Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Surface rupture detection with support vector machine classification: case study from Ridgecrest, CA
(USC Thesis Other)
Surface rupture detection with support vector machine classification: case study from Ridgecrest, CA
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
Surface Rupture Detection with Support Vector Machine Classification:
Case Study from Ridgecrest, CA
by
Jordan M. Cooper
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)
August 2021
Copyright © 2021 Jordan M Cooper
ii
To my beautiful wife and daughter
iii
Acknowledgements
Thank you to my advisor Dr. Fleming for his support, advice, and understanding throughout this
thesis. Also, thank you to my committee members for their recommendations, which helped
make this thesis project better. Additional thank you to Dr. Vos and the staff at the USC SSI for
their assistance and support during a trying time. The USGS and the various teams that collected
the data used in this thesis should also be thanked.
iv
Table of Contents
Dedication ....................................................................................................................................... ii
Acknowledgements ........................................................................................................................ iii
List of Tables .................................................................................................................................. v
List of Figures ................................................................................................................................ vi
Abbreviations ................................................................................................................................ vii
Abstract ........................................................................................................................................ viii
Chapter 1 Introduction .................................................................................................................... 1
1.1. Surface Rupture Mapping ...................................................................................................2
1.2. Classification of Remotely Sensed UAV Imagery .............................................................3
1.3. Study Area ..........................................................................................................................4
Chapter 2 Related Work.................................................................................................................. 7
2.1. Background .........................................................................................................................7
2.2. Linear Feature Detection.....................................................................................................8
2.3. Supervised Image Classification .........................................................................................9
2.4. Aerial Geologic Studies ....................................................................................................10
Chapter 3 Methods ........................................................................................................................ 13
3.1. Data Preparation................................................................................................................13
3.2. Classification.....................................................................................................................19
3.3. Spatial Analysis ................................................................................................................20
Chapter 4 Analysis ........................................................................................................................ 23
4.1. Test Results .......................................................................................................................23
4.2. Overall Results ..................................................................................................................33
Chapter 5 Discussion .................................................................................................................... 41
5.1. Future Work ......................................................................................................................42
References ..................................................................................................................................... 45
Appendix A: Test Run 2 Raster, Classification, and Error Matrix ............................................... 48
Appendix B: Test Run 3 Raster, Classification, and Error Matrix ............................................... 50
Appendix C: Test Run 5 Raster, Classification, and Error Matrix ............................................... 52
v
List of Tables
Table 1 Data Metadata .................................................................................................................. 13
Table 2 Test Run 1 Error Matrix................................................................................................... 27
Table 3 Test Run 4 Error Matrix................................................................................................... 32
Table 4 All Test Runs Error Matrix .............................................................................................. 39
Table 5 Test Run 2 Error Matrix................................................................................................... 49
Table 6 Test Run 3 Error Matrix................................................................................................... 51
Table 7 Test Run 5 Error Matrix................................................................................................... 53
vi
List of Figures
Figure 1 Ground View of Surface Rupture ..................................................................................... 2
Figure 2 Study Location.................................................................................................................. 4
Figure 3 Study Location.................................................................................................................. 5
Figure 4 Methods Flowchart ........................................................................................................ 13
Figure 5 Original Raster with Verified Surface Rupture Locations ............................................ 15
Figure 6 Raster Quality Report Sample ........................................................................................ 17
Figure 7 Test Run 1 Testing Pool ................................................................................................. 25
Figure 8 Test Run 1 Classification Results ................................................................................... 26
Figure 9 Test Run 1 Classified Rupture ........................................................................................ 27
Figure 10 Test Run 4 Testing Pool ............................................................................................... 30
Figure 11 Test Run 4 Classification Results ................................................................................. 31
Figure 12 Test Run 4 Classified Rupture ..................................................................................... 32
Figure 13 Results of All 5 Test Runs ............................................................................................ 35
Figure 14 Results of All 5 Test Runs With Verified Surface Rupture Locations......................... 36
Figure 15 Closeup of Classified Rupture Example....................................................................... 37
Figure 16 Closeup of Classified Rupture Example with Verified Surface Rupture Locations ... 38
vii
Abbreviations
GIS Geographic information system
GISci Geographic information science
SSI Spatial Sciences Institute
TIFF Tagged Image File Format
UAV Unmanned Aerial Vehicle
UAS Unmanned Autonomous System
USGS United States Geological Survey
USC University of Southern California
viii
Abstract
One possible costly outcome of large earthquakes is breakages in the ground surface, known as
surface ruptures. Surface ruptures can cause damage to human infrastructure as well as harm
humans. Detailed field studies that trace these structures take substantial time and effort because
they affect large regions, but this is reducible by collecting remote sensed imagery performing
supervised classification on the imagery with a GIS. For this study, unmanned aerial vehicles
(UAV) imagery from the Ridgecrest area in California was analyzed using the support vector
machine (SVM) classification method to attempt surface rupture detection in desert terrain. The
imagery covers a small area, so a K-Folds analysis method was used to attain statistically
significant results. This involved splitting the imagery into 5 random pools of 30-meter by 30-
meter squares before running the classification. The results of each classification were analyzed
by generating confusion matrices and visual inspection. An average of the five confusion
matrices was used for a final analysis. While this method did classify larger segments (> 0.5-
meter-wide) of surface rupture that was in the image, it missed most of the smaller surface
rupture segments (< 0.5-meter-wide). In addition, the technique misclassified parts of the
imagery as surface rupture, especially around the vegetation, road paths, and amongst a rock
field in the south-east corner. Further testing should be done with this method, including using it
on imagery with different land-cover. Based on the results of the further testing it may be ready
to try in practice scenarios for real-life earthquake disasters.
1
Chapter 1 Introduction
One of the natural disasters that cause the most significant damage to human-made
structures and loss of life is earthquakes. From 2009 to 2019, 27,193 deaths were attributable to
earthquakes, 47.7% of the 56,967 recorded natural disaster deaths occurred worldwide during
that period (CRED 2020). During that same period earthquake also caused 44.6 billion U.S.
dollars in economic losses. While earthquakes can cause damage in various ways, the most
damaging source is surface ruptures (Jin and Kim 2020). Surface ruptures are cracks in the
ground that can be anywhere from a few meters in length to several kilometers and can have
vertical offset up to six meters. The ruptures are associated with large earthquakes greater than
6.0 Mw. The San Fernando Earthquake in 1971 led to the Alquist-Priolo Earthquake Fault
Zoning Act in California, which prevents building occupied structures on active faults that can
cause surface ruptures. Surface ruptures were observable following earthquakes worldwide, such
as in the 1999 Chi-Chi, Taiwan earthquake, and the 2003 Bam, Iran earthquake. These
earthquakes caused damage to many buildings and had high fatality rates, with 2000 people who
died in Chi-Chi and at least 25000 died in Bam. Much of the destruction was associated with
surface ruptures.
In this thesis, remote sensed imagery was classified, mapping the potential location of
surface ruptures in imagery from a study site near Ridgecrest, California. This site experienced
an earthquake sequence that involved over 100,000 total earthquakes larger than 0.5 Mw.
Amongst all of these the earthquakes that were the most powerful were a 6.4 Mw on July 4
th
,
2019 and a 7.1 Mw on July 5
th
, 2019. While different satellite and aerial systems for collecting
remote sensed imagery exist, this thesis only used imagery from high spatial resolution cameras
mounted on unmanned aerial vehicles (UAV). The raster imagery used came from a collection of
2
imagery captured at two approximately 500-meter by 500-meter sites over six days between July
7
th
, 2019 and September 27
th
, 2019. The imagery, captured on July 11th, of a segment of the
surface rupture produced by the 7.1 Mw earthquake was used because of the data's good quality.
This thesis aims to classify the imagery to differentiate between surface ruptures and the
surrounding area. With a real-world application in mind for this project, the methods used were
selected for ease of use. Also, the methods should be able to analyze raster imagery and find
surface ruptures following large earthquakes to aid rescue crews and repair teams, as well as
policymakers.
