5 “(ms LIBRARY
;. C Michigan State
University

 

 

 

This is to certify that the
dissertation entitled

ASSESSING INVASIVE PLANT INFESTI'AION IN
FRESHWATER WETLANDS

presented by

Nathan M. Torbick

has been accepted towards fulfillment
of the requirements for the

Ph.D degree in Geography

 

 

 

V

Amior Professor’s Signature
10/28/07

 

Date

MSU is an afflnnative-action, equal-opportunity employer

oo---u-¢-o-o-a-o-.---.-,_ _ _ _ _ _

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DAIEDUE

DAIEDUE

DAIEDUE

 

072414

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6/07 p:lClRC/DateDue.indd-p.1

 

ASSESSING INVASIVE PLANT INFESTATION IN FRESHWATER WETLANDS.
By

Nathan M. Torbick

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY

Department of Geography

2007

ABSTRACT
ASSESSING INVASIVE PLANT INFESTATION IN FRESHWATER WETLANDS.
By

Nathan M. Torbick

Recent shifts in wetland ecosystem management goals have directed efforts
toward measuring ecological integn'ty, rather than only using physical and chemical
measures of ecosystems as health indicators. Invasive species pose one of the largest
threats to wetlands integrity. Resource managers can benefit from improved methods for
identifying invasive plant species, assessing infestation, and monitoring control measures.
The utilization of advanced remote sensing tools for species-level mapping has been
increasing and techniques need to be explored for identifying species of interest and
characterizing infestation.

The overarching goal of this research was to develop monitoring technologies to
map invasive plants and quantify wetland infestation. The first field-level objective was
to characterize absorption and reflectance features and assess processing techniques for
separating wetland species. The second field-level objective was to evaluate the abilities
of a shape filter to identify wetland invasive plant species. The first landscape-level
objective was to classify hyperspectral imagery in order to identify invasives of interest.
The second landscape-level objective was to quantify infestation within the study area.

Field-level hyperspectral data (350-2500nm) were collected for twenty-two
wetland plant species in a wetland located in the lower Muskegon River watershed in

Michigan, USA. The J effries-Matusita distance measure, continuum removal, and a

shape-filter were applied to hyperspectral species reflectance data to characterize spectral
features. Generally, continuum removal decreased separation distance for the invasive
species of interest. Using the shape-filter, Lythrum salicaria, Phragmites australis, and
Typha latifolia possessed maximum separation (distinguished from other species) at the
near-infrared edge (700nm) and water absorption region (1350nm), the near-infrared
down slope (1000 and 1100nm), and the visible/chlorophyll absorption region (500nm)
and near-infrared edge (650nm), respectively.

Airborne hyperspectral imagery was classified using a two-step approach in order
to obtain an optimal map (overall accuracy ~ 70%). Information in the near-infiared
enabled relatively accurate classification for Phragmites australis using the Spectral
Angle Mapper algorithm and image-derived training, while Typha latzfolia signatures
possessed high spectral overlap and required ISODATA clustering techniques. Landscape
pattern metrics relate infestation to disturbances and hydrological controls. The highest
levels of infestation and infestation patterns coincide with the most substantial levels of
hydrological modifications indicating human disturbances are correlated with Typha and
Phragmites percentages in the landscape.

Overall the approach was successful and increased the level of information
ultimately desired by decision makers. The rapidly advancing field of wetland remote
sensing science can obtain more meaningful information from hyperspectral imagery;
however, the data are challenging to work with and only the most precisely calibrated
datasets will provide utility. Combining these data with traditional wetland assessment
techniques can substantially advanced goals of preserving and restoring wetland

ecosystem integrity.

ACKNOWLEDGMENTS
Thanks to my committee, family, and friends for all their support. Additional thanks to

the organizations that funded this effort.

IV

TABLE OF CONTENTS

LIST OF TABLES ................................................................................... VI
LIST OF FIGURES ................................................................................. VII
CHAPTER 1: INTRODUCTION ................................................................... 1
1.1 Wetland Remote Sensing ............................................................... 3

1.2 Problem Statement ...................................................................... 6

1.3 Research Objectives ..................................................................... 7
GHAPTER 2: HYPERSPECTRAL REMOTE SENSING ..................................... 8
CHAPTER 3: METHODS ......................................................................... 12
3.1 Study Area ............................................................................... 12

3.2 Data ...................................................................................... 13

3.2.1 Field-level ................................................................... 13

3.2.3 Landscape-level ............................................................ 18

3.3 Quantifying Distance ................................................................... 19

3.4 Continuum Removal ................................................................... 20

3.5 Shape Filter ............................................................................. 22

3.6 Image Classification ................................................................... 22

3.7 Landscape Pattern Metrics ............................................................ 24
CHAPTER 4: RESULTS AND DISCUSSION ................................................. 29
4.1 Spectral Separation .................................................................... 29

4.2 Continuum Removal .................................................................. 34

4.3 Absorption Feature Characteristics .................................................. 38

4.4 Shape Filter ............................................................................. 40

4.5 Characterizing Infestation ............................................................ 45
CHAPTER 5: CONCLUSIONS .................................................................. 55
REFERENCES ...................................................................................... 59

LIST OF TABLES

Table 1. Dominant wetland plant species in study site ......................................... 16
Table 2. Airborne imagery configuration ......................................................... 20
Table 3. JM distance values ........................................................................ 30
Table 4. Change in JM distance values ........................................................... 36
Table 5. Accuracy assessment ..................................................................... 47
Table 6. Landscape pattern metric values ....................................................... 51

VI

LIST OF FIGURES

Figure 1. Wetlands percentage changes ............................................................ 1
Figure 2. Example spectral signatures .............................................................. 6
Figure 3. Study site watershed within Michigan ................................................. 12
Figure 4. Muskegon wetland complex ............................................................ 13
Figure 5. Data collection points ................................................................... 18
Figure 6. Reflectance factor for selected species ................................................ 33
Figure 7. Reflectance variability .................................................................. 41
Figure 8. Shape filter for invasives ............................................................... 43
Figure 9. ISODATA classification for image subset ........................................... 46
Figure 10. Accuracy example ..................................................................... 48
Figure 11. Patch size accuracy ..................................................................... 49
Figure 12. Study site transects ..................................................................... 50
Figure 13. Fractal and PLAND metric relations ................................................. 53
Figure 14. Infestation and disturbance metric .................................................... 54

VII

Chapter 1: INTRODUCTION

Global inventories estimate that approximately half of the world’s wetlands have
been lost due to human activities (Mitsch and Gosselink 2000). Over 50% of the 90
million hectares of wetlands estimated in the lower 48 states at the time of European
settlement have been lost due to human activities (Figure 1). In, the Great Lakes regions,
some states have lost 90% of the original wetlands primarily to intense agricultural

expansion, resource use, and urbanization (Dahl 2000).

I

 

 

Figure 1. Wetlands percentage loss in the past 150 years (Source: Dahl 2000).

The substantial loss of wetland area has gained attention over the past few
decades. The environmental and socioeconomic benefits of wetland ecosystems are now
well recognized. These benefits are in the form of services provided by the ecosystem

functions wetlands perform. These range from ground water recharge and flood control to

1

providing habitat and promoting biodiversity. Current policies and regulations at various
government levels require wetlands management practices to conserve and promote the
benefits wetlands provide. No net loss and assessment policies are in place which
requires monitoring and inventory to conserve the remaining wetlands.

Traditionally, wetland assessment techniques have largely relied on a few
physiochemical measurements as indicators of condition. Recent shifts in ecosystem
management goals have redirected efforts toward measuring ecological and biological
integrity, rather than only using physical and chemical measures of ecosystems as health
indicators (USEPA 1998, EMAP 2002). Here, integrity refers to a condition that is
determined to be characteristic of its natural condition considering history,
regionalization and scale, and levels of diversity and resiliency. Preserving and restoring
the ecological integrity of the remaining wetlands has become a priority (Danielson 2001 ,
NRC 1995, 2000). This shift is part of a growing consensus that wetland assessments
require a set of complementary indicators for a complete ‘check-up’ and assessment of
the threats to ecosystem fimctions (Paulsen et al. 1991, US EPA 1996, 1998).

Currently an overarching goal for wetlands assessment is to provide improved
information on invasive species (EMAP 2002, NRC 1995, 2000). Invasive species are
one of the largest threats to wetlands biodiversity and ecosystem functioning. In the US,
invasive species are estimated to cause $120 billion dollars per year in environmental
damage and associated control costs: Lythrum salicaria (purple loosestrife) alone is
estimated to cost $45 million per year as it spreads at a rate of 115,000 ha/yr across

wetlands in the US (Pimentel et al. 2005).

The term invasive is sometimes used interchangeably with exotic, alien, foreign,
introduced, and nonindigenous among others. The National Invasive Species Information
Center (USDA) defines an invasive species as non-native to an ecosystem and whose
introduction is likely to cause economic or environmental harm (USDA NRCS 2007).
The EPA and other organizations tend to focus on likelihood to cause ecological harm or
economic damage and not so much on alien categorization (EMAP 2002, NRC 2000).
The pre-industrial revolution (circa 1750) is often set as a benchmark for determining
whether a species is native or foreign. Currently there are approximately 7000 non-native
species introduced in the US, about 15% become invasive, and only a smaller percent of
those become a nuisance or cause harm (USDA NRCS 2007).

Ecologically, invasive refers to a set of characteristics that a species possess that
enable that species to establish, often aggressively, or invade. Those characteristics
include a plants reproduction ability (e. g., seed- abundance, persistence, vigor), growth
pattern (e.g., moisture use, anaerobic tolerance, density), and morphological adaptations
(e.g., foliage porosity, shape, form). The degree of invasiveness will vary by region and
environmental conditions. Generally, invasives have been found to alter ecosystem
functions and services such as hydrology, soil moisture, disturbance regimes, and
ecosystem resiliency. Invasives have a wide tolerance to environmental conditions (e. g.,
soil and water pH, temperature, and hydroperiod fluctuations), possess phenotypic
plasticity and unique life cycle features (e.g., reproduction, seed crop intervals), and can

often undergo rapid evolution.

1.1 Wetlands Remote Sensing

Wetlands mapping programs have largely focused on identifying extent or total
area. The National Wetlands Inventory (NW 1) program has relied extensively on fine-
scale aerial photography as its primary source of data. Human photo interpretation
techniques are utilized to delineate the extent of wetlands in a given area based on these
photos. This technique is by nature somewhat subjective as visual interpretation
approaches rely on an expert to delineate wetland boundaries. While acceptable accuracy
can be achieved for a region, the approach can be time consuming, expensive, and
extrapolation to other regions is problematic. The last large-area NW1 application is now
more than two decades old. Further, when using subjective approaches, discrepancies
exist between classifications (e. g., NW1 vs. Michigan Land Use Land Cover), program
objectives, and definitions used to map wetlands.

