University of Southern California Dissertations and Theses

Mixed forest image classification of paper birch: using AVIRIS bandwidths ranging from 530 to 745 nm

(USC Thesis Other)

Mixed forest image classification of paper birch: using AVIRIS bandwidths ranging from 530 to 745 nm

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

Mixed Forest Image Classification of Paper Birch:
Using AVIRIS Bandwidths Ranging from 530 to 745 nm

by

Kyle Benjamin Uhler

A Thesis Presented to the
Faculty of the USC Graduate School
University of Southern California
In Partial Fulfillment of the
Requirement for the Degree
Master of Science
(Geographic Information Sciences and Technology)

August, 2018

Copyright © 2018 by Kyle Uhler

To my wife Meghan

i

Table of Contents
List of Figures ................................................................................................................................ iii
List of Tables .................................................................................................................................. v
Acknowledgements ........................................................................................................................ vi
List of Abbreviations .................................................................................................................... vii
Abstract ........................................................................................................................................ viii
Chapter 1 Introduction .................................................................................................................... 1
1.1. Laurentian Mixed Forest Habitat ........................................................................................2
1.1.1. Paper Birch.................................................................................................................3
1.1.2. Habitat Change and Conservation .............................................................................5
1.2. Remote Sensing ..................................................................................................................7
1.2.1. Imagery Classification ...............................................................................................7
1.3. Methods Overview and Study Area ....................................................................................7
Chapter 2 Related Work................................................................................................................ 11
2.1. Background .......................................................................................................................11
2.1.1. Vegetation and Atmospheric Interactions with Electromagnetism .........................12
2.2. Image Classification..........................................................................................................15
2.1.1. AVIRIS Data Classification of Vegetation ..............................................................19
Chapter 3 Methods ........................................................................................................................ 21
3.1. Data Acquisition and Preparation .....................................................................................21
3.1.1. Data Acquisition ......................................................................................................21
3.1.2. Bandwidth Selection ................................................................................................25
ii

3.1.3 Data Correction .........................................................................................................26
3.1.3. Data Integration .......................................................................................................30
3.2. Image Classification..........................................................................................................32
3.3. Spatial Analysis ................................................................................................................34
Chapter 4 Results .......................................................................................................................... 36
4.1. Data Inspection & Correction Results ..............................................................................38
4.1.1. Image Georeferencing ..............................................................................................38
4.1.2. Atmospheric Scattering and Spatial Autocorrelation ..............................................39
4.2. Classification Results ........................................................................................................40
4.2.1. Training Site 4.3 m Resolution ................................................................................40
4.2.2. Validation Site 16.5 m Resolution ...........................................................................44
Chapter 5 Discussion and Conclusions ......................................................................................... 50
5.1. Methods Strengths ............................................................................................................50
5.2. Methods Limitation ...........................................................................................................52
5.3. Further Research ...............................................................................................................55
5.4. Applications of Research ..................................................................................................56
References ..................................................................................................................................... 58
Appendix A: Dataset Metadata ..................................................................................................... 63

iii

List of Figures
Figure 1 Paper birch leaf and seed morphologies (USDA) ............................................................ 4
Figure 2 Dead and dying paper birch trees on the North Shore (MN DNR) .................................. 5
Figure 3 Laurentian mixed forest seen in purple ............................................................................ 8
Figure 4 General methods schematic .............................................................................................. 9
Figure 5 Study area overview ....................................................................................................... 10
Figure 6 Common spectral reflectance of green vegetation (Gaussman 1977) ............................ 14
Figure 7 AVIRIS 4.3 m pixel resolution extent and ground truth collection locations ................ 24
Figure 8 Georeferenced GCP location example ........................................................................... 28
Figure 9 Paper birch locations with buffer (left image), then canopy area
rasterized (right image). .................................................................................................... 32
Figure 10 Linear unmixing model classification of the 16.5 m AVIRIS image with validation
forest plots ......................................................................................................................... 37
Figure 11 Intersection of dirt road and mowed tree line used as GCP. ........................................ 38
Figure 12 Screen procedure results for bands 15 – 43 in AVIRIS dataset, solid green line
(4.3 m dataset), dash blue line (16.5 m dataset) ................................................................ 40
Figure 13 Residual map of 4.3 m AVIRIS image ......................................................................... 42
Figure 14 Insert location in the 4.3 m training data site classification results and canopy
coverage ............................................................................................................................ 43
Figure 15 Residual map of 16.5 m AVIRIS image. ...................................................................... 46
Figure 16 Validation site (16.5 m) with average coverage of forest stand inventory plots. ......... 47
Figure 17 Insert map of the 16.5 m validation site displays two forest stand plots with
differencing levels of pixel coverage. ............................................................................... 48
iv

Figure 18 Image interpretation shows false return of paper birch spectral signature within
wetland on Westside of Paccini Lake. .............................................................................. 49
Figure 19 Typical paper birch ground truth sample...................................................................... 53

v

List of Tables
Table 1 SGCN study from the MBS ............................................................................................... 6
Table 2 Data requirements ............................................................................................................ 25
Table 3 Programs & uses .............................................................................................................. 25
Table 4 Resampled images GCPs used for the 4.3 m/ 16.5 m scenes .......................................... 29
Table 5 Output reference parameters ............................................................................................ 30
Table 6 Reclassified ranges categories ......................................................................................... 34
Table 7 Training data site (4.3 m) image analysis results. ............................................................ 41
Table 8 Confusion matrix for the 4.3 validation canopies with statistics ..................................... 44
Table 9 Validation data site (16.5 m) image analysis results. ...................................................... 45
Table 10 AVIRIS flight data ......................................................................................................... 63

vi

Acknowledgements
I wish to acknowledge the patience and guidance that Professor Dr. Laura Loyola and my thesis
advisor Dr. Elisabeth Sedano gave throughout this thesis. I am grateful to all my peers who
helped edit this report and facilitate discussion about this domain. Thank you to my thesis
committee members Dr. Andrew Marx and Dr. Steven Fleming for their input and advice. I want
to acknowledge my employer, the Minnesota Board of Water and Soil Resources, for facilitating
project-based learning concerning remote sensing. Finally, I would like to thank Jeff Walsh from
Peregrine Aerospatial LLC for assisting with this project’s direction and for insights into
commercial remote sensing applications.

vii

List of Abbreviations
AVIRIS Airborne Visible/ Infrared Imaging Spectrometer
CORS Continually operating reference station
DBH Diameter at breast height
DN Digital numbers
EM Electromagnetic
FSA Farm Service Agency
GCPs Ground control points
GIS Geographic information systems
LMF Laurentian mixed forest
MBS Minnesota Biological Survey
MN DNR Minnesota Department of Natural Resources
NAIP National Agriculture Imagery Program
NIR Near-infrared
NSFC North Shore Forest Collaborative
NSH North Shore Highlands
RGF Raster group files
RS Remote sensing
SGCN Specie of greatest conservational need

viii

Abstract
Paper birch (Betula papyrifera) is a dominant specie within Northern Minnesota’s Laurentian
mixed forest. Though these trees are common place, paper birch populations have been in
decline for the past couple decades along Lake Superior. Due to the reduced replacement rate of
this specie, organizations are implementing management strategies to promote healthy forests.
This thesis investigates remote sensing techniques to predict paper birch locations remotely. The
thesis tests individual specie level spectral signature effectiveness to classify community level
data of the same informational group. The project uses hyperspectral Airborne Visible/ Infrared
Imaging Spectrometer (AVIRIS) flights data for its imaging platform. Two spatial resolution
scenes were classified for this project. A 4.3 m pixel resolution image was used for defining
spectral signatures based on ground truth data. Then a lower resolution 16.5 m image was used to
apply the produced paper birch signature as a classifier to test functionality of these methods on
known paper birch communities. Pixels were used as the final classification unit. A linear
unmixing soft classification was utilized to produce percent signature contained within pixels.
The classification resulted in ~93% of forest plots containing some pixels with 95% similar
spectral signatures to paper birch. Total classified coverage of validation forest plots was low
with only ~30% covered with 75% similar signature or greater. The long-term objective for this
project is to automate specie identification to monitor trees. Further research is needed to
streamline classification and refine procedures, yet current findings can help forest managers and
conservationists identify priority sites to both map current specie distribution and implement
restoration activities.

1
Chapter 1 Introduction
Paper Birch (Betula papyrifera) populations on Minnesota’s North Shore Highlands (NSH) are
on the decline. Many factors influence the reduction of paper birch populations. Some
contributing variables that can directly affect paper birch are deer browsing, insect infestations,
soil compaction, and regional warming (NSFC 2017). Historic stands of white pine and cedars
were extensively harvested in the 1800s, altering the landscape into a successional habitat. In
those altered areas pioneer species like paper birch have thrived. Climate models used by the
Northern Institute of Applied Climate Sciences for forest adaptability assessments predict winter
temperatures will continue to increase. This increase in temperature could reduce the regional
snowpack, which makes up a majority of its annual precipitation. This spells a disaster for the
cold climate species of the North Shore like paper birch. Management of paper birch populations
depends upon identification of those populations. This paper demonstrates methods to identify
birch populations in mixed forest habitats using remotely sensed data.
This project is in response to the loss of paper birch on the landscape. This research
explores informational group classifications of paper birch in the Laurentian mixed forest (LMF)
habitats of the North Shore. It incorporates remote sensing techniques to better equip forest
managers to monitor paper birch remotely along the North Shore and use in the management
decision process. This project does not seek to model future vegetation, but to assist in
monitoring of past and current paper birch stands to give managers another tool to track plant
community change. The goals are to produce methods that maximize correct classification of
paper birch within the mixed forest and reduce time taken to model that plant community.

2
This study could help future researchers establish best practices for remote detection of
paper birch and plants with similar plant morphologies. Utilizing findings form this project could
help future monitoring efforts. This chapter introduces paper birch and the habitat it occupies. It
discusses habitat change and ongoing management of the landscape. The chapter then introduces
remote sensing applications for environmental research. In conclusion, an overview of this
project’s study area is described.
1.1. Laurentian Mixed Forest Habitat
Northern Minnesota is dominated by the LMF. This habitat type surrounds much of the
Great Lakes region. The LMF covers some 23 million acres of Minnesota and is divided into 26
subsections based on soil, climate, and vegetation (MN DNR 2006). The habitat is dominated by
conifer and deciduous broadleaf forests, swamps, and bogs with a variety of niche communities.
This landscape is marred by historic logging, land development, and more recently tourism.
Many factors have their effect on what this environment looks like today.
The LMF is a habitat that is influenced by historical geologic underpinnings. The
Laurentian name refers to the ancient continent and a precursor to North America. The bedrock
of the North Shore is situated along a mid-continental rift of this ancient landscape. The 1.1
billion-year-old basalt rift is greatly exposed, defining the province. The more recent geologic
events of glaciation have ground and eroded soils within the region. Due to these glacial events
this area is characterized by shallow coarse aggregate and topography with great relief.
The prevalent driver of climate to the near shore environments is Lake Superior. This is a
unique habitat that receives most of its annual precipitation in the form of snow. The lake acts as
a climate moderator (MNDNR 2006), both a source of heat in the winter and as an air

3
conditioner in the summer. The moderate cool annual climate gives rise to cold-weather
vegetative communities.
Both the presettlement and current vegetation communities are forest dominated. These
forests are fire-dependent communities that rely on episodic successional pioneer species. After
disturbance events stands of conifers are replaced by species like aspen, birch, and maple. This
project seeks to identify paper birch communities, which are deciduous-broadleaf species.
1.1.1. Paper Birch
Paper birch trees are easily identifiable specie on the northern landscape. They have
iconic bark that is thin, white, and peeling in nature. Paper birch trees are known to grow up to
30 m in height but are usually 20 m or shorter (Marshall 1785). Trees are most commonly found
with a singular main trunk but can be seen growing with multiple trunks. Multiple trunked trees
will usually be defined by shorter canopy heights. Paper birch trees are shallow-rooted with few
roots deeper than 60 cm below the soil surface (Safford 1990). A key factor in this research is the
average crown width. The North Dakota Tree Handbook states canopies at maturity can be
observed from 6 to 12 m in width. These birch crowns are comprised of simple leaves that are 2-
4 in. long with narrow 1 in. stems and can be seen in Figure 1.