1.1. Surface Rupture Mapping
Surface ruptures are cracks in the ground that are usually formed from earthquakes 6.0
Mw or greater in strength (see Figure 1). The ruptures are either primary ruptures, where the
earthquake fault extends to the surface, or secondary, where the earthquake fractures the surface
outside the earthquake fault plane. The creation of the ruptures can cause damage to utilities,
transportation infrastructure, and buildings with people in them. Rapid mapping of these
ruptures, along with infrastructure plans, would allow for quick response to earthquakes like the
2019 Ridgecrest Earthquake sequence.
Figure 1 Surface Rupture located near the maximum displacement from the Mw 7.1 earthquake
(Image Source: Jet Propulsion Laboratory, October 17, 2019).
3
The 2019 Ridgecrest Earthquake sequence generated two sets of surface ruptures when
the two largest earthquakes occurred, a 7.1 Mw mainshock and a 6.4 Mw foreshock. When
investigating the earthquakes, researchers used various methods, including satellite imagery,
aerial imagery, and field investigations. From their studies, the researchers established the
different characteristics of the ruptures generated from the two earthquakes. The 7.1 Mw
displacement was right lateral with up to 7.0 m of movement though displacement was mainly in
the range of 1.2-1.7 m. 66% of the displacement occurred around the earthquake's epicenter over
a 12 km section of the surface rupture. The 6.4 Mw was left lateral with up to 1.6 m of
movement on an individual strand, though most movement was in the range of 0.3-0.5 m.
1.2. Classification of Remotely Sensed UAV Imagery
Methods used during this thesis were based on a methodology previously used to classify
ground cracks in a coal mining area using UAV imagery (Zhang et. al. 2020). Traditionally
classification is performed by a researcher going through each image and marking the location of
the object of interest; in this case, it is surface ruptures. This can be a time-intensive endeavor.
With the advent of faster and more powerful computers, that task is simplified with automatic
classification methods, both supervised and unsupervised. During this project a supervised
method, SVM classification, was used in ArcGIS Pro. The software used training data to teach
the software the optical properties of the surface ruptures to classify it from the imagery that was
provided. An explanation of A further explanation on the classification methods is found in
Chapter 3.
In a GIS, Support Vector Machine algorithm is a supervised machine learning
methodology that is used to classify raster imagery. Classes are defined based on hyperplanes
that define the boundaries of the classes. A hyperplane in geometry is a subspace that has one
4
less dimension than the surrounding space, splitting the surrounding dimension into two parts
(See Figure 2). In machine learning, hyperplanes are the boundaries that separate different
classes. The hyperplane is plotted to fit between the groups of data points for the different classes
defined in the training data. Hyperplanes can usually fit several different orientations, but the one
that is selected fits between the closest data points on both sides in such a way that they will have
the greatest possible distance or margin from those data points (See Figure 2).
Figure 2 Datasets x and o separated by a hyperplane. Dashed lines represent the maximum
margins. Bold X and O’s are data points used to find the margins.
1.3. Study Area
The study area is a part of the Eastern California Shear Zone in the Ridgecrest-Trona
region of California. The site is located in the north-west corner of San Bernardino County,
approximately 200 kilometers north-east of Los Angeles, CA and about 234 kilometers south-
5
west of Las Vegas, NV (See Figure 3). Four mountain ranges are found in the region: the Cosos
in the north, the Sierra Nevada Range in the west, the Argus Range in the east, and the El Paso
Mountains in the south. In addition, five fault systems can be found in this region: the Owens
Valley fault zone, Panamint Valley fault zone, Garlock fault zone, Blackwater fault zone, and
San Andreas fault zone. Researchers are interested in this area because of the complicated
interaction occurring between the faults, especially the San Andreas and Garlock, both faults
which can produce large earthquakes. This area has limited vegetation and a small human
presence, making it easy to observe surface changes. Ridgecrest is the most populous city in the
area, with 27,616. Also, in the vicinity is the United States Navy's China Lake research facility,
where most of the damage occurred. Researchers had limited access to investigate, so not much
data exists from that area.
Figure 3 Location of the study site in north-west San Bernardino County.
6
The following chapter is a summary of research relevant to this thesis, starting with
background information on the site and location. This includes an exploration of field data and
remote sensed data collected from the Ridgecrest 2019 earthquake swarm. A detailed description
of what data has been collected and analyzed is necessary to show how this project fits into the
literature. Following is a look at the use of remote sensing and image classification of linear
features. This is looked at to see what other potential methods are available for surface rupture
detection. A short and broad look at supervised image classification comes next. Finally, the
usage of aerial studies related to the detection of geologic features is presented. The rest of this
thesis will be the following: methods used in this project, the results of the project, and the
analyst of the results.
7
Chapter 2 Related Work
This report, like all research, benefits from a look at the related body of work that has
been done in the past. To start, a look at the research that has been done related to the surface
ruptures and surface damage that occurred at Ridgecrest and other seismic events. After this,
linear feature detection is investigated. Following that will be a general look at supervised
classification methods with more in-depth coverage of support vector machine classification.
Finally, the last section of the chapter covers papers that used methods for the detection of linear
geologic features.
2.1. Background
The 2019 Ridgecrest earthquake swarm created a rare opportunity for researchers to
study surface ruptures with high-resolution equipment. Image data that was produced was
extensive, including Aerial LIDAR from airplanes, High-resolution multi-spectral imagery from
UAV's, and various forms of satellite imagery, both public and commercial. Each of these
capture methods produced imagery that contain various optical properties of the study area. This
makes it possible to classify the imagery through a GIS with different methodology.
Prior research into detecting damage from earthquakes through remote sensing means has
been concentrated on satellite imagery and airplane mounted cameras. These images have been
used to study the destruction of buildings and infrastructure following several major earthquakes
such as the 1999 Chichi, Taiwan Earthquake, the 2003 Bam, Iran Earthquake, and the 2010 Haiti
Earthquake (Dong and Shan 2013 and Fielding 2005). A common methodology that has been
used to detect such damage has been change detection, which is a technique that involves a
comparison of a before and after image. Early change detection algorithms would go through a
pixel-by-pixel comparison, while modern algorithms usually perform an object-oriented
8
comparison. This technique is helpful for detecting significant changes such as when a building
collapses, or a road becomes offset, but it is poor at discerning small changes such as surface
ruptures. In addition, it is not helpful for situations in which there is not a before image or the
before image is too much older than the after image. Based on this fact, new techniques that can
rely on a single image are needed to detect surface ruptures and other seismic damage.
2.2. Linear Feature Detection
Remote sensing imagery has existed since the 1950s since the first satellites were sent
into orbit by Russia and the United States. Following that time, the resolution of the imagery and
the types of imagery collected advanced, leading to the ability to automatically detect objects in
the 1970s. One of the early uses for the synthesis of remote sensing and automatic detection
methods was the detection of roads and highways, objects that are sufficiently large enough to be
detectable in the imagery of the time (Quackenbush 2004). Since the early days, the spatial
resolution has continued to advance, allowing for smaller objects to be observed with greater
accuracy.
In the utility industry, it has become commonplace to see linear object detection used. For
example, in the electrical industry, remote sensing has been used to map out the locations of
electrical towers and corridors in the early days. As new image capture methods have evolved,
the applications have changed for the industry, with uses in the mid-2010s aimed at monitoring
full grid status and fire safety along electrical corridors (Matikainen et. al. 2016). Methods used
include airborne laser scanning synthetic aperture radar (SAR) images, optical satellite and aerial
images, thermal images, airborne laser scanner (ALS) data, land-based mobile mapping data, and
unmanned aerial vehicle (UAV) data capture (Matikainen et. al. 2016 and Guo et. al. 2016). All
9
these various tools have been brought together to create complete 3-D views of the electrical
grid.
Another field that has benefited from linear object detection has been civil engineering.
Within civil engineering, there has been an increase in studies related to crack detection in
concrete (Mohan and Poobal 2018). These techniques are of particular interest to this thesis since
the goal of this thesis is to detect cracks in a different scenario. Cracks, in our situation, are
surface ruptures that should have similarities to those found in concrete, such as being linear and
darker colored than the surrounding material.
