between weather variables and the annual area burned in Alaska. Variation in the seasonality of the atmospheric circulation±®re linkage is addressed through an evaluation of both the East Paci®c teleconnection ®eld and a Paci®c Decadal Oscillation index keyed to an annual ®re index. In the MLR, seven explanatory variables and an interaction term collectively explain 79% of the variability in the natural logarithm of the number of hectares burned annually by lightning-caused ®res in Alaska from 1950 to 2003. Average June temperature alone explains one-third of the variability in the logarithm of annual area burned. The results of this work suggest that the Paci®c Decadal Oscillation and the East Paci®c teleconnection indices can be useful in determining a priori an estimate of the number of hectares that will burn in an upcoming season. This information also provides insight into the link between ocean±atmosphere interactions and the ®re disturbance regime in Alaska. Key words: Alaska boreal forest; East Paci®c teleconnection; ecological disturbance regimes; ®re regimes; multiple linear regression; Paci®c Decadal Oscillation; teleconnections.

INTRODUCTION tivities and developing a better understanding of how forecast climate change might impact the dominant dis- The boreal forest covers 12 ϫ 106 km2 of the northern turbance mechanism. hemisphere and contains roughly 40% of the world's Within the North American boreal forest, Interior reactive soil carbon, an amount similar to that held in Alaska (i.e., the region between the Alaska and Brooks the atmosphere (Melillo et al. 1993, McGuire et al. Ranges) contains 56 ϫ 106 burnable hectares and in- 1995, IPCC 2001). The biophysical phenomena af- cludes the largest national parks and wildlife refuges fecting carbon storage and high-latitude albedo make in the United States. Most of this huge area is roadless. the boreal forest an integral component of the global climate system (IPCC 2001). Fire-initiated succession For the period 1950±2003, wildland ®res burned an underlies the biophysical factors, and there is a pressing average of roughly 270 000 ha in Interior Alaska each need to characterize sensitivities and potential respons- year and they routinely threaten the lives, property, and es of the boreal forest disturbance regime to climatic timber resources of the sparse but growing population change (Schimel et al. 1997, Gower et al. 2001, Chapin (see Plate 1). Wildland ®res can threaten human values, et al. 2003). The impact of forecast climatic warming yet they play a crucial role in the maintenance of In- on ®re regimes in North America varies from a pre- terior Alaskan ecosystems. Despite the pervasive eco- diction for increased burning for Alaska and Canada nomic and ecological impacts, fundamental aspects of (Flannigan et al. 2000, 2001) to reduced ®re frequency the ®re regime in Interior Alaska are poorly understood. in eastern Canada (Carcaillet et al. 2001). Quanti®ca- Fire regimes consist of many components including tion of the links between climate and ®re in Alaska is frequency, duration, intensity, severity, seasonality, ex- not only essential for understanding the dominant land- tent, and spatial distribution. When complicating fac- scape-scale disturbance processes in Alaska, but it is tors such as interactions with other components of the also a valuable tool for planning ®re management ac- ecosystem (e.g., human impacts, weather, vegetation) and the importance of spatial and temporal scales are taken into account, the characterization of a ®re regime Manuscript received 12 May 2004; revised 4 August 2004; accepted 13 September 2004; ®nal version received 6 December requires a tremendous amount of data and appropriate 2004. Corresponding Editor: M. Friedl. analysis. One of the most basic aspects of a ®re regime 5 E-mail: [email protected] is the ®re cycle (i.e., the ®re recurrence interval for an 1317 1318 PAUL A. DUFFY ET AL. Ecological Applications Vol. 15, No. 4

PLATE 1. The Long Creek ®re of 2002, near Ruby, Alaska. Fire is the dominant landscape-scale disturbance mechanism in Interior Alaska. Photo credit: T. S. Rupp. area equivalent to the study area). There are only a few Alaska reveal that shifts in vegetation (i.e., increased ®eld studies from Interior Alaska that utilize ®re-scar dominance of Picea mariana) around 2400 yr BP are and/or tree age distributions to infer ®re cycle (Yarie associated with a corresponding shift in the ®re regime 1981, DeVolder 1999, Mann and Plug 1999). These (Lynch et al. 2003). The implication is that climatic studies were scattered over a region the size of Montana change occurring at decadal to centennial timescales and show that ®re cycles in Alaska are probably Ͼ250 in¯uenced the ®re regime through shifts in dominant years in the relatively moist, southern parts of the state vegetation. This yields the counterintuitive result that (D. H. Mann, personal communication), 80±100 years cooler and moister climate results in higher ®re fre- near Fairbanks in the central Interior (Mann et al. quency due to the increased dominance of the relatively 1995), and Ͻ80 years for the Porcupine River valley more ¯ammable Picea mariana. This response appears in the northeastern portion of the state (Yarie 1981). to extend outside Interior Alaska to forests south of the These estimates are consistent with the results of Kas- Alaska Range as well (Lynch et al. 2004). These studies ischke et al. (2002), who used ®re perimeter data (from provide evidence that climatically induced shifts in aerial photography and remotely sensed data) from the dominant vegetation within the boreal forest over a past ®ve decades to estimate ®re cycles for different period of decades to centuries can potentially modify ecoregions of the interior. The uncertainty associated the ®re regime. Outside of Alaska in the Canadian bo- with these ®re cycle estimates is unknown. real forest, there is evidence that, on timescales of hun- At longer temporal scales but somewhat more coarse dreds to thousands of years, climate has a more direct resolution, charcoal and pollen analyses from varved in¯uence on ®re regime (Carcaillet and Richard 2000, lake sediments reveal critical information about the ®re Carcaillet et al. 2001). Hence, climate differentially frequency and interactions between ®re, climate, and exerts in¯uences on both vegetation composition and vegetation. Due to the limited dispersal of charcoal ®re regime, depending on the location within the boreal particles that are used in these analyses, the results forest as well as the resolution of the timescale of in- apply to limited spatial scales. Within Interior Alaska, terest. there have only been a few studies that utilize sediment As a mechanism that modi®es atmospheric circula- cores to gain insight about the ®re regime. Pollen and tion patterns at large spatial scales, atmospheric tele- charcoal data from several sediment cores in Interior connections affect weather throughout the northern August 2005 FIRE AND CLIMATE LINKAGE IN ALASKA 1319 hemisphere (Hurrell et al. 2003). Teleconnections are correlated anomalies of geopotential height (Wallace and Gutzler 1981, Barnston and Livezey 1987) that impact regional weather through recurring and persis- tent shifts in pressure and circulation across large spa- tial scales. Links between disturbance and weather that are mediated by teleconnections include; droughts and ®re in Canada (Bonsal et al. 1993, Bonsal and Lawford 1999, Skinner et al. 2002, Girardin et al. 2004), ®res in the Paci®c Northwestern United States (Hessl et al. 2004) and ®res in the Southwestern United States (Swetnam and Betancourt 1990). In Alaska, deviations from synoptic weather patterns have been correlated with the Paci®c Decadal Oscillation (Papineau 2001,

