Indicators of burn severity at extended temporal scales: a decade of ecosystem response in mixed-conifer forests of western Montana

Additional keywords: char, hyperspectral remote sensing, multiple endmember spectral mixture analysis, QuickBird.

The potential of wildland fires to influence the cycle of carbon dynamics and ecosystem composition is widely recognised (Dale et al. 2001; Loehman et al. 2014; Rocca et al. 2014). Remote sensing is the most suitable method for monitoring these fire effects over extended spatial and temporal scales (Lentile et al. 2006; Morgan et al. 2014). Characterising severity and ecological condition of the post-fire environment (e.g. distance to seed source, vegetation structure and composition, likelihood of soil erosion) is important for forecasting future effects on ecological systems (Lentile et al. 2006; Hayes and Robeson 2011). The majority of severity assessments focus on either the immediate post-fire changes or effects in the year following the fire, especially when severity products from the Monitoring Trends in Burn Severity (MTBS) project are utilised (Cocke et al. 2005; Lentile et al. 2006, 2009; Kolden et al. 2015). Studies employing MTBS products are also spectrally and spatially constrained; spectrally, they only employ two spectral indices that have been widely critiqued (Roy et al. 2006; Kolden et al. 2015; Smith et al. 2016), while spatially, they are limited to data at the scale of Landsat pixels (30 m). In contrast, the post-fire environment is heterogeneous, with considerable spatial variability even within comparable burned patches at the scale of a Landsat pixel (Smith et al. 2005) that may or may not be captured depending on the scale and method of post-fire assessment (Smith and Hudak 2005; Holden et al. 2010; Kolden et al. 2012; Viedma et al. 2012). Such spatial heterogeneity is common, given mixed-severity fire regimes create a wide variation in fire effects that manifest as a fine-scale mosaic of stand-replacing and surface fire (Agee 1998; Arno et al. 2000). This variation is shaped by complex landscapes, including topography, moisture gradients, and the diversity and flammability of vegetation (Perry et al. 2011).

Vegetation response in the first post-fire year greatly depends on the extent of the damage to plants and soil above and below ground, which is a function of the severity of the fire and the resilience of the affected ecosystem (Neary et al. 1999; Diaz-Delgado et al. 2003; Abella and Fornwalt 2015). The surge of new vegetation growth in the first year may seem incongruent with the initial burn severity classification (Lentile et al. 2007a) due to a boost in plant-available nitrogen (Stephan et al. 2012) and nutrients remaining in the ash layer (Neary et al. 1999); thus, it is important to monitor vegetation and soil recovery characteristics over time (Sankey et al. 2013; Berryman et al. 2014). The resilience of the ecosystem to the fire disturbance depends on specific conditions of the pre-fire environment such as vegetation type and density, and forest health such as drought or insect kill, taken into consideration with the scale and intensity of effects of the fire on the vegetation and soil (Idris et al. 2005; Schoennagel et al. 2008; Bowman et al. 2009; Perry et al. 2011).

Mappable ground cover components of interest in the post-fire environment are often green vegetation (GV), brown (senesced or dead) non-photosynthetic vegetation (NPV), charred material (char), ash and exposed mineral soil (Sá et al. 2003; Lewis et al. 2007; Smith et al. 2007a, 2007b; Eckmann et al. 2008). Of particular interest immediately post fire are char and ash, which are measurable biophysical variables that can be scaled between fires, as well as between field and remotely sensed images (Smith et al. 2005). The amount of char present and its persistence through the first post-fire year have been suggested as a suitable index of burn severity (Lentile et al. 2006; Quintano et al. 2013, 2017; Fernandez-Manso et al. 2016). Smith et al. (2007a, 2007b), Hudak et al. (2007) and Lentile et al. (2009) collectively used linear spectral mixture analysis to estimate the subpixel fraction of char to classify burn severity in African savannas and in multiple north-western United States forest ecosystems. Lentile et al. (2009) determined that char fraction outperformed dNBR (differenced Normalized Burn Ratio), when calculated from 1-year post-fire imagery, in predicting many of the field measures at both canopy and subcanopy scales. Hudak et al. (2007) observed in a study of eight north-western Rocky Mountain fires that changes in fractional vegetation and char cover either outperformed or were not significantly different from common severity spectral indices. Beyond linear spectral mixture analysis (SMA), several studies have employed multiple endmember spectral mixture analysis (MESMA) to monitor trajectories in post-fire recovery following field assessments of severity (Quintano et al. 2013, 2017; Fernandez-Manso et al. 2016).

Monitoring ecosystem patterns and post-fire response over years and decades has been conducted by employing temporal series of remotely sensed datasets (e.g. Hicke et al. 2003), predictive modelling (e.g. Kashian et al. 2006) and long-term field data collections (e.g. Romme et al. 2011), although the combination of these is uncommon. In the present study, we assembled a unique suite of post-fire field and remotely sensed data from three 2003 western Montana wildfires over 10 years. The combination of these datasets gave us a comprehensive view of changing soil and vegetation conditions in the burned area, pre-fire through 10 years post fire. Collectively, these data were used to assess whether any immediate field or remotely sensed indicators could forecast recovery over longer time periods and at a finer spatial resolution, building on related studies (e.g. Smith et al. 2007a; Lentile et al. 2009). In the present study, we focus on a definition of burn severity (Lentile et al. 2006; Parsons et al. 2010) that describes the longer-term effects of fires on vegetation and soils (Morgan et al. 2014). Specifically, we define burn severity as the effects of the fire on the ground surface, as measured by plant mortality and char cover. Although post-fire ‘recovery’ can have many implications and definitions (Bartels et al. 2016), for the purposes of this study, we concentrate on vegetation regrowth and change in ground cover, particularly the reduction in char, over the 10-year study period.

