University of Southern California Dissertations and Theses

Geographic object based image analysis for utility scale photovoltaic site suitability studies

(USC Thesis Other)

Geographic object based image analysis for utility scale photovoltaic site suitability studies

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Geographic Object Based Image Analysis for Utility Scale Photovoltaic Site Suitability Studies
by
John McDermott
A Thesis Presented to the
FACULTY OF THE USC DORNSIFE COLLEGE OF LETTERS, ARTS AND SCIENCES
University of Southern California
In Partial Fulfillment of the
Requirements for the Degree
MASTER OF SCIENCE
(GEOGRAPHIC INFORMATION SCIENCE AND TECHNOLOGY)
May 2021
Copyright © 2021 John McDermott
ii

Acknowledgements
I want to thank my wonderful wife and family for always supporting my endevors, the JYG for
always being there for me, and I would also like to thank the faculty at USC SSI for all their help
along the way.
iii
Table of Contents
Acknowledgements ......................................................................................................................... ii
List of Tables ................................................................................................................................. vi
List of Figures ............................................................................................................................... vii
List of Abbreviations ................................................................................................................... viii
Abstract .......................................................................................................................................... ix
Chapter 1 Introduction .................................................................................................................... 1
1.1. Site Review Process ............................................................................................................2
1.1.1. Incorporating Object Based Image Analysis .............................................................6
1.2. Study Area ..........................................................................................................................7
1.2.1. Tennessee Valley Authority .......................................................................................7
1.2.2. Study Area Selection..................................................................................................8
1.3. Summary of Project Objectives ........................................................................................11
Chapter 2 Related Works .............................................................................................................. 13
2.1. Utility Scale Photovoltaic Construction Impediments ......................................................13
2.1.1. Roads........................................................................................................................14
2.1.2. Buildings ..................................................................................................................15
2.1.3. Vegetation ................................................................................................................15
2.2. Platforms ...........................................................................................................................16
2.2.1. Spatial Resolution ....................................................................................................17
2.3. OBIA Segmentation and Classification Workflows .........................................................18
2.3.1. Segmentation............................................................................................................19
2.4. Findings.............................................................................................................................20
Chapter 3 Data and Methods......................................................................................................... 22
3.1. Data Collection and Preparation .......................................................................................22
iv
3.1.1. Data Sets and Sources ..............................................................................................23
3.1.2. Wetlands ..................................................................................................................24
3.1.3. Floodplains ...............................................................................................................24
3.1.4. Slope ........................................................................................................................24
3.1.5. Parcels ......................................................................................................................25
3.1.6. Satellite Imagery ......................................................................................................25
3.2. Software ............................................................................................................................26
3.2.1. Cloudeo ....................................................................................................................26
3.3. Methodology and Workflow .............................................................................................26
3.3.1. Data Preparation.......................................................................................................27
3.3.2. Segmentation............................................................................................................29
3.3.3. Classification............................................................................................................31
3.3.4. Process Iteration .......................................................................................................32
Chapter 4 Results .......................................................................................................................... 34
4.1. Segmentation.....................................................................................................................34
4.2. Rules .................................................................................................................................35
4.2.1. Rule Development Process ......................................................................................35
4.3. Accuracy Assessment .......................................................................................................37
4.3.1. Process Iteration .......................................................................................................38
4.3.2. First Iteration ............................................................................................................40
4.3.3. Second Iteration .......................................................................................................41
4.3.4. Third Iteration ..........................................................................................................43
4.3.5. Fourth Iteration ........................................................................................................43
4.3.6. Fith Iteration.............................................................................................................44
4.3.7. Sixth Iteration...........................................................................................................45
v
4.3.8. Seventh Iteration ......................................................................................................46
4.4. Conclusions .......................................................................................................................47
Chapter 5 Discussion and Conclusions ......................................................................................... 48
5.1. Findings.............................................................................................................................48
5.2. Applications and Benefits .................................................................................................49
5.3. Limitations ........................................................................................................................51
5.3.1. Temporal ..................................................................................................................51
5.3.2. Regional ...................................................................................................................52
5.3.3. Complexity ...............................................................................................................52
5.3.4. Labor Investments ....................................................................................................53
5.4. Areas of Further Research ................................................................................................54
5.4.1. Objects .....................................................................................................................54
5.4.2. Scalability ................................................................................................................54
5.5. Conclusion ........................................................................................................................55
Refrences ………………………………………………………………………………………...56

Appendix A ................................................................................................................................... 61

vi
List of Tables
Table 1. Common Impediments and Related Works .................................................................... 11
Table 2: Datasets and Sources……….…………………………………………………………. 18
Table 3: Segmentation Parameters ………………………………………………………………21
Table 4: Rule 7…………………………………………………………………………………...37
Table 5: First Iteration ...………………………………………………………………………...37
Table 6: Second Iteration ...……………………………………………………………………...37
Table 7: Third Iteration ...…………………..…………………………………………………....38
Table 8: Fourth Iteration …………………………………………………………………….…..39
Table 9: Fifth Iteration ...………………………………………………………………………...40
Table 10: Sixth Iteration ...………………………………….…………………………………...40
Table 11: Seventh Iteration ...…………………………………………………………………...41
Table 12: Potential benefits ……………………………………………………………………..48

vii

List of Figures
Figure 1: Site Selection Process Inputs and Outputs……………………………………………. 4
Figure 2: Study Area within the Tennessee Valley Authority ….…………………………….... 8
Figure 3: Detailed Study Area ……………………………………………………….……….... 10
Figure 4: Site identification and mask creation process …………………………..…………… 28
Figure 5: Mask Creation………...……………………………………………………………… 30
Figure 6: Object identification and export………………...……………………………………. 32
Figure 7: Results of rule 7………………………………………………………………………. 36
Figure 8 Results of rule 7 modifications…………………………………………………………37
Figure 9: OBIA Process Iteration……….………………...……………………………………. 39
Figure 10: First iteration classification results………………………………………………….. 41
Figure 11: Second iteration classification results………………………………………………. 42
Figure 12: First iteration segmentation image………………………………………………….. 61
Figure 12: Second iteration segmentation image……………………………………………….. 62
Figure 12: Third iteration segmentation image…………………………………………………. 63
Figure 12: Fourth iteration segmentation image…………………………………………………64
Figure 12: Fifth iteration segmentation image………………………………………………….. 65
Figure 12: Sixth iteration segmentation image………...……………………………………….. 66
Figure 12: Seventh iteration segmentation image………………...…………………………….. 67

viii

Abbreviations
AVHRR Advance very high resolution radiometer
EPA Environmental Protection Agency
GIS Geographic information system
GISci Geographic information science
GUI Graphical user interface
MRS Multi resolution segmentation algorithm
MW Megawatt
OBIA Object based image analysis
PV Photovoltaic
RS Remote sensing
SEIA Solar Energy Industry Association
SSI Spatial sciences institute
TVA Tennessee Valley Authority
USC University of Southern California
USGS United States Geological Survey
USSE Utility scale solar energy
VHR Very high resolution

ix
Abstract
As our country grapples with the long term negative effects that traditional electrical generation
methods have on the environment, such as nuclear with a 50 year average decommissioning
time, natural gas and the methane emissions associated with it, and coal which is not clean, there
is a renewed focus on solar energy. This renewed focus is partially fueled by advancements in
photovoltaic cell technology and favorable regulatory conditions, resulting in a decrease of solar
energy production costs. This has led to the installed solar energy production capacity of the
United States to grow from 7.33 gigawatts in 2012 to 51.45 gigawatts in 2018. As the industry
matures and solar energy is adopted in new markets, the available land suitable for development
has subsequently been reduced. A result of this is the industry has shifted its focus to identify
suitable sites in areas that have been otherwise overlooked or discounted. To remain competitive,
potential sites must be screened to identify site conditions that can increase costs or render a site
undevelopable. This project identified that OBIA can successfully be used to identify a multitude
of features that are encountered during the development of USSE projects, but the complexity
and variability of the process makes it currently unsuitable to be deployed at scale. OBIA can
however, be used to assess current site suitability analyses by generating otherwise unknown
attribute data about a site to target locations wherein to look.

1

Chapter 1 Introduction
A combination of advancements in photovoltaic (PV) cell performance, shifting cultural attitudes
towards the use of fossil fuels, government mandates, and a reduction in capital expenditures has
created an energy market in which electricity generated by PV can compete on cost. In the first
quarter of 2020, a record 3.6GW of solar PV was installed with another 5.4GW of utility scale
projects being announced. Cumulatively 14.4GW of utility scale solar is expected to be installed
in the United States in 2020 (SEIA, 2020). Simultaneously, the cost to develop utility-scale PV
has decreased by two-thirds since 2009 (U.S. Department of Energy, 2017). With the increase in
installation numbers, comes increased competition to identify and acquire land suitable for
projects up to thousands of acres in size.
Typically, identifying ideal sites for USSE developments is a straightforward process
when there is ample land available. Criteria like aspect, slope, solar radiation, and zoning are
available as sources of data readily inputted into a GIS, however, this process does not apply to
all situations. When evaluating sites in areas categorized by less favorable conditions, such as
hilly terrain or areas prone to flooding, traditional methods used to identify sites with
conditionals suitable for the development of utility scale PV becomes a time-intensive process.
This is a result of the need to manually evaluate and vet each site for surface conditions and
possible natural or manmade features that may be present, but are not accounted for in the vector
data employed. Manual review is done to identify features on sites such as tree stands, wetlands,
buildings, or roads; that lead to increased construction costs and restricts where on the sites
USSE can be developed. While it may seem relatively straightforward to identify sites, a
decrease in ideal sites requires a more precise approach to site selection, which has historically
been done manually. This research aims to identify whether object-based image analysis can be
2

used to accurately identify common construction impediments encountered on less-viable USSE
sites, and if so, provide a conceptual workflow on how one could incorporate these findings into
their site selection process.
The term utility-scale has multiple definitions (SEIA, 2020; U.S. Department of Energy,
2017). The Solar Energy Industries Association (SEIA) defines utility-scale PV as any project
that has an offtake agreement with a utility, regardless of project size. Conversely, financiers of
PV development define utility-scale by the amount of investment required to construct the site.
For the purposes of clarity, this research defines utility-scale is as a PV development that is 5mw
or larger in size (US Department of Energy, 2012). Each megawatt (MW) produced requires, on
average, 8-10 acres of developable land (Mulvaney, 2019).
1.1. Site Review Process
A typical USSE site selection process is relatively standardized, which is convenient to
those involved. However it lacks accuracy by relying on datasets such as the National Wetland
Inventory (NWI) and the Federal Emergency Management Agency (FEMA) 100 Year
Floodplains that are dated. Since 1996 the Fish and Wildlife Service (FWS) has been updating
the NWI at a rate of 2% of the total land area of the lower 48 states per year (US Fish and
Wildlife Service, n.d.). Research that closely parallels this project, conducted in 2019, relies on
wetland datasets from 2002 and a landside risk dataset from 1991 in the site selection process
(Guaita-Pradas et al, 2019). These data sets have since become outdated, as the wetland
morphology has changed significantly in years since. As a result, inaccuracies are introduced
into models.
The site selection process begins with the exclusion of land deemed undevelopable. The
reason land can be deemed undevelopable can be grouped into two categories, topographic and
3

regulatory. Topographic constraints on development include aspect, slope, soil composition, and
wetlands. Regulatory constraints include protected lands, areas of environmental concern, and
zoning. Undevelopable land is removed from the original parcel geometry resulting in a new
parcel geometry that is representative of only the developable areas within a given site. This is
done by loading parcel data into a model created in ArcGIS Model Builder that erases wetlands,
floodplains and slope greater than 10%. In this example two parcels from the study area were
selected and used as the input to the model. The output of the model is the new parcel geometry
with only the developable land remaining. Figure one illustrates this process, it is important to
note that the total developable land decreased from 227.6 acres to 194.67 acres for Example A
and from 241.50 acres to 184.67 acres for Example B. This reduction of available developable
land only accounts for the above criteria and falls short of fully quantifying the sites conditions
such as roads, buildings, vegetation, irrigation channels and retention ponds.
4