Methods from industry are applicable to detect other linear features like surface ruptures
but with one caveat. First, the algorithms used for detecting and monitoring these features have
set rules which cannot be applied to ruptures. Pipelines, electric corridors, and roads have rules
that are common, such as a line ends in a node or a water pipeline should connect to other lines.
Surface ruptures also have rules that are as well established for where they occur. Many factors
affect where surface ruptures arise, such as the geometry of local faults, soil and rock properties,
and the strength of the earthquake that created it. Therefore, rules-based algorithms are not as
useful for the detection of not just surface ruptures but other geologic lineaments as well.
2.3. Supervised Image Classification
Supervised classification methods are methods that involve a user creating training data
with samples of the classifications that are of interest in remote sensed imagery. Supervised
classification methods having a higher initial set up than unsupervised and can be more time-
consuming, but it allows for some ground truthing to be incorporated into the classification
process. Traditionally classification was done by pixel-based methods, but object-based methods
10
have gained popularity (Lei Ma et. al. 2016). This method groups pixels together using
segmentation and classifies those segments based on what object the system predicts it to be.
Different object-based methods exist that are applicable to different scenarios, such as
classifying different environments like urban, forest, or agricultural areas (Tehrany, Pradhan, and
Jebuv 2013 and Duro, Franklin, and Dube 2012). Assorted methods are also best suited for work
with rasters captured by various imaging systems like multi-spectral data or hyperspectral
images. Amongst the many supervised classification techniques, Random Forest, K-Nearest
Neighbor, and SVM are considered to have the highest level of accuracies and are the most used
because of it (Thanh Noi and Kappas 2017). Amongst these three classification methods, the
overall accuracy and results of their classification was similar if they had sufficiently large
training sample size, around 0.25% of the research area. When below this value, SVM has been
found to be the least affected by the size of the training data. SVM was selected for this reason,
since the rupture lines are narrow areas of pixels in the rasters that are smaller than 0.25% of the
area. Because the SVM classification is a core component of this thesis it is discussed in the
following subsection.
2.4. Aerial Geologic Studies
Remote sensing has been a boon to the earth sciences since the first satellite images were
captured, changing the way research is done. Since the early days of geologic field studies,
research was done in person. Mapping of geologic features would be done this way; faults and
folds could be found as what was seen in the field was drawn on the map. This changed with the
development of one of the earliest forms of air travel, hot air balloons. A field mapper could go
up and see the geology from a bird's eye view. As more advanced forms of air travel developed
11
and better cameras, more of the earth could be studied this way. The airplane, the rocket, and
satellite, and now the UAV have all added to the available imagery of the earth's geology.
Studies of earthquake faults have used various remote sensing imagery to tease out the
structure of faults. This has included standard methods such as LIDAR and SAR, but also less
common field research tools such as UAVs that can detect changes in gravity anomalies or
magnetic field changes in the subsurface. These specialized systems tend to be costly, so
researchers use other cheaper methods such as InSAR. InSAR imagery is used to detect
deformation that occurs following major seismic events. InSAR is one of the tools that has been
used to create an image of the structure of the faults at depth. Surface ruptures occur along with
earthquake faults, so understanding what is happening at depth helps researchers develop a
complete image of what occurs at the surface, and the opposite is true too. For the Ridgecrest
Earthquake Swarm, InSAR was a significant tool that was used along with surface rupture
studies to find the geometry of the faults as well as the slip that occurred on them (Ross et. al.
2019)
Surface exposures of other linear geologic features have been detected as well, such as
fault scarps. Research was performed in Slovenia to detect hard to observe fault scarps that were
located in a forested area. To do this, researchers used a vegetation removal algorithm to remove
trees in order to get a clear look at the topography. From there, they were able to visually inspect
the topographic imagery and find fault scarps (Cunningham et. al. 2006). DEMs and DSMs have
been a common tool for locating faults, such as when DEMS and satellite imagery were used to
locate earthquake faults in the Black Forest, Germany (Meixner et. al. 2017). In another study,
researchers used a LIDAR data collection to detect fault scarps in Northern California (Sare,
Hilley, and DeLong 2019). This study used DEMs they generated in combination with template
12
matching algorithms to potentially detect segments of fault in the northern San Andreas Fault
system. All of these examples show the importance of remote sensed imagery for detecting fault
structures. A further consideration shows that combining it with various detection techniques
creates powerful tools for finding fault exposures and other lineaments in hard-to-reach areas
such as these.
UAV’s have been used to collect imagery in several different areas of research in both
the social sciences and the physical sciences. One area that is seeing increased usage as of 2020
is geologic studies. A significant component of geology is field research which is often done in
demote, hard-to-reach areas. To study some of these sites’ airplanes would be flown over to
collect images of the site. Since then, though, UAVs have started to gain usage because of
several advantages they have. UAVs are cheaper to fly than aircraft and can be used to image
the same area over a small timespan. They are also able to reach areas that are hard to get to or
hazardous to travel through, such as in a volcanic area or a landslide area. Another reason is that
they can collect a high-resolution image normal to the surface and cover a larger area than a
person on foot. The last reason is that cameras on some UAVs can be changed out, with each
camera being able to collect imagery in different wavelengths, which can then be used to study
various features.
13
Chapter 3 Methods
This study is composed of three general steps for the methods used: Data Integration,
Image Classification, and Spatial Data Analysis. The first step, Data Preparation, looks at what
data is used and what changes need to be made prior to classification. In the second step, (3.2)
Image Classification, the imagery will be classified by SVM, a supervised technique that is
readily available in ESRI's ArcPro 2.1. The last step, Spatial Data Analysis, (3.3) will involve an
explanation of error matrices for the classification results and how it is used for the analysis. An
overview of the steps can be seen in figure 4 below.
Figure 4 Flowchart of the steps involved in the methods used.
3.1. Data Preparation
This section covers: (3.1.1) the selection of the datasets for classification, (3.1.2)
processing of the data to make it into a usable format, and (3.1.3) preparing it for creating
training data for the object classification.
14
3.1.1. Data Selection
This thesis used two different files (see Table 1 below). The first data set is the UAV
imagery that was downloaded from Geo-Gateway.org for a study site that covers an area of
approximately 0.275 km
2
(Figure 5 below). This imagery contains a segment of surface rupture
from the 7.1 Mw that is observable with the naked eye in the image, as well as several smaller
strands that are harder to locate. Additional imagery was collected at this site on six other days,
but due to time and data storage constraints, they were not used for this project. The second
dataset is a line feature file of the surface ruptures produced by the United States Geologic
Survey. In this data set, the data is generalized and considered preliminary.
Table 1 Data Metadata
Data Name File
Type
Source Scale Precision Accuracy File
Size
Surface
Rupture
Lines
Surface_Rupture
Ridgecrest
_Prov_Rel_1
.kml/.shp USGS Unknown Unknown < 10 m 32
MB
Drone
Imagery
SVRL2019
0711C_
dsm_2cm
.laz/.las
https://archive.
Geo-gateway.
org:3000/
main.html
Multiscale .5 m to
1.72 m
2 cm 1.5
GB
15
Figure 5 High resolution imagery of the study site with confirmed surface ruptures marked.
Raster from Donnelan et al and lines from DuRoss et al.
16
Several remote sensed datasets related to the study site were found to be readily
available; these datasets were collected with various electromagnetic spectra and methods. It was
deemed acceptable for this project to use only one type of remote sensed data to simplify the
project. In this case, a UAV imagery file that covered approximately 500 meters of the surface
rupture from the Mw 7.1 earthquake. The UAV imagery has a high spatial resolution, with
resolution approximately at 2 cm and an average ground sample distance of 1.68 cm. This
dataset was taken on July 11
th
, seven days after the earthquake swarm began and five days after
the Mw 7.1 mainshock occurred. The UAV imagery data is available in two different formats, a
raw point cloud and a derivative digital surface model (DSM). Both file types have the same
accuracy and spatial resolution, but the DSM files were selected for the classification in this
project. The DSM file is preferred because it is a derivative of the point cloud that does not need
further processing prior to classification. In addition, the DSM is a smaller file, which is a benefit
when data storage is limited.