In this study, our objectives were to: (1) use data collected immediately post fire to confirm which factor(s) were most indicative of low, moderate and high burn severity; (2) evaluate the utility of high- and moderate-resolution remotely sensed imagery to discern post-fire char and other biophysical variables on the ground; (3) assess the contribution of the field measures to the degree and variability of burn severity and disturbance over a decade; and (4) analyse ecological change over 10 years using field data collected in 2013, and compare these data with the remotely sensed imagery to determine whether initial conditions are indicative of longer-term response.

The three study landscapes ranged from dry to wet mixed-conifer forest types in western Montana (Fig. 1) that were historically burned by mixed-severity fires at a fire return interval of ~80 years (Fischer and Bradley 1987). The driest landscape sampled was 2854 ha burned by the Black Mountain 2 Fire (henceforth referred to as Black Mountain) (centroid latitude and longitude: 46°50′29″N, 114°10′41″W; elevation: 1072–1743 m) on 8 August 2003. The Cooney Ridge Fire (centroid latitude and longitude: 46°40′10″N, 113°49′27″W; elevation: 1247–2167 m) also began on 8 August 2003 and burned 8589 ha of generally moist mixed-conifer forest. The third landscape containing the wettest mixed-conifer forest was burned by the Wedge Canyon Fire (centroid latitude and longitude: 48°54′22″N, 114°24′14″W; elevation: 1141–2414 m) that began on 18 July 2003 and burned 21519 ha.

Fig. 1.  The location of the 2003 wildfires in Montana with the site and plot layouts. Four sites were sampled on the Black Mountain Fire, seven at the Cooney Ridge Fire and five at the Wedge Canyon Fire. Each site had 9 plots and each plot had 15 subplots. (DBH, diameter at breast height.)

Warm, dry sites at lower elevations throughout most of the Black Mountain Fire and much of the Cooney Ridge Fire were co-dominated by ponderosa pine (Pinus ponderosa Lawson and C. Lawson) and Douglas-fir (Pseudotsuga menziesii [Mirb.] Franco) in the Pseudotsuga menziesii habitat type series (Pfister et al. 1977; Cooper et al. 1991). The higher-elevation forests of the Wedge Canyon Fire on the western border of Glacier National Park were characterised by the Abies lasiocarpa habitat type series (Pfister et al. 1977; Cooper et al. 1991), with subalpine fir (Abies lasiocarpa [Hook.] Nutt.), Engelmann spruce (Picea englemannii Parry ex Englem.) and lodgepole pine (Pinus contorta Douglas ex Loudon). Lodgepole pine and western larch (Larix occidentalis Nutt.) also commonly occurred at many sites across all three fires. The soils across the Black Mountain and Cooney Ridge Fires were predominately gravelly silt loams and gravelly loams, whereas the Wedge Canyon Fire soils were sandy and silty glacial tills (USDA Natural Resources Conservation Service (NRCS) Web Soil Survey, http://websoilsurvey.nrcs.usda.gov/app; accessed 5 June 2017). Mean annual precipitation varied from 40 to 130 cm (Pfister et al. 1977) across the sites, mainly as a function of elevation. July is historically the warmest month at all sites with average annual high temperatures of 28–30°C and December is the coldest month with average annual low temperatures of −11 to −7°C (Montana Climate Office, years 1981–2010; http://climate.umt.edu; accessed 5 June 2017).

Field data were collected in 2003 and in 2004 to document burn severity and initial post-fire change. We assessed the functionality of mapping these components using airborne hyperspectral (HS) imagery collected immediately post fire. Decadal field data (collected in 2013) were used to evaluate changes in vegetative cover relative to pre-fire conditions. Anniversary Landsat data were acquired in 2002, 2003, 2004, 2007 and 2013 to map initial pre- and post-fire conditions and to approximate a trendline of vegetation response (the change in ground cover components (understorey GV, NPV, soil and char) over time) on the sites. High-resolution satellite QuickBird (QB) imagery from 2007 was used as an additional midpoint approximation of recovery between field data campaigns.

Post-fire soil and vegetation data were collected between 11 September and 22 October 2003 at four sites on Black Mountain, seven sites on Cooney Ridge and five sites at Wedge Canyon (Table 1). These sites were sampled as soon as safety and logistics allowed after the fires were officially out following established post-fire protocol (Lentile et al. 2007b), and were selected using the initial post-fire Burned Area Reflectance Classification (BARC) map (RSAC 2005) as a guide, and sample plots were then randomly located within areas burned across a range of low, moderate and high burn severities (Hudak et al. 2007) based on the predominant overstorey canopy condition: live, green (low); scorched, brown (moderate); or charred, black (high) (Table 2).