Hartmann and Wendler 2003) as well as the El NinÄo/ FIG. 1. Map of Alaska, USA, identifying the seven cli- Southern Oscillation and Paci®c/North America pat- mate stations used in the statistical analyses. The enclosed terns (Hess et al. 2001). Speci®cally, the occurrence of areas with darker shading represent individual ®re perimeters large ®re years has been correlated with the presence from 1950 to 2003. of strong to moderate El NinÄo conditions (Hess et al. 2001). Our work moves a step further and quanti®es Cleve and Viereck 1983, Payette 1992, Mann and Plug the impact of these signals on the annual area burned 1999). in Alaska through the development of a statistical re- gression model. METHODS Experience and common sense dictate that ®re re- Fire data sponds to local weather conditions, but modeling re- sults indicate that the link between weather and ®re The Alaska Fire Service (AFS) maintains a database does not easily translate to the landscape scale (Flan- of ®res for the state of Alaska dating back to 1950. nigan and Harrington 1988, Hely et al. 2001, Wester- These data are commonly referred to as the Large Fire ling et al. 2002). Yet at large spatial scales, statistical Data Base (LFDB) and can be found at the Alaska Geospatial Data Clearinghouse (AGDC; available on- relationships quantifying these links at an annual tem- line)6 (Kasischke et al. 2002). The period of analysis poral resolution in Alaska have not been established. for our study is 1950±2003 (Fig. 1). For each year and It is important to stress the spatiotemporal scale at for each ®re recorded in the AGDC, there are many which this analysis is relevant, since the dynamics of pieces of information recorded, including ignition interactions between climate, ®re, and vegetation source (i.e., human or lightning) and a record of the change depend on the spatial and temporal scale of approximate size of each ®re. Estimates of the annual interest. This work quanti®es the relationship between number of hectares burned in Alaska that are used in monthly teleconnection indices, speci®cally the East this analysis come from summing the number of hect- Paci®c teleconnection and the Paci®c Decadal Oscil- ares burned from each lightning-caused ®re within a lation, and the annual area burned by lightning-caused given year. Because lightning is the natural ignition ®res in Alaska through the development of a statistical source, part of the climate±®re link includes the cli- regression model. To this end, a Multiple Linear Re- matic conditions that are favorable for lightning-caused gression (MLR) model is developed, with the natural ignitions (Johnson 1992, Nash and Johnson 1996, logarithm of the number of hectares burned annually Wierzchowski et al. 2002). Lightning-caused ®res are as the response variable and monthly climatic indices rare both north of the Brooks Range, and south of the for explanatory variables. We used a sequential selec- Alaska Range, and as a consequence the majority of tion procedure to evaluate linkages between telecon- ®res burn between these mountain ranges within the nection indices and monthly temperature and precipi- region known as Interior Alaska (Fig. 1). tation. In doing so, we characterize the linkages be- Human-caused ®res were excluded, since the anal- tween components of the climate system that exert an ysis is focused on the link between climate and ®re. in¯uence on short-term surface weather in Alaska. The Historically, the exclusion of human-caused ®res has statistical model presented in this work represents a a relatively small impact on the number of hectares simple ®rst step in quantifying the complex linkages burned annually. However, in 2001 and 2002, over 95% between climate and ®re in Interior Alaska, where ®re and 40% of the respective area burned was caused by is the dominant agent of landscape-level change (Van nonlightning ignition sources. Another factor to con- Cleve et al. 1991, Payette 1992) and dictates the com- sider regarding the use of the LFDB is that reliability position of the forest vegetation through determination of the successional trajectory (Zackrisson 1977, Van 6 ͗http://agdc.usgs.gov/data/blm/®re/index.html͘ 1320 PAUL A. DUFFY ET AL. Ecological Applications Vol. 15, No. 4 of source determination can reasonably be assumed to Additionally, we re®t our regression model using be more questionable for the early part of the record. monthly temperature and precipitation data from the Only ®res that burned an area Ͼ50 ha (0.5 km2) were Climate Research Unit (CRU) in place of the seven- included in the analysis. This was done so that the station average. The CRU data consist of data-based results of this work could be used in conjunction with model estimates of monthly temperature and precipi- the frame-based spatially explicit ecosystem model tation values for half-degree cells and are a distance- ALFRESCO (Star®eld and Chapin 1996, Rupp et al. weighted average of all available station data in Alaska. 2000), which operates on a spatial resolution of 1 km2. Hence, a different number of stations are used de- The exclusion of ®res Ͻ50 ha (0.5 km2) has a negligible pending on which data are available for a given month. impact on the output of the statistical model. We averaged the monthly values for cells covering the region of Interior Alaska where ®res burn (Fig. 1) to Climate data produce a CRU-based data set that is capable of driving Alaskan climate station data that are both homoge- our statistical model. The reduction in variability ex- neous and complete for the past half-century are sparse. plained by the statistical model when the CRU data set Data were obtained from the Western Region Climate is used was Ͻ5%. This shows that the results of the Center (WRCC; available online).7 There are fewer than regression model are robust with respect to the method a dozen climate stations in Alaska with suf®cient data of spatially integrating the climate data, and it also for our modeling purposes (Fig. 1). The requirement for shows that the simple seven-station average does a bet- selection of a climatic station was that no more than 5% ter job of representing the weather signals that explain of the monthly observations from 1950 to 2003 were interannual variability in area burned. missing. Based on this selection criterion and spatial representativeness with respect to ®res that have burned East Paci®c (EP) teleconnection data over the period of record, the following seven stations Monthly teleconnection indices were obtained from were used: Bettles, Delta Junction, Fairbanks, McGrath, the National Oceanic and Atmospheric Administration Nome, Northway, and Tanana (Fig. 1). (NOAA) Climate Prediction Center (CPC; available The question of how to represent both the temper- online).8 Each teleconnection identi®ed by the CPC ature and precipitation of Alaska based on these seven (e.g., the North Atlantic Oscillation, East Atlantic, East stations is not a trivial one. We are essentially deter- Atlantic Jet, East Atlantic/Western Russia, Scandina- mining from a small number of stations a spatial zone vian, Polar/Eurasia, Asian Summer, West Paci®c, East of weather in¯uence for the area burned annually in Paci®c, North Paci®c, Paci®c/North America, Tropical/ Alaska. Several methods for calculating representative Northern Hemisphere, and the Paci®c Transition) was weather were evaluated. Different weighting schemes evaluated for its ability to explain variability in the for the climate data were considered based on the spa- natural logarithm of the number of hectares burned tial location of ®res for a given year. As an example, annually in Alaska. The monthly indices for the East if the majority of ®res burned near Tanana in a given Paci®c (EP) teleconnection collectively explained the year, then a weighting scheme would give the data from greatest amount of variability. Since the EP has a strong Tanana greater weight than the other six stations when center near Alaska and a comparable but oppositely assembling the weather data for that year. Of the dif- signed locus between Hawaii and the Baja peninsula ferent methods evaluated for assembling the climate (Fig. 2, after Barnston and Livezy 1987), this ®nding data, the simplest and most effective in terms of ex- is physically plausible. Strong positive phases of the plaining the greatest variability was a simple average EP pattern are associated with upper air¯ow that is of the data for a given month from the seven stations. more meridional over the northeastern Paci®c. This re- This results in a ``statewide'' average of both temper- sults in enhanced ridging over Alaska and the western ature and precipitation for each month in a given year. coast of North America. Alternatively, negative phases This procedure essentially integrates any spatial infor- of the EP pattern are associated with increased zonal mation regarding intrastate variability over Interior ¯ow and strengthened westerlies in the Eastern North Alaska into a single estimate. For each station, the Paci®c as a consequence of negative height anomalies average monthly temperatures from the WRCC are cal- to the north and positive anomalies to the south of 40Њ± culated as the average of all the average daily tem- 45Њ N. The negative phase is also associated with an peratures in a given month. The monthly precipitation eastward displacement of the Aleutian Low (AL), in for each station was calculated as the sum of the pre- the North Paci®c. cipitation amounts recorded daily at each station for each day in the month. The average monthly precipi- Paci®c Decadal Oscillation tation for the MLR was then taken to be the average Signi®cant covariability exists between Sea Surface of the monthly precipitation recorded for each of the Temperatures (SSTs) and North American climate seven stations. 8 ͗http://www.cpc.ncep.noaa.gov/data/teledoc/telecontents. 7 ͗http://www.wrcc.dri.edu/summary/climsmak.html͘ html͘ August 2005 FIRE AND CLIMATE LINKAGE IN ALASKA 1321

opment of troughs is more likely (Bond and Harrison 2000). The cool phase is characterized by cool SSTs in the Gulf of Alaska (GOA) and warm temperatures in the North Paci®c. Alternatively, warm SSTs in the GOA and cool temperatures in the North Paci®c char- acterize the warm phase of the PDO. The strength and location of the AL is also correlated with the phase of the PDO. During cool phases of the PDO, the AL is, on average, located around the western extent of the Aleutian Islands. When the PDO shifts from a cool phase to a warm phase, the AL strengthens and moves to the east (Trenberth 1992, Niebauer 1998). Statistical model and spatiotemporal scaling The allowable number of explanatory variables in the MLR was ®xed at roughly 15% of the total number of data points. Standard subset selection techniques were used to evaluate signi®cance of monthly temper-

