Content uploaded by Kunwar K. Singh
Author content
All content in this area was uploaded by Kunwar K. Singh on Jan 23, 2020
Content may be subject to copyright.
71
4LiDAR and Spectral Data
Integration for Coastal
Wetland Assessment
Kunwar K. Singh, Lindsey Smart, and Gang Chen
4.1 INTRODUCTION
Coastal environments are unique, with highly diverse natural communities that
provide several benets to both humans and wildlife. These natural communities
are increasingly threatened by human population growth, amenity migration,
unsustainable land-use practices, and the increasing demand for natural resources.
Additional factors, such as direct inundation from sea-level rise and saltwater
intrusion, compound the threats to coastal environments (Gornitz 1995; Hackney and
Yelverton 1990; Pearsall and Poulter 2005; Poulter etal. 2009). Inundation from sea-
level rise is anticipated to have a signicant impact on coastal environments (Hauer
etal. 2016; Karegar etal. 2016; Moorhead and Brinson 1995). Saltwater intrusion,
another concern of sea-level rise, is the movement of saltwater into traditionally
freshwater areas through both natural and articial surface water conduits, such as
streams, rivers, ditches, or canals (Ardon etal. 2013). This saltwater intrusion has
resulted in a phenomenon referred to as marsh migration, which is the landward
movement of salt marsh into areas that were once primarily fresh forested wetlands,
leaving behind what many have come to refer to as “ghost forests” (areas of standing
CONTENTS
4.1 Introduction .................................................................................................... 71
4.2 Methods .......................................................................................................... 73
4.2.1 Study Area .......................................................................................... 73
4.2.2 Data Sets Used .................................................................................... 73
4.2.3 LiDAR Data Processing ..................................................................... 74
4.2.4 Spectral Data Processing .................................................................... 76
4.2.5 Data Integration and Analysis ............................................................77
4.3 Results .............................................................................................................79
4.3.1 Exploratory Data Analysis ..................................................................79
4.3.2 Random Forest Classication ............................................................. 80
4.4 Discussion ....................................................................................................... 81
4.5 Conclusion ......................................................................................................84
4.5.1 Opportunities and Challenges ............................................................ 85
References ................................................................................................................ 85
72 High Spatial Resolution Remote Sensing
dead trees within newly formed salt marshes) (Pearsall and Poulter 2005). This
transition from forest to marsh is highly variable both spatially and temporally, and
is dependent upon local factors such as vertical accretion, subsidence rates, and other
anthropogenic stressors. There is signicant uncertainty regarding the implications
of these impacts on the continued persistence of coastal wetland ecosystems and
their ability to continue to provide the benets upon which human communities
rely. The spatiotemporal variability of these processes makes it difcult to develop
sound land-use management practices and policies (Enwright etal. 2016). This
requires an adaptive land-use policy and management decisions, which depend on
accurate methodologies to quantify ecosystems (Hudak etal. 2012; Singh etal. 2016).
Consequently, quantifying ne-scale spatial patterns of coastal wetlands, particularly
those wetlands at risk from saltwater intrusion, is crucial for understanding the
continued persistence of coastal ecosystem services.
Integration of data from multiple remote sensing platforms is a potential solution
to quantifying coastal wetland. Different platforms can complement one another
and ll data gaps; for example, fusion of LiDAR and spectral imagery (Chen etal.
2004; Donoghue and Watt 2006; Hudak etal. 2002; Li etal. 2011; Singh etal.
2012). Spectral data capture the 2D spectral characteristics of the Earth’s surface,
whereas LiDAR measures the 3D arrangement of the Earth’s surface in point cloud
or continuous wave format; therefore, complementing the limitations of spectral
data with several advantages. First, LiDAR is a source of high-resolution and
highly accurate point cloud data. With a small footprint system, LiDAR has the
capacity to reach a level of accuracy as high as that of conventional eld surveys.
Second, the intensity component of LiDAR data is useful to either pan sharpen
spectral data or provide additional information for the segmentation of features in
the scene (Campos-Taberner etal. 2016; Singh etal. 2010). Third, environment and
atmosphere have little impact on LiDAR data, unlike spectral imagery, allowing
LiDAR data collection under a wide range of environmental conditions (Jensen
2007). Despite its ability to capture vertical structural components along with the
above advantages, LiDAR alone is not sufcient for estimating the overall condition
of ecosystems.
This makes data integration essential for utilizing the complementary
characteristics of both data sources for extraction and classication of Earth objects
(Zhang and Lin 2017). Data integration has been explored in natural systems for
mapping forest structural attributes, plant species distribution, and forest fuel
estimation (Alonzo etal. 2014; Chust etal. 2008; Elaksher 2008; Garcia etal. 2011;
Swatantran etal. 2011). In urban systems, spectral and LiDAR data fusion has been
applied to land-cover mapping, 3D modeling, and building footprint extraction
(Awrangjeb etal. 2010; Singh etal. 2012; Sohn and Dowman 2007). Multiple factors,
such as extent, targeted application, spatial and spectral resolution, temporal match,
and fusion methods, play a decisive role in the integration of spectral and LiDAR
data (Zhang and Lin 2017). For example, an evaluation of trade-offs between data
volume and thematic map accuracy in Singh etal. (2012)’s study suggests that a spatial
resolution of 5 m for LiDAR surface models offers a balance between computational
efciency and classication accuracy for mapping land cover over large and spatially
heterogeneousregions.
