The impacts and implications of an intensifying fire regime on Alaskan boreal forest composition and albedo

Climate warming and drying are modifying the fire dynamics of many boreal forests, moving them towards a regime with a higher frequency of extreme fire years characterized by large burns of high severity. Plot‐scale studies indicate that increased burn severity favors the recruitment of deciduous trees in the initial years following fire. Consequently, a set of biophysical effects of burn severity on postfire boreal successional trajectories at decadal timescales have been hypothesized. Prominent among these are a greater cover of deciduous tree species in intermediately aged stands after more severe burning, with associated implications for carbon and energy balances. Here we investigate whether the current vegetation composition of interior Alaska supports this hypothesis. A chronosequence of six decades of vegetation regrowth following fire was created using a database of burn scars, an existing forest biomass map, and maps of albedo and the deciduous fraction of vegetation that we derived from Moderate Resolution Imaging Spectroradiometer (MODIS) satellite imagery. The deciduous fraction map depicted the proportion of aboveground biomass in deciduous vegetation, derived using a RandomForest algorithm trained with field data sets (n=69, 71% variance explained). Analysis of the difference Normalized Burn Ratio, a remotely sensed index commonly used as an indicator of burn severity, indicated that burn size and ignition date can provide a proxy of burn severity for historical fires. LIDAR remote sensing and a bioclimatic model of evergreen forest distribution were used to further refine the stratification of the current landscape by burn severity. Our results show that since the 1950s, more severely burned areas in interior Alaska have produced a vegetation cohort that is characterized by greater deciduous biomass. We discuss the importance of this shift in vegetation composition due to climate‐induced changes in fire severity for carbon sequestration in forest biomass and surface reflectance (albedo), among other feedbacks to climate.