Satellite and airborne remote sensing has been utilized for nearly three decades to
inventory and identify wetlands (Hardinsky 1986). Similar to aerial photography, remote
sensing techniques are often less costly and time-consuming compared to conventional
field methods. Advantages of satellite and airborne data include regular overpass
intervals with capabilities for the mapping of wetlands and adjacent land use/covers,
monitoring human activities and impacts on wetlands, assessing wetland changes
seasonally, and extrapolation of information over large regions. Sensor instruments
record measurements in spectral regions outside of the visible spectrum. Additionally,
advances in technology and computer capabilities allow advanced modeling and
integration of remotely sensed data within a Geographic Information System (GIS).

Many wetlands mapping studies discuss the limitations of remote sensing

technology. Data resolution has been the foremost limiting factor in obtaining high

precision and detailed information on wetlands vegetation and biophysical characteristics
(Ozemsi and Bauer 2002). Coarse spatial resolutions have made delineations of wetland
boundaries challenging (Fortin et al. 2000, Torbick et al. 2006), while broad radiometric
and spectral resolutions prohibit species level separability (Becker et al. 2005, Schmidt
and Skidmore 2003). Generally fewer types, or categories, of wetlands can be mapped
due to complications with spatial resolution, temporal overpass constraints, and spectral
variability and overlap. In response to these limitations, the NWI program (in the USA)
has used color infrared aerial photography exclusively (Cowardin 1979, Tiner 1999).
New state of the art remote sensing technology can contribute to advancing
wetlands mapping and monitoring. Recent advances in sensor technology and remote
sensing science have promoted an interest in hyperspectral data for mapping wetlands at
the species level (Artigas and Yang 2006, Becker et al. 2005, Schmidt and Skidmore
2003, Thiemann and Kaufmann 2002, Thomas et al. 2002). Advanced spectroscopic
systems possess capabilities to capture data at narrow spectral bandwidths on the order of
three to ten nanometers (nm) contiguously covering large portions of the spectrum (e. g.,
350-2500nm). This allows for small variations in plant/substrate absorptance and
reflectance to be recorded (Figure 2). Incorporating such relatively high spectral detail
makes it possible to explore species separability and precise ecological process

monitoring (Schmidt and Skidmore 2003, Ustin et al. 2004).

 

 

— Phrag

— Salix
Lemna

— Scirpus

— Wet sand

— Sand 6'H20

 

 

 

 

350 650 950 1250 1550 1850 2150

 

 

 

Figure 2. Example plant signatures recorded using a hyperspectral instrument with

rescaled Landsat resolutions signatures overlaid.

1.2 Problem Statement

Traditionally most ecosystem assessment techniques have relied on a few physical
and chemical measures of systems as indicators of health. In order to meet firture goals
for enhancing wetland ecosystem integrity, resource managers and decision makers need
advanced information and improved methods for identifying wetland stressors and
ecological indicators, such as invasive plants, and for monitoring response to control
measures. Multispectral remote sensing has been a useful tool in monitoring and mapping
aquatic vegetation and stressors; however, the information provided is limited in
achievable detail by sensor resolutions.

Hyperspectral sensing technologies that capture narrow spectral and spatial
resolution data can advance assessment techniques for addressing complex wetland

assessment questions. A few studies have explored methods to identify wavelengths

possessing the greatest ability to differentiate wetland species (Becker et a1 2005,
Schmidt and Skidmore 2003). The ability of hyperspectral remote sensing technologies to

map wetland invasive plant infestation needs to be evaluated.

1.3 Research Objectives

The overarching goal was to develop monitoring technologies to map invasive
plants and quantify wetland infestation. The objectives were broken into two categories:
field-level and landscape-level. The first field-level objective was to characterize
absorption and reflectance features and assess processing techniques for separating
wetland species. The second field-level objective was to evaluate the abilities of a shape
filter, which requires unique absorption features for successful implementation, to
identify wetland invasive plant species. The overarching landscape-level goal was to map
invasive plant infestation. The first landscape-level objective was to classify
hyperspectral imagery in order to identify invasives of interest. The second landscape-

level objective was to quantify infestation within the study area.

Chapter 2: HYPERSEPCTRAL REMOTE SENSING

Satellite remote sensing has been a useful tool in providing general information
on wetlands types (Ozesmi and Bauer 2002); however, both spatial and spectral
resolutions have limited the level of detail ultimately required for comprehensive wetland
assessments. Recent advances in sensor technology and remote sensing science have
promoted an interest in hyperspectral data for mapping wetlands at the species level
(Artigas and Yang 2006, Becker et al. 2005, Hirano et al. 2003, Schmidt and Skidmore
2003, Thomas et al. 2002). Advanced spectroscopic systems possess capabilities to
capture data at narrow spectral bandwidths on the order of three to ten nanometers (nm)
contiguously covering large portions of the spectrum (e.g., 350-2500nm). This allows for
small variations in plant/substrate absorptance and reflectance to be recorded.
Incorporating such relatively high spectral detail makes it possible to explore species
separability and precise process monitoring (Schmidt and Skidmore 2003, Ustin et al.
2004)

The underlying theory of hyperspectral science is that increased spectral detail,
along with improved or adequate spatial resolution, can provide increased information
such as species-level mapping. Generally, the higher resolutions a sensor possess, the
greater the level of detail can be extrapolated. 1n the last few years several studies have
applied hyperspectral data for wetlands mapping. A primary goal in these investigations
was to evaluate and develop methods to utilize the increased level of data supplied via

hyperspectral instruments. The studies can be grouped into methods to identify

wavelengths of particular utility (processing techniques to extract and identify the most
useful bands) and evaluating classification algorithms to map species of interest.

Approaches to identify wavelengths that possess separation abilities are an
interest because of the large amount of spectral information and fact that much is
redundant. With hyperspectral instruments recording near-continuously, large portions of
the data tend to be highly correlated. At the same time the amount of storage space
required can be quite large as these sensors do in fact capture hundreds of narrow bands.
Processing protocols and limitations on spatial resolutions can be problematic due to the
large file sizes. Therefore identifying wavelengths of particular utility allows less spectral
information to be required simultaneously allowing for increases in spatial resolutions.

Becker et al. (2005) performed derivative analysis to identify unique points of
inflection along spectra for wetland plants in a Great Lakes coastal wetland and identified
eight bands as possessing the most utility for separation. The bands are located across the
visible (VIS) and near-infrared (NIR) portions of the spectrum and are affiliated with
domains that represent unique biophysical characteristics. The red-edge was highlighted
as having particular strength in separation. Artigas and Yang (2006) also identified the
NIR region using a discrimination metric (Cochrane 2000) and derivative analysis for a
New Jersey coastal wetland. This research concluded that monotypic stands of
Phragrnites could be identified by using the unique NIR response. Schmidt and Skidmore
(2003) found significant differences between salt marsh species using field-level
reflectance data from coastal Netherlands wetlands.

Developing and improving classifications has been a focus of hyperspectral

wetland remote sensing. Becker et al. (2007) examined the optimal spectral and spatial

resolutions for mapping Great Lakes coastal wetlands. A series of experiments tested
different bands, band combinations, and pixels sizes to identify the most advantageous
configurations to accurately map coastal wetlands. The results showed that narrow,
strategically located bands were necessary to achieve acceptable resiliency levels when
trying to limit the number of bands to maintain small pixel sizes. The research also found
that spatial resolutions of 1 meter or under were best situated to map coastal wetlands.
Obtaining pixels this size essentially requires use of airborne remote sensing platforms as
no satellite instruments possess hyperspectral spatial resolutions less than 20m.

Rosso et al. (2005) used Spectral Mixture Analysis (SMA) and Multiple End-
member (MESMA) techniques on airborne data to map a portion of coastal marshes in
central California. They found MESMA appropriate allowing for multiple end-members.
However, when similar end-members, defined by root-mean-square-error, were utilized
accuracy decreased substantially highlighting the need to carefully choose them for
image training. Rosso et al. (2005) further found that using less than half a dozen end-
members in a more homogeneous environment produced optimal results.

Some recent research focused on evaluating image processing techniques to map
wetlands has come to different conclusions. Underwood et a1 (2003) found increased
accuracy in using Minimum Noise Fraction (MNF), a data reduction technique, to map
Carpobrotus edulis in coastal California; while Artigas and Yang (2006) found less
separation ability after applying MNF for salt-marsh species in a New Jersey
meadowland. Underwood et a1 (2003) found band ratio techniques to perform moderately
well as an image processing technique to identify coastal Cortaderiajubuta, but it did not

outperform MNF enhancements. Artigas and Yang (2006) were able to distinguish

10

Phragmites australis from Spartina and Distichlis in the NIR region of imagery when
only these species were dominant. These papers highlight the fact that biological
community composition (number of species and level of heterogeneity) will largely

determine selection of image processing methodologies.

ll

Chapter 3: METHODS

3.1 Study Area

The investigation was carried out in an expansive wetland complex in the lower
Muskegon River Watershed (MRW) located on the western side of central Michigan
(W86° 09’ 45”, N43° 16’ 10”). The MRW is almost 7000 km2 in size, includes 94
tributaries and 183 stream segments (interrupted by 95 dams), and hundreds of lakes and
wetlands (Figure 3). Glacial deposits reach as much as 300m deep over bedrock and

create conditions with substantial ground-surface water exchange.

 

Figure 3. Location of MRW (left) within Michigan and overlay on false color (423:2)

Landsat ETM+ scene.

Fed from the Muskegon River and an extensive tributary network, the wetland
complex serves as the last land-water interface before draining into Lake Muskegon,
which flows directly into Lake Michigan (Figure 4). Lake Muskegon was recognized as

an Area Of Concern by the 1987 Great Lakes Water Quality Agreement due to poor

water quality conditions. Projections show that a 50% urban sprawl increase by 2040 is
likely in the coastal sub-watershed under current land use trends (Pijanowski et al. 2002).
Adjacent land uses in the immediate subwatershed are residential and urban
neighborhoods, industrial zones including a pulp and paper mill, chemical and
petrochemical companies, and recreation parks, along with expansive agriculture and
forest patches. NWI data classify the majority of the wetland complex as palustrine with
seasonally and semiperimanently flooded regions with scrub-shrub, forest, and emergent

COVCI’S.

 

Figure 4. Wetland complex resides adjacent to Lake Muskegon. Dashed red line

approximates hyperspectral imagery footprint overlaying 2006 aerial photo.