4

Paper birch form nearly pure pioneer communities on disturbed sites and are rare in late
successional forests (Uchytill 1991). In the LMF paper birch can be found in mixed established
communities due to the distribution of the tiny seeds on the wind. Compared to other tree species
birch are relatively short-lived species. Most birch trees will not grow past the age of 125 years
old (Day 1981). Paper birch have a fast grow rate, which is useful for long-term revegetation and
soil stabilization of severely disturbed sites (Uchytill 1991).
A multitude of pests can affect the health of birch trees, yet few pests are culpable for
much of birch harm. A major insect pest of birch trees is the bronze birch borer. Individual trees
compromised by environmental stresses are susceptible to infestation of the bronze birch bore
(Conklin 1969). Over-browsing can knock back birch populations too. Although there are many
browsing animals that affect birch, white-tailed deer eat substantial amounts of the trees’ foliage
later in the season, and heavy browsing by deer and moose populations have been shown to
prevent or setback forest regeneration (Irwin 1985).
Figure 1 Paper birch leaf and seed morphologies (USDA)

5
1.1.2. Habitat Change and Conservation
Minnesota’s NSH is uniquely situated along many environmental gradients. The area is at
the north edge of many warm-climate deciduous tree species and the southern limit of many
cold-weather species. Due to this fact, slight changes to climate gradients can alter the vegetative
community makeup. The changing forest can increase soil erosion, river turbidity, and sediment
deposition that affects the environment (Malgorzata 2010). Increasing habitat change throughout
the region has led conservation groups to develop restoration plans to combat the
transformations.
The North Shore Restoration Project is one program planned to manage 39,000 acres of
the Superior National Forest that is part of the NSH. It was implemented as a response to climate
change. It assessed the area and found that 80% of the birch stands are of mature age and dying.
This would be of no concern if the trees were being replaced by other age stands, but that is not
the case. The historically forested habitat is not regenerating and is being replaced by grasses as
seen in Figure 2.

Figure 2 Dead and dying paper birch trees on the North Shore (MN DNR)
The Superior National Forest assessment predicts an unsustainable future habitat due to
the lack of tree stand regeneration. Many projects are collaborating on reforesting efforts. The

6
North Shore Forest Collaborative (NSFC) is a promising initiative to promote healthy forests.
The NSFC strives to bring together a multitude of partners to restore a productive ecosystem and
they promote native forests. In 2015 alone, this project planted many thousands of trees, over
some 1000 acres of the North Shore. Attempts to reforest this region are underway.
The Minnesota Biological Survey (MBS) has been identifying species in greatest
conservational need (SGCN) for the NSH. The MBS found 84 wildlife SGCN that utilize this
region. Table 1 displays nine problems that contribute to specie decline and the percentage that
the problem influences the SGCN. Habitat loss and degradation lead the list, showing the
importance of managing these habitats appropriately. Monitoring this habitat can aid
management in restoring populations of greatest concern. Projects like these need a baseline
monitoring protocol to gauge progress of these habitat and implement management.
Table 1 SGCN study from the MBS

This project is motivated by a desire to help the restoration and management of
Minnesota’s NSH and other similarly endangered forests. Furthermore, this project’s scope
covers a wide-range of topics within the domain of remote sensing. Finally, this project brings
together areas of personal interest.

7
1.2. Remote Sensing
The domain of remote sensing (RS) is simplistically described as obtaining information
from a distance. The science of RS has been changing rapidly since the launch of Earth-orbiting
satellites in the 1970s. NASA’s Landsat mission’s objective was to make repeated Earth
observations through time. These missions gave researchers multispectral data over large extents
of the globe. The images from Landsat and other RS platforms allowed for image analysis on a
large scale. As RS advanced, so did geographic information systems (GIS). The ability to pair
geospatial material with RS increased the potential of geographic analysis.
1.2.1. Imagery Classification
Cover class description is an objective for researchers in the RS domain. Research and
development in classification methods have improved since early aerial photography. The early
classifications were accomplished through photo interpretation by trained specialists. Computer
algorithms now have the power to identify objects and even classify the unique characteristics of
features’ makeup. Classification of imagery pixels in this study is used to identify paper birch on
the landscape. Pixels contain values from multiple bands. Spectral difference between pixels can
be compared and assigned to informational groupings. This thesis investigates spectral
differences between pixels of known objects to define classes. Supervised classification of the
landscape was used to form a final project product. Supervised classifications require inputs to
identify pixels related to ground features. The spectral signature of paper birch is extracted with
use of an algorithm.
1.3. Methods Overview and Study Area
Image classification of paper birch within the Laurentian mixed forest habitat represents a
challenging scenario. Though paper birch represents a sizable portion of the individuals on the

8
landscape the trees’ morphological traits lead to mixed pixel outputs. Discrimination of this plant
due to this fact can be problematic.
This project investigates areas of the NSH subsection of the LMF. This subsection of the
LMF is adjacent to Lake Superior, only 25 miles inland at its furthest. The study area presents a
mixture of swamps, bogs, lakes, conifer, and hardwood forests. Figure 3 shows the extent of the
LMF in Minnesota (purple) and the 1.4-million-acre extent of the NSH subsection (orange).

The overall objective of this project was to develop individual level signature data and
apply those signatures to a community level scene to predict paper birch occurrence. This project
tested supervised classification using training data collected in the field. The training data
selection involved quality controls to ensure quality data to model the spectral class. The
supervised image classification section describes utilizing a linear unmixing technique to assess
Figure 3 Laurentian mixed forest seen in purple
and the North Shore Highlands subsection in orange (MN DNR)

9
percent similar pixel signature. This technique is a soft classification used in mixed pixel
situations.
In this thesis, two AVIRIS flight scenes were acquired over NE Minnesota. The
classification of both AVIRIS resolutions were tested for accuracy with ground truth data
collected and open source forest stand inventory data. Subpixel scale was used to represent
canopy locations that were based on percent spectral response. The data processing phase is
ongoing throughout this project and includes clipping data into workable datasets,
georeferencing, translating files into useable formats, and data exploration. Figure 4 displays a
general schematic of this project.

The study area within the NSH was chosen based on public access, historic survey
records, and availability of imagery. Areas between Two Harbors and Lutsen, MN exhibit those
factors. AVIRIS images located on the North Shore within 20 miles of one another were used for
classification. Figure 5 shows the locations of both images in relationship to one another and in
Minnesota. These areas have surveyed cover classes delineated in a MN Forest Inventory layer.
AVIRIS
Image
Data
Processing
Training
Data
Image
Classification
MN
Forest
Inventory
Classification
Analysis
Figure 4 General methods schematic

10

The following chapter examines background research in RS classification and specie
identification using hyperspectral AVIRIS datasets. This study utilizes prior investigations in
hyperspectral species classification to shape methodologies and those studies are discussed in the
literature review. Many aspects of remote sensing sciences were used to complete this project
and methods applied are described in Chapter 3. Then Chapter 4 presents results of both data
processing and image classification. Finally, Chapter 5 discusses recommendations for further
and future work as well as applications of this project.

Figure 5 Study area overview

11
Chapter 2 Related Work
Traditionally, remote sensing has used imagery to identify vegetative areas form non-vegetative
areas of the Earth’s surface. Many applications have utilized community-based observations
because individual plants can be difficult to verify with low satellite resolutions. The vegetative
areas imaged can be subdivided into classes of plants with similar reflections called spectral
classes. It is important to understand the foundations of remote sensing prior to completing this
project. The following sections describe related work in remote sensing that gives a background
to species level identification. It is crucial to understand the foundational research into remote
sensing systems and the electromagnetic (EM) interactions with our planet’s surface. Following
sections also investigate hyperspectral data uses in vegetation classification. Additionally,
knowledge for this project will utilize conventional classification methodologies of data and will
be discussed in this chapter.
2.1. Background
Remote sensing has many useful applications such as identifying land cover and more
specifically specie identification. Land cover classifications involve classifying large extents vs.
specie delineations happening on smaller scales e.g. stand level in forest ecology. Full cover
classification of specie level data is extremely limited. Using RS applications over more
traditional cover surveys can save investigators time in the field.
Digital imaging platforms coinciding with GIS has allowed raster pixels to be interpreted
as ground features. The individual pixels contain information beyond the georeferenced location
and area covered. Imaged pixels are commonly formatted as digital numbers (DN), which is the
measured intensity value. DN are usually a conversion from radiance captured at the sensor. DN

12
can be formatted as different binary digits or bits. The number of bits collected will regulate the
radiometric resolution of the image. For example, an 8-bit data represents pixels defined between
0-255, whereas 16 bit allows for values range of 0-65536. Many raster images are formatted as
8-bits, because it represents decent diversity for a scene. Higher radiometric resolution allows
greater detection of small variation in ground reflectance. In observing the radiation emitted by
these objects classification can be obtained. This project will exploit the difference between
known objects’ EM relative emittance and predict occurrence.
2.1.1. Vegetation and Atmospheric Interactions with Electromagnetism
The EM spectrum is divided into regions base on historic uses within disciplines. The
EM spectrum ranges from < 0.03 nm to > 30 cm. Any radiation captured with RS technology
inevitably must travel through the Earth’s atmosphere. This radiation doesn’t move unimpeded
in the atmosphere. The atmosphere can scatter, refract, and absorb incoming EM energy.
Observed scattering of particles is greatest near the blue end of the spectrum. Refraction of light
by atmospheric interactions will depend on humidity and thickness of atmosphere at image
capture time. Absorption of energy happens by molecules like ozone and carbon dioxide within
the image scene. There are multiple bandwidths in the atmosphere that absorbs energy; water
vapor absorption is centered at approximately 0.94, 1.14, 1.38 and 1.88 μm, an absorption band
for oxygen is at 0.76 μm, and a carbon dioxide absorption band is near 2.08 μm (Gao 2009). A
researcher may avoid use of these absorption points to limit effects on datasets.
The removal of atmospheric effects is necessary when preparing RS data. Remote
sensing sensors on airborne platforms can minimize atmospheric effects by lowering flight
altitude. This will mitigate some atmospheric effects by reducing the amount of atmosphere
images are captured through. Remotely sensed areas with sparse vegetation can omit both