Table 1.  Remotely sensed image characteristics and acquisition dates for each platform, fire and image

Table 2.  Field site characteristics and the corresponding remotely sensed estimates for each field campaign: (a) 2003; (b) 2004; and (c) 2013
Data are sorted by mean char cover on the ground in 2003. MTBS refers to the Monitoring Trends in Burn Severity project; NPV is uncharred non-photosynthetic vegetation; GV is green vegetation; HS refers to the hyperspectral imagery; nd indicates no data

Each site was centred a random distance between 80 and 140 m from the nearest access road, within a consistent stand and burn severity condition. As described in more detail by Hudak et al. (2007), sampling at each site was spatially nested, consisting of nine 8 × 8-m plots (Fig. 1) with each plot comprising 15 1× 1-m subplots, for a total of 135 subplots per site. Plot centres were geolocated with a global positioning system (GPS) and differentially corrected. Sites were revisited from 12 to 15 July 2004 on the Black Mountain and Cooney Ridge Fires, and 13 to 14 August 2004 on the Wedge Canyon Fire. Field data were collected at only the site centre (A plot, 15 subplots per site). Five subplots per site were resampled 10 years after the fire (6 to 14 August 2013): one each at the centre of the A plots, and at the B, D, F and H plots resituated 30 m away in the same orthogonal directions as in 2003 (Fig. 1).

At the subplot scale, the percent cover of understorey GV, rock, mineral soil, ash, litter and any large organic matter (logs, branches or stumps) was ocularly estimated. The percentage char, if any, of each ground cover component was also recorded. The cover measures of ground materials were constrained to sum to unity, whereas the charred fractions of the ground materials were summed separately to get an aggregate estimate of percentage char. At the plot scale, litter and duff depths were measured, canopy closure was estimated using a convex spherical densiometer, canopy vegetation condition (% green, brown or black) was estimated at one plot per site, and a digital photo was taken for reference. To ensure consistency and repeatability between field crews, sampling protocol was calibrated daily and there was at least one common person on the field crew for each of the three field campaigns.

Multiple endmember spectra of ash, soil, rock, GV, NPV and charred NPV materials (Fig. 2) were collected in 2003 and 2004 on cloud-free days within 2 h of solar noon within all burned areas using an ASD Pro-FR field spectroradiometer¹ (ASD 2002). Additional bright-target reference calibration spectra were acquired in a ~100-m² area of highly reflective light-coloured soil (such as a wide road intersection) on each fire. All field spectra were collected following the methods in Lewis et al. (2011), were compiled by fire and cover type, and are provided online for public use (http://www.frames.gov/partner-sites/assessing-burn-severity/spectral/; accessed 1 December 2016).

Fig. 2.  The four spectral libraries that were used for the multiple endmember spectral mixture analysis (MESMA). Atmospheric water absorption bands were removed at ~1400 and 1900 nm. GV is green vegetation; NPV is non-photosynthetic vegetation. The three soil samples were from different road intersections, all bright, uncharred soils. Blue, green, red and near-infrared (grey) Landsat and QuickBird bands are shaded, as well as two short-wave-infrared Landsat bands (grey).

Airborne hyperspectral imagery was collected shortly after fires were contained in the fall (autumn) of 2003 (Table 1). The Probe-1 whisk-broom sensor was flown at 2100 m above ground level and data were collected along tracks 2.3 km wide and up to 28 km long from one end of the fire perimeter to the other, corresponding to a 512-pixel swathe width with an approximately 5 × 5-m ground instantaneous field of view at nadir. Image acquisition was between 1100 and 1300 hours local time to minimise shadowing; the solar zenith angle within 1 h of solar noon in the fall at this latitude was between 36 and 46° (45° and lower is desirable to reduce shadowing). Thirteen of the sixteen field sites fell within two flight lines each on the Black Mountain and Cooney Ridge fires and one flight line on the Wedge Canyon fire, whereas three sites fell outside the airborne imagery.

We used the Landsat Thematic Mapper (TM)-based burn severity dataset provided through the MTBS program (Eidenshink et al. 2007) to assign a burn severity class post-hoc to each site for comparison with our initial classes. MTBS data provide a standardised classification of unburned, low, moderate and high severities for all large fires based on pre- and 1-year post-fire dNBR data. We also acquired Landsat 5 images from years 2002–04 and 2007 and Landsat 8 images from 2013 that corresponded with post-fire anniversary dates (Table 1).

QuickBird-2 satellite imagery was acquired in the summer of 2007 (Table 1) over several large areas that encompassed all three fires in our study. The QB satellite collected data over a 16.8-km swath in five bands: a panchromatic band and four visible and near-infrared bands (DigitalGlobe, Longmont, Colorado, USA). QB imagery (2.4-m multispectral; 0.65-m panchromatic at nadir) allows fine-scale spatial discrimination of ground components in the post-fire environment. Images at each wildfire were delivered as multiple geotiff files that were combined into radiometrically corrected and orthorectified image mosaics.

Image and field spectra processing steps are summarised in a flowchart in Fig. 3.

