Modeling a Historical Mountain Pine Beetle Outbreak Using Landsat MSS and Multiple Lines of Evidence

Full text

!
1!
!
Modeling a Historical Mountain Pine Beetle Outbreak Using Landsat MSS and Multiple
Lines of Evidence
TIMOTHY J. ASSALa, b*, JASON SIBOLDc, and ROBIN REICHd
aU.S. Geological Survey (USGS), Fort Collins Science Center, 2150 Centre Avenue, Fort
Collins, CO 80526,
bGraduate Degree Program in Ecology, Colorado State University, 1401 Campus Delivery, Fort
Collins, CO 80523, USA
cDepartment of Anthropology, Colorado State University, 1787 Campus Delivery, Fort Collins,
CO 80523, USA
dDepartment of Forest and Rangeland Stewardship, Colorado State University, 1472 Campus
Delivery, Fort Collins, CO 80523, USA
*Corresponding author at: USGS, Fort Collins Science Center, Fort Collins, CO 80526, USA.
Tel.: +1 970 226 9134.
E-mail address: assalt@usgs.gov (T.J. Assal).
Pre-print of published version.
Reference:
Assal, T.J., Sibold, J., and R. Reich. 2014. Modeling a Historical Mountain Pine Beetle
Outbreak Using Landsat MSS and Multiple Lines of Evidence. Remote Sensing of Environment
155:275-288.
https://www.sciencedirect.com/science/article/pii/S0034425714003435
Disclaimer:
The PDF document is a copy of the final version of this manuscript that was subsequently
accepted by the journal for publication. The paper has been through peer review, but has not
been subject to any additional copy-editing or journal specific formatting (so will look different
from the final version of record, which may be accessed following the DOI above depending on
your access situation)

!
2!
!
ABSTRACT
Mountain pine beetles are significant forest disturbance agents, capable of inducing widespread
mortality in coniferous forests in western North America. Various remote sensing approaches
have assessed the impacts of beetle outbreaks over the last two decades. However, few studies
have addressed the impacts of historical mountain pine beetle outbreaks, including the 1970s
event that impacted Glacier National Park. The lack of spatially explicit data on this disturbance
represents both a major data gap and a critical research challenge in that wildfire has removed
some of the evidence from the landscape. We utilized multiple lines of evidence to model forest
canopy mortality as a proxy for outbreak severity. We incorporate historical aerial and landscape
photos, aerial detection survey data, a nine-year collection of satellite imagery and abiotic data.
This study presents a remote sensing based framework to (1) relate measurements of canopy
mortality from fine-scale aerial photography to coarse-scale multispectral imagery and (2)
classify the severity of mountain pine beetle affected areas using a temporal sequence of Landsat
data and other landscape variables. We sampled canopy mortality in 261 plots from aerial photos
and found that insect effects on mortality were evident in changes to the Normalized Difference
Vegetation Index (NDVI) over time. We tested multiple spectral indices and found that a
combination of NDVI and the green band resulted in the strongest model. We report a two-step
process where we utilize a generalized least squares model to account for the large-scale
variability in the data and a binary regression tree to describe the small-scale variability. The
final model had a root mean square error estimate of 9.8% canopy mortality, a mean absolute
error of 7.6% and an R2 of 0.82. The results demonstrate that a model of percent canopy
mortality as a continuous variable can be developed to identify a gradient of mountain pine
beetle severity on the landscape.

!
3!
!
1. Introduction
Temperate forest ecosystems are subject to various ecological disturbances that can have
profound effects on the structure of the ecosystem for many years after the event (Turner & Dale,
1998) and influence the likelihood, severity and spread of subsequent disturbances (Veblen et al.,
1994). In western North America, native bark beetles are a major disturbance agent capable of
regional-scale forest mortality (Raffa et al., 2008). Remotely sensed imagery has been used to
characterize such widespread disturbance events over the last two decades (Wulder et al., 2006a).
However, very little research has employed these techniques to study insect disturbance prior to
the recent period of extended outbreak (~pre late 1990s). The northern Rocky Mountains
experienced a widespread mountain pine beetle outbreak in the late 1970s to early 1980s (Logan
& Powell, 2001). However, the lack of spatially explicit data on the extent and severity of this
outbreak limits our understanding of the influence that this disturbance had on the landscape. To
overcome this challenge, we utilized multiple lines of evidence to retrospectively characterize
forest canopy mortality from the outbreak by comparing temporal changes in archived satellite
imagery.
1.1 Mountain Pine Beetle Overview
The mountain pine beetle (Dendroctonus ponderosae) is a native species found in the
western United States and Canada that attacks and reproduces in live trees (Bentz et al., 2010).
The mechanisms with which populations switch to epidemic levels are complex (Bentz et al.,
2010; Raffa et al., 2008), but include suitable host availability (amount, vigor, age and density)
and condition (Fettig et al., 2007), along with beetle population survival and growth given
thermal conditions (Powell & Logan, 2005). Epidemic populations are capable of landscape-
scale forest mortality leading to cascading effects on forest structure, species composition and

!
4!
!
function (Raffa et al., 2008). Major host species include lodgepole pine (Pinus contorta),
ponderosa pine (P. ponderosa), and whitebark pine (P. albicaulis) (Bentz et al., 2010). Impacted
forests exhibit unique and visible characteristics at each stage of a mountain pine beetle attack
(Wulder et al., 2006a). Killed trees begin to show visible changes as the foliage changes from
green to yellow to red over the first year after the attack. The gray attack stage typically
commences three years after the attack, as most trees will have lost all needles at that time
(Wulder et al., 2006a).
1.2 Remote Sensing and Disturbance
Historical aerial photography is a valuable research tool providing detailed records of forest
landscapes over the last half century or more. Although limited in spatial extent, these records
provide a fine-scale snapshot of landscapes at one or multiple points in time. Previous studies
have successfully used aerial photos collected during two or more time periods to measure
changes in tree cover (Brown et al., 2006; Di Orio et al., 2005; Kadmon & Harari-Kremer, 1999;
Kennedy & Spies, 2004; Manier et al., 2005; Platt & Schoennagel, 2009; Strand et al., 2006).
The use of satellite multispectral imagery to map and monitor forest condition over larger
regions is also well documented (Cohen et al., 2001; Maselli, 2004; Nemani et al., 2009;
Schroeder et al., 2006; Townshend et al., 2012; Volcani et al., 2005) dating back to the early
1970s with the initiation of the Landsat program (NASA, 2013). Several studies have used aerial
photos as a surrogate for field data collection and then used that information to scale up to
satellite imagery. This technique has been accomplished to map various attributes including land
cover type (Parmenter et al., 2003), tree cover (Carreiras et al., 2006; Cohen et al., 2001; Homer
et al., 2007), and surface imperviousness (Homer et al., 2007). Photos can be used to sample
post-disturbance forest patterns, such as canopy mortality. The aerial photo reference data can be

!
5!
!
used to bridge the gap in scale between localized tree mortality measures and the more coarse
scale of satellite imagery (Meddens et al., 2013). This hybrid approach allows for detection of
fine-scale disturbance patterns captured in the aerial photos, while taking advantage of the
multispectral and multitemporal components of Landsat imagery at the landscape scale.
Furthermore, it provides a pathway to conduct a retrospective analysis.
Ecological disturbance alters ecosystem structure by both abrupt, conspicuous change and by
gradual, slow change over some period of time. Such impacts allow remote sensing to capture
the pre- and post-landscape, and in some cases, the duration of the event. Aerial photos have
been utilized to investigate the impacts of fire (Bebi et al., 2003; Johnson & Fryer, 1987), insect
damage (Bebi et al., 2003; White et al., 2005), extreme drought (Allen & Breshears, 1998), and
blowdown (Baker et al., 2002) on forest and woodland ecosystems. At regional scales,
multispectral satellite imagery has been employed to study diverse types of forest disturbance
including fragmentation (Fuller, 2001), fire (Turner et al., 1994), drought (Huang et al., 2010)
and insect induced mortality (DeRose et al., 2011; Vogelmann et al., 2009). Numerous studies
have utilized multispectral imagery to document the extent and severity of the recent mountain
pine beetle outbreak over the last decade. Efforts range from fine-scale satellite and aerial
multispectral imagery acquired from one time period (Coops et al., 2006; Dennison et al., 2010;
Hicke & Logan, 2009; Meddens et al., 2011), to moderate resolution sensors incorporating
multiple time periods (Goodwin et al., 2008; Meddens et al., 2013; Meigs et al., 2011; Wulder et
al., 2006b).
We found few studies in the literature that used the first generation of Landsat data to detect
mountain pine beetle outbreaks or other insect-driven forest disturbance. The Landsat
Multispectral Scanner System (MSS) sensor was carried onboard the first five Landsat satellites

!
6!
!
and provided imagery from 1972 until 1995 (NASA, 2013). Researchers in British Columbia
(Harris et al., 1978) used single date MSS imagery to detect damage caused by the Douglas-fir
tussock moth and western spruce budworm with little success. Weber et al. (1975) employed
single date MSS imagery to map mountain pine beetle damage in Ponderosa pine. Rencz and
Nemeth (1985) tested both a single date approach and a change detection approach over a six-
year period to map mountain pine beetle damage in British Columbia. Both mountain pine beetle
studies concluded that MSS imagery does not have the capability to detect beetle damage given
the spatial resolution of the imagery. However, the British Columbia study (Rencz & Nemeth,
1985) noted greater detection accuracy at sites with heavy, continuous damage, suggesting the
spatial resolution is less limiting in areas with high-severity outbreaks.
1.3 Outbreak Impacts to Forest Vegetation Spectral Properties
Living vegetation absorbs blue and red light energy, while radiation in the green and near-
infrared portion of the electromagnetic spectrum is reflected (Jones & Vaughan, 2010).
Therefore, color-infrared photos can be used to distinguish between areas of live trees and dead
trees. As the foliage of killed trees changes during the first year after the attack, the spectral
response also begins to change (Rencz & Nemeth, 1985). At the cellular level, mortality
contributes to a reduction in foliar moisture and chlorophyll, as other pigments and cellular
structure begin to break down (Mauseth, 1988). As a result, the spectral reflectance in the red
wavelength (630-690 nm) increases, whereas the reflectance in the green wavelength (520-600
nm) decreases (Ahern, 1988).
Disturbances where large portions of forest vegetation are removed from the landscape, such
as fire and clear cutting, create a drastic change in spectral reflectance. Conversely, subtle
changes in foliage color over time may prove more difficult to detect. Nevertheless, the

