Patterns of canopy and surface layer consumption in a boreal forest fire from repeat airborne lidar

Fire in the boreal region is the dominant agent of forest disturbance with direct impacts on ecosystem structure, carbon cycling, and global climate. Global and biome-scale impacts are mediated by burn severity, measured as loss of forest canopy and consumption of the soil organic layer. To date, knowledge of the spatial variability in burn severity has been limited by sparse field sampling and moderate resolution satellite data. Here, we used pre- and post-fire airborne lidar data to directly estimate changes in canopy vertical structure and surface elevation for a 2005 boreal forest fire on Alaska’s Kenai Peninsula. We found that both canopy and surface losses were strongly linked to pre-fire species composition and exhibited important fine-scale spatial variability at sub-30 m resolution. The fractional reduction in canopy volume ranged from 0.61 in lowland black spruce stands to 0.27 in mixed white spruce and broadleaf forest. Residual structure largely reflects standing dead trees, highlighting the influence of pre-fire forest structure on delayed carbon losses from aboveground biomass, post-fire albedo, and variability in understory light environments. Median loss of surface elevation was highest in lowland black spruce stands (0.18 m) but much lower in mixed stands (0.02 m), consistent with differences in pre-fire organic layer accumulation. Spatially continuous depth-of-burn estimates from repeat lidar measurements provide novel information to constrain carbon emissions from the surface organic layer and may inform related research on post-fire successional trajectories. Spectral measures of burn severity from Landsat were correlated with canopy (r = 0.76) and surface (r = −0.71) removal in black spruce stands but captured less of the spatial variability in fire effects for mixed stands (canopy r = 0.56, surface r = −0.26), underscoring the difficulty in capturing fire effects in heterogeneous boreal forest landscapes using proxy measures of burn severity from Landsat.


Introduction
Fire in the boreal region is the dominant agent of forest disturbance with both short and long-term impacts on carbon cycling and climate (Kasischke et Kasischke and Turetsky 2006), and increased fire size and intensity (Beck et al 2011, Gillett et al 2004. Such changes in the fire regime are one potential pathway for disturbance-mediated shifts in biome boundaries (Goetz et al 2005, Ju andMasek 2016).
Global impacts of boreal forest fires depend on the balance between cooling from increased albedo in burned areas and warming from the release of greenhouse gases and aerosols. Over time scales consistent with fire return intervals, cooling from increased spring and summer albedo ultimately surpasses the immediate warming effect of fire carbon emissions (Randerson et al 2006). However, with deep burning of the surface organic layer, particularly in black spruce forests and peatlands, the influence of fire carbon emissions may exceed the albedo cooling effect if fires consume very large amounts of soil carbon (Turetsky et al 2011).