Fig. 3.  An overview of the endmember selection and image processing steps. (ATREM, Atmospheric Removal Program; FLAASH, Fast Line-of-sight Atmospheric Analysis of Hypercubes; RTGC, radiative transfer ground calibration; ELC, empirical line calibration; MESMA, multiple endmember spectral mixture analysis; RMSE, root mean square error; EAR, endmember average RMSE; MASA, minimum average spectral angle; CoBI, count-based endmember selection index.)

Before delivery, the airborne hyperspectral data were converted to reflectance using FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes) (FLAASH Spectral Sciences, Inc., Burlington, Massachusetts, USA) (Black Mountain and Cooney Ridge Fires) or ACORN 4.0 (Atmospheric Correction Now; AIG 2002) (Wedge Canyon Fire) by the image provider. Atmospheric conditions were more favourable at the Wedge Canyon Fire situated more than 200 km north of the other two fires, where smokier conditions demanded different atmospheric correction techniques. All reflectance data were further refined with a Radiative Transfer Ground Calibration (RTGC) (Clark et al. 2002). To ensure that all images were comparable, reflectance data were also corrected using an Empirical Line Calibration (ELC), which minimises atmospheric and illumination effects by standardising them across all bands in the image (Richards and Jia 1999). Wavebands between 1340–1480 and 1810–1970 nm were eliminated from all images because of atmospheric water absorption in these bands, and band 2512 nm was removed for noise, leaving 119 bands for analysis. The excluded water absorption ranges were slightly wider for the Wedge Canyon imagery: 1280–1480 and 1735–1995 nm, leaving 108 bands for analysis.

On-board GPS and inertial measurement unit geolocation data, together with a 30-m digital elevation model, were used for georeferencing the imagery. On examination, the georeferenced images were distorted; therefore, we warped the imagery with a high density of ground control points (60–130 per flightline, proportional to area) via Delauney triangulation. Orthorectified digital orthoquads (1-m) were used as base imagery, necessitating resampling to 5-m resolution. No error statistics are reported for Delauney triangulation; however, accuracy was thoroughly checked between the warped images and the base images to visually confirm that features (such as road intersections) easily recognised on the ground and independently geolocated with a resource-grade GPS (i.e. with differential correction) spatially corresponded.

In linear SMA, the solution of a linear model enables the calculation of the relative proportions that a given cover type contributes to pixel reflectance (Roberts et al. 1993; Settle and Drake 1993). The linear mixture model (eqn 6 in Cochrane and Souza 1998) is defined by:

where R_i is the spectral reflectance of the ith spectral band of a pixel; R_i,j is the spectral reflectance of endmember j in band i; f_j is the fraction of endmember j; and e_i is the error in band i. MESMA utilises the same linear model as SMA, but the endmembers mapped vary on a pixel-by-pixel basis rather than the strict SMA constraint of each pixel containing the same endmembers (Roberts et al. 1998). MESMA code is part of VIPER (Visualisation and Image Processing for Environmental Research) Tools (VIPER Tools, Santa Barbara, CA, USA).

VIPER Tools have a built-in, iterative endmember process that allows the selection of the most representative endmembers relative to the objects of interest (Tompkins et al. 1997; Dennison and Roberts 2003). Related yet spectrally distinct endmembers (e.g. GV functional types) may be combined into a single endmember class. MESMA also allows for variations present within the same material, such as different soil types, which can all be accounted for within a single ‘soil’ spectral library. Computational intensity is increased with a high number of endmembers, whereas too few may result in large residuals (Dennison et al. 2004; Li et al. 2005; Roth et al. 2012).

It has been suggested that natural and disturbed systems are best modelled by two- or three-endmember models (Powell and Roberts 2008; Wang et al. 2012); thus, we used VIPER Tools to investigate two-, three- and four-endmember models to estimate char, soil, GV and NPV cover and photometric shade after wildfire. A large set of potential endmembers (59 unique endmembers, with 5–30 replicates of each) was iteratively narrowed down. Endmembers were first sorted into broad classes: GV, NPV, soil and char. Within the GV class, multiple species of deciduous and conifer trees, shrubs, forbs and grasses were considered to account for spectral variability, particularly in the sparsely vegetated post-fire environment (Okin et al. 2001). The NPV class consisted of various uncharred litter and bark materials. Uncharred soil and rock made up the soil class, and a separate char class had dark char, and charred soil and NPV. Two to four ‘best’ endmember signatures were selected for each category based on how well they spectrally represented the cover class; the same set of endmembers was used on all three fires (Fig. 2).

MESMA was applied to the 20 post-fire HS image strips. The minimum allowable endmember fraction, maximum allowable endmember fraction, and maximum allowable shade fraction were set to the default values of −0.05, 1.05 and 1.0 respectively, and the maximum root-mean-square error (RMSE) was 0.05. Typically, the shade fraction is set at 0.8, but because of low albedo in some areas of the HS imagery due to tall trees and topographic shading (Dennison et al. 2004; Yang et al. 2015), and the preponderance of dark char, we were able to extract more spectral information by increasing the shadow allowance. By analysing the models that mapped the greatest percentage of the area of interest, we were able to iteratively select the best endmembers and endmember combinations (Youngentob et al. 2011) until the best model results (i.e. accurate fractional estimates over the greatest area) were obtained.