!
7!
!
phenology associated with mortality caused by an outbreak will lead to a change in satellite-
detected reflectance of the forest canopy. An analysis of multiple years of moderate spatial
resolution imagery has the potential to capture reflectance patterns before, during and after
landscape-scale disturbance events (Goodwin et al., 2008; Wulder et al., 2006a).
Multiple types of spectral indices have been employed to detect the impacts of mountain pine
beetle disturbance over the last decade. Examples of indices include the Normalized Difference
Moisture Index (Goodwin et al., 2008, 2010; Meddens et al., 2013), the Tasseled Cap (Meddens
et al., 2013), the Enhanced Wetness Disturbance Index (Skakun et al., 2003; Wulder et al.,
2006b), the Normalized Burn Ratio (Meigs et al., 2011), the Red-Green Index (RGI) (Coops et
al., 2006; Hicke & Logan, 2009; Meddens et al., 2013), the Band 5/Band 4 Ratio (Meddens et
al., 2013), and the Normalized Difference Vegetation Index (Meddens et al., 2013). Various
levels of success were obtained with each index. Many of these indices are derived from Landsat
TM or ETM+ imagery. However, Landsat TM imagery is not available prior to 1984 (for the
study area) and Landsat ETM+ imagery is not available before 1999. Because the outbreak that
is the focus of this study erupted in the mid-1970s, Landsat MSS imagery represents the only
available satellite imagery. Given the four multispectral bands of MSS (Table 1), we were only
able to utilize a subset of these indices.
1.4 Aerial Detection Survey Data
The US Forest Service (USFS) has been conducting annual forest health aerial detection
surveys (ADS) since the middle of the 20th century. In summary, human observers record the
type and extent of abiotic and biotic disturbances and host species onto sketch maps (Meigs et
al., 2011). The sketch maps are hard copy maps used by human observers in planes that are later
converted to digital form. This data has successfully been integrated into remote sensing

!
8!
!
detection studies of insect disturbance (Meddens et al., 2012; Meigs et al., 2011). The Forest
Health Protection Aviation Program in USFS Region 1 (including Glacier National Park (GNP))
maintains digital files of the ADS data since 2000. Staff at GNP digitized the ADS data from
1962-1998. The data include information about insect species, host tree species, damage type,
and forest type. However, very few polygons contained information on the number of trees killed
per acre (severity), which is commonly included in contemporary ADS data and is critical to
relating outbreaks to forest processes and change. Furthermore, the disturbance polygons
identified in the ADS data were very large (e.g. > 70,000 ha). Although useful for broad-scale
monitoring, we suspect the ADS data do not represent the heterogeneous impacts of the
disturbance. Since we are interested in both the extent and severity of the disturbance, these
missing details heavily influenced the direction of this study.
1.5 Objectives
The goal of the study was to test an approach combining multiple lines of evidence to
reconstruct the extent and severity of a mountain pine beetle outbreak in a topographically
complex landscape. Furthermore, subsequent disturbance (fire) has removed evidence from large
areas of the study area. To accomplish this, we used a combination of aerial detection survey
data, historical aerial and landscape photos, National Park Service reports and a temporal
sequence of satellite imagery. Each data source has limitations in the spatiotemporal record.
However, by combining disparate sources of data across spatial and temporal scales, we aimed to
reduce the uncertainty associated with reconstructing outbreak parameters. Employing multiple
lines of evidence from independent data sources has the potential to extend the information
associated with each piece of data and create a robust composite picture of the outbreak
(Swetnam et al., 1999). Reference data was collected from aerial photos and scaled up to satellite

!
9!
!
imagery measurements over time. We hypothesized that the impacts of the disturbance to the
forest canopy (i.e. mortality) would be captured in spatiotemporal changes in reflectance. Finally
we sought to demonstrate a novel approach in the use of existing data to assess a historic
disturbance.
The objectives of this study are to:
1. Relate measurements of canopy mortality from fine-scale aerial photography to coarse-
scale multispectral imagery;
2. Classify the severity of mountain pine beetle affected areas using a temporal sequence
of Landsat data and other landscape variables.
2. Methods
2.1 Study Area
The study was located in Glacier National Park in northwestern Montana, USA (Figure 1)
and chosen because of the extensive mountain pine beetle epidemic that occurred there in the
1970s (Hamel et al., 1977; McGregor et al., 1975). The park encompasses 4,080 km2 (408,000
ha) of diverse terrain on either side of the Continental Divide. Mean average annual precipitation
is 73.1 cm, and average annual maximum and minimum temperatures are 11.9 ºC and -0.2 ºC,
respectively (1971-2000) (Western Regional Climate Center, West Glacier station, elevation:
970 m, http://www.wrcc.dri.edu; accessed 17 December 2012). The climate averages from this
station are consistent with stations on the east side of the park. Elevation ranges from ~ 950 m to
3184 m above sea level and major cover types include grasslands, conifer and deciduous forests,
lakes, wide glacial valleys and steep alpine zones. Forests are dominated by lodgepole pine
(Pinus contorta), western larch (Larix occidentalis), Engelmann spruce (Picea engelmannii) and
Douglas-fir (Pseudotsuga menziesii).

!
10!
!
Given the size and diverse landscape of the park, we limited the study area based on several
assumptions. First, vegetation cover types not susceptible to mountain pine beetle attack were
identified using ReGAP (Davidson et al., 2009) and omitted. Second, we calculated the
cumulative extent of mountain pine beetle damage identified by the ADS data between 1971 and
1987. The area not impacted by the mountain pine beetle outbreak during the buffered time
period was omitted from further analysis. The area of interest was also confined by the extent of
available satellite imagery used in the analysis. The confined area of interest is 1195 km2
(119,552 ha) and ranges in elevation from ~ 950 m to 2960 m above sea level (Figure 1).
2.2 Aerial and Landscape Photograph Processing
Six color infrared aerial photographs were obtained in digital format from the US Geological
Survey’s Earth Resources Observation and Science Center (Figure 1). Four of the photos were
acquired in 1982 (west of the Continental Divide), two in 1984 (east of the divide). All photos
have a scale of 1:58,000 and were scanned at a resolution of 1800 dots per inch. The photos were
orthorectified to a 2009 NAIP photo (National Agriculture Imagery Program) using ground
control points (GCPs) and a 30 m digital elevation model (DEM) (Leica Photogrammetry Suite,
Erdas, Inc., Norcross, GA, USA). The average root mean square error (RMSE) for each photo
was less than two meters. We independently assessed the average displacement between each of
the orthorectified images and the 2009 NAIP image at multiple locations within each image pair.
The average displacement between both sets of images was less than two meters and deemed
acceptable.
We searched two landscape photographic archives (the US Geological Survey Photographic
Library and Glacier National Park Research Library) to locate additional sources with evidence
of the disturbance. We obtained several color photos taken in the late 1970s or 1980 that

!
11!
!
contained evidence of the outbreak. In several cases the extent of the aerial color infrared photo
and the color landscape photo were congruent. We were able to match the two photos and
identify unique patterns and patches of mortality in each photo. Although this was a qualitative
analysis, the additional information provided us with concrete evidence of the disturbance in the
aerial photos (Figure 2).
2.3 Aerial Detection Survey Data
We obtained the digital version of the ADS data (1962-1998) from GNP and subset annual
shapefiles to correspond with the start of the outbreak (1971) and the last year before extensive
fires in the park (1987). We queried polygons associated with mountain pine beetle using the
Damage Causal Agent attribute code and clipped the shapefile to the extent of the park for each
year. Each annual shapefile was converted to an annual grid (30 m), snapped to the master
Landsat image, and aggregated to form a cumulative mountain pine beetle extent and used to
constrain the study area. We did examine the ADS data for other disturbance agents within the
park to ensure there were no unaccounted disturbances. However, we found very few disturbance
polygons, accounting for a very small area, within the analysis mask.
2.4 Satellite Data
Terrain corrected Level 1T MSS imagery were obtained from the USGS EarthExplorer
Archive (USGS, 2012) of the study area before, during and after the peak of the outbreak. The
imagery had been resampled by the USGS to a spatial resolution of 60 m in four spectral bands
(Table 1). We utilized nine scenes in the analysis (Table 2). Late summer data were used (late
August-September) due to availability of cloud-free imagery and the presumed relative
phenological stability of the forests during this time period (Vogelmann et al., 2009). Geometric