73LiDAR and Spectral Data Integration for Coastal Wetland Assessment
Numerous methods have been proposed for integrating these data sources.
Applications tend to be data driven, therefore dictating the data characteristics,
processing, and integration methods. For example, estimation of vegetation
parameters such as tree species, height, and forest types relies on spatial and spectral
resolution of spectral data and LiDAR point spacing. Since eld data are collected at
the plot level, multivariate and machine learning analyses are often applied to model
vegetation parameters. Zhang and Xie (2012) used a neural network approach with
tree-level survey data for detecting individual trees, estimating tree metrics, and
identifying their species types using LiDAR with hyperspectral imagery. Pixel-based
and object-based methods are often used to map land cover. Haala and Brenner (1999)
applied pixel-based techniques to multispectral and LiDAR-derived digital surface
models to identify buildings and trees. Gamba and Houshmand (2002) analyzed
synthetic aperture radar (SAR), LiDAR, and aerial imagery for extraction of land
cover. Likewise, Chen etal. (2009) employed an object-oriented classication using
multispectral pan sharpened QuickBird imagery and LiDAR data over an urban area.
We integrate LiDAR and spectral data to map coastal wetland land-cover types. The
main research objectives of this study were to (1) quantify the important vertical and
horizontal structural metrics for different wetland compositional types, (2) assess the
ability to differentiate wetland compositional types based on LiDAR and spectral remote
sensing metrics using the random forest algorithm, and (3) evaluate modelperformance.
4.2 METHODS
4.2.1 study AreA
Goose Creek State Park is located in Beaufort County, North Carolina (Fig ure 4.1). It
covers 6.77 km2 just off the Pamlico Sound in North Carolina’s Outer Coastal Plain.
The park contains extensive marshes, inlets, and creeks on the northern side of the
sound. It has been the site of timber production, commercial shing, and small-scale
subsistence farming. Lumber companies acquired extensive tracts of land along the
creeks and harvested vast stands of old growth bald cypress and longleaf pine. Much
of the land that is now part of the park was once clear-cut for timber. Evidence of the
timber industry, such as remnants of piers and loading docks up and down the creek,
remains at the park today. There are marshes at the shoreline that transition to riverine
and depressional swamp forests further from the shoreline at higher elevations.
4.2.2 dAtA sets used
Ecological training and validation data were obtained from North Carolina Division of
Coastal Management’s (NC DCM’s) rened version of the National Wetlands Inventory
(NWI) data set. Wetlands were developed with a minimum polygonal mapping unit of
1 acre, using NWI data, a hydrography line data set, county soils surveys, and 30 m
satellite imagery (North Carolina Department of Environmental Quality, http://deq.
nc.gov/about/divisions/coastal-management). In addition to the rened NWI data set,
we also used LiDAR and aerial color infrared (CIR) imagery collected by the National
Agriculture Imagery Program (NAIP) for the analysis (Table 4.1).
74 High Spatial Resolution Remote Sensing
4.2.3 lidAr dAtA Processing
Integrating LiDAR and spectral data requires preprocessing individual data sets to
fully utilize their complimentary information. LiDAR data preprocessing includes
(1) removal of artifacts (noise points), (2) classication of LiDAR points into ground
and nonground, and (3) data quality check. Removal of artifacts is recommended to
identify and eliminate noise points that are caused by transmission lines, atmospheric
aerosols, birds, or low-ying aircraft. These points are usually outside the range of
realistic elevations.
We used airborne mapping LiDAR that was acquired in LAS le format with
points classied as vegetation and ground (Table 4.2) to develop the digital elevation
Goose creek state park
0.500.5 KM0.25
FIGURE 4.1 Study area of Goose Creek State Park with an inset image of a typical salt-
affected forested wetland.
TABLE 4.1
Data Sets Used in the Analysis
Data Set Date Source Derived Metrics
LiDAR January to
February 2014
North Carolina Floodplain
Mapping Program
Vegetation and terrain
metrics
Color infrared
imagery
October 2014 NAIP, USDA Farm
Service Agency
Normalized difference
vegetation index
NWI data 1999–2010 North Carolina Division
of Coastal Management
Ecological training
and validation data
75LiDAR and Spectral Data Integration for Coastal Wetland Assessment
model (DEM) and normalized heights and metrics. LiDAR data processing was
completed using GRASS 7.0.1 (Neteler etal. 2012). The ground point locations were
imported, and the multiscale curvature-based classication algorithm (v.lidar.mcc)
was run over the ground points to remove any spurious classications of vegetation
as “ground,” which can often happen in areas with dense wetland vegetation (Evans
and Hudak 2007). Next, we imported the ground points using v.surf.rst at 3 m spatial
resolution to create the DEM. Multiple classes with different vegetation types were
merged using v.in.lidar and v.patch, respectively, to create a merged canopy point
layer. We exported the merged point layer to a text le using v.out.ascii. The merged
canopy le was brought back in using r.in.xyz, calculating each of the desired LiDAR
metrics at a 3 m resolution. The spatial resolution was determined by selecting the
nest resolution possible to detect ne-scale variability in vegetation, while also
maintaining the ability to characterize vegetation canopy. A resolution of 3 m was
deemed satisfactory because the number of vegetation returns for each 3 m cell
ranged from 0 to 76, with an overall average of 14 vegetation returns per cell. Then,
we subtracted the interpolated DEM from the height metric to normalize estimates
of height (Figure 4.2).
We derived metrics representing terrain characteristics from the LiDAR data.
