Severe fire weather and intensive forest management increase fire severity in a multi‐ownership landscape

Abstract

Many studies have examined how fuels, topography, climate, and fire weather influence fire severity. Less is known about how different forest management practices influence fire severity in multi‐owner landscapes, despite costly and controversial suppression of wildfires that do not acknowledge ownership boundaries. In 2013, the Douglas Complex burned over 19,000 ha of Oregon & California Railroad (O&C) lands in Southwestern Oregon, USA. O&C lands are composed of a checkerboard of private industrial and federal forestland (Bureau of Land Management, BLM) with contrasting management objectives, providing a unique experimental landscape to understand how different management practices influence wildfire severity. Leveraging Landsat based estimates of fire severity (Relative differenced Normalized Burn Ratio, RdNBR) and geospatial data on fire progression, weather, topography, pre‐fire forest conditions, and land ownership, we asked (1) what is the relative importance of different variables driving fire severity, and (2) is intensive plantation forestry associated with higher fire severity? Using Random Forest ensemble machine learning, we found daily fire weather was the most important predictor of fire severity, followed by stand age and ownership, followed by topographic features. Estimates of pre‐fire forest biomass were not an important predictor of fire severity. Adjusting for all other predictor variables in a general least squares model incorporating spatial autocorrelation, mean predicted RdNBR was higher on private industrial forests (RdNBR 521.85 ± 18.67 [mean ± SE]) vs. BLM forests (398.87 ± 18.23) with a much greater proportion of older forests. Our findings suggest intensive plantation forestry characterized by young forests and spatially homogenized fuels, rather than pre‐fire biomass, were significant drivers of wildfire severity. This has implications for perceptions of wildfire risk, shared fire management responsibilities, and developing fire resilience for multiple objectives in multi‐owner landscapes.

Full text

Severe fire weather and intensive forest management increase fire
severity in a multi-ownership landscape
HAROLD S. J. ZALD
1,3
AND CHRISTOPHER J. DUNN
2
1
Department of Forestry and Wildland Resources, Humboldt State University, 1 Harpst Street, Arcata, California 95521 USA
2
Department of Forest Engineering, Resources, and Management, Oregon State University, 280 Peavy Hall, Corvallis, Oregon 97331 USA
Abstract. Many studies have examined how fuels, topography, climate, and fire weather influence
fire severity. Less is known about how different forest management practices influence fire severity in
multi-owner landscapes, despite costly and controversial suppression of wildfires that do not acknowl-
edge ownership boundaries. In 2013, the Douglas Complex burned over 19,000 ha of Oregon & Cali-
fornia Railroad (O&C) lands in Southwestern Oregon, USA. O&C lands are composed of a
checkerboard of private industrial and federal forestland (Bureau of Land Management, BLM) with
contrasting management objectives, providing a unique experimental landscape to understand how
different management practices influence wildfire severity. Leveraging Landsat based estimates of fire
severity (Relative differenced Normalized Burn Ratio, RdNBR) and geospatial data on fire progres-
sion, weather, topography, pre-fire forest conditions, and land ownership, we asked (1) what is the rela-
tive importance of different variables driving fire severity, and (2) is intensive plantation forestry
associated with higher fire severity? Using Random Forest ensemble machine learning, we found daily
fire weather was the most important predictor of fire severity, followed by stand age and ownership,
followed by topographic features. Estimates of pre-fire forest biomass were not an important predictor
of fire severity. Adjusting for all other predictor variables in a general least squares model incorporat-
ing spatial autocorrelation, mean predicted RdNBR was higher on private industrial forests (RdNBR
521.85 18.67 [mean SE]) vs. BLM forests (398.87 18.23) with a much greater proportion of
older forests. Our findings suggest intensive plantation forestry characterized by young forests and
spatially homogenized fuels, rather than pre-fire biomass, were significant drivers of wildfire severity.
This has implications for perceptions of wildfire risk, shared fire management responsibilities, and
developing fire resilience for multiple objectives in multi-owner landscapes.
Key words: fire severity; forest management; Landsat; multi-owner landscape; Oregon; plantation forestry;
RdNBR.
INTRODUCTION
The wildfire environment has become increasingly compli-
cated, due to the unanticipated consequences of historical
forest management and fire exclusion (Weaver 1943, Hess-
burg et al. 2005, Ful
e et al. 2009, Naficy et al. 2010, Mer-
schel et al. 2014), an increasingly populated wildland urban
interface (Haas et al. 2013), and a rapidly changing climate
(Westerling and Bryant 2008, Littell et al. 2009, Jolly et al.
2015). These factors are resulting in more intense fire behav-
ior and increasingly negative ecological and social conse-
quences (Williams 2013, Stephens et al. 2014). Fuels
reduction via mechanical thinning and prescribed burning
have been the dominant land management response for miti-
gating these conditions (Agee and Skinner 2005, Stephens
et al. 2012), although there is an increasing recognition of
the need to manage wildfires more holistically to meet social
and ecological objectives. (North et al. 2015a, b). However,
overcoming these challenges is inhibited by numerous dis-
agreements in the scientific literature regarding historical
fire regimes and appropriate policies and management of
contemporary fire-prone forests (Hurteau et al. 2008, Han-
son et al. 2009, Spies et al. 2010, Campbell et al. 2012,
Odion et al. 2014, Collins et al. 2015, Stevens et al. 2016).
These factors and others have resulted in a nearly intractable
socioecological problem (Fischer et al. 2016); one that is
compounded by the fact that many fire-prone landscapes
consist of multiple owners and administrative jurisdictions
with varying and often conflicting land management
objectives.
Developing and prioritizing landscape fire management
activities (i.e., thinning, prescribed fire, wildland fire use,
and fire suppression) across jurisdictional and ownership
boundaries requires landscape-scale assessments of the fac-
tors driving fire severity (i.e., the fire behavior triangle of
fuels, topography, and weather). Researchers have focused
on the influence of bottom-up drivers such as topography
(Dillon et al. 2011, Prichard and Kennedy 2014, Birch et al.
2015), and fuels via fuel reduction effects (Agee and Skinner
2005, Raymond and Peterson 2005, Safford et al. 2009,
Prichard and Kennedy 2014, Ziegler et al. 2017), as well as
the top-down influence of weather on fire severity (Birch
et al. 2015, Estes et al. 2017). They have also focused more
broadly on how fire severity varies with vegetation and for-
est type (Birch et al. 2015, Steel et al. 2015, Reilly et al.
2017) and climate (Miller et al. 2012, Abatzoglou et al.
2017). While there is substantial value in further describing
how components of the fire behavior triangle influence fire
severity, we believe there is a need to account for these
known influences on fire behavior and effects to understand
Manuscript received 23 August 2017; revised 14 December 2017;
accepted 5 February 2018. Corresponding Editor: Bradford P.
Wilcox.
3
E-mail: hsz16@humboldt.edu
1068
Ecological Applications, 28(4), 2018, pp. 1068–1080
©2018 by the Ecological Society of America

how different management regimes interact with these con-
trolling factors, so appropriate landscape management
strategies can be developed to support social-ecological
resilience in fire-prone landscapes (Spies et al. 2014,
Schoennagel et al. 2017).
Understanding the relationships between forest manage-
ment regimes and fire severity is especially important in mul-
ti-owner landscapes, where wildfire governance systems
concerned about short-term property loss and public safety
can reinforce perceptions of wildfire risk and hazard, result-
ing in individual property owners being less likely to make
management decisions that reduce long-term risk exposure
(McCaffrey 2004, Fischer et al. 2016). This is particularly
important in landscapes that include intensive plantation
forestry, a common and rapidly expanding component of
forest landscapes at regional, national, and global scales
(Cohen et al. 1995, Landram 1996, Del Lungo et al. 2001,
Rudel 2009, FAO 2010, Nahuelhual et al. 2012). Research-
ers have hypothesized that intensive forest management
reduces fire behavior and effects (Hirsch et al. 2001,
Rodr