Introduction
Fire is a primary driver of long-term boreal forest dynamics including successional changes in vegetation composition (Viereck, 1973;Foote, 1983;Wurtz et al., 2006) and carbon cycling (Bond-Lamberty et al., 2007). In addition, fire generates an array of feedbacks between vegetation and climate, through combustion emissions during the fire event and through changes in carbon and energy exchange for decades afterwards due to fire impacts on soil moisture, temperature, drainage, decomposition, and vegetation composition (Chambers et al., 2005;Liu et al., 2005;Randerson et al., 2006). These feedbacks have the potential to modify northern hemisphere climate because the boreal biome is both expansive and dense in organic carbon, and because fire has substantial impacts on multiple forcing agents, including greenhouse gas fluxes, aerosols, black carbon deposition on snow and sea-ice, and surface albedo within burn perimeters (Bonan et al., 1995;Randerson et al., 2006).
The area of forest burned in Alaska has increased over the last four decades (Kasischke & Turetsky, 2006), and is projected to increase in the 21st century as a consequence of warming and drying during the fire season (Duffy et al., 2005;Balshi et al., 2009). The first, third, and seventh largest fire years since records began in the 1940s, were 2004, 2005, and 2009, respectively (Chapin et al., 2008;AICC, 2010). The annual area burned in Canada also shows an increasing trend during the second half of the 20th century (Gillett et al., 2004). Historical estimates from Russia are sparse (Conard et al., 2002) but recent fires have been anomalously large (Kasischke et al., 2000;UNECE, FAO, GFMC, 2005;van der Werf et al., 2010), as have tundra fires recorded in Alaska (Jones et al., 2009). This intensification of the fire regime, with extreme fire years characterized by large burns becoming more frequent, has been attributed to climate warming (Gillett et al., 2004) and is expected to accelerate over the course of the 21st century. For example, combining historical relationships between weather-estimated fire danger and the annual area burned with projections from two global circulation models, Flannigan et al. (2005) estimated a 74-118% increase in area burned in Canada by the end of this century. More recently, estimates based on an empirical fire model for Alaska and Western Canada showed the average area burned per decade doubling by 2041-2050, relative to 1991-2000, and increasing on the order of 3.5-5.5 times by the end of the 21st century (Balshi et al., 2009).
The same drivers that cause greater areas to burn annually are also expected to increase burn severity (Stocks et al., 2000;Conard et al., 2002;Amiro et al., 2006), which is commonly defined as the proportion of organic matter consumed by fire (Keeley, 2009). In boreal forests, soils store most of the organic carbon; thus, burn severity is largely a function of the degree to which a fire consumes soil organic matter (Boby et al., 2010). Moreover, because in boreal North America most fires result in close to 100% tree mortality , burn severity and its effects on postfire soil conditions, including its physical properties and suitability for seed survival, initially regulate postfire vegetation communities (Johnstone & Chapin, 2006). The shallowness or absence of a porous, organic layer after severe burning reduces the likelihood of moisture stress affecting seedlings. In Alaska deciduous tree species benefit from a shallow postfire organic layer because they, in contrast to spruce trees, lack large reserves stored in seeds and a taproot to buffer against moisture stress (Schopmeyer, 1974;Kay, 1993). As a result, increased deciduous seedling recruitment and an ensuing dominance of tall shrubs and deciduous trees, particularly willow (Salix spp) and trembling aspen (Populus tremuloides) (hereafter simply aspen), tend to characterize severely burned landscapes in the first decade after fire Johnstone et al., 2010b). In less severe burns, however, aspen are largely restricted to asexual resprouting (Viereck, 1973;Greene & Johnson, 1999) and evergreen seedlings have a higher recruitment success, relative abundance and persistence (Johnstone, 2005).
A similar pattern of high aspen recruitment in severely burned sites has been observed in other forested ecosystems of northwestern America such as Grand Teton and Yellowstone National Parks (Kay, 1993;Turner et al., 2003;Romme et al., 2005). In Yellowstone National Park, the intense ungulate browsing suppresses the growth of aspen seedlings in favor of coniferous trees such as lodgepole pine, preventing aspen trees from ever dominating the canopy except in areas where ungulate browsing is restricted (Turner et al., 2003). In interior Alaska, however, ungulates are unlikely to play a significant role in postfire succession because the region has a much lower density of ungulates: before wolf introduction in 1995, northern Yellowstone had 1.5-9 elk km À2 (White & Garrott, 2005), vs. 1-1.3 moose km À2 in a state-designated game management unit in interior Alaska (Wurtz et al., 2006).
Evidence suggests that in boreal forest of North America early postfire recruits, i.e. those that establish 3-6 years after fire, often dominate intermediate and mature boreal stand ages (Gutsell & Johnson, 2002;Johnstone et al., 2004). Yet specific differences in mortality can shift dominance patterns over a period of decades during regrowth (Gutsell & Johnson, 2002). The present ubiquity of spruce trees in interior Alaska despite the common presence of deciduous cohorts early in succession, for example, indicates that mortality of deciduous trees is higher than that of evergreen trees (Greene & Johnson, 1999;Yao et al., 2001). Despite this knowledge, there have been very few attempts to describe succession at decadal scales in response to boreal burn severity (Johnstone et al., 2004;Keane et al., 2004;Mack et al., 2008). This is due to the difficulty of measuring burn severity, particularly post hoc at the landscape level, and the paucity of long-term monitoring studies (see, however, Johnstone et al., 2004). Meanwhile, rates of biomass accumulation and the relative abundance of deciduous and evergreen species are pivotal to determining climate feedbacks associated with postfire succession (Viterbo & Betts, 1999;Lyons et al., 2008;Mack et al., 2008).
A range of legacy effects of burn severity on North American boreal forests have been hypothesized, with composition playing a central role in the direction and magnitude of productivity and albedo feedbacks to climate (Goetz et al., 2007;Mack et al., 2008). Where burning is severe, deciduous trees are expected to be more abundant for decades following fire disturbance. Since aspen, the dominant deciduous species in Alaska, generally show higher growth rates than black spruce (Yarie & Billings, 2002), more severely burned sites would accumulate biomass more quickly during intermediate stages of succession. Deciduous forests also have substantially higher albedo than evergreen forests, both during winter and summer (Amiro et al., 2006;Liu & Randerson, 2008;McMillan & Goulden, 2008) causing increased albedo with burn severity. Accumulation of soil organic carbon could change too because deciduous vegetation produces more litter than does evergreen vegetation, although deciduous litter is generally more readily decomposed than evergreen litter and could potentially suppress moss growth (Johnstone et al., 2010a).
The balance of different climate forcing agents associated with a shift in vegetation composition will determine the net effect of the boreal fire regime on the climate system (Hinzman et al., 2003;Randerson et al., 2006;Euskirchen et al., 2009). Thus far, however, little empirical work has been done to quantify the impacts of burn severity on species composition at a landscape scale or to assess the associated feedbacks to climate. Here, we investigate whether vegetation patterns across the current Alaskan landscape reflect the expected legacy effects of burn severity over the past six decades. Specifically, we tested whether severely burned areas display higher deciduous cover, biomass, and albedo than less severely burned areas during intermediate stages of succession (10-50 years since fire). To do so, we relied on a chronosequence of current stands with ages, i.e. time since last fire, ranging from 10 to greater than 50 years, an existing biomass map (Blackard et al., 2008), and newly developed maps of deciduous cover and albedo.

Materials and methods
We describe (a) mapping of albedo and deciduous fraction of vegetation, which is defined as the percentage of aboveground biomass (AGB) stored in deciduous species, as derived from remote sensing data; (b) creation of a series of chronosequences reflecting biomass regrowth and albedo following fire; (c) stratification of burns in the study region (interior Alaska) into high-and low-severity burns; (d) assessment of the stratification approach, focusing on data uncertainty and potentially confounding covariates of burn severity. An overview of the main processing steps is given in Fig. 1.

