You are on page 1of 13

Remote Sensing of Environment 114 (2010) 725–737

Contents lists available at ScienceDirect

Remote Sensing of Environment


j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / r s e

Fusion of LiDAR and imagery for estimating forest canopy fuels


Todd L. Erdody 1, L. Monika Moskal ⁎
Remote Sensing and Geospatial Analysis Laboratory and Precision Forestry Cooperative, School of Forest Resources, College of the Environment,
University of Washington, Seattle WA 98195-2100, United States

a r t i c l e i n f o a b s t r a c t

Article history: Due to increased fuel loading as a result of fire suppression, land managers in the American west are in need
Received 11 September 2009 of precise information about the fuels they manage, including canopy fuels. Canopy fuel metrics such as
Received in revised form 12 November 2009 canopy height (CH), canopy base height (CBH), canopy bulk density (CBD) and available canopy fuel (ACF)
Accepted 16 November 2009
are specific inputs for wildfire behavior models such as FARSITE and emission models such as FOFEM. With
finer spatial resolution data, accurate quantification of these metrics with detailed spatial heterogeneity can
Keywords:
Canopy fuel metrics
be accomplished. Light Detection and Ranging (LiDAR) and color near-infrared imagery are active and
Canopy height passive systems, respectively, that have been utilized for measuring a range of forest structure characteristics
Canopy base height at high resolution. The objective of this research was to determine which remote sensing dataset can
Canopy bulk density estimate canopy fuels more accurately and whether a fusion of these datasets produces more accurate
Available canopy fuel estimates. Regression models were developed for ponderosa pine (Pinus ponderosa) stand representative of
LiDAR eastern Washington State using field data collected in the Ahtanum State Forest and metrics derived from
Color near-infrared imagery LiDAR and imagery. Strong relationships were found with LiDAR alone and LiDAR was found to increase
Fusion
canopy fuel accuracy compared to imagery. Fusing LiDAR with imagery and/or LiDAR intensity led to small
Forest structure
increases in estimation accuracy over LiDAR alone. By improving the ability to estimate canopy fuels at
higher spatial resolutions, spatially explicit fuel layers can be created and used in wildfire behavior and
smoke emission models leading to more accurate estimations of crown fire risk and smoke related emissions.
© 2009 Elsevier Inc. All rights reserved.

1. Introduction for wildfire behavior modeling and reflect the true spatial heteroge-
neity of the forest. These metrics are canopy height (CH), canopy base
In the American west, increased ground and canopy fuel loading as height (CBH), canopy bulk density (CBD) and available canopy fuel
well as forest structure change due to fire suppression, timber man- weight (ACF). Semiempirical wildfire behavior modeling programs
agement and climatic effects (Running, 2006; Schmidt et al., 2002) such as FARSITE (Finney, 1998) and FlamMap (Finney, 2006) utilize
has amplified the frequency and extent of crown fires (Agee, 1993). spatial data including fuels, topography and weather to predict
Wildfires are major producers of CO2 (carbon dioxide), CO (carbon wildfire growth and crown fire initiation and propagation. Currently,
monoxide) and PM2.5 (particulate matter smaller than 2.5 μm most of these inputs are derived from layers of coarse resolution
diameter) which are harmful to human health (Hardy et al., 2001). which lead the pixels to be spatially homogeneous (Finney &
For predicting the initiation and propagation of crown fires in wildfire Andrews, 1994).
behavior modeling as well as estimating emissions, accurate canopy Previous estimates of canopy fuel metrics have been derived
fuel information is required. Improved emission estimates from wild- through the Landscape Fire and Resource Management Planning Tools
fires are necessary to estimate the impacts of these pollutants on Prototype Project (LANDFIRE), a nationwide effort to map fuels and
human health and the environment (Battye & Battye, 2002). vegetation at 30 m resolution for use by fire and landscape managers
Remote sensing technologies such as airborne LiDAR (Light (Rollins & Frame, 2006). For data layers created for LANDFIRE, canopy
Detection and Ranging) and aerial imagery are high spatial resolution fuel values of CBD and CBH were modeled based on Landsat imagery,
systems that can effectively improve estimates of canopy fuel metrics biophysical characteristics and topographic characteristics. Since
passive sensors cannot detect canopy depth well, the models have
low accuracy and tend to underestimate both CBD and CBH (Keane
et al., 2006). While these datasets are useful for general landscape
level fuel information, other datasets built on a local scale with greater
⁎ Corresponding author. Tel.: + 1 206 221 6391.
E-mail address: lmmoskal@uw.edu (L.M. Moskal).
accuracy should be utilized to complement LANDFIRE data or replace
1
Now at National Park Service, Southern Cascades Network, Klamath Falls, OR 97601, it completely (Reeves et al., 2006). By utilizing localized high
United States. resolution data instead of LANDFIRE data, more realistic estimations

0034-4257/$ – see front matter © 2009 Elsevier Inc. All rights reserved.
doi:10.1016/j.rse.2009.11.002
726 T.L. Erdody, L.M. Moskal / Remote Sensing of Environment 114 (2010) 725–737

of fire spread can be achieved (Krasnow et al., 2009) thus modeling Although utilizing imagery has the benefit of being easily available,
these fuel metrics at higher resolutions should increase estimation expertise is required for image interpretation, the costs for specific
accuracy and reduce model sensitivity. software are high and imagery does not capture what is beneath the
For this research, discrete-return airborne LiDAR and high reso- main canopy (Arroyo et al., 2008). What is needed is a sensor that can
lution color near-infrared imagery was used to estimate canopy fuel see through the canopy and represent the three-dimensional
metrics in a ponderosa pine and transitional mixed conifer forest of structure of the forest.
eastern Washington State. A double sampling approach using regres- The use of small-footprint (discrete-return) airborne LiDAR re-
sion models was utilized for ease of use and broad-scale applicability mote sensing technology includes applications for forestry by
using easily available analytical tools. LiDAR and imagery data were measuring various forest structure characteristics in three dimen-
used to create regression models in 3 different combinations: LiDAR sions. By allowing laser pulses to penetrate below the canopy, recon-
only, imagery only and LiDAR and imagery fusion. struction of the canopy can be achieved. LiDAR has been shown to
measure forest structure characteristics accurately in a variety of
1.1. Remote sensing of canopy fuels forest types. For example, tree heights were measured in Switzerland
in a mixed conifer forest (Morsdorf et al., 2004), canopy base heights
Canopy fuel metrics are difficult to measure for a number of in Sweden (Holmgren & Persson, 2004), crown volume in Spain
reasons. For uniform forest stands, it is assumed that canopy biomass (Riano et al., 2004), biomass in Washington and Alaska (Li et al.,
is uniformly distributed vertically, but this assumption does not hold 2008), leaf area index (LAI) in heterogeneous urban forest (Richard-
true in a complex forest stand. This is due to multiple layers in the son et al., 2009), forest stand characteristics in Oregon (Means et al.,
canopy, presence of ladder fuels and variation within tree species and 2000) and Colorado (Hall et al., 2005) and canopy fuel loading and
within the forest stand. Although destructive sampling is the most canopy bulk density in Yellowstone National Park (Halligan, 2007).
accurate way to measure canopy fuels, it is not a desirable or effective Similarly, canopy fuel metrics (CH, CBH and CBD) have been esti-
way of acquiring data. mated in mixed conifer/deciduous forests of Germany (Riano et al.,
Remote sensing technologies, including LiDAR (discrete-return 2003) and in Western Washington State (Andersen et al., 2005).
and full-waveform) and imagery (aerial and satellite) have been used Canopy base heights have been estimated using a voxel-based
to estimate canopy fuel metrics. Some of the LiDAR and imagery- approach (Popescu & Zhao, 2008). While most of these studies were
derived forest structure characteristics can be used as spatially explicit done on plot or stand scale, some obtained measurements at tree scale
data inputs for fire modeling software. Previous research suggests that (Coops et al., 2004; Riano et al., 2004), while others measured both
LiDAR-derived canopy metrics, used as data layers in FARSITE, could (Halligan, 2007).
serve as a more realistic prediction of fire spread and intensity Digital true color and false color near-infrared aerial imagery have
(Andersen et al., 2005). been used extensively for forest inventory and health monitoring and
In the context of forest fuels, imagery has mainly been utilized to the advancement of digital aerial imagery has extended the
determine fuel types through aerial photo interpretation (Bertolette & possibilities of using it in conjunction with other digital data sources
Spotskey, 1999; Scott et al., 2002) and Landsat image classification such as LiDAR. LiDAR alone cannot provide all the information about
(Kourtz, 1977; Van Wagtendonk & Root, 2003). For high resolution the canopy that is desired. Although LiDAR can accurately assess
sensors such as Quickbird and IKONOS, an object-based classification biomass and height metrics, the technology cannot discern tree
approach has been used to delineate fuel types due to the fact that the species very well, unless fused with optical sensors (Perrson et al.,
resolution is smaller than the objects of interest (Arroyo et al., 2006). 2004). By building on the strengths of each sensor, a fusion of the two
In forest inventory applications, both satellite and aerial optical could potentially improve canopy fuel estimations.
imagery have been proven to be good estimators of stand density, Studies have shown that a fusion of LiDAR and high resolution
volume and canopy height (De Wulf et al., 1990). optical imagery can improve accuracy of tree crown metrics
In terms of canopy fuel modeling and mapping, a variety of spatial (McCombs et al., 2003; Popescu & Wynne, 2004). LiDAR data, used
and spectral sensors has been utilized. Advanced Spaceborne Thermal in a height bin approach in conjunction with Quickbird imagery
Emission and Reflection Radiometer (ASTER) imagery has been used showed an increase in accuracy of fuel type mapping in Texas (Mutlu
to map fuel models and canopy bulk density in the conifer forests of et al., 2008a). Fuel type mapping was also achieved using a fusion of
northern Idaho (Falkowski et al., 2005) while Landsat imagery has LiDAR and AVIRIS data in Hawaii (Varga & Asner, 2008). Utilizing the
been utilized for estimating canopy fuel metrics for the LANDFIRE broad spatial coverage of Landsat Enhanced Thematic Mapper Plus
database (Rollins & Frame, 2006). Hyperspectral imagery in the form (ETM+) data and the high spatial resolution of LiDAR, regression and
of the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) has kriging methods were used to estimate and map canopy height
been used to estimate canopy cover and tree species fraction in the (Hudak et al., 2002). Forest height estimates using a fusion of LiDAR
ponderosa pine (Pinus ponderosa) and Douglas-fir (Pseudotsuga and IKONOS data using a least-squares linear regression model were
menziesii) forests of Colorado (Jia et al., 2006). Although imagery found to be significant (Donoghue & Watt, 2006). Fuel maps created
from these low to medium resolution sensors has been used for a from a LiDAR/imagery fusion have shown to be more accurate than
variety of canopy fuel attributes, canopy fuels have not been maps created from imagery alone and when used in FARSITE, produce
estimated using sub-meter resolution imagery. different fire perimeters and growth area (Mutlu et al., 2008b).
Some structural parameters of forests are often measured through Although different in many aspects, other studies have utilized
another important imagery derivative, image texture. Texture is a full-waveform LiDAR and interferometric synthetic aperture radar
statistical measure of the variation of the spectral response within an (InSAR) for the estimation of forest metrics. By fusing full-waveform
image and is used through applying occurance and co-occurance LiDAR data with either 30 m resolution Landsat, 2 m resolution
matrices to an image (Haralick, 1979). Since texture is measuring Quickbird imagery or InSAR, attributes of canopy height and biomass
the spectral and spatial variation of the pixels, which are affected are moderately improved than by LiDAR data alone (Hyde et al.,
by tree crown dimensions, it is similar to measuring the roughness 2006). By using InSAR alone, height estimations can be achieved with
of the canopy. Biomass estimations have been made using image slightly larger error rates but at lower cost than LiDAR (Andersen
texture (Wulder et al., 1998) as well as classification of species et al., 2008; Breidenbach et al., 2008).
composition and land cover types (Franklin et al., 2000) and forest By defining canopy fuels heterogeneously, more accurate assess-
structure for individual tree canopies and forest stands (Moskal & ments of canopy fuels can be achieved by utilizing high resolution
Franklin, 2002). remote sensing technologies. Fire managers will be able to make
T.L. Erdody, L.M. Moskal / Remote Sensing of Environment 114 (2010) 725–737 727

