Article

Predicting Potential Fire Severity Using Vegetation,

Topography and Surface Moisture Availability in a
Eurasian Boreal Forest Landscape
Lei Fang 1 ID
, Jian Yang 2, * ID
, Megan White 2 and Zhihua Liu 1
1 CAS Key Laboratory of Forest Ecology and Management, Institute of Applied Ecology,
Chinese Academy of Sciences, Shenyang 110016, China; fanglei@iae.ac.cn (L.F.); liuzh@iae.ac.cn (Z.L.)
2 Department of Forestry and Natural Resources, TP Cooper Building, University of Kentucky, Lexington,
KY 40546, USA; 12mgn11@gmail.com
* Correspondence: jian.yang@uky.edu; Tel.: +1-859-257-5820

Received: 16 January 2018; Accepted: 6 March 2018; Published: 8 March 2018

Abstract: Severity of wildfires is a critical component of the fire regime and plays an important
role in determining forest ecosystem response to fire disturbance. Predicting spatial distribution of
potential fire severity can be valuable in guiding fire and fuel management planning. Spatial controls
on fire severity patterns have attracted growing interest, but few studies have attempted to predict
potential fire severity in fire-prone Eurasian boreal forests. Furthermore, the influences of fire
weather variation on spatial heterogeneity of fire severity remain poorly understood at fine scales.
We assessed the relative importance and influence of pre-fire vegetation, topography, and surface
moisture availability (SMA) on fire severity in 21 lightning-ignited fires occurring in two different fire
years (3 fires in 2000, 18 fires in 2010) of the Great Xing’an Mountains with an ensemble modeling
approach of boosted regression tree (BRT). SMA was derived from 8-day moderate resolution imaging
spectroradiometer (MODIS) evapotranspiration products. We predicted the potential distribution of
fire severity in two fire years and evaluated the prediction accuracies. BRT modeling revealed that
vegetation, topography, and SMA explained more than 70% of variations in fire severity (mean 83.0%
for 2000, mean 73.8% for 2010). Our analysis showed that evergreen coniferous forests were more
likely to experience higher severity fires than the dominant deciduous larch forests of this region,
and deciduous broadleaf forests and shrublands usually burned at a significantly lower fire severity.
High-severity fires tended to occur in gentle and well-drained slopes at high altitudes, especially
those with north-facing aspects. SMA exhibited notable and consistent negative association with
severity. Predicted fire severity from our model exhibited strong agreement with the observed fire
severity (mean r2 = 0.795 for 2000, 0.618 for 2010). Our results verified that spatial variation of fire
severity within a burned patch is predictable at the landscape scale, and the prediction of potential
fire severity could be improved by incorporating remotely sensed biophysical variables related to
weather conditions.

Keywords: fire severity; surface moisture; remote sensing; spatial controls; boreal forest;
Great Xing’an Mountains

1. Introduction
Wildfires, ignited by human or natural agents, are crucial disturbances in the boreal forests of
Eurasia and North America [1–3]. Wildfires can strongly influence regional land surface processes
such as carbon cycling [4,5] and energy and water budgets [6,7]. Severe burns can result in tree
mortality and soil erosion, thereby degrading ecosystem functions [8]. Nevertheless, there is a
growing consensus that forest wildfires can also provide a unique opportunity for ecosystem

Forests 2018, 9, 130; doi:10.3390/f9030130 www.mdpi.com/journal/forests

Forests 2018, 9, 130 2 of 26

restoration [9,10]. In particular, fire severity can exert profound impacts on the successional trajectories
of the early post-fire vegetation [11,12], which may ultimately determine the future forest structure and
function [13–15]. Many recent studies have indicated that warming and drying climates tend to shift
the current fire regimes toward more frequent, large burns with high severities [16–18]. Understanding
the causal mechanisms of fire severity patterns is essential for mitigating the adverse effects of fires,
and for maintaining beneficial ecosystem functions and services.
Fire severity is generally defined as the magnitude of ecosystem changes caused by fire [19,20].
It is often based on metrics obtained from the field that represent short-term fire effects in the immediate
post-fire environment (e.g., tree mortality and soil organic matter loss) [19,21], or based on the validated
relationships between remotely sensed spectral indices (e.g., normalized burn ratio (NBR) and
differenced normalized burned ratio (dNBR)) and field-measured metrics [19–22]. It has been shown
that variations in fire severity patterns are at the heart of how an ecosystem will change in response
to fire events [23–25]. Spatially explicit fire severity maps can assist resource managers in evaluating
post-fire biomass loss and developing sound strategies for ecological restoration [26,27]. Numerous
studies that incorporate field investigation and remotely sensed images have been conducted to map
fire severity variability within burned patches [19,20,27,28]. However, a potential fire severity map
may be even more useful in allowing managers to anticipate the hotspot areas that are likely to burn
severely, and thus to prioritize resources for fuel treatments for those areas [29,30].
Fire behavior is regulated by three major controlling factors (i.e., weather, topography, and
fuels) that form a fire triangle at spatial scales ranging from a few hectares to thousands of square
kilometers [31]. Fuel composition and fuel loading interact with terrain features and fire weather
to influence burning duration and heat flux, which ultimately determine the spatial pattern of fire
severity [20,32]. Various methods have been developed to represent the cause-and-effect relationships
between environmental factors (drivers), fire behavior (process), and fire effects on an ecosystem
(patterns). Those methods generally fall into two categories: physical models (e.g., FIREHARM and
FOFEM) that simulate fire spread and effects based on fire spread physics developed from laboratory
experiments [29,33–35], and the empirical approaches that draw relationships from the analysis of
existing fire severity patterns [36–38]. In the past two decades, empirical relationships between fire
severity pattern and its spatial controls have gained considerable attention [37–40], with scientific
aims to identify the primary drivers and quantify their explanatory power on spatial heterogeneity of
fire severity.
Since fire severity can be evaluated across a range of spatial scales, environmental variables
in empirical models need to be consistently described at comparable spatial scales for maximum
predictive power. Variables representing fuels and topography are considered to be bottom-up
controls because their influences on fire patterns are pronounced largely at fine spatial scales [41].
Those variables can be easily observed and/or resampled at various scales without losing their
precision. For example, when fire severity is assessed in the field (site scale), the fuel properties and
topography measured in situ are often used to build relationships with fire severity [39,40]. When fire
severity is quantified at coarser scales (e.g., fire patch level), vegetation variables derived from remotely
sensed forest maps or forest inventory analysis databases, and topographic features developed from
digital elevation models (DEMs) are often re-sampled to a coarser spatial resolution (upscaling) [42–44].
Vegetation and topography have been widely recognized as the dominant controls on fire severity in
many types of forest ecosystems [37,38,42,43,45].
Fire weather is considered a top-down control of wildfires as its influences on fire behavior are
pronounced at broad scales. Although fire weather has been proven an influential driver of fire severity
at the fire patch level [37,42,43,46,47], its influences and predictive power on spatial heterogeneity
of fire severity remain poorly understood at fine scales. Current meteorological data are usually
collected at a coarse spatial resolution, which is inconsistent with the fine scale at which site-level fire
severity is assessed and predicted. The viewpoints that emphasize that fire weather is less important
than vegetation or topography on regulating fire severity may be problematic as the low-resolution
Forests 2018, 9, 130 3 of 26

fire weather may not represent the spatial variations of in situ meteorological conditions. Spatially
interpolated data from weather stations can mitigate this issue, but in regions where the weather
stations are scarce, this technique is inadequate. The lack of availability of credible and timely weather
data limits the understanding of spatial controls on fire severity.
Recent advances in remote sensing techniques have provided a new opportunity to examine
fine-scale fire weather effects on fire severity. Fire weather conditions are often characterized using
the metric of fuel moisture content, which reflects dryness of dead fuels and water deficit of live
biomass. Remotely sensed surface evapotranspiration (ET) products, which capture broad relationships
between surface moisture availability and fuel dryness and vegetation drought-stress [44,48,49],
have seldom been applied in modeling fire severity, even though they possess relatively high temporal
and spatial resolutions.
Furthermore, prediction of fire severity patterns based on the empirical understanding of spatial
controls has been well studied in North American boreal and western US forest landscapes [50–52],
but there is still a lack of comprehensive analysis in Eurasian boreal forests, which are expected
to become more fire prone with climate warming and drying [1,53,54]. This study conducts a
comprehensive analysis of within-patch fire severity variations in response to pre-fire vegetation,
topography, and surface moisture availability at fine spatial scales in a Eurasian boreal forest landscape.
The Great Xing’an boreal forest in northeastern China is an important forest ecosystem that stores
1.0–1.5 Pg C and provides 30% of the total timber yield in China [55]. It is located near the southern
frontier of Eurasian boreal forests where the fire regime is very sensitive to climate changes [43,54,56].
Since April 2014, commercial logging has been completely forbidden in this region in an effort to restore
and protect its valuable ecosystem services. Fire is the primary disturbance in this area, and there
is an increasing urgency to understand the driving mechanisms of fire severity patterns to mitigate
fire-induced ecological damage. In this paper, our objectives are three-fold. First, we investigate
how fire severity varies across the landscape in response to environmental factors characterized at
a consistent spatial resolution. Second, we verify the predictive power of remotely sensed pre-fire
surface moisture conditions on determining fire severity patterns. Third, we develop an empirically
based model to identify the distribution of potential high-severity burns.

2. Materials and Methods

2.1. Study Area

The boreal forest in the Great Xing’an Mountains of China is a fire-prone ecosystem that generally
experiences frequent, moderate- to low-severity surface burns, mixed with infrequent high-severity
crown fires. The climate is classified as mid-latitude continental cold-temperate type with short,
warm, humid summers and long, cold, dry winters [28]. The study area is mainly located in the
Huzhong Forestry Bureau (Figure 1), which is situated in the central part of Great Xing’an region and
represents a typical boreal forest landscape of northeastern China. It has a mean annual precipitation
of approximately 460 mm that mostly occurs between July and September, and a mean annual
temperature of approximately −4.7 ◦ C. The topography is mountainous, with elevations ranging
between 360 m and 1511 m above sea level. In contrast to the boreal forests dominated by evergreen
coniferous tree species in North America and Europe, the forests in this area are dominated primarily
by a deciduous coniferous tree species, Dahurian larch (Larix gmelinii (Rupr.) Rupr.), and mixed
with some evergreen coniferous tree/shrub species including Korean spruce (Picea koraiensis Nakai),
Scotch pine (Pinus sylvestris var. mongolica), and Siberian dwarf pine (Pinus pumila (Pall.) Regel),
and a few deciduous broadleaf species of birch (Betula platyphylla) and aspen (Populus davidiana and
Populus suaveolens Fisch.). The understory species are composed of evergreen shrubs (e.g., Ledum L.
and Vaccinium vitis-idaea L.), deciduous shrubs (e.g., Betula fruticose Pall. and Rhododendron dauricum
L.), and some herbaceous plants (e.g., Chamaenerion angustifolium (L.) and Carex appendiculata (Trautv.)
Kukenth.) [28], whose distributions are influenced by the topographic and soil conditions [57].
Forests 2018, 9, 130 4 of 26

Based on fire occurrence records published by the Chinese Forestry Science Data Center, there were
146 fires in the Huzhong Forestry Bureau’s jurisdiction between 1991 and 2010, of which 111 were
lightning-ignited
Forests 2018, 9, x FOR [58].
PEER Most
REVIEW fires occurred in June, July, and August, which suggests that summer4isofthe 26
primary fire season in our study area. Similar fire occurrence tendencies for the entire Great Xing’an
Great
MountainsXing’an Mountains
were reportedwere fromreported from in
1980 to 2005 1980
Fantoet2005 in Fan [54].
al. (2017) et al. Here
(2017)we[54]. Here we
focused on focused
21 fires
on 21 fires
(Table (Table 1)in
1) occurring occurring
two fire in two 2000
years: fire years: 2000
(3 fires) (3 2010
and fires)(18
andfires).
2010 These
(18 fires).
twoThese two exhibited
fire years fire years
exhibited
the greatest theburned
greatest burned
areas of any areas of any
other yearsother
in theyears
pastintwo
thedecades,
past twotogether
decades,accounting
together accounting
for about
for about 82.5% of the total burned area over a 20-year period [28,43]. Another
82.5% of the total burned area over a 20-year period [28,43]. Another important reason for selecting important reason for
selecting
these firesthese fires
is that is that
they weretheyall were all lighting-ignited
lighting-ignited in mid-to-late
in mid-to-late June and June and located
located withinwithin
similarsimilar
biotic
biotic and environments.
and abiotic abiotic environments.
In addition,In field
addition, field measurements
measurements of fire severityof and
fire forest
severity and forest
regeneration in
regeneration in the area of these fires have been conducted by our research
the area of these fires have been conducted by our research team since 2010 [28,59,60]. team since 2010 [28,59,60].

Figure 1. Location of study area (a) showing severity

severity of
of wildfires occurring
occurring inin 2000
2000 (blue
(blue perimeters)
perimeters)
and 2010 (pink perimeters). Most
Most of
of the
the study
study area is located in the Huzhong Forestry Bureau (b) in
the middle
middle of
of Great
GreatXing’an
Xing’anboreal
borealforests
forests(green
(greenpatch
patchininb band
andc)c)
which
which administratively
administrativelybelongs to
belongs
Heilongjiang province
to Heilongjiang in northeastern
province China
in northeastern (c). One
China (c). fire
Onewasfirelocated in E’lunchun
was located CountyCounty
in E’lunchun (a), which
(a),
belongs
which to the to
belongs Inner Mongolian
the Inner part part
Mongolian of Great Xing’an
of Great Xing’anboreal forests.
boreal forests.Forests
Forestswithin
withinthe
the Huzhong
Huzhong
Natural Reserve
Reserve are primarily natural
natural forests because
because of a strictly enforced ban on commercial and
salvage logging within the reserve since 1958, while forests outside the natural reserve experienced
severe cutting since the 1950s.