Figure 1: Example sites A and B’s parcel boundary (white) is used as the input for the model.
The model removes slope greater than 10%, NWI, and FEMA 100 year floodplain from the
parcel’s geometry.. The output of this process only leaves the remaining developable land (rust
color) remaining.
Figure 1 illustrates one problem with the current selection process. In Figure 1, Example
A has 194.67 acres of developable land after the initial screen compared to Example B with
184.69. However, the unaccounted for tree stands intersecting the developable land in Example
A create a break along the eastern edge of the parcel. To mitigate the impact of the trees, the
5

developer would need to clear them in advance. If this was not an option, the site would either
have to be designed around the trees or built at a smaller scale, by not developing the smaller of
the two areas. All of the options available increase capital expenditures or result in a smaller
project, thus producing less revenue. This is also true of buildings and roads, although in these
cases the impediments cannot always be moved or cleared. The earlier in the development
process these features are identified, the earlier they can be accounted for to allow for accurate
cost estimation and site planning to be conducted.
The number of potential sites remaining after the removal of the slope, wetlands, and
floodplains varies and is determined by the size of the area of the target market, the size of the
project, and the complexities of the local topography. A site for a 5MW project requires
approximately 50 developable acres, while a 50MW project requires upwards of 500 developable
acres. Due to this, there are fewer potential sites as a project size increases. That said, these
numbers run in the thousands. It is necessary to run the models at this scale because not every
landowner is willing to sell or lease their land for USSE development; it can be assumed that
only 5-10% of contacted landowners will be interested. Of the projects that make it past this
initial phase, still less will make it through the full development cycle and be placed in service
(Mulligan, 2020).
When faced with making decisions that are difficult to automate, a GIS technician uses
their spatial reasoning and logical assumptions to assess the impacts, not only of a single feature,
but all of a site’s features in relation to one another. Two sites of 100 acres can both have the
same number of buildings, roads, and trees, with only one being be suitable for development. For
example, one site may have a road running along the property line to a home in the corner of the
property, versus a site with a road that leads to a house surrounded by trees situated in the middle
6

of a property. In the first case, the layout of the features does not impact the overall site. By
relying on human reasoning and spatial analyses, the process gains precision and the GIS-based
site selection process is improved upon. However the process cannot be scaled or replicated,
which makes it time-consuming in nature, potentially causing delays in the selection process.
Further, human judgement, while it has the potential to increase precision, can also introduce the
element of human error. Ultimately, no two GIS technicians will evaluate a site’s suitability in
the same way, even though they may agree on many of the key characteristics and concepts of
the site selection process
1.1.1. Incorporating Object Based Image Analysis
During the site selection process, a GIS technician is identifying, categorizing, and
quantifying the objects present on the site based on shape, texture, spectral properties, and spatial
relationship to other objects present (Rizvi et al, 2019). The goal of this research is to identify if
this process can be mimicked by using OBIA to identify and classify objects, rather than being
done manually by a GIS technician. OBIA first groups pixels into objects based on their spectral,
textural, or spatial similarities; the objects are then identified using rule based classification.
OBIA was specifically chosen for this project as it “…applies a logic intended to mimic some of
the higher order logic employed by human interpreters, who can use the sizes, shapes, and
textures of regions, as well as the spectral characteristics used for conventional pixel based
classification.” (Campbell and Wynne 2012, 371). The assumption is, if OBIA can accurately
identify and classify construction impediments, similarly to how a technician would, this process
can be partially or fully automated. This could then be scaled in size to increase the overall
efficiency of the site selection process. The ensuing benefits of this improvement will be the
7

expedition the site selection process, granting a competitive advantage over developers
employing time consuming methods, and a decrease in the associated labor costs.
1.2. Study Area
Each environment presents its own unique conditions and challenges. In the desert
regions of the American Southwest, shallow bedrock leads to increased costs associated with
driving the pilings needed to support PV arrays. In the Pacific Northwest, there is extensive tree
cover and difficult topography. The study area select for this research is located in the southwest
corner of Weakly County, TN, within the Tennessee Valley Authority (TVA) electric service
territory. This area was purposely chosen for its diverse topography and the interplay between
agricultural land and forest or rural infrastructure, which increases the likelihood that
impediments will be found. Selecting a study area devoid of any impediments would not fulfill
the need of this research. Because the very high resolution (VHR) image that will be used is
being provided free of charge from a third party, this research was limited to acquiring only one
image. As such it was important to purposively select a study area that would provide enough
examples of impediments of interest.
1.2.1. Tennessee Valley Authority
The TVA service territory was selected for this study for multiple reasons. One being that
the market expected to see an increase in the construction of utility scale projects. Based on the
TVA’s integrated resource plan (IRP) published in 2019, the utility plans on installing an
additional 14 gigawatts of PV generating capacity over the next 20 years (St. John, 2019). The
TVA’s electric service territory has approximately 16,000 miles of transmission lines that span
roughly 80,000 square miles (TVA, 2020). This area covers all of Tennessee as well as parts of
Southern Kentucky, Southern Virginia, Eastern North Carolina, Northern Georgia, Northern
8

Alabama, and Northern Mississippi. Figure 2 shows TVA’s territory as well as the study area for
this project.

Figure 2 : TVA service territory and study area.
1.2.2. Study Area Selection
The selection of the study area was based on a number of factors. For a utility scale
project to succeed, it is necessary to be in proximity to load centers, or areas that consume large
amounts of energy such as cities, as well as a large amount rural or undeveloped land. The area
needed to have the topographic qualities associated with utility scale solar. These areas are
primarily flat and free of wetlands, floodplains, and high slopes.
9

The study area for this project lies within an area known as the Mississippi Valley Loess
Plains that comprises the western edge of Tennessee. This area is on average between 250 – 500
feet in elevation and known for loess deposits that can be up to 50 feet thick in some areas
(Environmental Protection Agency, n.d.). This area has a large number of river systems and
floodplains that transect the area. Historically the area was covered in oak and hickory tree
stands and floodplain forests, but much of this has been cleared and converted for agricultural or
livestock use (Environmental Protection Agency, n.d.). Figure three shows in greater detail the
location and features of the study area.
10

Figure 3: Study Area within the USGS 7.5 minute Garder, Tennessee Quadrangle.
11

After a visual review, the location of the study area was selected using base map satellite
imagery within ArcGIS. One critical criteria were that it contained the necessary features within
a 60 km
2
area. The size limitation was a result of the need for VHR imagery. Due to the high
costs of obtaining VHR imagery, 60 km
2
is the maximum size provided for academic use by the
vendor. This will be discussed in further detail in Chapters 2 and 3. The study area also had to
contain sites where a USSE project could feasibly be developed. After running an initial screen
on the whole state of Tennessee by removing slope greater than 10%, wetlands, and floodplains,
the forest cover of the sites was tabulated using the National Land Cover Database. The final
study area was selected because it contained potential sites with established forest cover within a
60 km
2
area.
1.3. Summary of Project Objectives
The objective of this project is to identify if OBIA can be used to more effectively
identify objects on a potential solar site that may cause additional time and cost despite what may
appear, via traditional methods, to be suitable sites. A literature review was conducted to identify
related works and research, and to ensure that this idea of research was a novel one. Related
works, discussed in Chapter 2, provide the framework on which this research project was
designed. The related works section informed the requirements of this project, such as the need
for VHR imagery and a means of measuring the accuracy of the process. Chapter 3 describes the
methodology used for this project. It explains and describes the sources for data used in the
project. In addition, the software utilized for OBIA, Harris Geospatial ENVI Feature Extraction,
will be discussed and explained. Chapter 4 discusses and assesses the accuracy of the OBIA
process. Chapter 5 concludes the paper by providing recommendations on how this research can
12

be expanded upon as well as its broader importance within the context of industrial solar
development, within the TVA, nationally, and globally.
13

Chapter 2 Related Works
This literature review did not identify previous research on the application of OBIA for USSE
site selection process. This is likely due to the fact that only recently have user interfaces been
developed that allow for routine or practical application of OBIA, while much the theoretical and
conceptual data was completed in previous decades (Campbell and Wynne, 2012). This chapter
introduces and defines common construction impediments encountered in the study area, reviews
research related to identifying features that share similarities with the features of interest to this
project, and the related research on how OBIA can be used to identify these features. It will also
discuss the spatial and spectral resolutions required for OBIA, and the importance they have in
achieving accurate results. The chapter concludes by providing sample workflows that can be
replicated or modified to suit the needs of other studies.
2.1. Utility Scale Photovoltaic Construction Impediments
There are many definitions of what constitutes a utility scale PV development. For the
purposes of this study, utility scale is defined as a PV development that is 5mw or larger in size
(US Department of Energy, 2012). Each megawatt (MW) produced requires, on average, 8-10
acres of developable land (Mulvaney, 2019). Using this formula, a 100mw PV development will
require 800-1000 acres of suitable land. Due to their size, utility scale PV developments are most
often found outside of urban centers in rural areas with large flat tracts of contiguous land, free
from impediments such as trees, buildings, roads, and water. Site impediments are conditions or
features on the ground that can increase construction costs or prohibit construction altogether
(Guaita-Pradas et. al, 2019). By identifying these features and assessing their impact on a
potential site, developers can make actionable decisions using temporally relevant data. Table 1
14

contains common impediments found within the study area, and previous research on using
OBIA for their identification.