13
atmospheric and background noise when using indices that account for those effects (Giannico
2004). The atmospheric effects can increase as spatial resolution decreases; and spectral
resolution increases. The energy that is not affected by atmospheric distortions are called
atmospheric windows. These are the regions that allow light to return to the sensor for capture.
Many methods for atmospheric correction have been established, ranging from simple
calibrations to complex models. The Gao et al. (2009) experiment reviewed multiple approaches
to correct atmospheric effects for routine processing of imaging data. They found that radiative
transfer model approaches are sufficient and can be used for preprocessing of hyperspectral data.
Radiative transfer models are tools to represent both scattering and absorption of radiation. When
multi-temporal images are used for image classification, atmospheric calibration is mandatory
(Lu 2007).
Once the images are corrected for atmospheric manipulation surface vegetation light
interaction can be measured. The reflectance of any plant depends on the leaf structure,
pigments, and water content within those parts. Species with similar leaf structures will produce
similar spectral response (Hoffer 1969). The typical spectral response of green vegetation is
shown in Figure 5. The troughs represent wavelengths at which absorption by pigment, then by
water occur within the plant. The human-visible spectrum occupies wavelengths between 380-
720 nm. In Figure 6 one can see pigment absorption near 400 and 650 nm. The infrared area of
the EM spectrum is a much larger section than visible light and is important to the identification
of plants. The near-infrared (NIR) wavelengths are longer than those within the visible spectrum
(720 nm -1300 nm). The air-spaces in leaves lower cross-sectional structure slows the incoming

14
light. Larger leaves air spaces have longer wavelengths in the NIR (Abrams 1990). These
variations in NIR response can be used to define plant species within a scene.
Variation in spectral reflectance from tree canopy level can be influenced by location,
specie, leaf structure, leaf angle, water content, background, varied irradiance and pigment
concentration (Ollinger 2011). It is difficult to reproduce canopy reflectance; attempts have been
made by stacking leaves on top of each other from below as in a canopy spectral measurement
(Coops 2003). Spectral differences within tree canopies is known in the literature and is
significant with respects to position in the canopy (Danson 1995). Spectral variation in the
visible wavelengths are due to chlorophyll concentrations (Cochrane 2000). Plant senescence in
the autumn changes the overall levels of chlorophyll contain in the leaf. Gitelson et al. (1994)
found that leaves with high chlorophyll concentrations had low pigment variation. Maximum
deviations of pigment where observed near 550-560 and 700-710 nm. These ranges are most
sensitive to pigment concentrations (Blackburn 2007). This project uses variations in these
ranges to classify paper birch.
Figure 6 Common spectral reflectance of green vegetation (Gaussman 1977)

15
Distinguishing variables characterize a plant’s functional type in the environment. The
plant’s functional type is defined as the vegetative structure, biochemical physiology, and
phenology. These plant variables affect the spectral return of the plant. Spectral returns that have
been identified at the leaf scale in studies have also been applied generally at both the canopy
and landscape scales (Zarco-Tejada 2000). This study will use canopy level spectral response
due to the spatial resolution of the raster images obtained.
Phenology does vary between plant species, but expression will fluctuate based on
temperature, moisture, and photoperiod (Sayn-Wittgenstein 1978). Although there is season
variability photosynthesis is highly reliant on temperature and photoperiod (Reed et al. 1994).
Thus, single year observations can generalize based on phenological differences.
These phenotypic seasonal differences are exploited in this project. In general, many
remote sensing studies focus on coarse resolution data to delineate forest communities and have
not used individual trees as a classification unit (Key et al. 2001).
2.2. Image Classification
Imaging landscapes for remote sensing purposes has long interested researchers.
Classification of the landscapes can produce a wide range of accuracies because of scene
complexity, informational groupings, image acquisition, and classification approach. The
challenges of RS can be overcome with insight about the study areas cover. When beginning
analysis of an imaged scene most researchers have some prior knowledge of the locational cover.
Furthermore, when using prior surveyed data for a site, the researcher can select sites with
individual spectral groups or plant communities of interest. If historic surveys acquired were
sampled within similar communities as the research questions, investigators can maximize
accuracies of classification (Lauver 1997).

16
Utilizing prior knowledge about surface cover to inform a classifier is considered a
supervised classification. Supervised classifications lean on the researcher’s prior knowledge of
the image location to reduce the errors in classification. The prior site knowledge is selected as
training data to produce similar information groups. In using supervised classifications, the
selection of training data is a crucial step in a project work flow and can affect overall accuracy
in capturing the attended informational group. In an elevation-based vegetation classification
study, Gartzia (2013) found that the proper selection of training data was the most principal
factor in improving accuracies. The problems in selecting enough quality data in mixed
landscapes is that it is difficult and is complicated when coarse spatial resolution datasets are
used for classification. These factors lead to increased mixed pixels (Lu 2007).
The habitat preferences of paper birch are commonly associated with many northern
forest species that don’t necessarily create homogenous timber stands. This leads to the increased
likelihood of mixed pixel outputs depending on resolution. Commonly the number of mixed
pixels increases as resolution decreases (Crapper 1984). Mixed pixels act to average the
brightness of a scene by subdivision. Mixed pixels are not inherently bad, if the spectral
signatures averaged are of similar groups the average signature of the pixel will not be affected
much. When the pixel overlaps distinctly different spectral signatures within a pixel the average
can change greatly (Franklin 2002). The mixing of vegetation and soil reflectance is an example
of possible distinct spectra in a mixed pixel composition. The combined signature of the distinct
groups may not match any of the groups captured in the pixel area (Campbell 1987).
The spectral signature contains all materials present in the training pixels and can neglect
the influence of the mixed pixels. The characteristics of each pixel spectra then represents the

17
objects acquired in a hyperspectral image (Campbell 2011). This thesis discriminates spectral
classes within mixed forest vegetation from other information classes within the imagery.
Mixed forest habitats by nature are difficult to distinguish informational groups that
exists within the community. Early studies like Franklin (1994) found that out of 12 classes used
on Landsat TM multi-spectral imagery of mixed deciduous forest they achieved accuracies of
only 77%. That study was using the entire community types and showed the difficulty of mixed
pixel identification. Leckie et al. (2005) sub-divided informational groups to increase the
representation of the group and bypass intra-group variation by environmental conditions.
Establishing a robust training data sample will limit interspecies variation. Fassnacht et
al. (2016) reviewed prior work and established six criteria on which training data selection
should successfully be based: classes much match the research question(s); data selected is
representative of study location; spatial scale should match the scale of the question; assumptions
underlying the methodologies applied; observed errors should be known and discussed for
impact; and samples should be spatially independent. The training data for this project will use
cover type class categories established by the MN DNR to distinguish areas to select the best
training data locations. In general, a good representative dataset for every class is fundamental in
carrying out a supervised classification. (Lu 2007).
The classification of objects with transitional boundaries are inherently difficult to
represent spatially. Fuzzy classifications or soft classifications offer a different approach to
classifying objects. The soft classifications result in a degree of membership to each
informational group. Unlike soft classifiers, hard classifications use well defined groups for each
pixel location. In mixed pixel images traditional hard classifiers are unsuitable (Atkinson et al.
1997). Long-established classifiers such as the maximum likelihood classifier utilizes the

18
spectral bands to resolve what class each pixel is most likely to belong to. The goal of a soft
classifier is not to categorize each pixel into one group but have it as a sum of its components
(Atkinson et al. 1997).
Soft classifiers are intrinsically difficult to develop accurately, but the produced pixel
mixtures can lead to more accurate representation of land-cover estimations compared to hard
classifiers (Ichoko et al. 2009). Linear unmixing is one such method that produces fuzzy
memberships. These categories can be hardened based on input criteria. Linear unmixing
functions have been found quick and useful for image classification. Linear unmixing models
every pixel as a linear function of the classes input. At its simplest linear unmixing is the
minimum difference between the training data spectra and the unknown spectra. This is based on
all possible composition of the training data. The process of spectral unmixing involves breaking
down pixel spectra into subpixel component spectra, commonly referred to as endmembers.
Studies have applied linear unmixing models to identifying rock substrates, crop cover
estimates, and land cover estimates. In arriving at land cover estimates, Foody and Cox (1994)
assumed pixels were pure spectra that defined endmember spectra. They found a strong
correlation between actual and predicted percent cover with a 99 percent confidence. Although
all endmember categories tested highly, grasses had the most accurate predictions. They noted a
tendency to underestimate the percent cover of trees in some pixels because trees tend to not
cover pixels homogeneously. A concern in using linear unmixing models is that in the study area
there may not exist a pure class spectra pixel, depending on image resolution. Selecting
appropriate training data for endmember spectra can increase accuracies (Atkinson et al. 1997).

19
2.1.1. AVIRIS Data Classification of Vegetation
Many early studies spotlight the use of broad spectral bandwidths in the visible and NIR
for use in vegetation classification. Over time innovative technologies increased the number of
bands reducing the spectral ranges. Recent work has targeted the emphasis of using narrow-band
regions like the red edge to distinguish image elements (Blackburn 2007). Hyperspectral images
consist of a hundred or more narrow spectral bands. Hyperspectral imaging is a specialized
discipline within RS sciences. The spectral bandwidths are contiguous throughout the EM
spectrum. Hyperspectral data offers a unique view of forest vegetation. This study takes
advantage of a narrow band range of hyperspectral raster imagery and will apply classification
methods to identify plants in the scene.
The AVIRIS platform has been utilized for countless classification studies. This is an
airborne image acquisition platform. Data acquired form the AVIRIS optical sensor captures
wavelengths from 400 to 2500 nanometers. This spectrum is obtained through 224 spectral
bands. These data are collected on a NASA airplane. Lower altitude fights can capture greater
than 3 m resolution. The higher resolution images are collected on smaller extent strips.
Disadvantages of limited spatial extent is overcome by increased spectral resolution to define
objects on the landscape. The pixels at sub-five meters begin to be able to isolate spectral
signatures of individual tree canopies and can give a better distinction between objects. Imaging
spectrometers collect images as contiguous spectral bands like AVIRIS. Complete reflectance
spectrum can be derived from the wavelength region covered in every pixel (Goetz 1985).
Hyperspectral imagery can be used to generate spectral signatures of vegetation species. Studies
have used hyperspectral data to identify and map forest species and have achieved a certain
degree of success (Ruiliang 2009). In nature, everything has a unique spectral characteristic that

20
can be exploited to identify information about the object (Parker 1965). The assumption for this
project is that spectral differences between species are greater than the variation with the target
specie.
Prior work has utilized many different classifiers at a multitude of scales. The increased
spectral resolutions obtained by hyperspectral datasets when remote sensing in adjacent narrow
bands can produce elevated levels of autocorrelation (Blackburn 2007). Autocorrelated data is
redundant in nature and is of low importance when identifying spectral variation. This needs to
be addressed in methodologies that use hyperspectral data. Misclassification of hyperspectral
imagery can be a result of many different variables. Problems can occur due to image timing
creating shadowing. Also, areas with high relief can cause issues with classification (Mehner
2004). Individual plant species can vary on reflective signatures independently of pigment
(Blackburn 2007). So, classification evolves to use both the average and diversity of each
spectral group or class. The advent of hyperspectral imagery allows more stratified ranges to be
analyzed.