better decisions based on improved mapping of canopy fuels and air 2.2. LiDAR and imagery data
quality personnel can benefit from more accurate estimates of
emissions. LiDAR data for Ahtanum (Table 1) are in Washington State Plane
Projection 4602 and units are in US survey feet. The high resolution
color near-infrared (NIR) aerial imagery was acquired on October 23,
2. Methods
2006 using an Intergraph DMC sensor flown on a Piper Navajo for the
Ahtanum area. The resolution is 61 cm in 3 bands (NIR, red and
The main research questions this study addresses are: Which data
green). No correction for atmosphere was needed due to the low
source (LIDAR, imagery, fusion) will be best at accurately describing
flying height. No normalization for reflectance was applied as it was
each tree canopy metric? Will there be improvements with data
deemed unnecessary and there were no readings taken from a
fusion? If there are improvements, how much better are they and is it
spectroradiometer on the ground at the time the imagery was flown.
worth having more than one data set to develop models?
No corrections were made for hillshade or shadowing. Imagery is in
The flowchart for methods used is shown in Fig. 1 and provides a
Washington State Plane Projection 4602 and units are in US survey
simple overview of what is described in detail within the next few
feet. Imagery and LiDAR for the Ahtanum were provided by the
sections. The plot data were processed to acquire the four canopy fuel
vendor already orthorectified to each other.
metrics (the dependent variables in regression). Raw LiDAR and high
resolution near-infrared imagery were independently processed to
2.3. Plot selection
produce a metrics representative of each data type that was then used
as the independent variables in regression. For model building, a total
Using existing methods, FUSION software (McGaughey, 2008) was
of 12 models were produced, 3 for each canopy fuel metric (LiDAR,
used to process the LiDAR data into 30.48 m resolution images of
imagery and LiDAR/imagery fusion).
canopy height and canopy cover for the project area. A brief outline of
methods for stand classification is discussed here as the methods are
2.1. Study area written up in detail elsewhere (Sullivan et al., 2009). A similar method
of classification based solely on a digital canopy height model was
Field data were collected in the Ahtanum State Forest, Washington used to delineate stands of Scots pine (Pinus sylvestris) in Spain with
State Department of Natural Resources, on the east slopes of the Cascades good results (Pascual et al., 2008). Stand delineation occurred through
in Washington State (Fig. 2). This forest is characteristic of Eastern object-based classification of the processed height and cover images
Cascade dry forests, dominated by ponderosa pine (P. ponderosa) and in SPRING software (Câmara et al., 1996). This produced a map of
associated mixed mesic conifers such as Douglas-fir (P. menziesii), grand forest stands that are similar in height and cover. By utilizing three
fir (Abies grandis), lodgepole pine (Pinus contorta), Engelmann spruce categories each of height and cover (low, medium and high), nine
(Picea engalmanii) and western larch (Larix occidentalis). Ahtanum is different stand types were produced. These categories were chosen
a working forest with active timber harvesting, leading to a wide variety arbitrarily based upon the relative grey scale of the image. This map
of forest stand types. The area is in the East Cascades ecoregion was then brought into ArcGIS (ESRI, Redlands, CA) to select plot
characterized by a dry climate with 50 cm of precipitation per year, locations using a stratified random sampling design. Plots were re-
mostly in the winter months as snow. Elevations within the study jected if placed on slopes of greater than 35° (70%) or if located in
area range from 960 to 1530 m. drainages. Plots located in areas of hillshade or extreme shadow were

Fig. 1. Flowchart for model building methods.


728 T.L. Erdody, L.M. Moskal / Remote Sensing of Environment 114 (2010) 725–737

Fig. 2. Location of Ahtanum State Forest in Washington State.

rejected due to issues of imagery analysis. This reduced the number of (dominant, co-dominant, intermediate and suppressed). Crown base
plots on north facing slopes (12 out of 57). height was defined as the lowest green foliage. In instances where the
canopy was unequal around the crown, the average lowest green
foliage was used. For trees under 7.62 cm DBH, only DBH was re-
2.4. Field data collection
corded. Heights and base heights for trees under 7.62 cm DBH were
then extrapolated using species specific allometric equations based on
Field data at Ahtanum were collected in July of 2008. A recreational
trees 7.62 to 20.32 cm DBH. Borderline trees were counted in if at least
hand-held GPS (Global Positioning System) was used to navigate to
50% of the trunk at DBH was inside the plot boundary. Data for dead
the plots. 57 0.0405 ha (404.69 m2) circular plots were used for the
trees were not collected nor included in analysis. The Impulse laser
study placed in the watersheds of the Middle and North Fork
range-finder was utilized for tree height and crown base height
Ahtanum Creeks and Foundation Creek (Fig. 3). A survey-grade
measurements. Since the LiDAR and imagery data are in US units (US
Maxor-GGD GPS (Javad Navigation Systems) was used to georefer-
survey feet) and the FuelCalc software produces most of its data in
ence plot centers, which acquired data for at least 15 min. GPS points
feet, field data collection was in feet.
were later post-processed and differentially corrected using the
WSRN (Washington State Reference Network) CORS (Continually
2.5. Derivation of variables
Operating Reference Station) YAKI base station in Yakima, WA using
Javad Pinnacle software. The base station was 40 km from the study
Plot level averages were derived for CH, CBH, CBD and ACF (Fig. 4)
area.
using FuelCalc software (Reinhardt et al., 2006a). This program uses
The plot boundaries (11.35 m radius) were measured using an
existing methods of calculating canopy fuels using published
Impulse laser range-finder (Laser Technology). Starting from the
allometric equations (Brown, 1978) and creates a canopy fuel profile
north and moving in a clockwise direction, the following parameters
as defined by Scott and Reinhardt (2001). Canopy bulk density (CBD)
were collected for each tree over 7.62 cm DBH: species, diameter at
is defined as the maximum 1.5 m running mean of canopy fuel density
breast height (DBH), height, crown base height and crown class
within a stand. In terms of crown fire initiation it is the mass per unit
volume of canopy biomass that would burn in a crown fire (twigs and
foliage b3 mm diameter). CBD is the most difficult to quantify due to
Table 1 multiple canopy layers (Perry et al., 2004) and has historically been
LiDAR specifications.
measured using inventory-based techniques and allometric equations
Collection date June, 2007 (Cruz et al., 2003; Scott & Reinhardt, 2001). CBD is an important
Aircraft Piper Navajo metric because programs such as FARSITE use a threshold value of
Sensor Optech 3100 EA
CBD for achieving and sustaining an active crown fire. Available
Laser pulse density N 4 points/m2
Laser pulse rate 100 kHz canopy fuel (ACF) is defined as foliage plus 50% of 0–3 mm live and 0–
Flying height 900 m 6 mm dead branchwood given in Mg/ha (Reinhardt et al., 2006b). ACF
Scan angle from nadir ± 15.5° is utilized in software such as the First Order Fire Effects Model
Horizontal accuracy ∼5 cm (FOFEM) to determine the proportion of the canopy that could
Vertical accuracy b 5–20 cm
potentially be consumed in a wildfire (Reinhardt et al., 1997, 2006a).
T.L. Erdody, L.M. Moskal / Remote Sensing of Environment 114 (2010) 725–737 729