Table 1. Common Impediments and Related Works

2.1.1. Roads
Roads are a common impediment found on sites within the TVA service territory and
cause problems to solar development interests by fragmenting the useable area on a site. The
fragmentation caused by roads is especially relevant in utility scale PV developments because it
is often necessary to create an assemblage of parcels with different owners to acquire the acreage
needed to develop a USSE project (US Department of Energy, 2012). Being that roads were put
in place without consideration for future development, their placement fragments the
developable area. This introduces the need to build around the roads, introducing gaps in the site
design where solar panels cannot be placed, leading to an increase in construction costs.
Removing or moving the location of the roadway will increase both engineering and construction
costs. Additionally, roads on private property are not readily available as a vector layer, making
identifying them individually on each site a laborious task. If OBIA can identify roads at scale
then they can be incorporated into the site screening process.
Impediment Examples Previous Research Method(s) Platform(s) Study Area
Road
• Street
• Driveway
• Access Road
• (Medhi and Saha, 2019)
• Nearest Neighbor
• Rule Based
• Multiresolution
Segmentation
Algorithm
• Resourcesat-II
• 5.8m Multispectral
• Kompsat
• 2.8m Multispectral
• 0.7m Panchromatic
• Jorhat District; Assam, India
Building
• House
• Barn
• Silo
• (Attarzadeh and Momeni, 2012)
• SEaTH
• SEperability
• THresholds
• QuickBird
• 2.4m Multispectral
• 0.6m Panchromatic
• Isfahan, Iran
Vegetation • Tree Stands
• (Chubey et al, 2006)
• (Rizvi et al, 2019)
• eCognition
• Multiresolution
Segmentation
Algorithm
• IKONOS-2
• 4.0m Multispectral
• 1.0m Panchromatic
• Rocky Mountains; Alberta,
Canada
15

2.1.2. Buildings
Buildings are impediments that contribute to construction costs for a number of reasons,
some shared and others specific to building type. Common buildings found on rural sites in the
TVA include houses, mobile homes, barns, stables, and silos. All buildings create a physical
impediment to construction, as the site must be built to accommodate their locations. Many
homes are occupied and have connecting infrastructure such as sewers lines, water lines, and the
electrical grid. The height of silos can often create large swaths of shading leading to areas with a
decreased production capacity. All buildings present unique challenges to a site and can be
constructed in short time, and the geometric shape and often uniform spectral returns make them
very compatible with OBIA methodologies.
2.1.3. Vegetation
Due to their size, the impact of utility scale solar developments on sites existing
vegetation must be considered. Developing utility scale solar involves the clearing and
disturbance of vegetation on the site. This can have negative effects on both the biological
environment and a create community opposition, resulting in the increased chance of a project
not making it to completion. These effects are manifested in a variety of ways and are related to
the type of vegetation in question.
From an environmental standpoint, the development of utility scale solar risks habitat
loss, effects groundwater runoff, pollution of local streams, and even alters the location where
aquatic insects lay their eggs (Mulvaney 2019, 171). The simple act of clearing wildflowers on a
site can negatively affect the local pollinator population. One study linked the deaths of between
16,200 and 59,400 birds in 2016 caused by the land use changes of utility scale solar
16

developments in Southern California (Walston et al, 2016). One of the most visible changes solar
can have on the landscape is the clear cutting of large swaths of trees.
The presence of trees on a site presents unique challenges with respect to the siting and
execution of a project. While solar developers traditionally target agricultural land, the conflict
between solar development and existing forests will only increase as less agricultural land
becomes available for development. For example, the Massachusetts Department of Energy
Resources estimates that 2,500 acres of trees have been cleared in the past 10-15 years within the
state to make way for solar development (LeMoult, 2019). Clear cutting can also generate
opposition within the community, leading to a project’s failure. This can also be the case even if
the clearing of the trees is not happening within the local community. Georgetown University’s
planned solar farm has faced sharp criticism from the student body for their plan to clear 240
acres of trees in in Charles County, MD to make way for the development of the project (Dance,
2019). Lastly, the costs associated with clear cutting forests to prepare a site for development,
while dependent of the thickness of the forest and its location, estimate between $3,000 and
$5,600 per acre (O’Keefe, 2020). For a site such as Georgetown University’s, this can equate to
between a $720,000 and $1,344,000 increase in construction costs.
2.2. Platforms
Selecting the platform to be used is determined by the size of the object being identified.
If the spatial resolution results in pixels larger than the object being analyzed, this requires per-
pixel or sub-pixel analysis and are not suitable for use in OBIA (Blashke, 2010). The launch of
IKONOS-1 in 1999 ushered in the era of commercially available VHR satellite imagery
providing 0.8m panchromatic and 4.0m multispectral spatial resolution (Chen and Hossain,
17

2019). Since IKONOS’ launch, advances in sensor technologies are now producing VHR images
with <1m spatial resolution, which is required for accurate OBIA.
2.2.1. Spatial Resolution
Accurate segmentation, or the grouping of individual pixels into objects, is dependent on
the spatial resolution of the input image. While image segmentation has been applied to remotely
sensed (RS) data since the launch of Landsat-1, it was the launch of IKONOS that finally
provided an opportunity for researchers to apply these techniques to VHR images (Chen and
Hossain, 2019). For the purposes of this study it is important to ascertain the spatial resolution
that will be suitable to use for segmentation on objects of varying size. Prior research has
established minimal thresholds required for accurate segmentation. Medhi and Salah (2019)
compared three segmentation techniques applied to images collected by the Resourcesat-II and
Kompsat satellites. An accuracy assessment was performed on the results and it was found that
by using a multiresolution segmentation algorithm (MRS) on Kompsat 0.7m Panchromatic
imagery resulted in 56% accuracy rate, compared to a 15% accuracy rate when using
Resourcesat-II 5.8m multispectral imagery (Medhi and Salah, 2019).
Past research has identified the minimum threshold required for segmentation and
identification. Attarzadeh and Momeni (2012) used QuickBird 2.4m multispectral imagery to
identify building outlines. Their workflow accurately identified 80% of the buildings in the
image. To achieve this accuracy, the authors ran multiple segmentation iterations to first identify
the proper scale level for segmentation. Their research also demonstrated the importance of rule
development; because of the variation in color, texture, and size, even among like objects such as
buildings, a rule designed to identify buildings with light color roofs can inadvertently exclude
buildings with dark roofs. Attarzadeh and Momeni note that their accuracy was limited as a
18

result of a rule they developed because the spectral threshold set excluded buildings with light
colored roofs. Additional research applied a semiautomated OBIA workflow to National Aerial
Image Program (NAIP) 1.0m multispectral images. This approach resulted in 95% of buildings
being accurately identified (Caggiano et. al, 2016).
Research on the use of OBIA for vegetation detection and classification has generally
been focused on plant species classification for environmental monitoring and inventory
(Blashke, 2010). Trees alone do not create an impediment to solar site construction, however,
tree stands and heavily forested areas do, as they require clearcutting and stump removal to
prepare the site for the driving of the piles and instillation of the solar panels. Chubey et. al
applied Trimble’s eCognition MRS module to IKONOS-2 4.0m multispectral imagery and were
able to identify 81 of 86 tree stand image objects in the dataset (Chubey et. al, 2006). This
suggests that 4.0m multispectral imagery can be suitable for tree stand identification using
OBIA.
2.3. OBIA Segmentation and Classification Workflows
There is a robust amount of research available on the topic of OBIA and the workflows
that provide the most accurate results in identifying specific feature within an image. This
research intended to create an iterative workflow suitable for identifying the most common
impediments found in the TVA service territory. To achieve this, a literature review was
conducted to identify workflows that have demonstrated to be able to achieve the desired
outcomes of this paper. A key determinate in selecting the workflows to be analyzed was there
was potential to be replicated using Harris Geospatial ENVI + IDL software and the
accompanying ENVI Feature Extraction Module (ENVI FX) (Xiaoying, 2009). The ENVI + IDL
and ENVI FX software will be discussed in detail in Chapter 3 of this paper.
19

2.3.1. Segmentation
The first step in OBIA is segmentation. Segmentation is the process of grouping pixels
with similar spectral properties into objects, which represent real world features. The accuracy of
the segmentation process has a direct effect on the accuracy of the classification process. The
primary segmentation algorithm used in the studies reviewed is the Multiresolution
Segmentation Algorithm (MRS). MRS has been shown to be more accurate than other
segmentation algorithms, such as the watershed transformation, by up to 18% (Kavzoglu and
Tonbul, 2017). MRS works by grouping pixels with similar spectral attributes until the variance
parameter threshold, also known as the scale parameter, is reached, at which point the
segmentation process ceases.
A workflow created by Belgui and Dragut involved using an MRS algorithm run in
eCognition to identify buildings. After the image was segmented, unsupervised classification
was ran with an overall accuracy rate of 82.3%. The same segmented image was run through a
supervised classification with overall accuracy of 86.4% (Belgui and Dragut, 2014). Being that
the difference in accuracy between the two classification methodologies is 4.1%, this workflow
shows promise for being semi-automated. MRS algorithms have shown promise not only in
identifying buildings, but across a range of features including water and trees.
MRS algorithms have been used to accurately segment water features in VHR images.
Moffett and Gorelick identified a workflow for water feature extraction using MRS ran in
eCognition. To optimize the accuracy of the segmentation process, the authors recommended
that a heavy emphasis is placed on spectral properties while de-emphasizing the object (Moffett
and Gorelick, 2012). Further research has shown that the use of MRS in OBIA workflows for
water feature extraction are more accurate that those that rely on pixel-based classification, with
one study achieving 90% accuracy (Kaplan and Avdan, 2017).
20

As previously discussed, MRS has proven to be successful in the identification of tree
stands using IKONOS-2 imagery (Chubey et al, 2006). Further research has expanded on the
work done by Chubey et al (2006) by using eCognition’s MRS algorithm on WorldView-2 0.5m
pan-sharpened imagery. While the researchers utilized imagery with considerably better spatial
resolution, their results were less accurate and attributed to over-segmentation (Sinaga et al,
2019). The over-segmentation of VHR imagery when using MRS is a problem identified in
previous research and is expected to be encountered in the course of this research (Chubey et al,
2006; Culvenor, 2003). To mitigate the problem of over segmentation, multiple segmentation
iterations are run at different scales and merge levels until the desired results are achieved. An
exhaustive literature review did not identify any automated solutions to over segmentation.
Harris Geospatial’s ENVI software offers the user the ability to preview their segmentation
results on a small subset of the image in real time, which is one reason this software was used for
this project.
2.4. Findings
The purpose of this literature review was to examine OBIA workflows and establish
which segmentation methodology would be best suited for the purpose of this research. Based on
the findings, it was determined that achieving the most accurate results across all of the listed
impediments would require sub-meter resolution multispectral images (Blashke, 2010). In all
cases, except for the identification of forest stands, it was shown that imagery with a higher
spatial resolution yield results with a higher rate of accuracy. The related works also identified
MRS to be a strong candidate to successfully segment the impediments this project focused on.
By using the ENVI + IDL and ENVI FX module, this study intends ascertain if OBIA can be
21

used to identify impediments to construction on potential utility scale solar sites at scale, to
replace the work currently performed by a human.