ıguez y Silva et al. 2014). However empirical results
have been mixed, with evidence that intensive forest manage-
ment can either reduce (Lyons-Tinsley and Peterson 2012,
Prichard and Kennedy 2014) or increase fire severity (Odion
et al. 2004, Thompson et al. 2007), and that reduced levels
of forest legal protection (a proxy for more active manage-
ment) have been associated with increased fire severity in the
western U.S. (Bradley et al. 2016). These conflicting results
further complicate the development of fire governance and
management strategies for increasing social-ecological resili-
ence in a rapidly changing fire environment.
The quality, spatial scale, and spatial correlation of
explanatory data (i.e., weather, topography, and fuels) are
major limitations to empirically understanding how forest
management activities influence fire severity across land-
scapes. Regional studies of fire severity often rely on spa-
tially coarse climatic data (Dillon et al. 2011, Miller et al.
2012, Cansler and McKenzie 2014, Kane et al. 2015, Harvey
et al. 2016, Meigs et al. 2016, Reilly et al. 2017), rather than
local fire weather that can be a significant driver of fire area
and severity (Flannigan et al. 1988, Bradstock et al. 2010,
Estes et al. 2017). This is in part because finer-scale fire
weather variables are often incomplete across the large spa-
tial and temporal domains of interest. Additionally, regional
studies often occur in areas with large elevation relief result-
ing in strong climatic gradients, while more local studies
often have less elevation relief and potentially weaker cli-
matic gradients. Perhaps more importantly, the geographic
distribution of different ownership types and management
regimes can confound quantification of the drivers of fire
severity. For example, high elevation forests in the Pacific
Northwest region of the United States are largely unman-
aged as National Parks and congressionally designated
wilderness areas, compared to intensively managed forests
at lower elevations, resulting in differences in topography,
weather, climate, forest composition, productivity, and his-
torical fire regimes between ownerships and management
regimes. While landscape studies of fire severity and man-
agement activities have used a variety of statistical tech-
niques to account for spatial correlation of both response
and predictor variables (Thompson et al. 2007, Prichard
and Kennedy 2014, Meigs et al. 2016), these techniques may
not overcome fundamental differences in response and pre-
dictor variables between management and/or ownership
types.
In this study, we examined the drivers of fire severity
within one large (~20,000 ha) wildfire complex that burned
within the Klamath Mountains, an ecoregion with a mild
Mediterranean climate of hot dry summers and wet winters
in southwestern Oregon, USA. The fire burned within a
checkerboard landscape of federal and private industrial for-
estry ownership. This spatial pattern of contrasting owner-
ship and management regimes provided a unique landscape
experiment where we quantified the effects of management
regimes after accounting for variation in well-known drivers
of fire behavior and effects. Leveraging geospatial data on
fire severity, fire progression, fire weather, topography, pre-
fire forest conditions, and past management activities, we
asked two questions: (1) What is the relative importance of
different variables driving fire severity? And (2) is intensive
plantation forestry associated with higher fire severity?
METHODS
Study site
In the summer of 2013, the Douglas Complex burned
19,760 ha of forestland in southwestern Oregon, USA
(Fig. 1). Starting from multiple lightning ignitions, individ-
ual small fires coalesced into two large fires (Dads Creek
and Rabbit Mountain) managed as the Douglas Complex.
FIG. 1. Location of and fire severity within the Douglas Com-
plex in Oregon, USA. Fire severity quantified using the Relative dif-
ferenced Normalized Burn Ratio (RdNBR).
June 2018 PLANTATION FORESTRY INCREASES SEVERITY 1069

This fire burned within the Oregon and California Railroad
Lands (hereafter O&C Lands). O&C Lands resulted from
19th century land grants that ceded every other square mile
(259 ha) of federally held land to railroad companies along
planned routes in Oregon and California to incentivize rail-
road development and homesteading settlement. The Ore-
gon and California Railroad Company received a total of
1.5 million ha, but failing to meet contractual obligations,
1.1 million ha were transferred back to federal ownership
under the Chamberlain-Ferris Revestment Act of 1916. The
USDI Bureau of Land Management (BLM) is currently
required to manage these lands for sustainable timber pro-
duction, watershed protection, recreation, and wildlife habi-
tat. Private industrial forestlands dominate the remaining
O&C landscape, and are managed intensively as native tree
plantations (primarily Douglas-fir, Pseudotsuga menziesii
var. menziesii) for timber production typically on 30–50 yr
harvest rotations. The Douglas Complex fires burned
10,201.64 ha of forests managed by the BLM, 9,429.66 ha
of private industrial forests, and 129.33 ha managed by the
Oregon Department of Forestry (ODF).
The Douglas Complex burned at elevations ranging from
213 to 1,188 m in mountainous terrain of the Klamath
Mountains Ecoregion. Climate in the ecoregion is character-
ized by hot dry summers and wet winters, with greater win-
ter precipitation at higher elevations and western portions of
the ecoregion. Vegetation types within the region include
oak woodlands and mixed hardwood/evergreen forests at
low to mid elevations, transitioning into mixed-conifer for-
ests at higher elevations (Franklin and Dyrness 1988). For-
ests within the Douglas Complex are dominated by
Douglas-fir, ponderosa pine (Pinus ponderosa), and white fir
(Abies concolor). Other conifer tree species present include
incense cedar (Calocedrus decurrens), sugar pine (Pinus lam-
bertiana), Jeffery pine (Pinus jefferyi), and knobcone pine
(Pinus attenuata). Hardwood species include Oregon white
oak (Quercus garryana), big-leaf maple (Acer macrophyl-
lum), Pacific dogwood (Cornus nuttallii), Pacific madrone
(Arbutus menziesii), canyon live oak (Quercus chrysolepis),
California black oak (Quercus kelloggii), golden chinkapin
(Chrysolepis chrysophylla), and tanoak (Lithocarpus densi-
flourus). Douglas-fir is the primary commercial timber spe-
cies managed on private and public lands, while fire
exclusion and historical management practices have
expanded the density and dominance of Douglas-fir across
much of the ecoregion (Franklin and Johnson 2012,
Sensenig et al. 2013).
Data sources
We analyzed fire severity in relation to eight predictor
variables representing topography, weather, forest owner-
ship, forest age, and pre-fire forest biomass (Fig. 2). We
quantified fire severity using the Relative differenced Nor-
malized Burn Ratio (RdNBR), a satellite-imagery-based
metric of pre- to post-fire change. Cloud-free pre-fire (3 July
2013) and post-fire (7 July 2014) images came from the
Landsat 8 Operational Land Imager. Normalized Burn
Ratio (NBR), which combines near-infrared and mid-infra-
red bands of Landsat imagery, was calculated for pre- and
post-fire images. Differenced Normalized Burn Ratio
(dNBR) was calculated by subtracting NBR
post-fire
from
NBR
pre-fire
values, and RdNBR was then calculated follow-
ing Miller et al. (2009), where:
RdNBR ¼dNBR
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Absolute Value ðNBRprefire=1;000Þ
p. (1)
We chose RdNBR over dNBR as our fire severity metric
because RdNBR removes, at least in part, the biasing effect
of pre-fire conditions, improving assessment of burn severity
across heterogeneous vegetation and variable pre-fire distur-
bances (Miller and Thode 2007). We used the continuous
RdNBR values as our response variable for fire severity at a
30-m resolution.
Elevation and other topographic variables were derived
from the National Elevation Dataset 30 m digital elevation
model (Gesch et al. 2002). We generated 30-m rasters of ele-
vation (m), slope (%), topographic position index (TPI), and
heat load (MJcm
2
yr
1
). TPI was calculated as the differ-
ence between elevation in a given cell and mean elevation of
cells within an annulus around that cell, calculated at fine
and coarse scales (TPI fine and TPI coarse) with 150–300 m
and 1,850–2,000 m annuli, respectively. We also originally
considered TPI at a moderate spatial scale (850–1,000 m
annuli), but rejected it as an predictor variable due to its
high correlation to TPI fine (r=0.64) and TPI course
(r=0.84). TPI course had strong linear correlations with
elevation (r=0.83) and TPI fine (r=0.46), so it was also
removed to avoid multi-collinearity in statistical analyses.
Heat load was calculated by least-squares multiple regres-
sion using trigonometric functions of slope, aspect, and lati-
tude following McCune and Keon (2002).
Rasters of daily fire weather conditions were generated by
extrapolating weather station data to a daily fire progression
map. We obtained hourly weather data for the duration of
active fire spread (7 July–20 August 2013) from the Calvert
Peak Remote Automatic Weather Station (NWS ID 352919;
42°46040″N 123°43046″W, 1,165 m), approximately 30 km
west-southwest of the Douglas Complex. We then subset
each 24-h period of weather data to the daily burn period
(10:00 to 18:00) when fire behavior is typically most active.
We then calculated the daily burn period minimum wind
speed (km/h), maximum temperature (°C), and minimum
relative humidity (%). For each daily burn period we also
calculated the mean energy release component (ERC),
spread component (SC), and burning index (BI) using
FireFamilyPlus Version 4.1 (Bradshaw and McCormick
2000). ERC is an index of fuel dryness related to the maxi-
mum energy release at the flaming front of a fire, as mea-
sured from temperature, relative humidity, and moisture of
1–1,000 h dead fuels. SC is a rating of the forward rate of
spread of a head fire, and is calculated from wind speed,
slope, and moisture of live fine and woody fuels (Bradshaw
et al. 1983). BI is proportional to the flame length at the
head of a fire (Bradshaw et al. 1983), calculated using ERC
and SC, thus incorporating wind speed and providing more
information than ERC and SC individually. ERC, SC, and
BI vary by broadly categorized fuel types. We calculated
ERC, SC, and BI using the National Fire Danger Rating
System Fuel Model G, which represents short-needled
1070 HAROLD S. J. ZALD AND CHRISTOPHER J. DUNN Ecological Applications
Vol. 28, No. 4