Severe fires in North American boreal forests can alter the course of post-fire succession (Amiro et al 2003, Johnstone et al 2010b, Mack et al 2008, Rogers et al 2015. Intense crown fires in black spruce forests can damage or destroy serotinous cones, leaving few seed trees and long distances from seed sources in neighboring unburned forests (Burton et al 2008, Frelich and Reich 1999, Johnstone et al 2010b. Exposure of mineral soil following a deep-burning surface fire increases soil temperature and decreases soil moisture (Bonan and Shugart 1989)-conditions that promote deciduous seedling recruitment (Johnstone and Kasischke 2005). On longer timescales, permafrost melt may result in a new, drier stable state that supports a long-term transition from spruce to deciduous forests (Chapin et al 2006, Hoy et al 2016, Jafarov et al 2013, Jones et al 2016. Fire effects vary as a function of pre-fire ecosystem structure, topography, moisture, and weather (Jain and Graham 2007). The combination of factors that influence fire effects can therefore be simultaneously spatially extensive and highly variable (Veraverbeke et al 2015). The remoteness of the boreal region, along with spatial and temporal variability in the variables that influence fire effects, make plot-based studies of preand post-fire conditions impractical. Landsat data are commonly used to evaluate fire effects, as changes in vegetation cover and moisture alter near and shortwave infrared reflectance in burned areas (French et al 2008, Benson 2006, Miller andThode 2007). However, relating spectral measures (e.g. differenced normalized burn ratio (dNBR)) to field measures of fire effects in the boreal region is difficult due to low sun angles, topographic variability and persistent cloud cover (French et al 2008, Hoy et al 2008. Moreover, it is not always clear how changes in surface reflectance relate to ecosystem impacts such as consumption of vegetation or surface material (Lentile et al 2006). Nevertheless, recent work suggests that dNBR, in combination with other variables, can provide important constraints on fire carbon emissions (Veraverbeke et al 2015).
Airborne lidar data are commonly used to estimate fine-scale variability in forest structure and terrain elevation, yet repeat lidar data that capture pre-and post-fire conditions are rare. Pre-fire stand structure (e.g. canopy bulk density, crown volume, canopy base height) and fire fuels have been well characterized using single-date lidar (e.g. Andersen et al 2005, Riaño et al 2003). Lidar has also been collected after fires to assess post-disturbance recovery based on vertical forest structure within the burn scar (Bolton et al 2015). However, very few studies directly estimate fire effects using repeat lidar (except see McCarley et al 2017). And, while lidar measurement of terrain elevation change is common (e.g. Jones et al 2013), only one previous study used repeat lidar to map surface layer removal due to combustion in a temperate wetland environment (Reddy et al 2015).
In this study we used airborne lidar data from 2004 and 2009 to directly estimate changes in canopy vertical structure and surface layer removal from the 2005 Fox Creek Fire on Alaska's Kenai Peninsula. A combination of forest inventory, repeat lidar, and Landsat data were used to explore three questions: (1) How does canopy consumption vary based on pre-fire structure and composition? (2) How does fire-induced surface elevation change vary by topographic position and covary with changes in forest structure? (3) Do Landsat estimates of burn severity track changes in ecosystem structure identified from repeat lidar? As a case study, this work seeks to bridge the scale gap between sparse field plots and extensive satellite-based assessments of fire effects in boreal forests to separate vegetation and surface changes that contribute to ecosystem, carbon cycle, and climate impacts from fire.