The UAV DSM imagery is composed of pixels that contain latitude and longitude data
along with the three-color bands that contain the optical properties for the image. In addition, the
surface elevation is also available in each pixel, which, along with the latitude and longitude
data, is based on the coordinates collected at GPS stations. Donnellan et al. collected the imagery
with a Parrot Anafi model UAV. The Anafi comes equipped with a 21-megapixel camera that
took the imagery in a forward-facing direction, 20 degrees up from the nadir. The height of the
flights was held at 45 m above ground level. Flights for the Mw 7.1 surface rupture segment
were performed in the early afternoon for each of the flight days except one which was
conducted in the morning. Like other common UAV’s the Anafi has a limited flight time, up to
approximately 25 minutes, that can cover only a small area.
17
Donnelan et al. (2020) used Pix4Dmapper software to create the orthomosaic and digital
surface model raster from the imagery that the UAV collected (Figure 6 below). Pix4Dmapper
created a quality report that contains information that is important to review when working with
any data. The UAV collected 2432 images in total but only 2049 were used to create the DSM.
The imagery was georeferenced using 11 ground control points with a RMS error of 0.016 m.
That is a small RMS error meaning relative to the ground control points the imagery is accurately
georeferenced. Figures showing the flight pattern can also be found in the quality report. These
show how often the same location was visited. More in-depth accuracy measures can be found in
the quality report.
Figure 6 Sample of pages from quality report generated by Pix4Dmapper
18
The line file was originally a kml file the USGS released with the outline of the
preliminary locations of the ruptures. Even though the line work is considered preliminary, it still
has a high level of accuracy and has been validated in the field by the USGS. The file was
imported into ArcGIS to use in the project as a a form of validation for the results of the
classification. It was also visually compared to the raster imagery and found to match up closely.
3.1.2. Raster Format
High-resolution imagery, such as what was used in this project, is usually large, often
reaching sizes greater than one gigabyte. The files were compressed to compensate for that and
make the files easier to transfer or download. In the case of the UAV imagery that was used in
this project, it was compressed with the laz compression format. The laz format is a standard
compression format used to compress las files. The las file format is a raster format for LIDAR
and high-resolution imagery. Las files contain 3-dimensional point cloud information, with each
point containing x, y, and z data as well as the optical information at that point. The data was
uncompressed in ArcGIS Pro with the Convert Las tool. All parameters in the tool were left as is.
This tool converted the laz file to a las file. ArcGIS Pro displays LAS files as dots, so to make
this usable, this format was then further processed with the LAS Dataset To Raster tool. Most of
the parameters were left as is except for the sampling value, which was defined as .02 m and the
file format selected was TIFF. TIFF is a common file format for rasters that ArcGIS Pro can
export natively.
3.1.3. Data Preparation
The raster covered a small section of the overall surface rupture system in the area, 0.5
km of the 50 km of surface ruptures produced by the Mw 7.1 earthquake. This limited coverage
of the site made it necessary to use a K-folds cross-validation method to ensure a statistically
19
significant number of tests were done. K-fold cross-validation is explained below in Section 3.3.
The DSM file was split up into 30 meters by 30 meters grid using the Raster Split tool to
generate the files that were used for the cross-validation. This yielded 274 separate raster files.
30-meters by 30-meters was chosen because it kept the number of the split rasters within a range
that was not too many that it would be time-consuming to work with but also not too few that a
statistical significance could not be reached.
3.2. Classification
A four-step process was used to perform the classification. First step was the
segmentation of the data. This is a technique that the GIS performs to create a new raster where
the pixels from the original raster are grouped together based on similar characteristics. The
second step was the creation of training data for each of the training groups in the Training
Sample Manager in ArcGIS Pro. When selecting training data, polygons were created that
overlay the pixels in the imagery that represent the three categories of classification, "rupture,"
"no rupture," and "vegetation." In initial tests, additional categories were attempted for the
classification schema, but these seemed to muddle the output and made each classification run
several hours longer. Visual inspections of the results showed a decreased performance of the
classification. The third step is to use the Train Support Vector Machine Classifier tool. This tool
creates an Esri classifier definition file (ecd), which is a type of file that contains the attributes of
the training results such as the hyperplane coordinates. The recommended max sample per class
setting of 1000 was found to be acceptable for this tool. To confirm this, various values for this
setting were tried on a small test area before the project was attempted. It was found that an
increase in number of sampling beyond 1000 did not produce a noticeable change in
classification but significantly increased the processing time for the tool. The final step was to
20
use the ecd file in the Classify Raster tool with the test samples. In this project, the tool uses the
training data to find whether a pixel belongs to the "rupture," "no rupture," or "vegetation"
category. This generated the final classified rasters that were then analyzed.
3.3. Spatial Analysis
The goal of this part of the thesis is to check the accuracy of the classification outputs
created by the methods used. To check for the accuracy quantitatively, error matrices were
created for each test run. Error matrices are table outputs that are created when one compares a
dataset with expected values to the classified dataset. Prior to an error matrix being created in
ArcGIS, random sample points are made using the Create Accuracy Assessment Points tool. This
generates points that contain the classified value at that location and requires the user to input the
expected value. These values are used in the calculation of the error matrix.
The error matrix table contains information on the results of the accuracy assessment
points as well as the producer accuracy, user accuracy, total accuracy, and kappa coefficient.
Producer accuracy is how likely is what is found in the original raster will be found in the
classified image. This is tied to errors of omission (Equation 1 below), which are those values
that were omitted by the classifier, so in this case, any surface rupture pixels that are missing.
User accuracy, on the other hand, is how likely what is found in the map matches with what is
observable in the field. This is tied to the errors of commission (Equation 2 below), which are
those values that were improperly classified as the classification of interest; in this case, it is
those pixels that are classified as "surface rupture" when they should have been "not surface
rupture" or "vegetation". Total accuracy is the overall accuracy of the classification, but it is an
average across all the classes. Because it is an average, one does not see if any individual classes
are more or less accurate than the other classes. This is why user accuracy and producer accuracy
21
are critical to check for each class. The last statistical measure looked at is the kappa coefficient.
Kappa is a measure of likeliness that results are due to the classification software or random
chance. Kappa is a value from -1 to 1, with a value closer to 0, meaning that the results are
random and a value closer to 1, indicating the results are due to the classification. The coefficient
is calculated by comparing the actual results to expected results (Equation 3 below).
Equation 1: Producer Accuracy = 100% - Omission Error
Equation 2: User Accuracy = 100% - Commission Error
Equation 3: Kappa Coeffiecent = (Actual Results - Predicted Results) / (1 – Predicted Results)
For this project, we are primarily interested in the correct classification of surface
ruptures from the imagery, but the overall accuracy of the classification of vegetation and non-
rupture surfaces is also investigated. Both the errors of commission and errors of omission were
looked at to check the accuracy to quantify how well the overall methodology worked. Five error
matrices were created, once for each of the 5-Fold cross validation runs. The Producer Accuracy
and User Accuracy from the five runs were then averaged out to find the resulting accuracy of
the method. Kappa coefficient is also looked at to check whether the results are due to the
classification algorithm or just a random result.
Usually when testing a classification method, such as SVM, it is necessary to run several
different tests to remove any potential biases that would occur in a single test run. This is not
possible with a small data set so the K-folds cross-validation is used instead. K-folds cross-
validation is a validation technique that is useful when working with a small data set, since it less
likely to be affected by biases (James 2017). When using K-folds the researcher randomly splits
22
the data into equal sized pools. The training data creation and classification is ran a number of
times equal to the number of pools. Each of these test runs will use a pool as a test group while
the rest will be used for training data. After a pool has been used a new pool is selected and the
last pool joins the training data. When all the data has been used once, the results are averaged
out creating the final assessment. As stated before, the averaged out results should limit the
effect of biases found in any one testing pool.