Mapping albedo and deciduous fraction
Total short-wave albedo for spring and summer were calculated from Moderate Resolution Imaging Spectroradiometer (MODIS) Bidirectional Reflectance Distribution Function (BRDF) parameters (MCD43A1, Schaaf et al., 2002) at 500 resolution for the periods March 14-21 and July 28-August 4, respectively, using observations from the years 2004-2008 as input. We first calculated black sky albedo (BSA) at the mean solar zenith angle at 3 hourly intervals using the RossThick-LiSparse-Reciprocal model (Lucht et al., 2000) with MODIS total short-wave BRDF parameters, since the operational MODIS BRDF/Albedo products only provides BSA at local solar noon. Daily all sky albedo was then calculated as a sum of 3 hourly BSA weighted by the ratio of 3 hourly direct incoming radiation over daily total incoming solar radiation plus white sky albedo weighted by the ratio of total daily diffuse radiation over total daily incoming solar radiation (Lewis & Barnsley, 1994). Radiation values were taken from the surface radiation budget data set provided by the World Climate Research Program/Global Energy and Water-Cycle Experiment (WCRP/GEWEX, Stackhouse et al., 2000).
The deciduous fraction (FD) of AGB was mapped using monthly composites of MODIS reflectance and field measurements using an ensemble regression tree model (Random Forest, Breiman, 2001). The field measurements of FD served as training data for the model and were collected in the summers of 2008 and 2009 in areas of homogenous vegetation cover that had not burned since 1987. The basal area of deciduous and evergreen trees and shrubs was measured in 269 plots measuring 10 and 20 m 2 and located along 100 m transects, with a maximum of 5 plots transect À1 . Using allometric equations, plot measurements were converted into deciduous and evergreen AGB from which FD was derived. Equations were produced from raw data available from the Bonanza Creek Long-Term Ecological Research Station (http://www.lter.uaf. edu/ascii/files/230_tree_biomass_data.csv) and augmented with those published by Yarie et al. (2007) and Bond-Lamberty et al. (2003). Plots were spread across 69 MODIS pixels, and data falling within identical MODIS pixels were grouped and regarded as single measurements in the analysis.
MODIS BRDF-adjusted reflectance data (NBAR, MCD43A4) from 2001 to 2008 were composited at 500 m spatial resolution to create monthly images of reflectance in seven spectral bands, based on the quality flags associated with the data (Baccini et al., 2008). Two data sets were created, one with data from different years processed separately, and one with data from the same month in different years pooled. The resulting images in the latter data set do not represent any changes in monthly reflectance since 2001, such as those due to disturbance or land cover, and were used to map FD. All band-month combinations between May and September, comprising 35 images in total, served as predictors in the RandomForest model calibrated using the field observations of FD. RandomForest models are ensembles of regression or classification trees, each using a bootstrap sample of the data where splits of the trees are determined by random subsets of the available predictors, randomly chosen at each node (Breiman, 2001). Furthermore, RandomForest is generally not likely to overfit the model to the data set used and can handle a large number of predictor variables, even when they are correlated, as is often the case with reflectance data. Crucially, performance metrics of RandomForest, such as the percentage of total variance in the reference observations explained by the model, are based on residuals associated with the data not included in the bootstrap sample used for calibration. Here, 500 trees with at least five observations in each terminal node comprised the RandomForest model. The predicted FD values were used to partition aboveground live forest biomass values mapped by Blackard et al. (2008) from a combination of MODIS, land cover, topography and climate images, into deciduous biomass (dBM) and evergreen biomass (eBM) in areas that burned between 1950 and 1991.
In addition to mapping FD in the 2001-2008 period, we also modeled the expected FD in absence of disturbance using growing season length (McKenney et al., 2006), permafrost extent (Jorgenson et al., 2008), geography and topography (elevation, slope, southern exposure, and wetness, Beven & Kirkby, 1979;Skidmore, 1989) as predictors in an environmental envelope model. In this case, FD values in 1600 uniformly distributed points in unburned areas (defined as areas more than 1 km but o10 km from any burns) served as training data for the RandomForest models.

Detecting burned areas
The database of Alaskan fire perimeters produced by the Bureau of Land Management, Alaska Fire Service (AFS), was acquired from the Alaska Geospatial Data Clearinghouse (http://agdc.usgs.gov/data/blm/fire) and used to define burns. Burn age was calculated with the year 2000 as reference, which is the start of the MODIS data record. Only burns between 10 and 50 years old were considered when constructing succession chronosequences to focus the study on intermediate stand ages (sensu Goetz et al., 2007;Mack et al., 2008), and to reduce effects of temporal uncertainty in the FD and albedo maps.
We focused our analysis on landscapes of interior Alaska that would normally be expected to follow a successional pathway ultimately leading from fire disturbance back to a black spruce forest, either directly or relayed via a deciduous phase. For the construction of the successional chronosequences, we excluded (i) fires that had burned areas smaller than 3200 ha, to minimize the effect of small patches in the landscape context and human-lit fires which are usually smaller than 500 ha in Alaska (DeWilde & Chapin, 2006); (ii) areas that were o1500 m from the burn perimeter, to reduce the effects of geographic misregistration in the database; (iii) areas that were o15 km from large roads or 7.5 km from smaller roads or the Alaska pipeline, to avoid effects of human disturbance and fire suppression activities, (iv) areas that had burned more than once since 1950 according to the fire database, to reduce effects of repeated burning, and (v) regions that had an expected FD 465%, as predicted by the environmental envelope model, in order to reduce the influence of areas where vegetation succession was not expected to culminate in evergreen-dominated forests, such as alpine areas.