Fig. 3. Ahtanum State Forest field plots.

This data can be then computed to determine emissions from transformed into a digital terrain model (DTM) grid using the
wildfires. GridSurfaceCreate tool. Using the ClipData tool, the DTM was then
The vertical distribution of canopy fuel is considered the zone in subtracted from the original point cloud to form a normalized point
which the canopy bulk density is greater than 0.012 kg/m3. Canopy cloud. The ClipData tool was used to clip out each plot from the
height (CH) is defined as the highest height at which the canopy bulk dataset at a radius of 11.35 m from the center point to match the field
density within a stand is greater than 0.012 kg/m3. Conversely, can- data. All LiDAR metrics were derived from these clipped and nor-
opy base height (CBH) is defined as the lowest height at which the malized point clouds.
canopy bulk density within a stand is greater than 0.012 kg/m3. In LiDAR and imagery metrics are shown in Table 2. LiDAR-based
terms of crown fire initiation CBH is the lowest height above the predictor variables used in similar studies (Andersen et al., 2005;
ground where there is enough canopy fuel to support a fire vertically Næsset & Bjerknes, 2001) were used for this research. These variables
through the canopy. While CBH is easy to measure for a single tree, it were calculated using the CloudMetrics tool in FUSION (McGaughey,
is difficult to measure for stand or plot level due to multiple canopy 2008). The use of these specific variables is preferred over the full set
layers and ladder fuels (Scott & Reinhardt, 2001). The discrepancy in of metrics derived from FUSION because they represent a simple yet
units between metric and US survey feet for FuelCalc outputs is a wide variety of explanatory variables that can be easily explained. A
function of the software settings. Output was delivered in US survey height break of 0.91 m was used to exclude all points below 0.91 m for
feet to be consistent with the units of the LiDAR and imagery data, the extraction of the variables. This removed any influence from
except CBD, which is delivered in metric by default. shrubs and vegetation which tended to be 0.91 m or less in height in
The Ahtanum LiDAR dataset used for this study was normalized for all plots.
slope using a ground model created from the all returns dataset using Nine LiDAR predictor variables were used and include maximum
FUSION software (McGaughey, 2008). The ground points were and mean height, quantile-based metrics of LiDAR height distribution
extracted from the point cloud using the GroundFilter tool and then (10th, 25th, 50th, 75th, 90th percentile heights), the coefficient of

Fig. 4. Canopy fuel metrics.


730 T.L. Erdody, L.M. Moskal / Remote Sensing of Environment 114 (2010) 725–737

Table 2 determination (R2) and the fit of the residuals. In the event of data
Metrics used as independent predictor variables in regression models. nonlinearity the Box–Cox procedure was used on the data to
LiDAR metric Description (units) determine which transformation was most appropriate (Kutner
et al., 2004). Variance of inflation tests (Fox & Monette, 1992) were
EMax Maximum height of point cloud (ft)
EMean Mean height of point cloud (ft) conducted for each selected model to look for colinearity between
ECV Coefficient of variation of point cloud heights (ft) independent variables. If colinearity was found between any
EP10 10th percentile point cloud height (ft) independent variables, another model was selected. Preferred models
EP25 25th percentile point cloud height (ft)
were chosen based on having a good balance between a high R2 value,
EP50 50th percentile point cloud height (ft)
EP75 75th percentile point cloud height (ft) a low root mean square error (RMSE), no colinearity and parsimony.
EP90 90th percentile point cloud height (ft) All best-fit models were validated using a leave-one-out cross-
D Density metric — # of first returns within the canopy (%) validation procedure using the PRESS statistic. From the PRESS sta-
tistic, an RMSE is derived (noted as RMSEcv for cross-validation) and
Imagery metric Description
is compared to the RMSE of the model. The RMSEcv will always be
B1Max NIR band — maximum DN value of plot larger than the RMSE but if the two values are close, it suggests that
B1Min NIR band — minimum DN value of plot
the model is not over fitting the data (Kutner et al., 2004). RMSEcv is
B1Mean NIR band — mean DN value of plot
B1SD NIR band — standard deviation DN values of plot the best indicator of model fit because the model has been tested and
B1CV NIR band — coefficient of variation DN values of plot validated.
B1Hom NIR band — homogeneity texture
B2Max Red band — maximum DN value of plot
3. Results and discussion
B2Min Red band — minimum DN value of plot
B2Mean Red band — mean DN value of plot
B2SD Red band — standard deviation DN values of plot The summary table of fuel and plot parameters for the Ahtanum
B2CV Red band — coefficient of variation DN values of plot data is shown in Table 3. The plots represented a wide amount of
B2Hom Red band — homogeneity texture
variation and gradation from east to west and in terms of elevation
B3Max Green band — maximum DN value of plot
B3Min Green band — minimum DN value of plot within the forest type.
B3Mean Green band — mean DN value of plot
B3SD Green band — standard deviation DN values of plot 3.1. Canopy fuel models
B3CV Green band — coefficient of variation DN values of plot
B3Hom Green band — homogeneity texture
NDVIMax NDVI — maximum DN value of plot The regression models organized by canopy fuel metric are shown
NDVIMin NDVI — minimum DN value of plot in Table 4. Both R2 and adjusted R2 values are shown. Adjusted R2, the
NDVIMean NDVI — mean DN value of plot coefficient of multiple determination, accounts for the number of
NDVISD NDVI — standard deviation DN values of plot
variables in the model and shows a truer representation of the model
NDVICV NDVI — coefficient of variation DN values of plot
without inflation due to additional variables. The R2 values will al-
ways be higher than the adjusted R2 values and the more parsi-
monious the model is, the less of a discrepancy between values. The
variation (CV) of LiDAR heights and canopy density. Canopy density, RMSE and RMSEcv values for CH are in original units. The RMSE and
D, is described as the number of first returns above a specified height RMSEcv values for CBH, CBD and ACF are in transformed values from
break divided by the total number of first returns (McGaughey, 2008). their original units. RMSE values in parentheses are back-transformed
Twenty-three imagery metrics were derived from the imagery values to show real differences in estimation accuracy between
using ArcGIS (ESRI, Redlands, CA) and ENVI (ITT, Boulder, CO) models.
software packages. These included plot level averages of the raw CBD and ACF models were log-transformed and CBH models were
maximum, minimum, mean, standard deviation (SD) and CV of the square-root transformed. When transformations occur in model
DN (digital number) values of each band (near-infrared, red and building, there are biases involved in the back-transformation, thus
green). The same metrics were calculated for a pseudo-NDVI correcting for bias can improve estimates of forest stand properties
(Normalized Difference Vegetation Index) using DN values of NIR (Andersen & Breidenbach, 2007). To get back-transformed values,
band and red band pixels. A co-occurance matrix-based texture with a correction factors were applied due to bias in the log back-transform
filter of 7 × 7 was run over the imagery to obtain a 2nd order (Sprugel, 1983) and the square-root back-transform (Miller, 1984).
homogeneity texture (Franklin et al., 2000; Moskal & Franklin, 2002; For CH (Table 4; Fig. 5), all models except the imagery model are
Wulder et al., 1998). A 7 × 7 window was chosen specifically for the comparable in their ability to explain most of the variance. The fusion
NIR band, which requires a smaller window size than the visible bands model has the highest R2 and lowest RMSEcv, but the LiDAR model
(Wulder et al., 1998) but for consistency the 7 × 7 window size was has the least discrepancy between RMSE and RMSEcv and is the most
used for all bands. These values were then averaged per plot and only
pixels with majority of their area within the plots were used for
analysis.
Table 3
2.6. Analysis Summary statistics of canopy fuel and plot parameters for Ahtanum (n = 57).