Table 1. Detailed information of 21 fires included in this study.

Occurrence Date DOY † Duration (Day) Longitude Latitude Burned Area (ha)
17 June 2000 168 7 122.830 51.891 8518.5
17 June 2000 168 5 123.175 51.314 2918.3
18 June 2000 169 3 123.294 51.724 1443.6
12 June 2010 163 1 123.092 52.003 207.4
12 June 2010 163 1 122.947 51.420 320.1
13 June 2010 164 1 122.844 52.036 394.8
15 June 2010 166 1 122.821 51.813 26.3
15 June 2010 166 1 123.579 51.583 29.0
15 June 2010 166 1 123.587 51.559 104.5
Forests 2018, 9, 130 5 of 26

Table 1. Detailed information of 21 fires included in this study.

Occurrence Date DOY † Duration (Day) Longitude Latitude Burned Area (ha)
17 June 2000 168 7 122.830 51.891 8518.5
17 June 2000 168 5 123.175 51.314 2918.3
18 June 2000 169 3 123.294 51.724 1443.6
12 June 2010 163 1 123.092 52.003 207.4
12 June 2010 163 1 122.947 51.420 320.1
13 June 2010 164 1 122.844 52.036 394.8
15 June 2010 166 1 122.821 51.813 26.3
15 June 2010 166 1 123.579 51.583 29.0
15 June 2010 166 1 123.587 51.559 104.5
20 June 2010 171 1 122.908 52.027 47.0
25 June 2010 176 1 123.513 51.577 17.4
26 June 2010 177 5 123.486 51.305 2891.5
26 June 2010 177 5 123.252 51.472 1926.1
27 June 2010 178 1 123.116 51.300 102.4
27 June 2010 178 3 123.182 51.431 255.1
27 June 2010 178 3 123.224 51.390 734.6
27 June 2010 178 3 123.108 51.435 258.8
28 June 2010 179 1 123.302 51.450 260.4
28 June 2010 179 1 122.784 51.459 536.0
28 June 2010 179 3 123.065 51.391 984.3
29 June 2010 180 1 122.922 51.879 670.8
† DOY: day of year corresponding to fire occurrence date.

2.2. Remote Sensing Imagery Processing

We obtained four L-1 terrain-corrected Landsat TM and ETM+ images (path-row 121/24) with
very good quality from 1999, 2000, 2007 and 2010 from the USGS website. To minimize spectral bias
caused by phenology differences, the four Landsat images selected for study were all acquired in
September. The raw digital number (DN) images of each spectral band were first calibrated into
at-satellite radiance using the sensor-specific parameters cited in Chander et al. [61]. A consistent
radiometric response between multitemporal Landsat images is critically important for regional fire
severity assessment over long time scales. Atmospherically-corrected Landsat surface reflectance
products recently have been provided by the USGS, but to our knowledge, a further radiometric
normalization can still be useful for consistent monitoring of forest changes. Thus, before using these
Landsat images, we applied the 6S atmospheric correction method [62] and an absolute radiometric
normalization approach, the iteratively reweighted multivariate alteration detection (IR-MAD) [63],
to eliminate atmospheric effects and improve radiometric consistency between the Landsat time-series
datasets. We selected a 2802 × 3483-pixel subset from each image (Figure 2), which covers all 21 fires,
to carry out the normalization procedure in ENVI/IDL 4.7 (ITT Industries Inc., White Plains, NY,
USA, 2009). We selected the 2007 image as the common reference image due to the least cloud
coverage and used the bandwise regression parameters generated by IR-MAD to normalize the other
three reflectance images. All the images were normalized to a consistent radiometric standard of the
2007 reference image. The dNBR is a well-known spectral index and has been proven to correlate well
with fire severity in many types of forest ecosystems, including the study area of this research [28].
We calculated the NBR and dNBR indices using Landsat bands 4 and 7 as proposed in Key and
Benson [19]:
TM4 − TM7
NBR = (1)
TM4 + TM7
dNBR = NBRpre − NBRpost (2)
Forests 2018, 9, 130 6 of 26

The two images acquired in 1999 and 2007 were used to calculate pre-fie NBR for the years
of 2000 and 2010, respectively, while the images acquired in 2000 and 2010 were used to calculate
post-fire NBR.
The 8-day moderate resolution imaging spectroradiometer (MODIS) ET product (MOD16A2)
includes actual ET (AET), latent heat flux, potential ET (PET), and potential latent heat flux data. It is
composed of daily canopy evaporation, plant transpiration, and soil evaporation and is calculated
as the average value of cloud-free ET during an 8-day period [64]. MOD16A2 products have
been evaluated based on flux tower measurements in many ecosystems and exhibit agreement
with the ground-measured ET in eastern Asian forests [64,65]. In this study, we assumed that the
MODIS-derived SMA index captures the broad relationship between remote sensing spectral signals
and fuel moisture content. Although we did not develop an empirical relationship model to retrieve
the actual pre-fire
Forests 2018, fuel
9, x FOR PEERmoisture
REVIEW content, many recent publications have tested this hypothesis 6 of 26 and
indicated that MODIS data could be operationally integrated into fire danger systems [48,49,66].
that MODIS
Currently, there aredatatwocould be operationally
versions of MOD16A2 integrated into available.
products fire dangerThe systems
latest[48,49,66].
(version Currently,
6) can provide
there are two versions of MOD16A2 products available. The latest
500 m ET observations, but it is not available for year 2000 as its collection began (version 6) can provide
in 2001.500Version
m ET 5 is
observations, but it is not available for year 2000 as its collection began in 2001. Version 5 is available
available for both fire years, but its spatial resolution is coarser (1 km) (Figure A1). In addition, because
for both fire years, but its spatial resolution is coarser (1 km) (Figure A1). In addition, because these
these two versions used different model input datasets (e.g., land cover product, leaf area index) and
two versions used different model input datasets (e.g., land cover product, leaf area index) and
algorithms
algorithms[67],[67],
thethe
final ETET
final outputs
outputsdiffer
differ between version5 5and
between version and version
version 6 (Figure
6 (Figure A2).A2). We obtained
We obtained
both both
version 5 and
version version
5 and version6 of6 of
MOD16A2
MOD16A2product
productforforour
our study areaand
study area andassessed
assessedthethe performance of
performance
bothofversions in our modeling. In our study, we mainly used the version 5 product
both versions in our modeling. In our study, we mainly used the version 5 product for a consistent for a consistent
comparison
comparison of the models
of the modelsforfor2000
2000and
and2010.
2010.We
Wealso
alsocompared modelperformance
compared model performance and
and thethe ability of
ability
of versions
versions 5 and 65 and 6 in explaining
in explaining spatial
spatial variation
variation of of fire
fire severityfor
severity forthe
the year
year 2010.
2010.

Figure
Figure 2. Pre-fire
2. Pre-fire (a) (a)
and and post-fire(b)
post-fire (b)false
false color
color Landsat
LandsatTM TMimages
images(R-TM5, G-TM4,
(R-TM5, and B-TM3)
G-TM4, of
and B-TM3) of
the study area. The light pink patches in (b) indicate fires occurred in 2000, while dark
the study area. The light pink patches in (b) indicate fires occurred in 2000, while dark red patches red patches
indicate fires occurred in 2010.
indicate fires occurred in 2010.

Based on occurrence dates of 21 fires (Table 1), we obtained five pre-fire MOD16A2 ET data (tile:
Based on
H25V03, occurrence
day of year fromdates
129 toof168)
21 for
fires
2000(Table 1),1 we
and six obtained
km pre-fire five pre-fire
MOD16A2 MOD16A2
ET data ET data
(tile: H25V03,
(tile: day
H25V03,
of year day
fromof129year from
to 176) for 129
2010to 168)
from thefor 2000 and
University six 1 km Numerical
of Montana’s pre-fire MOD16A2
Terra-dynamic ET data
(tile: Simulation
H25V03, group.
day ofThe
year sixfrom
500 m129MOD16A2
to 176) ET fordata
2010(tile:
fromH25V03, day of year of
the University from 129 to 176)Numerical
Montana’s were
obtained from
Terra-dynamic the USGS group.
Simulation Land Processes
The six Distributed
500 m MOD16A2 Active Archive
ET dataCenter. With the day
(tile: H25V03, assistance of from
of year
MOD16A2 Quality Control (QC) data, we only selected high-quality pixels (QC
129 to 176) were obtained from the USGS Land Processes Distributed Active Archive Center. With the equal 0) from each
8-day ET dataset and used them to generate two integrated pre-fire ET datasets for 2000 and 2010.
For a given burned pixel, the maximum of six observations with high quality were ranked by the
time since fire occurrence date. We gave preference to the observation whose acquisition time was
closest to the fire occurrence date to ensure the selection reflects the latest pre-fire surface ET
conditions.
Forests 2018, 9, 130 7 of 26

assistance of MOD16A2 Quality Control (QC) data, we only selected high-quality pixels (QC equal
0) from each 8-day ET dataset and used them to generate two integrated pre-fire ET datasets for
2000 and 2010. For a given burned pixel, the maximum of six observations with high quality were
ranked by the time since fire occurrence date. We gave preference to the observation whose acquisition
time was closest to the fire occurrence date to ensure the selection reflects the latest pre-fire surface
ET conditions.

2.3. Fire Severity Mapping

Burned pixels (30 m) were extracted based on a thresholding process for the forest disturbance
index following the protocol of the Landsat Ecosystem Disturbance Adaptive Processing System
(LEDAPS) [68]. The detailed description of this approach can be found in Fang et al. (2015) [43].
Twenty-one fires burned about 22,650 ha in total; the three fire events in 2000 burned over 12,880 ha
(about 251, 670 Landsat pixels), and the 18 fires in 2010 burned over 9760 ha (about 108, 450 Landsat
pixels). These burned pixels exhibited significantly different spectral features compared to the
unburned pixels in the Landsat images, as shown in Figure 2. Our previous study established a
quadratic polynomial relationship between the dNBR and field measured composite burn index (CBI):

CBI = −0.0425 + 2.753 × dNBR − 0.8142 × dNBR2 (3)

which was confirmed to explain 84.6% variance in 74 CBI plots of 2010 fires and to produce accurate
severity maps of the largest fire in 2000 (Kappa = 0.72) [28]. The CBI values of 1.1 and 2.0 were
applied as boundaries of three severity levels (i.e., low severity (0.1 ≤ CBI ≤ 1.1), moderate severity
(1.1 < CBI ≤ 2), and high severity (2 < CBI ≤ 3)) as they were corresponding to about 10% and 80% of
canopy mortality, respectively, in our ecosystem, where the live mature trees were very important for
vegetation restoration [28]. The two dNBR thresholds associated with CBI values of 1.1 and 2.0 were
0.484 and 1.099, respectively; these thresholds were applied to classify burned pixels of 2000 and 2010
into three severity levels.

2.4. Environmental Metrics

We used a suite of explanatory variables to describe various aspects of fuel, topography,
and surface moisture conditions, and modeled their relationships with fire severity (Table 2).
Using Landsat imagery, we developed two vegetation cover maps for two different years (1999 and
2007) and reconstructed the pre-fire vegetation conditions. Using a stratified decision tree classification
method, we combined the conspicuous differences of phenology and spectral characteristics among
vegetation types, and used Landsat surface reflectance, spectral indices, and components of Tasseled
Cap transformation to produce vegetation cover maps at 30 m resolution [69]. The whole area was
classified into six vegetated categories consisting of deciduous coniferous forest (DCF, i.e., larch forest),
deciduous broadleaf forest (DBF), evergreen coniferous forest (ECF), mixed forest (MF), grassland
(GRS), and shrublands (SRB), as well as five non-vegetation categories (water bodies, bare rock,
bare soil, urban land, and shade from mountains and clouds). Accuracy assessments were carried out
in Fang et al. (2015) [43], which reported that the overall accuracy and the Kappa coefficient of the
vegetated areas were 64% and 57% respectively. In detail, the mapping accuracy of DCF is 72%, DBF is
40%, ECF is 81%, MF is 49%, GRS is 82%, and SRB is 70%, indicating suitability for subsequent analysis.
We used a 30 m ASTER global digital elevation model (GDEM) product published by the
National Aeronautics and Space Administration (NASA) to derive a number of commonly used
topographic indices, including elevation (ELV), slope (SLP), aspect, and topographic wetness index
(TWI). Aspect was further converted to better represent the potential influence of solar radiation (PSR)
on site moisture conditions, following the equation:

PSR = cos((θ − 225) × π/180) (4)

Forests 2018, 9, 130 8 of 26

where θ is the aspect in degrees [70]. The higher PSR values represent higher potential solar insolation.
The TWI was designed based on the assumption that water movement is controlled by the topography
of the slopes [71]. TWI is defined as:

TWI = ln(α/ tan β) (5)

where α is the local upslope area draining through a certain point per contour length and tanβ is the
slope angle at the point [72]. TWI is often used to describe the spatial distribution of soil moisture and
surface saturation, and has shown good correlation with soil moisture and depth to ground water in
Eurasian boreal forests [72,73]. TWI values typically range from 3 to 30, where a higher value indicates
higher soil moisture potential.
Fuel moisture content is a good surrogate for representing the indirect impacts of weather
on fire behavior [74] because it influences many fire processes, such as ignition, combustion,
and smoldering [75]. Fuel moisture is also controlled by weather conditions such as precipitation
and ET and is closely associated with soil moisture [48,75]. Here, we calculated surface moisture
availability (SMA) as the indicator of live fuel moisture content using the equation:

SMA = AET/PET (6)

where AET is actual ET and PET is the potential ET [76]. The SMA is a critical parameter in governing
the partition between sensible and latent heat flux at the surface, which is the key determinant of soil
moisture, and also determines the water stress and water content of live plants [49,76]. In general,
when AET is equal to PET, the surface will reach moisture-saturated conditions and the SMA will be
equal to 1. However, when the SMA is below certain threshold values, the vegetation may begin to
suffer from drought stress and the soil may approach a limiting dryness [76].