The two most problematic invasive species in the study area are Phragmites
australis and T ypha latifolia. P. australis, is a perennial grass species that has strains
native to temperate wetlands for at least the past thousand years. Recently, however,

Saltonstall (2002) found that nonnative strains of P. australis have aggressively spread

13

throughout the northeastern US and Great Lakes region. T. latifolia is a herbaceous,
rhizomatous perennial plant and is common throughout the US and temperate and
tropical places worldwide. T. latzfolia is exploitative in its ability to clone rapidly and
produce a large litter amounts and biomass, which may contribute to its superior
competitive ability (Grace and Wetzel 1982). Every state in the US has now reported the
presence of Phragmites australis and T ypha latifolia.

The increasing rate of expansion is now a primary concern. Regionally, Lynch
and Saltonstall (2002) found that P. australis invasion in Lake Superior wetlands was
dominated by native strains, although nonnatives were present. Wilcox et al. (2004)
found that 90% of P. australis populations expanding into Long Point, Ontario on lake
Erie were nonnative strains of the species, and suggested that declining water levels in
the Great Lakes may have contributed to its expansion. Given that fluctuations in water
levels are associated with high plant diversity in coastal wetlands (Wilcox and Meeker
1991), stable water levels caused by management or natural climate change may favor the
spread of dominant invasive species. Normative or native classification is less relevant as
these two species are both considered a nuisance species because they are persistent,
produce large amounts of biomass, propagate easily, and are difficult to control with
mechanical, biological, or chemical means.

The effect of this rapid expansion is largely unknown, but is generally believed to
be negative (Roman et al. 1984; Marks et al. 1994). Plants can influence ecosystem
functions like nutrient cycling through their morphological, phenological, and metabolic
traits (Windham and Ehrenfeld 2003). Increasing cover of these invaders has known

multiple impacts on the ecology of wetlands, as it can reduce nutrient availability to

14

native species, shift community composition, and alter food webs (Mack and D’Antonio
2003, Saltonstall 2002, Uzarski and Burton 2005, Windham 2001, Windham and

Ehrenfeld 2003).

3.2 Data
3.2.1 Field-level data

A field campaign was conducted during mid-August (2005 and follow up 2006),
which generally represents the peak of the growing season for wetland species within the
study area. Capturing data during peak phonological cycles has been shown to increase
the separability of invasive wetland plants (Laba et al. 2005). Due to the expansive nature
of the wetland complex and the challenge of moving through a wetland, a compromise
between operational feasibility and statistical sampling rigor was required. Both logistical
constraints (equipment setup and takedown) and traveling throughout the wetland
complex required a substantial amount of time. Reconnaissance field work identified two
primary emergent pools where high biodiversity and ecologically noteworthy species of
interest (i.e., invasives) were present. Focusing our efforts around these two regions of
the wetland complex allowed the largest number of species spectra to be collected with
minimal time required between sampling sites. An airboat provided the most efficient
access for traveling around the wetland complex. Data acquisition focused on the
dominant terrestrial-, emergent-, and submergent- species. Dominance was qualitatively
identified during reconnaissance field work by evaluating percent cover and the
approximate size of a patch for a species. Eight measurements were averaged for one

spectrum and approximately nine spectra were collected in each homogeneous plot. A

15

total of twenty-two (Vallisneria americana repeated for two different water depth
conditions) wetland plant species were recorded (Table 1). Eight species are identified as
being invasive (USDA, NRCS 2006). Note that not all the species identified as invasive
are classified as exotic and the degree of ‘invasiveness’ can vary by region and

conditions.

Table 1. Plant species sampled in the study area.

 

 

 

 

 

Common Name Genus Sgecies Invasive
Arrowhead, broadleaf Sagittaria latzfolia Invasive
Bulrush, softstem Scirpus validus
Cattail, broad-leaved Typha latifolia Invasive
Cutgrass Leersia oryzoides
Eelgrass Vallisneria americana
Spikerush, beaked Eleocharis rostellata
Canadian waterweed Elodea canadensis Invasive
Bur-reed, branched Sparganium androcladum
Filamentous green algae
Grassleaf mudplatain, water star grass Heteranthera dubia
Iris, harlequin blue flag Iris versicolor
Common duckweed Lemna minor
Watermilfoil, whorl-leaf Myriophyllum verticillatum
Mowed field grass
Pennsylvania smartweed Polygonum pensylvanicum Invasive
Common reed Phragmites australis Invasive
Pickerelweed Pontederia cordata Invasive
Spiral pondweed Potamogeton spirillus
Purple loosestrife Lythrum salicaria Invasive
White water lily Nymphaea odorata Invasive
Willow Salix eriocephala
Yellow pond lily Nuphar lutea

 

 

 

 

 

 

We used a portable spectroradiometer (FieldSpec Pro FR®, Analytical Spectral
Devices, Inc., Boulder, Colorado) to collect in situ radiance between 350-2500 nm
(visible to shortwave infrared). Spectral resolution (full width half maximum) was
recorded at 3 nm in the visible wavelengths and 10 nm in the infiared region. The sensor

16

was equipped with a 24 degree field-of-view (FOV) optic and held approximately 1-
meter above the target at nadir for measurements representing field-canopy conditions.
Sun-target—sensor geometry was repeated as best as possible under these difficult field
conditions. The viewing geometry configuration approximately represents the spatial
resolution current airborne hyperspectral sensors can achieve. A reference Spectralon®
panel (Labsphere, Inc., North Sutton, New Hampshire) was used for calibration during
processing and atmospheric adjustments. During data acquisition, the sensor was first
placed over the reference panel to record the panel-reflected radiance. Then the sensor
was placed over the target to record the target-reflected radiance. Then, by ratioing the
radiance measurements, surface reflectance factor was calculated. By definition, the term
reflectance factor is the ratio of radiant emittance of a target (i.e., wetland plant) to that
reflected into the same reflected-beam geometry and wavelength range by an ideal and
diffuse standard surface (i.e., Spectralon calibration panel) irradiated under the same
conditions (Schaepman-Strub et al. 2006). The reflectance factor was calculated based on

the following equation:
p=tET/cpET I (1)
where p is in situ reflectance factor for target of interest
(wetland plant species), tE T is target (wetland plant
species) in situ radiance, and cpE T is the calibration panel

in situ radiance.

l7

Subsequent data processing in this study also removed wavelength regions severely
affected by atmospheric absorption in the spectral ranges of 1350-1480 nm, 1775-2000

nm, and >2400 nm (Thenkabail et al. 2004).

 

Figure 5. Ground control marking data collection locations. Polygons were stratified

throughout the complex and included a range of shapes and sizes.

Field-level ground control was collected during August 2005 and 2006 (Figure 5).
The 2005 field campaign was within a few days of the airborne imagery fly over but not
simultaneous. The ground control was collected in order to characterize the accuracy of
the classified airborne imagery. The ground control was focused on collecting reference
data for the invasive plant species of interest, primarily large T ypha latifolia and
Phragmites australis patches as these are the two primary invasives of concern in the
region. A Trimble Pro XR Global Positioning Systems (GPS) was utilized to record

ground polygons. A suite of polygons ranging in size, shape, and location were collected

Potential sites were identified using airborne photographs and on-the-fly sites were

collected while traversing the study area.

3.2.2 Landscape-level data

A flyover was performed on the 24th of August 2005. A light aircraft was
equipped with a push-broom Airborne Imaging Spectroradiometer for Applications
(AISA) and a GPS differential navigation system. The AISA sensor (16-bit) is a linear
array sensor (push-broom) that uses a system of semiconductive elements (e.g., charge-
coupled device array) to record one line of an image simultaneously. Concurrent
downwelling radiance was integrated using a Fiber Optic Downwelling Irradiance
System (FODIS) configured to match the upwelling radiance measurements. Twelve
overlapping flight lines and AISA configuration calculated apparent at sensor reflectance
covering the VIS—NIR domains (Table 2). AISA pre-processing (CaliGeo) included
geometric correction and image rectification. Flown at an above ground altitude of
approximately 1000m, one meter pixels were obtained for twenty bands in the visible

through near infrared.

19

Table 2. AISA hyperspectral configuration and resolutions for airborne data-collection

performed during August 2005.

 

 

 

nbr min. wvl I center wvl rn.ax wvl I ch. width I avg offset Ia av.g gain

1 434.45 438.43 442. 4 7. 95 21. 8771
2 451.94 455.92 459.89 7.95 0 17.8561
3 459.89 463.87 467.84 7.95 0 15.9328
4 499.64 503.62 507.59 7.95 0 7.1509
5 518.72 522.7 526.67 7.95 0 5.6548
6 575.56 579.73 583.91 8.35 0 3.382
7 585.58 589.75 593.93 8.35 0 3.2076
8 620.65 624.82 629 8.35 0 2.9989
9 644.03 648.2 652.38 8.35 0 2.9429
10 654.05 658.22 662.4 8.35 0 2.8382
11 674.09 676.6 679.1 5.01 0 3.9341
12 684.11 685.78 687.45 3.34 0 5.3284
13 689.12 691.62 694.13 5.01 0 3.4042
14 723.06 725.65 728.25 5.19 0 3.4821
15 731.71 734.31 736.9 5.19 0 3.5483
16 738.63 741.23 743.82 5.19 0 3.6101
17 757.66 760.25 762.85 5.19 0 4.0223
18 783.61 786.21 788.8 5.19 0 5.2317
19 833.78 836.38 838.97 5.19 0 5.8467
20 877.03 879.62 882.22 5.19 0 7.3689

 

 

3.3 Quantifying separation

A variety of separability measures can be used on spectral signatures to identify
similar and divergent spectra. The Jeffries-Matusita (JM) distance measure has been
utilized successfully to quantify the divergence of spectral signatures including wetland
vegetation (Schmidt and Skidmore 2003). Slight variations of the JM formula exist. The
Ayala and Menenti (2002) algorithm (Eq. 2 & 3) was used to evaluate spectral

separability and the impacts of processing techniques executed:

JMij = ,/2(1- e"“) (2,

20

(C. + 0,)“
2

1 _ 1 [C,.+C.]/2
a=§(U.-—U,-)T (U,—Uj)+—Ln I LI

2 ICi :1:ch

 

 

(3)

where Jsz is the Jeffiies-Matusita distance between
signatures i and j, T is transpose, C,- is the variance-
covariance matrix of signature i, Ui is the mean vector of

signature 1', I C,- I is the determinant of C,-,

3.4 Continuum removal

A processing technique that allows for the extraction and modeling of individual
spectral features (absorptance and reflectance) is continuum removal (Clark and Roush
1984). This technique is increasingly being implemented in hyperspectral vegetation
investigations as a processing technique to extract parameters of interest (Huang et al.
2004, Kokaly et al. 2003, Schmidt and Skidmore 2003, Underwood et al. 2003). Used
extensively in geological applications, continuum removal disregards albedo, or
background, to obtain individual features (absorption/reflectance) such as peak
wavelength position and depth (van der Meer 2004). Schmidt and Skidmore (2003) found
that applying continuum removal to salt-marsh vegetation spectra improved species
separation in the visible spectrum, but decreased it in the near-infrared (N IR) and
shortwave-infrared (SWIR) regions of the spectrum. This is important as the NIR domain
is often identified as an area of key utility for separation. Underwood et al. (2003) found
that continuum removal performed moderately well in improving hyperspectral image

classifications when trying to identify species with high leaf water content.