22

Chapter 3 Data and Methods
This project used a combination of vector data and VHR imagery to identify construction
impediments on potential USSE developments. Wetlands, FEMA 100 year floodplains, and slope
greater than 10% were erased from the parcels within the study area to remove land widely
considered unsuitable for development. The output of this process creates a mask, in the form of
a shapefile, that is used in the OBIA; this limits the segmentation and classification to only sites
that meet the minimum requirements for USSE developments. The mask created in the previous
step was uploaded, along with the VHR image of the study area, to Harris Geospatial’s ENVI.
Using ENVI’s Feature Extraction module, the image was first segmented and merged to create
objects. With the objects, rules were then created based on the attribute data created during the
segmentation process. Based on the rules, the classification process was initiated to identify all
similar objects within the image. The classification process creates a shapefile of all the objects
boundaries, which is then fed back into ArcMap to further delineate impediments within the
study area and the sample sites. This process and its inputs and outputs are described in detail
within this chapter.
3.1. Data Collection and Preparation
Prior to aggregating the data for this analysis, the researcher first identified a suitable
location for the study area within the area eliminated from future analysis via the basic site
suitability described in the last chapter. The further requirements of the study area are that it
contained a mixture of impediments on developable land within an area of less than 60 km
2
. The
size constraint on the study area was due to the costs associated with VHR imagery, which was
provided free of charge by Hexagon Geospatial via their Hexagon Imagery Program (HxIP). If
this project were forced to pay for the same image it would cost $251.48. The final study area
23

selected is 56.34 km
2
of primarily agriculture land within Weakley County, TN, and can be seen
in Figure 3. Traditionally, the first step in identifying PV sites is to map the electrical
infrastructure required for a utility scale project to interconnect. As the focus of this research was
on OBIA, this step was omitted in favor of finding an ideal study area from a topological
standpoint.
3.1.1. Data Sets and Sources
The data required for this project consisted of both vector and raster data. While the
majority of the datasets are publicly available, both the satellite imagery and parcel data are
proprietary datasets produced by private companies. Table 2 shows the datasets, their
description, and sources.
Table 2: Datasets and Sources

Dataset Data Type Description Source Location
National Wetland
Inventory
Polygon
Feature Class
This dataset contains delineated wetland
boundaries. It is produced by having a
investigator use satellite imagery to
identify and delineate wetland
boundaries and types.
US Fish
and
Wildlife Service
https://www.fws.gov/wetlands/data
National Flood
Hazard Layer
Polygon
Feature Class
This dataset contains the boundaries for
the Special Flood Hazard Area (SFHA).
SFHA are areas that have a 1% chance
of annual flooding, also known as 100-
year flood zones. This layer is produced
by FEMA for use in the Flood
Insurance Rate Map.
Federal Emergency
Management Agency
https://msc.fema.gov/portal/home
Slope
Polygon
Feature Class
This dataset is produced by merging the
study area DEMs into one raster that
covers the study area. A raster
calculator is used to group the slope into
buckets 0-5%, 5-10%, 10-15%, 15-20%,
and 20% +. The raster is then converted
into a feature class in ArcMap.
United States
Geological Survey
https://www.sciencebase.gov/catalog/i
tem/530f4226e4b0e7e46bd2c315
Parcel
Polygon
Feature Class
This proprietary dataset contains parcel
geometry and associated attribute data
such as owner name, parcel APN, tax
ID, address, use code, zoning, and
acreage.
Digital Map Products https://www.digmap.com/
Satellite Image GeoTIFF
Spatial Resolution: 30cm
Projection: NAD83
Format: GeoTIFF
AOI size: 56.34km
2
Hexagon Geospatial
https://www.hexagongeospatial.co
m/resources/resource-
library/content-providers/hexagon-
imagery-program
24

3.1.2. Wetlands
Due to their ecological sensitivity and physical characteristics, land that has been
identified as wetlands are not considered for development. The National Wetland Inventory
(NWI), produced by the US Fish and Wildlife Service (USFWS), is used to remove wetlands
from a parcel geometry. This is done by using the erase tool in ArcMap 10.6.
3.1.3. Floodplains
Floodplains, specifically Special Hazard Flood Areas (100-year floodplains), introduce
risk and delays into a project. The increased risk of flooding, and the insurance required to build
within these areas, increase project costs. These factors alone and in combination make
developing utility scale PV on floodplains impractical and difficult. Hence, these areas were also
excluded from development by using the erase tool in ArcMap 10.6.
3.1.4. Slope
The impacts that slope has on a potential site are difficult to quantify and are different
from site to site. Unlike wetlands and floodplains, there is no consensus as to the most
appropriate slope on which to develop industrial solar. For example, a site with an even 20%
slope with a southern aspect would, with the right conditions, warrant development. On the other
hand, a site that is flat but has small undulations in slope across the developable area would
require the site to be graded, increasing project costs. Similarly, a site with a steady 5% slope
with a northern aspect would be less favorable than a southern facing slope of 10%.
Additionally, the type of PV arrays used have different slope tolerances. Fixed axis arrays can be
built on undulating land, while a single axis tracker requires a site with much less undulation.
This problem is being addressed by manufactures such as Nevados Engineering who are
developing what they have coined as “all terrain trackers” (Nevados, 2020). Due to the
25

complexity of evaluating slope on a project by project basis, slope of any aspect above 10% was
removed from the parcel geometry. This seems to be the most popularly held threshold, despite
not being universal (cite).
3.1.5. Parcels
Property data, in the form of parcel geometry, was used as the basis for creating utility
scale PV sites. Because potential sites must adhere to real world boundaries, the parcels
themselves are used to create the developable area. Parcel geometry datasets can generally be
acquired from the county accessor. Weakley County, TN, where the study area is situated, only
provides parcel geometry as .pdf maps. As a result, this parcel geometry was sourced from
Digital Map Products (DMP). DMP assembles proprietary parcel datasets and provides them as a
feature class with over 300 attributes to choose from. For the purposes of this research, all that
was required was the parcel geometry and calculated acreage.
3.1.6. Satellite Imagery
The crux of this research is the satellite imagery. As previously discussed in Chapter 2,
VHR imagery was required for the success of this project. This is because this research hinges on
the ability to accurately identify site impediments on a scale beyond what is achieved through
traditional methods. The accuracy of the segmentation process is partially determined by the
spatial resolution of the imagery used. Because the costs of VHR imagery can be prohibitive,
imagery was provided by Hexagon Geospatial under an academic license. The academic license
provided one image, no larger than 60 km
2
, free of charge, with the only stipulation that it could
not be used for commercial purposes. Two images were provided, an RGB and a CIR for use.
The images have a spatial resolution of 30cm, and are provided as GeoTIFFs with an NAD83
projection.
26

3.2. Software
Two software platforms were required for this research. Esri’s ArcMap 10.6 was used to
remove undevelopable land from the parcel geometry, calculate acreage, identify potential sites,
and create the mask then used in the OBIA. Harris Geospatial’s ENVI FX was used for the
segmentation and classification process. ArcMap was provided by USC’s Spatial Science
Institute while ENVI FX was accessed using an academic license from Cloudeo.
3.2.1. Cloudeo
Cloudeo is a third party provider of Software as a Service (SaaS) and Data as a Service
(DaaS). Cloudeo was used to access ENVI FX, rather than acquiring it directly from Harris
Geospatial; this was a result of cost and license terms. Harris Geospatial only offers lifetime and
yearly license for access to ENVI, with the Feature extraction module being an additional cost. It
was estimated that it would take approximately 1 month to complete the work in ENVI making a
perpetual or yearly license was both cost prohibitive and unnecessary. Cloudeo provides 1 month
academic licenses for ENVI FX accessed through a remote desktop. This option provided this
project with ENVI FX at fraction of the costs of Harris Geospatial terms and the ability to extend
the license on a month to month basis if it was required. Table (X) below details the costs
associated with each license.
3.3. Methodology and Workflow
This section will discuss and describe in detail the methodology and steps taken to
complete the project. The workflow consists of two primary steps- data preparation and OBIA.
Data preparation was completed in ArcMap 10.6 and OBIA in ENVI FX 5.5.3.
27

3.3.1. Data Preparation
Data preparation consists of identifying potential sites using a traditional site selection
processes. Once the sites are identified, a mask is created so only areas of interest were included
during the OBIA process. Hard criteria included in the masking are discussed in sections x and y;
this creates a reduction in processing time and resource consumption when going through the
OBIA process.
3.3.1.1. Mask Creation
Mask creation was an important component to this workflow because it excludes
unwanted areas inclusion in the segmentation process resulting in faster processing times. The
process of creating the mask began in ArcMap, where wetlands, floodplains, and slope greater
than 10% were removed from a parcels geometry by using the erase tool. The output of this
process was a shapefile including only the developable area within the potential sites. The order
in which the constraints were removed does not affect the final remaining area, but does
determine the total acres lost in each category. Because of overlap in the constraints, areas that
are excluded because of wetlands will not be included in the number of acres for areas lost in the
different categories and vice versa. This research found that erasing wetlands first, followed by
floodplains, and lastly slope provided the quickest processing times. This process is outlined in
Figure 4.
28

Figure 4: Site identification and mask creation process.
The masking process was able to eliminate a combined 3010.25 acres of the study areas
13955.14 acres, or 21.57%. Of this 881.45 acres of wetlands, 1549.22 acres of floodplains, and
533.58 acres of slope were excluded leaving 10944.88 acres of developable land. This was
further reduced by only selecting sites with 100 or more developable acres remaining, leaving
29

only 1520.08 acres, or 10.76% of total study area. This results in reduced processing time during
segmentation and classification.
While a user can create a mask in ENVI, this workflow takes advantage of the fact that
the mask created is a byproduct of the site selection process itself. By using the parcel geometry
as the basis of the exclusion process, the sites remaining after removing all undevelopable land
can be exported as a shapfile and used as the mask in ENVI. This eliminates the need to repeat
this process in ENVI.
3.3.2. Segmentation
The process of OBIA is comprised of object identification and feature extraction. Object
identification beings with segmenting the image; the segmentation process defines the objects
and computes their spatial, spectral, and textural attributes. ENVI FX employs an edge-based
segmentation algorithm that, based on scale level, suppresses weak edges (Visual Information
Solutions, 2007). This is done by grouping pixels of like values into objects, which are defined
by their spectral attributes.. The boundaries of these objects are formed by the edges where there
are abrupt changes in the spectral gradient (Segmentation Algorithms background, n.d). The
scale level selected ran from 0-100, and determined the accuracy of the segmentation process. A
high scale level will result in less segmentation, conversely, a low scale level will result in
increased segmentation. Included in segmentation is the optional step of merging.
While merging is optional, it was used, with trial and error showing it yields better
segmentation results than without. Merging works by aggregating small segments that fall within
larger segments to account for over segmentation. Highly textured objects such as clouds and
trees are a cause of over segmentation.
30

Texture, in this context, is defined as the spatial variation of grayscale levels as a function
of scale (Texture Metrics Background, n.d.). The level is representative of pixel size of the box
used to compute the statistics. Thus, a scale level of 3 would equate to a 3 x 3 pixel box. The
kernel texture box creates the attribute information used in rules based classification. A box too
small will not capture enough variation among pixels for an accurate calculation. A box too large
will cause overlap leading to blending of texture across objects making creating rules based on
textural attributes unreliable.
The process began by uploading the image and mask to ENVI. The feature extraction
module will automaticity convert the shapefile into a single band raster to be used as the mask,
this can be seen in figure 5.