Materials and methods
2.1. Study site and field data The Fox Creek fire (hereafter FC2005) burned approximately 10 643 ha from July to September of 2005 along the southern shore of Lake Tustamena on Alaska's Kenai Peninsula (figure 1). The burn occurred in low elevation (30-300 m ASL) black spruce (Picea mariana) forest, with white spruce (Picea glauca) and broadleaf species (e.g. Betula papyrifera, Populus tremuloides) common at elevations >100 m and welldrained areas (Andersen 2009, Klein et al 2005. Field data from the USDA Forest Service Forest Inventory and Analysis (FIA) program were used for calibration and validation of remote sensing data. The location and species of individual trees in 105 subplots (7.3 m radius) were aligned with lidar data to estimate lidar metrics (supplemental figure S1 available at stacks.iop.org/ERL/12/065004/mmedia). Sub-plots were located throughout the Kenai Peninsula, and all plots were covered by the 2004 and 2009 lidar acquisitions. Eight sub-plots were located inside FC2005, but plot data both inside and outside the burn perimeter were only used to establish pre-fire forest composition.

Lidar data preparation and analysis
Airborne lidar data were collected in May 2004 (Andersen 2009) and September 2009. We analyzed 275 ha of lidar coverage within the FC2005 fire (figure 1), a sample that covers the full range of topographic variability and burn severity within the burn (further detail in supplemental material S2). Mean pulse densities were 4.5 m À2 and 3.7 m À2 for the 2004 and 2009 acquisitions, respectively. Lidar data were coregistered using a standard geoid model and clipped to a common spatial extent. Raster digital terrain models (DTM) and canopy height models (CHM) were generated at 1 m spatial resolution following standard processing methods (supplemental figure S2; Cook et al 2013). Canopy volume was used to summarize forest canopy structure in this study, calculated as the sum of 1 m canopy heights within 5 m or 30 m grid cells. Canopy volume depends on both tree height and fractional cover (calculated as the percent of 1 m pixels > 2 m in height within a 5 m or 30 m grid cell).
Tree crowns were delineated using watershed segmentation (Alonzo et al 2014). In boreal forests, and particularly in dense black spruce stands with small crowns (<2 m in diameter), segments may contain more than one stem. For this study, our goal was to isolate segments that were structurally and compositionally homogenous, but possibly containing >1 stem (Bortolot 2006).
Lidar and Landsat metrics were used to classify crown segments into three species classes: black spruce, white spruce, and broadleaf. Lidar canopy metrics were derived from the CHM (e.g. crown volume, crown diameter) and a set of terrain variables was assigned to each crown segment from the lidar DTM (elevation, slope, and aspect). To capture phenological information of pre-fire forest conditions, a forward finite difference noise metric was extracted from a dense Landsat NBR time series (158 May--September images spanning 1987-2004; Alonzo et al 2016). The phenology 'noise' metric was calculated from LEDAPS surface reflectance imagery (Masek et al 2006) and assigned to all crown segments in a given 30 m pixel. The set of variables maximizing species separability were used for training a canonical discriminant analysis classifier (see supplemental material S3 and Alonzo et al 2013 for details). Application of the calibrated discriminant functions across our study area resulted in a 1 m resolution map of species composition for each crown object within the lidar coverage.
Pre-and post-fire DTMs were standardized to estimate elevation changes within the burn scar. We constructed an empirical planar correction between acquisition dates, using unburned samples along the lidar transects both inside and outside the burn perimeter. The correction was subsequently applied to all areas. A road segment was used to evaluate the residual elevational error between the corrected terrain models, following methods outlined in Jones et al (2013). We use the term 'surface layer removal' to describe reductions in terrain elevation within the burn. Surface layer removal likely reflects consumption of surface litter and organic soils, but we cannot uniquely attribute elevation changes to combustion. Subsidence due to changes in subsurface hydrology, permafrost melt, or erosion could contribute to elevation losses identified using repeat lidar. Seasonal differences in lidar acquisitions between May 2004 (some areas leaf off) and September 2009 (all leaf on) likely result in a more conservative estimate of surface layer removal, as leaf-on conditions in 2009 could result in higher ground elevation.
All lidar data products (e.g. fractional cover, 2004 and 2009 canopy volume, tree species, surface layer removal) were gridded at both 5 m and 30 m resolution to consider fine-scale patterns of fire effects and facilitate spatial correlations (Pearson's r) with Landsat-based estimates of burn severity. For species, the mode 1 m classification was applied at both the 5 m Environ. Res. Lett. 12 (2017) 065004 and 30 m scales. Specifies-level analyses were performed using the 5 m grid since mixing of white spruce and broadleaf was common at the 30 m scale.

Burn severity products
Landsat dNBR and the burn severity classification from Monitoring Trends in Burn Severity (MTBS, www.mtbs.gov) were compared to lidar-based estimates of canopy and surface layer removal. We created analogs to the Landsat burn severity classification using fractional canopy removal and surface layer removal estimates at 30 m resolution. Both classifications (severities 1-4, as in MTBS) were executed using k-means clustering. Relative dNBR (RdNBR; Miller and Thode 2007) from Landsat was excluded from this study due to very high (r ¼ 0.99) correlation with dNBR within the FC2005 perimeter.

Tree species classification
Lidar and Landsat metrics accurately separated black spruce, white spruce, and broadleaf species. Mean cross-validation accuracy (n ¼ 212 crown segments) was 73%, or 82% when only separating conifer and broadleaf classes. Terrain elevation and Landsat-based phenology were the two most important variables for species classification. Combined, lidar and Landsat data captured the pre-fire distributions of black spruce, white spruce, and broadleaf forest cover across the NW, SW, and SE transects, highlighting the predominance of homogeneous black spruce stands throughout the study site (figure 2). Fine-scale information also illustrated the degree to which white spruce and broadleaf canopies mix, a key control on fire effects.