The same endmember bundles were used for MESMA modelling of the Landsat and QB images. QB only has four spectral bands; thus, we were constrained to two- or three-endmember models. Modelling efforts were most successful via a series of two-endmember models in which each endmember component was estimated separately (plus shade); the relative cover fractions of each pixel were calculated by combining the separate model results.

Pixel values from the fractional cover images were extracted at the field subplot locations (i.e. 135 subplots per site). These data were aggregated along with the field data at the plot (15 subplots per plot), and site (nine plots per site) scales. Most sample distributions were non-normal (via the UNIVARIATE procedure in SAS (SAS Institute Inc. 2003)); therefore, non-parametric statistical tests were used for subsequent analyses. The MESMA cover estimates of char, soil, GV and NPV from the remotely sensed imagery were compared with the 2003, 2004 and 2013 field data to evaluate the ability to make accurate estimations (2003 data) and predictions (2004 and 2013 data). Direct (e.g. 2003 image and field data) and predictive (e.g. 2003 image and 2013 field data) correlations among the variables were assessed using the non-parametric Spearman rank correlation test (SAS Institute Inc. 2003); Spearman correlation coefficients (ρ) are reported. A significance level of 0.05 was used for all statistical tests.

At the plot scale (~10 m), we characterised the variability of fire effects on soil and understorey vegetation and canopy condition. In 2003, plots classified as MTBS high burn severity had 92% average char cover and a trace of understorey GV, whereas moderate and low burn severity plots had less char (64 and 53% respectively) and sparse understorey GV cover (2%) (Table 3). Understorey GV was not a differentiating indicator of fire effects immediately after the fire. However, based on the 2004 extended assessment of coincidental plots, understorey GV increased considerably to 14% on moderate burn severity plots and 28% on low burn severity plots. Char was still most prominent on high burn severity plots (74%) in 2004 and less prevalent on the moderate (42%) and low (27%) burn severity plots (Table 3). Remaining canopy cover was low on all sites (25–56%), with high-severity sites dominated by charred canopy (23%), moderate-severity sites dominated by NPV (including dead and scorched) canopy (35%), and low-severity sites dominated by green canopy (23%).

Table 3.  Comparing the change in mean ground and canopy cover percentages at the spatially coincident A plots for all fires; standard error in parenthesis
The burn severity classes are based on the Monitoring Trends in Burn Severity (MTBS) extended assessment. In 2013, all plots but one were resampled (one moderate-severity plot was inaccessible). Cover fractions besides char sum to 100%. NPV is uncharred non-photosynthetic vegetation; GV is green vegetation. The sum of the canopy vegetation components in a row is the mean canopy cover of that burn severity class

At the site scale (n = 13 sites where HS imagery was available), there were significant positive correlations between three field components (2003) and their remotely sensed counterparts (Table 4a): char cover and the HS char estimate (ρ = 0.60); NPV cover and the HS NPV estimate (ρ = 0.68); and green canopy cover and the HS GV estimate (ρ = 0.75). The most significant correlation between Landsat and field data was canopy GV (ρ = 0.70). We also found correlations were split by inorganic and organic cover classes (e.g. HS and Landsat char were correlated with soil cover, and HS and Landsat GV were correlated with NPV cover), as the cover variables logically aligned by broader cover classes such as inorganic or organic, and charred or uncharred.

Table 4.  Direct (e.g. 2003 image and 2003 field data) and predictive (e.g. 2003 image and 2013 field data) Spearman rank correlation coefficients for: (a) 2003; (b) 2004; and (c) 2013 field cover measures
HS refers to hyperspectral imagery; NPV is uncharred non-photosynthetic vegetation; GV is green vegetation. Bold values are significant at P < 0.05; correlation coefficients with an absolute value of <0.1 are denoted by ‘<0.1’

In 2004, the most notable direct correlations were between Landsat GV with GV on the ground (understorey) and in the canopy (ρ = 0.72 and 0.84 respectively) (Table 4b). For the first post-fire year, HS char remained positively correlated with char (ρ = 0.74) and soil (ρ = 0.61) on the ground, and negatively correlated with GV on the ground or in the canopy (ρ = −0.66 and −0.78 respectively) (Table 4b). Hyperspectral GV was correlated with canopy GV (ρ = 0.75). The char and soil correlations remained mostly consistent from 2003 to 2004, whereas the Landsat GV correlations were stronger with vegetation regrowth.

The change in site characteristics over the first 4 years (2003–07) on three representative sites after the Cooney Ridge wildfire (Fig. 4) implied a significant shift in dominant cover from char to NPV and GV, regardless of the MTBS burn severity classification. Initially, char was prevalent on the high burn severity site, with more mixed components on the moderate burn severity site, and a prevalence of understorey GV on the low burn severity site. The transformation in site characteristics was apparent in the images: char was mostly replaced by soil and NPV, and on the high and moderate burn severity sites, GV (likely canopy GV from trees that later died) was replaced primarily by NPV. On the low burn severity site, GV was mixed with NPV rather than char after 4 years.