21
Chapter 3 Methods
Methodologies presented in this thesis work to classify paper birch’s likelihood of being
represented by any given pixel within the study areas scope. Methods used for this study can be
grouped into three general types of tasks: Data Acquisition and Preparation, Image
Classification, and Spatial Analysis. This chapter describes these tasks.
3.1. Data Acquisition and Preparation
Geospatial data and RS data are becoming more accessible in online open source formats.
Datasets obtained for this thesis project were all open-sourced or self-collected. Various
resolutions and extents could be obtained for this projects requirement.
3.1.1. Data Acquisition
This project relied on AVIRIS datasets to perform all subsequent work. Hyperspectral
datasets were acquired from NASA’s JPL AVIRIS platform for the study site. Data was
downloaded from NASA’s JPL AVIRIS data portal (https://aviris.jpl.nasa.gov/alt_locator/). All
files needed to be extracted with Zip7 multiple times to open the (.TAZ) compressed packaging.
Images were imported into TerrSet with use of the GDAL protocol. This procedure converted
ort.img to Idrisi native .rst format (RST Idrisi). The 224 band images could then be packaged as
raster group files (RGF). This format allows for multi- and hyperspectral image sets to be input
into a variety of protocols within TerrSet.
Two flight datasets were used for this project: f120930t01p00r11 (4.3 m resolution) and
f061002t01p00r11 (16.5 m resolution). The AVIRIS data utilized were captured on 30
September 2012 and 2 October 2006 respectively. The AVIRIS platform collects data in
continuous bandwidths ranging from 400 to 2500 nanometers (nm). AVIRIS uses a scanning

22
mirror as a “whisk broom” the scene over the 224 band detectors at the sensor. This project’s
methods only imported bandwidths or bands between 530 and 745 nm and bands 15 through 43.
These bands would cover the EM spectrum from green to near-infrared. Furthermore, it reduced
function processing time. There are multitudes of scanned regions on Minnesota’s North Shore.
The two flights chosen were due to similar time of year capture date, time of day, and azimuth
angle. The lower resolution scene also needed some level of accessibility for capturing training
data.
The classification analysis utilized previously identified forest community patches. Forest
stand data was needed to validate plant community type by percent coverage of classification
within the low resolution 16.5 m scene. Minnesota Department of Natural Resources (MN DNR)
Forest Stand Inventory was obtained through the Minnesota Geospatial Commons
(https://gisdata.mn.gov/dataset/biota-dnr-forest-stand-inventory). The dataset had all the
attributes needed for this project’s scope. The data is updated, and field checked by individual
administration areas throughout the state on an as-needed basis.
All datasets needed to be georeferenced to allow for data overlay and analysis. Imagery
from the Farm Service Agency (FSA) through the National Agriculture Imagery Program
(NAIP) was acquired to georeferenced the AVIRIS imagery. NAIP produces true color red,
green, and blue images with a near-infrared band. NAIP images covering the entire AVIRIS
images were downloaded through the USDA Geospatial Data Gateway
(https://gdg.sc.egov.usda.gov/GDGHome_DirectDownLoad.aspx). Minnesota’s Lake County
imagery was obtained for this project.
All self-collected field data were acquired between 10 and 24 September 2017. Ground
sampling was performed to establish training data locations within the high resolution 4.3 m

23
AVIRIS image. Ground truth data locations were selected based on proximity to trail access,
Forest Stand Inventory data, and Park Manager notes on specie distribution. Field observations
were noted on density of birch trees sampled and areas with homogenous stands to minimize
background signal. Areas with steep sloping terrain were avoided when sampling to reduce
terrain effects. Samples were distributed diversely through study images. Figure 7 shows ground
truth sample locations within the AVIRIS 4.3 m image. Trimble’s GeoExplorer 6000 series was
used to sample tree species. Pathfinder office was employed to create a data dictionary. The data
dictionary used generic point for vegetation format type. One could select for tree species of
concern or create comment notes about another feature collected. The Trimble unit employed
TerraSync software as a platform on which to collect GIS field data.

24

When performing field identification of tree species three reference sources were used:
Minnesota Flora: An Illustrated Guide to the Vascular Plants of Minnesota, Steve W. Chadde,
2013; Trees and Shrubs of Minnesota, 1
st
Ed. Welby R. Smith, 2008; Trees of Minnesota: Field
Guide, Stan Tekiela, 2002. Ground sampling utilized alive or recently dead trees with a diameter
at breast height (DBH) of equal or less than 20 cm. This would ensure that trees were in the
image scene during the acquisition date. Under these parameters, 57 validation tree locations
Figure 7 AVIRIS 4.3 m pixel resolution extent and ground truth collection locations

25
were identified. Table 2 summarizes the data used in this project. Data and layers were integrated
to perform classification and analysis. Table 3 show software needs and uses.
Table 2 Data requirements
Dataset Name Accuracy/
Precision
Source Collecti
on Date
Data
Size
Data
Format
Ground
truth
sample
N/A Only used
points
collected
under 4.3 m
accuracy
Self-
collected
9/10/17-
9/24/17
56.02
MB
Shp
MN
Forest
Stand
Inventor
y
MNDNR Forest Stand
Inventory
Community
based with
good
accuracy
MN
DNR
9/17/20
09
170
MB
Shp
AVIRIS f120930t01p00r11 and
f061002t01p00r11
4.3 m and
16.5 m
NASA-
JPL
9/30/12
and
10/2/06
3.67
GB
and
0.99
GB
Binary

FSA
NAIP
Imagery
ortho_1-
1_1n_s_mn031(75)_20
13_1
1m GSD FSA
through
NAIP
2013 3.83
GB
zip
Geotiff

3.1.2. Bandwidth Selection
In identification of plants remotely this project took advantage of the seasonal variation
between different species. Those differences are seen in the autumn as the leaf color changes due
Table 3 Programs & uses
Programs Uses
TerrSet Image classification, atmospheric correction
ArcGIS Spatial Analyst, figure creation
Pathfinder Data transfer, conversion, correction, and dictionary
TerraSync Data collection
Excel Plot spectra

26
to reducing chlorophyll concentrations. Not all bands can identify the change in spectra due to
leaf color change. A selective range of bands were exploited in this project to identify species
based on that difference. Band selection was based on part of the VIS and the “red edge “ of the
NIR regions. The EM spectrum within those regions is where chlorophyll is both absorbed and
reflects incoming radiation. The header file accompanying the downloaded AVIRIS image
dataset had band center points descriptions. When the header file was converted into a text file it
could be read. Only bands covering 530 to 745 nm were selected.
3.1.3 Data Correction
This project used a variety of methods to maintain high data quality standards. It was
important to the success of this thesis to georeference the image dataset and process the data for
noise. Addressing these factors is fundamental to a favorable outcome for the classifying
procedures tested. Data was corrected into the same orthorectified datum. Images used were
corrected geographically and atmospherically to allow for information extraction about ground
features. This allowed for the two different AVIRIS images to be compared with limited
variability. Preprocessing of atmospheric corrections was needed to accurately quantify ground
characteristics. TerreSet’s SCREEN procedure was utilized to rule out bandwidths that displayed
a high threshold of autocorrelation and low threshold due to particle scattering. This procedure
will find and eliminate bands that have significant atmospheric attenuation. Threshold 0.99 and
0.6 were used in this project. Bands 15-43 were grouped in a raster group file (RGF) and input
into the procedure.
Image classification required georeferencing, so pixels could be related to ground
features and would allow comparison of surveyed reference data. The images were corrected for
earth distortions with respect to shape. The raster data image can be manipulated based on

27
control points to correct the image to true orthometric. NASA’s JPL packages the AVIRIS data
as orthorectified, but for this project it is important to verify pixel position before analysis. Data
for this thesis was corrected in TerrSet using RESAMPLE. The nearest neighbor option was used
to resample the images. The nearest neighbor resample was chosen because it doesn’t change the
value of the input cells recorded at the sensor during image acquisition. NAIP tiled mosaic
images were used as the output reference file. This increased the number of pixels within the
small training data sample size. The NAIP images were clipped into ArcMap to the extent of the
4.3 m and 16.5 m images. This reduced working file size. The clipped extents were exported in
TIF format and imported into TerrSet with GDAL in RST format. Band 43 was used for the
input reference file in both image resolutions. The resample file specifications used the RGF of
Bands 15-43 and the output reference parameters were set to study area clipped. Once images
were georeferenced the data frame was registered to an appropriate datum. AVIRIS data was
projected in World Geodetic System 1984 and was reprojected into NAD 83 for this project. All
datasets used the reference coordinate system of UTM Zone 15 N NAD 83. Both field collected
vector data and AVIRIS raster data was validated against the continually operating reference
station (CORS) nearest to the study area. The NAIP imagery was tested against the Grand
Marais, MN CORS base station (GDMA). This location has a continuous position update every
second. Its ground position was: 47 44 54.75712 N, 090 20 28.47142 W and in NAD83.

28
Ground control points (GCP) were established to resample the AVIRIS images. The
AVIRIS and NAIP scenes were georeferenced based on easily identifiable ground features, such
as road crossings, field corners, and lake points. Figure 8 shows an example of ground feature
utilized to georeferenced the AVIRIS imagery. Rocks along the lake shore that were prominent
across historical images were used. Past glacial events removed much of the top soil within this
region exposing prominent bedrock outcroppings. Those outcroppings tend to remain static over
time and were used as GCPs. Lakeshores themselves change over time and were not used as
GCPs. The GCPs used were spread throughout the image capturing all four quadrants of the
image. The final selection of GCPs was based on trial and error. Points that produced the lowest
root mean squared value were kept.
The 4.3 m image established 12 GCPs to georeference the image. Readily identifiable
objects within both scenes were matched. From the 12, only GCPs with residual errors under 0.5
(image resolution (4.3)) were used. The AVIRIS 16.5 m image used 5 GCPs. Table 4 shows the
x and y locations used to register the AVIRIS images. The clipped output reference area can be
seen in Table 5. The AVIRIS images were clipped using ArcMap’s image analysis tool to a
smaller extent and saved as a .tif, which could be imported into TerrSet for use. The study area
Figure 8 Georeferenced GCP location example

29
extent was chosen based on MN Forest Stand Inventory polygons within the 16.5 m AVIRIS
scene and inside the NSH region. The clipped images served to reduce function processing time
and limit file size. The resampling type used was a nearest neighbor so to not alter original
values contained within the images. Furthermore, a first order linear map function was used to
reduce the time to process and reduce the final number of GCPs needed. The 16.5 m scene was
clipped before images were georeferenced.
Pixel Resolution/
ID
Input X Input Y Output X Output Y
4.3 / 1 622666.7988 5237020.9083 617885.5656 5237317.0642
4.3 / 2 622092.4844 5235131.7176 615934.8548 5237618.5533
4.3 / 3 623456.5614 5234542.7600 615547.6336 5236206.7594
4.3 / 6 622915.8458 5237848.4642 619633.4516 5237309.8924
16.5 / 1 596556.6298 5271691.7982 603985.7767 5283454.6238
16.5 / 3 593447.6751 5228298.3777 643381.8895 5265005.5165
16.5 / 4 602752.2926 5230394.3880 644763.9992 5274440.4801
16.5 / 5 600109.3798 5260353.2769 615848.3665 5282707.9713

Table 4 Resampled images GCPs used for the 4.3 m/ 16.5 m scenes

30
Pixel resolution 4.3 m 16.5 m
Number of columns 1022 785
Number of rows 1410 4305
Minimum X coordinate 620441.2 638528.8
Maximum X coordinate 624835.8 646094.2
Minimum Y coordinate 5233857.1 5265209.2
Maximum Y coordinate 5239920.1 5271638.0