Table 2. Abbreviations and descriptions of explanatory variables in this study.

Category Variable Description Mean ± SD (2000 & 2010)

Percentage of Landsat pixels classified into evergreen coniferous
0.158 ± 0.261
ECF trees within 240 m burned pixels, were primarily Pinus pumila
0.169 ± 0.278
shrublands and Larix gmelinii-Pinus pumila forest.
Percentage of larch forest. The three dominant larch forests are
0.615 ± 0.324
DCF Larix gmelinii-Ledum palustre L., Larix gmelinii-grass, and Larix
0.350 ± 0.302
gmelinii-Rhododendron dahurica L.
Vegetation † Percentage of broad leaf forest. The white birch and aspen are 0.017 ± 0.057
DBF
dominant broad leaf species. 0.177 ± 0.218
Percentage of mixed forest. Composited by broad leaf trees and 0.086 ± 0.145
MF
coniferous trees. 0.150 ± 0.218
0.047 ± 0.135
GRS Percentage of grassland.
0.086 ± 0.166
Percentage of shrublands, typically distributed in open land 0.035 ± 0.097
SRB
along the river and disturbed areas. 0.041 ± 0.134
Elevation (meters) derived from the aggregated ASTER GDEM 1089 ± 92.862
ELV
at 240 m spatial resolution. 1017 ± 107.316
Potential solar radiation. It ranges from −1 to 1, where high −0.024 ± 0.715
PSR
values represent xeric exposures. −0.173 ± 0.683
Topography
8.840 ± 4.090
SLP Slope (degree) computed from aggregated DEM.
8.958 ± 3.992
Topographic wetness index (unitless) is computed from the 13.953 ± 1.313
TWI
slope and the upslope contributing area per unit contour length. 14.024 ± 1.528
SMA is calculated from MOD16A2 and represents land surface
Surface 0.494 ± 0.099
SMA moisture availability. The higher SMA value indicates the
moisture 0.411 ± 0.153
wetter land surface.
† Statistics of five non-vegetation land cover types were not listed.
Forests 2018, 9, 130 9 of 26

2.5. Spatial Data Processing

To mitigate the issue of inconsistent spatial resolutions between various spatial datasets,
we conservatively selected a 240 m spatial resolution as the overall standard. This resolution (240 m) is
exactly eightfold the resolution (30 m) of the Landsat-derived vegetation maps and the ASTER GDEM
data. This resolution is particularly suitable for aggregating the fine-resolution categorical map to
the coarse-resolution continuous imagery. In addition, it is close to 250 m, which is the finest spatial
resolution of MODIS products, and many remote sensing studies use 250 m to conduct downscaling for
1 km MODIS ET products [77]. To downscale the SMA data, we used a nearest neighbor interpolation
approach to resample the data from 1 km to 240 m.
Three different spatial aggregation methods were employed to upscale the vegetation,
topographic, and fire severity data. Based on the Landsat-derived vegetation cover maps, we calculated
the fractional coverage of each vegetation type by calculating the proportion of occurrence within
an 8 × 8 pixel window (8-PW). Thus, the 30 m categorical vegetation maps were converted into
240 m continuous vegetation coverage images. For the numerical ASTER GDEM data, we calculated
an average value for each 8-PW and subsequently used the aggregated DEM data to calculate the
abovementioned topographic metrics.
To generate a fire severity value that can be consistently interpreted at different spatial scales,
we used an area-weighted average method to aggregate the 30 m categorical fire severity map to
240 m. We first analyzed the proportions of unburned, low, moderate, and high severity pixels within
each 8-PW; pixel proportions in the 8-PW serve as the area-weight values in the aggregation function.
For instance, if 32 pixels within a given 8-PW (64 pixels) were classified as high severity, the weight
value of high severity would be 50% (i.e., 32/64). If no pixel was classified as high severity in the 8-PW,
the weight value of high severity would be 0. Then, a set of fire severity rating integers from 0 to 3 was
used to indicate the fire severity gradient from unburned to high severity, respectively:

Severity = 0 × ωunburned + 1 × ωlow + 2 × ωmoderate + 3 × ωhigh (7)

where ω is the weight value of four severity gradients. The area-weighted average value of fire
severity ranged from 0 to 3, with higher values representing higher fire severity. The area-weighted
averaging process considers the relative importance of each severity level and provides a more balanced
interpretation of the data.

2.6. Statistical Modeling

A boosted regression tree (BRT) model was applied to explain the relationships between remote
sensed fire severity and selected environmental variables. The BRT model is a useful machine
learning approach that uses recursive binary splits and a boosting technique to combine a large
number of sequential trees to improve the fit and predictive performance [78,79]. It has the capacity
to handle complex relationships among numerical and categorical variables, quantify interactions
between explanatory variables and overcome inaccuracies associated with regression and classification
methods [78]. We applied the “gbm” package (version 2.1.3) in R 3.4.1 (R Development Core Team 2017,
Boston, MA, USA) to run the BRT analysis. We ran two sets of BRT models for the two different fire
years, 2000 and 2010. We used the same parameter settings and amounts of training data to ensure that
the outcomes of the two models were comparable. The random subsampling and bagging procedures
in a BRT model may introduce stochasticity into the model outcomes. To mitigate the stochastic errors
and create stable model outputs, we carried out 50 BRT modeling trials independently and calculated
an average as the final result.
Before running the BRT models, we carried out a data exploration procedure to assess whether
there was any conspicuous spatial autocorrelation or collinearity. Using functions in the “ape”
package [80], we calculated the Moran’s I index to examine spatial autocorrelation in fire severity and
explanatory variables. Moran’s I is similar to a correlation coefficient and represents the similarity of
Forests 2018, 9, 130 10 of 26

an observation to its nearby observations [81]. It ranges between −1 and 1. Higher positive values
indicate greater similarity, which can be interpreted as a spatially clustered distribution, while lower
negative values indicate stronger dissimilarity (i.e., a more dispersed spatial pattern), and the zero
value indicates a random spatial pattern (i.e., perfect randomness). We set a 300 m sample spacing
distance (>1 pixel) when computing the Moran’s I index. Although p-values were less than the
0.05 level, the Moran’s I values of both response and explanatory variables were small (mostly less than
0.20, see Table A1), suggesting weak spatial autocorrelations among sampling pixels. Similar or higher
Moran’s I thresholds for determining suitable sampling spaces have been used in wildfire-related
literature [36,45,50]. We also calculated the pairwise Pearson correlation coefficient (r) using the
“Hmisc” package to evaluate potential collinearity among predictor variables. All variables selected
for modeling had low pairwise Pearson correlations (|r| < 0.60), suggesting relatively low levels
of collinearity. We used a subset of 200 random sampling pixels to build the BRT model, while the
remaining 100 pixels were used for validation.
The optimization of a BRT model is jointly controlled by the learning rate, tree complexity, bagging
fraction, and number of trees. In this study, we set these parameters at 0.01, 5 and 0.5, respectively,
following the recommended model inputs of Elith et al. (2008) [79]. The number of trees in each BRT
trial was automatically selected based on a 5-fold cross-validation procedure to avoid over-fitting
problem. The relative importance of each explanatory variable to fire severity was measured by
averaging the frequency it was selected for splitting among all trees, and the importance was weighted
by the squared improvement to the model as a result of each split [79]. The relative importance values
of the predictors were scaled as a percentage, the sum of which equals 100%. Higher importance values
represented stronger impacts on controlling fire severity. In addition, we explored the dependency
relationships between several important variables and fire severity by plotting the effect of a specific
explanatory variable on the response variable after averaging the effects of remaining explanatory
variables in the same model.
The coefficient of determination (R2 ) reported by the BRT model was used to evaluate how well
the model fits the training data. We used the 100 samples that were independent of the training
data to assess the predictive power of our BRT models. We examined the goodness of fit between
simulated fire severity values and observed fire severity values by applying a linear regression model
and calculating the squared multiple correlation coefficients (denoted r2 ). To provide a standalone
evaluation of model performance, we used the error matrix method to evaluate the predicted fire
severity maps. We selected an optimal simulation image for each fire year based on the r2 value and
classified the burned pixels into low (fire severity ≤ 1), moderate (1 < fire severity ≤ 2), and high
(2 < fire severity ≤ 3) severity levels. The two aggregated fire severity images were also converted into
categorical severity maps using these same thresholds and were then used as reference maps. We used
all burned pixels from each fire severity level to ensure unbiased evaluation for both fire years.

3. Results

3.1. Evaluation of Model Performance

We found that the two sets of BRT models fit the training data very well, as they explained 83.0%
(R2 max = 85.0%, SD = 0.008) of the variation in fire severity in 2000 (Figure 3a) and 73.8% (R2 max = 81.3%,
SD = 0.035) of the variation in fire severity in 2010 (Figure 3b). Similarly, the 50 validation models
of the two years showed goodness of fits of 0.795 (r2 max = 0.820, SD = 0.040) (Figure 3c) and 0.618
(r2 max = 0.656, SD = 0.012) (Figure 3d), respectively, suggesting that the predicted fire severity results
were well correlated with the reference values.
The confusion matrices indicated that the optimal BRT model in 2000 achieved higher prediction
accuracy than the optimal model in 2010 (Table 3). The predicted severity map of 2000 produced fewer
commission errors in moderate (44.3% vs. 71.6%) and high (11.2% vs. 24.0%) severity pixels than 2010,
but it generated a high commission error for low severity (85.8%) pixels, indicating more pixels were
Forests 2018, 9, 130 11 of 26

erroneously predicted as low severity than were actually observed in the severity results. The predicted
severity map of 2010 generated high commission error (71.6%) for moderate severity level and high
omission error (64.6%) for low severity level, indicating overestimated moderate severity area and
underestimated low and high severity area (Figure A3). Overall, we found the prediction map of 2000
to represent very good agreement with the observed severity map, while the prediction map of 2010
Forests 2018, 9, x FOR PEER REVIEW 11 of 26
underestimated the severity of most of the 18 fires (Figure 4).
3.0

3.0
a b
2.5

2.5
Observed Fire Severity

Observed Fire Severity

2.0

2.0
1.5

1.5
1.0

1.0
0.5

0.5
R^2 = 0.85, p < 0.05 R^2 = 0.813, p < 0.05
0.0

0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0

Fitted Fire Severity Fitted Fire Severity

3.0

3.0
c d
2.5

2.5
Observed Fire Severity

Observed Fire Severity

2.0

2.0
1.5

1.5
1.0

1.0
0.5

0.5

r^2 = 0.82, p < 0.05 r^2 = 0.656, p < 0.05

0.0

0.0

0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
Predicted Fire Severity Predicted Fire Severity

FigureFigure
3. Two 3. Two examples
examples of linear
of linear relationshipsused
relationships used for
for validating
validating model
model performance
performance of 2000 (a,c) (a,c)
of 2000
and 2010
and 2010 (b,d)(b,d) based
based on on
200200 trainingsamples
training samples (a,b)
(a,b) and
and100
100verification samples
verification (c,d).(c,d).
samples Coefficients of
Coefficients
determination (R
2 2 for training samples), and the squared multiple correlation coefficients (r2 for2
of determination (R for training samples), and the squared multiple correlation coefficients (r for
verification samples) are also plotted. Blue solid lines show predicted linear regression fit. Black
verification samples) are also plotted. Blue solid lines show predicted linear regression fit. Black dashed
dashed lines represent 1:1 line.
lines represent 1:1 line.
Table 3. Accuracy assessment of predicted fire severity classification for fire year 2000 and 2010,
3. Accuracy assessment of predicted fire severity classification for fire year 2000 and
Table respectively.
2010, respectively.
Predicted Severity of 2000 Predicted Severity of 2010
Producer’s Producer’s
Severity Class Low Predicted
ModerateSeverity
Highof 2000 Low Predicted Severity
Moderate High of 2010
Accuracy Accuracy
LowClass 50 26 7 Producer’s
60.2% 196 263 94 Producer’s
35.4%
Severity Low Moderate High Low Moderate High
Accuracy Accuracy
Moderate 261 496 142 55.2% 95 276 115 56.8%
Low
High 50
42 26
369 7
1184 60.2%
74.2% 196
33 263
431 94662 35.4%
58.8%
Moderate
User’s Accuracy 261
14.2% 496
55.7% 142
88.8% 55.2%- 95
60.5% 276
28.5% 115
76.0% 56.8%-
High 42 369 1184 74.2% 33 431 662 58.8%
Overall
67.1% 52.4%
User’s Accuracy
Accuracy 14.2% 55.7% 88.8% - 60.5% 28.5% 76.0% -
Overall Accuracy 67.1% 52.4%
3.2. Relative Importance of Environmental Variables
The
3.2. Relative relative importance
Importance of individual
of Environmental Variablesenvironmental variables varied substantially in the
different fire years (Figure 5). For the fires in 2000, the six most important predictors of fire severity
The relative importance
in decreasing order were of individual
DBF, SRB, MF,environmental variables
TWI, SLP and ELV. Thesevaried substantially
predictors in to
contributed thea different
total
fire years (Figure 5). For the fires in 2000, the six most important predictors of fire severity in decreasing
of 76.8% of the relative importance. Three vegetation variables together contributed over 55% relative
order importance,
were DBF, SRB, MF, TWI,
indicating SLP and
the spatial ELV. These
pattern of firepredictors
severity incontributed to a total
2000 was largely of 76.8%
driven by theof the
relative importance.
distribution Three
of these vegetation
three vegetationvariables
types. Fortogether
the 2010contributed
burns, the sixover
most55% relativepredictors
important importance,
were ECF, SMA, SLP, PSR, ELV and TWI, which together contributed to 75.1% of the relative
importance. The ECF and SMA variables independently contributed as much as 23.3% and 12.8%
relative importance respectively. The 500 m version 6 SMA was found to have a slightly lower (11.0%)
Forests 2018, 9, 130 12 of 26

indicating the spatial pattern of fire severity in 2000 was largely driven by the distribution of these
three vegetation types. For the 2010 burns, the six most important predictors were ECF, SMA, SLP,
PSR, ELV and TWI, which together contributed to 75.1% of the relative importance. The ECF and SMA
variables independently contributed as much as 23.3% and 12.8% relative importance respectively.
The 500 m version 6 SMA was found to have a slightly lower (11.0%) relative importance than the
Forests 2018, 9, x FOR PEER REVIEW 12 of 26
1 km version 5 SMA product when modeling spatial variability of fire severity in 2010 (Figure A4).
An overall ranking of 11 variables
relative importance than thewas
1 kmcalculated
version 5 SMA based onwhen
product a weighted average
modeling spatial approach
variability of fire(see y-axis of
severity in 2010 (Figure A4). An overall ranking of 11 variables was calculated
Figure 5), which identified four vegetation variables (ECF, DBF, SRB, and MF) as the primary controls based on a weighted
average
Forests approach
2018, 9, x FOR(see
PEERy-axis
REVIEWof Figure 5), which identified four vegetation variables (ECF, DBF, SRB,
12 of 26
of fire severity,
andfollowed by SMA
MF) as the primary and oftopographic
controls variables.
fire severity, followed by SMA and topographic variables.
relative importance than the 1 km version 5 SMA product when modeling spatial variability of fire
severity in 2010 (Figure A4). An overall ranking of 11 variables was calculated based on a weighted
average approach (see y-axis of Figure 5), which identified four vegetation variables (ECF, DBF, SRB,
and MF) as the primary controls of fire severity, followed by SMA and topographic variables.