Fig. 4.  A comparison of the immediate post-fire images (2003) and those collected 4 years later (2007) and 10 years later (2013) across a range of burn severities. The multiple endmember spectral mixture analysis (MESMA) results are shown next to the true colour (red, green, blue) images. Resolution (pixel size) varied by image type: hyperspectral (5 m); QuickBird (2.6 m); Landsat (30 m). All three site subsets are from the Cooney Ridge (CR) Fire, sites 1, 5, and 2 with plot locations overlaid; site characteristics are shown in Table 2. (GV, green vegetation; NPV, non-photosynthetic vegetation.)

In 2007, when the QB and Landsat midpoint images were acquired, the trends from our field data (solid lines on Fig. 5) show an estimate of 40% understorey GV cover on moderate and low burn severity sites and closer to 20% on high burn severity sites. Understorey GV was seemingly underestimated in the QB imagery across all burn severities, whereas the Landsat data overestimated understorey GV on the low-severity sites (Fig. 6). Soil, char and NPV ground cover were closely approximated on the low- and moderate-severity sites by both QB and Landsat in 2007; on the high-severity sites, however, soil and char were underpredicted by the remotely sensed images and NPV was overpredicted (Fig. 6).

Fig. 5.  Means of the understorey ground cover field measures (solid lines) in 2003, 2004 and 2013 by Monitoring Trends in Burn Severity (MTBS) burn severity classes: (a) high; (b) moderate; and (c) low burn severity. The mean Landsat estimates from the multiple endmember spectral mixture analysis (MESMA) are shown as dashed lines. (GV, green vegetation; NPV, non-photosynthetic vegetation.)

Fig. 6.  Field site characteristics visually compared with 2002–04 and 2007 Landsat 5, 2013 Landsat 8, 2003 hyperspectral and 2007 QuickBird fractional cover image estimates. A 95% confidence interval is shaded around the field means (understorey ground cover) to illustrate how well the image estimates compare with the range of variability measured in the field. (GV, green vegetation; NPV, non-photosynthetic vegetation.)

Ten years after the fires in 2013, there was 10% or less char or exposed soil left on the plots, and the primary cover classes were NPV (20–34%) and understorey GV (~60–80%) (Table 3; Fig. 4). These data indicate a significant trend towards site recovery regardless of initial burn condition. Interestingly, the canopy data do not reflect as considerable change as do the ground data; canopy char on the high-severity sites (20–30%) and canopy GV on the low-severity sites (20–25%) are similar across 10 years, whereas the most notable change is the loss of canopy NPV as nearly all the scorched, brown needles fell from trees (Table 3).

The 2013 Landsat data underpredicted understorey GV and overpredicted char across all burn severity classes (Fig. 6). Although there was little char left on the ground, there was still nearly 30% charred (black) canopy remaining on the high-severity sites in 2013 (Table 3). Correspondingly, there was significant GV on the ground, but less than 25% in the canopy. Moderate-spatial-resolution sensors like Landsat will often detect occluding canopy rather than occluded surface condition (Hudak et al. 2007), which would account for the prediction discrepancies. Landsat estimates of NPV and especially soil cover were more accurate (Fig. 6).

The only significant direct correlation between 2013 Landsat 8 data and the ground data was between Landsat GV and canopy GV (ρ = 0.76); Landsat 8 char and NPV were negatively correlated with canopy GV (ρ ~−0.8). The 2003 Landsat data indicated a few predictive correlations: between char and soil measured on the ground and in the image (ρ ~0.6) and between Landsat GV and canopy GV (ρ = 0.76). Predictive correlations between the 2003 HS remotely sensed estimates and the 2013 field data were very similar to the 2013 Landsat 8 data, and were significant between canopy GV and HS char (ρ = −0.88), HS NPV (ρ = 0.65) and HS GV (ρ = 0.84) (Table 4c).

We were able to reasonably estimate char cover on all sites from the HS imagery because of its cover dominance and spectral separability (Figs. 4, 6). The significant correlations (ρ = 0.6–0.74) between image-predicted char and the field measures in the first post-fire year (Table 4a, b) combined with the prevalence and persistence of char lead us to confirm that char is a consistent indicator of burn severity, which is in agreement with prior studies in other ecosystems (French et al. 2004; Smith et al. 2005; Veraverbeke et al. 2012; Quintano et al. 2013). The data distributions in the first full year of the present study were bimodal, as the low and moderate sites had comparable cover fractions of char, understorey GV and NPV and soil, whereas the exclusively charred high burn severity sites were more distinctly separable (Table 3, Fig. 4). Quintano et al. (2017) also found post-fire data collected after a Mediterranean wildfire to more naturally separate between two classes: low–moderate and high, rather than the three classes commonly used (low, moderate and high).

There was more char remaining on the high burn severity plots 1 year after the fire (74%) than on the moderate burn severity plots immediately after the fire (64%). The amount of residual char through the first post-fire year and beyond is a measurable biophysical variable that can be scaled between fires (Smith et al. 2010), between field and remotely sensed images (Quintano et al. 2017), and may be used to forecast ecological response and future vulnerability, an area of remote sensing that will likely see further investigation (Smith et al. 2014). Lentile et al. (2009) suggested fractional char cover as estimated via spectral mixture analysis of Landsat Enhanced Thematic Mapper Plus (ETM+) data was the best correlate to 1-year post-fire field measures, thus providing a physical indicator of eventual site recovery. This result was also confirmed in later studies (Hudak et al. 2013).