Parameter Min Max Mean (SD)


Predictive models of estimates of CH, CBH, CBD and ACF were
Canopy height (m) 6.71 37.8 20.69 (7.09)
developed using LiDAR-derived variables, imagery-derived predictor Canopy base height (m) 0.3 13.72 3.73 (3.34)
variables and a combination of the two for the fusion models. All Canopy bulk density (kg/m3) 0.0249 0.2808 0.0897 (0.0617)
models used data that were collected in feet, but values were con- Available canopy fuel (Mg/ha) 1.97 27.4 9.52 (6.52)
Trees per acre (≥7.62 cm DBH) 60 860 244.2 (174.5)
verted to metric units for results and discussion purposes.
Trees per acre (b7.62 cm DBH) 0 1110 178.8 (228.6)
All statistical analyses were processed with the software package R Trees per acre total 60 1670 423.98 (344.34)
(R Development Core Team, 2005). All possible subset regression Basal area (m2/ha) 1.2 14.65 5.24 (2.89)
(Miller, 2002) was used to determine the best model fit for the data at Canopy cover (%) FuelCalc 21.55 86.94 49.24 (18.74)
α = 0.05. The preferred models were chosen based on multiple cri- Canopy cover (%) FUSION 21.14 95.17 57.45 (20.35)
Slope (%) 1.75 70.02 16.66 (17.51)
teria, including Bayesian Information Criteria (BIC), the coefficient of
T.L. Erdody, L.M. Moskal / Remote Sensing of Environment 114 (2010) 725–737 731

Table 4
Regression models.

Sensor Model R2 Adj. R2 P value RMSE RMSEcv

Height (m)
LiDAR 1.527+0.228(EP25)+ 0.935 0.931 b 0.0001 1.86 2.03
0.871(EP90)+0.111(D)
Imagery 62.541 + 143.944 0.415 0.393 b 0.0001 5.52 5.86
(B1Hom) − 0.983
(B2SD)
LiDAR + Imagery 18.356+0.457(EP50)+ 0.957 0.954 b 0.0001 1.52 1.73
0.59(EP90)+23.614
(B1CV)−0.531(B3SD)

Canopy base height (m): CBH is square-root transformed. RMSE for these models are in square-root transformed values. Correction factors are added to the end of each model
LiDAR (3.254 + 0.13(EP25) 0.783 0.771 b 0.0001 0.68 (1.63) 0.73
− 0.021(EP90)
− 0.039(D)) + 0.214
Imagery (19.139+0.152(B1SD) 0.309 0.241 0.002 1.24 (3.60) 1.43
−19.961(B1CV)
−0.305(B2SD)+
13.019(B2CV)−29.774
(NDVIMean))+0.233
LiDAR + Imagery (7.98− 3.458(ECV)+ 0.843 0.825 b 0.0001 0.59 (1.44) 0.68
0.088(EP25)− 0.038
(D) + 0.033(B1Min)
−0.021(B1Mean)
−5.907(B3Hom)) +
0.125

Canopy bulk density (kg/m3): CBD is natural log-transformed. RMSE for these models are in natural log-transformed values. Correction factors are added to the end of each model
LiDAR (−3.696 + 0.024 0.831 0.821 b 0.0001 0.25 (0.025) 0.29
(EP25) − 0.03(EP50) +
0.029(D)) ⁎ 1.032
Imagery (− 3.341 − 0.017 0.602 0.572 b 0.0001 0.39 (0.041) 0.42
(B1Min) − 8.119
(NDVIMax) + 12.16
(NDVIMean) + 5.335
(NDVICV)) ⁎ 1.078
LiDAR + Imagery (− 6.141 − 0.037 0.882 0.865 b 0.0001 0.22 (0.024) 0.25
(EMean) + 0.028
(EP25) + 0.03(D)
− 0.014(B1Min) +
0.012(B1Max) +
0.016(B3Mean)
− 0.026
(B3SD)) ⁎ 1.024

Available canopy fuel (Mg/ha): ACF is natural log-transformed. RMSE for these models are in natural log-transformed values. Correction factors are added to the end of each model
LiDAR (−0.536 + 0.031 0.879 0.877 b 0.0001 0.23 (2.29) 0.24
(D)) ⁎ 1.028
Imagery (1.092 − 0.024 0.671 0.638 b 0.0001 0.40 (3.61) 0.44
(B1Min) − 0.024
(B3SD) − 7.54
(NDVIMax) + 12.112
(NDVIMean) + 5.327
(NDVICV)) ⁎ 1.084
LiDAR + Imagery (− 0.467 + 0.032(D) 0.906 0.901 b 0.0001 0.21 (2.33) 0.23
− 0.02(B3SD) + 4.02
(NDVISD)) ⁎ 1.022

parsimonious model. The addition of imagery to LiDAR improved R2 parsimonious model. The addition of imagery to LiDAR improved R2
from 0.94 to 0.96 and reduced the RMSEcv from 2.03 to 1.73. from 0.88 to 0.91 and reduced the RMSEcv from 0.24 to 0.23.
For CBH (Table 4; Fig. 5), the fusion model has the highest R2 and For all metrics, the LiDAR models seem to be the most parsi-
lowest RMSEcv, but the LiDAR model has the least discrepancy monious, but the fusion models explain more variability and are the
between RMSE and RMSEcv and is the most parsimonious model. The best fit, based on RMSEcv and R2 values. As the results show, the
addition of imagery to LiDAR improved R2 from 0.78 to 0.84 and fusion models, for the most part, improve the LiDAR models in varying
reduced the RMSEcv from 0.73 to 0.68. degrees. The fusion models had higher R2 values and lower RMSEcvs
For CBD (Table 4; Fig. 5), the fusion model has the highest R2 and than the LiDAR or imagery models in all cases.
lowest RMSEcv and has the least discrepancy between RMSE and While no significance testing was performed to compare the
RMSEcv. Here, the addition of imagery to LiDAR improved R2 from models, the goal here is to present the models and associated meta-
0.83 to 0.88 and reduced the RMSEcv from 0.29 to 0.25. data so informed decisions can be made. If the goal is to get the lowest
For ACF (Table 4; Fig. 5), the fusion model has the highest R2 and RMSEcv then the fusion models should be chosen, depending on the
shares a low RMSEcv with the LiDAR model, but the LiDAR model has variable of interest. If simple, parsimonious models are preferred then
the least discrepancy between RMSE and RMSEcv and is the most the LiDAR only models should be chosen. While the LiDAR models do
732 T.L. Erdody, L.M. Moskal / Remote Sensing of Environment 114 (2010) 725–737

Fig. 5. Field measured versus predicted canopy fuel metrics: a) canopy height, b) square-root transformed canopy base height, c) natural log-transformed canopy bulk density, and
d) natural log-transformed available canopy fuel. Solid lines show 1:1 relationship.
T.L. Erdody, L.M. Moskal / Remote Sensing of Environment 114 (2010) 725–737 733

not have the lowest RMSEcv values, they possess small reductions in sensor as well as the optical characteristics of the lens leading to radial
accuracy and the extra error avoided by having less explanatory distortion. Tree canopies that are within a LiDAR plot boundary might
variables is beneficial. be considered out in an image if the tree is “leaning” outside of the
plot. This also works in reverse, with trees “leaning” into plots that do
3.2. Model summary not originate in the plot. This can lead to errors in the measures of DN
values for a plot and thus, errors in estimating canopy fuel metrics.
The LiDAR metrics clearly explained the majority of the variability Considering that imagery explains little variation in relation to what
in the models and averaged over all 4 metrics, the addition of imagery LiDAR explains, there is no need to utilize these data sets for canopy
to the LiDAR models explained an extra 3.4% of the variation. EP10 and fuel estimation, especially if cost-effective and time-efficient methods
EP25 are obvious predictors for CBH due to their vertical location are desired.
within the point cloud, which is similar to the range of vertical
locations of CBH within a forest stand. D, the canopy cover metric, was 3.3. Map creation from LiDAR models
instrumental in predicting significant amounts of variation for CBD
and ACF while adding significant explanatory power to CBH. This is Maps were made with a pixel resolution of 20 m over the study
consistent with other research that found D to be a key predictor of area, which is 66.89 km2. This consists of 45 LiDAR tiles that are each
biomass metrics (Krasnow et al., 2009; Li et al., 2008). 1.49 km2 in area. Fig. 6 shows the spatially explicit model images for
For imagery, the biomass fuel estimates were heavily influenced by canopy height, canopy base height, canopy bulk density and available
the various NDVI metrics. This is supported by other research which has canopy fuel. White areas on the images indicate open areas with no
found NDVI to be well correlated with woody biomass in a variety of canopy cover. Areas with both low CBH values (dark blue) and high
forest types (Dong et al., 2003; Zheng et al., 2004). Imagery data, while CBD or ACF values (dark red) are indicative of north facing slopes and
good for species identification, land cover and fuel classification, did not drainage bottoms. From a practical standpoint, the north facing slopes
estimate canopy fuels well in general. The lack of data beneath the are the most interesting as they suggest where fire managers should
canopy is a major reason why imagery is not a good estimator of three- be focusing fuel treatment efforts due to high fire risk. These areas
dimensional variables including canopy fuels. This is especially true for should also be analyzed with caution due to the sampling scheme of
CBH which is visible from the ground but not from above the canopy. the research which is discussed further in the next section.
Two trees that look similar from above could have different CBH values
that cannot be determined from the spectral signature of the tree 3.4. Sources of error
crowns. Shadows created by topography (i.e.-hillshade) and trees (i.e.-
self-shadowing) can impact the estimation of canopy fuels because DN 3.4.1. Field data collection
values are not consistent throughout the study area. In field measurement, every tree over 7.62 cm DBH was measured
Image parallax affects imagery data because tree crowns might be for height and base height. In dense plots and in plots with tall trees,
skewed about a plot. While LiDAR points are in three-dimensional getting accurate tree heights proved difficult due to interference from
space, they will always be in the same place relative to a center point other tree crowns. Every effort was made to re-measure if the initial
of a plot. In imagery, tree crown locations are impacted by the angle at measurement was incorrect. Since all trees less than 7.62 cm DBH
which the image was taken, the distance between the crowns and the only had DBH recorded, it was essential to use allometric equations to