Figure 4. Observed fire severity (a) aggregated from Landsat observations versus the modeled fire
Figure 4. Observed fire severity (a) aggregated from Landsat observations versus the modeled fire
severity (b) for 21 fires based on boosted regression tree models. All fire severity images were plotted
severity (b) for 21mfires
at 240
Figure based
spatial on
resolution.
4. Observed boosted
fire severity regression
(a) aggregated tree
from models.
Landsat All fire
observations versusseverity images
the modeled fire were plotted
at 240 m spatialseverity (b) for 21 fires based on boosted regression tree models. All fire severity images were plotted
resolution.
at 240 m spatial resolution.
GRS
RIMP2010
DCF RIMP2000
GRS
RIMP2010
PSR
DCF RIMP2000
TWI
PSR

ELV
TWI

SLP
ELV

SLP
SMA
SMA
MF

SRBMF
SRB
DBF
DBF
ECF
ECF
0 5 10 15 20 25 30 35 40
0 5 10 15
Relative 20
importance 25
(%) 30 35 40
Relative importance (%)
Figure 5. Relative importance proportion (RIMP) of explanatory variables for 50 boosted regression
Figure 5. Relative importance proportion (RIMP) of explanatory variables for 50 boosted regression
Figure 5. tree
Relativemodels (Mean + SD)proportion
importance of fire severity(RIMP)
in 2000 and 2010.
of2010. The y-axis (ECF
explanatory to GRS) for
variables shows
50an overall
boosted
regression
tree models (Mean + SD) of fire severity in 2000 and The y-axis (ECF to GRS) shows an overall
rank of 11 variables in descending order according to a weighted average of RIMP, which is calculated
rank of+11SD)
tree models (Mean variables in descending
of fire severity order
in according
2000 and to a2010.
weighted average
The of RIMP,
y-axis (ECFwhich
toisGRS)
calculated
shows an overall
rank of 11 variables in descending order according to a weighted average of RIMP, which is calculated
by multiplying two mean R2 (i.e., 0.830 for RIMPs of 2000, 0.738 for RIMPs 2010) with two mean
RIMP values (i.e., RIMP2000 and RIMP2010) of each variable. The ECF has the highest overall relative
importance while the GRS has the lowest value. See Table 2 for definition of variable abbreviations.
Forests 2018, 9, x FOR PEER REVIEW 13 of 26

by multiplying two mean R2 (i.e., 0.830 for RIMPs of 2000, 0.738 for RIMPs 2010) with two mean RIMP
values (i.e., RIMP2000 and RIMP2010) of each variable. The ECF has the highest overall relative
Forests importance
2018, 9, 130 while the GRS has the lowest value. See Table 2 for definition of variable abbreviations.13 of 26

When the 11 variables were grouped into three broad control types of vegetation, topography, or
SMA, When the 11 variables
the relative importance were grouped
of these into
three threetypes
control broadbecame
controlquite
typessimilar
of vegetation,
betweentopography,
the two fire
or
years. Vegetation consistently played a strong role in determining fire severity between
SMA, the relative importance of these three control types became quite similar in the twothefire
twoyears
fire
years.
becauseVegetation consistently
it contributed playedofatotal
about 40–80% strong role in
relative determining
importance fire severity
(Figure in thewith
6a). Together two 20–50%
fire yearsof
because it contributed about 40–80% of total relative importance (Figure 6a). Together with
the total relative importance contributed by topographic variables, the results revealed that potential 20–50% of
the
firetotal relative
severity importance
in our study areacontributed by topographic
is mostly controlled variables,
by vegetation andthe results revealed
topography. that potential
At the same time, we
fire severity
found SMA has in our study
similar area isimportance
relative mostly controlled by vegetation
to the maximum and topography.
importance At the same
value of topographic time,
variables,
we found SMA has similar relative importance to the maximum importance
indicating that SMA also has considerable predictive power over fire severity (Figure 6b). value of topographic
variables, indicating that SMA also has considerable predictive power over fire severity (Figure 6b).

Figure 6. Ternary plot (a) showing the total relative importance proportion (RIMP) of three groups
Figure 6. Ternary plot (a) showing the total relative importance proportion (RIMP) of three groups of
of explanatory variables (i.e., Vegetation, Topography and Surface Moisture Availability (SMA)) in
explanatory variables (i.e., Vegetation, Topography and Surface Moisture Availability (SMA)) in
controlling fire severity of 2000 (blue) and 2010 (red). In considering the difference of variable numbers
controlling fire severity of 2000 (blue) and 2010 (red). In considering the difference of variable
among three groups, a variable which has the maximum RIMP (MaxRIMP) of each group is selected
numbers among three groups, a variable which has the maximum RIMP (MaxRIMP) of each group is
and compared (b). Axis in ternary plot (b) representing relative proportions of MaxRIMP for three
selected and compared (b). Axis in ternary plot (b) representing relative proportions of MaxRIMP for
types of variables.
three types of variables.

3.3.
3.3. Relationships
Relationships between
between Environmental
Environmental Variables
Variablesand
andFire
FireSeverity
Severity
Partial
Partial dependence
dependence plots plots were
were helpful
helpful for
for visualizing
visualizing the the response
response of of fire
fire severity
severity to
to explanatory
explanatory
variables.
variables. We found that the relationships between fire severity and environmental variables were
We found that the relationships between fire severity and environmental variables were
mostly
mostly nonlinear
nonlinear andand varied
varied little
little inin different
different years
years (Figure
(Figure 7).7). In
In general,
general, high
high severity
severity fires
fires were
were
more
more likely
likely toto occur
occur inin dense
dense evergreen
evergreen forests
forests(Figure
(Figure7a), 7a),especially
especially on on well-drained,
well-drained, gentle
gentle slopes
slopes
at high altitudes (Figure 7f–h). On the other hand, the partial dependence
at high altitudes (Figure 7f–h). On the other hand, the partial dependence of fire severity generallyof fire severity generally
decreased
decreased withwithincreasing
increasingDBF DBFand andSRB SRB(Figure
(Figure7b,c),
7b and suggesting that the
c), suggesting fire-resistant
that traits of
the fire-resistant these
traits of
two vegetation types may mitigate the adverse effects of severe burning.
these two vegetation types may mitigate the adverse effects of severe burning. We found that We found that increased
coverage
increasedofcoverage
mixed forests
of mixedgenerally
forests had a negative
generally had aimpact
negativeon impact
fire severity
on fire(Figure 7d),(Figure
severity especially
7d),
in the year 2000, in which MF was recognized as an important variable.
especially in the year 2000, in which MF was recognized as an important variable. Fire severity inFire severity in deciduous
coniferous
deciduousforests was significantly
coniferous lower than inlower
forests was significantly evergreen
than coniferous
in evergreen forests (Figureforests
coniferous 7j vs. Figure
(Figure7a)7j
but higher than in the other forest types. DCF did not exhibit a strong relationship
vs. 7a) but higher than in the other forest types. DCF did not exhibit a strong relationship with fire with fire severity in
2010, but in
severity did exhibit
2010, butadid positive
exhibitrelationship
a positive with fire severity
relationship within 2000.
fire The GRS
severity variable
in 2000. The was
GRSfound to
variable
have little relative
was found to haveimportance to fire
little relative severity to
importance andfirerepresented
severity and a weak negativea impact
represented on fire severity
weak negative impact
in 2010 (Figure 7k). The south facing slopes and areas with high TWI
on fire severity in 2010 (Figure 7k). The south facing slopes and areas with high TWI values values typically experienced
typically
moderate
experienced severity fires severity
moderate (Figure 7h,i). Our results
fires (Figure 7h,i).also
Ourindicated thatindicated
results also higher probabilities
that higher of moderate-
probabilities
to high-severity fires were related to canopy or surface dryness (Figure 7e), especially when the 1 km
SMA ranged from 0.13 to 0.30 in the summer. In contrast, we found a strong positive association
between fire severity and 500 m SMA (Figure A5), as a result of inconsistent SMA values.
 Forests 2018, 9, x FOR PEER REVIEW 14 of 26

of moderate- to high-severity fires were related to canopy or surface dryness (Figure 7e), especially
Forestswhen
2018, the 1 km SMA ranged from 0.13 to 0.30 in the summer. In contrast, we found a strong positive14 of 26
9, 130
association between fire severity and 500 m SMA (Figure A5), as a result of inconsistent SMA values.

Figure
Figure 7. Partial
7. Partial dependence
dependence plotsofofexplanatory
plots explanatory variables
variablesononregulating
regulatingfirefire
severity in different
severity fire fire
in different
years (2000 in red and 2010 in cyan). This grouping of variables was following the rank as in
years (2000 in red and 2010 in cyan). This grouping of variables was following the rank as shown shown
y-axis of Figure 5. The curves demonstrate the smoothed means (solid line) and 95% confidence
in y-axis of Figure 5. The curves demonstrate the smoothed means (solid line) and 95% confidence
intervals (gray zone) of an ensemble of 50 models. Variable abbreviations were described in Table 2.
intervals (gray zone) of an ensemble of 50 models. Variable abbreviations were described in Table 2.
4. Discussion
4. Discussion
4.1. Environmental Influences on Fire Severity
4.1. Environmental Influences on Fire Severity
We found that the BRT method was effective in investigating the relationships between fire
We found
severity andthat the BRT method
environmental gradients,was
sucheffective in investigating
as vegetation the relationships
composition, terrain, between fire
and surface moisture
status. Fire severity is a complex function of these environmental gradients and such
severity and environmental gradients, such as vegetation composition, terrain, and surface moisture relationships
may
status. Firevary in different
severity years and
is a complex locations.
function With
of these thousands of gradients
environmental 240 m sampling
and such data from
relationships
representative historical fires, we found that the spatial distribution of fire severity could
may vary in different years and locations. With thousands of 240 m sampling data from representative be predicted
with adequate precision in a Great Xing’an boreal forest landscape. In our exploration of the relative
historical fires, we found that the spatial distribution of fire severity could be predicted with adequate
importance of these spatial controls on fire severity, we found that fuel conditions are the most
precision in a Great Xing’an boreal forest landscape. In our exploration of the relative importance of
influential predictor in determining the magnitude of fire severity. This finding is in accordance with
theseour
spatial controls
previous studyon[43],
fire severity,
which was weconducted
found that at fuel conditions
burned are the
patch level most
for the influential
entire predictor in
Great Xing’an
determining the magnitude of fire severity. This finding is in accordance with our previous
boreal forest, and studies in similar ecosystems, such as Canadian boreal forests [82,83] and subalpine study [43],
which was conducted
forests [37,84,85]. at burned patch level for the entire Great Xing’an boreal forest, and studies in
similar ecosystems, such as Canadian boreal forests [82,83] and subalpine forests [37,84,85].
Coniferous forests/shrubs experience higher severity fires than broadleaf forests and shrublands
in our ecosystem. Such result is generally consistent with observations in North America boreal forests
where deciduous tree dominated stands are found to be fire break and reduce landscape flammability
owing to higher foliage moisture and less surface fuels [86,87]. Larix gmelinii is the dominant coniferous
tree species in the boreal forests of Northeastern China (Figure 8a). Unlike the dominant evergreen
coniferous species (e.g., black spruce) in Northern American boreal forests or subalpine forests
(e.g., spruce-fir, ponderosa pine) in Western US, larch is a deciduous coniferous species; it is considered
 Forests 2018, 9, x FOR PEER REVIEW 15 of 26