!
12!
!
correction, calibration, atmospheric correction and image normalization procedures were applied
to the imagery and are fully described in Appendix A.
Given the four multispectral bands of MSS, we were only able to utilize three spectral indices
in the model evaluation process (Table 3). The GNDVI is sensitive to the presence of chlorophyll
since the green spectral region is used instead of the red region (Carreiras et al., 2006). We did
not use Band 3 as a covariate as it is often highly correlated with band 4 of MSS data. A
preliminary investigation identified that NDVI performed the best among spectral indices. In an
effort to limit redundancy in the data, we transformed the NDVI time series using principal
component analysis. The principal components were used as predictor variables in one of the
five models tested.
2.5 Sampling
We estimated beetle induced forest mortality using data collected from the aerial photos and
compared these measurements with changes in spectral values over time. We segregated the
landscape into 12 facets based on slope and aspect. These two variables influence forest
composition, tree vigor and subsequent susceptibility to mountain pine beetle (Raffa et al.,
2008). Furthermore, dividing the landscape into sub-regions of similar biophysical characteristics
can isolate spectral gradients (Homer et al., 2004). Both variables were derived from the
elevation dataset. Aspect was classified into four categories (north, east, south or west) while
slope was classified into three quantiles: low (<12%), moderate (12-29%) and high (>29%).
Initially 350 random points were proportionally allocated in each landscape class, and square
plots of 180 m x 180 m were delineated around the center of each point. The plot size was
chosen considering the spatial resolution of the satellite imagery, i.e. 3 x 3 Landsat MSS pixels.
A negative buffer was used to insure each plot was located completely within one landscape

!
13!
!
facet, and deleted (10%) if it fell within multiple facets. In addition, limitations due to
topographic shadow or image blur from the orthorectification process warranted the omission of
some plots (13%). As a result, each landscape facet did not contain the same number of sampling
plots.
An unsupervised classification in Erdas Imagine was conducted on each air photo resulting in
20 classes. We used an iterative approach to determine the number of unsupervised classes that
maximized spectral separation without generating an unwieldy number of classes. For each plot,
we manually interpreted the 20 classes and assigned each class to live forest, dead forest, or
shadow (Figure 3). We then calculated the ratio of dead canopy cover to total canopy cover in
each plot. We omitted shadow pixels as they represent unknown cover types.
2.6 Statistical Analysis
Regression analysis can be used to explain large-scale variability, while model residuals can
be used to describe small-scale variability in the data (Cressie, 1993). We used a generalized
linear model (GLM, Gaussian distribution, Identity link function) to identify a set of explanatory
variables to estimate canopy cover change on the sample plots over time. Predictor variables
included spectral indices derived from nine years of Landsat MSS data, topography (elevation,
slope, aspect and topographic position index), and variables derived from the ADS data (first
year detected, last year detected and total number of years detected). Aspect and the variables
derived from the aerial survey data were treated as indicator variables in the analysis. Aspect was
binned into four classes: North (0-45º; 315-365º), East (45-135º), South (135-225º), and West
(225-315º). Three categorical variables were derived from the aerial survey data: first year of
attack (early, mid, or late in the outbreak), last year of attack (early, mid, or late in the outbreak)
and total number of years recorded during the outbreak (low, moderate, or high).

!
14!
!
We tested five models in our analysis using different combinations of vegetation indices as
the primary biotic variables. For each model, a stepwise selection by Akaike’s Information
Criterion (AIC) was used to identify the best subset of independent variables to include in the
regression models (R Development Core Team, 2011). The aspect variable was allowed to
interact with the primary vegetation index in each model. We evaluated the models through
consideration of AIC, the mean absolute error of prediction (MAE) and the root mean square
error of prediction (RMSE). Furthermore, a ten-fold cross validation procedure (DAAG package
in R) was employed to calculate the prediction error of each model.
Residual error from the regression model can be utilized to describe the small-scale
variability in the data (Manier et al., 2005; Reich et al., 2011). Model residuals may still contain
useful information that can be utilized to gain precision in estimation (Pongpattananurak et al.,
2012). We modeled the residual error from the selected regression model using a binary
regression tree. We tested the residuals of the selected GLM model and the regression tree model
for spatial autocorrelation using the Moran’s I statistic (Legendre & Fortin, 1989). The sampled
plots were clustered on the landscape into three distinct groups based on the availability of the
aerial photos. We assumed points between each cluster were spatially independent and employed
a block diagonal spatial weights matrix (Upton & Fingleton, 1985) to account for the clustered
nature of the plots. We used inverse distance to define the neighborhood structure of the three
spatial weights matrices (one for each cluster).
The residuals of the GLM-CART model exhibited spatial autocorrelation. We addressed
the issue by running the regression analysis using a Generalized Least Squares (GLS) model. A
variogram was fit using the residuals of the GLM model to describe the degree of spatial
dependence in the residuals. A Gaussian variogram model was fit to the sample variogram using

!
15!
!
least squares to estimate the nugget, sill and range. The GLS regression was used to estimate the
parameters of the trend surface model in the presence of spatial autocorrelation. We allowed plot
location (east or west of the Continental Divide) to enter the model to test if the outbreak impacts
were different on either side of the divide. We used the residuals of the GLS to model the small-
scale variation in the data using binary regression trees as described above.
After parameterizing and validating the models, forest canopy change was projected to
the landscape area of interest in three steps. First a trend surface was created from the parameters
of the GLS model using the raster calculator in ArcGIS. Next, a surface of the residuals
generated from the regression tree model was created using a series of conditional statements in
the raster calculator. Finally, the trend and residual surfaces were added together to create a
continuous surface of forest canopy change scaled between 0 and 1. Areas of cloud cover, cloud
shadow and topographic shadows represent uncertainty and were omitted from the analysis. Only
two years of data (1978 and 1983) contained sparse clouds, but topographic shadows were
present in all years. We applied a NDVI threshold (< 0.2) to remove clouds and topographic
shadows (Hicke & Logan, 2009) and cloud shadows were manually delineated and removed.
3. Results
3.1 Aerial Detection Survey Data
Our analysis of the aerial survey data indicates the outbreak was first identified in 1971 in the
north-west portion of the park in very small isolated patches. The outbreak continued to spread
from these centers until the mid-1970s when it was reported widely across the western portion of
the park (Figure 4). There was no data available for 1975, and the following year the area
affected by beetles significantly expanded on the western side of the park. The aerial survey
continued to report large areas impacted from 1977 through 1980. The outbreak was first

!
16!
!
identified east of the Continental Divide in the north central and north east portion of the park in
1979. In the early 1980s, the area affected by beetles quickly decreased (Figure 5).
3.2 Determination of Tree Canopy Cover
A total of 261 plots were used to estimate tree canopy mortality from the air photo analysis
(Table 4). Initially, 282 plots were analyzed, but several were removed from the data set because
the photo plots fell within topographic shadows, cloud cover or cloud shadows in the satellite
imagery. The study area is dominated by west facing slopes, followed by south, east and north.
Each aspect class did not contain the same number of plots (see Section 2.5). However, the
number of plots in each aspect class is an adequate reflection of the percentage of the study area
in each aspect class. Plots ranged from very little mortality (4.4%) to nearly complete mortality
(99.8%). West-facing plots had the highest mean mortality (68.3%), while plots in the east aspect
class had the lowest mean mortality (49.4%) (Table 4). The majority of the data is concentrated
in mortality classes ranging from 40-90% (Figure 6). Given the severity and extent of the
outbreak, this is not an unexpected finding.
3.3 Model Adjustment and Validation
The model that employed NDVI and the Green Band (NDVI+G) (Table 5) provided the best
estimation of canopy change over time. This model had the lowest AIC (-237.55), MAE
(10.8%), and RMSE (13.6%) values while accounting for the greatest amount of explained
variability (65.4%) (Table 5). Furthermore this model had the lowest prediction error (15.4%) of
any model from the cross validation procedure. The incorporation of a green band resulted in a
stronger model than using NDVI alone (Table 5). The NDVI and PCA models had identical
coefficients of determination, and similar MAE and RMSE. However, the PCA model had

!
17!
!
substantially higher prediction error. The GNDVI model did not perform as well as the three
NDVI based models and the red-green index proved to be a poor indicator of mortality.
The NDVI+G model was selected to describe the large-scale variability of canopy change
over time. However, the residuals of the GLM model exhibited spatial autocorrelation (Moran’s I
test; p < 0.0001) indicating that the null hypothesis of spatial independence in the residuals be
rejected. The variables included in the NDVI+G model were then analyzed using a GLS model
that explained 62% of the variability with higher MAE (18%) and RMSE (21.1%) than the GLM
model (Table 6, Figure 7). However, the residuals of the GLS model did not exhibit spatial
autocorrelation (Moran’s I test; p =0.64). The green band from 1978 was the most important
predictor west of the Continental Divide, with a relative contribution to the model of 10.7%. The
green band from 1987 was the most important predictor east of the divide, with a relative
contribution to the model of 11.8%. Decreased values of green band reflectance indicated a
substantial increase in canopy mortality. NDVI from 1977 and 1981 were highly significant in
the model on the west side of the park (p<0.001) and also exhibited a negative relationship with
canopy mortality. NDVI from 1977 was also significant on the east side of the park.
The residuals were used to model the small-scale variation in the data using binary regression
trees. The initial regression tree identified 27 nodes and location (east or west of divide) did not
enter the analysis, so one tree was used to fit both sides of the Continental Divide. Given that
regression trees are prone to overfitting, we conducted a 10-fold cross validation on the data and
subsequently pruned the tree to 22 nodes. This simplified the model while still accounting for
spatial autocorrelation. The combined model (GLS + CART), which captures both the large- and
small-scale variability, had a lower rate of MAE (7.6%) and RMSE (9.8%) than the GLS model.
The combined model increased the amount of explained variability in the data by nearly 20% (R2