Figure 5: Mask utilization in ENVI
With the mask created, the scale and merge levels were set. Choosing the correct scale and
merge levels is an iterative process to identify ideal levels. This process was guided by
comparing each result to the previous iteration’s segmentation output. ENVI offers small
preview of the segmentation output based on the levels chosen prior to running the full
segmentation process. The parameters were developed by first starting with the default scale
level of 50, merge level of 0, and texture kernel size of 3. On the suggestion of the ENVI FX
31

tutorial the levels were adjusted, first in increments of 10, followed by increments of 1, and lastly
increments of 0.1 (Visual Information Solutions, 2007). This process was iterated through
numerous times, each time adjusting the parameters based on the previous iterations results until
the desired outcome was achieved. This example used a scale level of 60 and merge level of 80
and a kernel size of 3. These attributes were used to create the rules for the classification
process.
3.3.3. Classification
Rule based classification was used as it allows increased control over the classification
process. Rules are created based on the attribute information calculated in the segmentation
process using AND / OR logic. AND is used to combine multiple attributes within rule, and the
OR operator is used to combine multiple rules within one class. This process was rejected in
favor or using class and rules scores. The class score is a function of the rule score and is defined
as Class Score = ∑(Rule Score x Rule Weight). The rule score is defined as Rule Score =
∑(Attribute Score x Attribute Weight), where attribute score is the likelihood of an object meets
the conditions of an attribute. Attribute and Rule weights are defined by the user and must sum to
1 (Rule Classification Background, n.d). The rule created for this example used the spectral
mean of band 3 with a class threshold of .5 to identify tree stands. The classification process
creates a temporary .dat file containing the identified objects. This is converted a shapefile that is
fed back into ArcMap to account for the now identified impediments previously missed. This can
be seen in figure 6 below.
32

Figure 6: Object identification and export
3.3.4. Process Iteration
The methodology outlined in this section is just one example of multiple iterations completed in
the course of this research. The iteration process is time consuming, but necessary, as it can takes
a considerable amount of trial and error to identify the combination that provides the most
accurate results (Campbel and Wynne 2012, 372). Each iteration uses unique rules in an attempt
to identify those that yielded the most accurate results. The creation of the rules followed a
similar process to the development of the segmentation parameters. The rules were developed,
by first experimenting with the various attributes to identify those unique to trees. Using research
into the spectral returns of trees it was determined that spectral and textural attributes have
shown to be successful in identifying trees (Lin et al, 2013). Similarly to the preview window
provided in the segmentation process, ENVI also has a preview window that can be used to
display a rule confidence image while adjusting the rule parameters. The rule confidence image
demonstrates the relative confidence of an object belonging to a feature, the brighter the color the
greater the confidence, and vice versa (Visual Information Systems, 2007). As the rule
parameters were adjusted, the preview window was used to visualize the results of each subtle
33

change, allowing for on the fly adjustments and not having the run the entire classification
process to glean insight into the results.

34

Chapter 4 Results
The intent of this research was to discern if OBIA and rule-based classification could accurately
replicate and/or improve upon a manual utility scale solar site suitability prescreen workflow. To
achieve this, multiple iterations of the process were run, with adjustments made to the
segmentation parameters and/or the rules, in an effort to identify the most accurate combination
to identify tree stands within the study area. This chapter will discuss the results of each iteration,
the rules employed, and the justification for sampling technique and size.
4.1. Segmentation
As previously discussed, the accuracy of the classification process is largely influenced
by the accuracy of the segmentation process. For the purposes of identifying tree stands, the
scale level, set between 0-100, should be set at a level low enough to properly delineate trees
without causing over segmentation. To identify the correct scale and merge levels, 47
segmentation iterations were run, beginning with a scale level of 50 and a merge level of 0. A
general rule of thumb is, as the scale level decreases, the merge level should conversely increase
(ENVI, 2008). This process was assisted by ENVI’s segmentation preview window which allows
the user to view a subset of the results prior to initiating the segmentation process. As previously
noted, the selection of ENVI for this research was partially related to this specific feature. After
multiple iterations, the combination that most accurately delineated the trees from their
surroundings was iteration 7; with a scale level of 37, merge level of 97.5, and texture kernel size
of 3. The 7 iterations to be discussed can be found in Table 3 below. Once the segmentation
process was providing accurate results, only small adjustments of the input values were required
to alter the segmentation output. The resulting outputs required adjustments to the classification
35

rules to account for changes in the spectral, textural, and spatial attributes of the segmentation
image. The results of each segmentation can be found in the appendix.
Table 3: Segmentation Parameters

4.2. Rules
Similar to the segmentation process, the rules created for rule based classification were a
result of an iterative process guided biophysical characteristics of trees found within the study
area. A review of the attribute information generated during the segmentation process found that
the spectral and textural attributes of the trees were the most homogenous across the class. A
review of the spatial attributes found that there was too large of a range between the spatial
attributes of a single or small group, or trees to that of a large tree stand. As such, spatial
attributes were not used in the creation of the classification rules.
4.2.1. Rule Development Process
The development of the rules began by first identifying the biophysical characteristics
that make trees unique from their surroundings. The rules must be able to discern green grass
from green trees, or dark shadows from dark tree canopies, which can share similar spectral
returns for example. To achieve this, the rules were developed using an iterative process that
relied on the previous results to inform the changes required to further refine the rules. Take for
example, the rule used in the seventh iteration, seen below in table 4.
Iteration Scale Level Merge Level Texture Kernel Size
1 23 98 5
2 23 98 3
3 23 95 3
4 28.5 95 3
5 38 95 3
6 38 97 3
7 37 97.5 3
36

Table 4: Rule 7

Rule 7 uses 3 separate attributes in combination to define a tree. In this example, texture
entropy is being used to help differentiate between green, but otherwise smooth, grass from a
green, but highly textured tree canopy. When ran in combination, the results, seen in figure 7,
show how the rule was able to correctly classify trees while minimizing the misclassification of
grass as trees.

Figure 7: Results of Rule 7
When rule 7 is ran again, this time removing the texture entropy constraint the results are
vastly different with large swathes of not only grass, but also barren land misclassified as trees.
This can be seen in figure 8 below.
RULE 7 Min Max Band
Texture Entropy -0.59416 -0.53717 2
Spectral Max 66.71148 112.14201 3
Spectral Mean 22.21387 89 1
Class Threshold 0.75
37

Figure 8: Rule 7 Modification Results
This process also demonstrates how important a single constraint can be in a rule, as well as
inform the creator of the rule what constraints were eliciting the changes made when ran in
combination.

4.3. Accuracy Assessment
To assess the accuracy of each iteration, 545 randomly sampled ground truth points were
generated and manually classified as either 1 for trees or 0 for unclassified. This was done using
the same 30cm VHR image employed in the segmentation and classification process to eliminate
errors of registration (Campbell and Wynne, 2012, 416). The ground truth points were used to
generate error matrixes, also known as confusion matrixes, for each iteration by comparing the
reference, or ground truth data, to the classification results.
38

The accuracy of the results is the determining factor in judging this exercise to be a
success or not, as such the accuracy assessment employed was based on the framework created
by Alan Hay. While the confusion matrix provides the overall accuracy, this value alone does not
provide the context required to understand the results. To determine the accuracy, Hay proposes
using the error matrix to answer the following questions (Hay, 1979.)
1. What proportion of all the sample predictions proved to be correct?
2. What proportion of the sample predictions of a single category proved to correct?
3. What proportion of land, within a category, is correctly predicted?
4. Is the net effect, of numbers 2 and 3 above, for predictions to overestimate or
underestimate a given category?
5. If error occurs in either of the ways 2 and 3, is there any bias in these errors towards
specific categories?
4.3.1. Process Iteration
The need to run multiple iterations of both the segmentation and classification process to
identify the ones that yield the most accurate results is a common theme found the related works.
The ideal OBIA workflow put forth by Blaschke et al incorporates the iterative nature of the
process into the workflow demonstrating the need to iterate the process to refine the results.
Figure 7 outlines this workflow and shows that both the segmentation and classification steps
within the workflow are iterated through.
39

Figure 9: Example of the iterative nature of the idealized OBIA workflow (Blaschke et al 2012,
186)
The benefit of iterating though each step is it refines the outputs, the theory being that
each iteration will guide the GIS technician towards the most accurate results based on their
interpretations. A drawback of this process is that it requires iteration, and it cannot be assumed
that accurate results are achieved without it. This coupled with the differences in spectral,
textural, and spatial properties across VHR images makes it unfeasible to establish segmentation
parameters and classification rule sets that can be applied across different image sets. This
40

limitation and its effects on the ability to replicate the work presented in this chapter is discussed
in detail in Chapter 5.
4.3.2. First Iteration
The first iteration in this sequence was the result of multiple trial and error efforts to
identify an acceptable baseline to attempt to improve upon. Table 4 provides the segmentation
parameters, rule, and confusion matrix for the first iteration. The rule created to classify trees
relied on the spectral mean and textural range of band 3.
Table 5: First Iteration

The first iteration provided an overall accuracy of 0.9009, but only had a producer’s
accuracy (P_Accuracy) (also known as errors of omission) of 0.6886. In other words, trees were
only properly identified 68.86% of the time. This is not an acceptable level of accuracy if this is
to be used for actionable decision making. Additionally, the Kappa value of 0.7456 demonstrates
that the results of this iteration would achieve an accuracy that is 75.56% better than what would
be expected from a chance assignment of ground truth points to categories (Campbell and
Wynne, 2012, 420). A subset of the classification results can be seen below in Figure 8.
Iteration Scale Level Merge Level Texture Kernel Size
1 23 98 5
RULE 1 Min Max Band
Spectral Mean 33.68989 60 3
Texture Range 32 58.67765 3
Class Threshold 0.75
Class Name Unclassified Trees Total U_Accuracy Kappa
Unclassified 376 52 428 0.8785 0
Trees 2 115 117 0.9829 0
Total 378 167 545 0 0
P_Accuracy 0.9947 0.6886 0 0.9009 0
Kappa 0 0 0 0 0.7456
41

Figure 10: Classification results of the first iteration, the green represents areas classified as
trees, the grey represents masked areas.
4.3.3. Second Iteration
In an attempt to create more granular textural attributes, the second iteration reduced the
texture kernel size to 3, with all other segmentation parameters remaining the same (Table 5). To
account for the changes in the segmentation attribute, the range of the spectral mean was
decreased to 38.43266 – 59.7500. This did not achieve the desired results with the overall
accuracy reduced to 0.8092, the producer accuracy to 0.5389, and the kappa value to 0.5101. The
results of this change can be seen in figure 9; note the misclassified grasses adjacent to trees as
well as less trees being correctly classified.
42

Figure 11: Second iteration classification results, green represents areas classified as trees, green
represents masked areas.
Table 6: Second Iteration

Iteration Scale Level Merge Level Texture Kernel Size
2 23 98 3
RULE 2 Min Max Band
Spectral Mean 38.43266 59.75 3
Texture Range 32 58.67765 3
Class Threshold 0.75
Class Name Unclassified Trees Total U_Accuracy Kappa
Unclassified 351 77 428 0.8201 0
Trees 27 90 117 0.7692 0
Total 378 167 545 0 0
P_Accuracy 0.9286 0.5389 0 0.8092 0
Kappa 0 0 0 0 0.5101
43

4.3.4. Third Iteration
For the third iteration, the merge level was lowered to 95 with all other segmentation
parameters remaining the same. In an attempt to increase the accuracy of Rule 2, the texture
range was for band 2 was set to > 32.28139 as seen in table 6. This was done to attempt to
exclude grassy areas, which had similar spectral properties, but a lower textural range. This
change again reduced the overall accuracy, the producer accuracy, and the kappa value.
Table 7: Third Iteration

4.3.5. Fourth Iteration
For the fourth iteration, the scale level was increased to 28.5 with all other segmentation
parameters remaining the same. For this iteration, a new rule was created to focus on the textural
entropy of band 2 and the spectral max of band 3 (Table 7). These changes were implemented as
previous iterations were becoming less accurate when the rule parameters were adjusted. This
change in direction was added via the ENVI preview window, which allowed for “on the fly”
viewing of the classification results. This change increased accuracy over the previous iteration,
but was still less accurate than the baseline set in the first iteration.
Iteration Scale Level Merge Level Texture Kernel Size
3 23 95 3
RULE 3 Min Max Band
Spectral Mean 38.43266 59.75 3
Texture Range >32.28139 2
Class Threshold 0.75
Class Name Unclassified Trees Total U_Accuracy Kappa
Unclassified 378 114 492 0.7683 0
Trees 0 53 53 1 0
Total 378 167 545 0 0
P_Accuracy 1 0.3174 0 0.7908 0
Kappa 0 0 0 0 0.3921
44

Table 8: Fourth Iteration

4.3.6. Fith Iteration
After seeing a positive correlation between the changes made in the fourth iteration and
accuracy levels, the scale level was again increased, this time to 38, with all other segmentation
parameters remaining the same (Table 8). Further review of the classification results of the
fourth iteration identified that the RULE 4, in an attempt to exclude shadows, also excluded dark
trees. RULE 5 accounts for this by including the new attribute of Band 1 Spectral Mean <
74.75175. With this change, the spectral max of band 3 was similarly adjusted, increasing the
minimum to 66.71148 and the maximum to 112.14201. This change increased the overall
accuracy to 0.8826, the producer accuracy to 0.7126, and the kappa value to 0.7081, which
improves upon the fourth iteration.