Coniferous forests/shrubs experience higher severity fires than broadleaf forests and shrublands
in our ecosystem. Such result is generally consistent with observations in North America boreal 15 of 26
Forests 2018, 9, 130
forests where deciduous tree dominated stands are found to be fire break and reduce landscape
flammability owing to higher foliage moisture and less surface fuels [86,87]. Larix gmelinii is the
dominant coniferous tree species in the boreal forests of Northeastern China (Figure 8a). Unlike the
fire-tolerant because its self-pruning trait can reduce ladder fuels, and the open crown can reduce
dominant evergreen coniferous species (e.g., black spruce) in Northern American boreal forests or
the bulk density and connectivity of canopy fuels [88]. However, we found that larch forests usually
subalpine forests (e.g., spruce-fir, ponderosa pine) in Western US, larch is a deciduous coniferous
experience highit mortality
species; in flat
is considered areas and
fire-tolerant valleyitsbottoms
because self-pruningduetrait
to heat-induced
can reduce ladder rootfuels,
damageand (Figure
the 8c).
To adapt to shallow
open crown cansoil andthe
reduce bulk densityLarix
permafrost, gmelinii develops
and connectivity of canopylateral roots
fuels [88]. in the thick,
However, flammable
we found
moss and thatorganic
larch forests
soil usually
layers experience
to improve high mortalityand
adhesion in flat areas and
nutrient valley but
supply, bottoms
this due
alsotomeans
heat- it can
induced root damage (Figure 8c). To adapt to shallow soil and permafrost,
be easily injured or killed by surface fires [43]. Larch forests are also characterized by abundant Larix gmelinii develops
lateral roots in the thick, flammable moss and organic soil layers to improve adhesion and nutrient
understory plants due to their open-canopy environment, which can contribute higher fire severity
supply, but this also means it can be easily injured or killed by surface fires [43]. Larch forests are also
than thecharacterized
forests without an abundant understory component. Another important coniferous species in
by abundant understory plants due to their open-canopy environment, which can
this region Pinus pumila
contribute higher fire (Figure
is severity 8b),
than which is evergreen
the forests without an and usually
abundant mixed with
understory Larix gmelinii
component. Another in open
forests at altitudes
important of 800–1200
coniferous species minorthis
grows densely
region is Pinuson rocky
pumila ridges
(Figure 8b),at altitudes
which higherand
is evergreen than 1200 m [89].
usually
mixed with Larix gmelinii in open forests at altitudes of 800–1200 m or
Pinus pumila is highly flammable because it contains abundant volatile organic compounds in itsgrows densely on rocky ridges
needles,attwigs,
altitudes andhigher
seedsthan 1200Furthermore,
[90]. m [89]. Pinus pumila
windy is highly
and dry flammable because
conditions canitaccelerate
contains abundant
the spreading
volatile organic compounds in its needles, twigs, and seeds [90]. Furthermore, windy and dry
of fires on the ridges of mountains. Therefore, as demonstrated in our analysis, the increases in
conditions can accelerate the spreading of fires on the ridges of mountains. Therefore, as
Pinus pumila coverage
demonstrated may
in our considerably
analysis, increase
the increases in Pinusthe probability
pumila of high
coverage may severityincrease
considerably fires (Figure
the 8d).
Such finding is in consistent with Estes et al. (2017) [38], who reported that shrub
probability of high severity fires (Figure 8d). Such finding is in consistent with Estes et al. (2017) [38], species that
are favored by firesthat
who reported canshrub
generate
specieshigher
that arefire severity
favored by firesthan mixed hardwood/coniferous
can generate forests and
higher fire severity than mixed
hardwood/coniferous forests
hardwood forests in northern California. and hardwood forests in northern California.

Figure 8. Photographs of two unburned stands dominated by the deciduous coniferous tree Larix
Figure 8. Photographs of two unburned stands dominated by the deciduous coniferous tree
gmelinii (a) and the evergreen coniferous shrub Pinus pumila (b), and two 1-year post-fire stands
Larix gmelinii (a) and
previously the evergreen
dominated coniferous
by Larix gmelinii shrub
(c, surface Pinus
fire) pumila
and Pinus (b),(d,
pumila and two 1-year
canopy post-fire stands
fire) in Huzhong
previously dominated by Larix
Natural Reserve, China. gmelinii (c, surface fire) and Pinus pumila (d, canopy fire) in Huzhong
Natural Reserve, China.

Our results indicated that topographic factors had a considerable influence on fire severity. It is
well known that topography can influence fire behavior by impacting fuel moisture, local wind
patterns, fire spread direction, and vegetation composition [32], but quantitative demonstrations of
these relationships are still needed for optimal mitigation of adverse fire effects. Our results showed
that slope and elevation are the two most important topographic variables, followed by TWI and
PSR. This suggests that the primary pathway by which topography regulates fire severity in this
Siberian boreal ecosystem is by governing fuel moisture and by strongly interacting with vegetation
conditions. For example, we found that severe fires were more likely to occur in high altitude regions.
Forests 2018, 9, 130 16 of 26

A similar pattern was found in dry ponderosa pine forests of the western US and boreal forests in
Europe [42,51,91]. Surface fuels on the upper slopes could dry quickly due to efficient drainage
and greater degrees of solar exposure, which may increase the flammability of fuels and facilitate
more severe burns. The preheating effects of upslope fires on the adjacent fuels can also increase fire
severity [32]. In addition, it cannot be ignored that the ECF are generally densely distributed at high
altitudes in this region. In general, north-facing slopes possess higher fuel moisture and lower surface
temperature than south-facing slopes. However, in our ecosystem, we found that the north-facing
slopes burned more severely due to higher biomass coverage [60]. Similar findings were also reported
in studies conducted in boreal and subalpine ecosystems [42,43,84].
Fuel moisture interacts with many complex ecological and physical processes, making it important
yet difficult to represent in a modeling framework used to study its spatiotemporal dynamics and
influences on fire regime [75]. Topographic gradients could provide a partial explanation for the spatial
variation of fuel moisture, especially in light of our finding that TWI is inversely associated with
fire severity, as expected. Previous studies proved that TWI is closely associated with soil moisture
in European boreal forests [72,92]. Thus, we believe the drainage condition largely determines the
moisture gradient of a forest stand and further influences fuel moisture dynamics. Dead fuel moisture
dynamics are driven by three mechanisms—capillary forces, infiltration, and diffusion—among which,
infiltration and diffusion are the primary driving mechanisms and are both influenced by moisture
gradients [75]. Although live fuel moisture is driven by different mechanisms than dead fuel moisture
dynamics, soil water dynamics are still an important part of those mechanisms and can directly
influence plant transpiration.
The MODIS-derived 1 km SMA exerted considerable influences on model performance, and its
relationship with fire severity was negative, as expected. This finding aligns with the work of van
Mantgem et al. (2013) [18], which suggests that a pre-fire water deficit can increase fire severity
(tree mortality) because the drought-stressed trees are vulnerable to fire-induced injury. Similarly,
Xiao and Zhuang [93] found that drought directly affected fire activity in Canadian and Alaskan boreal
forests by enhancing fuel flammability and increasing ignitions. However, it should be noted that the
pre-fire SMA applied in this study can only represent the short-term temporal variability of the surface
moisture conditions, which may not necessarily reflect the long-term effects of drought stress on fire
severity. It has been reported that plant communities within a forest stand, especially understory
vegetation layers, may be influenced by the long-term drought stress that is regulated by topographic
and climatic factors [94].
Although the ranking of relative importance of vegetation, topography, and SMA was similar
between the two fire years (Figure 6), the relative importance of individual explanatory variables
differed between the two fire years (Figure 5). Such differences may be attributable to their different
pre-fire vegetation composition, structure, and disturbance history. By comparing the pre-fire
vegetation compositions (Figure 9), we found that fires in 2000 had lower proportions of DBF and MF
but higher proportions of DCF than the 2010 fires. Fires in 2000 burned more areas in the Huzhong
Natural Reserve where it is dominated by mature (>100 years) larch forest (DCF) due to the strictly
enforced cutting ban, and most DBF, MF, and shrubs are located in recently burned/disturbed areas
that carry significantly less fuels. Consequently, the proportion of DBF, MF, and shrubs were considered
more important than proportion of DCF in modeling severity of 2000 fires. In contrast, fires in 2010
burned more areas in the Huzhong Forestry Bureau jurisdiction that had been disturbed by clear
cutting since 1950s, and as recent as 2000s, leading to greater abundance of young stands irrespective of
forest type [95]. This could partly explain why fire severity of 2010 in the areas with high proportions of
DCF was similar to the areas with high proportion of DF or MF (Figure 7j vs. Figure 7b,d). In contrast,
because the highly flammable evergreen coniferous shrub species Pinus pumila is not an economically
viable species to cut, its fuel loading is generally higher than young stands of other forest types.
Consequently, the fire severity in areas with high proportion of ECF was high (Figure 7a), and ECF
 Forests 2018, 9, x FOR PEER REVIEW 17 of 26

irrespective of forest type [95]. This could partly explain why fire severity of 2010 in the areas with
high proportions of DCF was similar to the areas with high proportion of DF or MF (Figure 7j vs. 7b
Forests and
2018,7d). In contrast, because the highly flammable evergreen coniferous shrub species Pinus pumila
9, 130 17 of 26
is not an economically viable species to cut, its fuel loading is generally higher than young stands of
other forest types. Consequently, the fire severity in areas with high proportion of ECF was high
was considered
(Figure 7a),aandmore
ECFimportant vegetation
was considered a morevariable in vegetation
important driving overall fire
variable in severity variability
driving overall fire of
severity variability of 2010 fires.
2010 fires.

FigureFigure
9. An9.overview
An overview of the
of the pre-fire
pre-fire vegetationcomposition
vegetation composition within
withinfires of of
fires 2000 (a) (a)
2000 andand
20102010
(b). (b).
Variable
Variable abbreviations
abbreviations werewere described
described in Table
in Table 2. 2.

4.2. Prediction of Fire Severity

4.2. Prediction of Fire Severity
Our results revealed that landscape-level fire severity is primarily determined by fuel and
Our results
terrain revealed
features, that that
suggesting landscape-level
fire severity isfire severityat is
predictable primarily
this scale. Thedetermined by fuelis and
prediction accuracy
terraindetermined
features, suggesting
by the quality of the relevant predictors or proxies. Fuel conditions are accuracy
that fire severity is predictable at this scale. The prediction usually is
determined by the and
complicated quality of thetorelevant
difficult predictors
characterize, not onlyor proxies.
becauseFuel conditions
of the endogenous are usually
variety complicated
of plant
and difficult to characterize,
communities throughout the notlandscape
only becausebut alsoofbecause
the endogenous
of exogenousvariety
factors, of
suchplant communities
as disturbance
regime,the
throughout climate,
landscapeand anthropogenic
but also because activities [75]. Although
of exogenous various
factors, suchremote sensing methods
as disturbance regime,were climate,
proposed to improve fuel mapping, there are still challenges in
and anthropogenic activities [75]. Although various remote sensing methods were proposed accurately quantifying the critical fuel to
parameters that can regulate fire effects, such as fuel loading and canopy
improve fuel mapping, there are still challenges in accurately quantifying the critical fuel parameters bulk density [96,97].
Remotely sensed spectral information is usually applied as ancillary data or proxies when modeling
that can regulate fire effects, such as fuel loading and canopy bulk density [96,97]. Remotely sensed
fuel parameters, as it is strongly correlated with many biophysical vegetation parameters, such as
spectral information is usually applied as ancillary data or proxies when modeling fuel parameters,
biomass, leaf area index, and productivity [26,98,99].
as it is strongly
However,correlated
remotely with many
sensed biophysical
imagery vegetation
cannot detect surfaceparameters,
fuels obscured suchbyasthebiomass, leaf area
forest canopy
index,andandinsufficiently
productivity [26,98,99]. fine fuels from dead biomass pools due to the inconsistency between
distinguishes
However,
particle sizeremotely sensedresolution
and the spatial imageryofcannotthe image detect surface
[100]. Compared fuels obscured
with by complexity
the spatial the forest and canopy
and insufficiently
high variability distinguishes fine fuelsatfrom
of fuel parameters the dead biomass
landscape scale,pools due to cover
vegetation the inconsistency between
type is relatively
particleidentifiable
size and duethe to the unique
spatial spectral
resolution offeatures
the image of plant communities.
[100]. ComparedThe withLandsat-derived vegetation and
the spatial complexity
cover data can reflect the general variability of vegetation composition but may
high variability of fuel parameters at the landscape scale, vegetation cover type is relatively identifiable not necessarily depict
parameters related to fuel type or fuel particle size. Our results indicated
due to the unique spectral features of plant communities. The Landsat-derived vegetation cover data that vegetation coverage
could reliably explain the variability of fire severity at a 240 m spatial scale. The classification
can reflect the general variability of vegetation composition but may not necessarily depict parameters
accuracy of vegetation mapping could also affect the predictability of fire severity models. The spatial
relatedaggregation
to fuel typeprocess
or fuelmayparticle size. Our results indicated that vegetation coverage could reliably
confound the accuracy of the 240 m vegetation coverage data, thus
explain the variability
increasing of fire severity
the uncertainty at a 240 results.
of the modeling m spatial scale. The classification accuracy of vegetation
mapping could also affect the predictability
The large-scale terrain features of wildlands of fireare
severity
usuallymodels.
invariantThe
overspatial aggregation
long periods. process
In addition,
may confound
terrain features can be accurately characterized using various traditional or modern survey of
the accuracy of the 240 m vegetation coverage data, thus increasing the uncertainty
the modeling results.
technologies. Together with their close relationship with fire behavior, topographic variables are
commonly
The usedterrain
large-scale as predictors
featuresof fire severity. Digital
of wildlands elevation
are usually models provide
invariant over longessential
periods.topographic
In addition,
information from which aspect, slope, and other terrain features
terrain features can be accurately characterized using various traditional or modern can be derived. However, it should
survey
be noted that some topographic conditions are inherently scale-dependent. Thus, the upscaling
technologies. Together with their close relationship with fire behavior, topographic variables are
commonly used as predictors of fire severity. Digital elevation models provide essential topographic
information from which aspect, slope, and other terrain features can be derived. However, it should be
noted that some topographic conditions are inherently scale-dependent. Thus, the upscaling process
used for the ASTER GDEM data may filter out some detailed terrain features that would be reflected at
the 30 m spatial resolution. Because we focused on the 240 m spatial resolution, we did not examine the
sensitivity of the relationships between fire severity and topographic variables to the scaling process.
Many studies have reported scale dependency in the relationships between topographic characteristics
Forests 2018, 9, 130 18 of 26

and fire attributes, such as severity, frequency, burned area, and burn probability [101–104]. Despite
the good performance of our BRT models, we conclude that 240 m may not be the optimal spatial
resolution for predicting fire severity. Enlarging spatial scales (e.g., from 30 to 240 m) can be beneficial
in refining the relationships between fire severity and environmental gradients, but it may decrease
the visual effect of the prediction maps due to the coarse spatial resolution. Therefore, we suggest
that future applications should weigh the model performance against the practical applicability of the
prediction maps.