!
18!
!
= 0.819) (Figure 8). The residuals of the combined model are spatially independent (Lagrange
multiplier test; p=0.27) and the standardized mean square error (SMSE) of the combined model
is 0.996. An SMSE value of one indicates consistency between the estimation error variance and
the observed error variance in the model (Hevesi et al., 1992).
The combined model was then applied spatially to the study area as a continuous surface
with modeled canopy cover change scaled between 0 and 1. We binned the modeled data into
three categories based on natural breaks in the data (Figure 9). This classification resulted in 20%
of the project area in the low category (< 0.37 canopy change), 46% in the moderate and 34% in
the severe category (>0.62 canopy change). Pockets of high severity are found throughout the
park across the elevation gradient present. The three classes are generally represented across the
study area. However, it should be noted that pockets of low and severe impacts are clustered,
with the moderate severity often forming a transition between the classes.
To provide perspective on the classification, a color-infrared photo and corresponding
classification map is shown in Figure 10. Based on these visual comparisons, our model appears
to capture high levels of mortality associated with beetle attack areas as well as areas not as
heavily impacted. Furthermore, the gradient of impact on the landscape appears to be well
represented in the model. Example spectral trajectories of the three classes show clear
delineation during extent of the outbreak (Figure 11).
4. Discussion
A primary objective of our analysis was to develop a methodology to reconstruct the extent
and severity of the outbreak. We were able to identify a gradient of mortality on the landscape
using changes in NDVI and the green band reflectance over time. Our findings confirm the
outbreak was not homogenous across the landscape (Figure 9). The reported error metrics are

!
19!
!
reasonable given limitations in the data and comparable to related studies of insect impacts on
the forest canopy (Townsend et al., 2012). Error associated with the ADS data was not
quantified. Furthermore, this information was collected by observers presumably working under
difficult conditions. Therefore we suggest our model represents an unbiased view of the
disturbance. In addition, the modeling framework we applied in this study should be transferable
to other areas with similar forest disturbance characteristics.
This study builds on the ideology of many of the aforementioned studies which used
remotely sensed data to document various stages of the late 1990s-mid 2000s mountain pine
beetle outbreak. The common theme is the development of a time series imagery stack to assess
spectral changes over time (Goodwin et al., 2008; Meddens et al., 2013; Meigs et al., 2011). Our
normalization process gave us high confidence in the time series stack, given the consistent
reflectance values of the pseudo-invariant features over time (Figure 11). However, we were
unable to utilize many of the vegetation indices (e.g. Normalized Difference Moisture Index)
used in these studies. Given that our study objectives hinged around a historic disturbance that
occurred in the mid-1970s and early 1980s, we were unable to use imagery with the spectral
resolution needed for many of those indices. The major difference in our study and those
described in section 1.3, is that the disturbance we are interested in occurred in the 1970s and
early 1980s. This predates the advent of Landsat TM/ETM+ imagery and other finer scale
imagery employed in those studies.
There were two main differences between our study and those that used MSS data (Harris et
al., 1978; Rencz & Nemeth, 1985; Weber et al., 1975). First, we attempted to capture the
gradient of the disturbance on a continuous scale between 0 and 1. Second, we employed
multiple time periods of imagery to assess spectral changes at sites over time. Although Rencz

!
20!
!
and Nemeth (1985) used a change detection procedure, there was a gap of six years between
images. The use of just two images was likely insufficient to capture the full range of phenology
associated with the disturbance from pre-attack through the green, red and gray stages, followed
by the likely expansion of understory growth following canopy mortality. Conducting a
retrospective analysis afforded us several advantages over the prior MSS studies. The Landsat
archive is now readily available at no cost, removing the financial burden that inhibited prior
investigators from developing a time series imagery stack (Woodcock et al., 2008). Furthermore,
advances in radiometric calibration provide a basis for standardized comparison between images
acquired on different dates and by different sensors (Chander et al., 2009).
There are several strengths associated with our study that allowed us to overcome numerous
limitations. Overall, we provide an objective framework that can be applied to other areas, at
other time periods, involving other types of forest disturbance. The major limitation of
quantifying a disturbance over a large, topographically complex landscape where subsequent fire
has erased some of the evidence was overcome using existing data. The remote sensing archive
allowed us to extract information about the condition of the forest canopy across spatiotemporal
scales. By employing multiple lines of evidence, each independent data source contributed to a
composite picture of the disturbance (Swetnam et al., 1999). Several key factors led to a
successful analysis. The first was employing a mask to restrict the area of analysis (Garrity et al.,
2013) to forest types where mountain pine beetle had the potential to impact. The second critical
element was the development of a normalized time series of reflectance (Townsend et al., 2012;
Vogelmann et al., 2012) to characterize changes over time. We obtained many more images (24)
than we ultimately used (9), but this was necessary to conduct an exhaustive evaluation of
available imagery. The consistent level of pre-processing performed on the imagery by the

!
21!
!
USGS and our procedure to convert data to at-surface reflectance aided in the success.
Furthermore, the image acquisition dates were within a six-week window, which limited intra-
year differences. The final critical element was the development of a novel approach to measure
mortality in available aerial photos and scale up to multiple years of satellite imagery. This
procedure was crucial given the absence of field data.
The absence of validation data in this study precluded the use of the holdout method where
the data is separated into a training set and test set. We evaluated the suite of GLM models using
several common measures, including a ten-fold cross validation procedure. Cross validation
divides the data into k subsets of approximately equal size and the holdout method is repeated k
times (Salzberg, 1997). There are strengths and weaknesses to this method that is commonly
used in ecological studies. Cross validation is advantageous because it is less critical how the
data are divided, every point is used in the test set once, and each point is used in the training set
k-1 times. Disadvantages of the method include computation time and trade-offs between
variance and bias, dependent on the number of iterations (Kohavi, 1995). Furthermore, there is
only one data set involved in the cross validation, regardless of the number of subsets created
(Esbensen & Geladi, 2010).
4.1 Ecological Considerations
In areas where mountain pine beetle disturbance induces high mortality in the forest canopy
over a short time period, there will be a relatively quick change in NDVI. Therefore these areas
will have a heightened chance of detection by remote sensing methods. In addition, the release of
light, nutrients and moisture will occur at one time period. Therefore the flush of understory
growth will likely occur over a relatively short time period. This increases the likelihood of
obtaining a tight sequence of images to detect these rapid changes. Given the high severity of the

!
22!
!
impact, the model identified large negative relative contributions of the green band in the 1970s
on the west side of the divide, indicative of an increase in canopy mortality. However, the 1987
green band was significant, with a large positive relative contribution to the model. This can be
interpreted ecologically in that there was a sharp increase in canopy mortality during the late
1970s, but understory growth was prevalent in these high severity areas by the late 1980s. The
outbreak moved from the west to east over the divide. The 1987 green band had a large negative
contribution to the east side model, suggesting recent canopy mortality dominated the spectral
signature, while understory regrowth was likely not widespread.
However, the impacts of mountain pine beetle disturbance on the forest canopy do not
always exhibit characteristics that are easily identified by remote sensing methods. Areas that
have lower amounts of mortality will be composed of a mix of live and dead trees resulting in a
gradient of mortality over the duration of the disturbance. As trees die over this time period, they
will likely be interspersed with live trees. Given that the spectral response of a pixel is an
amalgamation of all elements present (Lefsky & Cohen, 2003), there will be a smaller change in
reflectance. Additionally, as individual trees die, the release of resources will impact a smaller
area of understory regrowth. The localized understory regrowth could offset or suppress the
change in reflectance associated with canopy mortality. This problem is manifested on the
landscape as the cycle of canopy mortality, resource release, and understory flush could be
occurring simultaneously in localized areas.
Several ecological phenomena could pose challenges to this methodology, particularly if the
recent disturbance history of the study area is unknown. Other disturbances could be identified
by this method, without being attributed to mountain pine beetle. We were able to incorporate
ancillary data about the mountain pine beetle outbreak such as ADS data, park reports and

!
23!
!
knowledge from park staff to supplement the primary imagery method. Harvest events typically
have sharp geometric boundaries (Goodwin et al., 2008) that often persist in reflectance patterns
for quite some time after the event. Unknown fires that are low severity or small in area could be
difficult to segregate from insect disturbance mortality, particularly if the event corresponds with
a gap in satellite imagery. Other insect disturbances such as mortality or defoliation events in the
study area could be detected as well (Meigs et al., 2011; Townsend et al., 2012). We analyzed
the Damage Causal Agent attribute code of the aerial survey data and found nearly no other
disturbance types recorded within the study area during the time periods 1971-87. Given that our
objective was to detect landscape-scale mortality associated with a widespread, high-severity
disturbance, we were not concerned with these small disturbances.
Periods of drought and fluctuations in hydrologic year (Oct.-Sept.) precipitation could impact
inter-annual indices of vegetation reflectance in areas of low mortality. However, our
normalization procedure should account for some of these differences between imagery years.
The establishment of appropriate reference conditions of an area remains a challenge in
ecological studies (Millar et al., 2007). Finally, all of the aforementioned challenges are made
more complex when attempting to conduct a retrospective analysis of historical forest
disturbance.
4.2 Technical Considerations
The technological challenges associated with this study are centered on the spatial, temporal
and spectral resolution of the aerial photographs and satellite imagery. Although we were
constrained to the use of best available data for the time period, consideration of some of the
shortcomings is necessary. We used aerial photographs collected in 1982 (four) and 1984 (two).
The scale of each photograph (1:58,000) is relatively coarse and does not allow for the

!
24!
!
identification of an individual tree crown. However, we believe the size of the photo plots (180
m x 180 m) was adequate to characterize the level of mortality within a stand. Given that our
objective was to measure canopy mortality, we were confined to using color-infrared
photographs. We would have considered natural color photographs if they had been available in
the archive. There were additional photographs available in the archive that were not selected
due to a combination of acquisition date, coarse resolution and gray scale film. Although
nominal, there are acquisition costs associated with historic aerial photos, and the
orthorecticfication process can be time consuming.
Additional landscape photographs would have been extremely helpful. However, we were
limited by those that were taken by park staff at the end of the outbreak and housed in the
National Park Service archive. Although they were not used in a quantitative analysis, they
provided valuable evidence of the impact of disturbance.
The Landsat MSS imagery employed in this study is also subject to spatial, temporal and
spectral constraints. Although we resampled the MSS imagery to 30 m to aid in the
georectification process, the native resolution is significantly lower. Pixel values are a mixture of
ground elements (Lefsky & Cohen, 2003), and given the coarse spatial resolution, MSS imagery
is limiting to the amount of mortality that can be detected at one pixel between multiple time
periods. Therefore, areas that experienced low mortality may have been underestimated by our
model. The temporal limitations of the image archive are two-fold. The study may have
benefited from a higher frequency of images collected every calendar year and additional image
years to establish pre-outbreak conditions. However, it was not tenable to alleviate these
constraints given the available imagery and the timing of the disturbance. The spectral resolution
of MSS imagery is limited compared to TM/ETM+ imagery. Many of the indices that have been