Negligible understorey GV was found immediately after the fire; however, more overstorey GV (tree canopy) remained: 23 and 11% on the low and moderate burn severity plots (Table 3). Moderate severity sites had 35% NPV canopy cover in 2003, which decreased to 22% by 2004, and almost completely disappeared by 2013. Scorched needles in the canopy and subsequently on the ground are a signature of moderate burn severity.

Because our study objectives concentrated on understorey vegetation and ground cover, we only have tree crown condition data (e.g. green, scorched, or charred) on one plot per site per field year. Although most of the canopy is consumed or blackened in high burn severity areas, tree crowns remaining on the low and moderate burn severity sites can occlude fire effects on soils, which may be disproportionately decoupled from the overstorey condition immediately post fire (e.g. prolonged soil heating due to a sustained, low-intensity surface fire that consumes much of the duff layer (Ryan 2002)). We observed these conditions at two Cooney Ridge field sites (4 and 6) that had 100% green crown yet 67 and 62% ground char cover (Table 2). Canopy occlusion of fine-scale ground effects likely further confounded our remotely sensed estimates on other sites as well.

Neither NPV nor soil cover were diagnostic indicators of burn severity immediately after the fire. Uncharred NPV cover was low (<15%) on all sites in 2003 and soil cover was somewhat ubiquitous (40–60%), regardless of burn severity (Tables 2 and 3). Quintano et al. (2013) also confirmed that the amount of soil cover did not change predictably among burn severity classes. Soil cover decreased by 20–50% in 2004, yet was still a primary cover factor (Table 3); however, 10 years after the fire, less than 10% exposed soil remained on any of the plots. The change in exposed soil due to a disturbance (or recovery from a disturbance) can be an indication of the degree of disturbance (Page-Dumroese et al. 2009), such as after a wildfire or logging.

The Landsat MTBS extended assessment classes suggested there were far more low burn severity sites than we initially assessed. Even though these sites were mostly charred in 2003 (>50%), significant vegetation recovery occurred in the first post-fire year, a trend that was sustained for 10 years (Figs. 5, 6). This is consistent with Bataineh et al. (2006), who found that even in lightly burned areas, removal of forest canopy or litter stimulated understorey vegetation growth for more than a decade following a wildfire in Arizona.

While soil and char cover decreased consistently over time, NPV and understorey GV increased; the rates of increase varied depending on the initial burn severity condition (Figs. 5, 6). Most notable was the increase in understorey GV on the high burn severity sites from 0 to 60% over the 10-year study period. This is comparable with findings of Shive et al. (2013), who found four-time greater understorey cover in areas burned at high severity compared with low severity 8 years after the Rodeo–Chediski wildfire in Arizona, and Wittenberg et al. (2007), who measured vegetation cover equivalent to pre-fire levels 5 years after a Mediterranean wildfire. Idris et al. (2005) reported following wildfires in Borneo and in China, the recovery time for areas burned at high severity was also highly dependent on the scale of the fire – larger patches took longer to recover. The relatively rapid understorey recovery we measured on these western Montana fires can partly be attributed to historical mixed-severity fire regimes in this region (Arno et al. 2000), which result in a mosaic of forest structure and plant communities, and often patchy fire effects.

As stated, we found inorganic variables changed more consistently and predictably during the post-fire period considered, while organic cover measures understorey GV and NPV responded non-linearly and exhibited greater discrepancies between the ground and canopy. Viedma et al. (1997) modelled asymptotic vegetation recovery over a decade in the Mediterranean using Landsat 5 data. The three field data points over a decade from our study are not sufficient to create a good fit, but it is possible to visualise that the green vegetation is increasing asymptotically (Figs. 5, 6).

Understorey GV and NPV cover replaced char and inorganic cover as primary ground cover 1 year after the fire on low burn severity sites (Fig. 5). We estimate this same benchmark of recovery occurred on the moderate burn severity sites approximately 2 to 3 years later, and another 3 to 4 years later (7 or more years after the fire) on the high burn severity sites (Fig. 5). As stated previously, the true shape of these recovery trajectories is almost certainly non-linear, as expected from a biological process (Peters et al. 2004), but importantly, the up-or-down tendencies in these basic post-fire projections are corroborated by the QB and Landsat estimates, substantiating the revegetation trends. At the scale of our study sites, recovery among the burn severity classes converged after 10 years in terms of major cover classes (Table 3, Fig. 5).

The shift and convergence of primary cover types over time emphasise the dynamic nature of the post-fire environment. Char is most prevalent immediately after the fire, but in the following decade, char is concealed by an abundance of NPV and understorey GV. Although none of the initial measures were statistically indicative of eventual understorey GV, HS char, HS NPV and HS GV estimates from 2003 were significantly correlated with the amount of canopy GV found in 2013. As expected, the absence of char and any presence of understorey GV immediately after the fire indicated a greater amount of understorey GV 10 years later (Table 4c). It is notable that initial remotely sensed HS estimates were sustained over a decade, suggesting their potential use for longer-term vegetation recovery predictions. The considerable presence of understorey GV and NPV 10 years post fire emphasises a productive vegetation response in the burned environments (Table 3).