3.1.3. Data Integration
Vector data collected in TerraSync was transferred to shapefile to pull the layer into
ArcMap for use in the classification of the raster image. Once downloaded the imaged flight data
from the AVIRIS platform had to be uncompressed from the .tar zipped file format. Once the file
is unzipped the metadata could be explored. The data needed a new header and extension file to
allow import of AVIRIS binary raster data into ArcMap.
All self-collected tree data was cleaned on return from the field. Only points collected
with positive identification and horizontal accuracies under pixel resolution were used in this
work. The collected tree data was split into training and validation sample points to avoid over
fitting of the classification model. This study used a simple split of points collected to assess
initial accuracies (70% training and 30% as validation). This was accomplished by creating a
new ascending number column joined to the selected paper birch data. A Random number set
generator was used to select the 70% training data vs. the 30% validation data. Surveyed Forest
Inventory was assessed to ground truth species present at validation locations. These cross-
Table 5 Output reference parameters

31
sections of birch habitat aligned with MNDNR forest inventory birch habitat. Areas within the
16.5 m image were delineated to validate present or absences of vegetation community type
stated in that layer. Areas within the 16.5 m stated as “Paper Birch” were bisected to record
species present. Only polygons with paper birch communities were used to perform image
analysis.
The MN Forest Stand Inventory layer was processed to reduce the size of the working
layer. Attributes were selected in Lake County, 2006 or later acquisition, cover type of paper
birch, specie type of paper birch, DBH category of 4 or greater, and areas within clipped 16.5 m
scene.
The North Dakota Tree Handbook states that paper birch canopies at maturity can be
observed from 6 to 12 m. A 9 m average canopy width was assumed and buffered to create
polygons to establish areas of training data pixels. For each birch tree location, a 4.5 m buffer
was used to account for the 9 m diameter canopy coverage to establish training pixel area. The
buffer of 4.5 m was applied to the 57 validation tree locations. Then the buffered tree canopy
area was rasterized. The produced raster was compared to the linear hyperunmix classification
coverage for analysis. Figure 9 shows an inset area of the 4.3 m scene with buffers applied, then
buffered areas rasterized.

32

3.2. Image Classification
This project grouped pixels in an image by similar spectral signatures. Using the ground
truth samples at the 4.3 m site, paper birch signatures were selected for areas with birch present.
Supervised classification utilized training points collected in the field. The supervised
classifications were performed across the entire 4.3 m raster image. The supervised class
produced served to find the optimal spectral signature to identify birch on the landscape. This is
the simplification or generalization of the true complexity of the natural scene.
Pure signatures of paper birch at this resolution were expected to be rare and most pixels
to exhibit a mixture of flora. Buffered polygons corresponding to the 70% training data was
imported as a vectored polygon into TerrSet. The training data polygon was used as the input to
base pixel selection in TerrSet’s HYPERUNMIX hyperspectral image classifier. All 4.3 m band
Figure 9 Paper birch locations with buffer (left image), then canopy area rasterized (right image).

33
scenes were grouped as an RGF and input into the classification. Other readily identifiable
habitats were selected in TerrSet’s MAKESIG to reduce residual areas of scene.
Image interpretation of community type was performed and utilized for this experiment.
Easily identifiable informational groups were chosen to develop a paper birch spectral signature.
A minimum of 100 pixels was needed to produce each in the selected informational group
polygons. Selected groupings included: evergreens, upland grasses, wetland/ marsh, open water,
sparse vegetation and shadows. These signatures were grouped with the produced paper birch
vector signature and was run in the HYPERUNMIX function. The produced paper birch
signature was used against the 16.5 m image. Only a clipped extent of the 16.5 m was classified
for this project based on output references.
The linear unmixing model assumes that pixels can be modeled as a linear function. The
linear function requires other classes to project classes against one another as a linear function.
Harrison et al. (1991) provide a simplistic calculation of the linear unmixing model. The
multispectral dataset in n layers (26 bands), with y endmember types (7 classes), x is the
observed reflectance (n *1), and f is the unknown endmember proportions (y *1). The goal is to
calculate f(x):
x=M*f+e
M is the endmember matrix (n*y), and e is recorded noise. Calculate f by standard least
squares fit:
(x – M*f)*(x – M*f)
In this system one cannot have more classes than bands. HYPERUNMIX performs these
calculations for each pixel to classify endmember proportions.

34
3.3. Spatial Analysis
To perform spatial analysis on both the 4.3 m and 16.5 m AVIRIS scene the validation
polygons needed to be rasterized. This allowed pixels to be directly compared. The 30 % of
ground truth data collected was input through the zonal statistics tool in ArcMap. This produced
area of pixels the birch trees occupied. The same was done for the selected MNDNR forest stand
inventory polygons. Reclassification of the produced classified raster images was performed to
group by percent of signature into ranges. Table 6 shows reclassified ranges in TerrSet.

1 0-0.75
2 0.75-0.85
3 0.85-0.95
4 0.95-1.1

Accuracy analysis for the classified AVIRIS images assess area of classifications
coverage compared to validation areas. A zonal statistic was again performed to produce pixel
coverage. Validation trees were also averaged for distance from 0.95 signature classification.
Validation results of the test polygons on both scene resolutions were based on whether a 95%
signature similarity was classified within the test polygon.
A confusion matrix was established for the 4.3 m validation site. This matrix compared
predicted classes with > 95 % similar signature as the training data to < 95 % signatures.
Stratified random points were sampled throughout the image. 50 points were randomly sampled
in both areas above and below 95 % similar signature. These points were the compared to known
Table 6 Reclassified ranges categories

35
reference data to complete the confusion matrix. Confusion matrix statistics were calculated
based on the completed matrix.

36
Chapter 4 Results
This chapter outlines the results of the linear unmixing classifier to predict known paper birch
locations. Once the spectral signatures were developed based on individual tree canopies the
focus of this project was to overlay the classifications and compare the resulting spatial
coverages. The results produced categories representing percent spectral signature similarities
and the corresponding coverages of the MN DNR forest stand inventory layer. Paper birch
classifications were based on both individual tree spectra captured from hyperspectral data and
field collected vector data. The methods were used to find outcomes when using individual tree
spectral signatures to predict known community level specie occurrence. This project
hypothesized these methods would produce elevated levels (90%) of classified pixel coverage of
known validation polygons based on literature review. The following sections of this chapter
give insight into produced results.
After all the methods were applied the linear unmixing classification produced low
coverage on analysis. Both coverage of the training data 4.3 m site and the 16.5 m validation site
experienced low coverages results ranging roughly from 7% - 30%. Validation results of the test
polygons were based on whether a 95% signature similarity was classified within the test
polygon. This led to higher levels of validation polygons identified as classifying the
informational group. This approach to polygon validation identified 44% of the individual tree
locations and an assumed higher rate of 93% of community level polygons identified. These
results show the ability of the methods to produce generalized coverage maps of birch trees.
Processing and analysis required substantial interactions from users and further
automation is needed to expedite these methods for use. These results can only be interpreted
through statistical analysis. Direct association between pixel results and ground features are not

37
assumed by the spatial results, but contributions of individual paper birch are assumed by pixel
mixtures produced. The resulting linear unmixing model classifier produced is shown in Figure
10. This map is of the 16.5 m resolution scene clipped to the working extent. The map shows the
coverage of the 16.5 m classification and canopy validation areas.

Figure 10 Linear unmixing model classification of the 16.5 m AVIRIS image with
validation forest plots

38
4.1. Data Inspection & Correction Results
AVIRIS imagery, NAIP imagery, and validation vector layers were needed to predict
paper birch coverage maps. All data for this project would require exploration and correction
when necessary.
4.1.1. Image Georeferencing
Alignment of AVIRIS data (4.3 m and 16.5 m pixel resolution) was fundamental to
assure classified pixels represented ground truth data. The AVIRIS scenes were georeferenced to
NAIP imagery. Ideal GCPs would have easily identifiable locations i.e. (road crossings,
structures, and monuments) within the image. The nature of this study area was remote.
Furthermore, this led to sparse numbers of permanent and well-defined GCP locations. Figure 11
show a typical GCP used at the 4.3 m training data site.

Figure 11 Intersection of dirt road and mowed tree line used as GCP.

39
Ground control points were established and described in section (3.1.2.). Out of the 12
points identified in the 4.3 m scene, 4 points were used (1, 2, 3, and 6). The set of GCPs had a
residual error of less than 0.5*(4.3 m) and the Root-Mean Square (RMS) of this set was
0.836196. The RMS is an estimate of the average error within the points that have been selected.
The RMS does not average the error of the entire image. To achieve an RMS of equal to or less
than half the scenes spatial resolution, GCPs with high residual errors were not selected for the
transformation.
The 16.5 m scene used GCPs 1, 3, 4, and 5. These GCPs had residual errors of less than
0.5*(16.5 m) and an RMS of 0.498495. Table 5 in section (3.1.2.) showed the GCPs and output
parameters used for both 4.3 m and 16.5 m images. The distribution of points for the resample
were chosen based on the lowest produced RMS. The images were then georeferenced into
geographic coordinate system UTM 15N, and datum North American Datum1983. Only bands
18 through 43 were resampled.
4.1.2. Atmospheric Scattering and Spatial Autocorrelation
Image classification can be altered due to atmospheric distortion; thus, it was important to
process the data for possible distortion. The screen procedures were used to remove bandwidths
with autocorrelation above 0.99 and lower than 0.6 due to atmospheric scattering. The screen
function was used on bands within the EM spectral range in question (bands 15-43). The
function resulted in 26 bands kept and 3 bands (bands 15- 17) removed for autocorrelation. This
procedure was run on only bands 18-43 of the 16.5 m AVIRIS because 15-17 were removed.
Results (see, Figure 12) found no bands were removed due to scattering, but 3 bands were
removed on the 4.3 m scene due to autocorrelation.

40
Both scenes images were acquired under clear conditions with no clouds or haze. Though
both scenes had a lack of clouds during image acquisition atmospheric distortions may still crop
up and are recovered through the autocorrelation testing performed. Those distortions can be
caused by particulate matter in the atmosphere.

4.2. Classification Results
Classification was split into two major groupings: the 4.3 m training site and the 16.5 m
validation site. Both resolution images were classified and assessed for coverage of known tree
locations.
4.2.1. Training Site 4.3 m Resolution
Data selection played an integral part in the project’s methodology. The training data
selection process produced 189 valid paper birch locations within the 4.3 m resolution image.
These 189-birch tree locations were used to develop signature files and validation data. This was
after selecting all trees with a DBH of greater or equal to 20 cm and GPS accuracies under pixel
resolution. A simple split of training and validation data was performed. Out of the 189-tree
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1.00
15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
Screen threshold
Bandwidths
Figure 12 Screen procedure results for bands 15 – 43 in AVIRIS dataset, solid green line (4.3 m
dataset), dash blue line (16.5 m dataset)

41
sampled 132 trees were used for training data vector locations, which equals ~ 70% of total birch
samples. These were the points used to establish a supervised classification of paper birch and
create a spectral signature group file. The other ~30% or 57 paper birch locations were used for
validation points. The total number of pixels in the buffered validation area was 3338.
The linear unmixing model returned pixels with percent similar signature as the training
dataset. The classification was reclassed to see areas of stratified pixels. Pixels with a match of
less than 75% similar signature were not investigated in this project.
The initial 4.3 m scene classification identified 25 out of 57 (44%) validation trees as a
95% signature match. Zonal statistics of the 4.3 m image produced 29.1% canopy coverage with
a signature match of 75% or greater. Lower classified coverages were produced as the percent
signature match increased. The 4.3 m scene had low percent coverage seen in Table 7.