conifer stands with heavy dead fuel loads. Daily fire weather
variables were then spatially extrapolated to the daily area
burned based on daily fire progression geospatial data cap-
tured during the fire (GeoMAC 2013).
Forest ownership was derived from geospatial data repre-
senting fee land title and ownership in Oregon (Oregon Spa-
tial Data Library 2015). We grouped ODF and BLM lands
as a single ownership type, because ODF lands were a small
component of the area burned and have management objec-
tives closer to federal vs. private industrial forests (Spies
et al. 2007). Pre-fire forest conditions were represented with
30-m rasters of live biomass (Mg/ha) and stand age, derived
from a regional 2012 map of forest composition and struc-
tural attributes developed for the Northwest Forest Plan
Monitoring Program (Ohmann et al. 2012, Davis et al.
2015). These maps were developed using the gradient nearest
neighbor method (GNN), relating multivariate response
variables of forest composition and structure attributes from
approximately 17,000 federal forest inventory plots to grid-
ded predictor variables (satellite imagery, topography, cli-
mate, etc.) using canonical correspondence analysis and
nearest neighbor imputation (Ohmann and Gregory 2002).
Biomass values are directly from the GNN maps, while we
quantified forest age as a two-step process. First, we calcu-
lated pre-fire forest age in 2013 based on years since each
pixel was disturbed in the Landsat time series (1985–2014)
from a regional disturbance map generated for the North-
west Forest Plan Monitoring Program using the LandTrendr
segmentation algorithm (Kennedy et al. 2010, Ohmann
et al. 2012, Davis et al. 2015). Second, for pixels where no
FIG. 2. Maps of response and predictor variables for Douglas Complex. TPI, topographic position index.
June 2018 PLANTATION FORESTRY INCREASES SEVERITY 1071

disturbance had occurred within the Landsat time series, we
amended forest age derived from the Landsat time series
using dominant and codominant tree age from the GNN
maps.
Statistical analyses
All statistical analyses were conducted in the R statisti-
cal environment version 3.3.3 (R Development Core Team
2017). We sampled the burned landscape using a spatially
constrained stratified random design, from which response
and predictor variables were extracted for analysis. Sample
points had to be at least 200 m apart to minimize short
distance spatial autocorrelation of response and predictor
variables. Our choice of minimum inter-plot distance to
reduce spatial autocorrelation was confounded by the
dominance of long distance spatial autocorrelation driven
by large ownership patches, which would have greatly
reduced sample size and potentially eliminated finer scale
variability in the sample. For these reasons we based our
200 m minimum inter-plot distance in part on prior
research (Kane et al. 2015), that found residual spatial
autocorrelation in Random Forest models of fire severity
in the Rim Fire of 2013 in the California Sierra Nevada
was greatly diminished when inter-plot distances were at
least 180 m apart. Additionally, point locations had to be
at least 100 m away from ownership boundaries to mini-
mize inter-ownership edge effects. Within these spatial
constraints, sample points were located in a stratified ran-
dom design, with the number of points proportional to
area of ownership within the fire perimeter, resulting in
571 and 519 points located in BLM and private industrial
forests, respectively. Mean response and predictor variables
were extracted within a 90 990 m plot (e.g., 3 93 pixels)
centered on each sample point location to minimize the
effects of potential georeferencing errors across data layers
and maintain a plot size comparable to the original inven-
tory plots used as source data in GNN maps as recom-
mended by Bell et al. (2015).
We observed high correlation between fire weather vari-
ables (mean absolute r=0.59), likely due to their temporal
autocorrelation during the fire event, which could result in
multi-collinearity in statistical analyses. Therefore, we evalu-
ated the relationships between each fire weather variable
and daily mean fire severity, selecting a single fire weather
variable as a predictor variable in subsequent analyses. We
based our variable selection on visual relationships to daily
RdNBR, variance explained in regressions of RdNBR and
fire weather variables, and Akaike information criterion
(AIC) scores of regressions of RdNBR and fire weather vari-
ables following Burnham and Anderson (2002).
The study’s strength rests in part on the implicit assump-
tion that the checkerboard spatial allocation of ownership
types is a landscape scale experiment, where predictor vari-
ables directly modified by management activities (e.g., pre-
fire biomass and forest age) are different between ownership
types, but fire weather and topographic variables are not.
We assessed this assumption by visualizing data distribu-
tions between ownerships using boxplots and violin plots,
and testing if variables were different between ownership
types using Mann–Whitney–Wilcoxon Tests.
To assess the relative importance and relationships
between predictor variables and RdNBR, we used Random
Forest (RF) supervised machine learning algorithm with the
randomForest package (Liaw and Wiener 2002). As applied
in this study, RF selected 1,500 bootstrap samples, each con-
taining two-thirds of the sampled cells. For each sample, RF
generated a regression tree, then randomly selected only
one-third of the predictor variables and chose the best parti-
tion from among those variables. To assess the relative
importance and relationships of predictor variables on
RdNBR across the entire study area and within different
ownerships, separate RF models were developed for all
1,090 sample plots across the entire burned area, as well as
separately for plots on BLM and private industrial lands.
For each of the three RF models, we calculated variable
importance values for each predictor variable as the percent
increase in the mean squared error (MSE) in the predicted
data when values for that predictor were permuted and all
other predictors were left unaltered. In addition to variable
importance values, we determined which predictor variables
should be retained in each RF model using multi-stage vari-
able selection procedures (Genuer et al. 2010). We applied
two-stage variable selection for interpretation to each RF
model using the VSURF package (Genuer et al. 2016).
Final RF models were then run including only the selected
variables. Predictive power of the final RF models were
assessed by calculating the variance explained, which is
equivalent to the coefficient of determination (R
2
) used with
linear regressions to assess statistical model fit for a given
dataset. Last, we visualized the relationships of individual
predictor variables on RdNBR in the final RF models using
partial dependency plots (Hastie et al. 2001).
Importance values in RF models are not the same as
quantifying the fixed effects of predictor variables, nor is
RF well suited to explicitly test hypotheses or quantify
effects of predictor variables while accounting for other vari-
ables in a model. To test if ownership type increased
RdNBR, we developed a generalized least squares (GLS)
regression model with an exponential spherical spatial corre-
lation structure using the nlme package (Pinheiro et al.
2017). The GLS regression used the distance between sam-
ple locations and the form of the correlation structure to
derive a variance–covariance matrix, which was then used to
solve a weighted OLS regression (Dormann et al. 2007).
Using the same response and predictor data as in the RF
model for the entire Douglas Complex, and a binary predic-
tor variable for ownership type, we developed a GLS model
from which we calculated the fixed effect of ownership on
RdNBR. We then predicted the mean and standard error of
RdNBR by ownership after accounting for the other predic-
tor variables in the GLS model using the AICcmodavg
package (Mazerolle 2017).
RESULTS
Fire weather variables
Regression models of fire weather variables (except maxi-
mum temperature) described a significant proportion of the
variance in daily mean RdNBR (Table 1; Appendix S1:
Fig. S1). SC described the most variance in daily RdNBR,
1072 HAROLD S. J. ZALD AND CHRISTOPHER J. DUNN Ecological Applications
Vol. 28, No. 4

had the lowest AIC score, and was most likely to be the best
model of those compared (w
i
=0.8250). However, BI
described a comparable amount of the variance in daily
RdNBR (R
2
=0.5815), had a substantial level of empirical
support (DAIC =3.3816), was the second most likely model
given the data (w
i
=0.1521), and contained additional
metrics that influence fire behavior (influence of temperature,
relative humidity, and drought on live and dead fuels) not
incorporated in SC. For these reasons, we choose to use BI as
the single fire weather variable in subsequent analyses,
acknowledging that it may describe slightly less variation in
RdNBR than SC.
RdNBR and predictor variable differences by ownership
The majority of predictor variables were not statistically
different by ownership, as expected given the spatial distri-
bution of ownership. Based on Mann-Whitney-Wilcoxon
tests, biomass and stand age were lower on private industrial
vs. BLM managed lands (Table 2; Appendix S1: Fig. S2).
TPI fine, heat load, slope, and BI were not different between
ownership types. Elevation was different between ownership
types, but only 44 m higher on BLM land across a range of
875 m for all sample plots. Mean RdNBR was higher
(536.56 vs. 408.75) on private industrial vs. BLM lands.
Random forest variable importance values and partial
dependency plots
Two-stage variable selection procedures retained seven,
five, and six predictor variables in the final RF models for
the entire Douglas Complex, BLM, and private forests,
respectively (Fig. 3). Across the entire Douglas Complex, BI
was the most important predictor variable of RdNBR
(increasing MSE by 138.4%), while BI was also the most
importance variable separately for BLM (105.4%) and pri-
vate forests (83.2%). Age and ownership were the next most
TABLE 1. Regression models of daily mean Relative differenced
Normalized Burn Ratio (RdNBR) in relation to daily burn
period fire weather variables.
Models R
2
AIC DAIC L(g
i
|x)w
i
RdNBR =SC
2
0.6532 210.0324 0.0000 1.0000 0.8250
RdNBR =BI
2
0.5815 213.4140 3.3816 0.1844 0.1521
RdNBR =min
wind speed
2
0.4542 218.1948 8.1624 0.0169 0.0139
RdNBR =log
(min relative
RH)
0.3800 220.4903 10.4579 0.0054 0.0044
RdNBR =ERC
2
0.3675 220.8497 10.8173 0.0045 0.0037
RdNBR =max
wind speed
2
0.2179 224.6700 14.6376 0.0007 0.0005
RdNBR =max
temperature
2
0.1069 227.0592 17.0268 0.0002 0.0002
RdNBR =null
model
0.0000 228.1855 18.1531 0.0001 0.0001
Notes: R
2
, adjusted Rsquared; AIC
c
, Akaike information crite-
rion corrected for sample size; DAIC
c
, AIC
c
differences; L(g
i
|x),
likelihood of a model given the data; w
i
, Akaike weights; SC, spread
component; BI, burn index; RH, relative humidity; ERC, energy
release component.
TABLE 2. RdNBR (mean with SE in parentheses) and predictor variables on sampled plots for Bureau of Land Management (BLM) vs.
private industrial (PI) ownership.
Variable BLM PI wP
RdNBR 408.75 (298.53) 536.56 (299.88) 111,124 <0.0001
Biomass (Mg/ha) 234.75 (87.24) 163.88 (74.47) 215,166 <0.0001
Age (yr) 108.81 (55.53) 52.18 (36.78) 236,021.5 <0.0001
BI (index) 62.99 (14.16) 63.64 (14.54) 142,575.5 0.2782
Elevation (m) 653.79 (153.48) 609.46 (161.62) 171,200 <0.0001
TPI fine 0.55 (32.51) 1.08 (32.12) 152,275 0.4296
Heat load (MJcm
2
yr
1
) 0.77 (0.2) 0.77 (0.2) 150,363 0.6734
Slope (%) 48.4 (13.4) 47.05 (14.01) 156,435 0.1115
Notes: The wvalues and associated Pvalues are from Mann–Whitney–Wilcoxon tests. TPI, topographic position index.
FIG. 3. Variable importance plots for predictor variables from Random Forest (RF) models of RdNBR for 1090 sample plots across the
entire Douglas Complex (left panel), 571 plots on Bureau of Land Management (BLM) forests (middle), and 519 plots on private industrial
(PI) forests (right). Solid circles denote variables retained in two-stage variable selection, open circles denote variables removed from the
final RF models during variable selection. BI, burning index; MSE, Mean Squared Error. [Correction added on May 1st 2018, after first
online publication: The x axis label was incorrectly labeled as “MSF”]
June 2018 PLANTATION FORESTRY INCREASES SEVERITY 1073