Fig. 6. LiDAR-derived canopy fuel metrics.


734 T.L. Erdody, L.M. Moskal / Remote Sensing of Environment 114 (2010) 725–737

estimate height and base height. FuelCalc adjusts an individual tree's Due to hillshade and extreme shadowing, the rejection of plots in
canopy weight based on its canopy class. As the majority of the trees these areas reduced the total number of plots on north facing aspects.
less than 7.62 cm DBH were in the suppressed canopy class, they This might have reduced the possible variation from these slopes
contributed little to overall plot weights. Even though these small leading to potentially inaccurate estimates on north facing slopes. If a
trees did not affect the plot level canopy fuel metrics significantly, not LiDAR only study was conducted, there would be no effect of hillshade
measuring the heights and base heights reduces the integrity of the on plot placement and thus, more equality in sampling. This would
estimations. As with previous research, sampling a representative have been the ideal way to sample more of the variation on the denser
selection of trees for height and base height within a plot and then north facing slopes. Due to the lack of plots in areas of hillshade, these
using regression to obtain the same measurements for the remaining specific areas on the map products in Fig. 6 should be interpreted
trees leads to potential unwanted variability (Andersen et al., 2005). carefully.
This can be addressed by measuring all heights and base heights for
trees over DBH height. 3.4.3. Forest structure
There were some issues with borderline trees that had canopies both The issue of how stand density affects the regression models is
in and out of the plot. If a tree was counted in then the whole tree crown, another potential source of error. As stand density increases, the
including foliage outside the plot was measured. This matters due to the likelihood that laser pulses penetrate below the canopy decreases and
clipping of the LiDAR point cloud for analysis. The plot point cloud is leads to occlusion of areas below the canopy. This can have a sig-
missing points for trees that were in that had foliage outside the plot and nificant impact on the density of points below the canopy which can
contains points for trees that were out but had foliage inside the plot. affect the metrics produced from the data. The biomass metrics, CBD
Assumptions are made that these points would cancel each other out. and ACF, could be affected greatly, especially if these metrics saturate
Additional error could be due to the one year gap between when at high values, as in Leaf Area Index and NDVI estimates (Richardson
the LiDAR was flown in June of 2007 and when field data were et al., 2009). If a stand is too dense, the number of ground points could
collected in July of 2008. For the Cascade region, tree height growth be reduced, leading to inaccurate computation of the LiDAR canopy
rates average about 37 cm/year for ponderosa pine and 39 cm/year for cover metric, D, which is present in many of the models. Although
Douglas-fir (Ritchie & Hann, 1989), but this is dependent upon there was no clear evidence of occlusion and saturation in this study,
elevation, soil, temperature and rainfall. This is most likely within the the effect of stand density on the models is important and needs to be
acceptable error considering that LiDAR routinely underestimates looked into further and can be corrected by transformation (Lefsky
height (Andersen et al., 2006; Lefsky et al., 2002b; Lim et al., 2003) et al., 2002a).
and that the output from FuelCalc is rounded to the nearest foot. Effects of species composition can affect how the models behave,
Limitations of using FuelCalc software include how the program especially with differences in crown architecture and reflectivity.
calculates the canopy fuel metrics using allometric equations. As with Utilizing InSAR data could likely produce more accurate estimates of
most biomass and forest metric estimations, allometric equations are biomass and canopy structure using texture metrics (Andersen et al.,
utilized for ease of use and offer quick results without destructively 2008), especially at longer wavelengths (Hyde et al., 2006). Although
sampling. Utilizing these equations will naturally add some error metrics from radar are highly sensitive to the viewing angle of the
because some trees might be out of the range of calculation (i.e., DBH sensor, its use would be beneficial for forest measurements as well as
is too large) or equations for one species might be used for the same cost-effective (Andersen et al., 2008).
species in another location that is growing in different conditions and
will have different properties. Additionally, FuelCalc does not account 3.4.4. Map interpretation
for dead trees in its equations. This is understandable as the focus of Caution should be used when interpreting data on map edges and
the program is to get information on live foliage, small branches and LiDAR tile edges due to the nature of how the maps were created and
canopy height parameters. While dead trees were encountered in merged. Edges between areas of open ground (scree slopes, grass) and
most of the plots, they were mostly only trunks without small forest stands or individual trees should be interpreted with caution due
branches and limbs. When a recently dead tree with brown foliage to pixels with minimal LiDAR points from overhanging tree canopies. In
was encountered, data was collected. This was due to the fact that some cases this has led to inflated values of CBH that are higher than
LiDAR was collected one year previously while the tree was either values of CH. This is obviously not feasible and to remedy this, the value
living or possessing similar amounts of dead foliage. The effect of of CBH was set to 0 if the value of CBH was higher than CH.
omitting dead trees from analysis of canopy fuel loadings needs to be The maps developed using the LiDAR models should be used with
looked at further to address these issues. caution in forest types not represented in the model building, areas on
the western boundary and areas above 1550 m elevation. The sub-
3.4.2. Remote sensing data alpine forests on the western edge and higher elevations (above
The imagery used in this study was not normalized for reflectance 1550 m) of Ahtanum state forest are dominated by sub-alpine fir (Abies
or shadow and only raw DN values were used. Because the imagery is lasiocarpa) and whitebark pine (Pinus albicaulus) which possess dif-
61 cm resolution, traditional methods to correct the imagery for ferent biophysical properties and stand structure than the study site.
shadow are difficult due to self-shading created by the trees them- Extrapolating these models for use in similar forest types should be
selves. This was found to be true for ponderosa pine tree crowns using used with caution. While these forests might contain similar tree
IKONOS imagery (Vierling et al., 2002). Because the imagery was not species, differences due to local and regional variability can affect forests
being used for change detection or physical models, reflectance values from bottom-up (i.e., nutrient and water availability) to top-down
did not need to be calibrated through correction. Proper correction for controls (i.e., climate). While extrapolation within the Ahtanum study
reflectance would require readings from a spectroradiometer at the area is reasonable due to the similarity of these controls, care should be
field site during the time of the flight or another suitable atmospheric taken for extrapolation to other regions in eastern Washington State.
correction method to translate the DN to reflectance values. Because
the imagery for this study was acquired from the vendor and originally 4. Conclusions
flown for general purpose no arrangements were made for spectral
data collection and reflectance calibration of the sensor. Due to the fact Studies have shown that remote sensing technologies including
that just raw DN values were used, the calculation of NDVI could only LiDAR can estimate canopy fuels (Andersen et al., 2005; Holmgren &
be considered a pseudo-NDVI. Atmospheric correction is needed for Persson, 2004; Riano et al., 2003). This research provides further
NDVI values to be properly analyzed (Xiong et al., 2006). evidence that this can be achieved in a fire-prone ponderosa pine
T.L. Erdody, L.M. Moskal / Remote Sensing of Environment 114 (2010) 725–737 735