Typical K-fold cross validation is done with somewhere from 5 to 10 test groups. In the
case of this thesis a 5-fold cross-validation method was used to cross-validate the data. This
means that every test was composed of a randomly selected mix of images, with 80% being used
to generate the ecd training files and 20% being used to test the training files and create the
classified rasters for validation. To do this, the split images were assigned a number, and those
numbers were randomly selected to make each test pool. Every image only participated in one
test pool.
23
Chapter 4 Analysis
Chapter 4 is separated into two sections. The first of the sections is the results for two of
the five test runs, which will be presented and analyzed (Section 4.1). The second section is the
evaluation of the overall results for the entire data set. (Section 4.2).
4.1. Test Results
The first four test runs were composed of 69 of the 30-meter by 30-meter rasters that
were split off from the original raster by the Split Raster tool. The fifth test run was composed of
68 raster portions. Features found throughout the original raster were observed to be distributed
between all five test runs. All the figures for this section can be found in Appendix A except for
the figures for Test Run 1 (4.1.1.) and Test Run 4 (4.1.2). Error matrices for Test Runs 2, 3, and
5 can be found in Appendix B. In each test run, the typical components of the original raster
were found throughout the group on split rasters. These components include the desert surface,
paths, dry streambeds, paved roads, rocks, and shrubs. In addition, some discoloration can be
found in the sections of the imagery towards the edges of the image. This discoloration shows up
as a blue tint on the eastern and western sections and orange in the northern section. There is also
a minor blue tint in the southern part of the map that is not as prevalent as the discoloration in the
other edges. It is relevant to be aware of the discoloration around the edges since that changes the
properties of the raster cells at those locations. When creating the training polygons, those
sections were avoided to prevent issues with placing the boundaries between the classes in the
training file.
Results from two of the five test runs are covered in this section (4.1.1 – 4.1.2.). These
two test runs were selected to represent the range of results where one consists of a poor result
and one that represents a good result. For each test run, the classified raster was visually
24
inspected prior to the creation of the error matrix. Observations from those inspections are
discussed first. Second, a small discussion of the results of the error matrix will follow.
4.1.1. Poor Test Results - Test Run 1
Portions of the original raster that were used in the first test can be seen below in Figure
7. For this test run, one of the crucial parts of the raster is the small segment of the main surface
rupture that can be found in the center of the study site, which is circled in figures 7 and 8, and a
larger version is in Figure 9. The rupture can be seen in the lower-left square as a line in the
desert surface with some dark soil around it in figure 9. In figure 9, there is a section of pink that
traces that same area, showing that the classification method did capture the main surface rupture
in that area. A large area appears to be captured, which includes disturbed soils around the
surface rupture as well as the rupture itself. Similarly, this can be seen in the square north-west
of those, providing additional evidence that the main surface rupture is detectable by the SVM
classification method. On the other hand, it does not appear to capture the thinner segments that
are present. It also seems to misclassify the paths and stream beds as "rupture." Much of the
shrubbery and non-surface rupture areas appear to be adequately classified, with obvious
exceptions being the aforementioned paths and streams, as well as the edges of the study area,
the rock field in the south-east corner, and the road to the north.
25
Figure 7 Randomized 30x30m rasters used for Test Run 1. Circled area shown in closer detail in
Figure 9.
26
Figure 8 Classified Results of Test Run 1, classified with SVM classifier. Circled area shown in
closer detail in Figure 9.
27
Figure 9 Comparison of the main surface rupture observed in the original raster (A) and the
classified raster (B) for Test Run 1.
The confusion matrix shows that the classification of Test Run 1 was a mixed result
(Table 2). The "no rupture" classification had a user accuracy of 96.7%, which means that only a
small amount of the other classes was classified as "no rupture." When looking at the producer
accuracy for the same class, one sees that many of the "rupture" cells were included, this in
combination with the small number of "vegetation" misclassification caused a drop in the
accuracy to 68.3%. "rupture" cells counted for only a small number of the total cells in the
classification, so any errors have a substantial effect on the accuracy. This is why the user
accuracy is at only 6.8% for the "rupture” class, with 105 "no rupture" and 17 "vegetation" cells
being classified as “rupture” versus only nine cells being correctly classified as "rupture." The
surface ruptures are narrow lines that limit the available cells to classify. Producer accuracy is
also affected by this fact, so even though it resulted in a 90.0%, it needs to be carefully
B A
28
considered when analyzing the results. The final class, "vegetation," had good results, with a
producer accuracy of 76.9% and user accuracy of 83.8%. "Vegetation" is distinct from the desert
surface, with the dark coloration of the vegetation being easy to distinguish from the light colors
of the desert. This helped in the consistent classification, though the edges of the vegetation
would blend with the desert surface immediately next to it, which is likely why the results were
average. Overall, the total accuracy was 70.6%, which was dragged down by the "rupture" user
accuracy results, as well as the producer accuracy results for "no rupture." Finally, the Kappa
coefficient can be looked at, which was .454. Kappa coefficients do not have exact breakdowns
for what each coefficient value means, though the fact that it is closer to zero means that the
results were only slightly better than random chance. All of these results point to Test Run 1 not
being successful at capturing the surface rupture in a way that is meaningful, especially for a user
of the map.
Table 2. Confusion Matrix (or Error Matrix) for Test Run 1
Class Value No Rupture Rupture Vegetation Total User Accuracy Kappa
No Rupture 261 1 8 270 96.7%
Rupture 105 9 17 131 6.9%
Vegetation 16 0 83 99 83.8%
Total 382 10 108 500
Producer
Accuracy
68.3% 90.0% 76.9%
70.6%
Kappa
0.454
4.1.2. Good Test Results - Test Run 4
Test Run 4 is composed of the portions of the original raster found below in Figure 10.
This test run, like all the test runs, has segments of the surface ruptures in various portions of the
map. A segment of the primary surface rupture is circled below in both the original raster in
29
figure 10 and in the classified raster seen in Figure 11. A larger version of the surface rupture is
found in figure 12. The fault is a dark line that cuts diagonally from north-west to south-east
across both image portions. In the matching classified image, the surface rupture is clearly
captured. There are small pockets of the “rupture” class also seen in the image along the paths,
but they do not stand out as much as the classified surface rupture. The SVM classification was
even able to capture the separate strands of the rupture in the upper left square. Test Run 4
appears to have been the best classification of all the tests, with not just the main surface rupture
classified but parts of the smaller ruptures as well. Like the other test runs, there was distortion
along the edges of the raster from the color tint, which caused widespread misclassification in
those areas. In addition, there was misclassification in the rock field area in the south-east.
30
Figure 10 Randomized 30x30m rasters used for Test Run 4. Circled area shown in closer detail
in Figure 12.
31
Figure 11 Classified Results of Test Run 4, classified with SVM classifier. Circled area shown in
closer detail in Figure 12.
32
Figure 12 Comparison of the main surface rupture observed in the original raster (A) and the
classified raster (B) for Test Run 4.
For the classification of Test Run 4, the error matrix had primarily good results (Table 3).
The "no rupture" class had a user accuracy of 98.9% and a producer accuracy of 89.6%, making
it the most accurately classified class in any of the tests. A small number of the "rupture" class
was thrown into both, which is expected since surface rupture is formed in the desert surface.
The training data used may have had a small amount of "rupture" pixels mixed in with "no
rupture," and vice versa as well. "Rupture" class results were okay at 18.8% for the user
accuracy, better than any of the other test runs. The producer accuracy was lower than the other
classes, though, at 69.2%. As previously stated, the results for the "Rupture" class need to be
carefully considered since the surface rupture is a small portion of the study site. “Vegetation”
was also highly accurate, with user accuracy of 82.5% and a producer accuracy of 83.3%. All
these results led to a good total accuracy of 87.8%. This is higher than any of the other test runs,
which in combination with the visual inspection of the classified raster, shows that this was a
successful test run. The final part of the confusion matrix to look at is the Kappa coefficient. Test
B A
33
Run 4 has the highest kappa coefficient of all test runs at 0.7. The score is close to 1, which
means that the classification results were not likely caused by random chance. Unlike Test Run
1, all of the results for Test Run 4 show that with good training data, large surface ruptures can
be captured with the SVM classification method.