4.3. Limitation and Uncertainty

Despite efforts to improve the prediction of fire severity by incorporating sound explanatory
variables, some knowledge gaps should be noted as they may influence the interpretation of the
modeling results. First, each MOD16A2 pixel contains the best possible daily ET estimation during
the 8-day period, which was selected based on the imaging conditions and observation coverage.
The arbitrary application of SMA as a proxy for fuel moisture may increase the alternative quantification
of land surface status for modeling fire severity, but its real relationship with actual fuel moisture
still needs to be validated in our ecosystem. Due to the lack of daily fire progression maps, and to
very sparse weather station coverage in our study area, we cannot address how day-to-day weather
impacts fire severity in this study. With the advantage of high-frequency MODIS observations,
many recent studies have begun to incorporate spatial interpolation approaches using MODIS data to
characterize daily progressions of large fires [105,106]. We believe such efforts can greatly improve
the understanding of spatial controls on fire severity. For example, based on Landsat-derived fire
progression maps and fire weather observations at 4-km spatial resolution, Birch et al. (2015) [37]
investigated the influences of vegetation, topography, and daily fire weather on severity patterns
of wildfires in the Western US and reported that vegetation cover had the greatest influence on fire
severity; this is quite similar to our findings in this study, as well as in our previous patch-scale
analysis [43]. They also acknowledged that the coarse weather conditions may not fully reflect the
influences of microscale meteorological conditions on severity patterns. The inconsistent temporal and
spatial resolution among daily weather observations, vegetation, and topography can obscure the real
effects of weather on fire severity.
Fire severity is a result of accumulated fire effects on forest ecosystems because the thick and moist
organic layers in boreal regions can prolong the fire duration. Although we tried to balance the spatial
resolution among explanatory variables, it is somewhat challenging to reflect environmental gradients
and fire activities at both fine spatial and temporal scales using the data sources available to us. At the
same time, because MODIS ET product updates led to considerable changes in SMA values, we found
that 500 m SMA has different relative importance and influences on fire severity. Our intentions
were not to arbitrarily justify which kind of ET product is more reliable for fire severity prediction,
especially without sufficient validation of 500 m ET products with site-based flux observation, but any
improvement in spatial resolution is valuable and further efforts are encouraged to verify the suitability
of these products for specific ecosystems.
Although our study area, the Huzhong Forestry Bureau jurisdiction, is a representative forest
landscape of Great Xing’an Mountains and shares a similar fire regime as other nearby areas [54,56],
we could not conclude that fire severity patterns of the entire region are following the same causal
mechanism at finer scales. The purpose of this study is not to establish a global prediction model for
fire severity that can be generalized for all Chinese boreal ecosystems. There were fires occurring in
meadow and wetland ecosystems, as well as in forests dominated by broadleaf trees in the southern
part of the Great Xing’an region. We believe our results may not be suitable for predicting the severity
of those kinds of wildfires. In addition, sampling data were extracted from fires occurring in summer;
although the vegetation coverage and topographic variables adopted in this study are insensitive
to intra-annual variability, the relationship between SMA and fire severity may change in different
seasons and should be further investigated.
Forests 2018, 9, x FOR PEER REVIEW 19 of 26

intra-annual variability, the relationship between SMA and fire severity may change in different
seasons and should be further investigated.
Forests 2018, 9, 130 19 of 26
5. Conclusions
Although the parallel comparison of the two models did not show strictly consistent modeling
5. Conclusions
relationships, the models generally demonstrated that fire severity was strongly controlled by the
Although
coverage the parallel
of certain vegetation comparison
types thatofhave
the twohighmodels did notor
flammability show
fire strictly consistent
resistance. modeling
The topographic
relationships, the models generally demonstrated that fire severity
conditions can help determine the distribution of flammable plant types and communities.was strongly controlled by the
coverage of certain
Topography can alsovegetation
directly types that fuel
influence havemoisture
high flammability
and createorfirebreaks
fire resistance. Thethe
through topographic
drainage
conditions can help determine the distribution of flammable plant types and
systems. Remotely sensed fuel moisture proxies (such as MODIS ET products) were also proven to communities. Topography
can also
play directlyroles
important influence fuel moisture
in modeling and create
fire severity. firebreaks
These findingsthrough
reveal the
thatdrainage systems.
fire severity Remotely
is predictable
sensed
at fuel moisture
the landscape scaleproxies
in our (such
studyas MODIS
area, and ET products) were
its prediction can bealso proven tobyplay
improved important roles
incorporating in
spatial
modeling fire severity. These findings reveal that fire severity is predictable
variables related to fire behavior. Our study provides an overview of the hotspot areas within theat the landscape scale in
our study area,
landscape where and its prediction
severe fires arecan be improved
most by incorporating
likely distributed. spatial variables
Such mapping related
capabilities can to fire
allow
behavior. Our
managers study provides
to optimize an overview
fuel treatment of theby
strategies hotspot areas within
considering the landscape
the vegetation, where severe
topography, and
fires are most likely distributed. Such mapping capabilities can allow
spatial patterns of land surface moisture. The modeling framework employed in our study can managers to optimize fuel
treatment
readily strategies new
incorporate by considering
observations theand
vegetation,
simulated topography, and spatial
spatial datasets, patterns
promoting theofmore
land reliable
surface
moisture. The modeling framework
prediction of fire severity in the future. employed in our study can readily incorporate new observations
and simulated spatial datasets, promoting the more reliable prediction of fire severity in the future.
Acknowledgments: This work is funded by the National Key R&D Program of China (2017YFA0604403 and
2016YFA0600804), theThis
Acknowledgments: work Natural
National is funded by theFoundation
Science National Key R&D (Project
of China ProgramNo.
of 31500387,
China (2017YFA0604403
31270511, and
and 2016YFA0600804), the National Natural Science Foundation of China (Project No. 31500387, 31270511,
31470517) and the CAS Pioneer Hundred Talents Program. We thank three anonymous reviewers and academic
and 31470517) and the CAS Pioneer Hundred Talents Program. We thank three anonymous reviewers and
editor for editor
academic comments that improved
for comments this manuscript.
that improved this manuscript.
Author Contributions:
Contributions: L.F.
L.F. and
andJ.Y.
J.Y.conceived
conceivedand
anddesigned
designed the
the study.
study. L.F.L.F.
andand M.W.
M.W. analyzed
analyzed data.
data. L.F.,L.F., J.Y.
J.Y. and
M.W.M.W.
and wrote and revised
wrote the manuscript.
and revised Z.L. contributed
the manuscript. to collecting
Z.L. contributed data data
to collecting and discussing the results.
and discussing the results.
Conflicts of
Conflicts Interest: The
of Interest: The authors
authors declare
declare no
no conflict
conflict of
of interest.
interest.

Appendix A
Appendix A

Figure A1. The

Figure A1. Theintegrated
integratedpre-fire
pre-fire surface
surface moisture
moisture availability
availability (SMA)(SMA) of(a)
of 2000 2000
and(a) and
2010 2010
(b,c) (b,c)
derived
derived from 8-day MODIS MOD16A2 Version 5 (b) and Version 6 (c) product. The good
from 8-day MODIS MOD16A2 Version 5 (b) and Version 6 (c) product. The good quality pixels selected quality
pixels selected
from five 8-day from
MOD16A2five 8-day MOD16A2
datasets datasets
(d–f) were (d–f) were
composited. composited.
Fire patches Fire
of 2000 (inpatches ofand
blue, a,d) 20002010
(in
blue, a,d) and 2010 (in red, b,d–f)
(in red, b,d–f) were also plotted. were also plotted.
Forests 2018,9,9,130
Forests2018, x FOR PEER REVIEW 20
20 ofof2626

Forests 2018, 9, x FOR PEER REVIEW 20 of 26

Forests 2018, 9, x FOR PEER REVIEW 20 of 26

FigureA2.
Figure A2.The
Thescatter
scatter plot
plot representing
representing inconsistent
inconsistent surface
surface moisture
moisture availability
availability (SMA)(SMA) values
values of theof
Figure A2. The scatter plot representing inconsistent surface moisture availability (SMA) values of
the whole
whole studystudy
Figure areaThe
A2. area (black
(black dot) dot)
and and burned
burned pixels pixels
of 2010 of 2010
fires fires
(red dot)(red dot) from
derived derived from
MOD16A2 MOD16A2
Verison
the whole studyscatter plot
area (blackrepresenting inconsistent
dot) and burned pixels ofsurface moisture
2010 fires availability
(red dot) derived (SMA) values
from MOD16A2 of
5 Verison
(V5)the
and5 Version
(V5)study
whole and6 Version
(V6) with6 (V6)
area (black dot) with
good good quality.
quality.
and burned pixels of 2010 fires (red dot) derived from MOD16A2
Verison 5 (V5) and Version 6 (V6) with good quality.
Verison 5 (V5) and Version 6 (V6) with good quality.
70 70
70 Low Moderate HighHigh
Low Moderate
60 60 Low Moderate High
60
(%)(%)
Pixel Proportion (%)

50 50
50
Proportion

40 40
Proportion

40
30
30 30
Pixel

20
Pixel

20 20
10
10 10
0
0 0 Observed_2000 Predicted_2000 Observed_2010 Predicted_2010
Observed_2000 Predicted_2000 Observed_2010
Observed_2000 Predicted_2000 Observed_2010 Predicted_2010
Predicted_2010
Figure A3. Comparison of pixel proportions of low, moderate and high severity levels between
Figure A3.
Figure Comparison
A3.fire
Comparison of pixel
of pixelproportions
proportions of low,
of low, moderate and high severity
moderate and high severity levels
levels between
between
observed
Figure A3. severity maps and predicted severity maps.
observed fire Comparison
observed severity maps
fire severity
ofand
maps
pixel proportions
predicted
and
of low,
predictedseverity
severitymaps.
maps.
moderate and high severity levels between
observed fire severity maps and predicted severity maps.
GRS
GRS RIMP2010_V6
DCF RIMP2010_V6
RIMP2010_V5
GRSDCF
PSR
RIMP2010_V6
RIMP2010_V5
DCFPSR RIMP2010_V5
TWI
PSR TWI
ELV
ELV
TWI
SLP
ELVSLP
SMA
SMA
SLP MF
MF
SMA SRB
SRB
MFDBF
DBF
SRB ECF
ECF
DBF 0 5 10 15 20 25 30 35
0 5 10 Relative
15 Importance
20 (%)25 30 35
ECF Relative Importance (%)

0 importance
Figure A4. The relative 5 10 of variables
(RIMP) 15 20
generated 2550 boosting
by 30 regression
35 tree (BRT)
Figure A4. The
Figure
models A4. relative
The
using importance
relative
two importance
different (RIMP)
versions(RIMP) of
ofvariables
(V5 and variables generated
generated
V6) ofImportance
surface byby
moisture5050 boosting
boosting regression
regression tree (BRT)
tree (BRT)
Relative (%) availability (SMA) derived from
models
models using
using two two different
different versions
versions (V5
(V5 and
and V6)
V6) of
of surface
surface moisture
moisture
MODIS ET products. Variable abbreviations were described in Table 2. availability
availability (SMA) derived
(SMA) fromfrom
derived
MODISMODIS ET products.
ET products. Variable
Variable abbreviationswere
abbreviations weredescribed
described in
inTable
Table2.2.
Figure A4. The relative importance (RIMP) of variables generated by 50 boosting regression tree (BRT)
models using two different versions (V5 and V6) of surface moisture availability (SMA) derived from
MODIS ET products. Variable abbreviations were described in Table 2.
Forests 2018, 9, 130 21 of 26
Forests 2018, 9, x FOR PEER REVIEW 21 of 26

FigureA5.
Figure A5.Partial
Partialdependence
dependence plots of of nine
nine variables
variableson onregulating
regulatingfire
fireseverity
severityininfire
fireyear
year2010 using
2010 using
differentversions
different versionsof of MODIS
MODIS derived
derived surface
surface moisture
moisture availability
availability (SMA)(SMA) (Version
(Version 5 in cyan 5 in
andcyan and 6
Version
inVersion 6 in grouping
red). This red). This of
grouping
variablesof was
variables wasasselected
selected the topas9 the top 9 variables
variables followingfollowing
the rank the rank asin
as shown
shown in y-axis of Figure 5. The curves demonstrate the smoothed means (solid
y-axis of Figure 5. The curves demonstrate the smoothed means (solid line) and 95% confidence intervalsline) and 95%
confidence intervals (gray zone) of an ensemble of 50 models. Variable abbreviations
(gray zone) of an ensemble of 50 models. Variable abbreviations were described in Table 2. were described
in Table 2.
Appendix B
Appendix B

Table
TableA1. The
A1. Moran’s
The I forI examining
Moran’s spatial
for examining autocorrelation
spatial of fireof
autocorrelation severity and explanatory
fire severity variables.
and explanatory
Variable abbreviations
variables. were described
Variable abbreviations were in Table 2. in Table 2.
described