Stratifying burned areas by burn severity
Landscape-level estimates of burn severity have largely relied on changes (differences) in the Normalized Burn Ratio (dNBR, Lopez Garcia & Caselles, 1991) which is derived from a comparison of infrared reflectance in remotely sensed imagery of the growing season before and following fire events. The absence of such imagery historically renders the dNBR method unsuitable to retrospectively estimate burn severity before the Earth observation satellite era. Hence, we had to rely on nonremote sensing or retrospectively derived attributes of the burn scars to infer burn severity. In theory, burns that occur later in the year are likely to be more severe since boreal soil dries out with time since snow melt and thus becomes more combustible . Analyses of in situ measured burn severity in five 2004 fires appear to confirm this pattern (Barrett et al., 2010). Combining this finding with the empirical evidence that the extent of a burn and its severity are modestly correlated (Duffy et al., 2007) and that the limited fire suppression in interior Alaska has a negligible effect on burn size in our study domain (DeWilde & Chapin, 2006), we hypothesized that large burns (460 000 ha) that occur later in the year (after DOY 160) had higher levels of burn severity. These two criteria, burn timing and size, were then used to stratify all Alaskan burns since 1950 into low-and highseverity burns. The hypothesis underlying the stratification was tested using fires that occurred between 2002 and 2007 and for which July and August MODIS reflectance allowed us to map dNBR as a reference estimate of burn severity. To do this, NBR was mapped yearly from mean July-August reflectance values in MODIS band 2 (841-876 nm) and band 7 (2105-2155 nm) as NBR 5 (band 2Àband 7)/(band 2 1 band 7). dNBR for a fire in year t was then calculated as NBR t 1 1 ÀNBR tÀ1 .
The MODIS dNBR maps for the years 2002-2007 not only allowed us to test the burn severity stratification, but also to test the quality of the burn perimeter database. We tested (i) if the burn perimeters in the database contained remnant unburned areas, which would have to be excluded from analyses of postfire succession, (ii) whether the dNBR in 2002-2007 burns supported the proposed stratification into low-and high-severity burns based on burn size and timing, and (iii) if the burn severity stratification was correlated with landscape characteristics other than burn severity, which might bias patterns in successional trajectories. These three tests are described in more detail below.

Assessment of the burn severity stratification
The influence of unburned areas within burn perimeters. Any observations of successional changes in vegetation characteristics after fire are potentially sensitive to spatial inaccuracies in the burn perimeter database. Although spatial buffering of the perimeters can eliminate uncertainties regarding the exact location of the burn perimeter, it cannot account for the presence of unburned areas within the burn perimeters drafted by the AFS. There is evidence from Canada that the fraction of unburned area within burn perimeters increases with burn size (Eberhart & Woodard, 1987) and that unburned areas tend to have greater deciduous cover (Kafka et al., 2001). Including unburned areas in our analysis could thus bias our observations of successional vegetation changes and their dependency on burn severity, which we partly estimated from burn size. To avoid this, we first mapped unburned areas within burn perimeters where remote sensing data predating the fire were available. We then developed a model to retrospectively map potentially unburned areas in older burn perimeters in interior Alaska relying largely on the controls topography exerts over fire dynamics, primarily through soil moisture (Barrett et al., 2010).
Using the fire perimeter database, we quantified for each post-2001 burn polygon the fraction that might not have burned (i.e. unburned remnant forest patches), by mapping (a) the MODIS dNBR in each polygon and comparing it to (b) the MODIS dNBR of areas outside the burn polygon (41500 and o5000 m from the burn perimeter and unburned since 1950). We then tagged as unburned the cells inside each burn perimeter where the dNBR was lower than the 95th percentile of the dNBR outside the burn (i.e. areas that were not significantly different from unburned areas). This allowed us to calculate for each burn since 2001 the unburned fraction within the burn perimeter. To assess the validity of this method for detecting unburned areas within the burn perimeters, we used light detection and ranging (LIDAR) data acquired by the Geoscience Laser Altimeter System (GLAS) onboard IceSat (Zwally et al., 2002). GLAS data are available since 2003 and are spatially discontinuous, with data recorded in 'shots' with a 65 m ellipsoidal footprint and spaced about 172 m apart along the overpass. GLAS data have previously been used to assess forest disturbance in Alaska (Goetz et al., 2010). We used the GLAS data here to derive estimates of vegetation height distribution (for GLAS processing stream, see Goetz et al., 2010). We used a GLAS metric that reflects the height at which half of the energy returned to the sensor is reached, commonly referred to as the Height of Medium Energy (HOME). HOME was used because we found it was less sensitive to algorithms designed to estimate canopy height by first detecting an initial return signal (from the top of the canopy) and then deconvolving a ground return (from the trailing edge of the LIDAR return). For the purpose of this study, we wanted to determine whether the distribution of HOME was lower in the year following fire than in the year before. Our assumption was that if HOME did not significantly change through time then it was unlikely that the location had experienced disturbance. In all, 12 622 GLAS measurements covering a total of 92 burns between 2004 and 2007 were processed to give pre-and postburn HOME estimates (n preburn 5 5232, n postburn 5 7390).
Once the unburned areas within the perimeters of post-2001 burns were mapped using dNBR, we characterized them by comparing their preburn land cover and their topography to that of the entire area within the burn perimeter. Land cover was estimated using a circa 2000 map at 0.011 spatial resolution [Global Land Cover (GLC) 2000, Bartholomé & Belward, 2005] reclassified into evergreen (GLC classes 4 and 11) vs. nonevergreen. The topographical variables southern exposure (Skidmore, 1989) and wetness [calculated as ln(a/tan b) where a is the local upslope area draining through a certain point per unit contour length and tan b is the local slope, Beven & Kirkby, 1979] were calculated from a digital elevation model and transformed to quantiles of their distribution within the burn. Statistical significance of the differences between unburned and burned areas was assessed within each burn perimeter using a w 2 -test for the evergreen fraction, and Wilcoxon's rank-sum tests for the other variables. Consistent patterns in the relative values of the topographical variables in unburned areas were then investigated across burn perimeters, by testing whether values were centered around 50%, for the quantiles, or 0%, for the differences in evergreen cover, using a Wilcoxon's signed-rank test. The aforementioned topographical parameters and the size of the burn were used to develop a predictive classification model, using Ran-domForest, to retrospectively detect areas that were unlikely to have burned within all burn perimeters since 1950. Finally, we tested if there was a relationship between the fraction of unburned areas in a burn and its size, and if excluding areas that were modeled as unburned changed this relationship.