21

Aspects of continuum removal not previously addressed in detail relate to the role
of plant canopy architecture and the impacts of continuum removal on separating
individual species of interest such as invasives. Schmidt and Skidmore (2003) suggest
that if continuum removal eliminates noise from the soil background, moisture content,
and canopy structure, then only the varying biogeochemical content of a species would
determine separability levels. The impact of continuum removal related to species canopy
architecture needs to be investigated to assess its contribution to distinguishing wetland
invasive species.

A modified continuum removal technique was developed for the wetland
vegetation spectra following methods outlined in Schmidt and Skidmore (2003). For
wetland spectra, amplitudes in the near-infrared can be large enough to disregard
absorption features in other spectral regions, such as the visible portion of the spectrum.
Therefore, a modified convex hull was forced to include seven primary spectral
reflectance maxirna distributed among spectral regions. Once the modified continuum
removal was forced, the continuum was removed by dividing the reflectance by the
convex hull (Clark and Roush 1984). The seven primary spectral regions that isolate the
major reflectance features were:

0 Visible domain and chlorophyll absorption region (350-67 5 nm)
0 Near-infrared edge (676-780 nm)

0 Near-infrared plateau (781-975 nm)

0 Near-infrared down slope (976-1190 nm)

0 Upper near-infrared shoulder (1191-1450 nm)

0 First shortwave infrared plateau (1451-2000 nm)

22

0 Second shortwave infrared plateau (2001-2400 nm)

3.5 Shape filter

In reality, plant canopy architecture and other factors create reflectance variability
and reflectance overlap between plant species that makes species-level identification
challenging. Vegetation reflectance varies across the spectrum with the visible domain
largely determined by the chlorophyll content, the NIR region a function of leaf structure
and biomass volume, and the SWIR region largely determined by leaf water content and
biomass volumes (Cochrane 2000, Danson 1995). Generally, differences in absorption
determine the amount of variation and spectral overlap between vegetation species. To
examine variability, a shape filter (Cochrane 2000) that incorporates species reflectance
variability was applied to evaluate separability of wetland invasive species in the study
area based on spectra shape. The maximum (Max) and minimum (Min) spectral

reflectance creates the shape-space for the species (Eq. 4).
Shape-space = Maxp : M lnp / ’1 (Eq.4)

3.6 Image classification

The primary objective associated with the imagery was to characterize infestation.
In order to complete this objective an accurate map was needed. Becker et al. (2007)
assessed classification algorithms and image resolutions for mapping coastal wetland
imagery. The research here is intended to build upon these efforts and not repeat

classification technique assessment. A series of classification algorithms were applied

23

and evaluated in order to achieve the most accurate depiction of the invasive plants of
interest within the landscape. An expert system approach was carried out that primarily
relied on two classification algorithms. A basic unsupervised clustering algorithm was
first executed and refined, then combined with a more advanced supervised algorithm
using image derived training data.

An unsupervised classification is generally considered a straightforward, basic
approach to group pixels based on spectral properties in n-dimensional space. The
Iterative Self-Organizing Data and Analysis Technique (IOSDATA) algorithm was
applied. Using a set of defined criteria that have been developed over the past two
decades, ISODATA approaches can be very effective and sophisticated. The approach
here included:

- The maximum number of classes was defined at forty based on
reconnaissance field work, ancillary data, and expert judgment

0 Statistical initialization was based on diagonal axis means which are
computed to be along a diagonal vector and are evenly distributed within the scaling

range for each band

0 The number of iterations was set at ten to recalculate cluster mean vectors
each iteration
- Convergence threshold was set to ninety-five percent. The convergence

threshold is the maximum percentage of pixels whose cluster assignments can go

unchanged between iterations

24

A second, more advanced supervised approach was carried out to classify the
species of interest in the landscape using ground—truthed training data rather than image
statistics. The Spectral Angle Mapper (SAM) algorithm (Kruse et al. 1993) was used in
this context. This algorithm relies on spectra shape and angles formed between a
reference spectrum and an unclassified pixel in n-dimensional space where n represents
the number of bands. The primary advantage of SAM is that the classifier is insensitive to
albedo and illumination effects (Sohn et al. 1999). The SAM approach has been shown to
correctly classify spectrally similar targets (South et al. 2004).

Using the ground-truthed GPS collected data, image training sites were
developed. Slight spatial adjustments were required due to registration and
georeferencing discrepancies. NRCS aerial color photographs were used as reference data
to assist in spatial adjustments. Spectral signatures were used to create a signature library.
Forty total signatures were created including the two primary species of interest- T ypha
latifolia and Phragmites australis. This library was developed within ENVI and fed into

the SAM protocol using imagery mosaics.

3.7 Landscape pattern metrics

Landscape-pattem metrics provide a means to quantify ecosystem characteristics
at various scales. In the past two decades an increasing number of investigations have
explored landscape-stressor relationships for wetland ecosystems using pattern metrics
(Liu and Cameron 2001, Lopez et al. 2002, Keams et al. 2005, Torbick et al. 2006). In
general, all landscape pattern metrics have strengths and weaknesses. When using metrics

to explore relationships and patterns, it is important to understand their limitations,

25

appropriate applications, inter-relationships, and possible interpretations of results
(Gustafson 1998, Hargis et al. 1998, Li and Wu 2004). Identifying specific thresholds of
indices to guide decision making and management practices individually has proven to be
a challenge when using landscape pattern measures. By comparing characteristics and
ecosystem attributes across the landscape at various scales using complementing
measures, a more thorough indication of relationships is possible. This can help focus
decision making and monitoring.

Several studies have used pattern metrics to indicate landscape patterns and
wetland stressors. The methods vary widely and range in results. The overarching goal of
using the landscape pattern metrics was to quantify and assess relationships between
hydrology and infestation. Simple metrics that describe the percentage and distribution of
the invasives of interest were generated using Fragstats v3.3 (McGarigal and Marks
1995) for the class (species) and landscape level. Composition measures, including
percent cover, were generated to calculate the total area covered by the invasives.
Configuration metrics, including the Aggregation Index and the Interspersion and
Juxtaposition Index (111), were generated to identify large homogenous patches and
assess dispersion. The Simpson Diversity Index (SDI), dependant only on the class of
interest, was chosen as the diversity metric.

Area-based metrics are a suite of measures quantifying fundamental information
on the ecological utility of a patch. Basic area metrics are perhaps the most important
metrics and provide information for many other metrics (Hargis et al 1998). Therefore
area and percent of landscape (PLAND) were generated to illustrate landscape

composition-invasive relations.

26

_ 1
Area — all (10,000) (Eq.5)

Area equals the area of a patch, class, or landscape. (1” is the area

of patch if.

n

2%-

PLAND = P,- = PIA (100) (Eq.6)

 

PLAND is the sum of the areas of all patches of the corresponding
patch type, divided by the total landscape area (converted into
units of preference). Pi=proportion of the landscape occupied by

patch type (class) i, aij = area of patch ij, A=total landscape area.

Aggregation Index is calculated from the number of like adjacencies of
corresponding classes (types), divided by the maximum possible number of like

adjacencies (McGarigal and Marks 1995).

g..
AI = "

g,-,- equals the number of like adjacencies (joins) between pixels of

 

patch type i based on the single-count method, maX-gri is
maximum number of like adjacencies (joins) between pixels of

patch type i based on the single-count method.

27

A disturbance metric was developed to assess relationships between infestations
and hydrogeomorphic modifications (HGMM) and human disturbances. The study area
was divided into contiguous 1km wide transects, or sections, for a total of 15 full
sections. Using a knowledge-based evaluation system, each section was assessed for level
of human disturbance and given a rating between 1 to 5 representing low to high
disturbance and/or activities. Human activities and/or disturbances were measured in
terms land use pressures and landscape modifications (agriculture, urban), intensity of
managements and disturbances (continuous crop fields, dense impervious road network,
etc...), and hydrological modification (river channelization, dredging). Following the
river continuum framework, adjacency and spatial dependence was incorporated into the
knowledge-based ratings. The river continuum framework emphasizes a holistic approach
considering system openness, or the notion of spatial scaling, in a landscape adjacency
context (Naiman et a1 1988, Weins 1989). For example, if a section had intense
agriculture along its boundary this increases the disturbance rating. Further, if a pristine

section was adjacent to a highly disturbed section, this would cause an increase.

28

Chapter 4: RESULTS AND DISCUSSION

4.1 Spectral separation

The wetland plant spectra displayed a range of JM distance values. Table 3
displays the JM distance values in a matrix against the invasive species (USDA, NRCS
2006). According to the JM values, Heteranthera dubia (grassleaf mudplain) and
Lythrum salicaria (purple loosestrife) are relatively easy to separate with the highest JM
value of 1.2645. Heteranthera dubia has the highest separability from most of the
invasives compared to the other species collected (Table 3). The invasive with the lowest
separation value from Heteranthera dubia is Elodea canadensis (Canadian waterweed) at

0.8599, still a moderate separation value.

29

Table 3. JM distance matrix for invasive species. Vallisneria americana-18 refers to a
collection of spectra made with Vallisneria americana located approximately 18inches

below the water surface (averaged spectra n=3 2). Low, medium, high detail three evenly

 

 

divided categories.
Wetland species Elodea Lythrum Nymphaea Phragmites
canadensis salicaria odorata australis
Eleocharis rostellata 0.4213 0.1 124 0.3245
Elodea canadensis
Filmacutee 0.3406 0.0969 0.2435

Heteranthera dubia

Iris versicolor

Leersia oryzoides 0.2365 0.3381
Lemna minor 0.2525 0.353
Lythrum salicaria - 0.3251 0.1093
Mowed field grass 0.1 132 0.2187 0.0051
Myriophyllum verticillatum

Nymphaea odorata 0.3251 - 0.2225
Nuphar lutea

Phragmites australis
Ponterderia cordata
Polygonum pensylvanicum
Potamogeton spirillus
Sagittaria lanfolia

Salix eriocephala

Scirpus validus
Sparganium androcladum
Typha latifolia

Vallisneria americana

 

 

Vallisneria americana-l8

30

Table 3 (cont).