Table 3. Confusion Matrix (or Error Matrix) for Test Run 4
Class Value No Rupture Rupture Vegetation Total User Accuracy Kappa
No Rupture 344 3 1 348 98.9%
Rupture 23 9 16 48 18.8%
Vegetation 17 1 85 103 82.5%
Total 384 13 102 499
Producer
Accuracy
89.6% 69.2% 83.3%
87.8%
Kappa
0.7
4.2. Overall Results
This part is similar to the previous sections, but it is a look at the overall results across all
five test runs. This includes a look at a map with the combined results, as well as averages of the
results of the five confusion matrix results, except for the kappa coefficient.
The results from all five test runs were brought together, as shown in Figure 13 below, to
get a better understanding of how effective the classification method is. Underneath is Figure 14,
which is Figure 13 with the addition of the linework from the USGS for comparison purposes.
There are several features that stand out when first looking over the complete picture. One
observation is that the road and paths in the original raster are traced out accurately in the image.
The paths stand out because of the lack of "vegetation" found in them, while the "no rupture"
class usually has "vegetation" dispersed throughout. This shows that the methodology used was
able to differentiate "vegetation" and "no rupture." Another observation is the misclassification
34
that occurred along the edges of the raster. This highlights the effect of the discoloration in the
original raster image. That discoloration has caused the cells in those areas to have the wrong
optical properties, so the classification placed them in a different class based on how the training
data was binned. In the south-east corner, there is also a large section of “vegetation," which is a
group of misclassified cells that should have been "no rupture." That area is a rock field that has
optical properties that are different from any other part of the raster. This could be corrected in
future runs of the methodology by creating an additional class, though this will increase
processing time. It is also important to point out the large areas of the raster that are classified as
"rupture" that should be placed into another class. Those are from Test Runs 1 and 2 and appear
to have occurred because of the training data that was used. This is based on the fact that all of
the settings and methods stayed the same between the tests, but a more significant amount of
training data was used for the last three test runs.
35
Figure 13 Classified Results of all 5 Test Runs. Close up of area in square in Figure 15 and 16
below.
36
Figure 14 Classified Results of all 5 Test Runs with surface rupture lines. Close up of area in
square in Figure 15 and 16.
37
Finally, the success of the method at classifying the surface ruptures needs to be
examined. Figures 15 and 16, found below, show a sample of the classification of the rupture and
the rupture lines laid over it. Looking at the image, one can see that the main rupture is well
classified and easily distinguishable from any other features in the image. It is even possible to
observe individual branches of the rupture. When the image is magnified further, some smaller
surface ruptures are also observable though not to the same degree. Groups of "rupture" pixels
form linear patterns along those lines, but there are breaks along the linear patterns, showing that
it is not capturing the entire rupture. Overall, all the methodology appears to work though the
results would be better if the first two test runs were replaced with test runs like the latter three.
Figure 15 Close up of surface rupture segment.
38
Figure 16 Close up of surface rupture segment with ground truthed surface rupture lines from
USGS.
39
Creating an average of the results of all the confusion matrices reveals the overall success
of the classification though some issues are found (Table 4). The "No rupture" class had a user
accuracy of 97.0% and 77.0%, which shows that the classification method was highly effective
at detecting and classifying the class. This was the largest class across the entire raster, as
evidenced by the number of sampling points correctly classified in the confusion matrix. The
results are trustworthy based on this fact. The second-largest class was “vegetation," which also
had good results, with a user accuracy of 78.0% and a producer accuracy of 75.6%. When
comparing the results of the classifications and the original raster, it can be seen how well the
SVM classification is able to differentiate between the classes. That is expected since the optical
properties are highly distinct between the two. The next class to look at is the “rupture” class, the
smallest class by a large degree. Only 47 sample points out of 2496 were classified correctly as
"rupture," so the results may be skewed because of this small sample size. User accuracy for this
class is exceptionally low at 10.0%, which would typically point to the method failing at
detecting the "rupture" cells, though this may not be true. It is likely that the results are
influenced by the poor results from Test 1 and Test 2, as well as the low number of sample points
in the confusion matrix. Producer accuracy is close in value to the other classes at 77.1%. Like
the individual test runs, this result needs to be carefully considered based on the low sampling.
The averaged total accuracy for all the tests is at 76.7%, which is a good result, but it does not
reflect the poor outcome of the "rupture" user accuracy. Results for the kappa coefficient were
not included since it cannot be averaged from the results of the five matrices.
40
Table 4. Confusion Matrix (or Error Matrix) for Combined Test Runs
Class Value Not Rupture Rupture Vegetation Total User Accuracy
Not Rupture 1463 11 33 1508
97.0%
Rupture 324 47 98 469 10.0%
Vegetation 112 2 405 519 78.0%
Total 1899 61 536 2496
Producer Acc. 77.0% 77.1% 75.6%
76.7%
41
Chapter 5 Discussion
The results of this project showed that the proposed methodology is only partially
applicable, as is, to this research site. In order to assess this, the classification results from each
test were compared to a dataset with the known locations of the surface ruptures. SVM was able
to capture the primary section of fault rupture but failed to capture the smaller strands of the
ruptures. In addition, it misclassified some of the paths and streams, or at least the edges, and the
areas around bushes as rupture. It is not clear why this is, but the areas around bushes likely have
similar optical properties to the soil disturbed by the surface rupture. This may be because the
shrubs are casting a shadow, or they may keep the soil moister, or it could be the decay of
material from the bushes. These are all possibilities but finding the cause is beyond the scope of
this project. Paths and streams were most likely misclassified because they share some
similarities with the ruptures. The ruptures probably share some level of linearity with them, as
well as similar coloration, likely caused by the fact that in these cases, the soil has been
disturbed, which reveals lower soils.
It is important to note that the training data used seems to have a substantial effect on the
results of this project. The first two test runs did not seem to get as accurate of a classification of
the main surface rupture as the last three test runs. While everything was kept consistent between
the five tests, the only difference was the number of training sample polygons used for training
the classifier. If Test Runs 1 and 2 had the same volume of training data, the resultant combined
user accuracy would have been higher. From this, one can infer that an overabundance of
training samples does not hurt classification in this case.
If this method had worked, it would have been a boon to response teams following
earthquakes and possibly other hazards. Despite it appearing to not work fully as intended,
42
similar methods, such as those used by Zhang et al, have worked before. This shows that the
method has potential that should be further researched. Anything that can accelerate the detection
of hazardous areas that rescue and repair crews can target following a natural disaster is essential
to help save people.
5.1. Future Work
5.1.1 Additional Data Sets
As mentioned previously, when the software was classifying the imagery, parts of the
study area were misclassified. This is because the optical properties the classification tool is
detecting are similar between the rupture and regions around the paths and bushes. One solution
for dealing with this is the inclusion of other datasets that cover different parts of the
electromagnetic spectrum. The best type of imagery dataset that may help with this is
hyperspectral imagery. While the multi-spectral imagery used in this study is composed of three
bands, hyperspectral imagery is usually composed of over one hundred bands. Each band is
composed of a small segment of the light spectrum. With hundreds of small bands, the software
would be able to bin a smaller range of values for the rupture classification and likely decrease
the number of misclassified pixels. A possible way to analyze the additional raster imagery is to
use Empirical mode decomposition (Adu-Gyamfi et. al. 2011). This method takes the imagery of
various resolutions and breaks them apart into smaller band components that represent unique
differences between the imagery. Those smaller components can be recombined or filtered to
bring out linear features.
5.1.2 Methods Outside of ArcGIS
ArcGIS is limited in the number of tools that are natively available for classification.
There are other methods that may be better suited for extracting the fault ruptures from imagery,
43
such as those covered in Mohan and Poobal’s (2018) review on crack detection in concrete.
These methods should be applicable since the properties of the concrete cracks should be like
those of fault ruptures. Both are linear features, usually not straight, and considerably darker than
the surrounding surface.