FIG. 4. Partial dependency plots showing relationships between each predictor variable and RdNBR in random forest models for all forests (BLM and PI, top panels), forests on Bureau of Land
Management (BLM, middle panels), and private industrial land (PI, bottom panels). Number within each panel shows variable importance (VI; mean squared error increase [%]) of each predictor in
the random forest model. Solid lines show trends in RdNBR in response to each predictor, histograms show the distributions of values for each predictor. Note there is no partial dependency plot
for the relationship between RdNBR and biomass for BLM forests, as biomass was not a significant predictor variable for BLM forests based on two-stage variable selection procedures.
1074 HAROLD S. J. ZALD AND CHRISTOPHER J. DUNN Ecological Applications
Vol. 28, No. 4

important predictor variables, increasing MSE across the
Douglas Complex by 56.7% and 53.2%, respectively. Age
was the second most important variable in the final RF
model for BLM forests (32%), but was the fourth most
important variable for private forests (18.2%). Pre-fire bio-
mass was the fourth most importance predictor variable in
the RF model of the entire Douglas Complex (33.9%), but
was not retained in the final RF model for BLM forests, and
was the least important variable (10.3%) in the final RF
model for private forests. Overall, topographic variables (TPI
fine, heat load, and slope) were less important than BI, own-
ership, and age, increasing MSE across the Douglas Complex
by 2.6–36.5%. RF models described 31%, 23%, and 25% of
the variability in RdNBR across the entire burned area,
BLM managed forests, and private forests, respectively.
Partial dependency plots displayed clear relationships
between RdNBR and predictor variables (Fig. 4). RdNBR
increased exponentially with BI across the entire Douglas
Complex as well as for BLM and private forests separately,
although RdNBR was shifted up by approximately 100
RdNBR on private forests vs. BLM forests for any given BI
value. RdNBR was consistently higher in young forests on
both ownerships. RdNBR declined rapidly on BLM forests
between stand ages of 20 and 80 yr old, and remained
roughly level in older forests. In contrast, RdNBR in private
forests declined linearly with age across its range, although
private lands had few forests greater than 100 yr old. RdNBR
on both BLM and private forests increased with higher eleva-
tions, higher TPI fine, and steeper slope. Heat load was nega-
tively correlated with RdNBR for all ownerships. Pre-fire
biomass was not included in the final RF model for BLM
lands, while, for the entire study and private lands, RdNBR
appeared to decline slightly in forests with intermediate pre-
fire biomass. However, the relationship between RdNBR and
pre-fire biomass is more tenuous on private lands because
they lacked forests with high pre-fire biomass.
Generalize least squares model
BI, age, ownership, TPI fine, and heat load were all signif-
icant predictors of RdNBR in the GLS model (Table 3).
Slope had a suggestive relation with RdNBR (P=0.0586),
while elevation (P=0.1769) and pre-fire biomass
(P=0.2911) were not a significant predictors. Relationships
between predictors and RdNBR were consistent with partial
dependency plots from RF models, with RdNBR increasing
with BI and TPI fine and declining with age and heat load.
Ownership had a fixed effect of increasing mean RdNBR by
76.36 22.11 (mean SE) in private vs. BLM. Adjusting
for all other predictor variables in the model, predicted
mean RdNBR was higher on private (521.85 18.67) vs.
BLM forests (398.87 18.23).
DISCUSSION
Quantifying fire severity in the unique checkerboard land-
scape of the O&C Lands, this study disentangled the effects
of forest management, weather, topography, and biomass on
fire severity that are often spatially confounded. We found
daily fire weather was the most important predictor of fire
severity, but ownership, forest age, and topography were also
important. After accounting for fire weather, topography,
stand age, and pre-fire biomass, intensively managed private
industrial forests burned at higher severity than older federal
forests managed by the BLM. Below we discuss how the dif-
ferent variables in our analysis may influence fire severity,
and argue that younger forests with spatially homogenized
continuous fuel arrangements, rather than absolute biomass,
was a significant driver of wildfire severity. The geospatial
data available for our analyses was robust and comprehen-
sive, covering two components of the fire behavior triangle
(i.e., topography, weather), with pre-fire biomass and age
serving as proxies for the third (fuel). However, we recognize
there are limitations to our data and analyses and describe
these below. We conclude by suggesting how our findings
have important implications for forest and fire management
in multi-owner landscapes, while posing important new
questions that arise from our findings.
Fire weather was a strong top-down driver of fire sever-
ity, while bottom-up drivers such as topography and
pre-fire biomass were less important. Across the western
United States, evidence suggests bottom-up drivers such as
topography and vegetation exert greater control on fire
severity than weather, although the quality of weather rep-
resentation confounds this conclusion (Dillon et al. 2011,
Birch et al. 2015). At the same time, it is recognized that
bottom-up drivers of fire severity can be overwhelmed by
top-down climatic and weather conditions when fires burn
during extreme weather conditions (Bradstock et al. 2010,
Thompson and Spies 2010, Dillon et al. 2011). Daily burn
period BI values were used in our analyses, but it is impor-
tant to place fire weather conditions for any single fire
within a larger historical context. We compared these daily
BI values to the historical (1991–2017) summer (1 June–30
September) BI data we calculated from the Calvert RAWS
data used in this study (3,296 total days). Within this his-
torical record, mean burn period BI during the Douglas
Complex for days with fire progression information was
above average (79th percentile), but ranged considerably for
any given day of the fire (15th–100th percentile). Fire sever-
ity was consistently higher on private lands across a range
of fire weather conditions for the majority of days of active
fire spread (Appendix S1: Fig. S3), leading us to conclude
that while fire weather exerted top-down control on fire
severity, local forest conditions that differed between own-
erships remained important, even during extreme fire
weather conditions.
TABLE 3. Coefficients of predictor variables in generalized least
squares model of RdNBR.
Variable Parameter estimate SE tP
Intercept 80.3321 90.4529 0.8881 0.3747
Age 1.0544 0.2132 4.9452 <0.0001
BI 6.1413 0.7618 8.0614 <0.0001
Ownership 76.3559 22.1111 3.4533 0.0006
Elevation 0.1179 0.0872 1.3512 0.1769
TPI fine 1.2839 0.2509 5.1169 <0.0001
Heat load 150.0098 39.5750 3.7905 0.0002
Slope 1.1321 0.5979 1.8933 0.0586
Biomass 0.1261 0.1194 1.0562 0.2911
June 2018 PLANTATION FORESTRY INCREASES SEVERITY 1075