 

 

 

Wetland species Polygonum Pontederia Sagittaria Typha
pe Ivanicum cordata lati oIia lati olia

Eleocharis rostellata 0. 3663 0.25 79 0. 3535

Elodea canadensis

F ilmacutee 0.2753 0.168 0.2672 0.3305

Heteranthera dubia

Iris versicolor

Leersia oryzoides 0.2949 0.4004 0.307 0.302
Lemna minor 0.309 0.4139 0.3218 0.4126
Lythrum salicaria 0.0763 0.1818 0.079

Mowed field grass 0.0693 0.0761 0.0438 0.1546
Myriophyllum verticillatum

Nymphaea odorata 0.2712 0.1585 0.2551 0.4629
Nuphar Iutea

Phragmites australis 0.0643 0.078 0.0388 0.6234
Ponterderia cordata 0. l 179 - 0.1 044 0.1224
Polygonum pensylvanicum - 0.1 179 0.0275
Potamogeton spirillus 0.3625 0.254 0.3499 0.318
Sagittaria lanfolia 0.0275 0. 1 044 - 0.0988
Salix eriocephala 0.2519 0.1389 0.2404

Scirpus validus

Sparganium androcladum

Typha latrfolia -
Vallisneria americana

 

Vallisneria americana-l8

Heteranthera dubia and Elodea canadensis are both perennial forbs with
somewhat similar morphology and physiology, growing at the water-surface with a mat-
like foliage texture. Elodea canadensis tends to grow at a higher density, with toothed
leaves (6-15mm) in whorls, mostly floating just under the water surface; whereas
Heteranthera dubia tends to grow along the water surface with long, linear leaves (10-
15cm) joined at the base to a tubular sheath wrapped around a stem. Therefore, both

species often have substantial amounts of water present in their spectra under field-

31

canopy conditions resulting in distinguishable signatures compared to the other emergent-
and upland- aquatic invasive species in the study area. The VIS, NIR, and SWIR
reflectance for Heteranthera dubia and Elodea canadensis spectra never surpassed ten
percent reflectance factor because of the high amounts of water absorbing energy in the

FOV.

32

 

Elodea canadensis

 

 

 

salicaria

 

 

 

 

 

 

australis

MH-

 

Figure 6. Reflectance factor at 50mm intervals for selected invasive plant species. Pre-
processing removed wavelength regions severely affected by atmospheric absorption in
the spectral ranges of 1350-1480 nm, 1775-2000 nm, and >2400 nm. Averaged spectra

(11:32).
33

Nymphaea odorata (water lily) had the lowest separation distance from all other
species, while Sagittaria latifolia (arrowhead) had the lowest separability value against
the other invasive species. Sagittaria latifolia had a relatively high separability measure
from submergent and emergent species such as the invasive Elodea canadensis at 1.0421,
but very low separation scores from the other six invasives with an average of 0.1006
indicating difficulty in identifying this species. Sagittaria latifolia had a very small,
separation distance from Polygonum pensylvanicum (Pennsylvania smartweed) and the
two are likely to cross-classify. These two plant species have relatively similar plant
architectures and inhabitat very similar niches in this ecosystem. Sagittaria latifolia is a
medium height perennial herb with erect orientation that can grow upwards of a few feet
from the water surface. Its structure has arrow-shaped, simple leaves (6-40cm) with small
white flowers arranged in whorls on the stalk (present during data collection) that fills out
to reach medium foliage texture with moderate porosity. Polygonum pensylvanicum is an
annual herb with lance-shaped leaves (25cm) that grows a few feet upward from the
water with flowering branches that has moderate porosity with medium foliage texture.
The similarity of these two species with respect to the JM index lends support to the
claim that plant canopy structure plays a large role in species separability. In contrast,
Sagittaria latifolia also had very low separation values from Phragmites australis
(common reed), an aggressive, very densely growing erect stalk with coarse texture that
extends upwards of 2m, and Nymphaea odorata, a floating-leaf forb with large, thick
circular leaves (25cm) that rest on the water surface, which suggests that plant canopy

structure does not play a singularly strong role in differentiating spectra.

34

4.2 Continuum removal

In theory, a normalization process based on continuum removal can remove
albedo, or background signal, from a spectral signature. In this study, the modified
continuum removal mostly decreased separation abilities (Table 4). The differentiation of
Lythrum salicaria from Sagittaria latifolia, Potamogeton spirillus (spiral pondweed), and
Polygonum pensylvanicum increased slightly, although these species had very high I M
separation values before the continuum removal was applied. The separability of
Potamogeton spirillus from five of the invasives also increased slightly. Potamogeton
spirillus can be submerged or float on the water surface with long (20cm), simple leaves.
Therefore, an increase in species separability for Lythrum salicaria, Nymphaea odorata,
Phragmites australis, Pontederia cordata (pickerelweed), and Sagittaria latifolia was
contrary to expected results based on plant architecture, while emphasizing the role of
leaf water content in separation. The decrease in separability from Elodea canadensis
further suggests that background signal and canopy architecture were indeed removed via

continuum removal.

35

Table 4. Change in J M distance values with modified continuum removal applied
Negative values indicate increase in separation values. Vallisneria americana-18 refers to
a collection of spectra made with Vallisneria americana located approximately 18inches

below the water surface (averaged spectra n=32).

 

 

 

 

 

 

Wetland species Elodea Lythrum Nymphaea Phragmites
canadensis salicaria odorata australis
Eleocharis rostellata 0.5615 0.3275 0.0344 0.1982
Elodea canadensis 0.7279 0.6421 0.7038
Filmacutee 0.5976 0.2086 .I... :9..91.l1. . ”II 0.0843
Heteranthera dubia 0.8599 1.2645 1.228 1.2528
Iris versicolor 0.3549 0.6905 0.5135 0.6105
Leersia oryzoides 0.8061 0.2127 0.5114 0.2968
Lemna minor 0.8418 0.2326 0.5038 0.3382
Lythrum saicaria 0.7279 0.3017 0.0765
Mowedfield grass 0.739 0.0626 0.1648
Myriophyllum verticillatum 0.0678 0.8029 0.6837 0.7336
Nymphaea odorata 0.6421 0.3017 0.1707
Nuphar Iutea 0.4064 0.6282 0.433 0.5388
Phragmites australis 0.7134 0.0401 0.1707
Pontederia cordata 0.7275 0.1201 0.1103 0.0087
Polygonum pensylvanicum 0.797 0.1822
Potamogeton spirillus 0.5914
Sagittaria latifolia 0.7686
Salix eriocephala 0.6238
Scirpus validus 0.4855 0.3898
Sparganium androcladum 0.5946 0.9468 1.1649 1.199
T ypha latifolia 0.3958 0.6254 0.4134 0.5229
Vallisneria americana 0.8959 0.7165 0.741 1
Vallisneria americana-18 0.6076 0.6523 0.8297 0.8302

 

36

Table 4 (cont).

 

 

 

 

 

 

Wetland species Polygonum Pontederia Sagittaria Typha
en Ivanicum cordata lati olia lati olia
Eleocharis rostellata 0.2157 0.1488 0.1879 0.4562
Elodea canadensis 0.797 0.7275 0.7686 0.3958
F ilmacutee 0.1221 0.0495 0.0859 0.2067
Heteranthera dubia 1.2552 1.255 1.2567
Iris versicolor 0.6689 0.6248 0.6279 0.5645
Leersia oryzoides 0.2194 0.3624 0.2326 0.2826
Lemna minor 0.2221 0.355 0.2523 0.3565
Lythrum saicaria 0.1201 I, . :0,0,08,_ _,I 0.6254
Mowedfield grass 0.0157 0.0362 0.0063 0.1058
Myriophyllum verticillatum 1.0843 1.0943 1.2034 0.8955
Nymphaea odorata 0.1822 0.1103 0.1641 0.4134
Nuphar lutea 0.5914 0.5437 0.5517 _ 0.6245
Phragmites australis 0.0087 £0,927}, -3 0.5229
Pontederia cordata 0.1179 0.0401 0.4901
Polygonum pensylvanicum 0.0748 0.5285
Potamogeton spirillus 0.3194
Sagittaria latifolia 0.0078
Salix eriocephala 0.2054 0.2036
Scirpus validus 0.5166 0.4748 0.4916 0.5062
Sparganium androcladum 0.7865 1.1033 1.2021 1.1462
T ypha latifolia 0.5285 0.4901 0.5072
Vallisneria americana 0.8655 0.8042 0.8221 0.8634
Vallisneria americana-I8 0.9187 0.8852 0.8752 0.6947

 

The continuum removal results suggest that the processing technique is not
necessarily useful for all vegetation applications. While continuum removal might be
effective in identifying absorption feature characteristics or particular wavelengths
associated with biophysical attributes, applying the technique for separating plant species
(or classifying image data) might be disadvantageous. Clearly, for wetland ecosystems,
continuum removal decreased abilities to distinguish invasives species. These results
further suggest that background and canopy architecture may contribute to improving
separation of wetland plant species. The results here advocate emphasizing plant canopy
architecture when attempting to map wetland invasive species.

37

Plant canopy architecture is not only the morphological and biophysical structure
of a species. The plant canopy architecture is also representative of the
microenvironment; the background removed via continuum removal. The background
signal, or local environment, is what often creates conditions that support hydrophytic
plants. The background signal includes variations in soil moisture or water content along
with understory debris and previous plant growth. These background factors provide
useful biophysical information that is well-known to be measured spectrally. Therefore,
when continuum removal techniques are applied, the loss of these background signals is
detrimental to spectral separation of species in many cases. In other ecosystems or
applications, such as geological and mineral identification, background signal may not be

usefirl; in wetland ecosystems these are critical.

4.3 Absorption feature characteristics

The peak reflectance and wavelength location in each of the seven spectral
regions varies by species. This is important because identifying the best wavelengths can
be arbitrary if those wavelengths are not unique to an invasive species of interest. The
SWIR plateaus (1451-2000 nm, 2001-2400 nm) have the largest variation in peak
reflectance wavelength. When normalized by the number of wavelengths in each spectral
domain (range of peak wavelengths/number of wavelengths in spectral domain), the
second SWIR plateau has the greatest range in peak wavelength locations, followed by
the NIR edge (781-975 nm) and first SWIR plateau (1451-2000 nm). This indicates that
these three spectral domains possess the largest variation in peak reflectance wavelength

location, which can assist in species discrimination. These results are consistent in

38

identifying the NIR edge as this wavelength region has been recognized as a useful
spectral domain for distinguishing wetlands plants using other statistical techniques,
namely second derivative analysis and Mann-Whitney U-testing (Artigas and Yang 2006,
Becker et al. 2005, Schmidt and Skidmore 2003).