!
25!
!
successfully applied to recent outbreaks are developed from a wider spectral range than that of
MSS. All of these factors may limit the sensitivity of the study to detect different levels of
mortality, especially low levels of mortality. However, given the scale and severity of the
disturbance, coupled with the dense imagery stack that was assembled, we were still able to
achieve acceptable results.
The Tasseled Cap transformation for Landsat data has been used to distill information from
Landsat imagery in forest disturbance mapping (Healey et al., 2005). However, we did not use
the Tasseled Cap transformation in our analysis. Unlike Landsat TM and ETM+, Tasseled Cap
coefficients have not been developed for MSS imagery that has been converted to reflectance
data (Schowengerdt, 2007). Our normalization process depended on normalized reflectance
values and not Digital Numbers. The established Tasseled Cap transformation can only be
applied to Landsat MSS imagery in Digital Numbers.
We chose to classify the continuous output into three categories based on natural breaks in
the data. Although relative differences are taken into account, the threshold between each class is
somewhat subjective. Prior investigators have used lower thresholds (low <= 10%, moderate
(11-29%), and severe > 30% of stands killed) (Aukema et al., 2006) or additional classes of
mortality severity (e.g. trace, light, moderate, severe, and very severe) (Meddens et al., 2012).
However, these two studies were considering ADS data which contained a measure of the
number of trees or the percentage of stand killed. This type of classification scheme does not
translate directly to our model. For example, if 15% of the trees were killed in a localized area, it
could have a large impact on the reflectance of those pixels and overestimate the severity. This
issue could be exacerbated by the coarse resolution of Landsat MSS pixels. Given that there is no
precedent for this type of analysis we opted for a natural break classification scheme.

!
26!
!
Our modeling framework was exhaustive in using multiple lines of evidence that represented
the best available data. Our model incorporated the full extent of available spectral reflectance in
MSS imagery (green, red and near infrared bands). Only band 3 was discarded given that it was
highly correlated with band 4. Furthermore, the spectral information used by the model can be
readily interpreted. NDVI is a commonly used index to assess ecological change (Pettorelli et al.,
2005) and its behavior can be reasonably predicted from plant physiology theory (Garrity et al.,
2013). Plant material containing chlorophyll reflects in the green wavelength. The reflectance in
the green band would be expected to decrease as the amount of chlorophyll in a pixel is reduced
from plant mortality. Therefore, the inclusion of the green band provides a measure of the
amount of chlorophyll present within a pixel over time.
5. Conclusions
We have presented a framework that incorporates multiple lines of evidence to
retrospectively characterize a landscape scale mountain pine beetle disturbance. Furthermore, we
have demonstrated that Landsat MSS data is a valuable tool to extend the moderate resolution
imagery record back to the early 1970s. We conclude that our approach is suitable to characterize
the extent and severity of the event despite initial data limitations. Key considerations of the
application of our model include the size and severity of the disturbance, as well as the timing
(first date, last date, and duration) of the satellite imagery. Our approach captures the
characteristics of a disturbance event that significantly impacts numerous ecological processes.
Given the availability of these data sources, the characterization of recent events will afford
investigators additional tools to study disturbance interactions and ecological legacies at the
landscape scale.
Acknowledgements

!
27!
!
This research was supported by an NPS grant through the Rocky Mountains Cooperative
Ecosystem Studies Unit (IMR H1200090004), logistical support was provided by the USGS Fort
Collins Science Center, and satellite imagery provided by the USGS Earth Resources
Observation Systems Data Center. We gratefully acknowledge Jeffrey Morisette and two
anonymous reviewers for their comments that helped to improve the manuscript. We thank the
following people for discussion on research direction and/or provision of data: Dennis Divoky,
Sheree West, Jeff Hicke, Michael O’Donnell, Dan Manier, Alexandra Urza and Sayat
Temirbekov.
Any use of trade, firm, or product names is for descriptive purposes only and does not imply
endorsement by the U.S. Government.
!

!
28!
!
Table 1. Spectral characteristics of Landsat MSS imagery (NASA, 2013).
Band
Wavelength
Spectral Region
1
500-600 nm
Green
2
600-700 nm
Red
3
700-800 nm
Near-infrared
4
800-1100 nm
Near-infrared
!
!
Table 2. Satellite imagery scene information and acquisition date used in the analysis.
Satellite
Scene Path/Row
Acquisition Date (year-month-day)
Landsat 1
44/26
19730910
Landsat 1
44/26
19740923
Landsat 2
44/26
19760921
Landsat 2
44/26
19770811
Landsat 3
44/26
19780902
Landsat 3
44/26
19790915
Landsat 2
44/26
19810913
Landsat 4
41/26
19830924
Landsat 5
41/26
19870911
Table 3. Spectral indices calculated with the Landsat MSS reflectance data; NDVI (Normalized
Difference Vegetation Index), RGI (Red Green Index), and GNDVI (Green Normalized
Difference Vegetation Index).
Spectral Index
Equation
Source
NDVI
NDVI = MSSBand4 – MSSBand2/MSSBand4 + MSSBand2
Rousse et al. 1974
RGI
RGI = MSSBand2/MSSBand1
Coops et al. 2006
GNDVI
GNDVI = MSSBand4 – MSSBand1/MSSBand4 + MSSBand1
Gitelson et al. 1996
!

!
29!
!
Table 4. Descriptive statistics of estimated tree canopy mortality from the aerial photo plots
grouped by aspect class (n=261).
Aspect
Number of
Plots
Tree Canopy Mortality Statistics
Mean
Minimum
Maximum
S.D.
North
46
54.9
4.4
91.8
23.6
East
47
49.4
12.8
93.9
24.0
South
75
54.1
17.0
99.2
20.8
West
93
68.3
12.7
99.8
21.3
Table 5. Comparison of model evaluation metrics.
Model
AIC
R2
MAE
RMSE
10-fold Cross
Validation Prediction
Error
NDVI + G
-237.55
0.65
10.8%
13.6%
15.4%
NDVI
-204.44
0.60
11.6%
14.5%
16.8%
PCA
-193.43
0.60
11.9%
14.6%
20.4%
GNDVI
-183.13
0.55
12.4%
15.6%
17.3%
RGI
-87.32
0.34
15.3%
18.8%
20.7%

!
30!
!
Table 6. Predictor variables used in the NDVI + G GLS model. Estimates of the model
parameters are listed for the west and east sides accordingly. The variables Aspect and Total # of
Years (low was the only category retained in the stepwise model) were treated as indicator
variables in the analysis. *P-value is significant at 0.05 or lower.
Variable
West Coefficient
East Coefficient
(Intercept)
3.18413*
3.43654*
Aspect
N
0.13478
-1.42706*
S
-0.24309
-
W
-0.58963*
-0.17668*
green.1973
-9.58080*
-
green.1974
-8.60109*
-
green.1977
-9.97942*
-
green.1978
-10.69192*
-12.20042
green.1979
-5.24848
-
green.1983
6.49274
-
green.1987
8.11468*
-11.77014*
Total # of Years - Low
0.15773*
-
ndvi.1973
-
-0.37092
ndvi.1974
-0.03773
-
ndvi.1976
0.41517*
-
ndvi.1977
-1.08326*
-1.74362*
ndvi.1978
-0.36746
-
ndvi.1979
0.40112
-
ndvi.1981
-1.42992*
-
ndvi.1983
-0.40449*
-0.50511
ndvi.1973 x N
-
2.01483
ndvi.1974 x N
-1.24923*
-
ndvi.1978 x N
1.69811*
-
ndvi.1978 x W
1.10049*
-
ndvi.1979 x N
-1.78641*
-
ndvi.1979 x S
-1.18434*
-
ndvi.1979 x W
-1.49155*
-
ndvi.1981 x S
1.52377*
-
ndvi.1981 x W
1.20919*
-
ndvi.1983 x N
0.9887*
-
!!

!
31!
!
!
Figure 1. Location of study area and extent of aerial photo coverage. Background image is
Landsat Thematic Mapper (TM) Imagery (bands 3, 2, 1) acquired on August 25, 2010.
Yellow polygons represent the location and extent of aerial photograph coverage; tan area
represents the confined study area.

!
32!
!
Figure 2. (Left) Landscape photo taken in the Summer of 1980 showing a mixture of live and
dead trees in the red attack stage in Waterton Valley (source: Glacier National Park
Research Library). (Right) A color-infrared aerial photo of the same area acquired in
October 1980 (source: NASA/Glacier National Park). The mosaic of live and dead forest
can be identified in both images. The letters correspond to the same area in each photo (A =
stream confluence, B = small patch of live trees, surrounded by dead forest, C = linear
ribbon of dead forest).

!
33!
!
Figure 3. (A) Plot used to sample aerial photos. The 180 m x 180 m plot size was chosen to
include a 3 x 3 block of Landsat MSS pixels. (B) Sampling plot overlaid on color infrared
photo at a low mortality site. (D) Sampling plot overlaid on color infrared photo at a high
mortality site. (C) Output classification from sampling plot in panel B (live canopy cover =
83%). (E) Output classification from sampling plot in panel D (live canopy cover = 10%).