Pre-fire structure and composition
Pre-fire forest structure varied with landscape position and dominant species class. Similar pre-fire fractional cover (fCov; landscape median ¼ 0.41) for all classes masked important variability in canopy volume from differences in tree heights among species (table 1). At the 30 m scale, black spruce (0.43) and white spruce (0.42) had similar fCov. However, taller crowns in white spruce and broadleaf stands (median heights of 12.4 m and 17.2 m, respectively) than black spruce stands (7.3 m) and greater tree spacing resulted in finescale variation in canopy volume (figure 2(a) and (b)).
Black spruce stands accounted for 55% of pre-fire canopy volume. Black spruce dominated flat lowland (100%) and flat upland sites (63%, table 1), yet lowland and upland black spruce stands had markedly different forest structure. In flat, lowland stands, tree heights were lower (7.3 m) and tree spacing was higher (fCov ¼ 0.41) compared to upland stands (8.0 m, fCov ¼ 0.58). In contrast, broadleaf and white spruce crowns were distributed evenly across all landscape positions with the exception of flat lowland areas, and mixed stands contributed the greatest proportion of canopy volume on all sloped terrain (table 1).

Fire effects: canopy consumption
The magnitude and variability of canopy volume (cVol) removal depended more on species and landscape position than pre-fire cVol (figure 2). Independent of species, pre-fire volume was a poor predictor of fractional cVol change (r 2 < 0.001). Black spruce in flat, lowland sites lost the highest fraction of cVol (0.61; table 1) while upland black spruce exhibited smaller changes (0.22). The white spruce fractional cVol loss of 0.42 was evenly distributed across landscape classes, but fine-scale variability was higher (coefficient of variation ¼ 0.31) compared with black spruce sites (cv ¼ 0.21). Remnant structure in mixed stands likely reflects unburned or partially-consumed canopies. Broadleaf areas had the lowest cVol loss (0.18), despite similar pre-fire cVol to that of white spruce, as crown fires are rare in broadleaf stands.
Broadly, MTBS burn severity classes were strongly correlated with reductions in fractional and absolute canopy removal from repeat lidar (figures 3(a) and (b)). Lack of structural complexity and compositional heterogeneity in black spruce stands allowed for spectral sensitivity to canopy removal at all burn severities. Both absolute and fractional cVol removal within the black spruce class correlated with dNBR variability (r ¼ 0.62 and 0.76, respectively; figures 3(a) and (b)). At MTBS burn severity classes 3 and 4 (only black spruce present), fractional cVol removal saturated at~0.8 while sensitivity of dNBR was possibly related to within-site variability in pre-fire biomass or exposure of underlying mineral soils ( figure 3(b)). In mixed stands, fine-scale heterogeneity in pre-fire volume and the patterns of pre-fire composition (figure 2(c)) weakened the relationship between dNBR and fractional cVol removal (figure 3(b), figure 4(b)). Post-fire cVol losses in these landscapes were patchy, suggesting a mosaic of fire effects from lower intensity surface fires, particularly in white spruce-dominated areas, and distinct patterns of residual structure compared to black spruce sites (figures 2(d) and (e)). Compared to black spruce, substantially more white spruce volume was lost at lower MTBS burn severities (figure 4) underscoring the challenges of applying spectral data in mixed stands and variable fire intensity ( figure 3(b)).