Peak reflectance wavelength locations for Lythrum salicaria are furthest from the
average of all other species in the first SWIR and NIR domains and to a lesser extent the
NIR down slope and second SWIR domains. The peak reflectance wavelengths in these
spectral domains are, therefore, more usefirl for discriminating Lythrum salicaria from
the other wetland species compared to the chlorophyll absorption region, NIR edge, and
the water absorption region. As such, the leaf moisture content and internal leaf structure
(Cochrane 2000, Danson 1995) of Lythrum salicaria are more useful characteristics for
separation than are plant pigmentation differences represented by the visible and
chlorophyll domain. However, when background was eliminated via continuum removal,
little increase in separation scores resulted.

The NIR regions (edge, plateau, and down slope) have the greatest range of
reflectance maxima between species. The visible domain and SWIR plateaus have
relatively narrow ranges between the peak reflectance of the wetland species. The
differences between species’ leaf structure, largely represented by the NIR and water
absorption regions, make the differences in reflectance amplitude a useful characteristic
for separation. However, for a particular species of interest, the most unique peak
reflectance feature might reside within a different wavelength domain. For Lythrum
salicaria, the water absorption region, followed by the NIR down slope, possess peak

reflectance features fiirthest from the average of all other species. The peak reflectance

39

features of the two SWIR plateaus have relatively equivalent separation abilities as the
NIR edge and NIR plateau for Lythrum salicaria (reflectance factor for Lythrum salicaria
— average reflectance factor for all species). The NIR edge is often emphasized as a
useful wavelength region for distinguishing species (Artigas and Yang 2006, Becker et a1
2005, Schmidt and Skidmore 2003, Thenkabail et al. 2004); however, for identifying
Lythrum salicaria other wavelength regions show larger distances in terms of peak
reflectance maxima and peak reflectance wavelength location. Thus, individual plant
absorption/reflectance features should receive some consideration when attempting to

map invasives or species of interest.

4.4 Shape filter

While the absorption/reflectance features from average species spectra provide
useful information, in reality the reflectance for individual wetland plant spectra display
considerable variation. Figure 7 illustrates the reflectance variability for Scirpus validus
(softstem bulrush), Phragmites australis, Lythrum salicaria, and T ypha latifolia (broad
leaved cattail). Recall that the shape filter method (Cochrane 2000) is intended to identify
species of interest, such as invasives, using the uniqueness of the absorption features and
reflectance variability. In essence, the more unique an absorption feature of a given

species is, the easier that species can be distinguished.

40

 

 

0.7

   

 

 

 

 

 

 

 

 

Scirpus validus Phragmites australis
.. 0.2 0-5 0.2
.3 0.5
.52 g 0.4
a) 0.1 0.3
‘c’ D o 2 0'1
to ‘D '
g 0. 1
50:) 0 0 0
or 35085013501me 350850135018502350
nm
0'7 Lythrum salicaria 0'7 Typha latifolia
0.6 02 0-6 02
0.5 ’ 0.5 '
0.4 0.4
0.3 0,1 0-3 0.1
0,2 0.2
0.1 0.1
O 0 0 O

   

 

 

 

 

351850135018502350 350$O135018502350

 

 

Figure 7. Reflectance factor (n=32) variability, or shape-space, (maximumo,
minimum A , range- A-) illustrated for Scirpus validus, Phragmites australis, Lythrum
salicaria, and T ypha australis at 50nm intervals. Pre-processing removed wavelength

regions severely affected by atmospheric absorption in the spectral ranges of 1350-1480

nm, 1775-2000 nm, and >2400.

The shape space varies by wavelength domain and by species. For example,
Scirpus validus had relatively less variation compared to Phragmites australis, Lythrum
salicaria, and T ypha latifolia. The variation Scirpus validus did possess occurred
primarily in the water absorption region, first SWIR plateau, and second SWIR plateau
while increasing respectively. This is likely due to the erect, small diameter structured

grth and the fact that Scirpus validus tends to occur as a transitional plant between

41

standing water and higher substrate on a microtopographic scale. Thus only minute
differences were detected in leaf water content and plant biomass volume compared to
variation in soil moisture and understory debris, again emphasizing the utility in
background signal for identification purposes. Phragmites australis and T ypha latifolia
have larger reflectance variability in the NIR down slope (976-1190nm) and the water
absorption region (1191-1450nm); however, both these regions have high separation

abilities when the shape filter was applied (Figure 8).

42

Elodea canadensis Lythrum salicaria

 

 

 

 

 

 

 

 

# of species separated

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 8. Number of plants species separated (displayed at 50nm intervals) by shape

filtering for the wetland invasive plant species.

43

The wavelengths identified as most useful for separating invasive species by the
shape filter vary by species (Figure 8). This is critical as classification and processing
techniques and/or choice of wavelengths might require evaluation based on the species on
interest. Further, the biophysical properties influencing reflectance become valuable as
background and plant canopy architecture vary by species thus potentially improving
identification. Polygonum pensylvanicum has the most separation around 605nm.
Compared to the other invasives this wavelength has low separation value using the
shape filter method. Nymphaea odorata is most separable from the other species in the
visible and chlorOphyll domain (350-600nm) likely due to the plant canopy architecture
possessing large (24cm), round leaves that float on the water surface. Using the shape
filter technique, the invasive Lythrum salicaria had between five (minimum at 2000nm)
and 20 (maximum at 700nm) species distinguished.

The results from applying the shape filtering technique confirm that information
provided by increased spectral data does increase abilities to distinguish plants of interest.
The wavelength domains of utility vary by species therefore data reduction and
wavelength selection methods need to consider evaluating species of interest and their
particular absorption/reflectance features. The concept of spectral libraries and
classification techniques based on shape filtering is promising for distinguishing invasive
species. In the wetland ecosystem where this study was conducted, even very similar

spectra were able to be filtered.

4.5 Characterizing Infestation

Complications due to shifts in aircraft flight path and multiple flight lines, lack of
simultaneous ground control, and high landscape heterogeneity required a multi-step
approach in order to obtain the most accurate map possible for the species of interest. A
series of classification runs were executed. The more advanced SAM approach using the
signature library developed from field-collected spectra resulted in very poor
classification accuracy results. SAM classifications carried out on irnage-derived training
data had satisfactory results for Phragmites, but the Typha delineation was poor.
Therefore, a simplified approach was executed to identify Typha in the form of
unsupervised algorithms. Qualitative inspections confirmed that the unsupervised
classification accurately depicted landscape structure and biological communities within
the study area. Figure 9 illustrates a subset of the 40-class unsupervised ISODATA
classification. Apparent is landscape structure with deep pools, biological communities,
forest regions, road networks, and tributary systems. Using GPS ground data, Typha

polygons were created and merged within the classification.

45

.r
6:; l

 

$2
‘14

light line gap

Figure 9. Unsupervised classification depicts landscape structure and man-made features

for adjacent buffer around Route 31 which bisects the wetland complex.

The SAM Phragmites classification was fused with the Typha unsupervised
classification for the optimal product. A total of twelve polygons were used to assess the
accuracy of the combined classification. Overall accuracies for each individual polygon
ranged from 39%-81% (Table 5). Figure 10 illustrates two Typha latifolia assessment
polygons located on opposite ends of the study area. P6 shows a high accuracy (73%)
while P3 shows a poor overall accuracy (39%). P6 was surrounded by lower strata
canopy (submergents/emergents) while the P3 assessment polygon area tended to have
similar strata canopy (rushes/sedges) which likely contributed to the confusion and
ultimately the misclassification. The overall accuracy (weighted average by # pixels) was
70%. Considering the level of detail, overlapping signatures, and imagery complications,
this is a fair (~industry standard is 85%) overall accuracy. When referencing the field-

level signatures, Typha tends to have common reflectance values within the wavelength

46

domains covered by the imagery. Phragmites, on the other hand, has a relatively unique

spike in the NIR.

Table 5. Accuracy assessment for the T ypha and Phragmites classes.

 

 

 

 

 

 

 

 

 

 

 

Polv # Qv_e_rgll # Pixles flgnTtfi
l 66.97% 445 Typha
2 80.00% 70 Typha
3 39.45% 844 Typha
4 62.67% 150 Typha
5 63.03% 238 Typha
6 73.33% 896 Typha
7 81.40% 328 Phrag
8 55.45% 101 Phrag
9 46.39% 761 Phrag
10 67.08% 814 Phrag
l l 76.56% 3720 Phrag
12 79.63% 2573 Phrag
Weighted average 70% 10665 Typha & Phrag_

 

47

 

 

Figure 10. Example accuracy assessment points for T ypha latifolia. P3 shows a poorly

captured patch whereas P6 illustrates a patch captured very accurately with the imagery.

Heterogeneous environments are challenging to classify and the desired level of
classification detail and pixel size often influence methodology, objectives, and overall
goals. Becker et al. (2007) evaluated classification algorithms and the optimal spatial
resolutions for mapping coastal wetlands. The research concluded that strategic
wavelengths and spatial resolutions of a few meters meter or under were required for
accurate maps. Li et al. (2004) found that increasing levels of landscape heterogeneity
required higher spatial and spectral resolutions. Large homogeneous stands have less
variance compared to smaller mixed stands and, essentially, large homogenous areas are
easier to map. Smaller mixed stands have greater spectral overlap and a larger amount of
species contributing toward the spectral signature. However, in this study no relationship

48

between patch size and accuracy was found. Both Typha and Phragmites were tested
individually and collectively, and no significant relationship was found using 1m size
pixels. Two related points can be extrapolated from the accuracy results. The first being
that the fine resolution imagery used in this study was not near the threshold in which
spatial resolution becomes a limiting factor. Also, these results suggest that the size of the
patch, also representing composition and variance, does not singularly determine

accuracy or abilities to delineate species.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

>. 70%
E 60% 4’!-
3 50% v]
8 40% ~ - ----- —
< 30% .

20% - —

10%

03/3 —r--—+~---l—~--i-~- I—--—t——-i--—+-—-I-- -—+—-—+’—l—--I~-—+~—+—i-‘-I

0 1000 2000 3000 4000
Size of assessment patch

 

Figure 11. No relationship between accuracy and patch size was found. Combined T ypha

and Phragmites ground control and overall accuracy for each polygon.