Comparison with dNBR
The validity of the burn severity stratification based on size and timing was tested using the dNBR. Because pre-and postfire remote sensing data are needed to calculate mean dNBR for a burn, only burns that occurred after 2001, and for which MODIS data were therefore available, were included in this test.

Covariates of burn size other than severity
After masking potentially unburned areas from burn perimeters since 1950 and prior to calculating chronosequences of regrowth after fire, we verified that the stratification of burn severity based on size and date of ignition did not introduce a bias in terms of topography (slope, southern exposure, and wetness) and bioclimate (growing season length and the expected deciduous fraction). If present, such a bias could influence the successional trajectories observed in low-and high-severity burns, independent of burn severity. To test for potential bias, both low-and high-severity burns were split into younger (o30 years) and older stands (430 years), and a Kruskall-Wallis rank sum test applied to detect dissimilarities in topography and bioclimate between the different burn groups (n low severity, young 5 125, n high severity, young 5 9, n low severity, old 5 93, n high severity, old 5 11).

Results
We first describe the deciduous cover map we derived from MODIS remote sensing observations, then assess our approach used to estimate historical burn severity, and finally describe the chronosequences of biomass and albedo in response to burn severity.

Mapping deciduous fraction of vegetation
The deciduous fraction of the vegetation was robustly predicted by the RandomForest model using the MODIS reflectance image products (% variance explained 5 71, n 5 69, Fig. 2a). Spatial patterns in the resulting maps of FD clearly reflect the distribution of burn scars as well as non-disturbance-related ecotones, such as alpine belts and the boreal forest-tundra transition (Fig. 2a, b  and d). The latter also are reflected by the map of expected FD in the absence of disturbance, predicted by a separate RF model (% variance explained 5 54, n 5 1600, Fig. 2c).