Reduction in terrain elevation
We used repeat lidar data to directly estimate the depth of surface layer removal within the FC2005 fire. The standard deviation of repeat lidar measurements along a main road (0.04 m) was used to characterize uncertainty in estimates of elevation loss. Additionally, the difference in mean elevation changes between 2004 Environ. Res. Lett. 12 (2017) 065004 and 2009 in severely burned areas (MTBS classes 3 and 4) and unburned areas (MTBS class 1 or outside burn perimeter) of 0.10-0.12 m were statistically significant (two sample t-test, p < 0.0001).
Removal of surface material varied with stand composition and topographic position. Inside the burn perimeter, median surface layer removal was 0.08 m (table 1) Evidence for covariation of canopy and surface layer removal was primarily driven black spruce stands (r ¼ 0.70, figure 4(a)), with lower correlations between canopy and surface changes in white spruce (r ¼ 0.54, figure 4(b)) and broadleaf stands (r ¼ 0.31). This result is consistent with the development of deeper organic soils in black spruce forests compared to broadleaf or mixed stands. Greater surface layer removal was associated with higher MTBS burn severity classes (figure 3(c), supplemental figure S5). The median depth of surface layer removal at low burn severity was 0.00 m while the removal depth at moderate to high severities was 0.20 m (table 1). The saturation of the lidar metric at higher dNBR values may be due to complete consumption of the surface organic layer (figure 3 (c)). Within moderately to severely burned areas (MTBS burn severity >2), flat lowlands areas had the greatest surface elevation losses (0.21 m), south-facing slopes lost the least (0.09 m), and, median flat upland loss rose considerably (0.18 m), largely due to the dominance of black spruce. As with fractional cVol removal, there was substantially more spatial variability of burn severity as measured by surface removal in the NW transect compared to the MTBS classification (figure 4).

Discussion
This study provides the first estimates of changes in canopy structure and surface elevation in a boreal forest fire based on repeat airborne lidar acquisitions. Fire effects were heterogeneous, even within species classes, highlighting the advantages of airborne lidar for estimating changes in ecosystem structure over large areas compared with sparse field plots or moderate resolution Landsat data. In black spruce landscapes, more homogeneous stand structure and topography led to high canopy consumption and deeper surface layer removal. Mixed species stands had heterogeneous patterns of crown and surface layer changes, with inconsistent relationships between losses in surface elevation and crown volume. In mixed stands, Landsat dNBR was largely insensitive to losses in canopy volume; without additional data, Landsat dNBR may lead to an underestimate of fire emissions and carbon stock changes in modeling studies. Lidar-based estimates of residual woody structure provide important insights regarding the delayed release of carbon in aboveground biomass Table 1. Pre-fire structure and losses in fractional cover (fCov), canopy volume (cVol), and surface layer from fire, summarized by topographic position following Turetsky et al 2011. Flat areas <6% slope and upland areas were >100 m elevation.

Landscape Class
Area (   Environ. Res. Lett. 12 (2017) 065004 following fire. Fire-induced changes in canopy and surface properties also influence albedo in burned areas (Randerson et al 2006) and contribute to finescale heterogeneity in understory light and moisture environments important for post-fire forest succession (Bonan and Shugart 1989, Hungerford and Babbitt 1987, Johnstone and Kasischke 2005.

Variability in fire effects
The fire community generally seeks reporting of fire effects in the form of direct, site-dependent impacts (e.g. vegetation consumption/mortality, surface scorching) that control post-fire recovery ( Lentile et al 2006). In contrast, airborne lidar data separated changes in fractional cVol and surface layer removal that may be confounded in Landsat data ( figure 4). The degree to which estimates of surface elevation changes in this study uniquely reflect combustion of surface litter and organic soils (e.g. Reddy et al 2015) merits further study with repeat lidar data collected at a range post-fire intervals.
Using lidar-derived measurements of canopy and surface layer removal, we quantified the extent to which pre-fire vertical structure, species composition, and topographic position contributed to spatial variability in fire effects. Similar to findings from Wulder et al (2009), pre-fire vertical structure was not strongly correlated with Landsat burn severity. Other studies point to a range of factors that influence fire effects, including structural variables and weather, moisture, vegetation composition, and topography (Jain and Graham 2007, Lentile et al 2006, Rogers et al 2015. Vegetation composition did influence cVol removal in the FC2005 fire; losses were highest in lowland black spruce areas, consistent with higher susceptibility to crown fires in dense black spruce stands (Rogers et al 2015, Johnstone et al 2010a). White spruce also suffered substantial fractional losses in canopy volume. In our study area, bark beetle infestation in the 1990s (Berg et al 2006) potentially exacerbated white spruce vulnerability, and may therefore have contributed to the patchwork pattern of canopy loss (Chapin et al 2006). Consistent with the spatial distribution of organic layer development (Bonan andShugart 1989, Johnstone andKasischke 2005), surface losses were substantial, homogeneous, and correlated with canopy losses in black spruce stands but surface loses were smaller and patchier in mixed stands (figure 4).