These included ow accumulation, topographic wetness index (TWI), and distance
to shoreline (Table 4.3). Flow accumulation represents the accumulated weight of
all cells owing into each downslope cell in the output raster. The elevation model
is a prerequisite for ow accumulation and TWI. TWI is a function of the slope and
upslope contributing area and the ow direction. The TWI is dened as
ln
tan
a
b
where a is the local upslope area draining through a certain point per unit contour
length, and tan b is the local slope (in radians).
TABLE 4.2
LiDAR Data Specifications
Content Specification
Survey data January 2014
Program North Carolina Floodplain Mapping Program
Altitude above ground 1676 m
Multiple returns Four returns per pulse
Swath width 1025 m
RMSEza11.7 cm
Average post spacing 0.7 m
Average point density 2.0/m2
a Root mean square error in z.
76 High Spatial Resolution Remote Sensing
We generated the distance to shoreline in two parts. First, we digitized the shoreline
using the elevation data as reference to identify the shoreline contour. Then, a simple
Euclidean distance from this linear shoreline feature was derived as the nal metric.
LiDAR-derived vegetation metrics were calculated by binning the laser return points
at the selected cell size resolution according to the univariate statistics of interest,
and were used to characterize vegetation canopy structure. The ability to detect
soil moisture or potential inundation is an important abiotic factor since it controls
wetland extent and function. Some research has explored the applicability of LiDAR
intensity values to detect soil saturation. Therefore, we imported intensity values at
the 3 m resolution. An enhanced Lee lter was applied to the raw intensity values in
an effort to reduce speckle but retain edges or transition zones in the intensity data.
The enhanced Lee lter was passed over the intensity values multiple times, with
increasing window size each time 5, 7, and 9 number of cells (Huang etal. 2014).
4.2.4 sPectrAl dAtA Processing
Spectral data preprocessing routines are necessary to improve the spectral quality
of imagery for accurate outcomes and for maximizing information through data
integration. Routines include geometric and radiometric corrections, and the
transformation of the image to a specic map projection system. Errors in the relative
position of pixels in imagery caused by the imaging system and imaging conditions
are addressed through geometric corrections. Radiometric correction addresses
errors in the imagery that occur due to sensor sensitivity, sun angle and topography,
and atmospheric interference (Jensen 2007). The CIR images were provided at 1 m
resolution and were resampled to 3 m resolution using a bilinear cubic interpolation to
match the resolution of the LiDAR data. We performed orthorectication by selecting
invariant features within the study area, specically road centerlines and parking,
Airborne
LiDAR data
Preprocessing
DEM
topographic data
Optical RS
data
NDVI calculation
Terrain metrics
Input data combinations
Ecological
training data Random forest classification
Classification maps
Ecological
validation data
Accuracy
assessment
Vegetation structure metrics
FIGURE 4.2 Overview of the data preparation and classication process.
77LiDAR and Spectral Data Integration for Coastal Wetland Assessment
and then using these features as tie points in both the LiDAR data and the spectral
data to ensure proper alignment of the two data sources. We then used CIR images
to calculate the normalized difference vegetation index (NDVI). NDVI utilizes the
near-infrared spectral band and the red spectral band, important spectral bands for
identifying characteristics of vegetation greenness and health. Finally, we classied
different input layer combinations using the random forest algorithm.
4.2.5 dAtA integrAtion And AnAlysis
Classication algorithms that are frequently used to integrate these data include
random forests, gradient boosting, classication and regression trees, and object-
based image analysis, among others (Chen and Hay 2011; Jakubowski etal. 2013).
We used the random forest algorithm to map land-cover classes of coastal wetlands
(Table 4.4). The random forest algorithm is a decision tree classication and has been
applied successfully in ecological research and land-cover mapping (Breiman 2001).
It is especially suited for classication of multisource remote sensing data because
TABLE 4.3
Random Forest Models, Input Variables, and Total Accuracy
Sources Variables Random Forest Models
12345678Final
Distance to shoreline x x x x x
Spectral-data-
derived metrics
Normalized difference
vegetation index (NDVI)
x xxxxx
LiDAR-derived
vegetation metrics
Minimum vegetation height x x x x x
Total vegetation returns x x x x x x
Coefcient of variation of
vegetation height
x x x x x x
Maximum vegetation height x x x x x
Mean vegetation height x x x x x x
Standard deviation of
vegetation heights
x x x x x x
Range of vegetation heights x x x x x x
Variance of vegetation
heights
x x x x x x
Skewness of vegetation
heights
x x x x x
25th percentile heights x x x x x x
50th percentile heights x x x x x x
75th percentile heights x x x x x x
Mean intensity x x x x x x
LiDAR-derived
terrain metrics
Elevation x x x x x x
Flow accumulation x x x x x x
Topographic wetness index x x x x x
Accuracy (%) 42.1 51.3 50.7 61.2 55.4 53.3 56.4 61.8 61.9
78 High Spatial Resolution Remote Sensing
it is insensitive to nonnormal and noisy data. It also ts a predetermined number of
classication trees, and then combines the predictions across all trees. The majority
rule across all decision trees determines the accuracy of output classication. The
number of input variables is user dened. We performed data manipulation and
analysis in R statistical software (R Core Team 2013). For the classication, we
used the randomForest R statistical software package (Lawrence etal. 2006). The
National Wetlands Inventory data set was used to identify and select training samples.
A bootstrapped sample of the original training data was used to train the model.