Variation in pre-fire forest conditions across ownerships
were clearly a significant driver of fire severity, and we
believe they operated at multiple spatial scales. Private
industrial forests were dominated by young trees, which have
thinner bark and lower crown heights, both factors known
to increase fire-induced tree mortality (Ryan and Reinhardt
1988, Dunn and Bailey 2016). At the stand scale, these plan-
tations are high-density single cohorts often on harvest rota-
tions between 30 and 50 yr, resulting in dense and relatively
spatially homogenous fuel structure. In contrast, public for-
ests were dominated by older forests that tend to have
greater variability in both tree size and spatial pattern vs.
plantations (Naficy et al. 2010), arising from variable natu-
ral regeneration (Donato et al. 2011), post-disturbance bio-
logical legacies (Seidl et al. 2014), and developmental
processes in later stages of stand development (Franklin
et al. 2002). Fine-scale spatial patterns of fuels can signifi-
cantly alter fire behavior, and the effects of spatial patterns
on fire behavior may increase with the spatial scale of
heterogeneity (Parsons et al. 2017), which would likely be
the case in O&C Lands due to the large scale checkerboard
spatial pattern of ownership types.
Management-driven changes in fuel spatial patterns at
tree and stand scales could also reconcile differences in
prior studies that have found increases (Odion et al. 2004,
Thompson et al. 2007) and decreases (Prichard and Ken-
nedy 2014) in fire severity with intensive forest manage-
ment. The two studies that observed an increase in fire
severity with intensive forest management were conducted
in the Klamath ecoregion of southwestern Oregon and
northwestern California, the same ecoregion as this study.
In contrast, Prichard and Kennedy (2014) examined the
Tripod Complex in north-central Washington State, where
harvests mostly occurred in low to mid elevation forests
dominated by ponderosa pine, Douglas-fir, lodgepole pine
(Pinus contorta var. latifolia), western larch (Larix occiden-
talis), and Engelmann spruce (Picea engelmannii). These
forests have lower productivity compared to those studied
in the Klamath ecoregion, with more open canopies and
longer time periods to reach canopy closure after harvest,
which likely results in more heterogeneous within stand
fuel spatial patterns. Furthermore, forest clearcut units
were relatively small in the Tripod Complex (mean 53 ha;
Prichard and Kennedy 2014), and while these harvest
units were spatially clustered, they were not large contigu-
ous blocks as found in the O&C Lands. Last, it is unclear
if the harvest units evaluated by Prichard and Kennedy
(2014) experienced the full distribution of fire weather or
topographic conditions compared to unharvested units, as
our study does, which may confound their conclusions
and our understanding of the relative importance of the
factors driving fire behavior and effects.
LIMITATIONS
Our study examined a landscape uniquely suited to disen-
tangling the drivers of wildfire severity and quantifying the
effects of contrasting management activities. Additionally,
we leveraged a robust collection of geospatial data to quan-
tify the components of the fire behavior triangle. However, it
is important to recognize the inherent limitations of our
study. First, this study represents a single fire complex,
instead of a regional collection of fires analyzed to elucidate
broader system behaviors (sensu Dillon et al. 2011, Birch
et al. 2015, Meigs et al. 2016). However, given the chal-
lenges of obtaining high quality fire weather information
and accurate daily fire progression maps for fires that have
occurred in landscapes with contrasting management
regimes, we believe the landscape setting of our study pro-
vides key insights into the effects of management on fire
severity that are not possible in large regional multi-fire
studies. Second, while Landsat imagery is widely used to
estimate forest conditions and fire severity, it has specific
limitations. The GNN maps used in this study to derive pre-
fire biomass and stand age are strongly driven by multi-spec-
tral imagery from the Landsat family of sensors, whose ima-
gery is known to saturate in forests with high leaf area
indices and high biomass (Turner et al. 1999). Third, GNN
maps of forest attributes used in this study were originally
developed for large regional assessments, and as such have
distinct limitations when used for analyses at spatial resolu-
tions finer than the original source data (Bell et al. 2015),
while application of GNN at fine spatial scales can underes-
timate GNN accuracy compared to larger areas commonly
used by land managers (Ohmann et al. 2014). We addressed
potential limitations of using GNN predictions at fine spa-
tial scales in two ways. First, our sample plots are 90-m
squares (3 93 30 m pixels) which more closely represents
the area of the inventory plots used as GNN source data
compared to pixel level analyses (Bell et al. 2015). Second,
we visually assessed GNN predictions of live biomass and
stand age within the Douglas Complex in relation to high
resolution digital orthoimagery collected in 2011 by the
USDA National Agriculture Imagery Program. From this
qualitative assessment we concluded that GNN predictions
characterize both between and within ownership variation
in pre-fire biomass and age (Appendix S1: Fig. S4). Fourth
and perhaps most fundamentally important, we relied on
pre-fire biomass and stand age as proxies for fuel, in part
because Landsat and other passive optical sensors have lim-
ited sensitivity to vertical and below-canopy vegetation
structure (Lu 2006). Accurate and spatially complete quanti-
tative information of forest surface and canopy fuels were
not available for the Douglas Complex. More broadly, there
are significant limitations to spatial predictions of forest
structure and fuels using GNN and other methods that rely
on passive optical imagery such as Landsat (Keane et al.
2001, Pierce et al. 2009, Zald et al. 2014), which is why we
relied on the more accurately predicted age and pre-fire bio-
mass variables as proxies. Surface and ladder fuels are the
most important contributors to fire behavior in general
(Agee and Skinner 2005), and surface fuels have been found
to be positively correlated to fire severity in plantations
within the geographic vicinity of the Douglas Complex
(Weatherspoon and Skinner 1995). Yet correlations between
biomass and fuel load can be highly variable due to site con-
ditions and disturbance history (i.e., mature forests with fre-
quent surface fires may have high live biomass but low
surface fuel loads, while dense young forests that have regen-
erated after a stand replacing wildfire will have low live bio-
mass but potentially high surface fuel loads as branches and
snags fall). Therefore, GNN predicted pre-fire biomass may
1076 HAROLD S. J. ZALD AND CHRISTOPHER J. DUNN Ecological Applications
Vol. 28, No. 4

represent the total fuel load, but not the available surface
and ladder fuels that have the potential to burn during a
specific fire, and this is supported by the low importance of
pre-fire biomass as a predictor of fire severity in our study.
Furthermore, it is important to recognize that in addition to
total surface and ladder fuels, the spatial continuity of these
fuels strongly influences fire behavior (Rothermel 1972,
Pimont et al. 2011). Fifth, while private industrial and BLM
forests in our study area had very different forest conditions
due to contrasting management regimes, ownership alone
misses management activities (e.g., site preparation, stock-
ing density, competing vegetation control, partial thinning,
etc.) that can influence fuels and fire behavior. Sixth, while
our spatial extrapolation of fire weather correlated well with
daily fire severity and area burned, it did not account for
topographic mediation of weather that can influence fine
scale fire behavior, nor did it examine the underlying
weather patterns such as temperature inversions that are
common to the region and may play a key role in moderat-
ing burning index (Estes et al. 2017). Finally, we were unable
to discern the effects of fire suppression activities and
whether they varied by ownership, since incident documen-
tation of suppression activities are generally not collected or
maintained in a manner consistent with quantitative or
geospatial statistical analyses (Dunn et al. 2017).
MANAGEMENT IMPLICATIONS
Although only one fire complex, the contrasting forest
conditions resulting from different ownerships within the
Douglas Complex are consistent with many mixed-owner-
ship or mixed-use landscapes, such that we believe our
results have implications across a much broader geographic
area. First, it brings into question the conventional view that
fire exclusion in older forests is the dominant driver of fire
severity across landscapes. There is strong scientific agree-
ment that fire suppression has increased the probability of
high severity fire in many fire-prone landscapes (Miller et al.
2009, Calkin et al. 2015, Reilly et al. 2017), and thinning as
well as the reintroduction of fire as an ecosystem process are
critical to reducing fire severity and promoting ecosystem
resilience and adaptive capacity (Agee and Skinner 2005,
Raymond and Peterson 2005, Earles et al. 2014, Krofcheck
et al. 2017). However, in the landscape we studied, intensive
plantation forestry appears to have a greater impact on fire
severity than decades of fire exclusion. Second, higher fire
severity in plantations potentially flips the perceived risk
and hazard in multi-owner landscapes, because higher sever-
ity fire on intensively managed private lands implies they are
the greater source of risk than older forests on federal lands.
These older forests likely now experience higher fire severity
than historically due to decades of fire exclusion, yet in com-
parison to intensively managed plantations, the effects of
decades of fire exclusion in older forests appear to be less
important than increased severity in young intensively man-
aged plantations on private industrial lands.
Furthermore, our findings suggest challenges and opportu-
nities for managing intensive plantations in ways that reduce
potential fire severity. Increasing the age (and therefore size)
of trees and promoting spatial heterogeneity of stands and
fuels is a likely means to reducing fire severity, as are fuel
reduction treatments in plantations (Crecente-Campo et al.
2009, Kobziar et al. 2009, Reiner et al. 2009). The extent and
spatial arrangement of fuel reduction treatments can be an
important consideration in their efficacy at reducing fire
severity at landscape scales (Finney et al. 2007, Krofcheck
et al. 2017). However, optimal extent and landscape patterns
of fuels reduction treatments can be hampered by a wide
range of ecological, economic, and administrative constraints
(Collins et al. 2010, North et al. 2015a, Barros et al. 2017).
In the past, pre-commercial and commercial thinning of
plantations (a potential fuel treatment) in the Pacific North-
west were common, economically beneficial management
activities that improved tree growth rates and size, but these
practices have become less common with improved reforesta-
tion success, alternative vegetation control techniques, and
shorter harvest rotations (Talbert and Marshall 2005). This
suggests there may be strong economic limitations to
increased rotation ages and non-commercial thinning in
young intensive plantation forests. More broadly, the devel-
opment of large-scale forest management and conservation
strategies can face legal and equitability challenges in multi-
owner landscapes given existing laws constraining planning
among private organizations (Thompson et al. 2004, 2006).
We believe two major questions arise from our findings
that are important to fire management in multi-owner land-
scapes, especially those with contrasting management objec-
tives. Plantations burned at higher severity, and this implies
they are a higher source of risk to adjacent forest owner-
ships. However, a more explicit quantification of fire severity
and susceptibility is needed to understand how risk is spa-
tially transmitted across ownership types under a variety of
environmental conditions. Second, we suggest the need for
alternative management strategies in plantations to reduce
fire severity at stand and landscape scales. However, the eco-
nomic viability of such alternative management regimes
remains poorly understood. Optimization models integrat-
ing spatial allocation of fuel treatments and fire behavior
with economic models of forest harvest and operations
could be used to determine if alternative management activi-
ties in plantations are economically viable. If alternative
management activities are not economically viable, but wild-
fire risk reduction is an important objective on lands adja-
cent to industrial forestlands, strategic land purchases or
transfers between ownership types may be required to
achieve landscape level goals. This may be particularly
important given the previously stated legal and equitability
challenges in multi-owner landscapes. Regardless of the
landscape-level objectives and constraints, it is clear that
cooperation among stakeholders will be necessary in multi-
ownership landscapes if wildfire risk reduction, timber har-
vesting, and conservation objectives remain dominant yet
sometimes conflicting objectives for these landscapes.
ACKNOWLEDGMENTS
Funding for this research was in part provided by the USDI
Bureau of Land Management (Cooperative Agreement no.
L11AC20137/L01540). We thank Krissan Kosel at the USDI BLM
Roseburg District for assistance providing Calvert Station RAWS
weather data, as well as thoughtful review and discussions of prior
versions of this manuscript. We also thank two reviewers for their
helpful suggestions on earlier versions of this paper for their insight-
ful and constructive comments.
June 2018 PLANTATION FORESTRY INCREASES SEVERITY 1077