FIG. 2. East Paci®c Pattern in January, adapted from ature/precipitation and teleconnection indices within Barnston and Livezy (1987). Values are correlations between the MLR model. By ®xing the number of parameters, 70 kPa (700 millibar) heights at grid points and the rotated we are essentially evaluating a subset of all possible principal-component time series of the East Paci®c telecon- models. In the context of information criteria (e.g., nection. The 70 kpa heights approximate the actual height of a pressure surface above mean sea level. These heights cor- AIC) we have ®xed the penalty function and are max- respond to the middle troposhere at roughly 3000 m. Height imizing the likelihood within this subset of models. anomalies at the 70-kPa level in¯uence the movement of sur- There are numerous methods with differing criteria that face weather systems. can be used to ®t models. The success of any model selection exercise can best be evaluated by crossvali- dation. The results of sequential crossvalidation along (Bonsal et al. 1993, Livezy and Smith 1999). Specif- with other model diagnostics (e.g., residual analyses) ically, there is correlation between SSTs in the North show that the regression model is relatively robust and Paci®c and weather in Alaska (Papineau 2001). The does not violate any of the regression assumptions. Paci®c Decadal Oscillation re¯ects differences in the The natural logarithm transformation is applied to circulation and location of anomalous warm/cool SSTs the data for annual area burned to both minimize het- across the Paci®c Ocean. (An image of typical win- eroscedasticity and decrease the potential for bias in tertime sea surface temperature, sea level pressure, and variable selection based on the relatively small number surface windstress anomaly patterns during both warm of years where the majority of area burned. This trans- and cool phases of the Paci®c Decadal Oscillation has formation results in more reliable tests of parameter been created by Steven Hare, International Paci®c Hal- signi®cance, since the variability of the response is no ibut Commission [available online].)9 To quantify the longer a nonconstant function of the expected value impact of North Paci®c SSTs on area burned in Alaska, estimated by the regression. All references to statistical indices for the Paci®c Decadal Oscillation (PDO; avail- signi®cance are made at the 0.10 Type I error level. able online)10 were examined as explanatory variables The spatial and temporal scales of the MLR need to in the MLR. Multiple temporal scales (i.e., one, two, be considered when interpreting the results. The sta- three, and four-month averages) of the PDO index were tistical model is essentially a point model in that it evaluated as explanatory variables in the MLR model. integrates values of climatic explanatory variables Some degree of statistical similarity exists between the across both space (i.e., Interior Alaska) and time. The signal for the PDO and El NinÄo events. Because of this model integrates across space through the calculation similarity, El NinÄo indices (both pressure-based and of a simple average of monthly climate indices for SST) were evaluated for their potential as explanatory weather stations that in this sense represent Interior variables in the statistical model. The metric used to Alaska (Fig. 1). This method represents a ®rst-order quantify the PDO was more statistically signi®cant. integration of spatial information regarding intrastate Depending on the phase of the PDO, certain atmo- variability into a single point estimate. The model in- spheric circulation patterns in the North Paci®c are tegrates across time through the use of monthly aver- favored. During cool phases of the PDO, ridging is ages for the temporal scale of explanatory variables more frequent, whereas in the warm phase, the devel- and the use of annual values for response variables. 9 ͗http://tao.atmos.washington.edu/pdo/img/pdo warm cool. The use of monthly indices lacks explicit consideration tif͘ of intramonthly variability. Extreme ®re behavior can 10 ͗http://jisao.washington.edu/pdo/PDO.latest͘ and often does occur at a less than monthly temporal 1322 PAUL A. DUFFY ET AL. Ecological Applications Vol. 15, No. 4

TABLE 1. Statistical output for the multiple linear-regression (MLR) model with the natural logarithm of the number of hectares burned annually as the response variable.

Explanatory variable Estimate SE tP Intercept Ϫ41.368 7.352 Ϫ5.626 1.12 ϫ 10Ϫ6 Average June temperature 0.617 0.098 6.302 1.11 ϫ 10Ϫ7 DEP Ϫ4.803 0.914 Ϫ5.254 3.93 ϫ 10Ϫ6 Average April temperature Ϫ0.199 0.041 Ϫ4.855 1.49 ϫ 10Ϫ5 DEP ϫ average April temperature 0.159 0.034 4.712 2.38 ϫ 10Ϫ5 Average May temperature 0.235 0.056 4.201 1.24 ϫ 10Ϫ4 Average PDO of January and February Ϫ0.465 0.151 Ϫ3.071 3.61 ϫ 10Ϫ3 Average July temperature 0.252 0.089 2.836 6.81 ϫ 10Ϫ3 Average June precipitation Ϫ1.195 0.458 Ϫ2.606 1.24 ϫ 10Ϫ2 ϭ ϭ ϭ ϭ ϫ Ϫ13 2 Notes: Residual SE 1.098; df 45. F8, 45 21.14, P 7.112 10 . The R value of 0.79 implies that the deterministic component of the MLR (consisting of the explanatory variables listed) model explains 79% of the total variability in the response variable. See Table 3 for a key to the abbreviations. scale (Flannigan and Harrington 1988, Alvarado et al. logarithm of number of hectares burned annually. The 1998). Nonetheless, important linkages between pres- explanatory variables in order of most to least signif- sure anomalies and ®re behavior on monthly/seasonal icant (Table 1) are Average June Temperature (AJT), timescales have been identi®ed in Canada (Johnson and Delta EP Teleconnection (DEP), Average April Tem- Wowchuk 1993, Skinner et al. 1999). perature, (AAT), the interaction between DEP and AAT, The goal of the statistical model is to identify var- Average May Temperature (AMT), Average PDO index iables that collectively explain the greatest amount of for January and February (PDOWIN), Average July variability in the natural logarithm of the number of Temperature (AJLT), and Average June Precipitation hectares burned for a given year. The structure and (AJP). DEP is de®ned as the January EP index minus spatiotemporal scaling of the model were selected to the April EP index. The low P value for the F test be both pragmatic and parsimonious. Pragmatically, it (Table 1) indicates that the deterministic component of is of interest to quantify the link between weather var- the statistical model explains a signi®cant percentage iables and the area burned for ®re management activ- of the variability in the natural logarithm of the number ities and a greater understanding of ®re±weather in- of hectares burned per year. Comparison between ob- teractions. Another pragmatic aspect is that the ante- served values of the number of hectares burned an- cedent nature of the various climatic variables gives nually (Fig. 3, solid circles) and transformed model the results application in long-range (monthly to sea- estimates (Fig. 3, solid diamonds with connecting lines) sonal) prediction of annual area burned in Alaska. As shows that the transformed estimates from the statis- a starting point, integrating both the response and ex- tical model provide a reasonable estimate of the number planatory variables across space provides a clear def- of hectares burned annually. The estimated values (Fig. inition of the spatial scale of interest and a simple 3, solid diamonds with connecting lines) are an ex- interpretation of the model results. Across the numer- ponentiation of the ®tted values produced by the MLR. ous temporal scales evaluated, the current state of the Since the regression was performed on the natural log- model maximizes the predictive relationship between arithm of the number of hectares burned annually, a ®re and climate. ``back-transformation'' was used to obtain estimates in the original space. There are numerous ways to ``back- RESULTS transform'' data and this one, although the simplest, Seven climatic variables and an interaction term col- does produce a negatively biased estimate under the lectively explain 79% of the variability in the natural assumption of lognormality. Hence, predictions made