Assessment of the burn severity stratification
The influence of unburned areas within burn perimeters. We relied on a comparison of MODIS dNBR in-and outside burn perimeters to detect unburned areas within the burn perimeters, as recorded since 2001 in the AFS database (supporting information Fig. S1). An example is shown for the Boundary Fire in Fig. 3. The comparison of GLAS-derived canopy heights, and the HOME metric in particular, confirmed that the areas identified as unburned based on the MODIS dNBR undergo no detectable structural change (supporting information Fig. S2). Pre-and postburn GLAS-HOME values did not differ in unburned areas (W 5 47 436, P 5 0.55, n preburn 5 271, n postburn 5 336), whereas they did in burned areas (W 5 15 248 949, Po0.001, n preburn 5 5021, n postburn 5 7099).
Unburned remnants within post-2001 burn perimeters tended to occupy more mesic areas than areas that burned (Fig. 4a, V 5 189, Po0.001, n 5 113) and often had a relatively northerly topographic aspect (Fig. 4b, V 5 683, P 5 0.06, n 5 113). In addition, unburned remnants tended to contain more deciduous vegetation before the fire than did areas that burned (Fig. 4c, V 5 458, P 5 0.001, n 5 113). If these areas were not accounted for, observations of deciduous fraction during postfire succession would tend to be biased towards showing greater shifts from evergreen to deciduous cover.
While prefire vegetation cover can only be mapped if prefire remote sensing data are available, topographic wetness and exposure can be mapped for historical as well as recent fires. We exploited the effects that topographical wetness and topographical exposure have on the likelihood of burning to detect potentially unburned remnants since 1950. Using burns since 2001, where dNBR made it possible to distinguish between burned and unburned areas, we calibrated a RandomForest classification model to predict potentially unburned remnants within burn perimeters based on topographical variables, expressed as quantiles, and fire size.
The post-2001 burns that were classified as high severity based on their size and date of ignition, had a larger proportion of unburned areas within their perimeter, based on the dNBR, than those classified as low severity (mean high sev. 5 7%, mean low sev. 5 3%, W 5 726, Po0.01, n high sev. 5 30, n low sev. 5 83). Using the Random-Forest classification algorithm to mask potentially unburned remnants these proportions and the difference between them were reduced (mean high sev. 5 4%, mean low sev. 5 2%, W 5 960, P 5 0.03, n high sev. 5 30, n low sev. 5 83). The algorithm used to mask unburned areas successfully detected 86% of the unburned areas in post-2001 burns. Yet, 74% of the area masked as unburned had actually burned. As a result, 15% of the burned area appeared to have been unnecessarily excluded from our analysis of succession. This deliberate choice in the trade-off between the errors of omission and commission ensured that any effects of unburned areas on the successional patterns were minimized. In low-and high-severity burns after 2001, respectively, 5% and 8% of the area was mapped as unburned based on the dNBR, on average, and 14% and 22% was predicted as unburned by the RandomForest model.
Comparison with dNBR. Mean MODIS dNBR in the burns that were classified as high severity, based on their size and date of ignition, was not statistically significantly higher than in those classified as low severity, (mean high sev. 5 0.38, mean low sev. 5 0.33, Student's t-test: t 5À1.86, P 5 0.06, n high sev. 5 26, n low sev. 5 87) but were less likely to show a low dNBR, while some burns classified as low severity did show a low dNBR (Fig. 5).
Covariates of burn size other than severity. The criteria used to identify high severity burns in the database since 1950, i.e. larger than 60 000 ha and ignition date after DOY 160, did not introduce a topographical or bioclimatic bias. Younger and older high-and low-severity burns did not show a different distribution along gradients of topography (P slope 5 0.52, P exposure 5 0.41, P wetness 5 0.07), growing season length (P 5 0.06), or expected deciduous fraction (P 5 0.33, supporting information Fig. S3). Ultimately, the data selection described above retained 2.6 million ha for the construction of the chronosequences, of which 1.5 million ha were within 218 low-severity burns, and 1 million ha within 20 highseverity burns (supporting information Table S1).

Effects of burn severity on composition, biomass and albedo of regrowth
The deciduous fraction of vegetation in high severity burns was higher than in low-severity burns during the successional period considered here, i.e. from 10 to $50 years after burning (Fig. 6). Deciduous vegetation-dominated biomass in the first 10-20 years of succession (mean DF 5 75%, n 5 81, W 5 3284, Po0.001, H 1 5 df450%). In lowseverity burns, dBM showed a steady decline after 25-30 years (W 5 4028, Po0.001, H 1 : dBM o30 years 4dBM !30 years , n o30 years 5 125, n 430 years 5 93), whereas it increased in high-severity burns (W 5 80, Po0.001, H 1 : dBM o30 years odBM ! 30 years , n o30 years 5 9, n ! 30 years 5 11). At the same time, i.e. after 30 years or more, high-severity burns showed a greater total AGB than low-severity burns (W 5 665, P 5 0.039, H 1 : AGB high sev. 4AGB low sev. , n high sev. 5 11, n low sev. 5 94), in large part due to the difference in deciduous growth (W 5 729, P 5 0.007, H 1 : dBM high sev. 4dBM low sev. ). Boreal biomass estimates calculated from DBH measurements can vary by almost 100% depending on the parameterization of the allometric equations used in the calculation, even when these are species specific, and biomass density can vary greatly within stands (Mack et al., 2008). The forest biomass values reported here were extracted from a map by Blackard et al. (2008)   A difference of À100 indicates that unburned areas were entirely covered by deciduous vegetation before the fire, whereas burned areas were entirely covered by coniferous vegetation. A difference of 100 indicates the opposite and a difference of 0 indicates that burned and unburned areas had an identical ratio of evergreen and deciduous cover before the fire. Black histogram sections indicate burn perimeters with statistically significant differences between burned and unburned areas based on a Wilcoxon's rank-sum test (a-b) or a w 2 -test (c), with a 5 0.05 for all tests. Grey sections indicate burn perimeters where differences are not statistically significant.
which covers the entire US and reportedly included live tree bole wood, stumps, branches, and twigs of trees 2.54 cm in diameter or larger. Biomass values in the ensuing chronosequences presented here ( Fig. 6b and c) correspond relatively well with the original transectlevel live biomass estimates measured in the field in stands of similar ages and used for the calculation of deciduous fractions [12 AE 3 Mg ha À1 (mean AE SE) in 22-30-year-old stands, n 5 13, and 37 AE 8 Mg ha À1 in 30-52year-old stands, n 5 18]. Because winter short-wave albedo was negatively correlated with evergreen fraction (r 5À0.71, Po0.001, df 5 340), i.e. albedo was lower in more densely forested evergreen areas, it decreased with stand age in low-severity burns (Fig. 7). A similar but weaker pattern was observed in summer (r 5À0.55, Po0.001, df 5 340). The higher deciduous fraction in high severity burns in all but the first decade after fire resulted in consistently higher summer albedo for at least four decades of succession (W 5 3702, P 5 0.002, H 1 : albedo high sev. 4albedo low sev. , n high sev. 5 20, n low sev. 5 218), which did not appear to decline with time (W 5 47, P 5 0.72, H 1 : summer albedo o30 years 4summer albedo ! 30 years , n o30 years 5 9, n 430 years 5 11). Differences between winter albedo in severe and nonsevere burns were apparent only after 20 years of succession (W 5 1357, P 5 0.032, H 1 : winter albedo high sev. 4winter albedo low sev. , n high sev. 5 15, n low sev. 5 142).