LITERATURE CITED
Abatzoglou, J. T., C. A. Kolden, A. P. Williams, J. A. Lutz, and A.
M. S. Smith. 2017. Climatic influences on interannual variability
in regional burn severity across western US forests. International
Journal of Wildland Fire 26:269.
Agee, J. K., and C. N. Skinner. 2005. Basic principles of forest fuel
reduction treatments. Forest Ecology and Management 211:83–
96.
Barros, A. M. G., et al. 2017. Spatiotemporal dynamics of simu-
lated wildfire, forest management, and forest succession in central
Oregon, USA. Ecology and Society 22:24.
Bell, D. M., M. J. Gregory, H. M. Roberts, R. J. David, and J. L.
Ohmann. 2015. How sampling and scale limit accuracy assess-
ment of vegetation maps: a comment on Loehle et al. (2015). For-
est Ecology and Management 358:361–364.
Birch, D. S., P. Morgan, C. A. Kolden, J. T. Abatzoglou, G. K. Dil-
lon, A. T. Hudak, and A. M. S. Smith. 2015. Vegetation, topogra-
phy and daily weather influenced burn severity in central Idaho
and western Montana forests. Ecosphere 6:art17.
Bradley, C. M., C. T. Hanson, and D. A. DellaSala. 2016. Does
increased forest protection correspond to higher fire severity in
frequent-fire forests of the western United States? Ecosphere 7:
e01492.
Bradshaw, L., J. Deeming, R. Burgan, and J. Cohen. 1983. The 1978
national fire-danger rating system. General Technical Report
INT-169. US Department of Agriculture, Forest Service, Inter-
mountain Forest and Range Experiment Station, Ogden, Utah,
USA.
Bradshaw, L., and E. McCormick. 2000. FireFamily Plus user’s
guide, version 2.0. Gen. Tech. Rep. RMRS-GTR-67WWW. U.S.
Department of Agriculture, Forest Service, Rocky Mountain
Research Station, Ogden, Utah, USA.
Bradstock, R. A., K. A. Hammill, L. Collins, and O. Price. 2010.
Effects of weather, fuel and terrain on fire severity in topographi-
cally diverse landscapes of south-eastern Australia. Landscape
Ecology 25:607–619.
Burnham, K. P., and D. R. Anderson. 2002. Model selection and
multi-model inference: a practical information-theoretic
approach. Springer Verlag, New York, New York, USA.
Calkin, D. E., M. P. Thompson, and M. A. Finney. 2015. Negative
consequences of positive feedbacks in US wildfire management.
Forest Ecosystems 2:9.
Campbell, J. L., M. E. Harmon, and S. R. Mitchell. 2012. Can fuel-
reduction treatments really increase forest carbon storage in the
western US by reducing future fire emissions? Frontiers in Ecol-
ogy and the Environment 10:83–90.
Cansler, C. A., and D. McKenzie. 2014. Climate, fire size, and bio-
physical setting control fire severity and spatial pattern in the
northern Cascade Range, USA. Ecological Applications 24:1037–
1056.
Cohen, W. B., T. A. Spies, and M. Fiorella. 1995. Estimating the age
and structure of forests in a multi-ownership landscape of western
Oregon, USA. International journal of Remote Sensing 16:721–746.
Collins, B. M., J. M. Lydersen, R. G. Everett, D. L. Fry, and S. L.
Stephens. 2015. Novel characterization of landscape-level vari-
ability in historical vegetation structure. Ecological Applications
25:1167–1174.
Collins, B. M., S. L. Stephens, J. J. Moghaddas, and J. J. Battles.
2010. Challenges and approaches in planning fuel treatments
across fire-excluded forested landscapes. Journal of Forestry
108:24–31.
Crecente-Campo, F., A. Pommerening, and R. Rodr

ıguez-Soalleiro.
2009. Impacts of thinning on structure, growth and risk of crown
fire in a Pinus sylvestris L. plantation in northern Spain. Forest
Ecology and Management 257:1945–1954.
Davis, R. J., et al. 2015. Northwest Forest Plan–the first 20 years
(1994–2013): status and trends of late-successional and old-
growth forests. U.S. Department of Agriculture, Forest Service,
Pacific Northwest Research Station, Portland, Oregon, USA.
Del Lungo, L., P. Vourinen, and J. Carle. 2001. Preliminary analysis
of global trends in forest plantation development, 1980–2000.
FAO, Rome, Italy.
Dillon, G. K., Z. A. Holden, P. Morgan, M. A. Crimmins, E. K.
Heyerdahl, and C. H. Luce. 2011. Both topography and climate
affected forest and woodland burn severity in two regions of the
western US, 1984 to 2006. Ecosphere 2:art130.
Donato, D. C., J. L. Campbell, and J. F. Franklin. 2011. Multiple
successional pathways and precocity in forest development: Can
some forests be born complex? Journal of Vegetation Science
23:576–584.
Dormann, C., et al. 2007. Methods to account for spatial autocor-
relation in the analysis of species distributional data: a review.
Ecography 30:609–628.
Dunn, C. J., and J. D. Bailey. 2016. Tree mortality and structural
change following mixed-severity fire in Pseudotsuga forests of
Oregon’s western Cascades, USA. Forest Ecology and Manage-
ment 365:107–118.
Dunn, C. J., D. E. Calkin, and M. P. Thompson. 2017. Towards
enhanced risk management: planning, decision making and moni-
toring of US wildfire response. International Journal of Wildland
Fire 26:551–556.
Earles, J. M., M. P. North, and M. D. Hurteau. 2014. Wildfire and
drought dynamics destabilize carbon stores of fire-suppressed for-
ests. Ecological Applications 24:732–740.
Estes, B. L., E. E. Knapp, C. N. Skinner, J. D. Miller, and H. K. Pre-
isler. 2017. Factors influencing fire severity under moderate burn-
ing conditions in the Klamath Mountains, northern California,
USA. Ecosphere 8:e01794.
Finney, M. A., R. C. Seli, C. W. McHugh, A. A. Ager, B. Bahro,
and J. K. Agee. 2007. Simulation of long-term landscape-level fuel
treatment effects on large wildfires. International Journal of Wild-
land Fire 16:712.
Fischer, A. P., et al. 2016. Wildfire risk as a socioecological pathol-
ogy. Frontiers in Ecology and the Environment 14:276–284.
Flannigan, M. D., J. B. Harrington, M. D. Flannigan, and J. B.
Harrington. 1988. A study of the relation of meteorological
variables to monthly provincial area burned by wildfire in Canada
(1953–80). Journal of Applied Meteorology 27:441–452.
Food and Agriculture Organization of the United Nations [FAO].
2010. Global forest resource assessment. FAO, Rome, Italy.
Franklin, J. F., and C. T. Dyrness. 1988. Natural vegetation of Ore-
gon and Washington. Oregon State University Press, Corvallis,
Oregon, USA.
Franklin, J. F., and K. N. Johnson. 2012. A restoration framework
for federal forests in the pacific northwest. Journal of Forestry
110:429–439.
Franklin, J. F., T. A. Spies, R. Van Pelt, A. B. Carey, D. A. Thorn-
burgh, D. R. Berg, D. B. Lindenmayer, M. E. Harmon, W. S. Kee-
ton, and D. C. Shaw. 2002. Disturbances and structural
development of natural forest ecosystems with silvicultural impli-
cations, using Douglas-fir forests as an example. Forest Ecology
and Management 155:399–423.
Ful
e, P. Z., J. E. Korb, and R. Wu. 2009. Changes in forest structure
of a mixed conifer forest, southwestern Colorado, USA. Forest
Ecology and Management 258:1200–1210.
Genuer, R., J.-M. Poggi, and C. Tuleau-Malot. 2010. Variable selection
using random forests. Pattern Recognition Letters 31:2225–2236.
Genuer, R., J.-M. Poggi, and C. Tuleau-Malot. 2016. VSURF:
variable selection using random forests. R package version 1.0.3.
https://cran.r-project.org/web/packages/VSURF/index.html
Geospatial Multi-Agency Coordination (GeoMAC). 2013. Douglas
complex fire perimeters. https://rmgsc.cr.usgs.gov/outgoing/Ge
oMAC/2013_fire_data/Oregon/Douglas_Complex/
Gesch, D. B., M. J. Oimoen, S. K. Greenlee, C. A. Nelson, M. J.
Steuck, and D. J. Tyler. 2002. The national elevation data set.
Photogrammetric Engineering and Remote Sensing 68:
5–11.
Haas, J. R., D. E. Calkin, and M. P. Thompson. 2013. A national
approach for integrating wildfire simulation modeling into
1078 HAROLD S. J. ZALD AND CHRISTOPHER J. DUNN Ecological Applications
Vol. 28, No. 4