FIG. 3. Time series plot (years 1950±2003) showing the number of hectares burned annu- ally (circles) and transformed estimates (dia- monds with connecting lines) based on the pre- dicted values from the multiple linear-regres- sion (MLR) model. The estimated values (dia- monds) in this plot are an exponentiation of the ®tted values produced by the MLR. Since the regression was performed on the natural loga- rithm of the number of hectares burned annu- ally, a ``back-transformation'' was used to ob- tain estimates in the original space. August 2005 FIRE AND CLIMATE LINKAGE IN ALASKA 1323

FIG. 4. Time series of East Paci®c (EP) teleconnection index and estimated partial autocorrelation function (PACF) for the EP index. Dashed lines in the PACF plot represent 95% con®dence intervals for the estimated correlation at a given lag. using this simple back-transformation will produce of the EP index in winter to a positive phase in spring slight underestimates under the assumption that the increases the likelihood of blocking ridges forming in data for area burned are lognormally distributed. the following summer. Hence negative values of DEP indicate a tendency toward drier-than-normal condi- Explanatory variables tions for the months of March through August (Fig. 5). Teleconnection indices.ÐThe time series of the EP Along with drier-than-normal conditions, negative val- index (Fig. 4) and the estimated partial autocorrelation ues of DEP are also associated with increased temper- function shows that the EP teleconnection is an auto- atures for the months of May and June (Fig. 6). The regressive process with maximal temporal correlations cumulative impact of the warmer and drier spring/sum- occurring at one- and three-month lags. DEP is de®ned mer conditions associated with negative values of DEP as the January EP index minus the April EP index, and is a greater area burned (Table 2). The negative value provides a metric of shifts in atmospheric circulation of the parameter estimate for DEP (Table 1) implies from winter to spring. Through its in¯uence on at- that a shift from negative to positive EP values between mospheric circulation, a shift from the negative phase

FIG. 5. Scatter plot of average precipitation for the FIG. 6. Scatter plot of average spring (May and June) months of March through August vs. DEP (January East Pa- temperatures vs. DEP (January East Paci®c teleconnection ci®c teleconnection index minus April East Paci®c telecon- index minus April East Paci®c teleconnection index). Shifts nection index). Shifts from negative values of the EP in Jan- from negative values of the EP in January to positive values uary to positive values of the EP in April (negative DEP) are of the EP in April (negative DEP) are associated with in- associated with decreased precipitation from March through creased temperatures in May and June. The least-squares re- August. The least-squares regression line and P value (cor- gression line and P value (corresponding to the test of slope responding to the test of slope equal to zero) are presented. equal to zero) are presented. 1324 PAUL A. DUFFY ET AL. Ecological Applications Vol. 15, No. 4

TABLE 2. DEP data have been partitioned into ``big'' and ``small'' ®re years with respect to 100 000 ha.

Minimum Median Mean Maximum Small ®re years Ϫ1.50 0.60 0.64 3.20 Big ®re years Ϫ2.60 Ϫ0.65 Ϫ0.60 1.50 Notes: Years with area burned Ͼ100 000 ha are classi®ed as ``big'' ®re years, and years with area burned Ͻ100 000 ha are classi®ed as ``small'' ®re years. Positive values for DEP signify increased zonal ¯ow from spring to summer, whereas negative values of DEP indicate more meridional ¯ow and increased potential for ridges to develop (Figs. 5 and 6). DEP values for big ®re years are signi®cantly different (and less than) DEP values for small ®re years (P Ͻ 0.0002 for Welsh's modi®cation (unequal variances) to a two-sided, two-sample t test).

January and April favors a greater area burned in the FIG. 7. Scatter plot of average precipitation for May and June vs. the average of the PDO index for the months of upcoming summer. January and February. Cool phases of the PDO (negative Monthly PDO indices were evaluated for use as ex- PDOWIN) are associated with decreased precipitation for the planatory variables (Table 3) in the MLR, and winter months of May and June. The least-squares regression line indices displayed the largest signal with respect to ex- and P value (corresponding to the test of slope equal to zero) plaining variability in the natural logarithm of the num- are presented. ber of hectares burned annually in Alaska. The months of January and February were the most signi®cant ex- is signi®cant negative correlation between precipitation planatory variables among the monthly PDO indices, in summer and precipitation in winter, but no corre- and since the estimates were of the same sign, the av- lation exists during the warm phase (Fig. 8). erage of these monthly indices (PDOWIN) was used Temperatures.ÐThe MLR analysis found the aver- as a single explanatory variable in the regression. The age temperatures for the months of April, May, June, negative value for the parameter estimate of PDOWIN and July (AAT, AMT, AJT, and AJLT, respectively) to indicates that cool phases of the PDO are associated be signi®cant explanatory variables for the logarithm with a greater area burned. There is positive correlation of the number of hectares burned (Table 1). Among the between PDOWIN and the average precipitation for parameter estimates for these monthly temperatures May and June (Fig. 7). Hence cool phases of the PDO only AAT has a negative value. AJT was the most are associated with drier summer conditions. Addi- signi®cant explanatory variable identi®ed by the MLR, tionally, the phase of the PDO is also associated with and by itself it explains Ͼ34% of the variability in the a shift in the correlation between winter and summer response. Attempts to combine the average tempera- precipitation. During the cool phase of the PDO, there tures from months (across two-, three-, and four-month

TABLE 3. List of abbreviations.

Abbreviation Variable AAT Average April Temperature AGDC Alaska Geospatial Data Clearinghouse AFS Alaska Fire Service AMT Average May Temperature AJLT Average July Temperature AJT Average June Temperature AJP Average June Precipitation AL Aleutian Low CPC Climate Prediction Center CRU Climate Research Unit DEP Delta EP ϭ January East Paci®c Teleconnection index minus the April East Paci®c Teleconnection index. This index measures changes in atmospheric circulation from winter to spring EP East Paci®c: a teleconnection index used to quantify atmospheric circulation GOA Gulf Of Alaska IPCC Intergovernmental Panel on Climate Change LFDB Large Fire Data Base MLR Multiple Linear Regression NOAA National Oceanic and Atmospheric Administration PDO Paci®c Decadal Oscillation PDOWIN Average PDO value for the months of January and February WRCC Western Region Climate Center August 2005 FIRE AND CLIMATE LINKAGE IN ALASKA 1325

FIG. 8. Scatter plots of July±August precip- itation vs. previous September±January precip- itation. Plots are partitioned based on the sign of the average PDO values for the months of January and February (PDOWIN). The negative correlation between winter and summer precip- itation exists only during cool phases of the PDO. Hence during cool phases of the PDO, it is less likely that a dry winter will be followed by a dry summer. The least-squares regression line and P value (corresponding to the test of slope equal to zero) are presented.