Some, if not all, of the methods that were used can be done in the Python programing
language with readily available libraries. ArcGIS already includes a python compiler and its own
python library. Python can be used to integrate ArcGIS functions with some of these various
libraries, such as those used for machine learning, like TensorFlow or Pytorch. Bringing python
into this would open up a large number of new methods for performing this research.
5.1.3 Computational Performance Increase
Data storage and time became issues when performing this project. This entire project
was performed on a virtual desktop that has limited storage and can only be accessed for 5 hours
at a time. These factors constrained the processed that could be performed to only those that fit
into that time frame and the allotted storage. If these limits were not in place, additional testing
on the imagery for the Mw 6.4 or performing the methods on the imagery from the six other days
could have been performed. This would have added to this project, creating more statistically
significant results to the project, but the space needed was far more than was available. So, in the
future, it would help to perform this on a desktop computer which can be left on for processes to
be completed and data issues are easier to manage. There were 14 total UAV images from the
two sites across the seven days that were available. A model could be created to run the steps
that were performed on the image used for this thesis without interruption. The only human input
needed would be the initial creation of the training data.
44
5.1.4 Additional Scenarios
Methods used in this project need to be tested in other locales besides a desert area like
Ridgecrest, CA. Imagery collected from other types of terrain will have different spectral
properties than those in the imagery used in this project. One type of terrain it would be
beneficial to test in is urban and suburban areas. These areas have more buildings, more
underground utilities, and higher populations. Being able to quickly locate areas that have been
damaged will help save lives. A forested area would also be an area to consider testing in. The
tree cover makes it hard to observe the ground surface in this sort of environment, but it is
possible to capture imagery of it with a LiDAR system. These terrains, and other settings, have
their own qualities that can hinder the observation of the fault ruptures. Working out how to deal
with these hindrances ahead of time is a criterion for applying this methodology to the real
world.
Testing it in a real-world scenario, or at least a simulation of one, would also be
insightful. One of the usages for this method is to use it for directing safety crews to the location
of the damage but to make sure it assists them with their duties; testing needs to be performed.
The simulation can include different situations, such as locating damaged roads, utilities, or
buildings and any individuals hurt by those hazards. It could be integrated into a dashboard that
can find the rupture and be used to estimate damage costs, track first responders, keep track of
hospital loads, or any other piece of information that policymakers, first responders, and other
members of government need.
45
References
Adu-Gyamfi, Y. O., N. O. Okine, Gonzalo Garateguy, Rafael Carrillo, and Gonzalo R. Arce.
“Multiresolution Information Mining for Pavement Crack Image Analysis.” Journal of
Computing in Civil Engineering 26, no. 6 (2012): 741–749.
Anniballe, Roberta, Fabrizio Noto, Tanya Scalia, Christian Bignami, Salvatore Stramondo,
Marco Chini, and Nazzareno Pierdicca. “Earthquake Damage Mapping: An Overall Assessment
of Ground Surveys and VHR Image Change Detection after L'Aquila 2009 Earthquake.” Remote
Sensing of Environment 210 (2018): 166–178.
Ci, Tianyu, Zhen Liu, Ying Wang, Qigen Lin, and Di Wu. “An Automated Technique for
Damage Mapping after Earthquakes by Detecting Changes between High-Resolution Images”
(2018).
CRED. Natural Disasters 2019. Brussels: CRED; 2020.
Cunningham, Dickson, Stephen Grebby, Kevin Tansey, Andrej Gosar, and Vanja Kastelic.
“Application of Airborne LiDAR to Mapping Seismogenic Faults in Forested Mountainous
Terrain, Southeastern Alps, Slovenia.” Geophysical Research Letters 33, no. 20 (2006).
Dong, Laigen, and Jie Shan. “A Comprehensive Review of Earthquake-Induced Building
Damage Detection with Remote Sensing Techniques.” ISPRS Journal of Photogrammetry and
Remote Sensing 84 (2013): 85–99.
Donnellan, Andrea, Gregory Lyzenga, Adnan Ansar, Christine Goulet, Jun Wang, and Marlon
Pierce. “Targeted High-Resolution Structure from Motion Observations over the Mw 6.4 and 7.1
Ruptures of the Ridgecrest Earthquake Sequence.” Seismological Research Letters 91, no. 4
(2020): 2087–2095.
DuRoss, Christopher B., Gold R., Dawson, T., Scharer, K.., Kendrick, K., Akciz, S., Angster, S.,
Bachhuber, J., Bacon, S., Bennett, S., Blair, L., Brooks, B., Bullard, T., Burgess, W.P., Chupik,
C., DeFrisco, M., Delano, J., Dolan, J., Frost, E., Graehl, N., Haddon, E., Hatem, A., Hernandez,
J., Hitchcock, C., Hudnut, K., Jobe, J.T., Koehler, R., Kozaci, O., Ladinsky, T., Madugo, C.,
McPhillips, D., Milliner, C., Morelan, A., Olson, B., Patton, J., Philibosian, B., Pickering, A.J.,
Pierce, I., Ponti, D., Seitz, G., Spangler, E., Swanson, B., Thomas, T., Treiman, J., Valencia, F.,
Williams, A., and Zinke, R.. Surface Displacement Observations of the 2019 Ridgecrest,
California Earthquake Sequence: U.S. Geological Survey data release. (2020).
Duro, Dennis C., Steven E. Franklin, and Monique G. Dubé. “Multi-Scale Object-Based Image
Analysis and Feature Selection of Multi-Sensor Earth Observation Imagery Using Random
Forests.” International Journal of Remote Sensing 33, no. 14 (2012): 4502–4526.
46
Duro, Dennis C., Steven E. Franklin, and Monique G. Dubé. “A Comparison of Pixel-Based and
Object-Based Image Analysis with Selected Machine Learning Algorithms for the Classification
of Agricultural Landscapes Using SPOT-5 HRG Imagery.” Remote Sensing of Environment 118
(2012): 259–272.
Fielding, E. J. “Surface Ruptures and Building Damage of the 2003 Bam, Iran, Earthquake
Mapped by Satellite Synthetic Aperture Radar Interferometric Correlation.” Journal of
Geophysical Research 110, no. B3 (2005).
Guo, Bo, Qingquan Li, Xianfeng Huang, and Chisheng Wang. “An Improved Method for Power-
Line Reconstruction from Point Cloud Data.” Remote Sensing 8, no. 1 (2016): 36.
Jain, Sanjay, and Ashok K Dubey. “Image Change Detection for Earthquake Damage
Assessment.” International Journal of Advanced Research in Computer Science 2, no. 5 (2011):
85–89.
James, Gareth, Daniela Witten, Trevor Hastie, and Robert Tibshirani. Essay. In An Introduction
to Statistical Learning: with Applications in R, 181–182. New York, New York: Springer, 2017.
Jin, Kwangmin, and Young-Seog Kim. “The Importance of Surface Ruptures and Fault Damage
Zones in Earthquake Hazard Assessment: a Review and New Suggestions.” Geological Society,
London, Special Publications (2019).
Lei Ma, Manchun Li, Xiaoxue Ma, Liang Cheng, Peijun Du, Yongxue Liu. “A review of
supervised object-based land-cover image classification.” ISPRS Journal of Photogrammetry and
Remote Sensing 130, (2017): 277-293.
Li, Zhengrong, Rodney Walker, Ross Hayward, and Luis Mejias. “Advances in Vegetation
Management for Power Line Corridor Monitoring Using Aerial Remote Sensing Techniques.”
2010 1st International Conference on Applied Robotics for the Power Industry (CARPI 2010)
(2010).
Matikainen, Leena, Matti Lehtomäki, Eero Ahokas, Juha Hyyppä, Mika Karjalainen, Anttoni
Jaakkola, Antero Kukko, and Tero Heinonen. “Remote Sensing Methods for Power Line
Corridor Surveys.” ISPRS Journal of Photogrammetry and Remote Sensing 119 (2016): 10–31.
Meixner, J., J.C. Grimmer, A. Becker, E. Schill, and T. Kohl. “Comparison of Different Digital
Elevation Models and Satellite Imagery for Lineament Analysis: Implications for Identification
and Spatial Arrangement of Fault Zones in Crystalline Basement Rocks of the Southern Black
Forest (Germany).” Journal of Structural Geology 108 (2018): 256–268.