4.2.
Linking changes in ecosystem structure to postfire recovery The ability to derive spatially-explicit measurements of canopy and surface layer removal from lidar presents opportunities to better evaluate post-fire ecosystem stability. In the FC2005 fire, lidar data suggest substantial variability in the spatial patterns of consumption, particularly between more uniformly burned black spruce and patchy consumption in mixed stands. Larger fires with spatially consistent intensity are generally more severe (Kolden et al 2012, Lentile et al 2006, leading to longer distances from seed sources within the burned area and possibly contributing to a shift in post-fire succession (Beck Across the Kenai, FIA plot data suggest that organic soils are shallow, and only 6% of inventory plots had organic soil depths >0.2 m (max ¼ 0.29 m). Lidarbased estimates of surface layer consumption in severely burned areas (0.21 m) may therefore indicate complete combustion of organic material and exposure of mineral soil. At least in black spruce landscapes, Landsat dNBR retained sensitivity to the depth of surface layer removal (figure 3(c)) and may therefore provide a means to better constrain organic layer consumption in emissions models (e.g. Veraverbeke et al 2015). Substantial relative surface layer removal (Johnstone and Kasischke 2005), along with consistently high levels of black spruce canopy consumption, generates site conditions that may support a shift in successional trajectory to broadleaf species (Johnstone et al 2010a). However, previous studies of post-fire recovery suggest that mesic conditions on the Kenai reinforce ecosystem stability (Kane et al 2007, Kasischke et al 2010. included sparse field measurements to calibrate carbon losses in black spruce forests, inadequate accounting for pre-fire structure, and limited ability to quantify post-fire residual structure (Veraverbeke et al 2015). In the FC2005 study area, spatially extensive repeat lidar suggest that residual forest structure in live and dead-standing carbon pools may be substantialup to 50% of pre-fire canopy volume. Even standing dead trees can delay the release of carbon in aboveground woody biomass for decades (Powers et al 2013). Connecting pre-fire species composition with patterns of residual structure and delayed decomposition further demonstrates important species-dependent variability in post-fire albedo (Rogers et al 2015).
Combustion of soil organic matter accounts for the greatest proportion of carbon emissions from boreal forest fires (Turetsky et al 2011). Repeat lidar data captured variability in depth of burn that has proven difficult to characterize with in-situ data (Turner et al 2003). Barrett et al (2010) used spatial modeling to infer the loss of soil organic matter compared to field plots, but empirical models have limited portability without local calibration sites. In the FC2005 fire, lidar-based estimates of surface elevation loss were consistent with depth-of-burn values from Turetsky et al (2011), but lidar captured spatial patterns of surface losses based on pre-fire species composition and topographic position. Complementary information from field, lidar, and spatial modeling approaches present opportunities for improved mapping of soil carbon losses to better constrain fire emission estimates for the boreal region.

Conclusion
Changes in ecosystem structure from fire in the boreal forests are spatially heterogeneous. Pre-fire information on forest structure, species composition, and topography from lidar provided important context that can be difficult to recreate using only post-fire observations. Canopy volume losses varied by forest type, but residual forest structure was substantial for all species classes, highlighting the delay in fire carbon emissions from aboveground biomass and contribution from woody structure to post-fire albedo. Surface elevation changes from repeat lidar were consistent with depth of burn estimates from boreal forest field sites. However, lidar data offer a broader landscape sample to evaluate fine-scale patterns of canopy and surface consumption. Future research on regrowth and successional dynamics following fire would benefit from the acquisition of lidar and spectral data at multiple intervals following fire. Solidifying the linkages between pre-fire structure and post-fire effects will improve our understanding of fire risk, fire behavior, and the mechanisms for larger scale biogeochemical and biophysical feedbacks in the Earth system.