Test set accuracy was determined using cross-validation of the remaining training
samples (out-of-the-bag samples). Variable importance was estimated by randomly
permuting the value of out-of-the-bag samples for variables. The error increases as
the particular variable is removed from the model, determining the importance of that
specic variable. Samples that were not used in the training were used as test samples
to perform accuracy assessment.
Preliminary exploratory data analysis was performed to initially identify
important variables and avoid overtting the model as well as to reduce collinearity
among predictor variables. One such exploratory test was the Kolmogorov-Smirnov
(KS) test, a nonparametric test of the equality of 1D probability distributions
TABLE 4.4
Target Land-Cover Classes and Descriptions
Land-Cover Types Descriptions
Bottomland hardwood forest Riverine forested or occasionally scrub/shrub communities
usually occurring in oodplains that are semipermanently to
seasonally ooded. In bottomland hardwoods, typical species
include oaks, sweet gum, river birch, and occasionally pines.
Depressional swamp forest Very poorly drained nonriverine forested or occasionally scrub/
shrub communities that are semipermanently or
temporarilyooded.
Estuarine forest A forested wetland community subject to occasional ooding by
tides, including wind tides.
Estuarine scrub/shrub Any scrub/shrub community subject to occasional ooding by
tides including wind tides.
Headwater swamp Forested systems along the upper reaches of rst-order streams.
Riverine swamp forest Riverine forested communities usually occurring in oodplains
that are semipermanently to seasonally ooded. In swamp
forest systems, typical species include cypress, black gum,
water tupelo, and red maple.
Salt-affected swamp forest Riverine forested communities as described above, but
inuenced by saltwater intrusion—either acutely or via
long-term inundation.
Salt/brackish marsh Any salt marsh or other marsh subject to regular of occasional
ooding by tides, including wind tides as long as this ooding
does not include hurricane or tropical storm waters.
Upland Nonwetland areas.
79LiDAR and Spectral Data Integration for Coastal Wetland Assessment
(Smirnov 1948). This test can be used to compare a sample with a reference
probability distribution or to compare two samples. The KS statistic quanties
a distance between the empirical distribution function of the sample and the
cumulative distribution function of the reference distribution, or between the
empirical distribution functions of two samples. The magnitude of the KS statistic
helped in identifying a subset of variables from all the input variables. Cumulative
distribution functions were drawn for these variables for each of the wetland types
being modeled to visualize differences across types.
4.3 RESULTS
4.3.1 exPlorAtory dAtA AnAlysis
Exploratory data analysis and KS tests showed that there were several spectral and
structural attributes that could be used to distinguish between wetland compositional
types. The KS statistics suggested that all of the potential predictor variables were
signicantly different between wetland types. The cumulative distribution functions
for each potential predictor variable and compositional type were evaluated for
discernable relationships between variables and type (Figure 4.3). For example, the
cumulative distribution functions for the distance to shoreline variable differ across
Proportion <= xProportion <= x
Proportion <=
xP
roportion <= x
10.0 30.0 0.2 0.6 1.0
20.0
Bottomland
Hardwood
Depressional
Swamp forest
Estuarine
Forest
Headwater
Swamp
Riverine swamp
forest
Salt affected
Swamp forest
Salt/brackish
Marsh
Up
land
Estuarine
Scrub/shrub
Mean vegetation height (m) Normalized difference vegetation index
1.0
0.0 0.0
1.0 1.0
1.02.0 3.0
Distance to shoreline (km) Elevation (m)
3.0
FIGURE 4.3 Cumulative distribution functions for wetland types and variables.
80 High Spatial Resolution Remote Sensing
wetland types (Figure 4.3). The salt or brackish marsh, which is associated with areas
very close to shore, is signicantly different from the upland or bottomland hardwood
forests, which tend to be located at a larger distance from the shore. Differences
such as these are also visible in the NDVI, minimum heights, and maximum heights
predictor variables.
4.3.2 rAndom Forest clAssiFicAtion
We observed signicant variability in mapping error rates across wetland types. While
estuarine scrub/shrub and riverine swamp forest classes had the lowest error (i.e., less
than 30% error), salt-affected riverine swamp forest and estuarine forest showed very
high error rates (∼80% error and over) (Table 4.5). The most important variable for
the classication was distance from the shoreline, followed by the terrain, LiDAR,
and spectral variables. Of the LiDAR variables, the total number of vegetation returns
(as a measure of vegetation density), the range in vegetation heights, and the mean
vegetation heights contributed the most (Figure 4.4).