dominated forest of the Pacific Northwest. Additionally, significant FARSITE and FlamMap models require landscape (.LCP) files that
minor improvements are seen within the models when utilizing raw consist of layers of raster files. Some of these layers, include elevation,
aerial imagery data and fusing them with LiDAR point data. slope, aspect and canopy cover can be acquired directly from LiDAR
Through estimation of canopy fuels by LiDAR and imagery, this data. By utilizing other methods to acquire fuel models from LiDAR
study showed that LiDAR is superior to near-infrared imagery at alone (Seielstad & Queen, 2003) or LiDAR fused with imagery (Mutlu
predicting canopy fuel metrics. Furthermore, the addition of imagery et al., 2008a), the whole suite of LCP data layers can be acquired from
to LiDAR aids in the estimation of canopy fuel metrics. By fusing LiDAR one LiDAR dataset.
with imagery, results were obtained that maximized the strengths of Research involving other data fusion could potentially be useful in
each sensor. A variety of recent studies has shown that the fusion of mapping canopy fuels that are suffering from poor health or insect
sensors have improved accuracies of measuring canopy height and attack. By utilizing LiDAR intensity, differences between healthy green
biomass as well as aiding in tree crown identification, tree species trees and recently dead trees (with needles) might be found. Fusion of
identification, and surface fuel mapping than by one sensor alone intensity values with LiDAR data have been used for individual species
(Hyde et al., 2006; McCombs et al., 2003; Mutlu et al., 2008a; Perrson identification and classification (Brandtberg, 2007; Holmgren &
et al., 2004). Persson, 2004; Kim et al., 2009) which could potentially pick up
Additionally, this research further reinforces the strengths of LiDAR differences in tree health. This would be beneficial in determining
relative to high resolution near-infrared imagery. Unlike imagery, forest areas prone to future insect attack and to existing fire risk while
there are no issues with shadow and tree self-shading and the nature assessing canopy fuel loading. Imagery could be used for health and
of LiDAR data enables one to partially “see” beneath the canopy. The insect issues and would require less processing time because changes
LiDAR metrics utilized in this study reflect this strength as they in forest health (i.e.-green versus red needles) can be clearly seen and
represent a vertical structure that cannot be obtained with imagery. delineated.
Conversely, issues with LiDAR data do exist but they are not as critical. Although canopy fuels represent a fraction of the total biomass
The main problem that can arise is that some areas might not get consumed in a wildfire (Campbell et al., 2007; Fahnestock & Agee,
scanned at all due to blocking of laser pulses by topography or 1983), this data, in conjunction with spatially heterogeneous ground
vegetation, especially if the scan angle is extremely shallow. Despite fuel and woody debris data could produce more accurate maps of total
this, LiDAR is still far superior to high resolution near-infrared imagery fuel loading. By using ground-based LiDAR technology, estimations of
in estimating canopy fuels. ground fuels in some forest types are being accomplished (Loudermilk
et al., 2007). This would lead to greater estimation accuracy of emiss-
4.1. Trade-offs: cost versus accuracy ions production on the whole fuel bed. Fire managers that utilize
models such as FARSITE can benefit from more accurate data down to
One important consideration to think about before acquiring 10 m resolution. Additionally, by resolving canopy fuels down to 1 m
remotely sensed data is a cost benefit analysis. Cost analysis should be resolution using voxel-based methods (Chasmer et al., 2004; Popescu
taken into consideration when asking if it is necessary to use two & Zhao, 2008), which are more suited as inputs into three-
different sensors for a marginal increase in accuracy when using just dimensional physical fire behavior models such as FIRETEC/HIGRAD
one (LiDAR) gives excellent results. At the time of this research, the (Linn, 1997; Linn et al., 2002), the possibilities of extremely accurate
cost of acquiring aerial LiDAR for a large survey area (50,000 ha+) and mapping are foreseeable in the future.
at a suitable laser pulse density (6 to 8 pulses/m2) is about US$3.00 While this research improves estimates of canopy fuels, the appli-
per ha plus US$10–20,000 mobilization costs (based on Puget Sound cations of LiDAR-derived fuel layers for wildfire behavior modeling
LiDAR Consortium 2008 cost). This is much higher compared to the and emission estimates are beyond the scope of this work. Future
cost of acquiring digital aerial imagery for the same area at just over 1$ research is needed to encompass these spatially explicit canopy fuel
per ha but similar mobilization costs. data into a FARSITE wildfire simulation to assess the accuracy of the
Even though the cost to obtain LiDAR is higher, one needs to data against a real wildfire. This has been previously done by com-
evaluate if it is worth paying extra for imagery if you already have paring the current suite of canopy fuel data from LANDFIRE versus
LiDAR data. For example, the core Ahtanum study area covered and models derived from locally available biotic and abiotic data. This
shown in Figs. 3 and 5 is 6689 ha. Ignoring mobilization costs, the study found that data from LANDFIRE are inadequate for accounting
LiDAR would cost about $20,067 and the imagery would cost about for local variability and that there is a need to map fuels on a local
$8026. For the additional cost of the imagery an increase in accuracy scale (Krasnow et al., 2009). By producing data at a local level instead
would be seen in the four canopy fuel metrics of 3 to 4%. It is unlikely at a larger scale (i.e., regional of national), more accurate estimations
that management decision and practices would drastically change of fuel loading, fire behavior and emissions can be achieved and fire
with the additional improvements in accuracy, thus, the cost to managers can have complete control and knowledge of their data.
accuracy trade-off suggests that imagery is unnecessary for this ap-
plication. However, if applications such as forest health mapping, Acknowledgements
knowledge of species, or vegetation health characterization are de-
sired, a different conclusion might be reached. Further applied re- This research has been supported by a grant from the U.S.
search in a range of applications is needed to finalize such conclusions. Environmental Protection Agency's Science to Achieve Results
In this study the LiDAR and imagery were acquired at two different (STAR) program and carried out in the University of Washington
times by two different vendors. If the LiDAR and imagery were (UW) Remote Sensing and Geospatial Analysis Laboratory (RSGAL).
acquired at the same time on the same flight, then the cost would be Additional funding was provided by the Byron and Alice Lockwood
reduced substantially (Renslow et al., 2000) although the small time Fellowship. LiDAR and imagery were provided by the Washington
frame when imagery can be flown will be a limiting factor to the State Department of Natural Resources (WADNR) and ImageTree
LiDAR data collection. If this were possible, then a good argument Corporation. The Precision Forestry Cooperative at UW is acknowl-
could be made for acquiring data from both sensors. edged for providing research resources and administrative assistance.
The authors would like to acknowledge the following individuals for
4.2. Future research their involvement in the research: Hans-Erik Andersen, Thomas
Hinckley, Jeffrey Richardson, Guang Zheng, Andy Cockle, Jacob Strunk,
LiDAR has been proven to estimate canopy fuels for fire models but Alicia Sullivan, Robert McGaughey, Steve Reutebuch, Clint Wright and
the uses of LiDAR for this application are not limited to this research. Ernesto Alvarado.
736 T.L. Erdody, L.M. Moskal / Remote Sensing of Environment 114 (2010) 725–737

References Haralick, R. M. (1979). Statistical and structural approaches to texture. Proceedings of