Post hoc estimation of burn severity
Our stratification of burns into low and high severity, based on their size and date of ignition, did not show a strong relationship with the dNBR of the burns. Burns in the high-severity stratum did have consistently high dNBR values, but so did some burns in the low-severity stratum (Fig. 5). This variability is not surprising because factors that operate at fine temporal and spatial scales are known to influence burn severity Barrett et al., 2010;Boby et al., 2010). The ability of the dNBR Trajectories were calculated from a chronosequence of stands of different ages since burning, starting in 1950, which were stratified by burn severity estimated from the burn size and ignition date. Total biomass in (b) and (c) estimates are extracted from an existing map (Blackard et al., 2008) and were partitioned into deciduous and evergreen biomass using the deciduous fraction mapped here from field data and 2001-2008 MODIS data. Lines are created through second order local polynomial regression fitting (span 5 1.2, Cleveland et al., 1992), where bold lines indicate the predicted mean and grey shading indicate estimated standard errors. MODIS, Moderate Resolution Imaging Spectroradiometer.
to reflect burn severity in boreal regions is limited, however. While the dNBR is useful for distinguishing burned from unburned areas (Lopez Garcia & Caselles, 1991), it is disproportionately sensitive to changes in the vegetation canopy, which in turn limits its sensitivity to the bulk of the carbon consumed by fire in boreal regions, i.e. in the soil organic layer (Murphy et al., 2008;Barrett et al., 2010). In contrast, a recent field survey of 178 distinct Alaskan burns, showed how the depth of burning, which is correlated with burn severity, increased with both burn size and the timing of the fire in the season, supporting the stratification of burns into low and high severity used here (Turetsky et al., 2011). The potential presence of severe burns in the lowseverity stratum, in combination with our averaging of observations within a single burn, most probably attenuated the actual differences in succession between burn severity regimes. Moreover, because we pooled the MODIS data acquired over the 2001-2008 period, in order to calculate FD (see Materials and methods), we limited the sensitivity of our FD predictions to a AE 4 year time window. This might have further limited the distinction between FD in low-and high-severity burns, particularly in very young stands.

Succession-mediated legacy effects of burn severity
Over the 50-year period of regrowth considered here, greater burn severity promoted AGB accumulation as a result of shifting composition towards deciduous trees which typically have higher growth rates (Goetz & Prince, 1996;Mack et al., 2008). Over a longer time period, however, severely burned sites might have lower biomass due to lower spruce cover and self-thinning of dBM. The combination of more decomposable litter from deciduous trees, and a smaller input of relatively recalcitrant litter from evergreen trees, would also limit the buildup of organic carbon in the duff and soil layers in more severe burns, increasing the risk of permafrost degradation (Harden et al., 2006).
Over a period of decades, the radiative forcing associated with successional shifts in plant species composition can offset the forcing caused by greenhouse gas emissions released during the burning of mature spruce forests (Randerson et al., 2006). This is due primarily to an early succession deciduous vegetation phase that has considerably higher albedo during spring and summer than mature conifer forests (Chapin et al., 2000;Amiro et al., 2006;Lyons et al., 2008;McMillan & Goulden, 2008). An important question in this context is the degree to which higher burn severities (which would generate more carbon and aerosol emissions and more short-term positive forcing) might be balanced on decadal timescales by increases in the net primary production, stand density, and canopy cover of deciduous trees that have biophysical cooling properties.
Another important biophysical difference, for example, between boreal deciduous and evergreen vegetation involves rates of mid-summer canopy conductance and latent heat fluxes. Reported latent heat fluxes (evapotranspiration) in deciduous forests during summer are substantially higher than in coniferous stands, whereas sensible heat fluxes show an opposite pattern (Baldocchi et al., 2000;Eugster et al., 2000;Amiro et al., 2006;Liu & Randerson, 2008). Higher mid-summer evapotranspiration fluxes in deciduous stands may be linked with high rates of stomatal conductance and leaf-level evapotranspiration (Ewers et al., 2005;Bond-Lamberty et al., 2009), whereas latent energy fluxes in spruce forests, with their lower productivity and canopy conductance, may have a larger contribution of evaporation from the moss layer. These contrasting carbon and energy balance dynamics are currently not adequately characterized relative to burn severity in process-based ecosystem models (Bond-Lamberty et al., 2007;Euskirchen et al., 2009) importance with respect to climate system feedbacks .