periods) yielded explanatory variables with compara- Precipitation.ÐAs the sole precipitation variable in tively lower explanatory capability. Hence, monthly the regression model, average June precipitation (AJP) temperatures exert differential impacts on the annual was the least signi®cant among the other variables (Ta- area burned. ble 1). Depending on the thaw depth of the organic The parameter estimate for AAT is negative, indi- layer, June precipitation data can yield information re- cating that lower April temperatures are associated with garding moisture status of organic layers. Temperature greater area burned (Table 1), but the interpretation is is also an integral component of moisture status, and made dif®cult by nonlinear interactions with DEP. in general, area-averaged anomalies of monthly tem- When zonal ¯ow is dominant in the spring (DEP Ͼ 0), perature and precipitation are negatively correlated in April temperatures are positively correlated with area high latitudes during summer. A common factor in burned, whereas when meridional ¯ow is building, anomalies of both variables is cloudiness, which is as- April temperatures are negatively correlated with area sociated with lower temperatures and greater precipi- burned. The reason for this shift is not fully understood, tation during summer. This correlation is primarily but is likely due to interactions with the precipitation driven by the prevalence of frontal low-pressure sys- signal associated with the EP. When zonal ¯ow is dom- tems that typically have large areas of cloud cover and inant, precipitation is more likely, and higher April relatively spatially homogenous precipitation. Alter- temperatures can signal the beginning of break-up and natively, convective storms associated with surface result in an earlier ®re season. When meridional ¯ow heating can result in highly localized cloud cover and is dominant, precipitation is less likely and elevated precipitation, which reduces the effectiveness of April temperatures may result in snowmelt before the monthly precipitation as an explanatory variable. organic horizon is thawed enough to accept the melt- Intramonthly variability of precipitation is also an water as recharge. The impact of spring snowmelt on important factor with respect to moisture status of the ®re dynamics is a subject that requires further study. organic layer (Flannigan and Harrington 1988). A Temperatures for the months of May, June, and July month receiving average precipitation that occurs on are positively correlated with area burned. May is the only several days can have a greater potential for ®re month when deciduous trees typically break dormancy than a month in which the same total rainfall is dis- in the Interior. In the period between snowmelt and tributed evenly throughout the month. In general, leafout, deciduous stands can have greater ¯ammability months with more precipitation will also have greater than conifers. This is partly because deciduous stands intramonthly variability. This hinders attempts to de- termine the relative importance of total monthly pre- are more prevalent on south-facing slopes, and respond cipitation vs. timing. The lack of information regarding rapidly when warm, dry spring conditions exist. Ele- intramonthly variability in precipitation is partly ob- vated temperatures in May decrease available moisture viated by the use of the EP monthly indices, which for ground cover and increase progressive drying of implicitly contain information about circulation and the deeper organic layers. Elevated May temperatures can potential for blocking highs associated with reduced be suf®cient to develop convective thunderstorm ac- precipitation (Johnson and Wowchuk 1993, Skinner et tivity, although the majority of lightning strikes in the al. 1999). Interior typically occur during June and July (Reap 1991). Average June and July temperatures probably DISCUSSION exert their in¯uence through drying of the organic layer This analysis provides a framework for understand- and development of convective activity. ing the in¯uence of ocean±atmosphere interaction on 1326 PAUL A. DUFFY ET AL. Ecological Applications Vol. 15, No. 4 the ®re regime of Interior Alaska. Our study is the ®rst mer. In general, warmer and dryer early summer con- to build a statistical model quantifying the link between ditions favor a progressive drying of deeper organic teleconnections, weather, and area burned in Interior layers. As the length of a dry spell increases, there is Alaska. Intuitively one assumes that, in the short term, greater potential for widespread, intense combustion less precipitation and higher temperatures increase ®re due to increased homogeneity of favorable fuel con- danger. This work extends such reasoning to the an- ditions. In the boreal forest of Alaska, the majority of tecedent ocean±atmosphere interactions that in¯uence ®res are dependent on combustion of the organic layer; short-term weather, and it quanti®es the in¯uences of however, during extreme ®re events ®res can burn in these factors on the annual area burned. Teleconnec- the canopy without consuming much soil organic mat- tions modify atmospheric circulation, and the statisti- ter. Hence DEP likely in¯uences the number of hectares cally signi®cant relationships between teleconnection burned in Interior Alaska through its effects on mois- indices and monthly weather provide plausible physical ture conditions in the organic layer. The cumulative mechanisms for the teleconnections to in¯uence area impact of the warmer and drier summer conditions as- burned. The relationships between teleconnections, sociated with negative values of DEP is a greater area weather, and conditions that are favorable for large burned (Table 2). areas burned are considered below. Like the EP teleconnection, the PDO also in¯uences the number of hectares burned in interior Alaska Teleconnection in¯uences on weather through modi®cation of weather. Of the PDO indices Positive phases of the EP teleconnection correspond evaluated, the index comprising of the average for the with upper air¯ow that is meridional, whereas negative winter months of January and February (PDOWIN) phases are associated with zonal circulation and explains the greatest amount of variability in the MLR. strengthened westerlies in the eastern North Paci®c. Like the EP teleconnection, the PDO dictates certain The meridional ¯ow is conducive to the formation of aspects of atmospheric circulation. Speci®cally, ridging blocking highs that impact short-term weather and ®re in the North Paci®c is more frequent during cool phases behavior in Interior Alaska. Blocking highs typically of the PDO, whereas in the warm phase the develop- persist for several days to several weeks, and in¯uence ment of troughs in this region is more likely (Bond and the ®re potential across regions of 100±1000 km (John- Harrison 2000). The PDOWIN index is positively cor- son and Wowchuk 1993). In summer, surface-blocking related to late spring precipitation (Fig. 7) in Alaska. high-pressure systems bring warm temperatures and The phase of the PDO also in¯uences correlation low precipitation that can cause deep drying of the between winter and summer precipitation amounts for organic layer. Hence, midtropospheric anomalies and Interior Alaska (Fig. 8). In the winter, there is a mono- the surface high-pressure systems that accompany them tonic decrease in the average monthly precipitation increase ®re danger at a landscape scale. Shifts from from October to May, and the majority of winter pre- positive to negative EP indices in the spring (DEP Ͼ cipitation comes in the months of October and Novem- 0) are associated with greater precipitation and cooler ber. Although the average temperature for October is temperatures across Interior Alaska. below freezing, precipitation events during this month With respect to the EP teleconnection, the antecedent can contribute to moisture content at the soil surface. atmospheric circulation pattern most strongly corre- Impacts of the interaction between winter and summer lated with a large area burned is a negative EP index precipitation on the area burned require further eval- in the winter that shifts to a positive EP index in the uation, but are possibly driven by soil moisture dy- spring. To quantify this, DEP is de®ned as the January namics in the organic horizons. Speci®cally, differ- EP index minus the April EP index. DEP measures the ences in snowpack can potentially in¯uence soil mois- change in atmospheric circulation from winter to ture in the following spring. For soils with organic spring. The negative value of the parameter estimate horizons, the ability of snowmelt water to percolate for DEP (Table 1) shows that a shift from a negative into the active layer depends on the quantity of ice in EP index in January to a positive EP index in April is soil pores (Kane 1980). Moisture content at the soil correlated with greater area burned in the following surface has been observed to increase throughout the summer. This shift represents a departure from the pos- winter, with this effect being most pronounced for wet itive seasonal correlation that is characteristic of the soils. Hence, higher soil moisture contents correspond autoregressive EP teleconnection (Fig. 4). As a con- to greater amounts of ice present in the frozen soil, sequence of the impact on atmospheric circulation, neg- which in turn reduces the in®ltration rate and the sat- ative values of DEP are associated with drier than nor- urated hydraulic conductivity (Kane 1980). Similar re- mal conditions for the months of March through August sults have been found for mineral soils in boreal forests (Fig. 5) and