Iteration Scale Level Merge Level Texture Kernel Size
4 28.5 95 3
RULE 4 Min Max Band
Texture Entropy -0.59416 -0.53717 2
Spectral Max 65.13619 101.27068 3
Class Threshold 0.75
Class Name Unclassified Trees Total U_Accuracy Kappa
Unclassified 377 83 460 0.8196 0
Trees 1 84 85 0.9882 0
Total 378 167 545 0 0
P_Accuracy 0.9974 0.5030 0 0.8459 0
Kappa 0 0 0 0 0.5798
45

Table 9: Fifth Iteration

4.3.7. Sixth Iteration
In the sixth iteration, the merge level was increased to 97 with all other segmentation
parameters remaining the same. Similarly to the results of the fifth iteration, the sixth iteration
again excluded to many dark trees in an effort to exclude shadows. To manage this, the spectral
mean was increased to < 88.23574. This change had the most significant positive change to the
accuracy of the classification. Overall accuracy was increased to 0.9174, producer accuracy to
0.8862, and kappa value to 0.8080 (Table 9).

Iteration Scale Level Merge Level Texture Kernel Size
5 38 95 3
RULE 5 Min Max Band
Texture Entropy -0.59416 -0.53717 2
Spectral Max 66.71148 112.14201 3
Spectral Mean < 74.75175 1
Class Threshold 0.75
Class Name Unclassified Trees Total U_Accuracy Kappa
Unclassified 362 48 410 0.8829 0
Trees 16 119 135 0.8815 0
Total 378 167 545 0 0
P_Accuracy 0.9577 0.7126 0 0.8826 0
Kappa 0 0 0 0 0.7081
46

Table 10: Sixth Iteration

4.3.8. Seventh Iteration
For the seventh iteration, the merge level was increased to 97.5 with all other
segmentation parameter remaining the same. Upon inspection of the results of the sixth iteration,
it was found that shadows were again being included and classified as trees. In an attempt to
mitigate this, the spectral mean of band 1 was changed from < 88.23574 to 22.21387 – 89.00, as
seen in Table 10. This was surprising because there was no change in the confusion matrix and
thus no change in accuracy.

Iteration Scale Level Merge Level Texture Kernel Size
6 38 97 3
RULE 6 Min Max Band
Texture Entropy -0.59416 -0.53717 2
Spectral Max 66.71148 112.14201 3
Spectral Mean <88.23574 1
Class Threshold 0.75
Class Name Unclassified Trees Total U_Accuracy Kappa
Unclassified 352 19 371 0.9488 0
Trees 26 148 174 0.8506 0
Total 378 167 545 0 0
P_Accuracy 0.9312 0.8862 0 0.9174 0
Kappa 0 0 0 0 0.8080
47

Table 11: Seventh Iteration

4.4. Conclusions
The hypothesis underlying this research was that OBIA can be used to accurately
replicate a manual USSE site review workflow. While each real world use case brings with it its
own unique environmental properties, the results of this research shows the potential for the use
of OBIA in USSE site suability analyses. A key finding is that the measurement of accuracy
must be put in the context of the desired outcome, namely with respect to what is determined to
be an acceptable error rate. For the purposes of using the classification outputs to make
actionable decisions a 90% accuracy rate would be desired. While an overall accuracy of 91.74%
was achieved, no more than 88.62% of the trees were correctly classified, resulting in a kappa
value of 0.8080. With more experience and iterations, it is assumed that the accuracy of a
classification technique such as this can be increased, but the measurement of the accuracy will
always be subjective as it is up to the end user to determine their confidence in the results.
Iteration Scale Level Merge Level Texture Kernel Size
7 38 97.5 3
RULE 7 Min Max Band
Texture Entropy -0.59416 -0.53717 2
Spectral Max 66.71148 112.14201 3
Spectral Mean 22.21387 89 1
Class Threshold 0.75
Class Name Unclassified Trees Total Comission Kappa
Unclassified 352 19 371 0.9488 0
Trees 26 148 174 0.8506 0
Total 378 167 545 0 0
Omission 0.9312 0.8862 0 0.9174 0
Kappa 0 0 0 0 0.8080
48

Chapter 5 Discussion and Conclusions
Chapter 5 discusses the results of the OBIA process, limitations of the study, and recommends
areas of further research. This project set out to investigate the use of OBIA to augment or
replace the time consuming process of visually inspecting each potential USSE site for
construction impediments. The results of this research demonstrated that OBIA using VHR
imagery can identify trees with 88.62% accuracy within the study area. This number was
achieved by running multiple iterations of OBIA, each time adjusting either the segmentation
parameters or class rules based on the previous iteration to increase accuracy.
5.1. Findings
The goal of this research was to identify if OBIA can be used to accurately identify
construction impediments on potential USSE sites, and if so, if a workflow could be developed
to achieve similar outcomes as if the sites were visually reviewed by a GIS technician. OBIA
was selected because it has the ability to apply logic that mimics some of the higher order logic
employed by GIS technicians (Campbell and Wynn, 2012). Success in this case is determined by
two factors; the accuracy of the OBIA results and the ability to use OBIA, at scale, to replicate
the site suitability analysis currently completed using human intervention.
The level of accuracy required to confidently make actionable decisions on is subjective
and determined by the needs and requirements of the end user. For this project, the desire was to
achieve a classification accuracy of greater than or equal to 90%. While an overall accuracy of
91.74% was achieved, the greatest accuracy achieved in identifying trees was 88.62%, which, in
this case, would be below the required threshold. With further iterations and development of
classification rules, it would be assumed that accuracy could be increased. This research has
49

further demonstrated that, when accurate rules are developed, OBIA can be applied to
homogenous areas to screen sites for unfavorable conditions and features. While there is
considerable upfront investment in time to develop the rules for classification, once they are
established they can be reused in future land acquisition campaigns within the same territory to
generate similar returns.
5.2. Applications and Benefits
The guiding idea behind the use of OBIA in USSE was that it had the capacity to
improve upon existing approaches to site selection, which are not scalable, time consuming, and
prone to human error. To assess the real world applications of this research requires complex
cost benefit analysis to determine if there is a favorable return on investment. This must be
conducted at the level of an individual organization to consider their ability to fund the
acquisition costs associated with VHR images and the specialized software required for OBIA.
As each organization has a different approach to quantifying and justifying their investments in
GIS, this research focused on the overall types of benefits that can come from the application of
this technology (Croswell, 2009). To evaluate the potential benefits of using OBIA in USSE site
selection a list of categories and descriptions created by Peter L Croswell to assess the impact of
GIS was used.

50

Table 12: Potential benefits gained from the use of OBIA in USSE site selection
(Croswell, 2009).

As seen in table 11, the benefits that can be realized by using OBIA to assist in the site selection
process manifest in different ways. OBIA can contribute to operational efficiency gains, cost
avoidance, revenue enhancement, and qualitative benefits. Cost savings, Non-Monetary
quantitative benefits, and difficult to predict benefits are can only be calculated by the
organization itself, as there are too many unknowns to make accurate predictions in these
Category Description OBIA Benefits
Operational Efficiency Gains
Expected Gains in current personnel efficiency and
productivity allowing work to be accomplished in
less time.
• Using OBIA to assist in the site selection
process works as a force multiplier allowing
more sites to be reviewed in less time and with
less personnel.
Cost Savings
Reduction in current expenses such as contract
costs and salaries.
• Unknown and calculated per organization.
Cost Avoidance
Reducing or eliminating costs that would be
incurred without the use of GIS technology, when
new programs, regulatory requirements, or other
new demands are placed on an organization.
• Using OBIA to asses, at scale, site conditions in
the early phases of development results in
more accurate cost estimations.
• Early identification of possible construction
impediments can disqualify a potential site
before more labor hours are invested in its
development.
Revenue Enhancement
Use of GIS technology and data in a manner that
results in increased revenue from existing or new
sources.
• Enhancement of unrealized future operating
revenue by decreasing capital expenditures for
site construction by favorable altering expected
return on investment.
Non-Monetary Quantitative
Benefits
Potential benefits that can be measured
quantitatively but do not translate precisely into
monetary terms.
• Unknown and realized on a per organization
level.
Difficult to Predict Benefits
Benefits that are driven by external events and
thus are not easily predictable or routine in nature
and that are not easily reflected in a return on
investment analysis.
• Unknown and realized on a per organization
level.
Qualitative Benefits
Benefits that are not easily quantified yet have a
positive impact on operations, decision making,
quality of service, social conditions, or economic or
environmental health.
• Quantified outputs of the OBIA process
creates data that is relevant to the organization
and can be shared to assist with decision
making.
51

categories. Any organization that intends to implement OBIA into their current operations should
conduct a thorough cost benefit analysis beforehand to accurately assess the impact it may have
on their operations.
This research’s intent was to validate the application of OBIA to assist in USSE site
selection processes, and provide a conceptual framework which an organization can use to
identify if OBIA has the potential to improve the site selection process. Due to OBIA being
highly influenced by the specific inputs used, it will be difficult to exactly replicate the results of
this research without identical inputs. While this can be done, it would be more advantageous for
an organization to first asses their current, if any, RS data to identify if it can be used for OBIA
or if VHR imagery must be acquired. By then following the framework provided in this research,
an organization can assess if they can achieve their desired accuracy, and if the investment of
capital required to extract value from OBIA is warranted.
5.3. Limitations
This project identified numerous limitations in the use of OBIA for the purposes of
identifying construction impediments. Some of the limitations of OBIA, such as the need for
VHR imagery, were expected and factored into the initial workflow. Unexpected limitations of
OBIA became evident during the classification process.
5.3.1. Temporal
OBIA relies on VHR imagery as its primary input, and the temporal relevance of the
image is extremely important to achieving accurate results. Rules written to classify trees during
the summer months would struggle to classify the same trees in fall or winter. The seasonal
transformation that many types of vegetation go through changes their physical appearance
which for the purposes of OBIA would require that rules are specific to seasonal changes. For
52