The linear unmixing classification produces a residual map seen in Figure 13. The higher
the value displayed in the residual map the less likely the pixels can be produced by the training
data. The residual map shows the NW corner of the unmix classification had a lower signature
match than some of the central regions of the image. The NW corner was not accessed during
ground truth data collection and could be a reason for the results displayed.

Signature match (%) Number of pixel Percent coverage
>75 971 29.1
>85 673 20.2
>95 428 12.8
Table 7 Training data site (4.3 m) image analysis results.

42

The 4.3 m linear unmixing classifciation produced three categories of spectral similarities
and each category was compared to pixel coverage of the 9 m diameter paper birch canopy
buffer (Figure 14).
Figure 13 Residual map of 4.3 m AVIRIS image

43

Figure 14 Insert location in the 4.3 m training data site classification
results and canopy coverage

Only percent polygon coverages were used in this project result due to only paper birch
locations sampled. Analysis was exclusively performed on pixels within paper birch canopies.
Furthermore, the actual pixels classified as paper birch inside the validation buffer should have
been 100%. Table 8 below shows a confusion matrix and statistics on accuracies. It shows a non-
significant kappa value of 0.13. This is repressed as a low agreement of predicted classification
compared to the referenced ground truth data.

44

Actual Cover Class
> 95 % < 95 % Total
Predicted Cover
Class
> 95 % 10 2 12
< 95 % 44 44 88
Total 54 46 100

Overall accuracy = 54 %
Error of omission = 81.5 %
Error of commission = 16.7 %
Producers accuracy = 18.5 %
User accuracy = 83.3 %
Kappa = 0.13

4.2.2. Validation Site 16.5 m Resolution
Validation polygons were obtained from the MN DNR with similar proximity to Lake
Superior as the training site and similar time of year image acquisition date. There were 14 forest
stand inventory plots used as validation polygons of paper birch communities. These validation
polygons covered an area of 128.1 ha. The developed spectral signatures were applied to the
community level polygons to validate previously sampled field data.
The 4.3 m paper birch spectral signature was applied to the 16.5 m scene and validated 13
out of 14 (93%) forest plots as having some picture elements with 95% similar signature. Though
a high percent of the forest plots were identified as having similar spectral signatures as paper
birch, the coverage of the validation plots was low as seen in Table 9. The total number of pixels
used was 5015 for the 16.5 m image. The overall forest plot coverage of the 16.5 m classification
was 30.1% with a signature match 75% or more.
Table 8 Confusion matrix for the 4.3 validation canopies with statistics

45

The produced residual map shown in Figure 15 is of the paper birch signature to emulate
similar pixel signatures. The low values represented in black are pixels with a good ability to
match those pixels and higher values in grey are a poor pixel match to represent that signature.
The map illustrates low ability to characterize cover classes: open water, spare vegetation,
evergreens, and upland grasses. This was an expected result. The resulting percent coverage was
mapped by mean area covered from low-high and shown in Figure 16. Then an insert map is
shown in Figure 17 that illustrates two validation polygons with differencing levels of cover
classification.
Table 9 Validation data site (16.5 m) image analysis results.
Signature match (%) Number of 16.5(m) pixel Percent coverage
> 75 1510 30.1
> 85 877 17.5
> 95 365 7.3

46

Figure 15 Residual map of 16.5 m AVIRIS image.

47

Figure 16 Validation site (16.5 m) with average coverage of forest stand inventory plots.

48

False returns were seen throughout the 16.5 m image by photo interpretation. Paper birch
signature was seen in pixels ranging from roads to wetlands. The occurrence of false pixel
classification was low and assumed to be similar to the 4.3 m classification due to the confusion
matrix. This needed to be noted and will need to be investigated further. Figure 18 shows paper
birch classification signature in a wetland West of Paccini Lake.
Figure 17 Insert map of the 16.5 m validation site displays two forest stand plots with
differencing levels of pixel coverage.

49

Figure 18 Image interpretation shows false return of paper birch spectral signature within
wetland on Westside of Paccini Lake.

50
Chapter 5 Discussion and Conclusions
Traditionally, large scale species delineation demands intensive field work. This includes
ancillary data collection and visual estimation. That process can be costly and time-consuming
due to accessibility (Lee et al. 1996). Remote sensing delivers economical means to discriminate
and estimate the physical properties of species. Due to the losses in paper birch population
regeneration it is important to utilize every reasonable measure to understand and monitor this
problem. Mapping specie distribution, quality, and quantity are critical tasks for management of
forests. Furthermore, it is necessary to regularly update spatial information about the extent and
quality of vegetation to manage these ecosystems effectively (He et al. 2005).
This project had both strengths and limitations. This chapter assesses the results of the
work and applications. The results should be viewed as exploratory and accuracies of methods
could be improved with further research that is described in this chapter.
5.1. Method ’s Strengths
This study’s methods resulted in overall low accuracies for the attempted identification.
These methodologies have not been applied to such a specific task before and results going into
analysis were unknown. The input considered in this project is an example of specific sets of
circumstances. The classification was produced to give the best likelihood of accurately
classifying paper birch. Though the produced accuracies were lower than sought there was much
to takeaway form this project.
The resulting paper birch signatures produced classifications of >95% membership that
on average were within 6 m of ground truth samples. Also, the user accuracy was calculated at
83 %. This shows how often the predicted class will be present on the ground. These methods

51
provide forest manager or researcher the ability to verify general paper birch locations. The
resulting data is usable to update current MN Forest Stand Inventory data.
Complex topography and convoluted mosaic of plants can affect sampling of field data.
The study are had areas inaccessible due to complex topography, but for standardizing spectra
responses flat terrain was needed. Avoidance of steep terrain was easy due to the flat nature of
the 4.3m sample sites. These methods were also able to exclude autocorrelation of pixels when
screened. This was a quick check for data quality. No scattering was observed in screening for
atmospheric interactions. This may be due to low altitude image acquisition flight path and no
reported haze on acquisition date. The preparation of data for this project was sound and could
setup a future project with different classification methodologies.
This study used limited training data to produce paper birch signatures. The process for
finding paper birch in the higher resolution AVIRIS image was random. If forest management
knew of paper birch locations prior to field work, they could quickly obtain GPS locations and
create a signature group. The other informational grouping used for classification were simple to
identify intentionally. This makes for rapid characterization of the imaged scene. Reduced
datasets used for this experiment increased turnaround time for forest managers too. The
Gitelson (1994) experiment showed that several functions of reflectance were directly
proportional to chlorophyll concentrations. Therefore, a reduced spectral resolution was used to
investigate the signature created by chlorophyll. This smaller dataset reduced file size and then
processing time for this project. This is an advantage to quickly producing a cover classification.
This project provided a learning opportunity to study the many facets of remote sensing.
It added variables that together were complicit in low accuracies. This shows that future work

52
needs to do more to limit the scope of paper birch classification to get a better handle on what
variables cause the most variation in the results.
5.2. Method ’s Limitations
Classification error can arise from several sources including: orbital position, clock error,
atmospheric affects, receiver noise, and multipath error (Campbell 2011). Limitations must be
considered as part of the cost and benefit analysis that all remote sensing specialists must
acknowledge. This project focused on limited variables to test methodologies. The reality is that
this classification methodology is an extreme simplification of the natural phenomena.
Limitations to this project are presented by both a short research window and level of data that
can be obtained.
The methods used in this project obtained lower than sought classification goals for paper
birch on the North Shore. Many factors have influenced the results of this project. The forest
stand inventory was not homogeneous in paper birch. This dataset identifies the plots of
dominate species and densities of that species. The forest areas sampled in the 4.3 m site
exhibited high species diversity with few homogenous groupings of paper birch. Terrestrial
plants will occur in large saturation compared to other community groupings (Elgadi 2010).
Largest samples of paper birch training data utilized were dense stand samples. This may
underestimate the true variability of paper birch on the NSH, due to spatial autocorrelation.
Furthermore, autocorrelation of trees was inevitable because paper birch can arise from asexual
parental rhizomes. This leads to pockets of paper birch groves.
This project utilized both reprojecting and resampling procedures. Further research into
AVIRIS images used showed that a simple rotation of the image was all that was needed and not
a reprojection. The resample procedure added error into this project considering the NAIP

53
images were used as the output standard. This increased pixel numbers within the canopy area
which increased variation of a “pure spectra”. These functions were used because of the
availability of the products within the Idrisi TerrSet software. They were not necessary for this
project and introduced error into ultimate results.
Crown density was not reviewed in this research and could introduce background noise
into signal. Figure 19 illustrates the typical paper birch canopy seen at the 4.3 m AVIRIS site. It
shows maple trees in the understory and a loose canopy morphology of paper birch trees.

Figure 19 Typical paper birch ground truth sample.
Tree selection was important to the overall success of this project. In researching the
historic aerial photos of the 4.3 m scene primitive roads were identified as access points to gather
data throughout the study area. The actual nature of the ground sampling was not as accessible.
Trees selected may not represent the true scene variation due to accessibility issues. Also, the
extent of the study area and the number of specific class used will affect the results of a
classification (Woodcock 1987). Sparse number of samples collected were usable due to quality

54
standards. The small number of training data samples led to not being able to stratify data and
identify best samples to base classification. The random number generator could have produced
different results by choice. Some trees will have a better or worst paper birch signature. Utilizing
a random number generator, the quality of the signature is chosen at random and accuracies will
be affected.
This project had an assumption of tree canopy buffer size. The diameter of tree could add
error to the classification. Tree canopies most likely had a variety of diameters. Averaging tree
canopy size most likely captures non-paper birch spectra. The application of these buffers could
have increased mixed pixels in the scene. Understory attributes were noted in a field notebook,
but not applied to methods. Mixing of vegetation and soil reflectance will occur with some
mixed pixels. Furthermore, the timing of this project’s data acquisition is narrow. Acquisition of
data just before or after leaf-off will lead to similarities between ground reference and dead leaf
reflectance. This project’s method is only applicable with fall image acquisition and need to be
considered in further work. Mixed pixels will diversify the true signature of birch on the
landscape and affect resulting classifications. Each pixel’s spectral response or “signature” can
be unique identifiers of pixels under ideal condition, but natural variation in the plants can vary
the spectral response (Campbell 2011).
Gradient and age of stands were not used, and the selection of a diverse age group of
birch introduced spectral variability into the experiment. All polygons used in the 16.5 m
validation scene had an average tree age of 70.7 years old. They also were noted to have paper
birch in decline at survey date. This influenced spectral output of specimens, but not examined.
The final classification methodologies used only 26 bands. Even with increased spectral
bands and reduced widths it is difficult to classify growth forms, because they also increase

55
fluctuation within and between spectral groups (Ustin 2010). Utilizing a reduced number of
bands was beneficial to data processing time, but with less bandwidth range was not appropriate
to resolve all endmembers involved in classifying. Further work utilizing the full range of
bandwidths may prove favorable for classifications accuracies. Many studies will remove bands
centered on both atmospheric water absorption and at the end of the EM spectrum due to noise
(Fassnacht et al., 2016). Spectral curves across species do not change that much, the level of
reflectance is the driver of specie contrast and can be exploited. This contrast between species
can be greatest in the NIR. The reflectance in the NIR is primarily controlled by leaf
morphology. This would need to be discussed if used for further work in this domain.
5.3. Further Research
This project shows inherent problems in classifying specie type on a landscape. That
means there is much opportunity to increase accuracies through adapting methodologies. Further
investigation into this domain will likely resolve part of the limitations. The use of remote
sensing tools like ArcGIS will have to be explored further. More research will be needed on
ground truthing surveys prior to going into the field.
Investigation of intra-specie variation is needed to be assessed if future work is done on a
subpixel level. Soft classifiers could be used to see if these classifiers do not assume homogeny.
Adding similar broadleaf species data to the linear modelling function could help narrow
signature to assess inter-species differences. Or comprehensive spectral libraries are needed to
delineate plant species under different conditions. Ground spectrometers could then be added to
project to capture best spectra for training data. The understory vegetation was not reviewed in
this project, but future classification could take into effect known understory vegetation.