!
34!
!
Figure 4. Mapped area impacted by mountain pine beetle according to the aerial detection survey
data. Note: there was no data available for 1975.

!
35!
!
Figure 5. Area impacted by mountain pine beetle annually based on aerial detection survey data.
Note: there was no data available for 1975.

!
36!
!
Figure 6. Histogram of canopy tree mortality (%) for all plots (n=261).

!
37!
!
Figure 7. The output of the NDVI+G GLS model used to estimate canopy change over time due
to mortality.

!
38!
!
Figure 8. The output of the combined GLS-CART model used to estimate canopy change over
time due to mortality.

!
39!
!
Figure 9. The output of the spatial model classified into three severity levels.

!
40!
!
Figure 10. (Left) Color-infrared photo (year acquired = 1982). (Right) Classified map result of
the same area (focal window applied). Black polygons correspond to spectral trajectories in
Figure 11 (A=Moderate, B=Severe, C=Low). Note: tick marks are spaced on a 2 km grid;
black polygons are 0.2 km2 (20 hectares) in size.
!

!
41!
!
Figure 11. Spectral trajectories of classified outbreak severity. The three trajectories correspond
to the polygons identified in Figure 10. Note: rock features are included to demonstrate the
success of the image normalization process and the stability of pseudo-invariant features over
time.
!

!
42!
!
References
Ahern, F. J. (1988). The effects of bark beetle stress on the foliar spectral reflectance of
lodgepole pine. International Journal of Remote Sensing, 9(9), 1451–1468.
Allen, C. D., & Breshears, D. D. (1998). Drought-induced shift of a forest-woodland ecotone:
rapid landscape response to climate variation. Proceedings of the National Academy of
Sciences, 95(25), 14839–42.
Aukema, B. H., Carroll, A. L., Zhu, J., Raffa, K. F., Sickley, T. A., & Taylor, S. W. (2006).
Landscape level analysis of mountain pine beetle in British Columbia, Canada:
spatiotemporal development and spatial synchrony within the present outbreak. Ecography,
29, 427–441.
Baker, W. L., Flaherty, P. H., Lindemann, J. D., Veblen, T. T., & Eisenhart, K. (2002). Effect of
vegetation on the impact of a severe blowdown in the southern Rocky Mountains, USA.
Forest Ecology and Management, 168(1-3), 63–75. doi:10.1016/S0378-1127(01)00730-7
Bebi, P., Kulakowski, D., & Veblen, T. T. (2003). Interactions between Fire and Spruce Beetles
in a Subalpine Rocky Mountain Forest Landscape. Ecology, 84(2), 362–371.
Bentz, B. J., Régnière, J., Fettig, C. J., Hansen, E. M., Hayes, J. L., Hicke, J. A., Kelsey, R. G.,
Negrón, J. F., & Seybold, S. J. (2010). Climate Change and Bark Beetles of the Western
United States and Canada: Direct and Indirect Effects. BioScience, 60(8), 602–613.
doi:10.1525/bio.2010.60.8.6
Bradley, B. A., & Fleishman, E. (2008). Relationships between expanding pinyon–juniper cover
and topography in the central Great Basin, Nevada. Journal of Biogeography, 35(5), 951–
964. doi:10.1111/j.1365-2699.2007.01847.x
Brown, K., Hansen, A. J., Keane, R. E., & Graumlich, L. J. (2006). Complex interactions
shaping aspen dynamics in the Greater Yellowstone Ecosystem. Landscape Ecology, 21(6),
933–951. doi:10.1007/s10980-005-6190-3
Carreiras, J. M. B., Pereira, J. M. C., & Pereira, J. S. (2006). Estimation of tree canopy cover in
evergreen oak woodlands using remote sensing. Forest Ecology and Management, 223(1-
3), 45–53. doi:10.1016/j.foreco.2005.10.056
Chander, G., Markham, B. L., & Helder, D. L. (2009). Remote Sensing of Environment
Summary of current radiometric calibration coefficients for Landsat MSS , TM , ETM + ,
and EO-1 ALI sensors. Remote Sensing of Environment, 113(5), 893–903.
doi:10.1016/j.rse.2009.01.007
Chavez, P. S. (1988). An improved dark-object subtraction technique for atmospheric scattering
correction of multispectral data. Remote Sensing of Environment, 24(3), 459–479.
doi:10.1016/0034-4257(88)90019-3

!
43!
!
Chavez, P. S. (1996). Image-Based Atmospheric Corrections - Revisited and Improved.
Photogrammetric Engineering & Remote Sensing, 62(9), 1025–1036.
Cohen, W. B., Maiersperger, T. K., Spies, T. A., & Oetter, D. R. (2001). Modelling forest cover
attributes as continuous variables in a regional context with Thematic Mapper data.
International Journal of Remote Sensing, 22(12), 2279–2310.
doi:10.1080/01431160121472
Coops, N. C., Johnson, M., Wulder, M. A., & White, J. C. (2006). Assessment of QuickBird high
spatial resolution imagery to detect red attack damage due to mountain pine beetle
infestation. Remote Sensing of Environment, 103(1), 67–80. doi:10.1016/j.rse.2006.03.012
Cressie, N. A. (1993). Statistics for Spatial Data (Revised., p. 928). New York, NY: Wiley-
Interscience. doi:978-0471002550
Davidson, A., Aycrigg, J., Grossmann, E., Kagan, J., Lennartz, S., McDonough, S., Miewald, T.,
Ohmann, J., Radel, A., Sajwaj, T., & Tobalske, C. (2009). Digital Land Cover Map for the
Northwestern United States. Moscow, ID: Northwest Gap Analysis Project: USGS GAP
Analysis Program.
Dennison, P. E., Brunelle, A. R., & Carter, V. A. (2010). Assessing canopy mortality during a
mountain pine beetle outbreak using GeoEye-1 high spatial resolution satellite data. Remote
Sensing of Environment, 114(11), 2431–2435. doi:10.1016/j.rse.2010.05.018
DeRose, R. J., Long, J. N., & Ramsey, R. D. (2011). Combining dendrochronological data and
the disturbance index to assess Engelmann spruce mortality caused by a spruce beetle
outbreak in southern Utah, USA. Remote Sensing of Environment, 115(9), 2342–2349.
doi:10.1016/j.rse.2011.04.034
Di Orio, A. P., Callas, R., & Schaefer, R. J. (2005). Forty-eight year decline and fragmentation
of aspen (Populus tremuloides) in the South Warner Mountains of California. Forest
Ecology and Management, 206(1-3), 307–313. doi:10.1016/j.foreco.2004.11.011
Esbensen, K. H., & Geladi, P. (2010). Principles of Proper Validation: use and abuse of re-
sampling for validation. Journal of Chemometrics, 24(3-4), 168–187. doi:10.1002/cem.1310
Fettig, C. J., Klepzig, K. D., Billings, R. F., Munson, A. S., Nebeker, T. E., Negrón, J. F., &
Nowak, J. T. (2007). The effectiveness of vegetation management practices for prevention
and control of bark beetle infestations in coniferous forests of the western and southern
United States. Forest Ecology and Management, 238(1-3), 24–53.
doi:10.1016/j.foreco.2006.10.011
Fuller, D. O. (2001). Forest fragmentation in Loudoun County , Virginia , USA evaluated with
multitemporal Landsat imagery. Landscape Ecology, 16, 627–642.

!
44!
!
Garrity, S. R., Allen, C. D., Brumby, S. P., Gangodagamage, C., McDowell, N. G., & Cai, D. M.
(2013). Quantifying tree mortality in a mixed species woodland using multitemporal high
spatial resolution satellite imagery. Remote Sensing of Environment, 129, 54–65.
doi:10.1016/j.rse.2012.10.029
Gitelson, A. A., Kaufman, Y. J., & Merzlyak, M. N. (1996). Use of a Green Channel in Remote
Sensing of Global Vegetation from EOS-MODIS. Remote Sensing of Environment, 58(3),
289–298.
Goodwin, N. R., Coops, N. C., Wulder, M. A., Gillanders, S., Schroeder, T. A., & Nelson, T.
(2008). Estimation of insect infestation dynamics using a temporal sequence of Landsat
data. Remote Sensing of Environment, 112, 3680–3689. doi:10.1016/j.rse.2008.05.005
Goodwin, N. R., Magnussen, S., Coops, N. C., & Wulder, M. A. (2010). Curve fitting of time-
series Landsat imagery for characterizing a mountain pine beetle infestation. International
Journal of Remote Sensing, 31(12), 3263–3271. doi:10.1080/01431160903186277
Hamel, D., McGregor, M., & Oakes, R. (1977). Status of Mountain Pine Beetle Infestation
Glacier National Park, 1976 (No. Report No. 77-4) (pp. 1–7). Missoula, MT.
Harris, J., Dawson, A. F., & Goodenough, D. (1978). Evaluation of LANDSAT Data for Forest
Pest Detection and Damage Appraisal Surveys in British Columbia. Report BC-X-182.
Victoria, BC: Pacific Forest Research Centre.
Healey, S., Cohen, W., Zhiqiang, Y., & Krankina, O. (2005). Comparison of Tasseled Cap-based
Landsat data structures for use in forest disturbance detection. Remote Sensing of
Environment, 97(3), 301–310. doi:10.1016/j.rse.2005.05.009
Hevesi, J. A., Istok, J. D., & Flint, A. L. (1992). Precipitation Estimation in Mountainous Terrain
Using Multivariate Geostatistics. Part I: Structural Analysis. Journal of Applied
Meteorology, 31, 661–676.
Hicke, J., & Logan, J. (2009). Mapping whitebark pine mortality caused by a mountain pine
beetle outbreak with high spatial resolution satellite imagery. International Journal of
Remote Sensing, 30(17), 4427–4441. doi:10.1080/01431160802566439
Homer, C., Dewitz, J., Fry, J., Coan, M., Hossain, N., Larson, C., Herold, N., Mckerrow, A.,
Vandriel, J. N., & Wickham, J. (2007). Completion of the 2001 National Land Cover
Database for the Conterminous United States. Photogrammetric Engineering & Remote
Sensing, 337–341.
Homer, C., Huang, C., Yang, L., Wylie, B., & Coan, M. (2004). Development of a 2001 National
Land-Cover Database for the United States. Photogrammetric Engineering & Remote
Sensing, 70(7), 829–840.