example, a deciduous tree loses it leaves in winter, resulting in a change in the spectral returns
because of the lack of green leaves. These changes also effect, in a similar fashion, the textural
and spatial attributes of an object. Additional seasonal changes, such as snow or drought, will
also affect classification rules by altering the physical landscape. For example if the image being
used is captured in winter, there can be snow on the ground and vegetation has gone dormant,
this changes the physical characteristics of vegetation and as result their spectral returns.
5.3.2. Regional
Similar to seasonal changes, region is also a factor. Changes in the physical environment
across a large utility territory would need to be accounted for in rule development. A service
territory such as the Electric Reliability Council of Texas (ERCOT) that spans the entire state of
Texas presents vastly different environments depending on the location of the state. Rules
developed for classification in the south east of the state, where there are vast wetlands and
rivers, would not necessarily apply to the arid lands of the Permian basin. This limitation
constrains the rules to the region they were designed for, making it difficult to create a ruleset
that would generate accurate results across a large area. This limitation can be mitigated, but
only by breaking down large regions into more homogeneous subregions. This would require
additional labor, but presumably would only need to be conducted once.
5.3.3. Complexity
OBIA’s ability to classify multiple objects of different origin in one pass is a result of a
complex process that means investigating the spectral, textural, and spatial attributes of an object
of interest. The development of the rules requires an understanding of remote sensing and
working with VHR imagery. The primary investigator required 3 months of studying, practicing,
and trial and error until they had a confident understanding of how the process worked, and how
53

to the adjust segmentation process and the rules to elicit the desired results. Even after spending
considerable time working with ENVI FX and studying OBIA and rule based classification there
is still much more to be learned to fully understand the process and underlying science. Because
of the complexity of OBIA, it would require an employee with knowledge of the process to
develop and implement the workflow, which is not a skill inherent in all GIS technicians.
The complexity of OBIA leads to another limitation imposed by the process, which is the
upfront labor required to develop the rules. As discussed, OBIA is an iterative process and
requires trial and error when identifying the correct scale and merge levels, and developing the
rules. This research focused on one impediment, trees, and spent 3 months researching to create
the simple rules employed in this project. To be completed at scale and capture the totality of a
site’s conditions, vastly more complex rules would need to be developed to identify a multitude
of different features each with their own unique attributes. This would require considerable
upfront capital expenditures and labor to achieve, something not always on hand.
5.3.4. Labor Investments
One noted limitation of this process is how labor intensive it is to set up. The requirement
to understand the biophysical characteristic of the target area local environment is necessary to
accurately segment and classify an image. This is further compounded by the limitations listed
previously, which dictate that this process must be redone for each region of interest. The total
man hours required to begin to achieve accurate results would, in this case, negate any time
savings incurred by using OBIA in the screening process. One key benefit that arises from
investing the time to develop rules is that they can be reused in the future should an area be
revisited.
54

5.4. Areas of Further Research
This project intended to classify a multitude of features such as trees, roads, wetlands,
and buildings. In the initial stages of the research it was determined that this would be difficult to
achieve in the time required for this project. As such, a focus was placed on developing rules for
only one class as a proof of concept. As such, there is still much research to be done to fully
develop the ideas presented in this paper. This section discusses the ways in which this can be
improved upon and further developed.
5.4.1. Objects
The first area that warrants further research is developing rules for the numerous site
conditions that can be encountered. While this paper only focused on a handful of features and
developed rules for only one, for OBIA to be truly effective in assisting site suability analysis
workflows it must be able to identify all features or conditions that create impediments to
construction. For this to be achieved properly, the individual developing the rules must
understand the prevailing site conditions in the region and the features that are expected to be
encountered. This process requires an upfront investment in time to study an area prior to
developing the rules. Without this upfront investment, the process described here is unlikely to
be successful.
5.4.2. Scalability
For OBIA to be successful it must be scalable. This relates to both area of interest and
computational resources. The study area selected for this study was limited to 60 km
2
as a
condition of the vendor who provided the VHR images for this project. For context, the
Tennessee Valley Authority is approximately 207199 km
2
, meaning only approximately .0002%
of the service territory was included for analysis. To scale this workflow to cover a whole
55

service territory would require an immense amount of capital to acquire VHR images for full
area coverage. Working with a dataset of this size would require immense computational
resources to process. These two factors would make this endeavor cost-prohibitive for many
users.
5.5. Conclusion
There is a growing consensus that clean, renewable sources of energy are necessary to
address global energy needs and address climate change. Solar energy provides passive clean
energy generation but not without costs. Utility scale solar requires large areas of land, can
displace local flora and fauna, encounters push-back from communities and environmental
groups, and requires large capital expenditures. This makes it all the more important to find new
ways to identify suitable locations for solar development.
This research explored the use of OBIA to assist in site suitability identification with a
desire to eventually scale the process to be used on a large scale. While OBIA provided
promising results, it also presented a number of obstacles that makes its use, at scale, a difficult
but worthwhile endeavor. This research lays a framework that can be built upon by anyone
willing to invest the time and resources, and identifies the opportunities and challenged in doing
so.

56

References
Attarzadeh, R., and M. Momeni. 2012. "Object-Based Building Extraction from High Resolution
Satellite Imagery." ISPRS - International Archives of the Photogrammetry, Remote
Sensing and Spatial Information Sciences XXXIX-B4 (July 27, 2012): 57-60.
doi:10.5194/isprsarchives-xxxix-b4-57-2012.

Belgiu, Mariana, and Lucian Drǎguţ. 2014. "Comparing Supervised and Unsupervised
Multiresolution Segmentation Approaches for Extracting Buildings from Very High
Resolution Imagery." ISPRS Journal of Photogrammetry and Remote Sensing 96 (2014):
67-75. doi:10.1016/j.isprsjprs.2014.07.002.

Blaschke, Thomas, Stefan M. Lang, and Geoffrey J. Hay. 2008. Object-based Image Analysis:
Spatial Concepts for Knowledge-driven Remote Sensing Applications. Berlin: Springer-
Verlag, 2008.

Blaschke, Thomas, Kasper Johansen, and Dirk Tiede. 2011."Object-Based Image Analysis for
Vegetation Mapping and Monitoring." Advances in Environmental Remote Sensing
Remote Sensing Applications Series, 2011, 241-71. doi:10.1201/b10599-13.

Brewer, Justin, Daniel P. Ames, David Solan, Randy Lee, and Juliet Carlisle. 2015. “Using GIS
Analytics and Social Preference Data to Evaluate Utility-Scale Solar Power Site
Suitability.” Renewable Energy 81 (2015): 825–36.
https://doi.org/10.1016/j.renene.2015.04.017.

Caggiano, Michael D., Wade T. Tinkham, Chad Hoffman, Antony S. Cheng, and Todd J.
Hawbaker. 2016. "High Resolution Mapping of Development in the Wildland-urban
Interface Using Object Based Image Extraction." Heliyon 2, no. 10 (2016).
doi:10.1016/j.heliyon.2016.e00174.

Campbell, J.B, and R.H Wynne. 2012. Introduction to Remote Sensing Fith Edition. New York:
Guilford Press, 2012.

Charabi, Yassine, and Adel Gastli. 2011. “PV Site Suitability Analysis Using GIS-Based Spatial
Fuzzy Multi-Criteria Evaluation.” Renewable Energy 36, no. 9 (2011): 2554–61.
https://doi.org/10.1016/j.renene.2010.10.037.

Chubey, Michael S., Steven E. Franklin, and Michael A. Wulder. 2006. "Object-based Analysis
of Ikonos-2 Imagery for Extraction of Forest Inventory Parameters." Photogrammetric
Engineering & Remote Sensing 72, no. 4 (2006): 383-94. doi:10.14358/pers.72.4.383.

Croswell, Peter L. 2009. The GIS Management Handbook. Frankfurt, KY: Kessey Dewitt
Publications

57

Curier, R.L. 2020. “Monitoring Sustainable Development: Climate and Energy Policy
Indicators.” Journal of Sustainability Research 2, no. 3 (2020).
https://doi.org/10.20900/jsr20200027.

Dance, Scott. 2019. “Go solar, or save the trees? Georgetowns University solar farm would clear
240-acres forest in Charles County.” January 31, 2019.
https://www.baltimoresun.com/news/environment/bs-md-georgetown-solar-trees-
20190131-story.html

“Eco-Regions of Tennessee .” . N.D .Eco-regions of Tennessee . USGS. Accessed June 13, 2020.
https://store.usgs.gov/assets/MOD/StoreFiles/Ecoregion/21632_tn_front.pdf.

Effat, Hala A. 2013. “Selection of Potential Sites for Solar Energy Farms in Ismailia
Governorate, Egypt Using SRTM and Multicriteria Analysis.” International Journal of
Advanced Remote Sensing and GIS 2, no. 1 (August 22, 2013): 205–20.
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.899.2425&rep=rep1&type=pdf

ENVI. 2008. ENVI Feature Extraction Module User’s Guide. 2008. Broomfield, CO. Harris
Geospatial.
http://www.harrisgeospatial.com/portals/0/pdfs/envi/feature_extraction_module.pdf

Environmental Protection Agency. n.d. “Ecoregions of Tennessee” Accessed June 13, 2020.
https://www.epa.gov/eco-research/ecoregion-download-files-state-region-4

Garcia-Pedrero, Angel; Gonzalo-Martín, Consuelo; Lillo-Saavedra, Mario; Rodríguez-
Esparragón, Dionisio. 2018. "The Outlining of Agricultural Plots Based on
Spatiotemporal Consensus Segmentation." Remote Sens. 10, no. 12: 1991.

Garni, Hassan Z. Al, and Anjali Awasthi. 2018. “Solar PV Power Plants Site Selection.”
Advances in Renewable Energies and Power Technologies, 2018, 57–75.
https://doi.org/10.1016/b978-0-12-812959-3.00002-2.

Gašparović, Iva, and Mateo Gašparović. 2019. “Determining Optimal Solar Power Plant
Locations Based on Remote Sensing and GIS Methods: A Case Study from Croatia.”
Remote Sensing 11, no. 12 (2019): 1481. https://doi.org/10.3390/rs11121481.

Georgiou, Andreas, and Dimitrios Skarlatos. 2016. “Optimal Site Selection for Sitting a Solar
Park Using Multi-Criteria Decision Analysis and Geographical Information Systems.”
Geoscientific Instrumentation, Methods and Data Systems 5, no. 2 (2016): 321–32.
https://doi.org/10.5194/gi-5-321-2016

Guaita-Pradas, Inmaculada, Inmaculada Marques-Perez, Aurea Gallego, and Baldomero Segura.
"Analyzing Territory for the Sustainable Development of Solar Photovoltaic Power
Using GIS Databases." Environmental Monitoring and Assessment 191, no. 12 (2019).
doi:10.1007/s10661-019-7871-8.

58

Hays, Alan. 1979. Sampleing Designs to Test Land Use Map Accuracy. Photogrammetric
Engineering and Remote Sensing., Vol 45, pp. 529-533

Hossain, Mohammad D., and Dongmei Chen. 2019 "Segmentation for Object-Based Image
Analysis (OBIA): A Review of Algorithms and Challenges from Remote Sensing
Perspective." ISPRS Journal of Photogrammetry and Remote Sensing 150 (2019): 115-
34. doi:10.1016/j.isprsjprs.2019.02.009.

Kavzoglu, Taskin, and Hasan Tonbul. 2017. "A Comparative Study of Segmentation Quality for
Multi-resolution Segmentation and Watershed Transform." 2017 8th International
Conference on Recent Advances in Space Technologies (RAST), June 19, 2017.
doi:10.1109/rast.2017.8002984.

“Landsat Mission.”. N.D. . Landsat Level-1 Processing Details. USGS . Accessed June 12, 2020.
https://www.usgs.gov/land-resources/nli/landsat/landsat-level-1-processing-details.