Mohan, Arun, and Sumathi Poobal. “Crack Detection Using Image Processing: A Critical
Review and Analysis.” Alexandria Engineering Journal 57, no. 2 (2018): 787–798.
47
Quackenbush, Lindi J. “A Review of Techniques for Extracting Linear Features from Imagery.”
Photogrammetric Engineering & Remote Sensing 70, no. 12 (2004): 1383–1392.
Ross, Zachary E., Benjamín Idini, Zhe Jia, Oliver L. Stephenson, Minyan Zhong, Xin Wang,
Zhongwen Zhan, et al. “Hierarchical Interlocked Orthogonal Faulting in the 2019 Ridgecrest
Earthquake Sequence.” Science 366, no. 6463 (2019): 346–351.
Sare, Robert, George E. Hilley, and Stephen B. DeLong. “Regional‐Scale Detection of Fault
Scarps and Other Tectonic Landforms: Examples From Northern California.” Journal of
Geophysical Research: Solid Earth 124, no. 1 (2019): 1016–1035.
Sofina, Natalia, and Manfred Ehlers. “Building Change Detection Using High Resolution
Remotely Sensed Data and GIS.” IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing 9, no. 8 (2016): 3430–3438.
Tehrany, Mahyat Shafapour, Biswajeet Pradhan, and Mustafa Neamah Jebuv. “A Comparative
Assessment between Object and Pixel-Based Classification Approaches for Land Use/Land
Cover Mapping Using SPOT 5 Imagery.” Geocarto International 29, no. 4 (2013): 351–369.
Thanh Noi, Phan, and Martin Kappas. “Comparison of Random Forest, k-Nearest Neighbor, and
Support Vector Machine Classifiers for Land Cover Classification Using Sentinel-2 Imagery.”
Sensors 18, no. 2 (2017): 18.
Zhang, Fan, Zhenqi Hu, Yaokun Fu, Kun Yang, Qunying Wu, and Zewei Feng. “A New
Identification Method for Surface Cracks from UAV Images Based on Machine Learning in Coal
Mining Areas.” Remote Sensing 12, no. 10 (2020): 1571.
48
Appendix A: Test Run 2 Raster, Classification, and Error Matrix
49
Class Value No Rupture Rupture Vegetation Total User Accuracy Kappa
No Rupture 227 3 8 238 95.4%
Rupture 136 13 27 176 7.4%
Vegetation 16 0 70 86 81.4%
Total 380 16 104 500
Producer
Accuracy
59.7% 81.3% 67.3%
62%
Kappa
0.358
50
Appendix B: Test Run 3 Raster, Classification, and Error Matrix
51
Class Value No Rupture Rupture Vegetation Total User Accuracy Kappa
No Rupture 322 1 10 333 96.7%
Rupture 40 6 19 65 9.1%
Vegetation 30 1 70 101 69.3%
Total 392 8 99 499
Producer
Accuracy
82.4% 66.7% 70.7%
79.8%
Kappa
0.531
52
Appendix C: Test Run 5 Raster, Classification, and Error Matrix
53
Class Value No Rupture Rupture Vegetation Total User Accuracy Kappa
No Rupture 309 4 6 319 96.6%
Rupture 20 10 19 49 20.0%
Vegetation 33 0 97 130 74.6%
Total 362 14 122 500
Producer
Accuracy
85.1% 71.4% 78.9%
83.5%
Kappa
0.645
Abstract (if available)
Abstract
One possible costly outcome of large earthquakes is breakages in the ground surface, known as surface ruptures. Surface ruptures can cause damage to human infrastructure as well as harm humans. Detailed field studies that trace these structures take substantial time and effort because they affect large regions, but this is reducible by collecting remote sensed imagery performing supervised classification on the imagery with a GIS. For this study, unmanned aerial vehicles (UAV) imagery from the Ridgecrest area in California was analyzed using the support vector machine (SVM) classification method to attempt surface rupture detection in desert terrain. The imagery covers a small area, so a K-Folds analysis method was used to attain statistically significant results. This involved splitting the imagery into 5 random pools of 30-meter by 30-meter squares before running the classification. The results of each classification were analyzed by generating confusion matrices and visual inspection. An average of the five confusion matrices was used for a final analysis. While this method did classify larger segments (> 0.5-meter-wide) of surface rupture that was in the image, it missed most of the smaller surface rupture segments (< 0.5-meter-wide). In addition, the technique misclassified parts of the imagery as surface rupture, especially around the vegetation, road paths, and amongst a rock field in the south-east corner. Further testing should be done with this method, including using it on imagery with different land-cover. Based on the results of the further testing it may be ready to try in practice scenarios for real-life earthquake disasters.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Evaluating machine learning tools for humanitarian road network mapping
PDF
An application of aerial drones in high definition mapping for autonomous vehicles
PDF
Assessing modern conflict to monitor human rights with remote sensing: Russia's war in Ukraine
PDF
A comparison of pixel-based, object-based, deep learning, and data fusion image classification approaches in delineating urban tree canopy in City Terrace
PDF
Object detection and digitization from aerial imagery using neural networks
PDF
Mixed forest image classification of paper birch: using AVIRIS bandwidths ranging from 530 to 745 nm
PDF
Operational optimization model for Hungry Marketplace using geographic information systems
PDF
Detection and accuracy assessment of mountain pine beetle infestations using Landsat 8 OLI and WorldView02 satellite imagery: Lake Tahoe Basin-Nevada and California
PDF
Analyzing earthquake casualty risk at census block level: a case study in the Lexington Central Business District, Kentucky
PDF
Improving wetland determination utilizing unmanned aerial systems
PDF
Geographic object based image analysis for utility scale photovoltaic site suitability studies
PDF
Classifying ash cloud attributes of Eyjafjallajökull volcano, Iceland, using satellite remote sensing
PDF
Development and evaluation of a USV based mapping system for remote sensing of eelgrass extent in southern California
PDF
Volumetric interactions between major ruptures and fault zones illuminated by small earthquake properties
PDF
Impacts of vegetation management on wildfire severity: a study of the 2021 Caldor fire
PDF
Geodatabase for archaeogenetics: ancient peoples and family lines
PDF
Exploring remote sensing and geographic information systems technologies to understand vegetation changes in response to land management practices at Finke Gorge National Park, Australia Between ...
PDF
An examination of close-range photogrammetry and traditional cave survey methods for terrestrial and underwater caves for 3-dimensional mapping
PDF
Wetland mapping and restoration decision making using remote sensing and spatial analysis: a case study at the Kawainui Marsh
PDF
Comparing Landsat7 ETM+ and NAIP imagery for precision agriculture application in small scale farming: a case study in the south eastern part of Pittsylvania County, VA
Asset Metadata
Creator
Cooper, Jordan M.
(author)
Core Title
Surface rupture detection with support vector machine classification: case study from Ridgecrest, CA
School
College of Letters, Arts and Sciences
Degree
Master of Science
Degree Program
Geographic Information Science and Technology
Degree Conferral Date
2021-08
Publication Date
07/18/2021
Defense Date
07/15/2021
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
earthquake swarm,Geology,image classification,OAI-PMH Harvest,remote sensing,Ridgecrest,support vector machine,surface rupture,SVM,UAV,unmanned aerial vehicle
Format
application/pdf
(imt)
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Fleming, Steven (
committee chair
), Chiang, Yao-Yi (
committee member
), Wu, An-Min (
committee member
)
Creator Email
jordan.coop49@gmail.com,jordanc@usc.edu
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-oUC15602350
Unique identifier
UC15602350
Legacy Identifier
etd-CooperJord-9765
Document Type
Thesis
Format
application/pdf (imt)
Rights
Cooper, Jordan M.
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 author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.
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
Repository Email
cisadmin@lib.usc.edu
Tags
earthquake swarm
image classification
remote sensing
Ridgecrest
support vector machine
surface rupture
SVM
UAV
unmanned aerial vehicle