Variables Moran’s I of Fires 2000 Moran’s I of Fires 2010

Variables Moran’s I of Fires 2000 Moran’s I of Fires 2010
Fire Severity 0.132 0.078
Fire Severity 0.132 0.078
SMA_V5 0.224 0.124
SMA_V5 0.224 0.124
ELV
ELV 0.192
0.192 0.236
0.236
PSR
PSR 0.061
0.061 0.066
0.066
SLP
SLP 0.120
0.120 0.075
0.075
TWI
TWI 0.054
0.054 0.06
0.06
ECF 0.078 0.092
ECF 0.078 0.092
DCF 0.178 0.107
DCF
DBF 0.178
0.041 0.107
0.19
DBF
MF 0.041
0.134 0.19
0.07
MF
GRS 0.134
0.104 0.07
0.056
SRB
GRS 0.035
0.104 0.156
0.056
SRB 0.035 0.156
Forests 2018, 9, 130 22 of 26

References
1. De Groot, W.J.; Flannigan, M.D.; Cantin, A.S. Climate change impacts on future boreal fire regimes.
For. Ecol. Manag. 2013, 294, 35–44. [CrossRef]
2. Flannigan, M.; Stocks, B.; Turetsky, M.; Wotton, M. Impacts of climate change on fire activity and fire
management in the circumboreal forest. Glob. Chang. Biol. 2009, 15, 549–560. [CrossRef]
3. Stephens, S.L.; Burrows, N.; Buyantuyev, A.; Gray, R.W.; Keane, R.E.; Kubian, R.; Liu, S.; Seijo, F.;
Shu, L.; Tolhurst, K.G.; et al. Temperate and boreal forest mega-fires: Characteristics and challenges.
Front. Ecol. Environ. 2014, 12, 115–122. [CrossRef]
4. Conard, S.G.; Solomon, A.M. Effects of wildland fire on regional and global carbon stocks in a changing
environment. Dev. Environ. Sci. 2008, 8, 109–138.
5. Harden, J.; Trumbore, S.; Stocks, B.; Hirsch, A.; Gower, S.; O’neill, K.; Kasischke, E. The role of fire in the
boreal carbon budget. Glob. Chang. Biol. 2000, 6, 174–184. [CrossRef]
6. Bond-Lamberty, B.; Peckham, S.D.; Gower, S.T.; Ewers, B.E. Effects of fire on regional evapotranspiration in
the central Canadian boreal forest. Glob. Chang. Biol. 2009, 15, 1242–1254. [CrossRef]
7. Jin, Y.; Roy, D. Fire-induced albedo change and its radiative forcing at the surface in northern Australia.
Geophys. Res. Lett. 2005, 32. [CrossRef]
8. Keane, R.E.; Agee, J.K.; Fulé, P.; Keeley, J.E.; Key, C.; Kitchen, S.G.; Miller, R.; Schulte, L.A. Ecological effects
of large fires on US landscapes: Benefit or catastrophe? Int. J. Wildland Fire 2008, 17, 696–712. [CrossRef]
9. Anderson-Teixeira, K.J.; Miller, A.D.; Mohan, J.E.; Hudiburg, T.W.; Duval, B.D.; Delucia, E.H. Altered
dynamics of forest recovery under a changing climate. Glob. Chang. Biol. 2013, 19, 2001–2021. [CrossRef]
[PubMed]
10. Bond, W.J.; Keeley, J.E. Fire as a global ‘herbivore’: The ecology and evolution of flammable ecosystems.
Trends Ecol. Evol. 2005, 20, 387–394. [CrossRef] [PubMed]
11. Johnstone, J.F.; Hollingsworth, T.N.; Chapin, F.S.; Mack, M.C. Changes in fire regime break the legacy lock
on successional trajectories in Alaskan boreal forest. Glob. Chang. Biol. 2010, 16, 1281–1295. [CrossRef]
12. Turner, M.G.; Hargrove, W.W.; Gardner, R.H.; Romme, W.H. Effects of fire on landscape heterogeneity in
Yellowstone National Park, Wyoming. J. Veg. Sci. 1994, 5, 731–742. [CrossRef]
13. Amiro, B.; Orchansky, A.; Barr, A.; Black, T.; Chambers, S.; Chapin, F., III; Goulden, M.; Litvak, M.; Liu, H.;
McCaughey, J. The effect of post-fire stand age on the boreal forest energy balance. Agric. For. Meteorol. 2006,
140, 41–50. [CrossRef]
14. Beck, P.S.; Goetz, S.J.; Mack, M.C.; Alexander, H.D.; Jin, Y.; Randerson, J.T.; Loranty, M. The impacts and
implications of an intensifying fire regime on Alaskan boreal forest composition and albedo. Glob. Chang. Biol.
2011, 17, 2853–2866. [CrossRef]
15. Rogers, B.M.; Soja, A.J.; Goulden, M.L.; Randerson, J.T. Influence of tree species on continental differences in
boreal fires and climate feedbacks. Nat. Geosci. 2015, 8, 228–234. [CrossRef]
16. De Groot, W.J.; Cantin, A.S.; Flannigan, M.D.; Soja, A.J.; Gowman, L.M.; Newbery, A. A comparison of
Canadian and Russian boreal forest fire regimes. For. Ecol. Manag. 2013, 294, 23–34. [CrossRef]
17. Running, S.W. Is global warming causing more, larger wildfires? Science 2006, 313, 927–928. [CrossRef]
[PubMed]
18. Van Mantgem, P.J.; Nesmith, J.C.; Keifer, M.; Knapp, E.E.; Flint, A.; Flint, L. Climatic stress increases forest
fire severity across the western United States. Ecol. Lett. 2013, 16, 1151–1156. [CrossRef] [PubMed]
19. Key, C.H.; Benson, N.C. Landscape Assessment (LA). Firemon: Fire Effects Monitoring and Inventory System;
U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station: Fort Collins, CO,
USA, 2006.
20. Lentile, L.B.; Holden, Z.A.; Smith, A.M.S.; Falkowski, M.J.; Hudak, A.T.; Morgan, P.; Lewis, S.A.; Gessler, P.E.;
Benson, N.C. Remote sensing techniques to assess active fire characteristics and post-fire effects. Int. J.
Wildland Fire 2006, 15, 319–345. [CrossRef]
21. De Santis, A.; Chuvieco, E. Geocbi: A modified version of the composite burn index for the initial assessment
of the short-term burn severity from remotely sensed data. Remote Sens. Environ. 2009, 113, 554–562.
[CrossRef]
Forests 2018, 9, 130 23 of 26

22. French, N.H.F.; Kasischke, E.S.; Hall, R.J.; Murphy, K.A.; Verbyla, D.L.; Hoy, E.E.; Allen, J.L. Using landsat
data to assess fire and burn severity in the north American boreal forest region: An overview and summary
of results. Int. J. Wildland Fire 2008, 17, 443–462. [CrossRef]
23. Hollingsworth, T.N.; Johnstone, J.F.; Bernhardt, E.L.; Chapin, F.S. Fire severity filters regeneration traits to
shape community assembly in Alaska’s boreal forest. PLoS ONE 2013, 8, e56033. [CrossRef] [PubMed]
24. Liu, Z.; Yang, J. Quantifying ecological drivers of ecosystem productivity of the early-successional boreal
Larix gmelinii forest. Ecosphere 2014, 5, 1–16. [CrossRef]
25. Turner, M.G. Disturbance and landscape dynamics in a changing world. Ecology 2010, 91, 2833–2849.
[CrossRef] [PubMed]
26. Rollins, M.G.; Keane, R.E.; Parsons, R.A. Mapping fuels and fire regimes using remote sensing, ecosystem
simulation, and gradient modeling. Ecol. Appl. 2004, 14, 75–95. [CrossRef]
27. Morgan, P.; Keane, R.E.; Dillon, G.K.; Jain, T.B.; Hudak, A.T.; Karau, E.C.; Sikkink, P.G.; Holden, Z.A.;
Strand, E.K. Challenges of assessing fire and burn severity using field measures, remote sensing and
modelling. Int. J. Wildland Fire 2014, 23, 1045–1460. [CrossRef]
28. Fang, L.; Yang, J. Atmospheric effects on the performance and threshold extrapolation of multi-temporal
landsat derived dNBR for burn severity assessment. Int. J. Appl. Earth Obs. Geoinf. 2014, 33, 10–20. [CrossRef]
29. Keane, R.E.; Drury, S.A.; Karau, E.C.; Hessburg, P.F.; Reynolds, K.M. A method for mapping fire hazard and
risk across multiple scales and its application in fire management. Ecol. Model. 2010, 221, 2–18. [CrossRef]
30. Stephens, S.L.; Moghaddas, J.J.; Edminster, C.; Fiedler, C.E.; Haase, S.; Harrington, M.; Keeley, J.E.;
Knapp, E.E.; McIver, J.D.; Metlen, K.; et al. Fire treatment effects on vegetation structure, fuels, and potential
fire severity in western U.S. Forests. Ecol. Appl. 2009, 19, 305–320. [CrossRef] [PubMed]
31. Whitlock, C.; Higuera, P.E.; McWethy, D.B.; Briles, C.E. Paleoecological perspectives on fire ecology:
Revisiting the fire-regime concept. Open Ecol. J. 2010, 3, 6–23. [CrossRef]
32. Whelan, R.J. The Ecology of Fire; Cambridge University Press: Cambridge, UK, 1995.
33. Cary, G.J.; Keane, R.E.; Gardner, R.H.; Lavorel, S.; Flannigan, M.D.; Davies, I.D.; Li, C.; Lenihan, J.M.;
Rupp, T.S.; Mouillot, F. Comparison of the sensitivity of landscape-fire-succession models to variation in
terrain, fuel pattern, climate and weather. Landsc. Ecol. 2006, 21, 121–137. [CrossRef]
34. Karau, E.C.; Keane, R.E. Burn severity mapping using simulation modelling and satellite imagery. Int. J.
Wildland Fire 2010, 19, 710–724. [CrossRef]
35. Keane, R.E.; Cary, G.J.; Davies, I.D.; Flannigan, M.D.; Gardner, R.H.; Lavorel, S.; Lenihan, J.M.; Li, C.;
Rupp, T.S. A classification of landscape fire succession models: Spatial simulations of fire and vegetation
dynamics. Ecol. Model. 2004, 179, 3–27. [CrossRef]
36. Kane, V.R.; Cansler, C.A.; Povak, N.A.; Kane, J.T.; McGaughey, R.J.; Lutz, J.A.; Churchill, D.J.; North, M.P.
Mixed severity fire effects within the rim fire: Relative importance of local climate, fire weather, topography,
and forest structure. For. Ecol. Manag. 2015, 358, 62–79. [CrossRef]
37. Birch, D.S.; Morgan, P.; Kolden, C.A.; Abatzoglou, J.T.; Dillon, G.K.; Hudak, A.T.; Smith, A.M.S. Vegetation,
topography and daily weather influenced burn severity in central Idaho and western Montana forests.
Ecosphere 2015, 6, 1–23. [CrossRef]
38. Estes, B.L.; Knapp, E.E.; Skinner, C.N.; Miller, J.D.; Preisler, H.K. Factors influencing fire severity under
moderate burning conditions in the Klamath Mountains, northern California, USA. Ecosphere 2017, 8.
[CrossRef]
39. Alexander, J.D.; Seavy, N.E.; Ralph, C.J.; Hogoboom, B. Vegetation and topographical correlates of fire
severity from two fires in the Klamath-Siskiyou region of Oregon and California. Int. J. Wildland Fire 2006,
15, 237–245. [CrossRef]
40. Lentile, L.B.; Smith, F.W.; Shepperd, W.D. Influence of topography and forest structure on patterns of mixed
severity fire in ponderosa pine forests of the South Dakota Black Hills, USA. Int. J. Wildland Fire 2006, 15,
557–566. [CrossRef]
41. Peters, D.P.; Pielke, R.A., Sr.; Bestelmeyer, B.T.; Allen, C.D.; Munson-McGee, S.; Havstad, K.M. Cross-scale
interactions, nonlinearities, and forecasting catastrophic events. Proc. Natl. Acad. Sci. USA 2004, 101,
15130–15135. [CrossRef] [PubMed]
42. Dillon, G.K.; Holden, Z.A.; Morgan, P.; Crimmins, M.A.; Heyerdahl, E.K.; Luce, C.H. Both topography and
climate affected forest and woodland burn severity in two regions of the western US, 1984 to 2006. Ecosphere
2011, 2, 1–33. [CrossRef]
Forests 2018, 9, 130 24 of 26