increased temperatures for the months of (Zhao and Gray 1999). Additionally, after winters with May and June (Fig. 6). Shifts in the EP index from little snowfall, it is possible for organic layers at high negative phases in winter to positive phases in spring latitudes to remain frozen into July (Nyberg et al. modify atmospheric circulation and increase the like- 2001). Hence the importance of summer precipitation lihood of blocking ridges forming in the following sum- can depend on the development and persistence of win- August 2005 FIRE AND CLIMATE LINKAGE IN ALASKA 1327 ter snowpack. Wetter soils during freeze-up in the fall from cool to warm phase of the PDO occurred. One of coupled with above-average precipitation during the the consequences of this type of shift was that the Aleu- winter can potentially result in drier soil organic layers tian Low intensi®ed and moved southeast of its former in the following spring. Soil moisture of organic layers position. This shift causes a more easterly (less south- plays a critical role in the smoldering combustion of erly) ¯ow component across Interior Alaska and is as- deeper organic horizons (Miyanishi and Johnson 2002). sociated with regional summer droughts. The years of Fires that smolder in deeper organic horizons are more 1957, 1969, 1977, 1990, and 1999 had the ®rst, fourth, likely to persist through rainfall events that may occur. ®fth, second, and 12th largest ®re years (respectively) Although the correlation between winter and summer of the period from 1950 to 2003. One of the most precipitation is signi®cantly different between warm extreme cases of this was in the summer of 1977, when and cool phases of the PDO, none of the fall or winter the Kotzebue weather station recorded 0.03 inches of monthly precipitation variables were signi®cant in the total precipitation for the months of June and July; regression. Hence if the relationship between the PDO (available online).11 Not surprisingly, many ®res that and the correlation between winter and summer pre- burned that summer were located in the northwest part cipitation in¯uences area burned in the subsequent year, of the state near Kotzebue. the mechanism is likely complex and has yet to be de®nitively characterized. Model development The majority of summer precipitation in Alaska A key component of any model development lies in comes in the months of July and August, and during testing of the ®nal model. This is especially important this time area-averaged anomalies of monthly temper- when developing a statistical model through the se- ature and precipitation are negatively correlated. As a quential elimination of explanatory variables based on consequence, elevated temperatures typically accom- signi®cance tests. The potential always exists for the pany low precipitation and both are conducive to in- chance selection of statistically signi®cant variables creased area burned. During cool phases of the PDO, that are of little practical importance. By the de®nition there is a strong negative correlation between cumu- of statistical signi®cance, the probability of this hap- lative precipitation amounts for October±November pening increases as more potential explanatory vari- and July±August (Fig. 8). Hence in cool phases, dry ables are evaluated. A ®rst step in exploring the po- winters are likely to be followed by wet summers and tential for the selection of explanatory variables that dry summers are likely preceded by relatively wet win- are not practically signi®cant is to consider plausible ters. Simply put, it is less likely to have both a wet mechanisms for the selected variables to in¯uence the winter and a wet summer during cool phases of the response. All of the variables included in the regression PDO than it is during warm phases of the PDO. Fol- model (Table 1) have plausible physical mechanisms lowing the reasoning in the previous paragraph, the for in¯uencing annual area burned (Figs. 5±8, Table relationship between winter and summer precipitation 2), and the identi®cation of these variables will help during the cool phase of the PDO can potentially pro- guide future exploration of the link between climate duce more favorable conditions for greater area burned, and ®re in different boreal systems. since wet winters are likely to be followed by dry sum- Another potential issue is that of correlation between mers (Fig. 8). Alternatively, during the warm phase of explanatory variables, or collinearity. When dealing the PDO, there is negligible correlation between winter with monthly temperature data, correlation between and summer precipitation (Fig. 8), and it is less likely months should be expected, since months are somewhat that a warm summer will follow a wet winter. Shifts arbitrary designations for discretizing the year. For ex- in sea-ice dynamics off the west coast of Alaska have ample, a high-pressure ridge bringing warm and dry been linked to changes in the phase of the PDO (Nie- weather at the end of May will likely have a similar bauer 1998) and it is possible that the shift in intra- in¯uence on the weather in the beginning of June. The annual correlation between precipitation signals is re- collinearity that exists between explanatory variables lated to the impact of the PDO on sea-ice extent. can in¯uence model output in several ways, including The Paci®c Decadal Oscillation the selection of variables with little practical signi®- and the Aleutian Low cance. A quantitative way to assess the impact of both collinearity and the potential for selection of variables The phase of the PDO is positively correlated with with little practical signi®cance is through cross-vali- spring precipitation (Fig. 7) in Alaska. One mechanism dation. If explanatory variables of little practical sig- that can explain these short-term weather anomalies is ni®cance have been selected due to collinearity or impacts of the PDO on the location and intensity of chance, cross-validation routines often reveal this. the Aleutian Low (Niebauer 1998, Papineau 2001). In Based on our cross-validation results, the impact of the period used for this study, there were only six times collinearity on model predictions seems negligible, and (1957±1958, 1969±1970, 1976±1977, 1990±1991, 1999±2000) where substantial (greater than a 1.6 unit shift in the PDOWIN metric for sequential years) shifts 11 ͗http://www.wrcc.dri.edu/cgi-bin/cliMAIN.pl?akkotz͘ 1328 PAUL A. DUFFY ET AL. Ecological Applications Vol. 15, No. 4 there do not appear to be any variables selected that pattern are associated with strengthened westerlies in are not signi®cant both practically and statistically. the Eastern North Paci®c as a consequence of a more zonal upper air¯ow over the region south of Alaska. Forecasting with the model The shift in sign of the EP teleconnection over a period The results of this paper can be used for long-range of several months in winter exerts a signi®cant signal (monthly-to-seasonal) forecasts of the magnitude (area on both temperature and precipitation during the spring burned) of the upcoming ®re season. As an example, and summer in Interior Alaska. consider two forecasts for the upcoming 2004 ®re sea- The Paci®c Decadal Oscillation exerts a direct in- son, one at the end of April and one at the end of May. ¯uence on winter temperatures and summer precipi- Estimates of the other necessary monthly variables are tation and also modi®es the correlation between winter obtained by interpreting the Climate Prediction Center and summer precipitation. During cool phases of the prognostic discussion for long-lead seasonal outlooks PDO, there is strong negative correlation between win- (http://www.cpc.ncep.noaa.gov/products/predictions/ ter and summer precipitation. One possible explanation 90day/fxus05.html). Based on the prognostic discus- for the importance of above-average winter precipita- sions, the 60th percentile of the May, June, and July tion is that decreased percolation due to ice in the soil temperatures will be used for the ®rst forecast (made pores can potentially leave the organic layers more sus- at the end of April) and the 70th percentile will be used ceptible to warm, dry temperatures in May and June. for the June and July temperature for the second fore- This scenario is more likely during cool phases of the cast (made at the end of May). The median June pre- PDO, and consequently 69% of the area burned for the cipitation will be used for ®rst forecast and the 40th period of this study occurred when the PDOWIN metric percentile will be used for the second. Based on this was Ͻ0 (i.e., cool phases of the PDO). The interaction information, the ®rst forecast (end of April) is for between moisture content of the soil organic layers, 41 000 ha and the second forecast (end of May) is for winter precipitation, and ®re extent in the following 135 000 ha. Both of these are well below average year remains unclear and merits further study. (270 000 ha). In fact, as of the end of July 2004, the This work represents a ®rst step in quantifying the total area burned had already reached roughly link between weather and ®re in Alaska. The multiple 1 780 000 ha (second overall since 1950). In Fairbanks, linear regression model used to characterize this link June 2004 was the second hottest in the past 100 years. is a simple tool for taking this ®rst step. The model In addition to the hot weather, lightning activity was lacks a spatially explicit component, as monthly tem- record-breaking as well. On the 14th of June, 8589 peratures and precipitation represent averages from a lightning strikes were recorded throughout the state, relatively small number of sites located throughout the and one month later, on