the IEEE, 67, 786−804.
Agee, J. K. (1993). Fire ecology of Pacific Northwest forests. Washington, D.C.: Island Press. Hardy, C. C., Ottmar, R. D., Peterson, J. L., Core, J. E., & Seamon, P. (2001). Smoke
Andersen, H. -E., & Breidenbach, J. (2007). Statistical properties of mean stand biomass management guide for prescribed and wildland fire: 2001 edition.Boise, ID: National
estimators in a LIDAR-based double sampling forest survey design. International Wildfire Coordinating Group 226 pp.
Society for Photogrammetry and Remote Sensing Istanbul, Turkey. 7 pp. Holmgren, J., & Persson, A. (2004). Identifying species of individual trees using airborne
Andersen, H. -E., McGaughey, R. J., & Reutebuch, S. E. (2005). Estimating forest canopy laser scanner. Remote Sensing of Environment, 90, 415−423.
fuel parameters using LIDAR data. Remote Sensing of Environment, 94, 441−449. Hudak, A. T., Lefsky, M. A., Cohen, W. B., & Berterretche, M. (2002). Integration of LiDAR
Andersen, H. -E., Mcgaughey, R. J., & Reutebuch, S. E. (2008). Assessing the influence of and Landsat ETM+ data for estimating and mapping forest canopy height. Remote
flight parameters, interferometric processing, slope and canopy density on the Sensing of Environment, 82, 397−416.
accuracy of X-band IFSAR-derived forest canopy height models. International Hyde, P., Dubayah, R., Walker, W., Blair, J. B., Hofton, M., & Hunsaker, C. (2006). Mapping
Journal of Remote Sensing, 29, 1495−1510. forest structure for wildlife habitat analysis using multi-sensor (LiDAR, SAR/InSAR,
Andersen, H. -E., Reutebuch, S. E., & McGaughey, R. J. (2006). A rigorous assessment of ETM+, Quickbird) synergy. Remote Sensing of Environment, 102, 63−73.
tree height measurements obtained using airborne LiDAR and conventional field Jia, G. J., Burke, I. C., Kaufmann, M. R., Goetz, A. F. H., Kindel, B. C., & Pu, Y. (2006).
methods. Canadian Journal of Remote Sensing, 32, 355−366. Estimates of forest canopy fuel attributes using hyperspectral data. Forest Ecology
Arroyo, L. A., Healey, S. P., Cohen, W. B., Cocero, D., & Manzanera, J. A. (2006). Using and Management, 229, 27−38.
object-oriented classification and high-resolution imagery to map fuel types in a Keane, R. E., Frescino, T., Reeves, M. C., & Long, J. L. (2006). Mapping wildland fuel across
Mediterranean region. Journal of Geophysical Research, 111. large regions for the LANDFIRE prototype project. In M. G. Rollins, & C. K. Frame (Eds.),
Arroyo, L. A., Pascual, C., & Manzanera, J. A. (2008). Fire models and methods to map fuel The LANDFIRE Prototype Project: nationally consistent and locally relevant geospatial
types: The role of remote sensing. Forest Ecology and Management, 256, 1239−1252. data for wildland fire management. Gen. Tech. Rep. RMRS-GTR-175 (pp. 367−396).
Battye, W., & Battye, R. (2002). Development of emissions inventory methods for Fort Collins, CO: U.S. Department of Agriculture, Forest Service, Rocky Mountain
wildland fire. U.S. Environmental Protection Agency Contract No. 68-D-98-046. Research Station.
91 pp. Kim, S., McGaughey, R. J., Andersen, H. -E., & Schreuder, G. (2009). Tree species
Bertolette, D. R., & Spotskey, D. B. (1999). Fuel model and forest type mapping for differentiation using intensity data derived from leaf-on and leaf-off airborne laser
FARSITE. The joint fire science conference and workshop Boise, ID: University of Idaho scanning data. Remote Sensing of Environment, 113, 1575−1586.
and International Association of Wildland Fire. Kourtz, P. H. (1977). An application of Landsat digital technology to forest fire fuel type
Brandtberg, T. (2007). Classifying individual tree species under leaf-off and leaf-on mapping. 11th International Symposium on Remote Sensing of Environment Ann
conditions using airborne LiDAR. ISPRS Journal of Photogrammetry and Remote Arbor, MI1111-1115.
Sensing, 61, 325−340. Krasnow, K., Schoennagel, T., & Veblen, T. T. (2009). Forest fuel mapping and evaluation
Breidenbach, J., Koch, B., Kandler, G., & Kleusberg, A. (2008). Quantifying the influence of LANDFIRE fuel maps in Boulder County, Colorado, USA. Forest Ecology and
of slope, aspect, crown shape and stem density on the estimation of tree height at Management, 257, 1603−1612.
plot level using LiDAR and InSAR data. International Journal of Remote Sensing, 29, Kutner, M. H., Nachtsheim, C., & Neter, J. (2004). Applied linear regression models.
1511−1536. Boston; New York: McGraw-Hill/Irwin.
Brown, J. K. (1978). Weight and density of crowns of Rocky Mountain conifers. Res. Pap. Lefsky, M. A., Cohen, W. B., Harding, D. J., Parker, G. G., Acker, S. A., & Gower, S. T. (2002).
INT-197.Ogden, UT: U.S. Department of Agriculture, Forest Service, Intermountain LiDAR remote sensing of above-ground biomass in three biomes. Global Ecology &
Forest and Range Experiment Station 56 pp. Biogeography, 11, 393−399.
Câmara, G., Souza, R. C. M., Freitas, U. M., & Garrido, J. (1996). Spring: Integrating remote Lefsky, M. A., Cohen, W. B., Parker, G. G., & Harding, D. J. (2002). LiDAR remote sensing
sensing and gis by object-oriented data modelling. Computers & Graphics, 20, 395−403. for ecosystem studies. Bioscience, 52, 19−30.
Campbell, J., Donato, D., Azuma, D., & Law, B. (2007). Pyrogenic carbon emission from a Li, Y., Andersen, H. -E., & McGaughey, R. (2008). A comparison of statistical methods for
large wildfire in Oregon, United States. Journal of Geophysical Research, 112. estimating forest biomass from light detection and ranging data. Western Journal of
Chasmer, L., Hopkinson, C., & Treitz, P. (2004). Assessing the three-dimensional Applied Forestry, 23, 223−231.
frequency distribution of airborne and ground-based LiDAR data for red pine and Lim, K., Treitz, P., Wulder, M., St-Onge, B., & Flood, M. (2003). LiDAR remote sensing of
mixed deciduous forest plots. Proceedings of the ISPRS working group VIII/2 ‘Laser- forest structure. Progress in Physical Geography, 27, 88−106.
Scanners for Forest and Landscape Assessment’ Freiburg, Germany: International Linn, R. (1997). Transport model for prediction of wildfire behavior. Scientific Report LA
Archives of Photogrammetry, Remote Sensing And Spatial Information Sciences. 13334-T. : Los Alamos National Laboratory.
Coops, N. C., Wulder, M. A., Culvenor, D. S., & St-Onge, B. (2004). Comparison of forest Linn, R., Reisner, J., Colman, J. J., & Winterkamp, J. (2002). Studying wildfire behavior
attributes extracted from fine spatial resolution multispectral and LiDAR data. using FIRETEC. International Journal of Wildland Fire, 11, 233−246.
Canadian Journal of Remote Sensing, 30, 855−866. Loudermilk, E. L., Singhania, A., Fernandez, J. C., Hiers, J. K., O'Brien, J. J., Cropper, W. P.,
Cruz, M. G., Alexander, M. E., & Wakimoto, R. H. (2003). Assessing canopy fuel stratum Jr., et al. (2007). Application of ground-based LIDAR for fine-scale forest fuel
characteristics in crown fire prone fuel types of western North America. Interna- modeling. In B. W. Butler, & W. Cook (Eds.), The fire environment — Innovations,
tional Journal of Wildland Fire, 12, 39−50. management, and policy: Proceedings RMRS-P-46CD Destin, FL: U.S. Department of
De Wulf, R. R., Goossens, R. E., De Roover, B. P., & Borry, F. C. (1990). Extraction of forest Agriculture, Forest Service, Rocky Mountain Research Station, Fort Collins, CO
stand parameters from panchromatic and multispectral SPOT-1 data. International 662 pp.
Journal of Remote Sensing, 11, 1571−1588. McCombs, J. W., Roberts, S. D., & Evans, D. L. (2003). Influence of fusing LiDAR and
Dong, J., Kaufmann, R. K., Myneni, R. B., Tucker, C. J., Kauppi, P. E., Liski, J., et al. (2003). multispectral imagery on remotely sensed estimates of stand density and mean
Remote sensing estimates of boreal and temperate forest woody biomass: Carbon tree height in a managed Loblolly Pine Plantation. Forest Science, 49, 457−466.
pools, sources, and sinks. Remote Sensing of Environment, 84, 393−410. McGaughey, R. J. (2008). FUSION/LDV: Software for LIDAR data analysis and visualization. :
Donoghue, D. N. M., & Watt, P. J. (2006). Using LiDAR to compare forest height estimates USDA Forest Service, Pacific Northwest Research Station.
from IKONOS and Landsat ETM+ data in Sitka spruce plantation forests. Interna- Means, J. E., Acker, S. A., Fitt, B. J., Renslow, M., Emerson, L., & Hendrix, C. J. (2000).
tional Journal of Remote Sensing, 27, 2161−2175. Predicting forest stand characteristics with airborne scanning LiDAR. Photogram-
Fahnestock, G. R., & Agee, J. K. (1983). Biomass consumption and smoke production by metric Engineering and Remote Sensing, 66, 1367−1371.
prehistoric and modern forest fires in Western Washington. Journal of Forestry, 81, Miller, A. (2002). Subset selection in regression. : Chapman & Hall/CRC.
653−657. Miller, D. M. (1984). Reducing transformation bias in curve fitting. The American
Falkowski, M. J., Gessler, P. E., Morgan, P., Hudak, A. T., & Smith, A. M. S. (2005). Statistician, 38, 124−126.
Characterizing and mapping forest fire fuels using ASTER imagery and gradient Morsdorf, F., Meier, E., Kotz, B., Itten, K. I., Dobbertin, M., & Allgower, B. (2004). LIDAR-
modeling. Forest Ecology and Management, 217, 129−146. based geometric reconstruction of boreal type forest stands at single tree level
Finney, M. A. (1998). FARSITE: Fire area simulator-model development and evaluation. RMRS- for forest and wildland fire management. Remote Sensing of Environment, 92,
RP-4.Ogden, UT: USDA, Forest Service, Rocky Mountain Research Station 47 pp. 353−362.
Finney, M. A. (2006). An overview of FlamMap fire modeling capabilities. In P. L. Moskal, L. M., & Franklin, S. E. (2002). Multistory forest stand discrimination with
Andrews, & B. W. Butler (Eds.), Fuels management—How to measure success: multiscale texture from high spatial detail airborne imagery. Geocarto International,
Conference proceedingsRocky Mountain Research Station Proceedings RMRS-P-41. 17, 53−66.
(pp. 213−220) Portland, OR: USDA Forest Service. Mutlu, M., Popescu, S. C., Stripling, C., & Spencer, T. (2008). Mapping surface fuel models
Finney, M. A., & Andrews, P. L. (1994). The FARSITE fire area simulator: Fire management using LiDAR and multispectral data fusion for fire behavior. Remote Sensing of
applications and lessons of summer 1994. Proceedings of the International Fire Council Environment, 112, 274−285.
Meeting ID: Coeur D'Alene. Mutlu, M., Popescu, S. C., & Zhao, K. (2008). Sensitivity analysis of fire behavior
Fox, J., & Monette, G. (1992). Generalized collinearity diagnostics. Journal of the American modeling with LIDAR-derived surface fuel maps. Forest Ecology and Management,
Statistical Association, 87, 178−183. 256, 289−294.
Franklin, S. E., Hall, R. J., Moskal, L. M., Maudie, A. J., & Lavigne, M. B. (2000). Næsset, E., & Bjerknes, K. -O. (2001). Estimating tree heights and number of stems in
Incorporating texture into classification of forest species composition from young forest stands using airborne laser scanner data. Remote Sensing of
airborne multispectral images. International Journal of Remote Sensing, 21, 61−79. Environment, 78, 328−340.
Hall, S. A., Burke, I. C., Box, D. O., Kaufmann, M. R., & Stoker, J. M. (2005). Estimating Pascual, C., García-Abril, A., García-Montero, L. G., Martín-Fernández, S., & Cohen, W. B.
stand structure using discrete-return LiDAR: An example from low density, fire (2008). Object-based semi-automatic approach for forest structure characteriza-
prone ponderosa pine forests. Forest Ecology and Management, 208, 189−209. tion using LiDAR data in heterogeneous Pinus sylvestris stands. Forest Ecology and
Halligan, K. (2007). Mapping forest canopy fuels in Yellowstone National Park using Management, 255, 3677−3685.
LiDAR and hyperspectral data. Dissertation thesis, University of California, Santa Perrson, A., Holmgren, J., Soderman, U., & Olssen, H. (2004). Tree species classification of
Barbara. individual trees in Sweden by combining high resolution laser data with high
T.L. Erdody, L.M. Moskal / Remote Sensing of Environment 114 (2010) 725–737 737