Is boreal Alaska experiencing a biome shift?
The patterns of vegetation regrowth after fire in interior Alaska since 1950 indicate that severe burning promotes deciduous tree species for several decades. We expect the recent episodic fire years 2004, 2005 and 2009, in which a total of 5.7 million ha burned across the region ( $ 5% of interior Alaska), to follow a similar trajectory. Extensive plot level field work in these burns supports this view (Johnstone et al., 2009). Directionality in the extent and severity of the fire regime and the consequences for postburn vegetation successional trajectories we documented here also support expectations of a shift towards a more deciduous-dominated landscape in Alaska in the near future (e.g. at least the next 50 years). These changes have implications not only for climate feedbacks, but also for game management and numerous other land management and resource uses of forest products in the region. We note that even in the absence of increased burn severity, a shortening of fire return interval will limit the development of mature spruce forests, increasing the presence of young deciduous vegetation (Johnstone et al., 2009). Is this trend likely to continue in the future? Climate warming is projected to further intensify the fire regime, generating a higher frequency of extreme fire years characterized by large burns of high severity. At the same time, negative feedbacks from increasingly deciduous land cover have the potential to modify and mitigate future changes. Increasing the area burned annually will gradually reduce the amount of forest that has not recently burned across the landscape, and will reduce the mature conifer forest cover across interior Alaska. Deciduous forests that replace the fire adapted conifer-dominated stands are less likely to burn as a result of their architecture, lower flammability of tree tissues, and limited accumulation of fuels on the forest floor due to more rapid turnover of detritus (Dyrness et al., 1986;Hély et al., 2000;Krawchuk et al., 2006). This mechanism indicates increased deciduous cover may ultimately limit the intensification of the fire regime via a negative feedback effect, especially if warmer temperatures promote the longevity of deciduous trees.
The paleological record shows evidence of a strong black spruce-mediated biological control on the fire regime at centennial to millennial time scales: Between 6500 and 5600 years before present, following millennia when wildfires had been rare, black spruce expanded in interior Alaska, replacing less flammable vegetation communities dominated by birch, white spruce and alder (Lynch et al., 2003). This expansion of coniferdominated vegetation appears to have been largely driven by an increased occurrence of wildfires resulting from, as well as promoting, increased black spruce cover, since there is no evidence of simultaneous climate change that would have intensified the fire regime (Lynch et al., 2004;Hu et al., 2006). However, the current climate in interior Alaska is moving beyond the optimal growth conditions for spruce growth (Beck et al., 2011) and is intensifying the fire regime at decadal timescales to the extent that it favors deciduous rather than evergreen regrowth, thus creating a negative feedback. Therefore, a new tipping point may be approaching where the boreal forest moves from evergreen dominance towards a greater cover of deciduous species. Such a biome shift would be accelerated by the other effects of ongoing climate change on spruce trees, such as drought stress (Barber et al., 2000;Beck et al., 2011) and increased susceptibility to pest outbreaks (Malmströ m & Raffa, 2000). Additional research on burn severity and its influence on vegetation succession in the boreal landscape is needed, particularly with respect to the magnitude and directionality of net feedbacks to climate and associated disturbance regimes.

Supporting Information
Additional Supporting Information may be found in the online version of this article: Figure S1. (a) The 95th percentile of the Moderate Resolution Imaging Spectroradiometer (MODIS) difference Normalized Burn Index (dNBR) observed in an unburned area around fire polygons (41500 m and o5000 m from the burn perimeter and unburned since 1950) was used to screen the fire polygons as registered in a database of Alaskan fires produced by the Bureau of Land Management, Alaska Fire Service (AFS). For areas that burned between 2002 and 2007, areas within the fire polygon that had a dNBR below this associated threshold dNBR value were marked as potentially unburned. b) This screening resulted in an increase of the mean dNBR of the burned areas and c) a reduction in the size of the areas considered burned. Only burns where sets of at least 100 (500 m) MODIS pixels were available to determine the threshold (unburned) dNBR value and the mean dNBR in the burned area are included in the plots. Figure S2. Distribution of vegetation heights as quantified by the Height of Medium Energy (HOME) recorded by the Geoscience Laser Altimeter System (GLAS) onboared Icesat. Height distributions are shown before (grey shading) and after (black lines) fire events in areas marked as a) unburned and b) burned inside the burn perimeters registered in a database of Alaskan fires produced by the Bureau of Land Management, Alaska Fire Service (AFS). Unburned areas were distinguished from burned areas based on the comparison of the difference Normalized Burn Ratio (dNBR) recorded by the Moderate Resolution Imaging Spectroradiometer (MODIS) inside and outside the fire polygons. Figure S3. Topographic slope, exposure, and wetness, meteorological growing season length and the expected deciduous fraction predicted by a bioclimatic envelope model for low severity (LS) and high severity (HS) burns, that are younger and older than 30 years (labelled as young and old, respectively, LS-young: n 5 124; LS-old: n 5 94; HS-young: n 5 9; HS-old: n 5 11). Table S1. Names, sizes and year of burning as registered in a database of Alaskan fires produced by the Bureau of Land Management, Alaska Fire Service (AFS), of the fires classified here as 'high severity' in the creation of deciduous and evergreen biomass chronosequences.
Please note: Wiley-Blackwell is not responsible for the content or functionality of any supporting materials supplied by the authors. Any queries (other than missing material) should be directed to the corresponding author for the article.