July 15th, Ͼ9000 strikes were Alaskan Interior. The use of broader circulation indices recorded (courtesy of the National Weather Service). reduces the reliance on, and possible concerns about, These represent two of the largest single-day outbreaks the representativeness of local (single-point) surface- on record for the state. With the values for June inserted based measurements made in a heterogeneous land- in the statistical model (and using the 70th percentile scape. The interactions between weather and ®re char- for July temperature), the estimate is 2 100 000 ha acterized in this work were developed conditional on burned. This exercise demonstrates the importance of the current spatial distribution of vegetation across In- June temperature in the statistical model. Attempts to terior Alaska. Debate exists as to the importance of use this model for planning should incorporate multiple age-dependent ¯ammability in the boreal forest (John- plausible scenarios for June temperatures in order to son 1992, Johnson et al. 1998, 2001, Hely et al. 2000a, appropriately plan for the upcoming season. b, Ward et al. 2001). If stand ¯ammability does indeed change as a function of age, the spatial distribution of CONCLUSION such vegetation would mediate the relationship be- Climate, ®re, and vegetation in the boreal forest of tween climate and ®re. Cross-validation of the model Interior Alaska interact on multiple spatial and tem- suggests that collinearity and spurious variable selec- poral scales. It is clear from this work that telecon- tion are not serious issues and model predictions can nections operating on multiple timescales in¯uence the provide input to ®re management of®cials charged with annual area burned across Interior Alaska. The results developing resource allocation plans for upcoming ®re presented in this paper provide evidence linking the seasons. Due to the strong dependence of area burned East Paci®c teleconnection and the Paci®c Decadal Os- on weather, forecasts of area burned produced by the cillation to several weather variables that are directly model are only as reliable as the forecasts for temper- related to the annual area burned in Interior Alaska. ature and precipitation used in the model. Speci®cally, The most likely, ultimate mechanisms for these link- June temperature plays a critical role in the magnitude ages are shifts in atmospheric circulation. Strong pos- of area burned. Future attempts to forecast area burned itive phases of the EP pattern are conducive to the in Alaska should focus on identifying those atmospher- development of blocking highs that impact short-term ic mechanisms that most strongly in¯uence June tem- weather and ®re behavior. Negative phases of the EP perature. A next step is to apply similar MLR proce- August 2005 FIRE AND CLIMATE LINKAGE IN ALASKA 1329 dures for tundra vs. forest vegetation to identify the Flannigan, M. D., B. J. Stocks, and B. M. Wotton. 2000. climatological characteristics that drive the burning of Climate change and forest ®res. The Science of the Total Environment 262:221±229. different vegetation types. Girardin, M.-P., J. Tardif, M. D. Flannigan, B. M. Wotton, and Y. Bergeron. 2004. Trends and periodicities in the ACKNOWLEDGMENTS Canadian Drought Code and their relationships with at- Funding for this research was provided by the USDA/USDI mospheric circulation for the southern Canadian boreal for- Joint Fire Science Program and the Bonanza Creek Long-Term est. Canadian Journal of Forest Research 34:103±119. Ecological Research (LTER) program (funded jointly by NSF Gower, S. T., O. Krankina, R. J. Olson, M. Apps, S. Linder, grant DEB-0423442 and USDA Forest Service, Paci®c North- and C. Wang. 2001. Net primary production and carbon west Research Station grant PNW01-JV111261952-231). The allocation patterns of boreal forest ecosystems. Ecological University of Alaska, School of Natural Resources and Agri- Applications 11:1395±1411. cultural Sciences Doctoral Fellowship supported P. A. Duffy Hartmann, B., and G. Wendler. 2003. Manifestation of the during this work. A portion of this work was also supported Paci®c Decadal Oscillation shift of 1976 in Alaskan cli- by the Center for Global Change and Arctic System Research. matology. Seventh AMS Conference on Polar Meteorology Thanks to David McGuire and Eric Kasischke for thoughtful and Climatology, Hyannis, Massachusetts. American Me- reviews. Thanks also to Terry Chapin for facilitation of this teorological Society, Boston, Massachusetts, USA. group of scientists. Randi Jandt and Mary Lynch of the Alaska Hely, C., Y. Bergeron, and M. D. Flannigan. 2000a. Coarse Fire Service also provided valuable assistance. All statistical woody debris in the southeastern Canadian boreal forest: analyses were performed using the R language and environ- composition and load variations in relation to stand re- ment for statistical computing and graphics. For more infor- placement. Canadian Journal of Forest Research 30:674± mation see ͗http://www.r-project.org/͘. 687. Hely, C., Y. Bergeron, and M. D. Flannigan. 2000b. Effects LITERATURE CITED of stand composition on ®re hazard in mixed-wood Ca- Alvarado, E., D. V. Sandberg, and S. G. Pickford. 1998. Mod- nadian boreal forest. Journal of Vegetation Science 11:813± eling large forest ®res as extreme events. Northwest Sci- 824. ence 72(Special Issue):66±74. Hely, C., M. Flannigan, Y. Bergeron, and D. McRae. 2001. Barnston, A. G., and R. E. Livezey. 1987. Classi®cation, Role of vegetation and weather on ®re behavior in the seasonality and persistence of low-frequency atmospheric Canadian mixedwood boreal forest using two ®re behavior circulation patterns. Monthly Weather Review 115:1083± prediction systems. Canadian Journal of Forest Research 1126. 31:430±441. Bond, N. A., and D. E. Harrison. 2000. The Paci®c decadal Hess, J. C., C. A. Scott, G. L. Hufford, and M. D. Fleming. oscillation, air±sea interaction and central North Paci®c 2001. El Nino and its impact on ®re weather conditions in winter atmospheric regimes. Geophysical Research Letters Alaska. International Journal of Wildland Fire 10:1±13. 27:731±734. Hessl, A. E., D. McKenzie, and R. Schellhaas. 2004. Drought Bonsal, B. R., A. K. Chakravarti, and R. G. Lawford. 1993. and Paci®c Decadal Oscillation linked to ®re occurrence Teleconnections between North Paci®c SST anomalies and in the inland Paci®c Northwest. Ecological Applications growing season extended period dry spells on the Canadian 14:425±442. prairies. International Journal of Climatology 13:865±878. Hurrell, J. W., Y. Kushnir, G. Ottersen, and M. Visbeck. 2003. Bonsal, B. R., and R. G. Lawford. 1999. Teleconnections The North Atlantic Oscillation: climatic signi®cance and between El NinÄo and La NinÄa events and summer extended environmental impact. Geophysical Monograph Series. dry spells on the Canadian prairies. International Journal American Geophysical Union, Washington, D.C., USA. of Climatology 19(13):1445±1458. IPCC (Intergovernmental Panel on Climate Change). 2001. Carcaillet, C., Y. Bergeron, P. J. H. Richard, B. Frechette, S. Climate Change 2001: Technical Summary of the Working Gauthier, and Y. T. Prairie. 2001. Changes of ®re frequency Group I Report. WMO/UNEP,Cambridge University Press, in the Eastern Canadian boreal forest during the Holocene: Cambridge, UK. does vegetation composition or climate trigger the ®re re- Johnson, E. A. 1992. Fire and vegetation dynamics: studies gime? Journal of Ecology 89:930±946. from the North American boreal forest. Cambridge Uni- Carcaillet, C., and P. J. H. Richard. 2000. Holocene changes versity Press, Cambridge, UK. in seasonal precipitation highlighted by ®re incidence in Johnson, E. A., K. Miyanishi, and S. R. J. Bridge. 2001. eastern Canada. Climate Dynamics 16:549±559. Wild®re regime in the boreal forest and the idea of sup- Chapin, F. S.,III, T. S. Rupp, A. M. Star®eld, L. DeWilde, E. pression and fuel buildup. Conservation Biology 156: S. Zavaleta, N. Fresco, J. Henkelman, and A. D. McGuire. 1554±1557. 2003. Planning for resilience: modeling change in human± Johnson, E. A., K. Miyanishi, and J. M. H. Weir. 1998. Wild- ®re interactions in the Alaskan boreal forest. Frontiers in ®res in the western Canadian boreal forest: landscape pat- Ecology and the Environment 1(5):255±261. terns and ecosystem management. Journal of Vegetation DeVolder, A. (1999). Fire and climate history of lowland Science 9:603±610. black spruce forests, Kenai National Wildlife Refuge, Alas- Johnson, E. A., and D. R. Wowchuk. 1993. Wild®res in the ka. Dissertation. Northern Arizona University, Flagstaff, southern Canadian Rocky Mountains and their relationship Arizona, USA. to mid-tropospheric anomalies. Canadian Journal of Forest Flannigan, M. D., I. Campbell, B. M. Wotton, C. Carcaillet, Research 23:1213±1222. P. Richard, and Y. Bergeron. 2001. Future ®re in Canada's Kane, D. L. 1980. Snowmelt in®ltration into seasonally fro- boreal forest: paleoecology results and general circulation zen soils. Cold Regions Science and Technology 3(2±3): modelÐregional climate model simulations. Canadian 153±161. Journal of Forest Research 31:854±864. Kasischke, E. S., D. Williams, and D. Barry. 2002. Analysis Flannigan, M. D., and J. B. Harrington. 1988. A study of the of the patterns of large ®res in the boreal forest region of relation of meteorological variables to monthly provincial Alaska. International Journal of Wildland Fire 11:131±144. area burned by wild®re in Canada (1953±80). Journal of Livezey, R. E., and T. M. Smith. 1999. Covariability of as- Applied Meteorology 441±452. pects of North American climate with global sea surface 1330 PAUL A. DUFFY ET AL. Ecological Applications Vol. 15, No. 4