We observed the highest model performance when LiDAR, multispectral remote
sensing, and terrain metrics were combined (nal model), with an increase of 6%
accuracy over the spectral and terrain model (i.e., Model 5) (Table 4.3). The LiDAR
model alone (Model 2) performed the best of the three single data source models
(Model 1, spectral imagery alone; Model 3, terrain alone). One hundred permutations
of the nal random forest tted model were run to analyze the permutation importance
of all predictor variables in the model (Figure 4.5). The contribution of predictor
variables toward the model overall were analyzed by evaluating the mean decrease
in accuracy with the removal of each variable. The predictor variable importance
was also evaluated for each land-cover classication. Permutation results suggested
that distance to shoreline, elevation, NDVI, and total vegetation returns were
consistently the most important to overall accuracy in all permutations. This was
statistically signicant as represented in Figure 4.5. Comparison of the importance
of the predictor variables for the highest and the lowest accuracy class reveals that
TABLE 4.5
Land-Cover Classifications, Associated Areas, and Class Errors
Vegetation Type Area (acres [km2]) Class Error
Bottomland hardwood forest 38.1 (0.15) 0.66
Depressional swamp forest 2.1 (0.01) 0.52
Estuarine forest 1.19 (0.01) 0.97
Estuarine scrub/shrub 162.8 (0.66) 0.28
Headwater swamp 48.24 (0.20) 0.74
Riverine swamp forest 534.93 (2.16) 0.21
Salt-affected swamp forest 1.66 (0.01) 0.76
Salt/brackish marsh 21.85 (0.088) 0.33
Upland 527.38 (2.13) Not available
81LiDAR and Spectral Data Integration for Coastal Wetland Assessment
the lowest accuracy class relied on distance to shoreline and elevation along with
mean intensity for prediction accuracy. This differed signicantly from the highest
accuracy class, whose accuracy was determined most consistently by LiDAR-derived
structural metrics. Figure 4.6 shows a map of the best model.
4.4 DISCUSSION
According to the random forest outputs, the forested wetland types and salt-affected
swamp forests were the most difcult to classify. Variability in species composition
and spatial distribution may have caused difculties in the classication of wetland
types. Such difculty in mapping forested wetlands has been cited in other analyses
in similar coastal areas within North Carolina (Allen et al. 2013). When the
classication was run on the six broader NWI wetland classications present in
the study area (e.g., estuarine shrub/scrub, palustrine forest wetland), accuracy
of the most parsimonious model reached 82%. In the random forest classication
of the NC DCM wetlands, the distance to shoreline and elevation variables were
most important. This is understandable, since distance from the shoreline as well
as elevation patterns are both important drivers of wetland extent and function.
Although the LiDAR data were collected in January, during leaf-off conditions, to
optimize the mapping of ground elevation, highly useful information on canopy
structure can still be extracted from the nonground returns. Several studies have
shown that leaf-off LiDAR data can be used to characterize vegetation canopy and
structure (Sexton etal. 2009; Smart etal. 2012). Although the exact height of the
canopy may be inuenced by the season (due to leaf-on and leaf-off differences),
the overall structure and variability remains identiable even in leaf-off LiDAR
data. Therefore, we concluded that the importance of the elevation and distance to
shoreline variables were due to the inherently dynamic nature of wetlands that can
appear both spectrally and structurally similar but may differ hydrologically and
in their classication due to slight changes in elevation, inundation patterns, and
distance to water.
0.050 0.1 0.15 0.2 0.25
Distance to shoreline
Elevation
Total vegetation returns
Range vegetation heights
Mean vegetation height
Std. de
v. vegetation heights
NDVI
75th percentile heights
50th percentile heights
Variance vegetation height
Mean intensity
Coeff. variation heights
25th percentile heights
FIGURE 4.4 Variable importance for predictor variables used in the random forest
classications.
82 High Spatial Resolution Remote Sensing
Mean decrease in accuracy
Distance to shoreline
Distance to shoreline
Elevation
Elevation
Total vegetation returns
Total vegetation returns
NDVI
NDVI
Mean intensity
Mean intensity
Range vegetation heights
Range vegetation heights
Variance vegetation height
Variance vegetation height
Mean vegetation height
Mean vegetation height
Coeff. variation heights
Coeff. variation heights
Coeff. variation heights
Std. dev. vegetation heights
Std. dev. vegetation heights
25th percentile heights
25th percentile heights
50th percentile heights
50th percentile heights
75th percentile heights
75th percentile heights
Distance to shoreline
Elevation
Total vegetation returns
NDVI
Mean intensity
Range vegetation heights
Variance vegetation height
Mean vegetation height
Std. de
v. vegetation heights
25th percentile heights
50th percentile heights
75th percentile heights
Coeff. variation heights
Distance to shoreline
Elevation
Total vegetation returns
NDVI
Mean intensity
Range vegetation heights
Variance vegetation height
Mean vegetation height
Std. dev. vegetation heights
25th percentile heights
50th percentile heights
75th percentile heights
0
0
–10 –5 20151050
25
200 400 600 800 1000 1200 1400
20 40 60 80 100 120 140 160 180
020406080 100 120
Mean decrease in gini
Estuarine forest
Riverine swamp forest
FIGURE 4.5 Permutation importance of predictor variables after 100 permutations of the
random forest tted model. Aside from the mean decrease in Gini and overall accuracy,
shown in the gure are predictor variable importance for one of the highest accuracy classes
(riverine swamp forest) and one of the lowest accuracy classes (estuarine forest). Black color
shows statistically signicant across all permutations, whereas gray signies nonsignicant
impacts on overall accuracy.
83LiDAR and Spectral Data Integration for Coastal Wetland Assessment
Bottomland hardwood
Depressional swamp forest
Estuarine forest
Random forest
Wetland classification
0
0.475 0.95 1.9 km
Estuarine scrub/shrub
Headwater swamp
Riverine swamp forest
Salt/brackish marsh
Upland
Salt-affected swamp forest
NWI-NC DCM
Wetland classification
N
(a
)
(b
)
FIGURE 4.6 Costal wetland maps. (a) NC DCM—NWI wetland classication. Polygons
from NC DCM wetland classication provided the target land cover for the random forest
classication. (b) A map from the top-performing random forest classication model. White
areas within the study area bounds represent areas of insufcient data for prediction. Note:
Salt-affected swamp forests only appear in the random forest classication because this is
not a land-cover class present in the NC DCM wetlands. Ancillary training data were made
available and were used to supplement the NC DCM wetlands so that the algorithm could be
used to explore the ability to identify salt-impacted wetlands.