Wildland Urban Interface risk assessments within the United
States. Landscape and Urban Planning 119:44–53.
Hanson, C. T., D. C. Odion, D. A. DelleaSala, and W. L. Baker.
2009. Overestimation of fire risk in the Northern Spotted Owl
recovery plan. Conservation Biology 23:1314–1319.
Harvey, B. J., D. C. Donato, and M. G. Turner. 2016. Drivers and
trends in landscape patterns of stand-replacing fire in forests of
the US Northern Rocky Mountains (1984–2010). Landscape
Ecology 31:2367–2383.
Hastie, T., R. Tibshirani, J. Friedman, and J. Franklin. 2001. The
elements of statistical learning: data mining, inference, and pre-
diction. Springer, New York, New York, USA.
Hessburg, P. F., J. K. Agee, and J. F. Franklin. 2005. Dry forests and
wildland fires of the inland Northwest USA: contrasting the land-
scape ecology of the pre-settlement and modern eras. Forest Ecol-
ogy and Management 211:117–139.
Hirsch, K., V. Kafka, C. Tymstra, R. McAlpine, B. Hawkes, H.
Stegehuis, S. Quintilio, S. Gauthier, and K. Peck. 2001. Fire-smart
forest management: a pragmatic approach to sustainable forest
management in fire-dominated ecosystems. Forestry Chronicle
77:357–363.
Hurteau, M. D., G. W. Koch, and B. A. Hungate. 2008. Carbon pro-
tection and fire risk reduction: toward a full accounting of forest
carbon offsets. Frontiers in Ecology and the Environment 6:493–
498.
Jolly, W. M., M. A. Cochrane, P. H. Freeborn, Z. A. Holden, T. J.
Brown, G. J. Williamson, and D. M. Bowman. 2015. Climate-
induced variations in global wildfire danger from 1979 to 2013.
Nature Communications 6:7537.
Kane, V. R., C. A. Cansler, N. A. Povak, J. T. Kane, R. J.
McGaughey, J. A. Lutz, D. J. Churchill, and M. P. North. 2015.
Mixed severity fire effects within the Rim fire: relative importance
of local climate, fire weather, topography, and forest structure.
Forest Ecology and Management 358:62–79.
Keane, R. E., R. Burgan, and J. van Wagtendonk. 2001. Mapping
wildland fuels for fire management across multiple scales: inte-
grating remote sensing, GIS, and biophysical modeling. Interna-
tional Journal of Wildland Fire 10:301–319.
Kennedy, R. E., Z. Yang, and W. B. Cohen. 2010. Detecting trends
in forest disturbance and recovery using yearly Landsat time ser-
ies: 1. LandTrendr—temporal segmentation algorithms. Remote
Sensing of Environment 114:2897–2910.
Kobziar, L. N., J. R. McBride, and S. L. Stephens. 2009. The effi-
cacy of fire and fuels reduction treatments in a Sierra Nevada pine
plantation. International Journal of Wildland Fire 18:791.
Krofcheck, D. J., M. D. Hurteau, R. M. Scheller, and E. L. Louder-
milk. 2017. Prioritizing forest fuels treatments based on the prob-
ability if high-severity fire restores adaptive capacity in Sierran
forests. Global Change Biology. https://doi.org/10.1111/gcb.13913
Landram, M. 1996. Status of reforestation on National Forest lands
within the Sierra Nevada Ecosystem Project Study Area. Page
Sierra Nevada Ecosystem Project: Final Report to Congress (Vol-
ume 3). University of California, Centers for Water and Wildland
Resources, Davis, CA USA.
Liaw, A., and M. Wiener. 2002. Classification and regression by ran-
domForest. R News 2:18–22.
Littell, J. S., D. McKenzie, D. L. Peterson, and A. L. Westerling.
2009. Climate and wildfire area burned in western US eco-
provinces, 1916–2003. Ecological Applications 19:1003–1021.
Lu, D. 2006. The potential and challenge of remote sensing-based
biomass estimation. International Journal of Remote Sensing
27:1297–1328.
Lyons-Tinsley, C., and D. L. Peterson. 2012. Surface fuel treatments
in young, regenerating stands affect wildfire severity in a mixed
conifer forest, eastside Cascade Range, Washington, USA. Forest
Ecology and Management 270:117–125.
Mazerolle, M. J. 2017. AICcmodavg: model selection and multi-
model inference based on (Q) AIC (c). R package version, 2.1.
https://cran.r-project.org/web/packages/AICcmodavg/index.html
McCaffrey, S. 2004. Thinking of wildfire as a natural hazard. Soci-
ety and Natural Resources 17:509–516.
McCune, B., and D. Keon. 2002. Equations for potential annual
direct incident radiation and heat load. Journal of Vegetation
Science 13:603.
Meigs, G. W., H. S. Zald, J. L. Campbell, W. S. Keeton, and R. E.
Kennedy. 2016. Do insect outbreaks reduce the severity of subse-
quent forest fires? Environmental Research Letters 11:045008.
Merschel, A. G., T. A. Spies, and E. K. Heyerdahl. 2014. Mixed-
conifer forests of central Oregon: effects of logging and fire exclu-
sion vary with environment. Ecological Applications 24:
1670–1688.
Miller, J. D., H. D. Safford, M. Crimmins, and A. E. Thode. 2009.
Quantitative evidence for increasing forest fire severity in the
Sierra Nevada and southern Cascade Mountains, California and
Nevada, USA. Ecosystems 12:16–32.
Miller, J. D., C. N. Skinner, H. D. Safford, E. E. Knapp, and C. M.
Ramirez. 2012. Trends and causes of severity, size, and number of
fires in northwestern California, USA. Ecological Applications
22:184–203.
Miller, J. D., and A. E. Thode. 2007. Quantifying burn severity in a
heterogeneous landscape with a relative version of the delta nor-
malized burn ratio (dNBR). Remote Sensing of Environment
109:66–80.
Naficy, C., A. Sala, E. G. Keeling, J. Graham, and T. H. DeLuca.
2010. Interactive effects of historical logging and fire exclusion on
ponderosa pine forest structure in the northern Rockies. Ecologi-
cal Applications 20:1851–1864.
Nahuelhual, L., A. Carmona, A. Lara, C. Echeverr