Assessing infestation required an expert system framework that depicted a variety
of disturbances and scale. Using a knowledge-based evaluation system, each section was
assessed for level of human disturbance and given a rating between 1 to 5 representing

low to high disturbance and/or activities. Human activities and/or disturbances were

49

measured in terms land use pressures and landscape modifications (e.g., agriculture,
urban), intensity of managements and disturbances (continuous crop fields, dense
impervious road network, etc...), and hydrological modification (river channelization,
dredging). Following the river continuum framework, adjacency and spatial dependence
was incorporated into the knowledge-based ratings. If a section had intense agriculture
along its boundary, the disturbance rating will increase. Further, if a pristine section was
adjacent to a highly disturbed section, this would cause an increase. Figure 12 displays
transects across the study area. Transects were aligned N—S in order to complement the

imagery and locate the Rte 31 bridge into one transect.

 

Figure 12. Illustration of infestation transects used to assess T ypha and Phragmites

distribution and develop the disturbance metric.

The expert-system values are displayed in table six. Transects start upstream,
which is the eastern most portion of the classification in this case. Transects one and two
received the highest hydrogeomorph rating because the river in these transects has
undergone substantial channelization altering the hydrological flow of the main branch of

50

the Muskegon. As the river resumes natural meandering toward transect three,
hydrogeomorph disturbance ratings decrease. As the river branches into its delta-like
pattern in transect five ratings receive their lowest scores. At transect eleven substantial
modification is present as the Route 31 bridge bisects the wetland complex. A channel
was dredged running parallel to the bridge creating an artificial, slow moving branch.
Dredging and miniature dikes for several rail bridges and old transport channels are

present closer toward the City of Muskegon resulting in increased modification ratings.

Table 6. Disturbance ratings for the study area east to west.

 

 

Transect LU/LAN DP HGMM Rating
1 5 5 5
2 2 5 3.5
3 2 2 2
4 1 1 1
5 2 2 2
6 3 1 2
7 3 1 2
8 3 2 2.5
9 3 2 2.5
10 2 2 2

11 5 5 5
12 4 4 4
13 4 4 4
14 5 5 5
15 5 5 5

 

 

 

 

 

 

The human disturbances in transects one and two are relatively moderate. As the
wetland complex resides in essentially a valley, adjacency disturbances play a role. The
northern buffer of the study area remains relatively consistent. Clearings for power lines

aiming westward starting in transect six and seven increased the rating. Transect eleven

51

and the western most transect have the highest amounts of impervious surfaces causing
the highest possible rating.

For replication, quantitative approaches are desired. Therefore, a fractal metric
was derived to quantify the shape of the Muskegon River and its tributaries to assess
changes in hydrogeomoprh across the study area. Figure 13 displays the area-weighted
fractal metric for the water class against Phragmites australis (A) and Typha latifolia (B).
Although sample size was limited to the number of transects (15), moderately strong (.5
and .44 respectively) relationships are evident with decreasing PLAND values as the
shape complexity of the water class increases. This supports the interpretation that as the
river becomes channelized and impacted by human activities infestation is higher, and as
the river flow reflects the delta-like hydrology of the complex, Phragmites PLAND

decreases.

52

 

 

 

 

1.2 1.25 1.3 1.35 1.4 1.45

 

 

 

 

14

12 R2=0.437
10 - ~

 

 

own-moo

1.2 1.25 1.3 1.35 1.4 1.45

 

 

B.

 

Figure 13. Moderately strong relationships exist between Phragmites australis (A)

and Typha latifolia (B) PLAND metric and the shape complexity of the water class.

The spatial pattern of T ypha and Phragmites are closely correlated in addition to
the percentage of areal coverage these two invaders occupy. Their area-weighted fractal
was strongly positive correlated (.69) across the study site and these two species appear
to mimic the other species pattern and distribution. The shape complexity, as measured
by area-weighted fractal, of T ypha was highly correlated (.76) to the aggregation index of
Phragmites patches across the landscape. As T ypha patch shape becomes more complex,

Phragmites tends to become more aggregated and decrease in complexity within the

53

landscape. Figure 14 displays the assessment metric showing the relationship between

disturbance and infestation. The quantitative fractal metric indicates that the hydrological

modifications are significantly related to pattern of infestation.

 

 

 

 

 

 

 

 

 

A
\ \ I
.50 \\ I \ ’I 3.5
:E \ 1' \j 3'
\‘ I ‘

_ \ I, A U
0 \ a‘
> \ I n
.3 t, l S
c \ I g
9 “ I, 3
a . ~ . a

8 \ /\ ' " ‘4 g
a \v I <
_ \ 1
\ I
\ I
\ I .—
o
8 s
_l
1 2 3 4 5 6 7 8 9 10 11 12 13 14
Disturbance Metric Phragmites PLAND Typha PLAND

Figure 14. displays the relationship between infestation level and disturbance ratings for

the transects contiguously covering the study site.

54

Chapter 5: CONCLUSIONS

The research carried out in this investigation was focused on distinguishing
wetland invasive plant species of interest and mapping infestation. Field-level
hyperspectral reflectance factor was characterized to identify unique wavelength regions
for dominant invasive plant species within the study area. Processing techniques were
evaluated and showed varying degrees of utility. Relatively accurate maps for Typha
latifolia and Phragmites australis were developed from airborne hyperspectral imagery.
The mouth and western portion of the study site were identified as having higher levels of
infestation. These infested regions coincide with relatively higher human land use

intensities and slow-moving hydrological conditions.

Specific and general conclusions are as follows:

1. Characterizing absorption/reflectance features for particular species of interest
should be considered when determining processing and classification techniques.
When an overarching goal is to identify and map species of interest, such as
Phragmites australis, techniques should consider absorption/reflectance features
of that species response, since wavelengths of utility might vary. ‘Universally’
applied processing techniques are not always applicable as shown in this research.
Shape filtering, which is a relatively straightforward approach, is one useful
method to isolate wavelengths and/or biophysical attributes of interest for
individual species. Useful wavelengths, according to shape-filtering approaches,

can be species specific as shown in this work.

55

2. Continuum removal largely decreased separation abilities. The understory and
previous plant growth are critical for identifying wetland species and play a role
in creating unique signatures. Removing these features ultimately removed
biophysical characteristics that aided in identification. Unless an absorption
feature is identified as having a particular utility in aiding identification,
continuum removal decreases separation and should not be used. Plant canopy
structure had mixed results in terms of whether continuum removal disregarded
structure, yet it did emphasize the utility of background signal.

3. In this study, basic pattern metrics were most applicable. While more complex
metrics exist, describing infestation is really a factor of basic area, percentages
covered, and distribution of those species across the study site. Straightforward
metrics that describe percentage of landscape and aggregation measures
illustrated infestation which was a primary goal. The infestation maps developed
through this work can help assist management plans and efforts to control
invasive species in the future. The framework developed in this work should be
applicable to many regions and many species.

4. The concept of signature libraries (or building a database of signatures to feed
into imagery for classification) for species-level mapping in wetland
environments with current technologies is very challenging. In environments
where species reflectance ranges widely and landscape heterogeneity is high,
classification accuracies will rarely achieve levels above the industry standard
threshold of 85%. Ground control and collection must occur simultaneously (i.e.,

radiometric calibration) to overpass for successful implementation. The highest

56

quality imagery is also necessary as slight reflectance differences between species
are often smaller than signal to noise ratios of imagery. If species have unique
spectral characteristics, if those characteristics are distinguishable in the
wavelength captured by the imagery, if simultaneous ground control is collected,
if the landscape has lower levels of dominant species and low levels of mixed
pixels, and imagery is of the highest caliber, it will improve mapping accuracies
and signature library concepts might be capable of discriminating covers.

. The highest quality data is required. Issues related to multiple flight lines and no
simultaneous ground control created calibration problems and limited imagery
accuracy. Field work in diverse, wetland environments is extremely challenging
and obtaining a large ground control dataset is extremely useful. In this study,
several trips to the field were required and more data was always useful.

. As eluded to in conclusion four and five, operational monitoring using
hyperspectral wetland imagery for invasives monitoring is challenging. Data
quality issues and landscape conditions produce ‘thresholds of utility ’ where the
framework developed here is and is not useful. Operational monitoring might be
possible with lower diversity in simpler landscapes where large, dominant patches
of spectrally unique species are the plants of interest.

. Future work needs to explore fusing advanced hyperspectral remote sensing
technologies with other sensors such as LiDAR (providing topographic
information) and RADAR (providing moisture information). Hyperspectral
sensing provides advanced canopy and biological community (e.g., species)

information. By combining these data with geomorphology and hydroperiod data

57

increased assessment information can possibly be provided. Wetlands health
assessment approaches that can link all these technologies together and link with

traditional field-based measures (e.g., HGM) is the next step in this field.

58

REFERENCES

Artigas, F. and J. Yang. 2006. Spectral Discrimination of Marsh Vegetation Types in the
New Jersey Meadowlands, USA. Wetlands, 26, 271-277.

Ayala, R. and M Menenti. 2002. Evaluation methodology for classification process of
digital images. IEEE Geosciecne and Remote Sensing Symposium, IGARSS,
3363-3365.

Becker, B., Lusch, D. and J. Qi. 2005. Identifying optimal spectral bands from in situ
measurements of great Lakes coastal wetlands using second derivative analysis.
Remote Sensing of Environment, 97, 23 8-248.

Becker, B.L., D.P. Lusch and J. Qi, 2007. A classification-based assessment of the
optimal spectral and spatial resolutions of coastal wetland imagery. Remote
Sensing and Environment, 108, l 1 1-120.

Clark, R. and T. Roush. 1984. Reflectance spectroscopy: quantitative analysis techniques
for remote sensing applications. Journal of Geophysical Research, 89, 6329-6340.

Cochrane, M. 2000. Using vegetation reflectance variability for species level
classification of hyperspectral data. International Journal of Remote Sensing, 21,
2075-2087.

Cowardin, L.M., Carter, V., Golet, F. and E. LaRoe. 1979. Classification of Wetlands and
Deepwater Habitats of the United States. FWS/OBS-79/31. US Fish and Wildlife
Service, Washington, DC.

Dahl, T. 1990. Wetlands losses in the United States 1780's to 1980's. US. Department
of the Interior, Fish and Wildlife Service, Washington, DC. 13 pp.

Dahl, T. 2000. Status and trends of wetlands in the conterminous United States 1986 to
1987. US Department of the Interior. Fish and Wildlife Service, Washington,
DC. 82pp.

Danielson, T. 2001. Methods for Evaluating Wetland Condition: Introduction to Wetland

Biological Assessment. Office of Water, US. Environmental Protection Agency,
Washington, DC, USA.EPA 822-R-01-007a.

Danson, F. 1995. Developments in the remote sensing of forest canopy structure. In F.
Danson and S. Plummer (Eds), Advances in Environment Remote Sensing (pp.53-
69). John Wiley and Sons.