84 High Spatial Resolution Remote Sensing
The random forest method was selected because it is a nonparametric technique
that can handle both continuous and categorical independent variables. Random forest
runs efciently on large databases, can handle many input variables without variable
deletion, and maintains high predictive power even when a signicant proportion of
the data is missing. It works well with large and noisy remote sensing data sets. It
has also been shown to reduce overtting and achieve higher accuracies compared
to traditional classication trees (Breiman 2001). The training data are the key to the
performance of the random forest method. In future analyses, the training data can
be improved to include acquisitions of more recent eld vegetation data. The data
points used to train the model in this study were drawn from polygonal data sets
with a relatively coarse minimum mapping unit. The inclusion of more accurate and
thorough eld-based analyses is considered a source of potential improvement, and
may improve overall accuracies if the model were to be trained on a different data
set. In addition, accuracy may be improved with an ensemble modeling approach.
Potentially, a combination of LiDAR and another active remote sensing technology,
SAR, could be included. This may improve the ability to differentiate the marshes that
appear spectrally similar but differ in elevation (i.e., high and low marsh). Results from
initial studies combining these active remote sensing technologies prove promising
(Allen 2014; Allen etal. 2013).
It was expected that the addition of LiDAR-derived vegetation metrics would
increase the accuracy of the classication signicantly, but results suggested it
increased the accuracy by 6%. LiDAR was able to distinguish between different
wetland types that had characteristic patterns of vertical structure (i.e., differentiating
between marsh, scrub/shrub, and forest), particularly in cases where the types might
have appeared spectrally similar. However, there are also cases wherein wetland
types (i.e., estuarine forest and riverine forest) appear both spectrally and structurally
similar, and the main drivers of species differentiation are related to terrain factors
(i.e., distance to shoreline and elevation). Despite the relatively small contribution of
LiDAR in this particular study, because of the complexity and diversity of coastal
wetland composition, a multidata source approach (one that combines terrain,
spectral, and vertical measures) appears to be the most appropriate.
4.5 CONCLUSION
The ability to classify and discern wetland types is necessary in order to quantify
changes in wetland ecosystems through time. The vegetation extent and habitat
classication maps show the utility of combining LiDAR and spectral imagery in
the mapping of coastal wetlands. Analyses of the classications indicate that elevation
(along with distance to shoreline) is by far the most important variable for mapping
wetland extent and function. Vegetation metrics also contributed to the overall
classication accuracy. Results of this research suggest that these remote sensing
variables hold promise for assessing wetland composition and differentiating between
wetland types. In coastal ecosystems, further analysis is needed to understand the
relationships between wetland composition and saltwater intrusion. Future research
will use multitemporal LiDAR data to detect changes in wetland ecosystems resulting
from saltwater intrusion using spectral data sources and eld data. Results from this
85LiDAR and Spectral Data Integration for Coastal Wetland Assessment
analysis would provide a better understanding of the complex interactions between
wetland ecosystems and sea level rise induced saltwater intrusion.
4.5.1 oPPortunities And chAllenges
In order to exploit all the potential offered by the new generations of LiDAR and
spectral sensors to map coastal wetlands, we must address several issues. The rst
issue is temporal disparity. It is a rare occurrence that all data used in an analysis
are collected at the same time. Temporal resolution of a satellite sensor may vary
from hours to days, whereas in the case of airborne sensors it may range from days
to years, depending on mission planning. Landscapes are not static—they show high
spatiotemporal variabil ity due to processes such as forest growth and urba nization. This
dynamic nature can lead to erroneous results from data registration and integration.
Identication of such disparities and then detrending them (e.g., removing an aspect
from the data that is causing some kind of distortion) will require research. The
second issue is data registration. Registration of LiDAR and spectral images data with
low possible errors is necessary to harness the full potential these data offer. Multiple
approaches are available for data registration; however, new generations of sensors
require the development of new registration algorithms to achieve accurate registration
(Zhang and Lin 2017). The third issue is the limited availability of standardized
data processing approaches. Despite promising prospects, the limited availability
of multisensor data and the lack of standardized data processing algorithms hamper
their widespread integration in analyses. For example, standardized processing and
analysis approaches for multitemporal LiDAR and return intensity would help spur
research (Eitel etal. 2016) and might improve data integration for Earth and ecological
sciences. The fourth issue is the increasing need for high-performance computing.
Fusion studies are typically limited to small spatial extents due to LiDAR’s high
cost, small footprint, and large data volume. While cost will decrease, we expect
increases in LiDAR data volumes with the expansion of the spectral and temporal
dimensions of LiDAR data (Eitel etal. 2016). The increasingly high dimension of
LiDAR and spectral data and the complexity of the processing algorithms require the
design of adequate algorithms and architectures. The nal issue is the development
of appropriate and accurate mapping algorithms. Data fusion offers great potential to
map and model environmental problems. This requires an improved understanding
of the physics of LiDAR and spectral data in order to design optimal classication
and mapping algorithms. With very high point spacing and resolutions, one needs an
in-depth understanding of the involved physics of the sensors for accurate calibration
and mapping of environmental conditions and processes.
REFERENCES
Allen, T.R. 2014. Advances in remote sensing of coastal wetlands: LiDAR, SAR, and object-
oriented case studies from North Carolina. In Remote Sensing and Modeling, 405–428.