resolution near-infrared digital images. Proceedings of the ISPRS WG VIII/2, October Tech. Rep. RMRS-GTR-175.Fort Collins, CO: U.S. Department of Agriculture, Forest
2004 (pp. 204−207). Freiburg, Germany. Service, Rocky Mountain Research Station 416 pp..
Perry, D. A., Jing, H., Youngblood, A., & Oetter, D. R. (2004). Forest structure and fire Running, S. W. (2006). CLIMATE CHANGE: Is global warming causing more, larger
susceptibility in volcanic landscapes of the Eastern High Cascades, Oregon. Con- wildfires? Science, 313, 927−928.
servation Biology, 18, 913−926. Schmidt, K. M., Menakis, J. P., Hardy, C. C., Hann, W. J., & Bunnell, D. L. (2002). Deve-
Popescu, S. C., & Wynne, R. H. (2004). Seeing the trees in the forest: Using LiDAR and lopment of coarse-scale spatial data for wildland fire and fuel management. General
multispectral data fusion with local filtering and variable window size for estimating Technical Report RMRS-GTR-87.Fort Collins, CO: U.S. Department of Agriculture,
tree height. Photogrammetric Engineering and Remote Sensing, 70, 589−604. Forest Service, Rocky Mountain Research Station 41 pp.
Popescu, S. C., & Zhao, K. (2008). A voxel-based LiDAR method for estimating crown base Scott, K., Oswald, B., Farrish, K., & Unger, D. (2002). Fuel loading prediction models
height for deciduous and pine trees. Remote Sensing of Environment, 112, 767−781. developed from aerial photographs of the Sangre de Cristo and Jemez mountains of
R Development Core Team. (2005). R: A language and environment for statistical computing New Mexico, USA. International Journal of Wildland Fire, 11, 85−90.
Vienna, Austria: R Foundation for Statistical Computing http://www.R-project.org. Scott, J., & Reinhardt, E. (2001). Assessing crown fire potential by linking models of surface
Reeves, M. C., Kost, J. R., & Ryan, K. C. (2006). Fuels products of the LANDFIRE project. In and crown fire behavior. Res. Pap. RMRS-RP-29.Fort Collins, CO: U.S. Department of
P. L. Andrews, & B. W. Butler (Eds.), Fuels management—How to measure success: Agriculture, Forest Service, Rocky Mountain Research Station 59 pp.
conference proceedings. RMRS-P-41 (pp. 239−252). Portland, OR: U.S. Department Seielstad, C. A., & Queen, L. P. (2003). Using airborne laser altimetry to determine fuel
of Agriculture, Forest Service, Rocky Mountain Research Station. models for estimating fire behavior. Journal of Forestry, 101, 10−15.
Reinhardt, E., Keane, R.E., & Brown, J.K. (1997). First order fire effects model: FOFEM 4.0, Sprugel, D. G. (1983). Correcting for bias in log-transformed allometric equations.
user's guide. General Technical Report INT-GTR-344.: 72 pp. Ecology, 64, 209−210.
Reinhardt, E., Lutes, D., & Scott, J. (2006). FuelCalc: A method for estimating fuel Sullivan, A., McGaughey, R., Andersen, H. -E., & Schiess, P. (2009). Object-oriented
characteristics. In P. L. Andrews, & B. W. Butler (Eds.), Fuels management—How to classification of forest structure from light detection and ranging data for stand
measure success: Conference proceedings. March 28–30, 2006 (pp. 273−282). Portland, mapping. Western Journal of Applied Forestry, 24(4), 198−204.
OR: U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station. Van Wagtendonk, J. W., & Root, R. R. (2003). The use of multi-temporal Landsat Normalized
Reinhardt, E., Scott, J., Gray, K., & Keane, R. (2006). Estimating canopy fuel Difference Vegetation Index (NDVI) data for mapping fuel models in Yosemite
characteristics in five conifer stands in the western United States using tree and National Park, USA. International Journal of Remote Sensing, 24, 1639−1651.
stand measurements. Canadian Journal of Forest Research, 36, 2803−2814. Varga, T. A., & Asner, G. P. (2008). Hyperspectral and LiDAR remote sensing of fire fuels
Renslow, M., Greenfield, P., & Guay, T. (2000). Evaluation of multi-return LIDAR for in Hawaii Volcanoes National Park. Ecological Applications, 18, 613−623.
forestry applications. RSAC-2060/4810-LSP-0001-RPT1.: US Department of Agricul- Vierling, L. A., Rowell, E., Xuexia, C., Dykstra, D., & Vierling, K. (2002). Relationships
ture Forest Service — Remote Sensing Applications Center 19 pp. among airborne scanning LiDAR, high resolution multispectral imagery, and
Riano, D., Chuvieco, E., Condes, S., Gonzalez-Matesanz, J., & Ustin, S. L. (2004). ground-based inventory data in a ponderosa pine forest. Geoscience and Remote
Generation of crown bulk density for Pinus sylvestris L. from LiDAR. Remote Sensing Sensing Symposium, 2002. IGARSS '02. 2002 IEEE International (pp. 2912−2914).
of Environment, 92, 345−352. Wulder, M. A., LeDrew, E. F., Franklin, S. E., & Lavigne, M. B. (1998). Aerial image texture
Riano, D., Meier, E., Allgower, B., Chuvieco, E., & Ustin, S. L. (2003). Modeling airborne information in the estimation of northern deciduous and mixed wood forest leaf
laser scanning data for the spatial generation of critical forest parameters in fire area index (LAI). Remote Sensing of Environment, 64, 64−76.
behavior modeling. Remote Sensing of Environment, 86, 177−186. Xiong, J., Wu, B., Yan, N., Zhang, L., & Zhang, B. (2006). The influence of atmospheric
Richardson, J. J., Moskal, L. M., & Kim, S. -H. (2009). Modeling approaches to estimate correction on remote sensed parameters in multi-temporal evapotranspiration
effective leaf area index from aerial discrete-return LIDAR. Agricultural and Forest monitoring. 27th Asian conference on remote sensing Ulanbaatar, Mongolia.
Meteorology, 149, 1152−1160. Zheng, D., Rademacher, J., Chen, J., Crow, T., Bresee, M., Le Moine, J., et al. (2004). Estimating
Ritchie, M. W., & Hann, D. W. (1989). Equations for predicting the 5-year height growth of aboveground biomass using Landsat 7 ETM+ data across a managed landscape in
six conifer species in southwest Oregon.Corvallis, OR: Forest Research Laboratory, northern Wisconsin, USA. Remote Sensing of Environment, 93, 402−411.
Oregon State University 14 pp.
Rollins, M. G., & Frame, C. K. (2006). The LANDFIRE prototype project: nationally
consistent and locally relevant geospatial data for wildland fire management. Gen.