ıa, and M. E.
Gonz
alez. 2012. Land-cover change to forest plantations: proxi-
mate causes and implications for the landscape in south-central
Chile. Landscape and Urban Planning 107:12–20.
North, M. P., S. L. Stephens, B. M. Collins, J. K. Agee, G. Aplet, J.
F. Franklin, and P. Z. Ful
e. 2015a. Reform forest fire manage-
ment. Science 349:1280–1281.
North, M., A. Brough, J. Long, B. Collins, P. Bowden, D. Yasuda, J.
Miller, and N. Sugihara. 2015b. Constraints on mechanized treat-
ment significantly limit mechanical fuels reduction extent in the
Sierra Nevada. Journal of Forestry 113:40–48.
Odion, D. C., E. J. Frost, J. R. Strittholt, H. Jiang, D. A. DellaSala,
and M. A. Moritz. 2004. Patterns of fire severity and forest condi-
tions in the western Klamath Mountains, California. Conserva-
tion Biology 18:927–936.
Odion, D. C., et al. 2014. Examining historical and current mixed-
severity fire regimes in Ponderosa pine and mixed-conifer forests
of western North America. PLoS ONE 9:e87852.
Ohmann, J. L., and M. J. Gregory. 2002. Predictive mapping of for-
est composition and structure with direct gradient analysis and
nearest-neighbor imputation in coastal Oregon, USA. Canadian
Journal of Forest Research 32:725–741.
Ohmann, J. L., M. J. Gregory, and H. M. Roberts. 2014. Scale con-
siderations for integrating forest inventory plot data and satellite
image data for regional forest mapping. Remote Sensing of Envi-
ronment 151:3–15.
Ohmann, J. L., M. J. Gregory, H. M. Roberts, W. B. Cohen, R. E.
Kennedy, and Z. Yang. 2012. Mapping change of older forest with
nearest-neighbor imputation and Landsat time-series. Forest
Ecology and Management 272:13–25.
Oregon Spatial Data Library. 2015. Oregon land management—
2015. http://spatialdata.oregonexplorer.info/geoportal/catalog/sea
rch/resource/details.page?uuid=%7B9B644E0F-7A7D-4124-A50F-
6B35C05626AE%7D
Parsons, R., R. Linn, F. Pimont, C. Hoffman, J. Sauer, J. Win-
terkamp, C. Sieg, and W. Jolly. 2017. Numerical investigation of
aggregated fuel spatial pattern impacts on fire behavior. Land
6:43.
Pierce, K. B., J. L. Ohmann, M. C. Wimberly, M. J. Gregory, and J.
S. Fried. 2009. Mapping wildland fuels and forest structure for
land management: a comparison of nearest neighbor imputation
June 2018 PLANTATION FORESTRY INCREASES SEVERITY 1079

and other methods. Canadian Journal of Forest Research
39:1901–1916.
Pimont, F., J.-L. Dupuy, R. R. Linn, and S. Dupont. 2011.
Impacts of tree canopy structure on wind flows and fire propa-
gation simulated with FIRETEC. Annals of Forest Science
68:523–530.
Pinheiro, J., D. Bates, S. DebRoy, D. Sarkar, and R. C. Team. 2017.
nlme: linear and nonlinear mixed effects models. https://cran.
r-project.org/web/packages/nlme/index.html
Prichard, S. J., and M. C. Kennedy. 2014. Fuel treatments and land-
form modify landscape patterns of burn severity in an extreme
fire event. Ecological Applications 24:571–590.
R Core Team. 2017. R: a language and environment for statistical
computing. R Foundation for Statistical Computing, Vienna,
Austria.
Raymond, C. L., and D. L. Peterson. 2005. Fuel treatments alter the
effects of wildfire in a mixed-evergreen forest, Oregon, USA.
Canadian Journal of Forest Research 35:2981–2995.
Reilly, M. J., C. J. Dunn, G. W. Meigs, T. A. Spies, R. E. Kennedy, J.
D. Bailey, and K. Briggs. 2017. Contemporary patterns of fire
extent and severity in forests of the Pacific Northwest, USA
(1985–2010). Ecosphere 8:e01695.
Reiner, A. L., N. M. Vaillant, J. Fites-Kaufman, and S. N. Dailey.
2009. Mastication and prescribed fire impacts on fuels in a 25-
year old ponderosa pine plantation, southern Sierra Nevada. For-
est Ecology and Management 258:2365–2372.
Rodr
ıguez y Silva, F., J. R. Molina Mart
ınez, and A. Gonz
alez-
Cab
an. 2014. A methodology for determining operational
priorities for prevention and suppression of wildland fires. Inter-
national Journal of Wildland Fire 23:544.
Rothermel, R. 1972. A mathematical model for predicting fire
spread in wildland fuels. Page 40. Research Paper INT-115,
USDA Forest Service, Intermountain Forest and Range Experi-
ment Station, Ogden, Utah USA.
Rudel, T. K. 2009. Tree farms: driving forces and regional patterns
in the global expansion of forest plantations. Land Use Policy
26:545–550.
Ryan, K. C., and E. D. Reinhardt. 1988. Predicting postfire mortal-
ity of seven western conifers. Canadian Journal of Forest
Research 18:1291–1297.
Safford, H. D., D. A. Schmidt, and C. H. Carlson. 2009. Effects of
fuel treatments on fire severity in an area of wildland–urban inter-
face, Angora Fire, Lake Tahoe Basin, California. Forest Ecology
and Management 258:773–787.
Schoennagel, T., et al. 2017. Adapt to more wildfire in western
North American forests as climate changes. Proceedings of the
National Academy of Sciences USA 114:4582–4590.
Seidl, R., W. Rammer, and T. A. Spies. 2014. Disturbance legacies
increase the resilience of forest ecosystem structure, composition,
and functioning. Ecological Applications 24:2063–2077.
Sensenig, T., J. D. Bailey, and J. C. Tappeiner. 2013. Stand development,
fire and growth of old-growth and young forests in southwestern Ore-
gon, USA. Forest Ecology and Management 291:96–109.
Spies, T. A., K. N. Johnson, K. M. Burnett, J. L. Ohmann, B. C.
McComb, G. H. Reeves, P. Bettinger, J. D. Kline, and B. Garber-
Yonts. 2007. Cumulative ecological and socioeconomic effects of
forest policies in coastal Oregon. Ecological Applications 17:5–17.
Spies, T. A., et al. 2010. Underestimating risks to the Northern
Spotted Owl in fire-prone forests: response to Hanson et al.
Conservation Biology 24:330–333.
Spies, T., et al. 2014. Examining fire-prone forest landscapes as cou-
pled human and natural systems. Ecology and Society 19:9.
http://dx.doi.org/10.5751/ES-06584-190309
Steel, Z. L., H. D. Safford, and J. H. Viers. 2015. The fire frequency-
severity relationship and the legacy of fire suppression in Califor-
nia forests. Ecosphere 6:1–23.
Stephens, S., R. Boerner, C. Fettig, J. Fontaine, B. Hartsough,
P. Kennedy, and D. Schwilk. 2012. The effects of forest fuel-
reduction treatments in the United States. BioScience 62:549–
560.
Stephens, S. L., et al. 2014. Temperate and boreal forest mega-fires:
characteristics and challenges. Frontiers in Ecology and the Envi-
ronment 12:115–122.
Stevens, J. T., et al. 2016. Average stand age from forest inventory
plots does not describe historical fire regimes in Ponderosa pine
and mixed-conifer forests of western North America. PLoS ONE
11:e0147688.
Talbert, C., and D. Marshall. 2005. Plantation productivity in the
Douglas-fir region under intensive silvicultural practices: results
from research and operations. Journal of Forestry 103:65–70.
Thompson, J. R., M. D. Anderson, and K. N. Johnson. 2004.
Ecosystem management across ownerships: the potential for colli-
sion with antitrust laws. Conservation Biology 18:1475–1481.
Thompson, J. R., K. N. Johnson, M. Lennette, T. A. Spies, and P.
Bettinger. 2006. Historical disturbance regimes as a reference for
forest policy in a multiowner province: a simulation experiment.
Canadian Journal of Forest Research 36:401–417.
Thompson, J. R., and T. A. Spies. 2010. Factors associated with
crown damage following recurring mixed-severity wildfires and
post-fire management in southwestern Oregon. Landscape Ecol-
ogy 25:775–789.
Thompson, J. R., T. A. Spies, and L. M. Ganio. 2007. Reburn sever-
ity in managed and unmanaged vegetation in a large wildfire.
Proceedings of the National Academy of Sciences USA
104:10743–10748.
Turner, D. P., W. B. Cohen, R. E. Kennedy, K. S. Fassnacht, and J.
M. Briggs. 1999. Relationships between leaf area index and Land-
sat TM spectral vegetation indices across three temperate zone
sites. Remote Sensing of Environment 70:52–68.
Weatherspoon, C. P., and C. N. Skinner. 1995. An assessment of fac-
tors associated with damage to tree crowns from the 1987 wild-
fires in northern California. Forest Science 41:430–451.
Weaver, H. 1943. Fire as an ecological and silvicultural factor in the
Ponderosa-pine region of the pacific Slope. Journal of Forestry
41:7–15.
Westerling, A., and B. Bryant. 2008. Climate change and wildfire in
California. Climatic Change 87:231–249.
Williams, J. 2013. Exploring the onset of high-impact mega-fires
through a forest land management prism. Forest Ecology and
Management 294:4–10.
Zald, H. S. J., J. L. Ohmann, H. M. Roberts, M. J. Gregory, E. B.
Henderson, R. J. McGaughey, and J. Braaten. 2014. Influence of
lidar, Landsat imagery, disturbance history, plot location
accuracy, and plot size on accuracy of imputation maps of forest
composition and structure. Remote Sensing of Environment
143:26–38.
Ziegler, J. P., C. Hoffman, M. Battaglia, and W. Mell. 2017. Spatially
explicit measurements of forest structure and fire behavior follow-
ing restoration treatments in dry forests. Forest Ecology and
Management 386:1–12.
SUPPORTING INFORMATION
Additional supporting information may be found online at: http://onlinelibrary.wiley.com/doi/10.1002/eap.1710/full
DATA AVAILABILITY
Data available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.3gv5c78
1080 HAROLD S. J. ZALD AND CHRISTOPHER J. DUNN Ecological Applications
Vol. 28, No. 4