!
45!
!
Huang, C., Asner, G. P., Barger, N. N., Neff, J. C., & Floyd, M. L. (2010). Regional
aboveground live carbon losses due to drought-induced tree dieback in piñon–juniper
ecosystems. Remote Sensing of Environment, 114(7), 1471–1479.
doi:10.1016/j.rse.2010.02.003
Johnson, E. A., & Fryer, G. I. (1987). Historical vegetation change in the Kananaskis Valley,
Canadian Rockies. Canadian Journal of Botany, 65(5), 853–858. doi:10.1139/b87-116
Jones, H. G., & Vaughan, R. H. (2010). Remote Sensing of Vegetation: Principles, Techniques,
and Applications (p. 400). Oxford, UK: Oxford University Press.
Kadmon, R., & Harari-Kremer, R. (1999). Studying Long-Term Vegetation Dynamics Using
Digital Processing of Historical Aerial Photographs. Remote Sensing of Environment, 68(2),
164–176. doi:10.1016/S0034-4257(98)00109-6
Kennedy, R., & Spies, T. (2004). Forest cover changes in the Oregon Coast Range from 1939 to
1993. Forest Ecology and Management, 200, 129–147. doi:10.1016/j.foreco.2003.12.022
Kohavi, R. (1995). A Study of Cross-Validation and Bootstrap for Accuracy Estimation and
Model Selection. International Joint Conference on Artificial Intelligence, 14(2), 1137–
1145.
Lefsky, M., & Cohen, W. (2003). Selection of remotely sensed data. In M. A. Wulder & S.
Franklin (Eds.), Remote Sensing of Forest Environments: Concepts and Case Studies (pp.
13–47). Boston: Kluwer Academic Publishers.
Legendre, P., & Fortin, M. (1989). Spatial Pattern and Ecological Analysis. Vegetatio, 80(2),
107–138.
Logan, J. A., & Powell, J. A. (2001). Ghost Forests, Global Warming, and the Mountain Pine
Beetle (Coleoptera: Scolytidae). American Entomologist, 47(3), 160–173.
Manier, D. J., Hobbs, N. T., Theobald, D. M., Reich, R. M., Kalkhan, M. A., & Campbell, M. R.
(2005). Canopy dynamics and human caused disturbance on a semi-arid landscape in the
Rocky Mountains, USA. Landscape Ecology, 20(1), 1–17. doi:10.1007/s10980-004-3987-4
Maselli, F. (2004). Monitoring forest conditions in a protected Mediterranean coastal area by the
analysis of multiyear NDVI data. Remote Sensing of Environment, 89(4), 423–433.
doi:10.1016/j.rse.2003.10.020
Mauseth, J. (1988). Plant Anatomy (First Ed., p. 560). Menlo Park, CA: The
Benjamin/Cummings Publishing Company.
McGregor, M., Hamel, D., Lood, R., & Meyer, H. (1975). Status of Mountain Pine Beetle
Infestation Glacier National Park, 1974 (pp. 1–7). Missoula, MT.

!
46!
!
Meddens, A. J. H., Hicke, J. A., & Ferguson, C. A. (2012). Spatiotemporal patterns of observed
bark beetle-caused tree mortality in British Columbia and the western United States.
Ecological Applications, 22(7), 1876–1891.
Meddens, A. J. H., Hicke, J. A., & Vierling, L. A. (2011). Evaluating the potential of
multispectral imagery to map multiple stages of tree mortality. Remote Sensing of
Environment, 115(7), 1632–1642. doi:10.1016/j.rse.2011.02.018
Meddens, A. J. H., Hicke, J. A., Vierling, L. A., & Hudak, A. T. (2013). Evaluating methods to
detect bark beetle-caused tree mortality using single-date and multi-date Landsat imagery.
Remote Sensing of Environment, 132, 49–58. doi:10.1016/j.rse.2013.01.002
Meigs, G. W., Kennedy, R. E., & Cohen, W. B. (2011). A Landsat time series approach to
characterize bark beetle and defoliator impacts on tree mortality and surface fuels in conifer
forests. Remote Sensing of Environment, 115(12), 3707–3718.
doi:10.1016/j.rse.2011.09.009
Millar, C. I., Stephenson, N. L., & Stephens, S. L. (2007). Climate change and forests of the
future: managing in the face of uncertainty. Ecological Applications, 17(8), 2145–51.
Moran, M. S., Jackson, R. D., Slater, P. N., & Teiuet, P. M. (1992). Evaluation of Simplified
Procedures for Retrieval of Land Surface Reflectance Factors from Satellite Sensor Output.
Remote Sensing of Environment, 41, 169–184.
NASA. (2013). The Multispectral Scanner System. Retrieved April 15, 2013, from
http://landsat.gsfc.nasa.gov/about/mss.html
Nemani, R., Hashimoto, H., Votava, P., Melton, F., Wang, W., Michaelis, A., Mutch, L., Milesi,
C., Hiatt, S., & White, M. (2009). Monitoring and forecasting ecosystem dynamics using
the Terrestrial Observation and Prediction System (TOPS). Remote Sensing of Environment,
113(7), 1497–1509. doi:10.1016/j.rse.2008.06.017
Parmenter, A. W., Hansen, A. J., Kennedy, R. E., Cohen, W. B., Langner, U., Lawrence, R.,
Maxwell, B., Gallant, A., & Aspinall, R. (2003). Land Use and Land Cover Change in the
Greater Yellowstone Ecosystem: 1975 – 1995. Ecological Applications, 13(3), 687–703.
Pettorelli, N., Vik, J. O., Mysterud, A., Gaillard, J. M., Tucker, C. J., & Stenseth, N. C. (2005).
Using the satellite-derived NDVI to assess ecological responses to environmental change.
Trends in ecology & evolution, 20(9), 503–10. doi:10.1016/j.tree.2005.05.011
Platt, R. V., & Schoennagel, T. (2009). An object-oriented approach to assessing changes in tree
cover in the Colorado Front Range 1938–1999. Forest Ecology and Management, 258(7),
1342–1349. doi:10.1016/j.foreco.2009.06.039

!
47!
!
Pongpattananurak, N., Reich, R. M., Khosla, R., & Aguirre-Bravo, C. (2012). Modeling the
Spatial Distribution of Soil Texture in the State of Jalisco, Mexico. Soil Science Society of
America Journal, 76(1), 199. doi:10.2136/sssaj2011.0180
Powell, J. A., & Logan, J. A. (2005). Insect seasonality: circle map analysis of temperature-
driven life cycles. Theoretical population biology, 67(3), 161–79.
doi:10.1016/j.tpb.2004.10.001
R Development Core Team. (2011). R: A language and environment for statistical computing.
Vienna, Austria: R Foundation for Statistical Computing. Retrieved from http://www.r-
project.org/
Raffa, K. F., Aukema, B. H., Bentz, B. J., Carroll, A. L., Hicke, J. A., Turner, M. G., & Romme,
W. H. (2008). Cross-scale Drivers of Natural Disturbances Prone to Anthropogenic
Amplification: The Dynamics of Bark Beetle Eruptions. BioScience, 58(6), 501.
doi:10.1641/B580607
Reich, R. M., Aguirre-Bravo, C., Bravo, V. A., & Briseño, M. M. (2011). Empirical evaluation
of confidence and prediction intervals for spatial models of forest structure in Jalisco,
Mexico. Journal of Forestry Research, 22(2), 159–166. doi:10.1007/s11676-011-0144-1
Rencz, A. N., & Nemeth, J. (1985). Detection of Mountain Pine Beetle Infestation Using Landsat
MSS and Simulated Thematic Mapper Data. Canadian Journal of Remote Sensing, 11(1),
50–58.
Rousse, J. W., Hass, R. H., Schell, J. A., Deering, D. W., & Harlan, J. C. (1974). Monitoring the
vernal advancement of retrogradation of natural vegetation. NASA/GSFC, Type III, Final
Report. Greenbelt, MD, p. 371.
Salzberg, S. L. (1997). On Comparing Classifiers: Pitfalls to Avoid and a Recommended
Approach. Data Mining and Knowledge Discovery, 1, 317–328.
Schott, J. R., Salvaggio, C., & Volchok, W. J. (1988). Radiometric Scene Normalization Using
Pseudoinvariant Features. Remote Sensing of Environment, 26, 1–16.
Schowengerdt, R. A. (2007). Remote Sensing: Models and Methods for Image Processing (Third
Ed., p. 560). San Diego: Academic Press, San Diego, CA.
Schroeder, T. A., Cohen, W. B., Song, C., Canty, M. J., & Yang, Z. (2006). Radiometric
correction of multi-temporal Landsat data for characterization of early successional forest
patterns in western Oregon. Remote Sensing of Environment, 103, 16 – 26.
doi:10.1016/j.rse.2006.03.008
Skakun, R. S., Wulder, M. A., & Franklin, S. E. (2003). Sensitivity of the thematic mapper
enhanced wetness difference index to detect mountain pine beetle red-attack damage.
Remote Sensing of Environment, 86(4), 433–443. doi:10.1016/S0034-4257(03)00112-3