Cham: Springer.
Allen, T.R., Wang, Y., & Gore, B. 2013. Coastal wetland mapping combining multi-date SAR
and LiDAR. Geocarto International, 28, 616–631.
86 High Spatial Resolution Remote Sensing
Alonzo, M., Bookhagen, B., & Roberts, D.A. 2014. Urban tree species mapping using
hyperspectral and LiDAR data fusion. Remote Sensing of Environment, 148, 70–83.
Ardon, M., Morse, J.L., Colman, B.P., & Bernhardt, E.S. 2013. Drought-induced saltwater
incursion leads to increased wetland nitrogen export. Global Change Biology, 19,
2976–2985
Awrangjeb, M., Ravanbakhsh, M., & Fraser, C.S. 2010. Automatic detection of residential
buildings using LiDAR data and multispectral imagery. ISPRS Journal of
Photogrammetr y and Remote Sensing, 65, 457–467.
Breiman, L. 2001. Random forests. Machine Learning, 45, 5–32.
Campos-Taberner, M., Romero-Soriano, A., Gatta, C., Camps-Valls, G., Lagrange, A., Le
Saux, B., Beaupere, A., Boulch, A., Chan-Hon-Tong, A., Herbin, S., Randrianarivo,
H., Ferecatu, M., Shimoni, M., Moser, G., & Tuia, D. 2016. Processing of extremely
high-resolution LiDAR and RGB data: Outcome of the 2015 IEEE GRSS data fusion
contest—Part A: 2-D contest. IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing, 9, 5547–5559.
Chen, G., & Hay, G.J. 2011. An airborne LiDAR sampling strategy to model forest canopy
height from Quickbird imagery and GEOBIA. Remote Sensing of Environment, 115,
1532–1542.
Chen, X.X., Vierling, L., Rowell, E., & DeFelice, T. 2004. Using LiDAR and effective LAI data
to evaluate IKONOS and Landsat 7 ETM+ vegetation cover estimates in a ponderosa
pine forest. Remote Sensing of Environment, 91, 14–26.
Chen, Y.H., Su, W., Li, J., & Sun, Z.P. 2009. Hierarchical object oriented classication using
very high resolution imagery and LiDAR data over urban areas. Advances in Space
Research, 43, 1101–1110.
Chust, G., Galparsoro, I., Borja, A., Franco, J., & Uriarte, A. 2008. Coastal and estuarine
habitat mapping, using LiDAR height and intensity and multi-spectral imagery.
Estuarine Coastal and Shelf Science, 78, 633–643.
Donoghue, D.N.M., & Watt, P.J. 2006. Using LiDAR to compare forest height estimates from
IKONOS and Landsat ETM+ data in Sitka spruce plantation forests. International
Journal of Remote Sensing, 27, 2161–2175.
Eitel, J.U.H., Höe, B., Vierling, L.A., Abellán, A., Asner, G.P., Deems, J.S., Glennie, C.L.,
Joerg, P.C., LeWinter, A.L., Magney, T.S., Mandlburger, G., Morton, D.C., Müller, J., &
Vierling, K.T. 2016. Beyond 3-D: The new spectrum of LiDAR applications for earth
and ecological sciences. Remote Sensing of Environment, 186, 372–392.
Elaksher, A.F. 2008. Fusion of hyperspectral images and LiDAR-based DEMs for coastal
mapping. Optics and Lasers in Engineering, 46, 493–498.
Enwright, N.M., Grifth, K.T., & Osland, M.J. 2016. Barriers to and opportunities for
landward migration of coastal wetlands with sea-level rise. Frontiers in Ecology and
the Environment, 14, 307–316.
Evans, J.S., & Hudak, A.T. 2007. A multiscale curvature algorithm for classifying discrete
return LiDAR in forested environments. IEEE Transactions on Geoscience and Remote
Sensing, 45, 1029–1038.
Gamba, P., & Houshmand, B. 2002. Joint analysis of SAR, LiDAR and aerial imagery for
simultaneous extraction of land cover, DTM and 3D shape of buildings. International
Journal of Remote Sensing, 23, 4439–4450.
Garcia, M., Riano, D., Chuvieco, E., Salas, J., & Danson, F.M. 2011. Multispectral and LiDAR
data fusion for fuel type mapping using support vector machine and decision rules.
Remote Sensing of Environment, 115, 1369–1379.
Gornitz, V. 1995. Sea-level rise—A review of recent past and near-future trends. Earth Surface
Processes and Landforms, 20, 7–20.
Haala, N., & Brenner, C. 1999. Extraction of buildings and trees in urban environments. ISPRS
Journal of Photogrammetr y and Remote Sensing, 54, 130–137.
87LiDAR and Spectral Data Integration for Coastal Wetland Assessment
Hackney, C.T., & Yelverton, G.F. 1990. Effects of human activities and sea level rise on
wetland ecosystems in the Cape Fear River Estuary, North Carolina, USA. In Wetland
Ecology and Management: Case Studies, 55–61. Dordrecht: Springer.
Hauer, M.E., Evans, J.M., & Mishra, D.R. 2016. Millions projected to be at risk from sea-level
rise in the continental United States. Nature Climate Change, 6, 691–695.
Huang, C., Peng, Y., Lang, M., Yeo, I.Y., & McCarty, G. 2014. Wetland inundation mapping
and change monitoring using Landsat and airborne LiDAR data. Remote Sensing of
Environment, 141, 231–242.