LeMoult, Craig. 2019. “Some Massachusetts Forestland is Being Clear-cut to put up Solar
Farms” April 26, 2019. https://www.wgbh.org/news/local-news/2019/04/26/some-
massachusetts-forestland-is-being-clear-cut-to-put-up-solar-farms

Li, Dongrong. 2013. “Using GIS and Remote Sensing Techniques for Solar Panel Installation
Site Selection.” Using GIS and Remote Sensing Techniques for Solar Panel Installation
Site Selection. UWSPACE, 2013. https://uwspace.uwaterloo.ca/handle/10012/7960.

Lin, Yi, Eetu Puttonen, and Juha Hyyppä. "Investigation of Tree Spectral Reflectance
Characteristics Using a Mobile Terrestrial Line Spectrometer and Laser
Scanner." Sensors 13, no. 7 (2013): 9305-320. doi:10.3390/s130709305.

Lomax, K. 1943. THE TENNESSEE VALLEY AUTHORITY : An Experiment in Regionalism.
Nature 151, 592–593 (1943). https://doi.org/10.1038/151592a0

Malof, Jordan M., Kyle Bradbury, Leslie M. Collins, and Richard G. Newell. 2016. “Automatic
Detection of Solar Photovoltaic Arrays in High Resolution Aerial Imagery.” Applied
Energy 183 (2016): 229–40. https://doi.org/10.1016/j.apenergy.2016.08.191.

Medhi, Ankita & Saha, Ashis. 2019. Rural Road Extraction using Object Based Image Analysis
(OBIA): A case study from Assam, India. Advances in Cartography and GIScience of the
ICA. 1. 1-8. 10.5194/ica-adv-1-13-2019.

Moffett, Kevan B., and Steven M. Gorelick. 2012 "Distinguishing Wetland Vegetation and
Channel Features with Object-based Image Segmentation." International Journal of
Remote Sensing 34, no. 4 (2012): 1332-354. doi:10.1080/01431161.2012.718463.

Mulligan, Ean. 2020. Interviewed by author September 16, 2020.

59

Mulvaney, Dustin. Solar Power: Innovation, Sustainability, and Environmental Justice.
Oakland, CA: University of California Press, 2019.

Nevados Engineering “All terrain trackers”. 2020. Accessed September 12, 2020.
https://nevados.co/

Nazari, Mohammad Alhuyi, Alireza Aslani, and Roghayeh Ghasempour. 2018. “Analysis of
Solar Farm Site Selection Based on TOPSIS Approach.” International Journal of Social
Ecology and Sustainable Development 9, no. 1 (2018): 12–25.
https://doi.org/10.4018/ijsesd.2018010102.

Omitaomu, Olufemi A., Nagendra Singh, and Budhendra L. Bhaduri. 2015. “Mapping Suitability
Areas for Concentrated Solar Power Plants Using Remote Sensing Data.” Journal of
Applied Remote Sensing 9, no. 1 (2015): 097697. https://doi.org/10.1117/1.jrs.9.097697

Poursanidis, Dimitris, Nektarios Chrysoulakis, and Zina Mitraka. 2015. “Landsat 8 vs. Landsat
5: A Comparison Based on Urban and Peri-Urban Land Cover Mapping.” International
Journal of Applied Earth Observation and Geoinformation 35 (2015): 259–69.
https://doi.org/10.1016/j.jag.2014.09.010.

Rizvi, R. H., R. Newaj, S. Srivastava, and M. Yadav. 2019. "Mapping Trees On Farmlands
Using Obia Method And High Resolution Satellite Data: A Case Study Of Koraput
District, Odisha." ISPRS - International Archives of the Photogrammetry, Remote
Sensing and Spatial Information Sciences XLII-3/W6 (2019): 617-21. doi:10.5194/isprs-
archives-xlii-3-w6-617-2019.

Renewable Resource Coalition. n.d. “About Renewable Resources”. Accessed June 16, 2020.
"Solar Energy Disadvantages: The Top Drawbacks of Solar Power." Renewable
Resources Coalition. https://www.renewableresourcescoalition.org/solar-energy-
disadvantages/.
“Rule Classification Background”. n.d L3Harris Geospatial. Accessed October 13, 2020.
https://www.l3harrisgeospatial.com/docs/BackgroundRuleClassification.html#Attribut

“Segmentation Algorithm Background” n.d. L3Harris Geospatial. Accessed October 12, 2020.
https://www.l3harrisgeospatial.com/docs/BackgroundSegmentationAlgorithm.html

SEIA. 2020. “Solar Market Insight Report 2020 Q2” Accessed October 22, 2020.
https://www.seia.org/research-resources/solar-market-insight-report-2020-q2

SEIA. 2020. “Climate Change.” Accessed June 15, 2020.
https://www.seia.org/initiatives/climate-change#:~:text=Electric Sector&text=Through
Q1 2020, the U.S.,tons of carbon dioxide emissions.

Sinaga, Siti Martha Uly, and Muhammad Kamal. 2019. "Image Segmentation for Vegetation
Types Extraction Using WorldView-2: A Case Study in Parts of Dieng Plateau, Central
Java." Sixth Geoinformation Science Symposium, 2019. doi:10.1117/12.2543480.
60

St. John, Jeff. 2019. "Tennessee Valley Authority Plans for Up to 14GW of Solar by 2038."
Greentech Media. July 02, 2019. Accessed August 14, 2020.
https://www.greentechmedia.com/articles/read/tennessee-valley-authority-plans-for-up-
to-14gw-of-solar-5gw-of-storage-by.

“Texture Metrics Background”. n.d L3Harris Geospatial. Accessed October 12, 2020.
https://www.l3harrisgeospatial.com/docs/backgroundtexturemetrics.html

US Fish and Wildlife Service. n.d. “NWI Program Overview”. Accessed December 10, 2020.
https://www.fws.gov/wetlands/nwi/Overview.html

US Department of Energy, Office of Energy Efficiency and Renewable Energy, National
Renewable Energy Laboratory. 2012. Utility-Scale Concentrating Solar Power and
Photovoltaics’ Projects: A Technology and Market Overview, by Michael Mendelsohn,
Travis Lower, and Brendan Canavan.

Uyan, Mevlut. 2013. “GIS-Based Solar Farms Site Selection Using Analytic Hierarchy Process
(AHP) in Karapinar Region, Konya/Turkey.” Renewable and Sustainable Energy
Reviews 28 (2013): 11–17. https://doi.org/10.1016/j.rser.2013.07.042.

Visual Information Solutions. 2007. “ENVI Feature Extraction Module User’s Guide” Accessed
July 22, 2020.
http://www.harrisgeospatial.com/portals/0/pdfs/envi/Feature_Extraction.pdf

Watson, Joss J.w., and Malcolm D. Hudson. 2015. “Regional Scale Wind Farm and Solar Farm
Suitability Assessment Using GIS-Assisted Multi-Criteria Evaluation.” Landscape and
Urban Planning 138 (2015): 20–31. https://doi.org/10.1016/j.landurbplan.2015.02.001.

Xiaoying, Jin. 2009. Segmentation-Based Image Processing System. US Patent US
2009/0123070 A1, filed November 14, 2007, and issued May 14, 2009.

Xiong, Xiaoxiong. 2010. “Using the Sonoran and Libyan Desert Test Sites to Monitor the
Temporal Stability of Reflective Solar Bands for Landsat 7 Enhanced Thematic Mapper
plus and Terra Moderate Resolution Imaging Spectroradiometer Sensors.” Journal of
Applied Remote Sensing 4, no. 1 (January 2010): 043525.
https://doi.org/10.1117/1.3424910.

61

Appendix A

Figure 12: First Iteration Segmentation Image
62

Figure 13: Second Iteration Segmentation Image
63

Figure 14: Fourth Iteration Segmentation Image

64

Figure 15: Fifth Iteration Segmentation Image

65

Figure 16: Sixth Iteration Segmentation Image

66

Figure 17: Seventh Iteration Segmentation Results

Abstract (if available)

Abstract As our country grapples with the long term negative effects that traditional electrical generation methods have on the environment, such as nuclear with a 50 year average decommissioning time, natural gas and the methane emissions associated with it, and coal which is not clean, there is a renewed focus on solar energy. This renewed focus is partially fueled by advancements in photovoltaic cell technology and favorable regulatory conditions, resulting in a decrease of solar energy production costs. This has led to the installed solar energy production capacity of the United States to grow from 7.33 gigawatts in 2012 to 51.45 gigawatts in 2018. As the industry matures and solar energy is adopted in new markets, the available land suitable for development has subsequently been reduced. A result of this is the industry has shifted its focus to identify suitable sites in areas that have been otherwise overlooked or discounted. To remain competitive, potential sites must be screened to identify site conditions that can increase costs or render a site undevelopable. This project identified that OBIA can successfully be used to identify a multitude of features that are encountered during the development of USSE projects, but the complexity and variability of the process makes it currently unsuitable to be deployed at scale. OBIA can however, be used to assess current site suitability analyses by generating otherwise unknown attribute data about a site to target locations wherein to look.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Suitability analysis for wave energy farms off the coast of Southern California: an integrated site selection methodology

PDF

Site suitability analysis for implementing tidal energy technology in southern California

PDF

Site suitability analysis: small-scale fixed axis ground mounted photovoltaic power plants in Fresno, CA

PDF

Beyond visual line of sight commercial unmanned aircraft operations: site suitability for landing zone locations

PDF

Onshore wind power systems (ONSWPS): a GIS-based tool for preliminary site-suitability analysis

PDF

Urban areas and avian diversity: using citizen collected data to explore green spaces

PDF

The use of site suitability analysis to model changes in beach geomorphology due to coastal structures

PDF

Utilizing existing museum collections and GIS for paleontological site assessment and management

PDF

An application of aerial drones in high definition mapping for autonomous vehicles

PDF

Building a spatial database of biochar research and practice with Web-GIS

PDF

Risk assessment to wildlife from Ohio on-shore wind farm development: a landscape model approach

PDF

A model for placement of modular pump storage hydroelectricity systems

PDF

Utilizing online data sources to improve existing military aircraft systems

PDF

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

PDF

Scenario-based site suitability analysis and framework for biodiversity conservation: agricultural zone, Galapagos Archipelago, Ecuador

PDF

Web-based relational spatiotemporal geodatabase of glaciers in California

PDF

Exploring the relationship between urban forests and ethnicity in West Covina using GIS and object-based image analysis

PDF

A spatial investigation of New York City's historical shoreline

PDF

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

PDF

Site location suitability analysis for a smart grid network

Asset Metadata

Creator McDermott, John Michael (author)

Core Title Geographic object based image analysis for utility scale photovoltaic site suitability studies

School College of Letters, Arts and Sciences

Degree Master of Science

Degree Program Geographic Information Science and Technology

Publication Date 04/24/2021

Defense Date 01/26/2021

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

Tag GIS,OAI-PMH Harvest,OBIA,object based image analysis,remote sensing,solar

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Bernstein, Jennifer (committee chair), Marx, Andrew (committee member), Wu, An-Min (committee member)

Creator Email jmmcderm@usc.edu,jmmcdermott1985@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-449699

Unique identifier UC11668936

Identifier etd-McDermottJ-9526.pdf (filename),usctheses-c89-449699 (legacy record id)

Legacy Identifier etd-McDermottJ-9526.pdf

Dmrecord 449699

Document Type Thesis

Rights McDermott, John Michael

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