56
The USGS Spectroscopy lab studies methods for classification through spectroscopic
remote sensing. There they have full spectra libraries of spectral signatures of minerals,
vegetation, and more. The cover classification produced in this project could be compared to
groupings with spectral curves mapped. Paper birch is not in their digital spectral library yet, but
species with similar morphology to birch like aspen have been investigated. They have recorded
spectra for all major seasonal variations within the specie. This could be used as training data for
future similar work given the timing of image acquisition.
Finally, future assessments could place more stratified random points. The strata would
be based on the collected ground truth refence data or prior known surveys. A traditional
confusion matrix would then be applied to the harden classifications. This could be produced
within the MN Forest Stand Inventory data. These stratified points would be buffered to the 9 m
diameter, turned into raster, and compared directly to canopy coverages. Future work could use
the soft classification of paper birch in the multiple image format and automate the process as a
hard classification. This would assign the pixels with the highest percentage of paper birch
signature into one membership grouping.
5.4. Applications of Research
Northeastern Minnesota is seen by the NFSC as a potential area of refuge for boreal
species. Adapting to climate change will need action plans that can anticipate and respond to
specie decline. There are many partners involved in managing and protecting these forests. Both
the MN DNR and the USDA Forest Service have some of the largest tracks of land within the
NSH. This project would have the greatest impact if utilize by them due to the resources
involved for both of those public organizations. This research helps to identify general paper
birch stand locations. This project recommends that those stand locations should be cross-

57
referenced against the best sites to retain paper birch populations for long-term protection e.g.
topography, hydrology, soil type. Pure spectral signature of paper birch is hard to come by within
the NSH. This project showed there is still a dominate percent of this habitat occupied by paper
birch. Areas classified as greater than 95 percent paper birch signature should be considered as
dense habitat and identified as priority locations for paper birch refuge. This would facilitate the
protection of paper birch by prioritizing management allocation on the landscape.
Forest communities are traditionally grouped into dominant specie type. This linear
unmixing approach assumes that large pixel resolution is diverse with multiple additive spectral
responses. The species boundaries are gradual and can be present in more than one community
type. The soft classifier is a valid approach to establishing methods for classifying the diverse
forest landscapes. This project’s method shows a way in which not to obtain high accuracies.
This knowledge will help future researchers with methods to more appropriately predict mixed
forest tree species.
Further GIS analysis and overlay would help forest managers to assess paper birch
change over time. The land cover change over time is detected by the comparison of two dates of
imagery. Birch tree loss overlay mapping could compare variables that affect paper birch i.e.
drought, soil compactness, historically logged areas, and deer population browsing. A “hot spot”
analysis could be produced with percent coverage of classification. Such a map would be
significant to forest restoration groups for funding opportunities and land cover change
monitoring. Being able to produce accurate specie locations based on spectral signatures would
greatly reduce cost of monitoring and resource exploration.

58
References
Abrams, Marc and Mark Kubiske. 1990. “Leaf structural characteristics of 31 hardwood and
conifer tree species in central Wisconsin: Influence of light regime and shade-tolerance
rank.” Forest Ecology and Management 31, no. 4 (April): 245-253.

Atkinson, P. M., M. E. J. Cutler and H. Lewis. 1997. “Mapping sub-pixel proportional land
cover with AVHRR imagery.” International Journal of Remote Sensing 18, no. 4
(November): 917-935.

Blackburn, George A. 2007. “Hyperspectral remote sensing of plant pigment.” Journal of
Experimental Botany 58, no. 4 (March): 855-867. doi: 10.1093/jxb/erl123.

Campbell, James B. and Randolph H. Wynne. 2011. Introduction to Remote Sensing. 5
th
ed. New
York: the Guildford Press.

Cochrane, M.A. 2000. “Using vegetation reflectance variability for species level classification of
hyperspectral data”. International Journal of Remote Sensing 21, no. 10 (November): 2075-
2087.

Conklin, James G. 1969. “Insect enemies of birch.” In Proceedings, birch symposium. 151-154.
USDA Northeastern Forest Experiment Station, Upper Darby, PA.

Coops, N.C., J.L. Lovell, D.L.B. Jupp, and D.S. Culvenor. 2003.“Using airborne and ground-
based ranging lidar to measure canopy structure in Australian forests.” Canadian Journal of
Remote Sensing 29, no. 5 (June), 607-622.

Crapper, P. F. 1984. “An estimate of the number of boundary cells in a mapped landscape
coded to grid cells.” Photogrammetric Engineering and Remote Sensing 50, no. 10: 1497-
1503.

Danson, F. M. 1995. “Developments in the remote sensing of forest canopy structure”. In
Advances in Environmental Remote Sensing. edited by F.M. Danson and S. E. Plummer
(Chichester: John Wiley & Sons): 53–69.

Day, R. J., and E. M. Harvey. 1981. “Forest dynamics in boreal mixedwood.” Lecture, Thunder
Bay, Ontario, September 16.

Elhadi, Adam, Mutanga Onisimo, and Denis Rugege. 2010. “Multispectral and hyperspectral
remote sensing for identification and mapping of wetland vegetation: a review.” Wetland
Ecol Manage 18, (December): 281-296.

59
Fassnacht, Fabian Ewald, Hooman Latifi, Kryzstof Sterenczak, Aneta Modzelewska, Michael
Lefsky, Lars T. Waser, Christopher Straub, and Aniruddha Ghosh. 2016. “Review of studies
on tree species classification from remotely sensed data.” Remote Sensing of Environment
186, (December): 64-87.

Foody, G.M., and D. P. Cox. 1994. “Sub-pixel land cover composition estimation using a linear
mixture model and fuzzy membership functions.” Remote Sensing 15, no. 3 (March): 619-
631.

Franklin, S.E., and M.A. Wulder. 2002. “Remote sensing methods in medium spatial resolution
satellite data land cover classification of large areas.” Progress in Physical Geography 26,
no. 2 (June): 173–205.

Franklin, S.E., B. Titus, and R. Gillespie. 1994. “Remote sensing of vegetation cover at forest
regeneration site.” Global Ecology and Biogeography Letters 4, no. 2 (March): 40-46.

Gao, Bo-Cai, Alexander Goetz, Marcos J. Montes, and Curtiss O. Davis. 2009 “Atmospheric
correction algorithms for hyperspectral remote sensing data of land and ocean.” Remote
Sensing in the Environment 113, no. 1 (September): S17-S24.

Gartzia, Maite, Concepcion Alados, Fernado Perez-Cabello, and C. Guillermo. 2013 “Improving
the Accuracy of Vegetation Classifications in Mountainous Areas: A Case Study in the
Spanish Central Pyrenees.” Mountain Research and Development 33, no. 1 (December): 63-
74.

Gaussman, H.W. 1977. “Reflectance of leaf components.” Remote Sensing in the Environment 6,
no. 1: 1-9.

Gitelson, Anatoly and Mark N. Merzlyak. 1994. “Spectral Reflectance Changes Associated with
Autumn Senescence of Aesculus hippocastanum L. and Acer platanoides L. Leaves.
Spectral Features and Relation to Chlorophyll Estimation.” Journal of Plant Physiology
143, no. 3 (March): 286-292.

Giannico, Chiara 2004. “Remote sensing of vegetation in the Calabrian region.” Acta Astronaut
60, no. 2 (January): 119-131.

Goetz, A.F.H., and V. Srivastava. 1985. “Mineralogical Mapping in the Cuprite Mining District,
Nevada,” Proc. of the Airborne Imaging Spectrometer Data Anal. Workshop; 22-31
conference paper Jet Propulsion Lab. Pasadena, CA

Harrison, A. R., S. J. Hurcom, M. Taberner. 1991. “Examining spectral mixture modelling using
in situ spectroradiometric measurements.” Spatial Data 2000 (Nottingham: Remote Sensing
Society), 242-249.

60
He, C., Q. Zhang, Y. Li, X. Li, and P. Shi. 2005. “Zoning grassland protection area using remote
sensing and cellular automata modeling—A case study in Xilingol steppe grassland in
northern China.” Journal of Arid Environments 63, no. 4 (December): 814-826.

Hoffer, R. M., and Chris J. Johannsen. 1969. “Ecological potentials in spectral signature
analysis.” In P. L. Johnson ed., Remote sensing in ecology, University of Georgia, Athens,
Georgia.

Ichoku, Charles, and Arnon Karnieli. 2009. “A review of mixture modeling techniques for sub ‐
pixel land cover estimation.” Remote Sensing Reviews 13, no. 3-4 (October): 161-186.

Irwin, Larry L. 1985. “Foods of moose, Alces, and white-tailed deer, Odocoileus virginianus, on
a burn in boreal forest.” Canadian Field-Naturalist 99, (February): 240-245.

Jensen, J.R., 1996. Introduction to Digital Image Processing: A remote sensing perspective. 2nd
ed. Piscataway, NJ: Prentice Hall.

Jordan, James S. and Francis M. Rushmore. 1969. “Report on animal damage to birch.” In
Proceedings, birch symposium. 155-163. USDA Northeastern Forest Experiment Station,
Upper Darby, PA.

Key, Thomas, Timothy A. Warner, James B. McGraw, and Mary Ann Fajvan. 2001. “A
Comparison of Multispectral and Multitemporal Information in High Spatial Resolution
Imagery for Classification of Individual Tree Species in a Temperate Hardwood Forest”.
Remote Sensing of Environment 75, no. 1 (January): 100-112.

Lauver, Chris. 1997. “Mapping species diversity patterns in Kansa shortgrass region by
integrating remote sensing and vegetation analysis.” Journal of Vegetation Sciences 8, no. 3
(February): 387-394.

Leckie, Donald G., Sally Tinis, Trisalyn Nelsen, Charles Burnett, Francois A. Gougeon, Ed
Cloney, and Dennis Paradine. 2005. “Issues in species classification of trees in old growth
conifer stands.” Canaidian Journal of Remote Sensing 31, no. 2 (June): 175-190.

Lee, K.H. and R.S. Lunetta. 1996. Wetland Detection Methods. In: Lyon J.G. and McCarthy J.
(eds). Wetland and Environmental Applications of GIS. Lewis Publishers, New York, 249–
284.

Lu, D., and Q. Weng, 2007. “Survey of image classification methods and techniques for
improving classification performance.” International Journal of Remote Sensing 28, no. 5
(March): 823–870.

Malgorzata V.W., and N. Egyetem, 2010. “Data acquisition and integration, Remote Sensing.”
University of West Hungary Faculty of Geoinformatics.