Hudak, A.T., Lefsky, M.A., Cohen, W.B., & Berterretche, M. 2002. Integration of LiDAR
and Landsat ETM plus data for estimating and mapping forest canopy height. Remote
Sensing of Environment, 82, 397–416.
Hudak, A.T., Strand, E.K., Vierling, L.A., Byrne, J.C., Eitel, J.U.H., Martinuzzi, S., &
Falkowski, M.J. 2012. Quantifying aboveground forest carbon pools and uxes from
repeat LiDAR surveys. Remote Sensing of Environment, 123, 25–40.
Jakubowski, M.K., Li, W.K., Guo, Q.H., & Kelly, M. 2013. Delineating individual trees from
LiDAR data: A comparison of vector- and raster-based segmentation approaches.
Remote Sensing, 5, 4163–4186.
Jensen, J.R. 2007. Remote sensing of the environment: An Earth resource perspective. (2nd
ed.). Upper Saddle River, NJ: Pearson Prentice Hall.
Karegar, M.A., Dixon, T.H., & Engelhart, S.E. 2016. Subsidence along the Atlantic Coast of
North Amer ica: Insights from GPS and late Holocene relative sea level data. Geophysical
Research Letters, 43, 3126 –3133.
Lawrence, R.L., Wood, S.D., & Sheley, R.L. 2006. Mapping invasive plants using hyperspectral
imagery and Breiman Cutler classications (randomForest). Remote Sensing of
Environment, 100, 356–362.
Li, A.M., Huang, C.Q., Sun, G.Q., Shi, H., Toney, C., Zhu, Z.L., Rollins, M.G., Goward,
S.N., & Masek, J.G. 2011. Modeling the height of young forests regenerating from
recent disturbances in Mississippi using Landsat and ICESat data. Remote Sensing of
Environment, 115, 1837–1849.
Moorhead, K.K., & Brinson, M.M. 1995. Response of wetlands to rising sea level in the lower
coastal plain of North Carolina. Ecological Applications, 5, 261–271.
Neteler, M., Bowman, M.H., Landa, M., & Metz, M. 2012. GRASS GIS: A multi-purpose open
source GIS. Environmental Modelling & Software, 31, 124–130.
Pearsall, S., & Poulter, B. 2005. Adapting coastal lowlands to rising seas. Principles of
Conservation Biology, 3, 366–370.
Poulter, B., Feldman, R.L., Brinson, M.M., Horton, B.P., Orbach, M.K., Pearsall, S.H., Reyes, E.,
Riggs, S.R., & Whitehead, J.C. 2009. Sea-level rise research and dialogue in North Ca rolina:
Creating windows for policy change. Ocean & Coastal Management, 52, 147–153.
R Core Team. 2013. R: A language and environment for statistical computing. Vienna,
Austria: R Foundation for Statistical Computing.
Sexton, J.O., Bax, T., Siqueira, P., Swenson, J.J., & Hensley, S. 2009. A comparison of
lidar, radar, and eld measurements of canopy height in pine and hardwood forests of
southeastern North America. Forest Ecology and Management, 257, 113 6 –1147.
Singh, K.K., Chen, G., Vogler, J.B., & Meentemeyer, R.K. 2016. When big data are too
much: Effects of LiDAR returns and point density on estimation of forest biomass.
IEEE Journal of Selected Topics in Applied Earth Obser vations and Remote Sensing,
9, 3210 –3218.
Singh, K.K., Vogler, J.B., Meng, Q., & Meentemeyer, R.K. 2010. Mapping Land Use Patterns
in an Urbanizing Landscape Using LiDAR Intensity Data. In Optical Remote Sensing
of the Environment, OMC2. Tucson, AZ: Optical Society of America.
Singh, K.K., Vogler, J.B., Shoemaker, D.A., & Meentemeyer, R.K. 2012. LiDAR-Landsat
data fusion for large-area assessment of urban land cover: Balancing spatial resolution,
88 High Spatial Resolution Remote Sensing
data volume and mapping accuracy. ISPRS Journal of Photogrammetry and Remote
Sensing, 74, 110–121.
Smart, L.S., Swenson, J.J., Christensen, N.L., & Sexton, J.O. 2012. Three-dimensional
characterization of pine forest type and red-cockaded woodpecker habitat by small-
footprint, discrete-return lidar. Forest Ecology and Management, 281, 100 –110.
Smirnov, N. 1948. Table for estimating the goodness of t of empirical distributions. Annals
of Mathematical Statistics, 19, 279–279.
Sohn, G., & Dowman, I. 2007. Data fusion of high-resolution satellite imagery and LiDAR
data for automatic building extraction. ISPRS Journal of Photogrammetry and Remote
Sensing, 62, 43–63.
Swatantran, A., Dubayah, R., Roberts, D., Hofton, M., & Blair, J.B. 2011. Mapping biomass
and stress in the Sierra Nevada using LiDAR and hyperspectral data fusion. Remote
Sensing of Environment, 115, 2917–2930.
Zhang, C.Y., & Xie, Z.X. 2012. Combining object-based texture measures with a neural
network for vegetation mapping in the Everglades from hyperspectral imagery. Remote
Sensing of Environment, 124, 310–320.
Zhang, J., & Lin, X. 2017. Advances in fusion of optical imagery and LiDAR point cloud
applied to photogrammetry and remote sensing. International Journal of Image and
Data Fusion, 8, 1–31.