temperatures on interannual to interdecadal timescales. Schimel, D. S., B. H. Braswell, andVEMAP participants. Journal of Climate 12:289±302. 1997. Continental scale variability in ecosystem processes: Lynch, J. A., J. S. Clark, N. H. Bigelow, M. E. Edwards, and models, data, and the role of disturbance. Ecological Mono- B. P. Finney. 2003. Geographic and temporal variations in graphs 67:251±271. ®re history in boreal ecosystems of Alaska. Journal of Geo- Skinner, W. R., M. D. Flannigan, B. J. Stocks, D. L. Martell, physical Research 108(d1±1851):8-1±8-17. B. M. Wotton, J. B. Todd, J. A. Mason, K. A. Logan, and Lynch, J. A., J. L. Hollis, and F. S. Hu. 2004. Climatic and E. M. Bosch. 2002. A 500 hPa synoptic wildland ®re cli- landscape controls of the boreal forest ®re regime: holocene matology for large Canadian forest ®res, 1959±1996. The- records from Alaska. Journal of Ecology 92:477±489. oretical and Applied Climatology 71:157±169. Mann, D. H., C. L. Fastie, E. L. Roweland, and N. H. Bigelow. Skinner, W. R., B. J. Stocks, D. L. Martell, B. Bonsal, and 1995. Spruce succession, disturbance and geomorphology A. Shabbar. 1999. The association between circulation on the Tanana River ¯oodplain, Alaska. Ecoscience 22: anomalies in the mid-troposphere and the area burned by 184±199. wild®re in Canada. Theoretical and Applied Climatology Mann, D. H., and L. J. Plug. 1999. Vegetation and soil de- 63:89±105. velopment at an upland taiga site, Alaska. BioScience 6(2): Star®eld, A. M., and F. S. Chapin,III. 1996. Model of tran- 272±285. sient changes in arctic and boreal vegetation in response McGuire, A. D., J. M. Melillo, and L. A. Joyce. 1995. The to climate and land use change. Ecological Applications 6: role of nitrogen in the response of forest net primary pro- 842±864. duction to elevated atmospheric carbon dioxide. Annual Swetnam, T. W., and J. L. Betancourt. 1990. Fire-Southern Review of Ecology and systematics 26:472±503. Oscillation relations in the Southwestern United States. Melillo, J. M., A. D. McGuire, D. W. Kicklighter, B. MooreIII, Science 249(4972):1017±1020. C. J. Vorosmarty, and A. L. Schloss. 1993. Global climate Trenberth, K. E. 1992. Climate system modeling. Cambridge change and terrestrial net primary production. Nature 63: University Press, Cambridge, UK. 234±240. Van Cleve, K., F. S. ChapinIII, C. T. Dyrness, and L. A. Miyanishi, K., and E. A. Johnson. 2002. Process and patterns Viereck. 1991. Element cycling in Taiga forests: state-fac- of duff consumption in the mixedwood boreal forest. Ca- tor control. BioScience 41(2):78±88. nadian Journal of Forest Research 32:1285±1295. Van Cleve, K., and L. A. Viereck. 1983. A comparison of Nash, C. H., and E. A. Johnson. 1996. Synoptic climatology successional sequences following ®re on permafrost-dom- inated and permafrost-free sites in interior Alaska. Pages of lightning-caused forest ®res in subalpine and boreal for- 1286±1291 in Permafrost: Proceedings of the Fourth In- ests. Canadian Journal of Forest Research 26:1859±1874. ternational Conference. National Academy Press, Wash- Niebauer, H. J. 1998. Variability in Bering Sea ice cover as ington, D.C., USA. affected by a regime shift in the North Paci®c in the period Wallace, J. M., and D. S. Gutzler. 1981. Teleconnections in 1947±1996. Journal of Geophysical Research 103(c12): the geopotential height ®eld during the Northern Hemi- 27171±27737. sphere winter. Monthly Weather Review 109:784±812. Nyberg, L., M. Stahli, P.-E. Mellander, and K. H. Bishop. Ward, P.C., A. G. Tithecott, and B. M. Wotton. 2001. ReplyÐ 2001. Soil frost effects on soil water and runoff dynamics a re-examination of the effects of ®re suppression in the along a boreal forest transect: 1. Field investigations. Hy- boreal forest. Canadian Journal of Forest Research 31: drological Processes 15:909±926. 1467±1480. Papineau, J. M. 2001. Wintertime temperature anomalies in Westerling, A. L., A. Gershunov, D. R. Cayan, and T. P. Alaska correlated with ENSO and PDO. International Jour- Barnett. 2002. Long lead statistical forecasts of area burned nal of Climatology 21:1577±1592. in western US wild®res by ecosystem province. Interna- Payette, S. 1992. Fire as a controlling process in the North tional Journal of Wildland Fire 11:257±266. American boreal forest. Pages 144±168 in H. H. Shugart, Wierzchowski, J., M. Heathcott, and M. D. Flannigan. 2002. R. Leemans, and G. B. Bonan, editors. A systems analysis Lightning and lightning ®re, central cordillera, Canada. In- of the global boreal forest. Cambridge University Press, ternational Journal of Wildland Fire 11:41±51. Cambridge, UK. Yarie, J. 1981. Forest ®re cycles and life tables: a case study Reap, R. M. 1991. Climatological characteristics and objec- from interior Alaska. Canadian Journal of Forest Research tive prediction of thunderstorms over Alaska. Weather and 11(3):554±562. Forecasting 6(3):309±319. Zackrisson, O. 1977. In¯uence of forest ®res on the North Rupp, T. S., A. M. Star®eld, and F. S. ChapinIII. 2000. A Swedish boreal forest. Oikos 29:22±32. frame-based spatially explicit model of subarctic vegetation Zhao, L., and D. M. Gray. 1999. Estimating snowmelt in®l- response to climatic change: comparison with a point mod- tration into frozen soils. Hydrological processes 13:1827± el. Landscape Ecology 15:283±400. 1842.