High-spatial- and spectral-resolution HS imagery allowed effective discrimination of fine-scale post-fire ground cover, especially char, as others have found (van Wagtendonk et al. 2004; Kokaly et al. 2007; Robichaud et al. 2007). The time lag between site burning, field sampling and image acquisition as well as the difference in conditions between sites (e.g. smoke) likely affected correlations between the image and field data across all sites, as described in detail by Hudak et al. (2007). Ideally, all data would be collected as temporally coincidentally as possible; however, field campaigns and tasked image collections are often challenging to coordinate, and we felt fortunate to have collected all initial data within 3 months of burning.

As supplied from the vendor, the HS image data were distorted even after georeferencing. To compare the field data with the image data at coincidental locations, warping and resampling the images was necessary. Although these intensive adjustments may have affected data at the subplot (1–5-m) scale, all field and image data were aggregated to the site scale (0.01 km²) and we feel confident that the spatial variability and true means of the data were captured at this resolution on aggregation to the site level. The potential for misaligned data was further overcome by the unusually intensive ground sampling effort undertaken in the initial assessment, with 15 times as many 1-m² subplots per plot in 2003 compared with 2004 and 2013, over an area equivalent to nine Landsat pixels in 2003 and five pixels in 2004 and 2013. The initial sites were also situated randomly but well within large (0.1-km²) patches of a given burn severity condition to minimise edge effects confounding the relationships to Landsat resolution imagery (Hudak et al. 2007).

We had some difficulty with shadowing and low albedo in the both the HS and QB images. Shadows are common in forested areas, and these study areas had significant topographic relief. Shadows were apparent from both tall trees and steep terrain, and the HS imagery was acquired from a fixed-wing aircraft, which added to off-nadir viewing angles, especially compared with satellite imagery. Additionally, the burned sites averaged 65% char cover in 2003, which decreased the overall albedo of the images and thus spectral contrast (Rogan and Yool 2001). We compensated for some of the darkness and shadowing by allowing more shadow to be mapped as well as increasing the maximum allowable RMSE to 0.05 or 0.1 as needed. Overall mean RMSE for the MESMA modelling was low: 0.01–0.02. The Cooney Ridge QB image had the highest RMSE at 0.06. Adjusting the constraints of the MESMA model within VIPER Tools allowed more pixels to be mapped in several images.

QB data were acquired from the DigitalGlobe image archive (https://browse.digitalglobe.com/imagefinder; accessed 16 September 2016). QB is more often used as a ‘ground truth’ dataset in ecological studies (Willis 2015) and somewhat less often for high-level analysis (such as spectral unmixing) because of its broadband spectral resolution. It is worth noting that there is no conceptual reason why SMA cannot be applied to any type of calibrated, multispectral imagery. Arguably, the estimation of char, understorey GV and other ground-cover components is a more scalable method in the context of burn severity assessments than the much more commonly considered Normalized Burn Ratio (NBR) (Veraverbeke and Hook 2013). Indeed, the absence of a short-wave infrared (SWIR) band in QB and most other high-resolution imagery precludes the calculation of NBR. However, our primary intention in this study was to use the midpoint images (QB and Landsat) to validate the field data trends (Fig. 6), and both image types served that purpose well.

The rate of post-fire environmental change (e.g. increase in understorey GV and decrease in exposed char and soil) is high in the first year and slows with time. The initial post-fire conditions were highly spatially heterogeneous at the scale of our study, with the exception of sites burned at high severity, which were primarily charred. We found char cover to have considerable persistence; it remained the primary cover factor for high burn severity 1 full year after the fire. Although ~25–75% char remained on site in the first post-fire year across all burn severities, the ratio of understorey GV and NPV to char and soil cover was much higher after a growing season, indicating resurgence of the vegetative strata and consequent protection of the soil strata. An increase in NPV cover reduces the risk of soil erosion and runoff, helps maintain soil moisture and provides microsites for vegetation establishment.

Substantial vegetation regrowth was measured 10 years post fire across all fires and burn severities. Understorey GV, NPV and inorganic cover proportions were nearly equal on all sites 10 years post fire regardless of whether they were initially classified as low, moderate or high burn severity. The convergence of cover fractions between burn severity classes after 10 years may constitute the single best indicator of post-fire recovery. The shift in associations from char (immediately post fire) to understorey GV and litter (10 years post fire) occurred at varying rates depending on the initial degree of disturbance, and represents a continuum of responses that we were able to infer from the initial post-fire HS images and confirm from the 2007 QB and 2007 and 2013 Landsat images. The proportional cover combinations of char and soil, understorey GV and NPV capture the dynamics of burn severity and vegetation recovery. This study adds to the growing evidence that scalable fractional cover components have a firmer biophysical basis than field (e.g. Composite Burn Index, CBI) or remotely sensed (e.g. dNBR) burn severity indices with less interpretability (Roy et al. 2006).

The authors declare that they have no conflicts of interest.