!
48!
!
Strand, E., Smith, A., Bunting, S., Vierling, L., Hann, D., & Gessler, P. (2006). Wavelet
estimation of plant spatial patterns in multitemporal aerial photography. International
Journal of Remote Sensing, 27(10), 2049–2054. doi:10.1080/01431160500444764
Swetnam, T. W., Allen, C. D., & Betancourt, J. L. (1999). Applied Historical Ecology: Using the
Past To Manage for the Future. Ecological Applications, 9(4), 1189–1206.
doi:10.1890/1051-0761(1999)009[1189:AHEUTP]2.0.CO;2
Townsend, P. A., Singh, A., Foster, J. R., Rehberg, N. J., Kingdon, C. C., Eshleman, K. N., &
Seagle, S. W. (2012). A general Landsat model to predict canopy defoliation in broadleaf
deciduous forests. Remote Sensing of Environment, 119, 255–265.
doi:10.1016/j.rse.2011.12.023
Townshend, J. R., Masek, J. G., Huang, C., Vermote, E. H., Gao, F., Channan, S., Sexton, J. O.,
Feng, M., Narasimhan, R., Kim, D., Song, K., Song, D., Song, X., Noojipady, P., Tan, B.,
Hansen, M. C., Li, M., & Wolfe, R. E. (2012). Global characterization and monitoring of
forest cover using Landsat data: opportunities and challenges. International Journal of
Digital Earth, 5(5), 373–397.
Turner, M. G., & Dale, V. H. (1998). Comparing Large, Infrequent Disturbances: What Have
We Learned? Ecosystems, 1, 493–496.
Turner, M. G., Hargrove, W. W., Gardner, R. H., & Romme, W. H. (1994). Effects of Fire on
Landscape Heterogeneity in Yellowstone National Park, Wyoming. Journal of Vegetation
Science, 5, 731–742.
Upton, G. J. G., & Fingleton, B. (1985). Spatial data Analysis by Example. Volume 1: Point
Pattern and Quantitative Data (p. 410). Chichester: John Wiley & Sons.
USGS. (2012). EarthExplorer archive. Retrieved March 01, 2012, from
http://edcsns17.cr.usgs.gov/NewEarthExplorer/
Veblen, T. T., Hadley, K. S., Nel, E. M., Kitzberger, T., Reid, M., & Villalba, R. (1994).
Disturbance regime and disturbance interactions in a Rocky Mountain subalpine forest.
Journal of Ecology, 82, 125–135.
Vicente-Serrano, S., Perezcabello, F., & Lasanta, T. (2008). Assessment of radiometric
correction techniques in analyzing vegetation variability and change using time series of
Landsat images. Remote Sensing of Environment, 112(10), 3916–3934.
doi:10.1016/j.rse.2008.06.011
Vogelmann, J. E., Tolk, B., & Zhu, Z. (2009). Monitoring forest changes in the southwestern
United States using multitemporal Landsat data. Remote Sensing of Environment, 113(8),
1739–1748. doi:10.1016/j.rse.2009.04.014

!
49!
!
Vogelmann, J. E., Xian, G., Homer, C., & Tolk, B. (2012). Monitoring gradual ecosystem
change using Landsat time series analyses: Case studies in selected forest and rangeland
ecosystems. Remote Sensing of Environment, 122, 92–105. doi:10.1016/j.rse.2011.06.027
Volcani, A., Karnieli, A., & Svoray, T. (2005). The use of remote sensing and GIS for spatio-
temporal analysis of the physiological state of a semi-arid forest with respect to drought
years. Forest Ecology and Management, 215(1-3), 239–250.
doi:10.1016/j.foreco.2005.05.063
Weber, F. P., Roberts, E. H., & Waite, T. H. (1975). Forest Stress Detection: Ponderosa Pine
Mortality from Mountain Pine Beetle. In Evaluation of ERTS-1 Data for Forest and
Rangeland Surveys (pp. 44–61). Berkeley, California: USDA Forest Service Research Paper
PSW-112.
White, J., Wulder, M. A., Brooks, D., Reich, R., & Wheate, R. (2005). Detection of red attack
stage mountain pine beetle infestation with high spatial resolution satellite imagery. Remote
Sensing of Environment, 96(3-4), 340–351. doi:10.1016/j.rse.2005.03.007
Woodcock, C., Allen, R., Anderson, M., Belward, A., Bindschadler, R., Cohen, W., Gao, F.,
Goward, S., Helder, D., Helmer, E., Nemani, R., Oreopoulos, L., Schott, J., Thenkabail, P.,
Vermote, E., Vogelmann, J., Wulder, M. A., & Wynne, R. (2008). Free Access to Landsat
Imagery. Science, 320, 1011.
Wulder, M. A., Dymond, C., White, J. C., Leckie, D. G., & Carroll, A. L. (2006a). Surveying
mountain pine beetle damage of forests: A review of remote sensing opportunities. Forest
Ecology and Management, 221, 27–41.
Wulder, M. A., White, J. C., Bentz, B. J., Alvarez, M. F., & Coops, N. C. (2006b). Estimating
the probability of mountain pine beetle red-attack damage. Remote Sensing of Environment,
101, 150–166.
Yuan, D., & Elvidge, C. D. (1996). Comparison of relative radiometric normalization techniques.
Photogrammetry and Remote Sensing, 51(3), 117–126. doi:10.1016/0924-2716(96)00018-4

!
50!
!
Appendix A. Satellite Image Processing
A.1. Geometric Correction
Twenty GCPs were established to compare the spatial accuracy between the 2009 NAIP
photo and a 2010 Landsat Thematic Mapper (TM) image of the study area. We used the
AutoSync module in Erdas Imagine to georectify the image to the 2009 photo (RMSE < 0.5
pixel). The process was repeated to georectify each of the nine Landsat MSS images to the 2010
TM image. Each MSS image had an RMSE < 0.4 pixel and was resampled to 30 m during the
georectification process using a nearest neighbor transformation to minimize geometric offsets in
the image stack (Goodwin et al., 2008). However, the spatial resolution of the data is still
considered 60 m.
A.2. Calibration
Radiometric calibration of imagery is an important step for creating a consistent, high-
quality temporal image series. We converted the four bands of each image from Digital Numbers
to absolute units of at-sensor spectral radiance using the formula (Chander et al., 2009):
Lλ = (LMAXλ – LMINλ / Qcalmax – Qcalmin) * (Qcal – Qcalmin) + LMINλ (A1)
where
Lλ = Spectral radiance at the sensor’s aperture [W/(m2 sr µm)]
Qcal = Quantized calibrated pixel value [DN]
Qcalmin = Minimum quantized calibrated pixel value corresponding to LMINλ [DN]
Qcalmax = Maximum quantized calibrated pixel value corresponding to LMAXλ [DN]
LMINλ = Spectral at-sensor radiance that is scaled to Qcalmin [W/(m2 sr µm)]
LMAXλ = Spectral at-sensor radiance that is scaled to Qcalmax [W/(m2 sr µm)]

!
51!
!
The spectral radiance values were converted to Top-Of-Atmosphere (TOA) reflectance
to account for differences in sensor and viewing angle using the formula (Chander et al., 2009):
ρλ = π * Lλ * d2 / ESUNλ * cosθs (A2)
where
ρλ = Planetary TOA reflectance [unitless]
π = Mathematical constant equal to ~3.14159 [unitless]
Lλ = Spectral radiance at the sensor’s aperture [W/(m2 sr µm)]
d = Earth-Sun distance [astronomical units]
ESUNλ = Mean exoatmospheric solar irradiance [W/(m2 µm)]
θs = Solar zenith angle [degrees]
Inputs used in the formulas above were supplied by the header file (.MTL) for each scene
and Chander et al. (2009).
A.3. Atmospheric Correction
Each image was snapped to the reference image (1979 image) in ArcGIS to ensure each
30 m pixel for every year was exactly congruent with the master image. An absolute
normalization was applied to the 1979 master image using a dark object subtraction technique
(Chavez 1988). The minimum pixel value of each band (recorded in at least 1000 pixels),
representing deep glacial lakes and shadows, was identified (Chavez, 1996). The theoretical
radiance of a dark object is assumed to have 1% reflectance (Chavez, 1996; Moran et al., 1992)
so the minimum identified pixel value was multiplied by 0.99 to generate the presumed dark
object of each image band.

!
52!
!
A.4. Relative Normalization
The remaining images were normalized to the master image using a relative
normalization technique. This procedure removes non-surface noise and improves the temporal
homogeneity between images so that spectral change associated with surface phenomena can be
detected (Yuan & Elvidge, 1996). Pseudo-Invariant Features (PIFs) are targets in each image that
are not expected to change between image dates (Schott et al., 1988). Relative normalization is
based on the assumption that a linear relationship exists between the reference image and the
image to be normalized (Schott et al., 1988; Yuan & Elvidge, 1996). This technique has been
applied in many studies to analyze vegetation change (Bradley & Fleishman, 2008; Schroeder et
al., 2006; Vicente-Serrano et al., 2008). We identified 60 PIFs that encompassed a range of
pseudo-invariant reflectance values in each band. Each PIF was 32,400 m2 in size; equivalent to
a 3x3 block of 60 m Landsat MSS pixels. The mean of the reflectance values at these sites were
used to fit an ordinary least squares regression model between the image to be normalized for
each year and the reference image for each of the four bands. We tested the residuals for spatial
autocorrelation using the Moran’s I statistic and the Likelihood Ratio Test (Legendre & Fortin,
1989). Inverse distance was used to define the neighborhood structure of the spatial weights
matrix. If spatial autocorrelation was detected, a spatially autoregressive model was used to fit
the data (Cressie, 1993). In all cases, the fit of lines used to spectrally align the images had R2
values > 0.92. Statistical analysis was conducted using the r package (R Development Core
Team, 2011) and the linear regression was applied to each image in Erdas Imagine.