Environmental Monitoring and Assessment Program. 2002. Research Strategy. Office of
Research and Development. US EPA. EPA 620/R-02/002.

59

Fortin, M., Olson, R., Ferson, S., Iverson, F., Hunsaker, 0, Edwards, G., Levine, D.,
Butera, K. and V. Klemas. 2000. Issues related to the detection of boundaries.
Landscape Ecology, 15, 453-466.

Grace, J ., and R. Wetzel. 1982. Variations in grth and reproduction within
populations of two rhizomatous plant species: Typha latifolia and Typha
angustifolia. 0ecologia, 53, 258-263.

Gustafson, E. 1998. Quantifying landscape spatial pattern: what is the state of the art?
Ecosystems, 1, 143—156.

Hardisky, M., Gross, M., Klemas, V. 1986. Remote sensing of coastal wetlands.
BioScience, 36, 453-460.

Hargis, C., J. Bissonette, and J. David. 1998. The behavior of landscape metrics
commonly used in the study of habitat fragmentation. Landscape Ecology, 13, 67-
186.

Hirano, A., Madden, M. and R. Welch. 2003. Hyperspectral image data for mapping
wetland vegetation. Wetlands, 23, 436-448.

Huang, Z., Turner, B., Dury, S., Wallis, I. and W. Foley. 2004. Estimating foliage
nitrogen concentration fiom HYMAP data using continuum removal analysis.
Remote Sensing of Environment, 93, 18-29.

Keams, F., N. Kelly, J. Carter, and V. Resh. 2005. A method for the use of landscape
metrics in freshwater research and management. Landscape Ecology, 20, 113—
1 25.

Kokaly, R., Desoain, D., Clark, R. and K. Livo. 2003. Mapping vegetation in
Yellowstone national Park using spectral feature analysis of AVIRIS data. Remote
Sensing of Environment, 84, 437-456.

Kruse, A., Lakeoff, B., Boardman, B., Heidebrecht, B., Shapiro, P., A. Goetz. 1993. The
Spectral Image Processing System [SIPS)-Interactive visualization and analysis of
imaging spectrometer data. Remote Sensing of Environment, 44, 145-163.

Laba, M., Tsai, F.,Ogurcak, D., Smith, S. and M. Richmond. 2005. Field determination of
optimal dates for the discrimination of invasive wetland plant species using
derivative spectral analysis. Photogrammetric Engineering and Remote Sensing,
71, 603-611.

Li, H. and J. Wu. 2004. Use and misuse of landscape indices. Landscape Ecology,
19, 389—399.

Liu, A. and G. Cameron. 2001. Analysis of landscape patterns in coastal wetlands of

60

Galveston Bay, Texas. Landscape Ecology, 16, 581-595.

Liu, A. and G. Cameron. 2001. Analysis of landscape patterns in coastal wetlands of
Galveston Bay, Texas. Landscape Ecology, 16, 581—595.

Lopez, R., C. Davis, and M. Fennessy. 2002. Ecological relationships between landscape
change and plant guilds in depressional wetlands. Landscape Ecology, 00, 1—14.

Mack, M. and CM. D’Antonio. 2003. Exotic grasses alter controls over soil nitrogen
dynamics in a Hawaiian woodland. Ecological Applications, 13, 154-166.

Marks, M., Lapin, B., and J. Randall. 1994. Phragmites australis (P. communis): threats,
management and monitoring. Natural Areas Journal, 14, 285 - 294

McGarigal, K. and B. Marks. 1995. FRAGSTATS: Spatial Pattern Analysis Program for
Quantifying Landscape Structure. USDA Forest Service, Pacific Northwest
Research Station, Portland, OR, USA. Gen. Tech. Report PNW-GTR-351.

Mitsch, W. and J. Gosselink. 2000. Wetlands, third edition. John Wiley & Sons, Inc.,
New York, NY, USA.

Naiman, R., Pastor, H., Johnston, C. 1988. The potential importance of boundaries to
fluvial ecosystems. J. N. Am. Benthol. Soc., 7, 289-306.

National Research Council. 1995. Wetlands: Characteristics and Boundaries. National
Academy Press. Washington, DC. 328pp.

National Research Council. 2000. Ecological Indicators for the Nation. National
Academy Press. Washington, DC. 180 pp.

Ozesmi, S., Bauer, M. 2002. Satellite remote sensing of wetlands. Wetlands Ecology and
Management, 10, 381-402.

Paulsen, S. G., Larsen, D. P. Kaufinann, P. R. Whitier, T. R. Baker, J. R. Peck, D. V.
McGue, J. Hughes, R. M. McMullen, D. Stevens, D. Stoddard, J. L. Lazorchak, J.
Kinney, W. Selle, A. R. and Hjort R. 1991. Environmental Monitoring and
Assessment Program: EMAP-surface waters monitoring and research strategy-
fiscal year 1991. EPA/600/3-91/022. United States Environmental Protection
Agency. Washington, D. C.

Pijanowski, B., Shellito, B., Pithadia, S. and K. Alexandris. 2002. Forecasting and
assessing the impact of urban sprawl in coastal watersheds along eastern Lake

Michigan. Lakes and Reservoirs: Research and Management, 7, 271-285.

Pimentel, D., Zuniga, R. and D. Morrison. 2005. Update on the environmental and
economic costs associated with alien-invasive species in the United States.

61

Ecological Economics, 52, 273-288.

Roman, C., Niering, W. and Warren, RS. 1984. Salt marsh vegetation change in response
to tidal restriction. Environmental Management, 8, 141-150

Rosso, P., Ustin, S., and A. Hastings. 2005. Mapping marshland vegetation of San
Francisco Bay, Califonria, using hyperspectral data. International Journal of
Remote Sensing, 26, 5169-5191.

Saltonstall, K. 2002. Kryptic invasion by non-native genotypes of the common reed,
Phragmites australis, into North America. Proc. Nat. Acad. Sci., 99, 2445-2449.

Schaepman-Strub, G., Schaepmen, M., Painter, T., Dangel, S. and J. Martonchik. 2006.
Reflectance quantities in optical remote sensing — definitions and case studies.
Remote Sensing of Environment, 103, 27-42.

Schmidt, K. and A. Skidmore. 2003. Spectral discrimination of vegetation types in a
coastal wetland. Remote Sensing of Environment, 85, 92-108.

Sohn, Y., E. Moran, and F. Gurri. 1999. Deforestation in north-central Yucatan (1985-
1995): Mapping secondary succession of forest and agricultural land use in Sotuta
using the cosine of the angle concept. Photogrammetric Engineering & Remote
Sensing, 65, 947-958.

South, 8., Qi, J ., Lusch, D. 2004. Optimal classification methods for mapping agricultural
tillage practices. Remote Sensing of Environment, 91, 90-97.

Thenkabail, P., Enclona, B., Ashton, M. and F. van der Meer. 2004. Accuracy
assessments of hyperspectral waveband performance for vegetation analysis
applications. Remote Sensing of Environment, 91, 354-376.

Thiemann, S. and H. Kaufinann. 2002. Lake water quality monitoring using hyperspectral
airborne data- a semiempirical multisensor and multitemporal approach for the
Mecklenburg Lake District, Germany. Remote Sensing of Environment, 81, 228-
237.

Thomas, V., Treitz, P. Jelinski, D., Miller, J ., Lafleur, P. and J. McCaughey. 2002. Image
classification of a northern peatland complex using spectral and plant community
data. Remote Sensing of Environment, 84, 83-99.

Tiner, R. 1999. Wetland indicators: a guide to wetland identification, classification, and
mapping. Boca Raton. Lewis Publishers.

Torbick, N., P. Lawrence, and K. Czajkowski. 2006. Application and assessment of an

expert system for wetland identification in northwest Ohio, USA. Journal of
Environmental Management and Restoration, 2, 26-37.

62

Torbick, N., Qi, J ., Rolof‘f, G., Stevenson, J. 2006. Investigating impacts of land-use land
cover change on wetlands in the Muskegon River watershed, Michigan, USA.
Wetlands, 26, 1103-1 1 13.

Underwood, E., Ustin, S. and D. DiPietro. 2003. Mapping nonnative plants using
hyperspectral imagery. Remote Sensing of Environment, 86, 150-161.

US. Department of Agriculture, Natural Resource Conservation Service. (2006). The
PLANTS Database (http://plants.usda. gov. 1 August 2006). National Plant Data
Center, Baton Rouge, LA 70874-4490 USA.

US. Department of Agriculture, Natural Resource Conservation Service. (2007). The
NISIC Database (1 August 2007). National Invasive Species Information Center,
Baton Rouge, LA 70874-4490 USA.

U. S. Environmental Protection Agency. 1996. Strategic Plan for the Office of Research
and Development. EPA/600/R-96/059. Washington, D. C.

U. S. Environmental Protection Agency. 1998. Guidelines for Ecological Risk
Assessment. EPA/630/R-95/002F. United States Environmental Protection
Agency. Washington, D. C.

Ustin, S., Roberts, D., Gamon, J ., Asner, G. and O. Green. 2004. Using imaging

spectroscopy to study ecosystem processes and properties. BioScience, 54, 523-
534.

Uzarski, D., Burton, T., Cooper, M., Ingram, J ., and S. Tirnmermans. 2005. Fish Habitat Use
Within and Across Wetland Classes in Coastal Wetlands of the Five Great Lakes:
Development of a Fish Based Index of Biotic Integrity. Journal of Great Lakes
Research, 31, 171-187.

van der Meer, F. 2004. Analysis of spectral absorption features in hyperspectral imagery.
International Journal of Applied Earth Observation, 5, 5-68.

Wiens, J. 1989. Spatial scaling in ecology. Functional Ecology, 3, 385-397.

Wilcox, D. and J. E. Meeker. 1991. Disturbance effects on aquatic vegetation in
regulated and unregulated lakes in northern Minnesota. Canadian Journal of
Botany, 69, 1542-1551.

Wilcox, D. 2004. Implications of hydrologic variability on the succession of plants in
Great Lakes wetlands. Aquatic Ecosystem Health & Management, 7, 223-231.

Windham, L. 2001. Comparison of biomass production and decomposition between
Phragmites australis (common reed) and Spartina patens (salt hay) in brackish
tidal marsh of New Jersey. Wetlands, 21, 179-188

63

Windham , L., and J. Ehrenfeld. 2003. Conflicting effects and the net impact of a plant
invasion on nitrogen cycling processes within brackish tidal marshes. Ecol.
Applic., 13, 883-896.

64

,4 .9. ,r.-. .. .. - ., ; .3 'I'?‘a: ; - In" ."-::' "-v "urn-“'- .~.. ' -~

 

lllilllllllllilllllllllllllllll‘llllllllll
3 1293 02956 3875‘