43. Fang, L.; Yang, J.; Zu, J.; Li, G.; Zhang, J. Quantifying influences and relative importance of fire weather,
topography, and vegetation on fire size and fire severity in a Chinese boreal forest landscape. For. Ecol. Manag.
2015, 356, 2–12. [CrossRef]
44. Parks, S.A.; Parisien, M.A.; Miller, C.; Dobrowski, S.Z. Fire activity and severity in the western US vary
along proxy gradients representing fuel amount and fuel moisture. PLoS ONE 2014, 9, e99699. [CrossRef]
[PubMed]
45. Viedma, O.; Quesada, J.; Torres, I.; De Santis, A.; Moreno, J.M. Fire severity in a large fire in a pinus pinaster
forest is highly predictable from burning conditions, stand structure, and topography. Ecosystems 2014, 18,
237–250. [CrossRef]
46. Cansler, C.A.; McKenzie, D. Climate, fire size, and biophysical setting control fire severity and spatial pattern
in the northern Cascade Range, USA. Ecol. Appl. 2014, 24, 1037–1056. [CrossRef] [PubMed]
47. Clarke, P.J.; Knox, K.J.E.; Bradstock, R.A.; Munoz-Robles, C.; Kumar, L.; Ward, D. Vegetation, terrain and fire
history shape the impact of extreme weather on fire severity and ecosystem response. J. Veg. Sci. 2014, 25,
1033–1044. [CrossRef]
48. Qi, Y.; Dennison, P.E.; Spencer, J.; Riano, D. Monitoring live fuel moisture using soil moisture and remote
sensing proxies. Fire Ecol. 2012, 8, 71–87.
49. Yebra, M.; Dennison, P.E.; Chuvieco, E.; Riaño, D.; Zylstra, P.; Hunt, E.R.; Danson, F.M.; Qi, Y.; Jurdao, S. A
global review of remote sensing of live fuel moisture content for fire danger assessment: Moving towards
operational products. Remote Sens. Environ. 2013, 136, 455–468. [CrossRef]
50. Barrett, K.; Kasischke, E.; McGuire, A.; Turetsky, M.; Kane, E. Modeling fire severity in black spruce stands
in the Alaskan boreal forest using spectral and non-spectral geospatial data. Remote Sens. Environ. 2010, 114,
1494–1503. [CrossRef]
51. Holden, Z.A.; Morgan, P.; Evans, J.S. A predictive model of burn severity based on 20-year satellite-inferred
burn severity data in a large southwestern US wilderness area. For. Ecol. Manag. 2009, 258, 2399–2406.
52. Kane, V.R.; Lutz, J.A.; Alina Cansler, C.; Povak, N.A.; Churchill, D.J.; Smith, D.F.; Kane, J.T.; North, M.P.
Water balance and topography predict fire and forest structure patterns. For. Ecol. Manag. 2015, 338, 1–13.
[CrossRef]
53. Ponomarev, E.; Kharuk, V.; Ranson, K. Wildfires dynamics in siberian larch forests. Forests 2016, 7, 125.
[CrossRef]
54. Fan, Q.; Wang, C.; Zhang, D.; Zang, S. Environmental influences on forest fire regime in the Greater Hinggan
Mountains, northeast China. Forests 2017, 8, 372. [CrossRef]
55. Fang, J.; Chen, A.; Peng, C.; Zhao, S.; Ci, L. Changes in forest biomass carbon storage in China between 1949
and 1998. Science 2001, 292, 2320–2322. [CrossRef] [PubMed]
56. Liu, Z.; Yang, J.; Chang, Y.; Weisberg, P.J.; He, H.S. Spatial patterns and drivers of fire occurrence and its
future trend under climate change in a boreal forest of northeast China. Glob. Chang. Biol. 2012, 18, 2041–2056.
[CrossRef]
57. Wang, C.; Gower, S.T.; Wang, Y.; Zhao, H.; Yan, P.; Lamberty, B.P. The influence of fire on carbon distribution
and net primary production of boreal larix gmelinii forests in north-eastern China. Glob. Chang. Biol. 2001, 7,
719–730. [CrossRef]
58. Liu, Z.; Yang, J.; He, H.S. Identifying the threshold of dominant controls on fire spread in a boreal forest
landscape of northeast China. PLoS ONE 2013, 8, e55618. [CrossRef] [PubMed]
59. Cai, W.; Yang, J.; Liu, Z.; Hu, Y.; Weisberg, P.J. Post-fire tree recruitment of a boreal larch forest in northeast
China. For. Ecol. Manag. 2013, 307, 20–29. [CrossRef]
60. Kong, J.; Yang, J.; Chu, H.; Xiang, X. Effects of wildfire and topography on soil nitrogen availability in a
boreal larch forest of northeastern China. Int. J. Wildland Fire 2015, 24, 433–442. [CrossRef]
61. Chander, G.; Markham, B.L.; Helder, D.L. Summary of current radiometric calibration coefficients for landsat
MSS, TM, ETM+, and EO-1 ALI sensors. Remote Sens. Environ. 2009, 113, 893–903. [CrossRef]
62. Vermote, E.F.; Tanre, D.; Deuze, J.L.; Herman, M.; Morcette, J.J. Second simulation of the satellite signal in
the solar spectrum, 6S: An overview. IEEE Trans. Geosci. Remote Sens. 1997, 35, 675–686. [CrossRef]
63. Canty, M.J.; Nielsen, A.A. Automatic radiometric normalization of multitemporal satellite imagery with the
iteratively re-weighted mad transformation. Remote Sens. Environ. 2008, 112, 1025–1036. [CrossRef]
64. Mu, Q.; Zhao, M.; Running, S.W. Improvements to a modis global terrestrial evapotranspiration algorithm.
Remote Sens. Environ. 2011, 115, 1781–1800. [CrossRef]
Forests 2018, 9, 130 25 of 26

65. Kim, H.W.; Hwang, K.; Mu, Q.; Lee, S.O.; Choi, M. Validation of modis 16 global terrestrial evapotranspiration
products in various climates and land cover types in Asia. KSCE J. Civ. Eng. 2012, 16, 229–238. [CrossRef]
66. Wang, L.; Hunt, E.R.; Qu, J.J.; Hao, X.; Daughtry, C.S.T. Remote sensing of fuel moisture content from ratios of
narrow-band vegetation water and dry-matter indices. Remote Sens. Environ. 2013, 129, 103–110. [CrossRef]
67. Running, S.W.; Mu, Q.; Zhao, M.; Moreno, A. Modis Global Terrestrial Evapotranspiration (ET) Product
(NASA MOD16A2/A3) NASA Earth Observing System Modis Land Algorithm; NASA: Washington, DC,
USA, 2017.
68. Masek, J.G.; Huang, C.; Wolfe, R.; Cohen, W.; Hall, F.; Kutler, J.; Nelson, P. North American forest disturbance
mapped from a decadal landsat record. Remote Sens. Environ. 2008, 112, 2914–2926. [CrossRef]
69. Friedl, M.A.; Brodley, C.E. Decision tree classification of land cover from remotely sensed data.
Remote Sens. Environ. 1997, 61, 399–409. [CrossRef]
70. Yang, J.; He, H.S.; Shifley, S.R.; Gustafson, E.J. Spatial patterns of modern period human-caused fire
occurrence in the Missouri Ozark Highlands. For. Sci. 2007, 53, 1–15.
71. Quinn, P.; Beven, K. Spatial and temporal predictions of soil moisture dynamics, runoff, variable source
areas and evapotranspiration for Plynlimon, Mid-Wales. Hydrol. Process. 1993, 7, 425–448. [CrossRef]
72. Sörensen, R.; Zinko, U.; Seibert, J. On the calculation of the topographic wetness index: Evaluation of
different methods based on field observations. Hydrol. Earth Syst. Sci. Discuss. 2006, 10, 101–112. [CrossRef]
73. Qin, C.; Zhu, A.; Yang, L.; Li, B.; Pei, T. Topographic wetness index computed using multiple flow direction
algorithm and local maximum downslope gradient. In Proceedings of the 7th International Workshop of
Geographical Information System, Beijing, China, 12–14 September 2007.
74. Wotton, B.M. Interpreting and using outputs from the Canadian forest fire danger rating system in research
applications. Environ. Ecol. Stat. 2009, 16, 107–131. [CrossRef]
75. Keane, R.E. Wildland Fuel Fundamentals and Applications; Springer: Berlin, Germany, 2015.
76. Carlson, T.N. Regional-scale estimates of surface moisture availability and thermal inertia using remote
thermal measurements. Remote Sens. Rev. 1986, 1, 197–247. [CrossRef]
77. Ha, W.; Gowda, P.H.; Howell, T.A. A review of downscaling methods for remote sensing-based irrigation
management: Part I. Irrig. Sci. 2012, 31, 831–850. [CrossRef]
78. De’ath, G. Boosted trees for ecological modeling and prediction. Ecology 2007, 88, 243–251. [CrossRef]
79. Elith, J.; Leathwick, J.R.; Hastie, T. A working guide to boosted regression trees. J. Anim. Ecol. 2008, 77,
802–813. [CrossRef] [PubMed]
80. Paradis, E.; Claude, J.; Strimmer, K. Ape: Analyses of phylogenetics and evolution in R language.
Bioinformatics 2004, 20, 289–290. [CrossRef] [PubMed]
81. Legendre, P. Spatial autocorrelation: Trouble or new paradigm? Ecology 1993, 74, 1659–1673. [CrossRef]
82. Ferster, C.; Eskelson, B.; Andison, D.; LeMay, V. Vegetation Mortality within Natural Wildfire Events in the
Western Canadian Boreal Forest: What Burns and Why? Forests 2016, 7, 187. [CrossRef]
83. Whitman, E.; Parisien, M.-A.; Thompson, D.; Hall, R.; Skakun, R.; Flannigan, M. Variability and drivers of
burn severity in the northwestern Canadian boreal forest. Ecosphere 2018, 9, e02128. [CrossRef]
84. Bigler, C.; Kulakowski, D.; Veblen, T.T. Multiple disturbance interactions and drought influence fire severity
in Rocky Mountain subalpine forests. Ecology 2005, 86, 3018–3029. [CrossRef]
85. Schoennagel, T.; Veblen, T.T.; Romme, W.H. The interaction of fire, fuels, and climate across Rocky Mountain
forests. BioScience 2004, 54, 661–676. [CrossRef]
86. Rupp, T.S.; Starfield, A.M.; Chapin, F.S.; Duffy, P. Modeling the impact of black spruce on the fire regime of
Alaskan boreal forest. Clim. Chang. 2002, 55, 213–233. [CrossRef]
87. Johnstone, J.F.; Rupp, T.S.; Olson, M.; Verbyla, D. Modeling impacts of fire severity on successional trajectories
and future fire behavior in Alaskan boreal forests. Landsc. Ecol. 2011, 26, 487–500. [CrossRef]
88. Kobayashi, M.; Yury, P.N.; Olga, A.Z.; Takuya, K.; Yojiro, M.; Toshiya, Y.; Fuyuki, S.; Kaichiro, S.; Takayoshi, K.
Regeneration after forest fires in mixed conifer broad-leaved forests of the amur region in far eastern Russia:
The relationship between species specific traits against fire and recent fire regimes. Eurasian J. For. Res. 2007,
10, 51–58.
89. Zhao, F.; Shu, L.; Wang, M.; Liu, B.; Yang, L. Influencing factors on early vegetation restoration in burned
area of Pinus Pumila—Larch forest. Acta Ecol. Sin. 2012, 32, 57–61. [CrossRef]
90. Zhao, F.; Shu, L.; Wang, Q.; Wang, M.; Tian, X. Emissions of volatile organic compounds from heated needles
and twigs of Pinus Pumila. J. For. Res. 2011, 22, 243–248. [CrossRef]
Forests 2018, 9, 130 26 of 26

91. Ryan, K.C. Dynamic interactions between forest structure and fire behavior in boreal ecosystems. Silva Fenn.
2002, 36, 13–39. [CrossRef]
92. Kopecký, M.; Čížková, Š. Using topographic wetness index in vegetation ecology: Does the algorithm
matter? Appl. Veg. Sci. 2010, 13, 450–459. [CrossRef]
93. Xiao, J.; Zhuang, Q. Drought effects on large fire activity in Canadian and Alaskan forests. Environ. Res. Lett.
2007, 2, 044003. [CrossRef]
94. Parks, S.A.; Miller, C.; Nelson, C.R.; Holden, Z.A. Previous fires moderate burn severity of subsequent
wildland fires in two large western US wilderness areas. Ecosystems 2014, 17, 29–42. [CrossRef]
95. Li, X.; He, H.S.; Wu, Z.; Liang, Y.; Schneiderman, J.E. Comparing effects of climate warming, fire, and timber
harvesting on a boreal forest landscape in northeastern China. PLoS ONE 2013, 8, e59747. [CrossRef]
[PubMed]
96. Arroyo, L.A.; Pascual, C.; Manzanera, J.A. Fire models and methods to map fuel types: The role of remote
sensing. For. Ecol. Manag. 2008, 256, 1239–1252. [CrossRef]
97. Tian, X.; Mcrae, D.J.; Shu, L.; Wang, M.; Li, H. Satellite remote-sensing technologies used in forest fire
management. J. For. Res. 2005, 16, 73–78.
98. Pierce, A.D.; Farris, C.A.; Taylor, A.H. Use of random forests for modeling and mapping forest canopy fuels
for fire behavior analysis in Lassen Volcanic National Park, California, USA. For. Ecol. Manag. 2012, 279,
77–89. [CrossRef]
99. Song, C. Optical remote sensing of forest leaf area index and biomass. Prog. Phys. Geogr. 2012, 37, 98–113.
[CrossRef]
100. Keane, R.E.; Burgan, R.; van Wagtendonk, J. Mapping wildland fuels for fire management across multiple
scales: Integrating remote sensing, GIS, and biophysical modeling. Int. J. Wildland Fire 2001, 10, 301–319.
[CrossRef]
101. Cyr, D.; Gauthier, S.; Bergeron, Y. Scale-dependent determinants of heterogeneity in fire frequency in a
coniferous boreal forest of eastern Canada. Landsc. Ecol. 2007, 22, 1325–1339. [CrossRef]
102. Parisien, M.A.; Parks, S.A.; Krawchuk, M.A.; Flannigan, M.D.; Bowman, L.M.; Moritz, M.A. Scale-dependent
controls on the area burned in the boreal forest of Canada, 1980–2005. Ecol. Appl. 2011, 21, 789–805.
[CrossRef] [PubMed]
103. Parks, S.A.; Parisien, M.-A.; Miller, C. Multi-scale evaluation of the environmental controls on burn
probability in a southern Sierra Nevada landscape. Int. J. Wildland Fire 2011, 20, 815–828. [CrossRef]
104. Wu, Z.; He, H.S.; Bobryk, C.W.; Liang, Y. Scale effects of vegetation and topography on burn severity under
prevailing fire weather conditions in boreal forest landscapes of northeastern China. Scand. J. For. Res. 2014,
29, 60–70. [CrossRef]
105. Veraverbeke, S.; Sedano, F.; Hook, S.J.; Randerson, J.T.; Jin, Y.; Rogers, B.M. Mapping the daily progression of
large wildland fires using modis active fire data. Int. J. Wildland Fire 2014, 23, 655–667. [CrossRef]
106. Parks, S.A. Mapping day-of-burning with coarse-resolution satellite fire-detection data. Int. J. Wildland Fire
2014, 23, 215–223. [CrossRef]

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).