61
Marshall, Arbust. Amer. 19. 1785 “Betula papyrifera,” Flora of North America 3, Accessed June
23, 2017 http://www.efloras.org/florataxon.aspx?flora_id=1&taxon_id= 233500260.

Mehner, Henny, Mark Cutlert, David Fairbairn, and Gillian Thompson. 2004. “Remote sensing
of upland vegetation: the potential of high spatial resolution satellite sensors.” Global
Ecology and Biogeography 13, no. 4 (June): 359-369.

North Dakota Tree Handbook. North Dakota State University, Centennial Tree Program project.

MN DNR. 2006. North Shore Highlands; Subsection Profile, an Action plan for Minnesota
Wildlife. Accessed June 16, 2017.
http://files.dnr.state.mn.us/assistance/nrplanning/bigpicture/cwcs/profiles/north
_shore_highlands.pdf.

North Shore Forest Collaborative. “Why are the Birch Dying?” Accessed June 16, 2017.
http://northshoreforest.org/what-happened-to-the-forest/.

Ollinger, S.V. 2011. “Source of variability in canopy reflectance and the convergent properties
of plants.” New Phytologist 189, no. 2 (November): 375-394.

Parker, D.C., and M.F. Wolff. 1965. “Remote Sensing” International Science and Technology
43, (July): 20-31.

Reed, Bradley C., Jesslyn F. Brown, Darrel VanderZee, Thomas R. Loveland, James W.
Merchant, and Donald O. Ohlen. 1994. “Measuring phenological variability from satellite
imagery” Journal of Vegetation Science 5, no. 5 (October): 703-714.

Ruiliang, Pu, 2009. “Broadleaf species recognition with in situ hyperspectral data.” International
Journal of Remote Sensing 30, no. 11 (June): 2759-2779. Accessed September 18, 2017.
DOI: 10.1080/01431160802555820.

Safford, L. O., John C. Bjorkbom, and John C. Zasada. 1990. Betula papyrifera Marsh. paper
birch In: Burns. Silvics of North America. Hardwoods. Agric. Handbook. U.S. Department
of Agriculture, Forest Service. Washington, DC. 654 (2): 158-171.

Sayn-Wittgenstein, L. 1978. Recognition of Tree Species on Aerial Photographs. Canadian
Forest Service Forest Management Institute Information Report, FMR-X-118, Environment
Canada. 97.

Uchytil, Ronald J. 1991. “Betula papyrifera.” In: Fire Effects Information System. USDA,
Rocky Mountain Research Station, Fire Sciences Laboratory.

Ustin, Susan, and John Gamon. 2010. “Remote sensing of plant functional groups.” New
Phytologist 186, no. 4 (May): 795-816.

62
Woodcock, Crutis E., and Alan H. Straher. 1987. “The factor of scale in remote sensing.”
Remote Sensing of the Environment 21, no. 3 (April): 311-332.

Zarco-Tejada, P.J., A. Berjon, and J.R. Miller. 2000. “Stress Detection in Crops with
Hyperspectral Remote Sensing and Physical Simulation Models.” Institute for Sustainable
Agriculture Universidad de Valladolid, Spain Dept. Earth and Space Science &
Engineering, York University.

63
Appendix A: Dataset Metadata
AVIRIS Metadata:

Abstract: The AVIRIS sensor collects data that can be used for characterizing
Earth's surface and atmosphere from geometrically coherent spectro-radiometric measurements.
The AVIRIS instrument has 224 spectral bands with wavelengths from 400 to 2500 nanometers
(visible, near infrared and mid-infrared portions of the spectrum). The AVIRIS instrument is
flown on a NASA ER-2 aircraft. Data archived at the JPL AVIRIS Data Facility are available in
both reflectance and radiance units. Scale and resolution depend on flight statistics.
Flight Name F120930t01p00r11 F061002t01p00r11
Date/ Time 9/30/2012 UTC 18:33 10/2/2006 UTC 18:45
Site Name Aspen 4, MN Arrowhead 1, MN
Investigator Phil Townsend Phil Townsend
Comments Alt = 17.5 kft, SOG = 100
kts, Clouds = Clear, Direction
= 91.6
100% clear
# Samples 1022 758
# Lines 1410 4305
Pixel Size 4.3 m 16.5 m
Solar Elevation 38.84 37.59
Solar Azimuth 192.05 195.94
Rotation -82 69
File Size 0.71 GB 1.67 GB

Table 10 AVIRIS flight data

64
FSA NAIP Imagery:

National Agriculture Imagery Program (NAIP) natural color .6-meter pixel resolution. The
imagery was collected statewide. This data set contains imagery from the National Agriculture
Imagery Program (NAIP). The NAIP acquires digital ortho imagery during the agricultural
growing seasons in the continental U.S. A primary goal of the NAIP program is to enable
availability of ortho imagery within one year of acquisition. The NAIP provides two main
products: 1-meter ground sample distance (GSD) ortho imagery rectified to a horizontal accuracy
within +/- 5 meters of reference digital ortho quarter quads (DOQQ's) from the National Digital
Ortho Program (NDOP) or from the National Agriculture Imagery Program (NAIP); 1 meter or
60cm GSD ortho imagery rectified within +/- 6 meters to true ground. The tiling format of NAIP
imagery is based on a 3.75' x 3.75' quarter quadrangle with a 300-pixel buffer on all four sides.
The NAIP imagery is formatted to the UTM coordinate system using the North American Datum
of 1983 (NAD83). The NAIP imagery may contain as much as 10% cloud cover per tile.
Constraints: Not to be used for navigation, for informational purposes only.

Ground Truth Dataset:

This data was collected by Kyle Uhler. Data collected is an inventory of paper birch locations
near roads within the AVIRIS image F120930t01p00r11. Areas accessed are shown in figure 7.
Only paper birch with DBH greater or equal to 20 cm were sampled. Horizontal accuracies
above pixel resolution of 4.3 m were not kept for projects analysis. This data was collected for
the methodologies outlined in this thesis document. Data is up-to-date as 09/24/2017. All data
was collected in UTM Zone 15 N coordinate system and North American Datum 1983 (meters).

MN Forest Stand Inventory Metadata:

This dataset is operated by Minnesota Department of Natural Resources (MNDNR) - Division of
Forestry. This layer is a digital inventory of individual forest stands. The data are collected by
MNDNR Foresters in each MNDNR Forestry Administrative Area, and is updated on a
continuous basis, as needed. There are 50 classification categories in this layer. Most stands are
field checked, and their characteristics described. The MN DNR uses internal MNDNR
classification schema. This data originates from the MNDNR's "Forest Inventory Management"
system (also referred to as FIM). The data are collected for Forest Resource Planning, harvest
plans, treatment plans, wildlife habitat assessment, biotic community mapping support, historical
vegetation studies. The data are up-to-date as of 07/14/2017. Content date indicates the date
which the user can be confident of accuracy and completeness of the dataset. All polygons are in
UTM Zone 15 N NAD83 (meters).

Abstract (if available)

Abstract Paper birch (Betula papyrifera) is a dominant specie within Northern Minnesota’s Laurentian mixed forest. Though these trees are common place, paper birch populations have been in decline for the past couple decades along Lake Superior. Due to the reduced replacement rate of this specie, organizations are implementing management strategies to promote healthy forests. This thesis investigates remote sensing techniques to predict paper birch locations remotely. The thesis tests individual specie level spectral signature effectiveness to classify community level data of the same informational group. The project uses hyperspectral Airborne Visible/ Infrared Imaging Spectrometer (AVIRIS) flights data for its imaging platform. Two spatial resolution scenes were classified for this project. A 4.3 m pixel resolution image was used for defining spectral signatures based on ground truth data. Then a lower resolution 16.5 m image was used to apply the produced paper birch signature as a classifier to test functionality of these methods on known paper birch communities. Pixels were used as the final classification unit. A linear unmixing soft classification was utilized to produce percent signature contained within pixels. The classification resulted in ∼93% of forest plots containing some pixels with 95% similar spectral signatures to paper birch. Total classified coverage of validation forest plots was low with only ∼30% covered with 75% similar signature or greater. The long-term objective for this project is to automate specie identification to monitor trees. Further research is needed to streamline classification and refine procedures, yet current findings can help forest managers and conservationists identify priority sites to both map current specie distribution and implement restoration activities.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

A comparison of pixel-based, object-based, deep learning, and data fusion image classification approaches in delineating urban tree canopy in City Terrace

PDF

From ruins to pixels: using remote sensing and GIS to analyze, document, and visualize archaeological sites and ancient roadways in Chaco Canyon, New Mexico

PDF

Finding the green in greenspace: an examination of geospatial measures of greenspace for use in exposure studies

PDF

Species distribution modeling to predict the spread of Spartium junceum in the Angeles National Forest

PDF

Surface rupture detection with support vector machine classification: case study from Ridgecrest, CA

PDF

Classifying ash cloud attributes of Eyjafjallajökull volcano, Iceland, using satellite remote sensing

PDF

Impacts of vegetation management on wildfire severity: a study of the 2021 Caldor fire

PDF

Using Landsat and a Bayesian hard classifier to study forest change in the Salmon Creek Watershed area from 1972–2013

PDF

Climate conservation priorities: using MCDA to identify future refuge for the Joshua Tree Forest in California

PDF

Assessing modern conflict to monitor human rights with remote sensing: Russia's war in Ukraine

PDF

Use of remotely sensed imagery to map sudden oak death (Phytophthora ramorum) in the Santa Cruz Mountains

PDF

Using aerial imagery to assess tropical forest cover surrounding restoration sites in Costa Rica

PDF

Exploring remote sensing and geographic information systems technologies to understand vegetation changes in response to land management practices at Finke Gorge National Park, Australia Between ...

PDF

An examination of close-range photogrammetry and traditional cave survey methods for terrestrial and underwater caves for 3-dimensional mapping

PDF

Deep convolutional neural networks for remote sensing investigation of looting of the archeological site of Al-Lisht, Egypt

PDF

Alaska Hike Search: designing a mapping application for Alaskan trails and user contributed hazards

PDF

Building better species distribution models with machine learning: assessing the role of covariate scale and tuning in Maxent models

PDF

The role of precision in spatial narratives: using a modified discourse quality index to measure the quality of deliberative spatial data

PDF

Assessing the transferability of a species distribution model for predicting the distribution of invasive cogongrass in Alabama

PDF

Evaluating machine learning tools for humanitarian road network mapping

Asset Metadata

Creator Uhler, Kyle Benjamin (author)

Core Title Mixed forest image classification of paper birch: using AVIRIS bandwidths ranging from 530 to 745 nm

School College of Letters, Arts and Sciences

Degree Master of Science

Degree Program Geographic Information Science and Technology

Publication Date 08/13/2018

Defense Date 05/09/2018

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag AVIRIS,image classification,linear unmixing model,OAI-PMH Harvest,paper birch,remote sensing

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Sedano, Elisabeth (committee chair)

Creator Email kuhler@usc.edu,uhler.kyle@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-65233

Unique identifier UC11668665

Identifier etd-UhlerKyleB-6738.pdf (filename),usctheses-c89-65233 (legacy record id)

Legacy Identifier etd-UhlerKyleB-6738.pdf

Dmrecord 65233

Document Type Thesis

Format application/pdf (imt)

Rights Uhler, Kyle Benjamin

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA