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
/
Use of remotely sensed imagery to map sudden oak death (Phytophthora ramorum) in the Santa Cruz Mountains
(USC Thesis Other)
Use of remotely sensed imagery to map sudden oak death (Phytophthora ramorum) in the Santa Cruz Mountains
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
USE OF REMOTELY SENSED IMAGERY TO MAP SUDDEN OAK DEATH
(PHYTOPHTHORA RAMORUM) IN THE SANTA CRUZ MOUNTAINS
by
Trinka Gillis
A Thesis Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
MASTER OF SCIENCE
(GEOGRAPHIC INFORMATION SCIENCE AND TECHNOLOGY) May 2014
Copyright 2014 Trinka Gillis
ii
DEDICATION
I dedicate this work to Dr. Wayne Caplinger, whose enthusiasm for life was only equaled by its
unexpected brevity, and to all whose lives are cut short before their time.
iii
ACKNOWLEDGEMENTS
I would like to thank my advisor, Dr. John Wilson, for his valuable feedback and faith in my
work, and committee members Dr. Su Jin Lee and Dr. Tarek Rashed, whose advice and
suggestions were indispensable in carrying out this project. I would also like to thank my
professors in the GIST program, whose lessons made this thesis possible. Last but not least, I
would like to acknowledge the support of my husband, who gave up a year of weekends so that I
could pursue this dream.
iv
TABLE OF CONTENTS
Dedication....................................................................................................................................... ii
Acknowledgements........................................................................................................................ iii
List of Tables ................................................................................................................................. vi
List of Figures............................................................................................................................... vii
List of Abbreviations ................................................................................................................... xiii
Abstract......................................................................................................................................... xv
Chapter One: Introduction .............................................................................................................. 1
1.1 Sudden Oak Death............................................................................................................. 1
1.2 Role of Remote Sensing in SOD Studies .......................................................................... 4
1.3 Research Questions ........................................................................................................... 5
1.4 Thesis Outline.................................................................................................................... 7
Chapter Two: Related Work ........................................................................................................... 9
2.1 Biology of SOD................................................................................................................. 9
2.2 Remote Sensing of Tree Health....................................................................................... 12
2.3 Mapping of SOD ............................................................................................................. 21
Chapter Three: Data and Methodology......................................................................................... 23
3.1 Description of Study Area............................................................................................... 23
3.2 Data ................................................................................................................................. 25
3.2.1 Landsat Satellite Data .............................................................................................. 26
3.2.2 National Land Cover Database Classified Raster Data ........................................... 29
3.2.3 Fire and Resource Assessment Program Vector Data ............................................. 30
3.2.4 USDA Aerial Detection Survey Vector Data .......................................................... 31
3.2.5 Google Earth Historical Aerial Imagery.................................................................. 32
3.2.6 Land Cover Points ................................................................................................... 32
3.2.7 Results Validation Points......................................................................................... 33
v
3.3 Methodology ................................................................................................................... 34
3.3.1 Refining the Study Area .......................................................................................... 35
3.3.2 Identifying the Most Effective Index and Mapping 2011 Affected Areas .............. 43
3.3.3 Accuracy Assessment .............................................................................................. 44
3.4 Summary ......................................................................................................................... 47
Chapter Four: Results ................................................................................................................... 48
4.1 Mapping Results.............................................................................................................. 48
4.1.1 Map of Serious Tree Death in 2011......................................................................... 48
4.1.2 Serious Tree Death Change From Previous Years .................................................. 51
4.1.3 Area of SOD infestation .......................................................................................... 56
4.1.4 2011 Comparison to ADS and SODMAP ............................................................... 56
4.2 Accuracy Evaluation ....................................................................................................... 60
Chapter Five: Discussion and Conclusions................................................................................... 63
References..................................................................................................................................... 69
Appendix A: Hosts regulated for Phytophthora ramorum ........................................................... 74
Appendix B: Index Analysis Histograms for Land Cover Points................................................. 76
Appendix C: Predicted Maps of Sudden Oak Death .................................................................... 84
Appendix D: Histograms for Results Validation Points............................................................. 102
vi
LIST OF TABLES
Table 1: Multi-spectral indices used to detect vegetation health 14
Table 2: Data used for this analysis 25
Table 3: Landsat 5 TM bands 26
Table 4: Landsat 5 TM scenes used in project (path 44 row 34) 28
Table 5: NLCD land cover classes 29
Table 6: Land Cover Points classification categories 33
Table 7: Results Validation Points within study area 34
Table 8: Indices and value ranges tested to determine which one best masked
shrub without masking healthy trees. The ranges indicate the values
used to map healthy trees. 41
Table 10: Error matrix for SWIR/NIR index applied to Results Validation Points
with range of 400-570 classified as SD 45
Table 11: Error matrix for NBR index applied to Results Validation Points with
range of 575-725 classified as SD 46
Table 12: Error matrix for OR Combination of NBR 575-725 and SWIR/NIR
400-570 indices applied to Results Validation Points 46
Table 13: Number of pixels within study area identified as containing serious
levels of tree mortality 55
Table A1: Proven host plants regulated for Phytophthora Ramorum 74
vii
LIST OF FIGURES
Figure 1: Sudden Oak Death at Mescal Ridge, Carmel, Big Sur, Monterey County
2012......................................................................................................................... 2
Figure 2: Predicted spread risk for P. ramorum in northern California ................................. 3
Figure 3: Coast live oak showing the bleeding ulcers typical of Sudden Oak Death .......... 10
Figure 4: California bay laurel (Umbellularia californica) infected with Ramorum
blight ..................................................................................................................... 11
Figure 5: Risk for SOD distribution on the west coast. The value indicates the
number of models used to predicted risk in that area. The inset shows
confirmed infections as of 2007............................................................................ 19
Figure 6: A sample image from SODMAP by Oakmapper, displaying PR-infected
trees in the Santa Cruz Mountains ........................................................................ 22
Figure 7: The area of interest in the Santa Cruz Mountains of central California ............... 24
Figure 8: The methodology for data extraction and analysis that was created and
used for this project............................................................................................... 35
Figure 9: The NLCD classification did not adequately remove non-forested areas
from the study area. Areas outlined in white were identified as non-forest
classes using the NLCD 1992 and 2011 layers. This sample shows many
non-forested areas were not correctly classified................................................... 37
Figure 10: The NBR Land Cover Points histogram for shrubs .............................................. 38
Figure 11: The NBR Land Cover Points histogram for healthy trees .................................... 38
Figure 12: The SWIR/NIR Land Cover Points histogram for shrubs .................................... 39
Figure 13: The SWIR/NIR Land Cover Points histogram for healthy trees .......................... 39
Figure 14: The NBR Land Cover Points histogram for shrubs .............................................. 40
Figure 15: The NBR Land Cover Points histogram for healthy trees .................................... 40
Figure 16: The NLCD non-forest layers, shown in solid white, were not sufficient to
limit the mask to only forested areas. Landsat 1992 and 1993 data, used
to supplement the mask, removed all white-outlined areas from the study
area. Figure 9 shows the NLCD non-forest layers without the
supplemental mask................................................................................................ 42
Figure 17: The 2011 SD extent in the study area. Areas with lower rates of tree
death – Areas A, B and C – are marked in green.................................................. 49
viii
Figure 18: 2010 SD and new 2011 SD points (at least 31 meters away from a
2011 point)............................................................................................................ 52
Figure 19: A point density map showing the areas with the most new SD in 2011
compared to 2010. Detail can be seen in Figure 18.............................................. 53
Figure 20: Detail of new SD points in 2011........................................................................... 54
Figure 21: The 2011 SD pixels compared to ADS-identified SOD areas.............................. 58
Figure 22: The 2011 SD pixels compared to SODMAP confirmed SOD infections............. 59
Figure B1: The SWIR/NIR Land Cover Points histogram for serious tree death. 76
Figure B2: The SWIR/NIR Land Cover Points histogram for healthy trees. This
index is a good option because the peaks of the two histograms are
separated and don’t overlap with large quantities in the other graph.
This index was one of three considered to differentiate healthy from
SD points, and the range of 400-570 was ultimately used.................................... 76
Figure B3: The NBR Land Cover Points histogram for serious tree death. ........................... 77
Figure B4: The NBR Land Cover Points histogram for healthy trees. This index
was not chosen because the left HLH peak overlaps with three of the
SD points............................................................................................................... 77
Figure B5: The NDMI Land Cover Points histogram for serious tree death. ......................... 78
Figure B6: The NDMI Land Cover Points histogram for healthy trees. This index
was considered for use to differentiate the healthy and SD points.
Although many of the points are in overlapping ranges, the peaks of the
histograms are separated and do not overlap with points in the other class.. ....... 78
Figure B7: The NDVI Land Cover Points histogram for serious tree death........................... 79
Figure B8: The NDVI Land Cover Points histogram for healthy trees. This index
is not ideal because the main curve of each histogram is so close to the
other, and most of the points are in overlapping ranges. ...................................... 79
Figure B9: The RGI Land Cover Points histogram for serious tree death.............................. 80
Figure B10: The RGI Land Cover Points histogram for healthy trees. This index was
not used because here is too much overlap between the ranges and the
peaks. .................................................................................................................... 80
Figure B11: The Tasseled Cap Bright Land Cover Points histogram for serious
tree death............................................................................................................... 81
ix
Figure B12: The Tasseled Cap Bright Land Cover Points histogram for healthy
trees. This index was not used because there was no large a range of
values in each chart, no clear histogram peak, and too much overlap.................. 81
Figure B13: The Tasseled Cap Green Land Cover Points histogram for serious
tree death............................................................................................................... 82
Figure B14: The Tasseled Cap Green Land Cover Points histogram for healthy
trees. This index was not used because the values are spread over
too great a range with too much overlap between the two classes. ...................... 82
Figure B15: The Tasseled Cap Wet Land Cover Points histogram for serious
tree death............................................................................................................... 83
Figure B16: The Tasseled Cap Wet Land Cover Points histogram for healthy trees.
This index was considered for use to differentiate the healthy and SD
points. Although there is undesireable overlap between the ranges,
the histogram peaks are separated and there are few overlapping
points at the peak ranges. ...................................................................................... 83
Figure C1: The extent of serious tree death in 1994. Each pixel classified as SD is
displayed as a single yellow point. Non-forested areas, masked out, are
shown in brown. 84
Figure C2: The extent of serious tree death in 1995. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 85
Figure C3: The extent of serious tree death in 1996. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 86
Figure C4: The extent of serious tree death in 1997. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 87
Figure C5: The extent of serious tree death in 1998. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 88
Figure C6: The extent of serious tree death in 1999. This year’s data showed
the fewest pixels classified as SD. Each pixel classified as SD is
displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 89
Figure C7: The extent of serious tree death in 2000. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 90
x
Figure C8: The extent of serious tree death in 2001. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 91
Figure C9: The extent of serious tree death in 2002. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 92
Figure C10: The extent of serious tree death in 2003. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 93
Figure C11: The extent of serious tree death in 2004. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 94
Figure C12: The extent of serious tree death in 2005. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 95
Figure C13: The extent of serious tree death in 2006. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 96
Figure C14: The extent of serious tree death in 2007. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 97
Figure C15: The extent of serious tree death in 2008. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown........................................................................................ 98
Figure C16: The extent of serious tree death in 2009. This year’s data showed the
most pixels classified as SD. Each pixel classified as SD is displayed
as a single yellow point. Non-forested areas, masked out, are shown
in brown. ............................................................................................................... 99
Figure C17: The extent of serious tree death in 2010. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown...................................................................................... 100
Figure C18: The extent of serious tree death in 2011. Each pixel classified as SD
is displayed as a single yellow point. Non-forested areas, masked
out, are shown in brown...................................................................................... 101
xi
Figure D1: The SWIR/NIR Results Validation Points histogram for serious
tree death............................................................................................................. 102
Figure D2: The SWIR/NIR Results Validation Points histogram for healthy
trees. The main range for the SD points in Figure D1 is shifted
slightly to the right in comparison to the HLH points. ....................................... 102
Figure D3: The NBR Results Validation Points histogram for serious tree
death.................................................................................................................... 103
Figure D4: The NBR Results Validation Points histogram for healthy trees.
The main curve for SD points, in Figure D3, is shifted slightly
to the left in comparison to the HLH points. ...................................................... 103
Figure D5: The NDMI Results Validation Points histogram for serious tree
death.................................................................................................................... 104
Figure D6: The NDMI Results Validation Points histogram for healthy trees.
The main range of the SD points in Figure D5 is shifted slightly
to the left in comparison to the HLH graph. ....................................................... 104
Figure D7: The NDVI Results Validation Points histogram for serious tree
death.................................................................................................................... 105
Figure D8: The NDVI Results Validation Points histogram for healthy trees.
The main body of SD points in Figure D7 is shifted slightly to
the right in comparison to HLH. ......................................................................... 105
Figure D9: The RGI Results Validation Points histogram for serious tree
death.................................................................................................................... 106
Figure D10: The RGI Results Validation Points histogram for healthy trees.
There is correlation between the main sets of points in these
ranges so this index is not useful for differentiating the classes......................... 106
Figure D11: The Tasseled Cap Bright Results Validation Points histogram
for serious tree death........................................................................................... 107
Figure D12: The Tasseled Cap Bright Results Validation Points histogram
for healthy trees. The values are spread over too wide a range
to be able to use this index to identify a change in spectral
signature.............................................................................................................. 107
Figure D13: The Tasseled Cap Green Results Validation Points histogram for
serious tree death................................................................................................. 108
xii
Figure D14: The Tasseled Cap Green Results Validation Points histogram for
healthy trees. The values are spread over too wide a range to be
able to use this index to identify a change in spectral signature......................... 108
Figure D15: The Tasseled Cap Wet Results Validation Points histogram for
serious tree death................................................................................................. 109
Figure D16: The Tasseled Cap Wet Results Validation Points histogram for
healthy trees. The values are spread over too wide a range to be
able to use this index to identify a change in spectral signature. ........................ 109
xiii
LIST OF ABBREVIATIONS
ADAR Airborne Data Registration
ADS USDA Forest Health Monitoring Program Aerial Detection Survey
AOI Area of Interest
BLM Bureau of Land Management
CalFire California Department of Forestry and Fire Protection
CALVEG Classification and Assessment with Landsat of Visible Ecological Groupings
CDR Climate Data Record
COMTF California Oak Mortality Task Force
FRAP Fire and Resource Assessment Program
GAP California Gap Analysis Project
HLH Healthy Tree Land Cover
MODIS Moderate Resolution Imaging Spectroradiometer
NAIP National Agricultural Imagery Program
NASA National Aeronautics and Space Administration
NBR Normalized Burn Ratio
NDII Normalized Difference Infrared Index
NDMI Normalized Difference Moisture Index
NDVI Normalized Difference Vegetation Index
NLCD National Land Cover Database
NPS National Park Service
PR Phytophthora ramorum
RGI Red Green Index
RMSE Geometric Root Mean Square Error
SD Serious Tree Death
xiv
SLC Scan Line Corrector
SOD Sudden Oak Death
SWIR/NIR Short Wave Infrared/Near Infrared
TC Bright Tasseled Cap Brightness
TC Green Tasseled Cap Greenness
TC Wet Tasseled Cap Wetness
TM Thematic Mapper
USDA US Department of Agriculture
USGS US Geological Survey
xv
ABSTRACT
This project sought a method to map Sudden Oak Death distribution in the Santa Cruz
Mountains of California, a coastal mountain range and one of the locations where this disease
was first observed. The project researched a method to identify forest affected by SOD using 30
m multi-spectral Landsat satellite imagery to classify tree mortality at the canopy-level
throughout the study area, and applied that method to a time series of data to show pattern of
spread. A successful methodology would be of interest to scientists trying to identify areas which
escaped disease contagion, environmentalists attempting to quantify damage, and land managers
evaluating the health of their forests. The more we can learn about the disease, the more chance
we have to prevent further spread and damage to existing wild lands.
The primary data source for this research was springtime Landsat Climate Data Record
surface reflectance data. Non-forest areas were masked out using data produced by the National
Land Cover Database and supplemental land cover classification from the Landsat 2011 Climate
Data Record image. Areas with other known causes of tree death, as identified by Fire and
Resource Assessment Program fire perimeter polygons, and US Department of Agriculture
Forest Health Monitoring Program Aerial Detection Survey polygons, were also masked out.
Within the remaining forested study area, manually-created points were classified based on the
land cover contained by the corresponding Landsat 2011 pixel. These were used to extract value
ranges from the Landsat bands and calculated vegetation indices. The range and index which
best differentiated healthy from dead trees, SWIR/NIR, was applied to each Landsat scene in the
time series to map tree mortality. Results Validation Points, classified using Google Earth high-
resolution aerial imagery, were created to evaluate the accuracy of the mapping methodology for
the 2011 data.
xvi
Results indicated three areas which had largely escaped Sudden Oak Death infestation
and one area with high tree mortality that was not previously identified as Sudden Oak Death.
However, the methodology identified widespread tree death throughout the study area, including
in 1994, when little tree death should have been found. This indicated that healthy tree canopy
was able to produce a spectral signature matching that of pixels containing some dead trees. In
addition, the number of pixels classified as containing tree death varied widely from year to year,
suggesting that seasonal variation plays a much larger role in the spectral signature than
anticipated. Finally, an analysis of the Results Validation Points showed a high rate of false
positives, with only 24 percent mapping accuracy for tree death. This demonstrated conclusively
that the methodology and mapping results were unreliable.
The project demonstrates that Landsat data did not work for this study due to spectral
confusion and seasonal variation. Results might have been improved if a custom index was
devised to remove some of the false positives, if the definition of serious tree death was limited
to pixels containing a greater percentage of dead tree canopy, and/or if seasonal differences in
rainfall and temperature had been considered when choosing Landsat scenes to represent each
year.
1
CHAPTER ONE: INTRODUCTION
In the 1990s, oak trees suddenly began dying in the San Francisco Bay area of California.
Twenty years later, much is understood about the disease and cause, but no effective means of
control exists. This project sought to increase knowledge of Sudden Oak Death (SOD) by finding
a method to map tree morality over the 20-year period that the disease has been active in
California. Methods for mapping the disease and its pattern of spread could be adapted to other
areas to evaluate forest health, quantify the scale of the disease, and evaluate success at disease
management. The more we learn about the disease, the more tools we have to prevent further
spread and damage to existing wild lands.
1.1 Sudden Oak Death
The first observed occurrences of SOD were in 1994, in Mill Valley and Santa Cruz, California
(Garbelotto, Svihra and Rizzo 2001, Mascheretti et al. 2008). In 2000 the aerially-dispersed
oomycete, water mold, Phytophthora ramorum(PR) was identified as the causal agent(Rizzo,
Garbelotto and Hansen 2005). PR is believed to have been introduced to California via infected
rhododendron (Rhododendron spp.) and viburnum (Viburnum spp.) ornamental plants. The
fungus was transmitted via wind and rain to nearby wild lands, where it became established in
additional susceptible hosts.
On the central California coast, SOD primarily kills California black oak (Quercus
kelloggii), coast live oak (Quercus agrifolia) and tanoak (Lithocarpus densiflorus)(Barrett et al.
2006), but is transmitted by many other trees and shrubs, as well as by human activity. New SOD
outbreaks like those shown in Figure 1 are discovered each year. In the US, nurseries are closely
monitored for infected plants, and stream baiting techniques are used to determine where PR
2
might spread next. No cure exists for SOD, but preventive applications of anti-fungal agent have
successfully been applied to individual uninfected susceptible trees after nearby infected trees
were removed. The wide variety of carrier vegetation, low natural resistance to the disease, and
the natural means of contagion make it a difficult infection to fight.
Figure 1 Sudden Oak Death at Mescal Ridge, Carmel, Big Sur, Monterey County 2012
Photograph by Tom Coleman, US Forest Service
The total acreage affected by SOD in California has not been estimated. As of 2013, SOD
affects areas from southern Oregon to the San Luis Obispo area of California. Figure 2 shows
vulnerable habitat in northern California, and Meentemeyer et al. (2004) calculated that 10,305
km
2
(2.5 percent of California’s land area) are at high or very high risk for SOD infection.
Additional at-risk environments include the Sierra Nevada and the Appalachian mountain ranges
in the US, as well as parts of Europe and Asia. As of October 2013, efforts to contain PR
3
infection in the UK are responsible for the felling of over 16,000 ha(39,536 acres) of Japanese
larch(Forestry Commission UK 2014), an economically important timber tree. The large-scale
loss of California oak trees, which play a vital role in the coastal forest ecosystem, will have
serious consequences for wildlife and the landscape. Today’s forests will be unrecognizable to
future generations. More knowledge about the infection, including locations and patterns of
spread, may aid in attempts to minimize contagion.
Figure 2 Predicted spread risk for P. ramorum in northern California
Source: Meentemeyer et al. (2004)
4
1.2 Role of Remote Sensing in SOD Studies
Remote sensing has been used to increase SOD knowledge through modeling risk, identifying
infected trees, and mapping current disease locations. Modeling risk areas is valuable to manage
and prevent the spread of the disease, but mapping infected locations is necessary to evaluate
success at disease management.
The California central coast is widely accepted as the proper host environment for SOD,
but models based on remotely-sensed data have identified other susceptible regions. These
studies used variables such as vegetation, slope, aspect, precipitation, temperature and humidity
to map areas with a hospitable environment and vegetation likely to be infected. These risk
models covered large areas, from parts of California to the entire world. One worldwide study
(Kluza et al. 2007) predicted wide distribution of SOD from San Diego to Vancouver Island,
throughout the Sierra Nevada mountains, and in the southeastern US, with additional risk areas
in South America, southern Africa, southern Europe, the UK, and Asia, especially South Korea
and Japan. Risk maps are valuable to identify areas where land managers need to watch out for
SOD, but they highlight the fact that the disease has the potential to infect very large areas which
are not easily monitored from the ground.
Remotely-sensed imagery has been used to identify SOD infections in small areas using
canopy-level tree health as an indicator of disease presence. In the Big Sur area of central
California, south of this project’s study area, Meentemeyer et al. (2008) manually digitized dead
trees from 0.33 m resolution aerial imagery to calculate 20 percent mortality in infected areas. A
similar study(Liu, Kelly and Gong 2006) used 1 m Airborne Data Registration (ADAR) images
to classify land cover and map SOD in China Camp State Park, north of this project’s study area.
5
Both of these studies used high-resolution aerial imagery, the data storage needs and cost of
which often prevent application over a large area.
The only program which attempts to map SOD infection on a large scale is the United
States Department of Agriculture (USDA) Forest Health Monitoring Program Aerial Detection
Survey (ADS)(Heath et al. 2012) which uses aerial monitoring to map forest health hazards over
large areas. The USDA studies produce generalized maps of threats, focus on different hazards
and regions from year to year, and rely heavily on observation of changes from a moving
aircraft.
SOD has the potential to quickly infect large areas. The scale of the infestation and
inaccessible and remote locations prevent individual tree observation, so remote sensing is a
promising tool for disease monitoring. Many land management agencies have limited budgets, so
an effective monitoring method needs to use low-cost data in order to be applicable by others.
Moderate resolution data, such as Landsat, offers the opportunity to monitor large areas
repetitively with minimal data cost and minimal data storage requirements.
1.3 Research Questions
The biology of SOD has been extensively studied, and maps have been created showing
individual confirmed infections, but no methods exist to map SOD on a large scale. The goals of
this project were to find the best index to map SOD in the selected study area using Landsat data,
and apply that method to a series of annual images to analyze the rate and pattern of spread from
1994 to the present. A simple method to map SOD locations over a large area will give us an
estimate of the scope of the disease, and may attract more attention to the disease and funding for
research. The same method, applied to multiple years of imagery, would produce a view of the
6
disease over time, including the rate and pattern of spread. This could identify previously
unnoticed contaminated areas, identify areas which escaped damage, and lead to a better overall
understanding of the disease.
This is a pilot project which analyzed the accuracy of different vegetation indices at
differentiating tree mortality from healthy forest in the Santa Cruz Mountains on the central
California coast. The hope was that a successful method could be adapted and applied in other
areas where SOD threatens, to form a picture of the scope of SOD infestation throughout the
state. SOD spreads quickly, and many of its prime habitats are difficult to access. Remote
sensing provides the only practical monitoring approach in these areas. Landsat moderate-
resolution (30 m) satellite data, available from the US Geological Survey (USGS) is high-
quality, free and readily available for most of the world, which makes it the ideal tool for
organizations with limited budgets which cannot purchase higher resolution data. Landsat data is
re-collected every 16 days so the likelihood of finding suitable cloud-free data is high, and the
pixel size strikes a balance which requires some adjustments for mixed pixels, but minimizes
storage space and processing time.
Canopy-level tree health provides a glimpse of changes that are occurring at multiple
levels of the forest. A disease like SOD affects old and young trees, and when the tallest and
most well-established trees die, the devastation in the understory has already taken place.
Mapping SOD using canopy-level tree health as an indicator of disease extent has been
performed with high resolution imagery on a local scale, but high-resolution imagery prevents
large-scale analysis because of data costs and storage needs. This study attempted to map disease
extent using moderate-resolution data over a large area. Although all tree mortality in the study
area is not due to SOD, this study masked out areas where other circumstances were known to
7
have affected large numbers of trees. This study used canopy-level tree mortality as a proxy to
indicate areas where SOD has had the strongest impact on the environment.
The goals of this project were to find a method to identify SOD-affected forest in the
study area using Landsat data, and use a time series data set to identify the rate and pattern of
spread. Questions that needed to be addressed include:
1. Can canopy-level tree death indicating SOD infestation be accurately identified
and mapped using Landsat remotely-sensed imagery?
2. Are vegetation indices alone sufficient to differentiate tree health and tree
mortality from other types of land cover?
3. Will existing supplemental land cover data sets improve or simplify the
classification process?
The resulting methodology to evaluate SOD contagion in California would be of interest to
scientists, environmentalists and land managers who might adapt this methodology to map SOD
damage in other areas.
1.4 Thesis Outline
The next chapter provides background on SOD, a synopsis of studies analyzing tree health after
insect infestation, and a summary of the ways that remote sensing data has been used to further
our knowledge of SOD. An understanding of the origin, spread and control of SOD is necessary
to understand the need for additional information about the disease. Satellite data alone has never
been used to map SOD, but the tree health studies demonstrate successful techniques that were
used to map changes in tree health, which could be adapted to this study of SOD. In addition,
8
chapter two includes a discussion of the ways that remote sensing techniques or data have been
used to study SOD, and how this project will add to our knowledge.
Chapter 3 describes the data sources and methodology used in this study, which relied on
aerial imagery to classify land cover points, land classification raster and several vector layers to
refine the study area, in addition to Landsat scenes, which were the primary data source. This
chapter describes how the study area was built, as well as how the most effective index was
chosen to differentiate healthy trees from tree death. This chapter also describes the method used
to evaluate the accuracy of the final classification results.
Chapter 4 presents maps of the results, including an analysis of new areas of tree
mortality, and a comparison to other SOD map data. This chapter concludes with an analysis and
discussion of the accuracy of the mapping results.
Chapter 5 discusses the implications of the project results, offers suggestions for further
work, and reflects on success at responding to the project’s research questions.
9
CHAPTER TWO: RELATED WORK
When SOD was first discovered, scientists focused on identifying the origin and spread methods
of the disease in an effort to learn how to control it. Twenty years after the first reported sighting,
mapping the current extent of this still uncontrolled pathogen is our next feasible step in
understanding the disease. The hope is that effective techniques which used satellite imagery to
map declining tree health in other forest landscapes can be adapted to fill this gap in our
knowledge of SOD.
2.1 Biology of SOD
SOD spreads naturally and rapidly, and there is no effective means of control. Understanding the
biology of SOD helps explain the long-term effect on our forests, and why controlling it is so
difficult.
SOD earned its name because infected trees can change appearance from healthy to dead
in as little as three months, but SOD is not the only result of PR infection. PR may infect
susceptible plants in three possible ways(Hansen, Parke and Sutton 2005, Rizzo, Garbelotto and
Hansen 2005). Sudden Oak Death, shown in Figure 3, the most serious infection, causes bleeding
ulcers on the trunk and limbs which weaken the tree and lead to death(Parke et al. 2007, Collins
et al. 2009). In California, only three types of trees develop SOD: California black oak (Quercus
kelloggii), coast live oak (Quercus agrifolia) and tanoak (Lithocarpus densiflorus). Trees
infected with SOD do not produce spores or spread the disease(Mascheretti et al. 2008). PR
spores are produced by vegetation infected with the non-lethal Ramorum Blight, evidenced by
leaf discoloration(Mascheretti et al. 2008), as shown in Figure 4. California bay laurel
(Umbellularia californica) is the primary sporulator for PR in California(Davidson et al. 2005),
10
but as of 2012 more than 40 species, listed in Appendix A, are known to host the disease. Tanoak
trees(Lithocarpus densiflorus) are unusual because they have been observed with both types of
infection: they develop SOD in California and Ramorum Blight in Oregon(Grünwald et al.
2012).
PR spores are spread naturally and by human action. Wet, warm spring seasons provide
optimal conditions for PR spore creation(Davidson et al. 2005), and the spores are dispersed by
rain splash and wind-driven rain. Infected soil transported via nursery plants(Cushman and
Meentemeyer 2008) and via tires and shoes of recreational forest users(Davidson et al. 2005) cause long-distance dispersal of PR. Although stream monitoring has detected PR up to 20 km
downstream of a known infection site(Davidson et al. 2005, Sutton et al. 2009), research does
not indicate that infection patterns follow streams.
Figure 3 Coast live oak showing the bleeding ulcers typical of Sudden Oak Death
Photo: Steve Tjosvold, UC Cooperative Extension
11
Figure 4 California bay laurel (Umbellularia californica) infected with Ramorum blight
Photo: John Bienapfl, University of California Davis
Attempted methods for controlling PR include burning of host trees, chemical sprays,
antimicrobial applications, and creation of a disease-break area that contains no possible hosts
(Filipe et al. 2012, McGinnis 2008). There is no effective means to cure an infected tree, and
preventive methods need to be applied to individual trees in order to be most effective. Fire as
disease suppression is only effective if all PR-infected California bay laurel trees are completely
destroyed(Beh et al. 2012). No current control methods can avert large-scale infection.
Ramage, O’Hara and Forrestel (2011) believe that PR will remain permanently in
California forests because numerous species host the disease without succumbing to it. Tanoak
(Lithocarpus densiflorus), whose acorns are an important animal food source, may become
extinct because it has no genetic resistance and it can both transmit the disease and die from it
(Ramage, O’Hara and Forrestel 2011, Maloney et al. 2005, Ramage and O’Hara 2010).
12
PR poses a danger to the long-term existence of all SOD-susceptible trees. We cannot
predict what California forests will look like 50 years from now: although the first SOD-infected
trees died more than a decade ago, new vegetation has been slow to take over. Mapping disease
locations could expose a disease-resistant strain of trees, or demonstrate environmental variables
which impede SOD spread, and might lead scientists to discover new means of control.
2.2 Remote Sensing of Tree Health
Remote sensing data describes data collected at a long distance from the object or location being
viewed. For the purposes of this study, remote sensing data refers to aerial- or satellite–collected
data, including images. Collecting data from a distance makes it possible to quickly collect
information which encompasses a large land area while sacrificing some detail.
Remotely sensed images can be collected aerially or via satellite. High resolution aerial
photography provides fine detail for land change analysis, but it is expensive to collect, each
image covers a relatively small area, and it is difficult to obtain historical data meeting
specifications. Satellites’ high altitude means the imagery is generally of lower resolution than
aerial photography, but their orbit around the Earth allows consistent repeated collection over
multiple years, so it is a good choice for historical land change analysis.
Satellite data sources can be compared based on spatial, spectral or temporal resolution.
Spatial resolution describes the land area covered by a single pixel. The smaller the land area
covered by each pixel, the finer the spatial resolution. Spectral resolution expresses the number
of bands collected for each scene. Collecting more bands requires more instruments and data
storage on the satellite. Temporal resolution describes the time lapse between repeated
collections in the same area, and is a function of the land area covered by each image. No
13
existing satellites have high resolution in all three areas, so choosing a satellite data source
requires compromise.
The Landsat satellite missions, a joint program between the USGS and National
Aeronautics and Space Administration (NASA) have collected data covering all non-polar
regions of the world since 1972. This project primarily uses Landsat 5 TM data, which has a 30
m spatial resolution, 16-day repeat cycle, and seven bands of data. Landsat Climate Data Record
(CDR) reflectance images, corrected to remove atmospheric distortion, were used to facilitate
comparison between data sets collected under different weather conditions and at different times
of day. Each Landsat scene covers approximately 100 square miles and the data is available for
free download from the USGS. The frequent collection over the same areas makes it likely that a
researcher can obtain cloud-free images of a study area, and the low cost of data makes it
inexpensive for agencies to adapt this methodology to their needs.
Multi-spectral image analysis applies algorithms which highlight information contained
in separate spectral bands. Vegetation analysis relies primarily on the green, red, near infrared
and shortwave infrared bands to classify vegetation and measure vegetation health. Landsat’s
seven bands are:(1) blue;(2) green;(3) red;(4) near infrared;(5) shortwave infrared;(6) thermal; and (7) a second shortwave infrared band. The most common indices used to categorize
vegetation are shown in Table 1.
There are no recent studies of SOD contagion using satellite imagery, but a Mahon et al.
(2002) study attempted to map potential SOD infection on the California central coast in 2000.
By comparing image changes from 1996 to 2000 using the Tasseled Cap transformation applied
to Landsat imagery, with supplements of aerial imagery, field study, and other remotely sensed
and land classification data, they were able to identify large potential SOD concentrations in
14
Table 1 Multi-spectral indices used to detect vegetation health
Index Landsat Formula Notes
Short Wave Infrared/Near Infrared
(SWIR/NIR) Band5 / Band4 Detects moisture content, an
indication of health, water stress or
drought.
Normalized Burn Ratio (NBR)(Band4-Band7) /
Band4 + Band7) Identifies burned areas and quantifies
severity. Occasionally used to assess
vegetation health.
Normalized Difference Moisture
Index (NDMI)(also known as
Normalized Difference Infrared
Index(NDII)(Band4-Band5) /
Band4+Band5) Detects water content, a measure of
vegetation health.
Normalized Difference Vegetation
Index (NDVI)(Band4-Band3) /
Band4+Band3) Detects green plant canopies
indicating vegetation health and
density. Used to quantify
photosynthetic capacity.
Red Green Index (RGI) Band3 / Band2 As tree health declines, the red value
increases, causing this index to
increase. Useful for detecting dead or
dying trees.
Tasseled Cap brightness(TC Bright)(also known as Kauth-Thomas
transformation) 0.2043*Band1 +
0.4158*Band2 +
0.5524*Band3 +
0.5741*Band4 +
0.3124*Band5 +
0.2303*Band7
Measures image brightness. Used to
differentiate dark and light soils.
Tasseled cap greenness(TC Green)(also known as Kauth-Thomas
transformation) -0.1603*Band1 –
0.2819*Band2 –
0.4934*Band3 +
0.7940*Band4 –
0.0002*Band5 –
0.1446*Band7
Detects greenness, a measure of
vegetation density and
photosynthetically-active vegetation.
Tasseled cap wetness(TC Wet)(also known as Kauth-Thomas
transformation) 0.0315*Band1 +
0.2021*Band2 +
0.3102*Band3 +
0.1594*Band4 –
0.6806*Band5 –
0.6109*Band7
Detects wetness, indicating surface
moisture, vegetation density and dried
vegetation.
15
three coastal counties and small concentrations in three other counties, confirmed with aerial
data. At the time of publication, field verification was incomplete but preliminary results
indicated that the change detected was potential SOD. The researchers admitted that pixel-level
analysis likely overestimated mortality, but an aggregation would result in missed positives.
They concluded that change detection methods were able to successfully identify subtle changes
in canopy cover.
Many studies similar to this project have used Landsat scenes to assess insect infestation.
Studies of beetle-caused mortality used a variety of vegetation indices, including Tasseled Cap,
NDMI, SWIR/NIR and NBR to identify areas of infestation and evaluate severity. Their methods
give us clues to effective ways to map SOD.
One project in the Santa Fe National Forest(Vogelmann, Tolk and Zhu 2009), for
example, evaluated conifer tree health over an 18-year period using eight Landsat scenes, and
found an original way to validate results using the ADS data. Using the SWIR/NIR index, more
sensitive to conifer health than NDVI, researchers evaluated increases and decreases in the
SWIR/NIR index indicating forest health changes. When researchers compared their results to
the ADS insect defoliation maps, they found annual variability in the quality of the ADS data.
However, by combining multiple years of ADS data, disregarding areas which had only been
identified as damaged in a single year, areas which their study identified as experiencing
consistently decreasing tree health correlated with areas reported as damaged by the USDA over
multiple years. Their study concluded that this Landsat time series effectively captured decline in
forest health and that Landsat data is particularly well-suited to studies over large areas.
Another study in Colorado compared the accuracy of different vegetation indices and
band combinations applied to single- and multiple-date Landsat data to map bark-beetle caused
16
tree mortality(Meddens et al. 2013). Seven vegetation indices were calculated: RGI, NDVI,
NDMI, SWIR/NIR and the three Tasseled cap indices (see Table 1 for additional details). Land
cover classifications created from a 30 cm aerial image were aggregated to the Landsat pixel size
to classify the portions of each land cover within an equivalent Landsat pixel. This study
obtained 91 percent overall accuracy using single-date image classification with the Tasseled
Cap indices to map red-stage (dying) trees, and 89 percent accuracy with multi-date image
analysis and the SWIR/NIR index. Although the single-date method was more accurate with high
mortality, the researchers concluded that both methods resulted in high classification accuracy
using Landsat data.
Several other studies used a variety of indices with Landsat data to evaluate insect impact
on forest health. In an unusual choice, the NBR, primarily used by fire sciences, was used to
evaluate beetle infestation impact on ground fuels which might affect the severity of forest fires
(Meigs, Kennedy and Cohen 2011) in Oregon. They found a correlation between the presence of
coarse woody detritus and Landsat spectral change, consistent with Landsat’s sensitivity to
vegetation cover. This study successfully used NBR to map both short- and long-term spectral
change, and concluded that methods which only focus on short-term changes miss many of the
signs of insect infestation.
Similarly, a Canadian forest study(Goodwin et al. 2008) successfully used the NDMI
with multiple years of imagery to identify forest stands declining due to beetle infestation, with
71 to 86 percent accuracy. These researchers believed that analyzing imagery from more than
two years might be more accurate. These researchers found significant variation in NDMI values
for healthy forest, approximately 50 percent variation around the mean. As the infestation
worsened, the mean and the extremes decreased and mapping accuracy increased. The
17
researchers concluded that NDMI showed greater changes in beetle-infested areas than in the
healthy forest, but spectral confusion was a problem in areas with low levels of infestation.
These studies indicate that Landsat imagery can successfully detect changes in tree
health. Spectral confusion is a concern, and accuracy increases as tree mortality increases, but
both single- and multi-date imagery produce successful results under a variety of circumstances.
These beetle-damage studies are different from SOD studies in several important ways. The trees
in these examples were all conifers, but the three species killed by SOD are broadleaf, so the
same vegetation indices may not be the most effective. Another important difference is the way
the foliage changes. Beetle-induced death takes several years, and dying trees can be recognized
by red foliage indicating tree stress. Trees infected with SOD display bleeding lesions for several
years, but at the final stage the foliage changes from green to yellow to brown within a few
weeks(Alexander and Swain 2010), which may be difficult to capture on imagery.
Like these studies, the SOD study will be evaluating health of sub-pixel-sized features.
Several studies above observed that results were more accurate in areas with greater mortality, so
this study differentiates between pixels containing lower and higher levels of mortality. This
project also minimizes the mixed-pixel dilemma by masking out non-forested areas, decreasing
the incidence of pixels containing non-forest land cover types with conflicting spectral
responses. Several of these researchers were measuring tree health change over time, including
both decline and improvement. Trees infected with SOD have no chance of recovery, so this
study evaluates index values in a binary approach: pixels containing only healthy trees or
containing some dead trees. The index values chosen will determine how sensitive the formula is
at detecting areas with small occurrences of tree death relative to overall canopy.
18
These studies indicate that there are multiple effective methods and indices to effectively
measure tree health using Landsat imagery and additional data. Whether a study used a single
year or multiple years of remotely sensed imagery, they were all able to successfully identify
decreasing tree health. This study will draw on several techniques used in the studies above,
particularly an evaluation of the accuracy of different vegetation indices, and a comparison of
results to ADS data.
Remotely sensed data has been used to study SOD by providing inputs for disease risk
models, aerial imagery has been used to identify infected trees in localized areas, and aerial
monitoring has been used to identify generalized disease locations.
Remotely sensed data, such as temperature, slope, aspect, weather or vegetation is often
an input in the creation of risk maps. Risk maps can be useful for land managers to learn which
areas to monitor for signs of the disease, but inexact models can cause certain areas to be
unmonitored, or give a false sense of security to residents, who fail to take simple precautions to
prevent introduction of the disease. Many different remotely sensed variables can be weighted,
resulting in dramatically different models. Unfortunately, the accuracy of risk models can usually
only be determined many years after their creation.
A 2007 SOD risk model used remotely sensed data, including topographic, climatic data,
and raster data to create a worldwide view of disease potential. Kluza et al. (2007) predicted
wide distribution of SOD on the west coast of North America, from San Diego to Vancouver
Island and throughout the Sierra Nevada mountains, as shown in Figure 5. In addition, this model
showed large risk in the southeastern US, with additional risk areas in South America, southern
Africa, southern Europe, the UK, and Asia, including South Korea and Japan. This model puts
19
risk in many areas where PR infection has not yet been detected, but the accuracy of this model
cannot be evaluated at this time.
Two early SOD risk models for California used land cover data created from remotely
sensed raster data, which may have caused inaccuracies in their final result. A 2004 model
Figure 5 Risk for SOD distribution on the west coast. The value indicates the number of models
used to predicted risk in that area. The inset shows confirmed infections as of 2007.
Source: Kluza et al. (2007)
20
(Meentemeyer et al. 2004), which used Classification and Assessment with Landsat of Visible
Ecological Groupings (CALVEG) and California Gap Analysis Project(GAP) data, was found to
have underestimated the risk for many areas due to inexact vegetation mapping used as an input
to the model. A 2005 study(Guo, Kelly and Graham 2005) also used the GAP dataset as one of
their inputs, so may have similar problems.
High-resolution remotely sensed imagery can be useful for detailed land cover analysis. It
can minimize mixed pixels and increase detail, but its cost can prevent use by organizations with
small budgets, and the relatively small area covered by a single image makes it difficult to use
for analyzing disease extant over a large scale. Two SOD studies were conducted using high-
resolution aerial imagery to determine the severity of the SOD outbreak in localized areas. In the
Big Sur area, Meentemeyer et al. (2008) manually digitized dead trees from 0.33 m resolution
aerial imagery and calculated 20 percent mortality in infected areas. After field study, the
researchers determined that counting dead trees at the canopy level underestimated overall
mortality due to dead trees below the canopy and trees that had fallen over. In China Camp State
Park in Marin County, Liu, Kelly and Gong (2006) used 1 m resolution Airborne Data
Registration (ADAR) images to build a land cover classification model which would accurately
identify dead trees due to SOD, but they discovered that bare areas and dead trees displayed
similar spectral signatures, and that newly-leafed oaks have a low NIR value which can make
them appear dead. Although high-resolution imagery minimizes the problem of mixed pixels, the
researchers chose to smooth the results to minimize noise in the output. The moderate-resolution
Landsat data used in this study have larger scene areas than aerial images, to better facilitate
landscape analysis over an extended area, and a low cost for its use by cash-strapped
organizations, but requires adjustment for mixed pixels with sub-pixel-sized features.
21
Remote sensing aerial monitoring is used to map current disease locations through the
ADS. These surveys are collected by technicians in a moving aircraft who visually identify areas
of decreasing forest health and mark them on a hand-held device. The quality of the results
varies from year to year due to different operators collecting the data, and not all areas are
updated annually. In addition, the data is only collected for regions where forest health changes
are anticipated, so these surveys are unlikely to detect SOD outbreaks in new areas. However,
these are a useful input to a study like this for results comparison.
2.3 Mapping of SOD
Very few maps exist online which show the extent of SOD. Those that do exist are small-scale
and quickly outdated. One notable dynamic map which must be mentioned is the SODMAP by
OakMapper, a sample of which is shown in Figure 6(Kelly and Tuxen 2001, Kelly and Tuxen
2003, Kelly, Tuxen and Kearns 2004). This Google-Earth-based map displays locations of
confirmed SOD infections maintained by the California Oak Mortality Task Force (COMTF).
Data is updated annually and covers all of California. Unfortunately, the scale of the SOD
epidemic is many times greater than the SODMAP shows. Since SODMAP relies on human
observation and laboratory analysis, the majority of infections it shows are near roads or other
easily-accessible areas. The results of a successful project like this could be used to build an
online map showing the spread of SOD disease extent in the study area over the past 20 years,
and the methodology could be adapted to create a map of SOD infection for the entire state.
Although the locations would not be confirmed SOD infections, they would more realistically
depict the extent of the epidemic.
22
Figure 6 A sample image from SODMAP by Oakmapper, displaying PR-infected trees in the
Santa Cruz Mountains
Source: Kelly and Tuxen (2003) Remote sensing data and techniques have been used to build models of SOD disease risk,
identify individual dead trees and conduct aerial surveys, but satellite data has not been used to
map disease extent recently or identify progress over an extended period of time. This presents a
unique opportunity to expand our knowledge of SOD using techniques like those described
above. This study endeavored to find a methodology effective at identifying dead trees at the
canopy level using remotely sensed data. A successful methodology, adapted to other areas, will
make it possible to evaluate forest health on a large scale, quantify damage, discover SOD in
areas where it was not previously noticed, identify risk areas that remain untouched, and lead to
new discoveries.
23
CHAPTER THREE: DATA AND METHODOLOGY
This project’s primary goal was to identify an index and range that successfully differentiated
Landsat pixels containing healthy trees at the canopy level from those containing tree death, as a
proxy for mapping trees killed by SOD, and to use that to map SOD spread since its discovery.
This chapter describes the study area, the data and the methodology used for this research.
Satellite scenes are the primary data source, supplemented by aerial images and additional raster
and vector data which define the study area. The data section of this chapter discusses each of
these sources in detail, as well as the limitations of the data, an understanding of which is
necessary to avoid drawing unreasonable inferences from the results. The methodology followed
is then described in detail, including data preparation, refining the study area, and choosing the
most effective index for mapping SOD. A thorough understanding of the data and methodology
will assist other researchers in interpreting these results, understanding their limitations, and
adapting this process to map SOD in other regions.
3.1 Description of Study Area
This project’s study area is the Santa Cruz Mountains in Central California(37° 08’ N, 122° 8’
W), a coastal mountain range on the San Francisco Peninsula, shown in Figure 7, with the dense
urban areas of San Francisco and San Jose to the north, east and south. The climate is
Mediterranean, with dry summers and wet winters, but fog often occurs on the west side of the
range in summer. Elevation ranges from sea level to 3,786 feet (1,154 m), with many moderately
steep valleys. Hilltops often are covered with grass and drought-resistant bushes, while valleys
and lower elevations include conifers and broadleaf forests. Half of the area is state and county
parkland and open space preserves, which enjoy heavy recreational use. The remainder is
24
privately owned, with scattered residences and isolated agricultural areas. The area of interest is
1,834 km
2
, of which approximately 706 km
2
is forested. Many areas are steep and densely
vegetated, with few roads or trails.
Figure 7 The area of interest in the Santa Cruz Mountains of central California
25
3.2 Data
This study utilizes raster and vector data, as shown in Table 2, to map the locations and spread of
SOD from 1994 to 2011. The primary data source is Landsat CDR, supplemented by other
sources which help refine the study area. NLCD raster data classifies land cover into forested and
non-forested areas. FRAP polygons identify locations of recent wildfires, and ADS polygons
map locations of tree damage due to multiple causes. These three data sets were used to create
the study area, limiting it to forested areas without other known causes of change. In addition, for
the purposes of sampling Landsat raster values, Land Cover Points and Results Validation Points
were created and categorized based on May 2011 Google Earth Historical Imagery. All data was
referenced to the North American Datum 1983 HARN California Teale Albers projected
coordinate system (m), except for data displayed in Google Earth, which used the World
Geographic System 1984. This section describes each data set used in this project and errors that
the data may contain.
Table 2 Data used for this analysis
Description Type Data sets used Creator
Landsat CDR Reflectance images Raster 1992-2011 USGS
National Land Cover Database (NLCD) polygons Raster 1992, 2011 USGS
Fire and Resource Assessment (FRAP) polygons Vector 1985-2012 CalFire
Aerial Detection Survey (ADS) polygons Vector 2005-2013 USDA
Google Earth Historical Imagery Raster May 1, 2011 Google
Land Cover Points Vector May 1, 2011 Trinka Gillis
Results Validation Points Vector May 1, 2011 Trinka Gillis
26
3.2.1 Landsat Satellite Data
Landsat CDR images were the primary data source for this project. Landsat 5 Thematic Mapper
(TM) satellite images collect seven bands of data, detailed in Table 3, covering non-polar Earth
with repeat collection every 16 days(DOI-USGS 1989-2011). Bands 1-5 and 7 were used for
this study. CDR images are created from Landsat Level 1 images, which are systematically
Table 3 Landsat 5 TM bands
Band Wavelength (µm) Resolution (m) Common Uses
Band 1 - Blue 0.45-0.52 30 Measures visible blue light.
Useful for mapping sediment,
coastal habitats, water depth, and
distinguishing
soil/rock/vegetation.
Band 2 – Green 0.52-0.60 30 Measures visible green light.
Useful for identifying vegetation
and measuring plant health
Band 3 – Red 0.63-0.69 30 Measures visible red light.
Vegetation absorbs red light, so it
produces low values in this band.
Useful to distinguish vegetation,
soil, and vegetation health.
Band 4 – Near
Infrared
0.77-0.90 30 Water absorbs most light in this
wavelength, producing low
values, while soil and vegetation
produce high values. Useful for
distinguishing water, vegetation
varieties, soil/crop/water contrasts
Band 5 - Short-
wave Infrared
1.55-1.75 30 Sensitive to moisture. Useful to
measure soil and vegetation
moisture content, clouds and
snow.
Band 6 – Thermal
Infrared
10.40-12.50 120 Measures surface temperature.
Useful for geology, measures
plant heat stress, locates clouds.
Band 7 – Short-
wave Infrared
2.09-2.35 30 Similar to band 5, measures
moisture. Distinguishes
water/soil/rock.
Source: Quinn (2001), DOI-USGS (1989-2011), Geospatial Innovation Facility UC Berkeley and
Center for Biodiversity and Conservation (2003)
27
corrected for geometric and radiometric accuracy using ground control points, with a DEM used
for topographic accuracy(DOI-USGS 2014). CDR images are further corrected to surface
reflectance values using Moderate Resolution Imaging Spectroradiometer (MODIS) formulas to
remove atmospheric distortion caused by water vapor, ozone, geopotential height, aerosol optical
thickness and elevation(DOI-USGS 2013). The CDR images used in this project are all
classified as LIT, indicating standard terrain correction. Reflectance images facilitate comparison
between data sets collected at different times of day and under different weather conditions. This
study used WRS2 path 44 row 34 images, which contain the entire study area, and favored April
and May images to avoid spectral confusion caused by early-senescing California buckeye and
newly-leafed oaks, both of which can be spectrally confused with dead trees. Images from 1992-
2011 were used, as shown in Table 4. The number of ground control points used for the
geometric correction, and the resulting Root Mean Square Error (RMSE) are also summarized in
Table 4. All images had maximum 10 percent cloud cover.
Unfortunately, 2011 was the last year that acceptable springtime Landsat 5 images are
available for this area. If this methodology were to be applied to later data, Landsat 7 and
Landsat 8 scenes would need to be used. Landsat 7 data is available from 1999 to the present in
CDR format, but scenes collected after May 2003 have a striping issue due to a failed Scan Line
Corrector (SLC). Landsat 8 was launched in 2013, but is not yet available in CDR format. This
project therefore relied exclusively on Landsat 5 data.
The 2011 Landsat image was used as the basis for all this project’s calculations. To
facilitate comparison and analysis, eight indices were calculated using ArcMap Raster
Calculator. Values for five of the indices – NDVI, SWIR/NIR, NDMI, RGI and NBR – were
28
Table 4 Landsat 5 TM scenes used in project(path 44 row 34) Imagery Date Ground Control Points Geometric RMSE (m) April 6, 1992 143 4.387
June 12, 1993 218 3.525
March 11, 1994 153 4.583
March 30, 1995 154 3.798
May 3, 1996 183 3.687
April 4, 1997 194 3.446
June 26, 1998 221 3.151
June 29, 1999 242 2.793
April 28, 2000 184 3.575
May 1, 2001 203 3.501
June 5, 2002 223 3.235
April 5, 2003 N/A N/A
April 23, 2004 197 3.511
April 10, 2005 169 3.919
June 16, 2006 216 3.443
April 16, 2007 190 3.776
June 5, 2008 215 3.193
May 7, 2009 182 3.825
April 24, 2010 155 3.766
April 27, 2011 151 3.766
multiplied by 1,000 to preserve data precision. Values for Tasseled Cap indices were calculated
and used in their original form.
The 2011 Landsat image was used as the basis for all this project’s calculations. To
facilitate comparison and analysis, eight indices were calculated using ArcMap Raster
Calculator. Values for five of the indices – NDVI, SWIR/NIR, NDMI, RGI and NBR – were
multiplied by 1,000 to preserve data precision. Values for Tasseled Cap indices were calculated
and used in their original form.
Although Landsat CDR images are geometrically and radiometrically corrected, pixels
from different years do not align and pixel values vary for the same location over different years.
Measures of geometric distortion for each image are indicated by the RMSE values in Table 4.
Pixels could have been aligned, and values could have been normalized for this project, but to do
29
so would introduce additional error. To facilitate location comparison of pixels, all were
converted to centroid points and were evaluated using proximity functions. Other common errors
with Landsat data include missing pixels and saturated bands.
3.2.2 National Land Cover Database Classified Raster Data
This project used 1992 and 2011 National Land Cover Database (NLCD) classified raster data to
construct a study area. The NLCD is produced by the USGS to characterize land cover and
monitor changes throughout the US, and is now updated every five years. Using Landsat data as
its primary source, the NLCD categorizes land cover into 16 classes listed in Table 5, using a
decision tree model with training points, and an algorithm which merges cover types to preserve
land cover logic.
Table 5 NLCD land cover classes
Land Cover Classification
Open Water
Perennial Ice/Snow
Developed, Open Space
Developed, Low Intensity
Developed, Medium Intensity
Developed, High Intensity
Barren Land (Rock/Sand/Clay) Deciduous Forest
Evergreen Forest
Mixed Forest
Shrub/Scrub
Grassland/Herbaceous
Pasture/Hay
Cultivated Crops
Woody Wetlands
Emergent Herbaceous Wetlands
Note: Four additional classes were used in Alaska only
Source: Jin et al. (2013)
30
There are several sources of inaccuracy in the NLCD. Although the classification is
conducted at the pixel level, classification of mixed pixels is a problem, and smoothing the result
to create more consistent land cover decreases accuracy. Errors can be observed by laying the
NLCD data over a high resolution aerial photo. The creators have allowed for ambiguity,
however, in the category descriptions. For example, forest categories are described as containing
a minimum of 20 percent total vegetation of trees greater than 5 m tall, with more than 75
percent of species matching the category description. This allows up to 80 percent of land cover
to be small trees and up to 25 percent to be types not described by the category. Because of the
generalized categorization and observed inaccuracies, this data set was not sufficient to define
the study area without supplemental resources.
3.2.3 Fire and Resource Assessment Program Vector Data
This project also used the Fire and Resource Assessment Program (FRAP) polygons to identify
areas burned in wildfires and prescribed fires in 1985 or later, and remove them from the study
area. This step was included to prevent the changed spectral response of dead trees in burned
areas from incorrectly being mapped as SOD. The California Department of Forestry and Fire
Protection (CalFire) maintains a comprehensive FRAP database, last updated in 2012, which
includes fire data gathered by CalFire, the USDA Forest Service Region 5, the Bureau of Land
Management (BLM), the National Park Service (NPS), certain counties, and other agencies.
Although the FRAP database attempts to track all wildfires and prescribed fires 10 acres
or larger on all public and private lands in California, inaccuracies occur due to the differing
standards of the reporting agencies. The year that an agency began collecting data, the minimum
31
size of fire that they report, the data creation method and accuracy of the polygons vary.
Observed errors include perimeters which truncate at administrative boundaries and duplicate
fires reported by multiple agencies.
3.2.4 USDA Aerial Detection Survey Vector Data
The ADS polygons indicating tree damage due to causes other than SOD were used to refine the
study area. ADS polygons are created by the USDA through summer aerial surveys which detect
and map tree mortality and damage. Surveyors in an aircraft flying 1,000 feet above ground level
mark areas of tree damage and cause on a hand-held device. The resulting polygons are used to
report on status of known threats, and to estimate acreage and number of trees affected(USDA-
FS Forest Health Monitoring Program 1978-2013, Heath et al. 2012, USDA-FS Pacific
Southwest Region 2014). The USFS decides annually which regions or diseases to focus on, but
SOD is often a main concern. ADS data is available for 1978-2013, but collection was erratic
and polygons were highly generalized prior to 2005, so this project only uses data from 2005
onwards.
The ADS polygons and attributes associated with them are fuzzy, subject to the
experience, training and subjective opinion of the surveyor, as well as the sun angle and viewing
window. Polygons from different years often overlap. Researchers mapping tree health in New
Mexico(Vogelmann, Tolk and Zhu 2009) found that areas mapped as damaged by the ADS for
multiple years correlated to damage severity, but areas mapped by the ADS as damaged in only a
single year were unreliable, possibly due to the use of inexperienced mapping technicians.
32
3.2.5 Google Earth Historical Aerial Imagery
This project also used a Google Historical Imagery dataset from May 1, 2011, with sub-meter
natural color imagery, to verify the land cover type and tree health in spring 2011. Google
maintains a database of historical imagery within Google Earth, displayed with a terrain model.
The type of imagery, the quality, and the area covered varies. Images with no source identified in
Google Earth are Google copyright and are often high quality. Images cannot be exported, but
ArcMap shape files can be converted to the WGS84 projection and displayed within Google
Earth. The capture date of this Google high-quality aerial imagery is ideal for visualizing the
canopy land cover at the time of the April 27, 2011 Landsat scene used as the base data set for
this project.
Misalignments of up to 2 m have been observed when comparing Google imagery from
different dates, but these displacements are not large enough to affect classification for this
project.
Although National Agricultural Imagery Program (NAIP) aerial images created by the
USDA Farm Service Agency could have been used, preliminary work revealed that those images
were often overexposed or captured with a low sun angle, both of which caused color distortion
that caused healthy trees to appear dead.
3.2.6 Land Cover Points
To identify spectral ranges for each land cover type in the study area, Land Cover Points were
created, randomly and manually, at least 45 m apart. The points were projected to WGS84 to
display in Google Earth with May 2011 imagery. The Google Earth imagery was used to
categorize each point into one of the classes shown in Table 6, with the location of the 2011
33
Landsat pixel relative to the point taken into consideration when assigning a classification.
Locations with mixed pixels other than those examples mentioned below were discarded. The
distinction between minor and serious tree death(SD) was necessary to obtain values which
clearly distinguished healthy areas from dead trees in later processing. Although no pixels
contained more than 40 percent dead trees, pixels which contained less than 10 percent dead
trees produced a spectral signature similar enough to healthy forest that including them in a
single land cover type with a larger canopy area of dead trees blurred the distinction between the
classes. This project’s goal was to identify and map only the serious tree death.
Although care was taken to consistently and correctly classify the land cover types, the
classification process was, inevitably, subject to some operator error.
3.2.7 Results Validation Points
To evaluate the efficacy of the Landsat image classification, 300 results validation points were
randomly generated throughout the refined study area. These points were displayed over the May
Table 6 Land Cover Points classification categories
Code Description Points
10 Agricultural land 4
20 Barren land 31
30 Herbaceous 44
40 Healthy forest 68*
51 Minor death (forest with <10% dead trees as percentage of pixel area) 26
52 Serious death (forest with >10% dead trees as percentage of pixel area) 33
60 Shrubbery 66
70 Buildings (mixed pixels included) 24
80 Road (mixed pixels included) 20
90 Water 16
Total 301
* A random subset of 33 points was used to compare to an equal number of points classified as
serious tree death.
34
2011 Google Earth image, as described above, and assigned a land cover code shown in Table 7,
taking into account the relative location of the 2011 Landsat pixel. Approximately 10 percent of
the randomly generated points were classified as SD and 68 percent were classified as healthy
tree land cover(HLH).
Although care was taken to correctly classify the land cover types, the classification
process was subject to operator judgment on the approximate location of the pixel and the
percent of pixel covered by dead trees.
Table 7 Results Validation Points within study area
Class Code Number of Points
Healthy trees 0 204
Minor death (< 10% of pixel area) 1 37
Serious death (>10% of pixel area) 2 31
Mixed land cover classes 9 28
Total 300
3.3 Methodology
In addition to data sourcing and preparation, work on this project consisted of refining the study
area to remove non-forested areas from consideration, identifying the index most effective at
distinguishing dead tree areas from healthy forest, validating results, and extrapolating the
chosen index and range to the entire study area. An outline of the workflow that was deployed is
shown in Figure 8. To refine the study area, a mask was created from three types of data –
recently burnt areas, areas identified as having tree diseases other than SOD, land cover
identified as non-forest – and supplemented by raster data identifying non-forest land cover. An
analysis of Landsat band ratios was used to determine the most effective index for identifying
35
Figure 8 The methodology for data extraction and analysis that was created and used for this
project
dead trees, and this result was applied to multiple years of imagery to map the spread of SOD
over time.
3.3.1 Refining the Study Area
This project’s Area of Interest was the Santa Cruz Mountains. Preliminary results demonstrated
that shrub land cover shared spectral ranges with areas of dead trees. To minimize false
identification of dead trees, this project needed to develop an accurate land cover mask that
would limit the study area to only forested land.
The basis for the mask was the 1992 and 2011 NLCD classified raster data, from which
the non-forest categories were converted to polygons. The combination of early and late NLCD
data sets was used to more effectively mask out land use changes that occurred during the period
of the study. These layers failed to remove all non-forested areas from the study area.
36
Large-area land cover classifications, such as NLCD, seek to classify all land cover in the
US into a small number of classes, and seek to minimize noise. Because of this, small areas of
one type of land cover, pixels on the edge of other land covers, and mixed pixels are often
merged with other classes, all of which created problems for this study. Observations of the
NLCD data overlaid over an aerial image, a sample of which is shown in Figure 9, reveals that
many non-forested areas were incorrectly classified as forest. Edges of land cover types are
generalized and often contain vegetation which could cause inaccurate results. As a result, a
mask created from NLCD non-forest categories does not remove all non-forested areas from the
study area. This project used the NLCD as the primary land cover classification, but it needed to
be supplemented by other data to remove shrub from the study area and increase confidence in
the final results. This was accomplished using Landsat raster data.
To remove as much shrubbery from the study area as possible, it was necessary to
identify the index applied to Landsat data which best differentiated healthy trees and shrubs in
this environment. Using the Land Cover Points described above, the Extract Values to Points and
Zonal Statistics as Table tools were used to extract raster values for each index based on the
Landsat 2011 scene. The mean and standard deviation values of each land cover class produced
significant overlap. Instead, the histograms representing healthy trees and shrub land cover types
were visually compared, as in Figures 10 and 11, to determine which ranges captured the most
values. Histograms for NBR, SWIR/NIR and TC Wet, displayed in Figures 10-15, showed the
best separation between the two classes. An analysis of possible differentiation ranges, shown in
Table 8, showed that NBR with values greater than 640 was the most effective range to separate
healthy forest from shrub. When applied to the Land Cover Points set, this value removed 92
percent of the shrub points and only 4 percent of the healthy tree points.
37
Figure 9 The NLCD classification did not adequately remove non-forested areas from the study
area. Areas outlined in white were identified as non-forest classes using the NLCD 1992 and
2011 layers. This sample shows many non-forested areas were not correctly classified.
38
Figure 10 The NBR Land Cover Points histogram for shrubs
Figure 11 The NBR Land Cover Points histogram for healthy trees
39
Figure 12 The SWIR/NIR Land Cover Points histogram for shrubs
Figure 13 The SWIR/NIR Land Cover Points histogram for healthy trees
40
Figure 14 The NBR Land Cover Points histogram for shrubs
Figure 15 The NBR Land Cover Points histogram for healthy trees
41
Table 8 Indices and value ranges tested to determine which one best masked shrub without
masking healthy trees. The ranges indicate the values used to map healthy trees.
NBR =>640 SWIR/NIR
=<510
SWIR/NIR
=<460
TC WET =>-450 TC WET =>-560
Shrub masked 92% 85% 94% 97% 82%
HLH masked 4% 1% 7% 12% 3%
Using the ranges identified, the 1992 and 1993 CDR Landsat images, which had already
been converted to NBR, were reclassified to identify all pixels with values lower than 640, and
this area was converted to new polygons to be added to the mask. An example of this is shown in
Figure 16. Although land cover categories representing dead trees fall between the ranges
indicating healthy trees and shrub, and may be partially masked out by this method, these values
were applied to Landsat images representing years before SOD is believed to have taken hold. If
SOD was present in the forest before its discovery in 1994, it was in insignificant amounts. The
layers created from this process were added to the mask to remove shrubbery and refine the
study area.
Fire data and areas of other tree damage were also part of the mask. To remove recently
burned areas from the study area, likely to contain dead trees which are not due to SOD, FRAP
polygons were used. Two files were created: one containing all wild fires in the area which
occurred in 1985 or later, and one containing all prescribed fires in the area, none of which
occurred prior to 1985. To remove other causes of tree death from the study area, ADS 2005 to
2013 polygons with tree threat attributed to causes other than SOD were added.
The final study area was defined as the AOI less the areas masked out: non-forest NLCD
data, the 1992 and 1993 Landsat shrub layers, FRAP fire polygons, and ADS disease polygons.
By carefully shaping the study area, it was hoped that the final results would be more accurate.
42
Figure 16 The NLCD non-forest layers, shown in solid white, were not sufficient to limit the
mask to only forested areas. Landsat 1992 and 1993 data, used to supplement the mask, removed
all white-outlined areas from the study area. Figure 9 shows the NLCD non-forest layers without
the supplemental mask.
43
3.3.2 Identifying the Most Effective Index and Mapping 2011 Affected Areas
The premise of this project was that one could differentiate healthy tree canopy from areas of SD
using Landsat data with an appropriate analysis index. Once the final study area was determined,
it was necessary to select the formula and index to map tree death. The final step was to apply
the chosen formula to the study area to map SOD infection.
The process of choosing an index to differentiate SD from HLH was similar to that used
for shrubs. The mean and standard deviations showed too much overlap, so histograms were
created to determine the index and ranges which captured the most points from each class while
minimizing points from other classes. This project sought a formula that would accurately
identify at least 60 percent of points in each class. Test value ranges were based on a comparison
of histograms, shown in Appendix B, graphing spectral signatures of the full set of 33 SD points
with a randomly-chosen subset of 33 HLH points. The SWIR/NIR,(Figures B1 and B2), NBR
(Figures B3 and B4), NDMI(Figures B5 and B6), NDVI (Figures B7 and B8) and TC Wet
(Figures B15 and B16) indices showed the best possibilities for differentiating these two classes
with a minimum of overlap. Test ranges were compared to index raster values to identify the
index and range which captured the maximum number of SD points and a minimum number of
healthy points. Over 50 ranges, indices and combinations were tested and the three best results
are shown in Table 9. Composite formulas which used two or more indices based on the OR
function (if either index marks this point as SD, mark it as SD) increased false positives when
compared to formulas using a single index, and composite formulas based on AND (mark this
point as SD only if all other indices mark it as SD) increased false negatives. Formulas based on
narrower ranges, even when combined with other formulas, created more false negatives because
44
the narrow ranges missed the same outliers. The best formulas were found by looking closely at
the histograms and testing many ranges and indices against the index values.
Indices SWIR/NIR, NBR and a composite of the two had test ranges with errors less than
20 percent(Table 9), but the NBR index and range showed the most accurate SD classification
with a low level of false positives, correctly categorizing 88% of points in both classes. These
three indices and ranges were then compared to the Results Validation Points to measure
accuracy and confirm that the same range was still the best choice when applied to the entire
2011 data set.
Table 9 Three indices and value ranges compared to determine which one best differentiated SD
from HLH
SWIR/NIR 400-570 NBR 575-725 OR Combination
% SD Correct 85% 88% 91%
% HLH Correct 85% 88% 85%
SD Omission 15% 12% 9%
HLH Omission 15% 12% 15%
SD Map Accuracy 74% 78% 79%
HLH Map Accuracy 74% 78% 78%
3.3.3 Accuracy Assessment
The mapping results were assessed for accuracy by applying the formula to the Results
Validation Points, and the most accurate index and range was applied to the time series to map
SD change over time.
The Results Validation Points, described above, were used to extract the value of the
containing pixel from the 2011 Landsat image with each index calculated. Tables 10, 11 and 12
show the error matrices for these three formulas upon application to the Results Validation
Points. Although the NBR formula looked slightly more accurate when applied to the Land
45
Cover Points, it shows more false negatives and false positives when compared to the SWIR/NIR
formula. The combination formula has more false positives than either of the other formulas. The
SWIR/NIR formula is the best of the three, with 69% tree health mapping accuracy and 24% SD
mapping accuracy. Based on these matrices, this project used the SWIR/NIR index with a range
of 400-570 to map SD in the study area.
The SWIR/NIR index and range described above was used to classify the study area
using the Landsat 2011 image. To facilitate comparison of non-aligned raster pixels, all pixels in
the range were converted to centroid points, and all points that fell outside of the mask were
considered to represent SD. The resulting 2011 SD layer, shown later in Figure 17(as part of
Chapter 4), is the map of SOD extent in 2011 produced using the methodology that was created
and deployed for this project. The SWIR/NIR index was calculated for the 1994-2010 Landsat
Table 10 Error matrix for SWIR/NIR index applied to Results Validation Points with range of
400-570 classified as SD
Observed Data
Classification Tree Health Serious Tree
Death
Row Total User
Accuracy
Commission
Error
Tree Health 148 10 158 94% 6%
Serious Tree
Death 56 21 77 27% 73%
Column Total 204 31 235
Producer’s
Accuracy 73% 68%
Omission
Error 27% 32%
Map Accuracy – Tree Health 69%
Map Accuracy – Serious Tree Death 24%
46
Table 11 Error matrix for NBR index applied to Results Validation Points with range of 575-725
classified as SD
Observed Data
Classification Tree Health Serious Tree
Death
Row Total User
Accuracy
Commission
Error
Tree Health 143 12 155 92% 8%
Serious Tree
Death 61 19 80 24% 76%
Column Total 204 31 235
Producer’s
Accuracy 70% 61%
Omission
Error 30% 39%
Map Accuracy – Tree Health 66%
Map Accuracy – Serious Tree Death 21%
Table 12 Error matrix for OR Combination of NBR 575-725 and SWIR/NIR 400-570 indices
applied to Results Validation Points
Observed Data
Classification Tree Health Serious Tree
Death
Row Total User
Accuracy
Commission
Error
Tree Health 135 7 142 95% 5%
Serious Tree
Death 69 24 93 26% 74%
Column Total 204 31 235
Producer’s
Accuracy 66% 77%
Omission
Error 34% 23%
Map Accuracy – Tree Health 64%
Map Accuracy – Serious Tree Death 24%
CDR scenes, and the same 400-570 value range was applied to map the extent of SOD for each
year. The final result was 18 maps, shown in Appendix C, one for each year 1994-2011.
47
3.4 Summary
This chapter presented the data and methodology used in this study, so other researchers can
understand the preparation, processes followed, and methods of obtaining results. Crucial to the
explanation of data sources was the discussion of sources of data error. No GIS data set is
without errors, and understanding the limitations of each data set is important for interpretation
of results. Construction of an accurate study area was complex, but was critical to clarifying the
accuracy of the results and the accompanying interpretation of their significance for the problem
at hand.
The next chapter will present the results of this process, including maps, an analysis of
accuracy, and a discussion of trends and patterns.
48
CHAPTER FOUR: RESULTS
Chapter three described the methodology followed in this project, which created maps of
canopy-level serious tree death for the Santa Cruz Mountains from 1994-2011. The project
produced 18 maps, one for each year of the study, showing areas identified as canopy level tree
death based on the SWIR/NIR index created from Landsat CDR scenes. The complete sequence
of maps is provided in Appendix C. This chapter discusses the mapping highlights and evaluates
accuracy of the final product.
4.1 Mapping Results
This project created maps for the first 18 years that SOD was known in the study area(Appendix
C). These maps were used to identify areas of tree death as a means to discover areas infected
with SOD, approximate the extent of contagion, and to compare areas of infection identified by
the project results with known locations of SOD infection.
4.1.1 Map of Serious Tree Death in 2011
The 2011 map of serious tree death, in Figure 17, shows tree death is present throughout most of
the study area, and especially dense on the more populated east side of the study area. Three
areas – Areas A, B and C – contain notably fewer dead trees than other areas. These areas will be
discussed in more detail below.
49
Figure 17 The 2011 SD extent in the study area. Areas with lower rates of tree death – Areas A,
B and C – are marked in green.
50
This methodology classified 27 percent of pixels in the study area as containing dead
trees in 2011, with the highest density on the east side of the study area. These slopes are
generally east-facing, which may influence moisture or vegetation which makes the ecosystem
more susceptible to SOD. These areas are also more densely populated than other parts of the
study area. Consequently, they may be at increased risk of SOD due to more accessible, therefore
heavier, recreational use, and proximity to homes in the interface zone which introduced the
infection via nursery plants. A 27 percent infection rate is much higher than expected and may be
a sign of inaccurate results.
Three areas stand out in this map for their relative lack of dead trees detected by this
methodology. Area A, shown in Figure 17, straddles the ridge between Redwood City and the
Pacific coast. This west-facing slope is parkland and preserved land, and this east-facing slope
abutting Crystal Springs Reservoir is owned by the San Francisco Public Utilities Commission
and closed to the public. Area B is on the coastal side of the mountain range, and is also
parkland. Area C is Nisene Marks State Park, a park that is difficult to access except by a few
trails which allow bicycles. With the exception of Highway 35 along the ridge in Area A, these
areas are steep, largely inaccessible, and are among the most densely vegetated areas on the
peninsula. Neither SODMAP nor ADS show any known SOD infestations in these three areas.
These areas may have escaped SOD infestation due to forest monitoring and rapid removal of
infected trees, because inaccessibility has prevented human transmission, or because the steep
canyons have provided an inhospitable environment for the disease. It is also possible that these
areas contain taller older-growth trees which obscure the view of dead trees lower in the canopy.
51
If this map is accurate, then tree death is widespread through the study area, and only a
few areas have managed to escape. This would indicate that SOD’s effects on the forest were
extensive and rapid, with devastating consequences within twenty years of first detection.
4.1.2 Serious Tree Death Change From Previous Years
Although the 2011 map of serious tree death shows tree mortality throughout the study area, the
result is not significantly different from 2010 or even 1994(Appendix C). In every year, this
methodology shows infestation throughout the study area, with no areas appearing to show more
rapid spread of the disease than others. Overall, the number of pixels classified as SD varies from
11 percent in 1999 to 51 percent of the study area in 2009, detailed in Table 13.
The methodology classified 165,000 pixels as SD in 2010, and 208,000 pixels in 2011, a
26 percent increase. To facilitate comparison of multiple years of SD classification data, pixels
were converted to centroid points. A comparison of new SD from 2010 to 2011, shown in Figure
18, shows 82,000 SD points in 2011 that were at least 31 m away from the nearest 2010 SD
point. An inspection of these changes shows 2011 damage in the same areas as the 2010 damage,
but covering a slightly greater areal extent.
A point density map of the new SD points, reproduced in Figure 19, shows the strongest
increase in the southern portions of the study area. Figure 20 provides detail on one particular
area which showed a significant number of pixels classified as new SD. Due to edge effects from
the complex study area, larger continuous areas may show disproportionally strong results with
the point density function.
52
Figure 18 2010 SD and new 2011 SD points(at least 31 meters away from a 2011 point).
53
Figure 19 A point density map showing the areas with the most new SD in 2011 compared to
2010. Detail can be seen in Figure 20.
54
Figure 20 Detail of new SD points in 2011
55
Table 13 Number of pixels within study area identified as containing serious levels of tree
mortality
Year Day SD
Pixels
Percent of
Total Area*
Increase from
Previous Year
New SOD Pixels
Since Previous Year
Percent of
Total Area
1994 070 176,057 23%
1995 089 265,448 34% 51% 75,552 10%
1996 124 120,302 15% -55% 18,961 2%
1997 094 143,181 18% 19% 54,572 7%
1998 177 120,702 15% -16% 41,773 5%
1999 180 89,280 11% -26% 23,322 3%
2000 119 175,114 22% 96% 96,287 12%
2001 121 130,225 17% -26% 36,047 5%
2002 156 194,596 25% 49% 69,067 9%
2003 095 230,463 29% 18% 89,315 11%
2004 114 204,170 26% -11% 75,741 10%
2005 100 235,000 30% 15% 47,920 6%
2006 167 142,706 18% -39% 21,616 3%
2007 106 200,631 26% 41% 69,928 9%
2008 157 159,868 20% -20% 31,210 4%
2009 127 396,994 51% 148% 192,582 25%
2010 114 165,210 21% -58% 7,537 1%
2011 117 208,452 27% 26% 82,067 11%
* Based on average of total pixels within study area for 1994, 1995, 2010 and 2011.
This project expected the 1994 SD map to show little tree death but, as can be seen in
Appendix C Figure C1, it instead shows large areas of dead trees throughout the study area.
Although SOD was not discovered until 1994, it may have been present in the forest before then,
limited to a few small areas. In that case, the widespread tree mortality depicted would be due to
causes other than SOD, but this is unlikely because anything causing tree death on this scale
would have been noticed. More likely, it is an indication of unreliable results.
The widespread tree death shown in the map of each year’s results was unexpected and
brings the accuracy of the results into doubt. This issue is explored with some tabular data
covering the period 1994-2011 in the next section.
56
4.1.3 Area of SOD infestation
One method to evaluate the change in SOD infestation over time is to look at the number of
pixels classified as SD for each year. This value would be expected to be near 0 in 1994 and
increase gradually, with some fluctuation for seasonality. Data captured early or later in the
spring could show different total calculations due to different seasonal spectral responses. Values
could also decrease if dead trees fall over and cease to be mapped.
Unfortunately, the pixel count shown in Table 13 does not conform to the logical
expectation for a spreading infection. Initial pixel counts are much higher than expected, and
show large annual variation. It may be notable that the lowest pixel count, 89,280, was for the
image captured latest in the season, on day 180 of 1999, and the second highest pixel count,
265,448, was for an image captured early in the season.
This unexpected fluctuation in pixels classified as SD may be an indication that seasonal
variation plays a much stronger role in the spectral signature than was anticipated in this project,
and annual differences in rainfall and temperature may further confound the interpretation of the
results. The results summarized in Table 13 are therefore a strong indicator that the results
produced by this methodology are not reliable.
4.1.4 2011 Comparison to ADS and SODMAP
ADS and SODMAP are the most thorough SOD datasets publicly available. This study
found SD throughout the study area, but current ADS and SODMAP data show SOD infection
primarily in the central-east portion of the study area. If the methodology described in this
project were reliable, then a comparison of this project’s results with these datasets could
highlight new areas of dead trees which might be SOD infection.
57
Figure 21 shows 2011 SD overlaid with the ADS polygons classified as SOD, and Figure
22 shows 2011 SD overlaid with the SODMAP confirmed SOD-positive infections. Both of
these datasets show SOD primarily near the ridgeline in the central part of the study area.
SODMAP collects data by field study, so most of their samples are near roads. ADS identifies
areas of tree death by flying over them. Without field study, ADS is forced to identify a cause of
tree death using other known information. If ADS is relying on SODMAP to tell them where
SOD is present, then SOD tree death in other areas may be mistakenly assigned to other causes
or may be undetected. A successful methodology to detect dead trees using satellite imagery
could identify areas of concern that are not easily accessible by SODMAP field testers or that
ADS hasn’t prioritized for fly over.
The 2011 map of serious tree death produced by this project reveals one region that
should be looked at more closely. In the southeast portion of the study area, between Los Gatos
and Morgan Hill, is a large area of east-facing private land which was mapped by this
methodology as containing large amounts of tree death, but has not been identified by either
ADS or SODMAP as a known area of SOD infection. SODMAP collected eight samples in this
area in 2013, in two locations, and all tested negative for SOD. The ADS flew this area in both
2012 and 2013 and found tree damage which they attributed to insects. Although analysis of this
project’s results shows that mapping results may be inaccurate, it appears that this area is already
on the radar as a possible area for SOD, and will probably be watched closely for new infections.
An analysis of the maps produced by this project indicated that densely forested, steep,
inaccessible areas are less susceptible to SOD, and identified an area in the southeast portion of
the study area that should be looked at more closely as possibly infected by SOD. However,
58
Figure 21 The 2011 SD pixels compared to ADS-identified SOD areas
59
Figure 22 The 2011 SD pixels compared to SODMAP confirmed SOD infections
60
several factors indicate that these results may be unreliable, including the unexpectedly large
numbers of pixels classified as dead trees, and the unforeseen seasonal variation. The reliability
and accuracy of the results is taken up in the next section.
4.2 Accuracy Evaluation
From an analysis of the maps and tables showing this project’s results, and the error matrices in
the previous chapter, we have already seen indications that bring into doubt the reliability of the
output – too many pixels are being classified as SD. The error matrix which evaluated the
accuracy for the SWIR/NIR index and range 400-570 applied to the Results Validation Points
measured that 68 percent of SD points were classified correctly and 73 percent of healthy trees
were classified correctly. However, large numbers of pixels that appeared to contain only healthy
trees were classified as SD, and the relatively larger number of healthy tree points within the
sample strongly affected the overall mapping accuracy. Tree health was accurately mapped 69
percent of the time, but SD was only mapped correctly in 24 percent of the cases, as shown in the
error matrix in Table 10 (Chapter 3). This means that approximately three quarters of pixels
mapped as tree death are actually healthy trees. Unfortunately, this mapping method and results
should therefore be considered unreliable.
A visual comparison of points marked falsely positive does not reveal any obvious reason
for the misclassification. They do not contain small amounts of tree death, and they do not
contain a particular type of tree. This indicates that the misclassification is due to spectral
variation. Similar types of vegetation were observed producing values over a wide range, which
caused an overlap in the index values produced by each land cover type. This is further
confirmed by areas where SD-classified pixels appear and disappear over different years.
61
Although fluctuation over different years is partially due to seasonal changes, the variation in
pixel values for the same land cover type created more variability than could be captured or
excluded with a specific value range.
To rule out misclassification, both sets of points were verified to confirm accuracy. Both
Land Cover Points and Results Validation Points SD points conform to the minimum standard of
10% canopy dead trees, but the non-random Land Cover points have an average of 21% dead
tree cover, and the random Results Validation points have an average of 12%. This means that
the index and range based on the Land Cover points is not optimized to pick up the low levels of
SOD displayed in the Results Validation points.
A histogram analysis was applied to the Results Validation Points to determine which
index and range would have mapped these points more accurately, and ranges were tested. To
facilitate comparison, a randomly-chosen subset of the HLH points was used to equalize the
numbers of HLH and SD points. The histograms of these points, shown in Appendix D, did not
show differentiation as clearly as the Land Cover Points. There was only slight offset between
the peaks of SD and HLH with the SWIR/NIR (Figures D1 and D2), NBR (Figures D3 and D4),
NDMI(Figures D5 and D6) and the NDVI (Figures D7 and D8), and no offset for the other
indices. No indices or ranges were found which produced better results than the SWIR/NIR
formula used in this project.
To determine if an index and range could be more effective at identifying Results
Validation Points with greater than 20% canopy SD, the four Results Validation Points meeting
that threshold were tested, but no indices or ranges were found which could encompass all four
SD points while excluding all HLH points. The best formula, when applied to the entire Results
62
Validation dataset, only picked up 20% (6 points) of the entire Results Validation SD point
dataset. Even a classification based on 20% SD would still have captured HLH points.
Initial analysis of these project results seemed to indicate three areas which escaped SOD
infestation, and one new area which should be looked at more closely. However, the number of
pixels mapped as SD each year is not consistent with a spreading disease, and the analysis of the
Results Classification Points found the mapping of SD to be only 24 percent accurate. Because
of these reasons, these project results must be considered unreliable.
The next chapter discusses the broader implications of this interpretation, makes
suggestions for further work, and draws conclusions about the project as a whole.
63
CHAPTER FIVE: DISCUSSION AND CONCLUSIONS
This project attempted to classify dead trees at the canopy level using 30 m raster imagery, as a
proxy for mapping the spread of Sudden Oak Death since its discovery in 1994, and concluded
that there is too much spectral variation to accomplish this reliably. The project sought a simple
method to map SOD over a large area that could be adapted to other areas to form a complete
picture of the state of SOD infestation on the west coast of North America. The motivation for
this project was that these data could have been useful to quantify the disease impact, to evaluate
success at disease management, and could have been adapted to map SOD in other susceptible
areas. Unfortunately, the methodology described in this project identified too many false
positives to be considered a dependable means to map SOD.
This study expected to find tree death spreading outwards from existing infected areas,
with occasional instances of SOD popping up in new places. The possibility for a successful
result seemed likely based on the promising Land Cover Points results. Unfortunately, the
Results Validation points showed that many areas with healthy trees were misidentified as
containing dead trees. Although the methodology was effective at mapping serious tree death at
the canopy level, it is picking up too many false positives to produce reliable results to describe
the phenomena at hand. The failure to accurately map tree death is due to a variety of spectral
signatures produced by similar vegetation which led to overlap in each land cover type’s value
range. Unfortunately, within the confines of this project’s goals – to find a simple method to map
SOD over a large area – a more refined analysis was not possible. The remainder of this chapter
discusses the implications of the project results, and makes suggestions for further research that
might overcome some of the problems that this project experienced.
64
Preliminary analysis of results implied that dense forest was less susceptible to SOD, and
that the southeast portion of the study area contained high levels of tree death not previously
noted, but the poor results from the Results Validation Points showed that we could not
substantiate these conclusions were accurate. Although this project failed to find a method to
map SOD infestation, we can determine what caused the failure and what, if anything, could
have been done to improve results.
The project identified pixels containing dead trees with a 24 percent accuracy rate. This
low accuracy rate was due to large numbers of pixels containing healthy trees being falsely
identified by the methodology as containing dead trees. This is due to a large variation in spectral
signatures for the same land cover type. Even within small areas of healthy forest, SWIR/NIR
index values ranged by up to 25 percent. While non-forest land cover areas were removed from
the study area, there was still sufficient spectral signature variation within classes to cause
confusion between pixels containing only healthy trees and those containing some dead trees.
Although this project expected to address issues caused by sub-pixel sized features, and so
approached this analysis by analyzing land cover at the canopy layer instead of at the tree-level,
this wide range of pixel index values made it difficult to identify a value range which would
exclude other land cover types. This project did not find a reliable method to map the locations
of SOD based on 30 m raster data.
Similar projects analyzing tree health using Landsat imagery claim success, but they
measured success differently than this project. Vogelmann, Tolk and Zhu (2009) used correlation
with the ADS data as an indicator of success. Meddens et al. (2013) evaluated success based on
correct classification of selected homogenous pixels. Meigs, Kennedy and Cohen (2011) and
Mahon et al. (2002) performed field testing to evaluate results. It is likely that a pixel-based
65
validation such as that used by this project would have shown a comparable poor performance.
This method of evaluation is objective, but is much more stringent than that applied to other
studies of sub-pixel sized features. Despite the measured inaccuracies, the methodology
described in this project may be useful at identifying hotspots of new infestation, but the large
number of false positives makes it difficult to interpret the specific outcomes.
This project’s methodology could have been improved with certain techniques.
Specifically, a means could be found to address both the seasonal variability in pixel values
across years, and the false positives. To remove or minimize annual variability, seasonal rainfall
and temperature data would need to be incorporated into the analysis. This could be used to limit
the analysis to data which represented land cover in roughly similar phenological conditions.
This would necessarily require using fewer images to conduct the time series analysis because
there would not be sufficient cloud-free images during the desired collection period. To remove
false positives, two different methods could have been employed. Currently, the methodology
identifies serious tree death as greater than 10 percent of pixel area containing dead trees. A
higher percentage would have produced a more focused index value range and likely have
decreased spectral overlap with pixels containing healthy trees. Another possibility would be to
devise a custom index or identify an additional band which differentiated false positive tree death
from actual tree death. Both of these were experimented with in this project without
improvement to results, but a more intense focus on these two changes might reveal a solution.
Although the aforeementioned methods might decrease the error rate found by this project, if the
new methods proved complicated, it might make the methodology too difficult to be adapted and
used by other users.
66
Some other methods which might have produced successful results include higher
resolution images, additional Land Cover Points, a smaller study area or multi-date analysis.
These options were considered but were discarded for reasons described below. Higher
resolution data, used in place of or in conjunction with Landsat, would have improved feature
resolution and project results. This was not considered because one project goal was to develop a
method to map tree death over a large area, that could be adapted to other areas. Although many
studies use data fusion methods to improve the feature differentiation of low-resolution satellite
data, a methodology which required higher resolution data might make this methodology
inadaptable to areas where such data was unavailable, or too costly to acquire, process, store
and/or interpret. Another solution considered was to create additional Land Cover Points to more
precisely identify each land cover type index value range. This was not pursued because a simple
method which could be adapted by others could not rely on such a careful creation of sample
points. The initial evaluation of the classification method showed satisfactory results using the
Land Cover Points, so it was reasonable to proceed. Refining a SD mapping method using a
smaller, more homogenous study area was considered. Decreased vegetation variation that would
be typical of many smaller study areas might have made this methodology successful, even with
low-resolution data. However, this project sought to develop methodology that could be adapted
to map SOD throughout California, and only a large study area would capture the topographic
variation necessary. A method which mapped SOD accurately in a homogenous area would
likely fail if applied to large areas with more varied topography. Several successful studies of
tree health used multi-date analysis to map subtle changes over multiple years. Based on the
minor offsets visible in the Results Validation Points histograms shown in Appendix D, a
method which detected slight changes in spectral signature might have been more effective.
67
However, this comparison would need to be performed between three or more sets of data, to
eliminate outlier values, because of the pixel value variation observed. These methods discussed
could have improved this project’s results, but complex methodologies run the risk of
discouraging others from adapting it. Although certain changes might lead to successful mapping
of SOD on a small or large scale, in a way that this project did not, the root problem is large
spectral variation for a single land cover type and overlap between land cover classes.
This project concludes that Landsat data was not able to accurately map tree death in this
densely vegetated study area. In addition to unexpected seasonal variation, there was wide
variation in pixel values for similar land cover types, which created overlap between classes
which could not be contained or excluded with a specific value range.
This project had three original research questions:
1. Can canopy-level tree death indicating SOD infestation be accurately identified
and mapped using Landsat remotely-sensed imagery?
2. Are vegetation indices alone sufficient to differentiate tree health and tree
mortality from other types of land cover?
3. Will existing supplemental land cover data sets improve or simplify the
classification process?
Although pixels containing dead trees could be identified using Landsat imagery, there was too
much spectral signature overlap with other land cover types to produce an accurate result.
Spectral signature overlap between dead trees and scrub meant that vegetation indices alone were
not sufficient. Supplemental land cover data sets used to mask out non-forest and areas with
other known causes of death were essential to ensure that the methodology was only applied to
the target land cover type.
68
Although the research questions have been addressed, a successful methodology has not
been found. A method is still needed to quantify SOD damage and assess forest health on a large
scale. Until then, the ADS surveys are the best data available to assess SOD infestation.
Ironically, the current drought in California which is killing trees from lack of water is likely to
be the most effective treatment so far to slow the spread of SOD.
69
REFERENCES
Alexander, J. M. & S. V. Swain. 2010. Sudden Oak Death Integrated Pest Management in the
Landscape. http://www.ipm.ucdavis.edu/PMG/PESTNOTES/pn74151.html(accessed
Mar 13, 2014).
Barrett, T. M., D. Gatziolis, J. S. Fried & K. L. Waddell (2006) Sudden Oak Death in California:
What Is the Potential? Journal of Forestry, 104, 61-64.
Beh, M. M., M. R. Metz, K. M. Frangioso & D. M. Rizzo (2012) The Key Host for an Invasive
Forest Pathogen Also Facilitates the Pathogen's Survival of Wildfire in California
Forests. New Phytologist, 196, 1145-1154.
Collins, B. R., J. L. Parke, B. Lachenbruch & E. M. Hansen (2009) The Effects of Phytophthora
Ramorum Infection on Hydraulic Conductivity and Tylosis Formation in Tanoak
Sapwood. Canadian Journal of Forest Research, 39, 1766-1776.
Cushman, J. H. & R. K. Meentemeyer (2008) Multi scale Patterns of Human Activity and the
Incidence of an Exotic Forest Pathogen. Journal of Ecology, 96, 766-776.
Davidson, J. M., A. C. Wickland, H. A. Patterson, K. R. Falk & D. M. Rizzo (2005) Transmission of Phytophthora Ramorum in Mixed-evergreen Forest in California.
Phytopathology, 95, 587-596.
DOI-USGS. 1989-2011. "Landsat 4/5 TM Multispectral Imagery." Earth Explorer.
http://earthexplorer.usgs.gov/(accessed May 6, 2014).
---. Product Guide Landsat Climate Data Record (CDR) Surface Reflectance. US Geological
Survey. US Geological Survey, December 2013.
---. 2014. Landsat Processing Details. http://landsat.usgs.gov/Landsat_Processing_Details.php
(accessed May 14, 2014).
Filipe, J. A. N., R. C. Cobb, R. K. Meentemeyer, C. A. Lee, Y. S. Valachovic, A. R. Cook, D. M.
Rizzo & C. A. Gilligan (2012) Landscape Epidemiology and Control of Pathogens with
Cryptic and Long-distance Dispersal: Sudden Oak Death in Northern Californian Forests.
PLoS Computational Biology, 8, e1002328.
70
Forestry Commission UK. 2014. Phytophthora Ramorum - Frequently Asked Questions.
http://www.forestry.gov.uk/forestry/INFD-5UBESN(accessed Feb 9, 2014).
Garbelotto, M., P. Svihra & D. Rizzo (2001) New Pests and Diseases: Sudden Oak Death
Syndrome Fells Three Oak Species. California Agriculture, 55, 9-19.
Geospatial Innovation Facility UC Berkeley and Center for Biodiversity and Conservation. 2003.
Landsat Spectral Band Information. Berkeley, California.
http://gif.berkeley.edu/documents/Landsat%20Band%20Information.pdf(accessed May
27, 2014).
Goodwin, N. R., N. C. Coops, M. A. Wulder, S. Gillanders, T. A. Schroeder & T. Nelson (2008) Estimation of Insect Infestation Dynamics Using a Temporal Sequence of Landsat Data.
Remote Sensing of Environment, 112, 3680-3689.
Grünwald, N. J., M. Garbelotto, E. M. Goss, K. Heungens & S. Prospero (2012) Emergence of
the Sudden Oak Death Pathogen Phytophthora Ramorum. Trends in Microbiology, 20,
131-138.
Guo, Q., M. Kelly & C. H. Graham (2005) Support Vector Machines for Predicting Distribution
of Sudden Oak Death in California. Ecological Modelling, 182, 75-90.
Hansen, E., J. Parke & W. Sutton (2005) Susceptibility of Oregon Forest Trees and Shrubs to
Phytophthora Ramorum: A Comparison of Artificial Inoculation and Natural Infection.
Plant Disease, 89, 63-70.
Heath, Z., J. Moore, B. Oblinger & M. Woods. 2012. "2012 Aerial Survey Results: California."
USDA-FS Pacific Southwest Region Forest Health Protection,.
http://www.fs.usda.gov/Internet/FSE_DOCUMENTS/stelprdb5414677.pdf(accessed
August 25, 2013).
Jin, S., L. Yang, P. Danielson, C. Homer, J. Fry & G. Xian (2013) A Comprehensive Change
Detection Method for Updating the National Land Cover Database to Circa 2011. Remote
Sensing of Environment, 132, 159-175.
Kelly, M. & K. Tuxen. 2001. Oakmapper. Berkeley, California. http://www.oakmapper.org/
(accessed 19 April 2013).
71
---(2003) WebGIS for Monitoring “Sudden Oak Death” in Coastal California. Computers,
Environment and Urban Systems, 27, 527-547.
Kelly, M., K. Tuxen & F. Kearns (2004) Geospatial Informatics for Management of a New
Forest Disease: Sudden Oak Death. Photogrammetric Engineering and Remote Sensing,
70, 1001-1008.
Kluza, D. A., D. A. Vieglais, J. K. Andreasen & A. T. Peterson (2007) Sudden Oak Death:
Geographic Risk Estimates and Predictions of Origins. Plant Pathology, 56, 580-587.
Liu, D., M. Kelly & P. Gong (2006) A Spatial–temporal Approach to Monitoring Forest Disease
Spread Using Multi-temporal High Spatial Resolution Imagery. Remote Sensing of
Environment, 101, 167-180.
Mahon, L., C. Fischer, L. Levien & J. Mai. The Use of Remotely Sensed Data for the Detection,
Mapping and Monitoring of Sudden Oak Death. United States Department of Agriculture
United States Forest Service. USDA-FS, 2002.
Maloney, P. E., S. C. Lynch, S. F. Kane, C. E. Jensen & D. M. Rizzo (2005) Establishment of an
Emerging Generalist Pathogen in Redwood Forest Communities. Journal of Ecology, 93,
899-905.
Mascheretti, S., P. J. P. Croucher, A. Vettraino, S. Prospero & M. Garbelotto (2008) Reconstruction of the Sudden Oak Death Epidemic in California Through Microsatellite
Analysis of the Pathogen Phytophthora Ramorum. Molecular Ecology, 17, 2755-2768.
McGinnis, L. (2008) Tree Heartwood Extracts Halt Sudden Oak Death. Agricultural Research,
56, 10-11.
Meddens, A. J. H., J. A. Hicke, L. A. Vierling & A. T. Hudak (2013) Evaluating Methods to
Detect Bark Beetle-caused Tree Mortality Using Single-date and Multi-date Landsat
Imagery. Remote Sensing of Environment, 132, 49-58.
Meentemeyer, R., D. Rizzo, W. Mark & E. Lotz (2004) Mapping the Risk of Establishment and
Spread of Sudden Oak Death in California. Forest Ecology and Management, 200, 195-
214.
72
Meentemeyer, R. K., N. E. Rank, D. A. Shoemaker, C. B. Oneal, A. C. Wickland, K. M.
Frangioso & D. M. Rizzo (2008) Impact of Sudden Oak Death on Tree Mortality in the
Big Sur Ecoregion of California. Biological Invasions, 2008, 1243-1255.
Meigs, G. W., R. E. Kennedy & W. B. Cohen (2011) A Landsat Time Series Approach to
Characterize Bark Beetle and Defoliator Impacts on Tree Mortality and Surface Fuels in
Conifer Forests. Remote Sensing of Environment, 115, 3707-3718.
Parke, J., E. Oh, S. Voelker, E. Hansen, G. Buckles & B. Lachenbruch (2007) Phytophthora
Ramorum Colonizes Tanoak Xylem and is Associated with Reduced Stem Water
Transport. Phytopathology, 97, 1558-1567.
Quinn, J. W. 2001. Band Combinations. Portland State University.
http://web.pdx.edu/~emch/ip1/bandcombinations.html(accessed September 25, 2013).
Ramage, B. S. & K. L. O’Hara (2010) Sudden Oak Death-Induced Tanoak Mortality in Coast
Redwood Forests: Current and Predicted Impacts to Stand Structure. Forests, 1, 114-130.
Ramage, B. S., K. L. O’Hara & A. B. Forrestel (2011) Forest Transformation Resulting from an
Exotic Pathogen: Regeneration and Tanoak Mortality in Coast Redwood Stands Affected
by Sudden Oak Death. Canadian Journal of Forest Research, 41, 763-772.
Rizzo, D. M., M. Garbelotto & E. M. Hansen (2005) Phytophthora Ramorum: Integrative
Research and Management of an Emerging Pathogen in California and Oregon Forests.
Annual Review of Phytopathology, 43, 309-335.
Sutton, W., E. Hansen, P. Reeser & A. Kanaskie (2009) Stream Monitoring for Detection of
Phytophthora Ramorum in Oregon Tanoak Forests. Plant Disease, 93, 1182-1186.
USDA-FS Forest Health Monitoring Program. 1978-2013. "Aerial Detection Monitoring."
United States Department of Agriculture, United States Forest Service.
http://www.fs.usda.gov/detail/r5/forest-grasslandhealth/?cid=fsbdev3_046696(accessed
May 6, 2014).
USDA-FS Pacific Southwest Region. 2014. Aerial Detection Survey: Methodology. Vallejo, CA.
http://www.fs.usda.gov/detail/r5/forest-grasslandhealth/?cid=stelprdb5429568(accessed
August 25, 2013).
73
USDA Animal and Plant Health Inspection Service. 2012. APHIS List of Regulated Hosts and
Plants Proven or Associated with Phytophthora Ramorum.
http://www.aphis.usda.gov/plant_health/plant_pest_info/pram/downloads/pdf_files/usdap
rlist.pdf(accessed August 8, 2013).
Vogelmann, J. E., B. Tolk & Z. Zhu (2009) Monitoring Forest Changes in the Southwestern
United States Using Multitemporal Landsat Data. Remote Sensing of Environment, 113,
1739-1748.
74
APPENDIX A: HOSTS REGULATED FOR PHYTOPHTHORA RAMORUM
Table A1: Proven host plants regulated for Phytophthora Ramorum
Scientific Name Common Name(s) Acer macrophyllum Bigleaf maple
Acer pseudoplatanus Planetree maple
Adiantum aleuticum Western maidenhair fern
Adiantum jordanii California maidenhair fern
Aesculus californica California buckeye
Aesculus hippocastanum Horse chestnut
Arbutus menziesii Madrone
Arctostaphylos manzanita Manzanita
Calluna vulgaris Scotch heather
Camellia spp. Camellia - all species, hybrids and cultivars
Castanea sativa Sweet chestnut
Cinnamomum camphora Camphor tree
Fagus sylvatica European beech
Frangula californica (Rhamnus californica) California coffeeberry
Frangula purshiana(Rhamnus purshiana) Cascara
Fraxinus excelsior European ash
Griselinia littoralis Griselinia
Hamamelis virginiana Witch hazel
Heteromeles arbutifolia Toyon
Kalmia spp. Mountain laurel - all species, hybrids and
cultivars
Lithocarpus densiflorus Tanoak
Lonicera hispidula California honeysuckle
Laurus nobilis Bay laurel
Magnolia doltsopa (Michelia doltsopa) Michelia
Maianthemum racemosum (Smilacina
racemosa) False Solomon’s seal
Parrotia persica Persian ironwood
Photinia fraseri Red tip photinia
Pieris spp. Andromeda, Pieris - all species, hybrids and
cultivars
Pseudotsuga menziesii var. menziesii Douglas fir
75
Table A1(continued) Scientific Name Common Name(s) Quercus agrifolia Coast live oak
Quercus cerris European turkey oak
Quercus chrysolepis Canyon live oak
Quercus falcata Southern red oak
Quercus ilex Holm oak
Quercus kelloggii California black oak
Quercus parvula var. shrevei Shreve’s oak
Rhododendron spp. Rhododendron (including
azalea) – all species, hybrids and cultivars
Rosa gymnocarpa Wood rose
Salix caprea Goat willow
Sequoia sempervirens Coast redwood
Syringa vulgaris Lilac
Taxus baccata European yew
Trientalis latifolia Western starflower
Umbellularia californica California bay laurel, pepperwood, Oregon
myrtle
Vaccinium ovatum Evergreen huckleberry
Viburnum spp. Viburnum – all species, hybrids and cultivars
Source: USDA Animal and Plant Health Inspection Service (2012)
76
APPENDIX B: INDEX ANALYSIS HISTOGRAMS FOR LAND COVER POINTS
Figure B1 The SWIR/NIR Land Cover Points histogram for serious tree death.
Figure B2 The SWIR/NIR Land Cover Points histogram for healthy trees. This index is a good option because the peaks of the two
histograms are separated and don’t overlap with large quantities in the other graph. This index was one of three considered to
differentiate healthy from SD points, and the range of 400-570 was ultimately used.
77
Figure B3 The NBR Land Cover Points histogram for serious tree death.
Figure B4 The NBR Land Cover Points histogram for healthy trees. This index was not chosen because the left HLH peak overlaps
with three of the SD points.
78
Figure B5 The NDMI Land Cover Points histogram for serious tree death.
Figure B6 The NDMI Land Cover Points histogram for healthy trees. This index was considered for use to differentiate the healthy
and SD points. Although many of the points are in overlapping ranges, the peaks of the histograms are separated and do not overlap
with points in the other class.
79
Figure B7 The NDVI Land Cover Points histogram for serious tree death.
Figure B8 The NDVI Land Cover Points histogram for healthy trees. This index is not ideal because the main curve of each histogram
is so close to the other, and most of the points are in overlapping ranges.
80
Figure B9 The RGI Land Cover Points histogram for serious tree death.
Figure B10 The RGI Land Cover Points histogram for healthy trees. This index was not used because here is too much overlap
between the ranges and the peaks.
81
Figure B11 The Tasseled Cap Bright Land Cover Points histogram for serious tree death.
Figure B12 The Tasseled Cap Bright Land Cover Points histogram for healthy trees. This index was not used because there was no
large a range of values in each chart, no clear histogram peak, and too much overlap.
82
Figure B13 The Tasseled Cap Green Land Cover Points histogram for serious tree death.
Figure B14 The Tasseled Cap Green Land Cover Points histogram for healthy trees. This index was not used because the values are
spread over too great a range with too much overlap between the two classes.
83
Figure B15 The Tasseled Cap Wet Land Cover Points histogram for serious tree death.
Figure B16 The Tasseled Cap Wet Land Cover Points histogram for healthy trees. This index was considered for use to differentiate
the healthy and SD points. Although there is undesireable overlap between the ranges, the histogram peaks are separated and there are
few overlapping points at the peak ranges.
84
APPENDIX C: PREDICTED MAPS OF SUDDEN OAK DEATH
Figure C1 The extent of serious tree death in 1994. Each pixel classified as SD is displayed as a
single yellow point. Non-forested areas, masked out, are shown in brown.
85
Figure C2 The extent of serious tree death in 1995. Each pixel classified as SD is displayed as a
single yellow point. Non-forested areas, masked out, are shown in brown.
86
Figure C3 The extent of serious tree death in 1996. Each pixel classified as SD is displayed as a
single yellow point. Non-forested areas, masked out, are shown in brown.
87
Figure C4 The extent of serious tree death in 1997. Each pixel classified as SD is displayed as a
single yellow point. Non-forested areas, masked out, are shown in brown.
88
Figure C5 The extent of serious tree death in 1998. Each pixel classified as SD is displayed as a
single yellow point. Non-forested areas, masked out, are shown in brown.
89
Figure C6 The extent of serious tree death in 1999. This year’s data showed the fewest pixels
classified as SD. Each pixel classified as SD is displayed as a single yellow point. Non-forested
areas, masked out, are shown in brown.
90
Figure C7 The extent of serious tree death in 2000. Each pixel classified as SD is displayed as a
single yellow point. Non-forested areas, masked out, are shown in brown.
91
Figure C8 The extent of serious tree death in 2001. Each pixel classified as SD is displayed as a
single yellow point. Non-forested areas, masked out, are shown in brown.
92
Figure C9 The extent of serious tree death in 2002. Each pixel classified as SD is displayed as a
single yellow point. Non-forested areas, masked out, are shown in brown.
93
Figure C10 The extent of serious tree death in 2003. Each pixel classified as SD is displayed as
a single yellow point. Non-forested areas, masked out, are shown in brown.
94
Figure C11 The extent of serious tree death in 2004. Each pixel classified as SD is displayed as
a single yellow point. Non-forested areas, masked out, are shown in brown.
95
Figure C12 The extent of serious tree death in 2005. Each pixel classified as SD is displayed as
a single yellow point. Non-forested areas, masked out, are shown in brown.
96
Figure C13 The extent of serious tree death in 2006. Each pixel classified as SD is displayed as
a single yellow point. Non-forested areas, masked out, are shown in brown.
97
Figure C14 The extent of serious tree death in 2007. Each pixel classified as SD is displayed as
a single yellow point. Non-forested areas, masked out, are shown in brown.
98
Figure C15 The extent of serious tree death in 2008. Each pixel classified as SD is displayed as
a single yellow point. Non-forested areas, masked out, are shown in brown.
99
Figure C16 The extent of serious tree death in 2009. This year’s data showed the most pixels
classified as SD. Each pixel classified as SD is displayed as a single yellow point. Non-forested
areas, masked out, are shown in brown.
100
Figure C17 The extent of serious tree death in 2010. Each pixel classified as SD is displayed as
a single yellow point. Non-forested areas, masked out, are shown in brown.
101
Figure C18 The extent of serious tree death in 2011. Each pixel classified as SD is displayed as
a single yellow point. Non-forested areas, masked out, are shown in brown.
102
APPENDIX D: HISTOGRAMS FOR RESULTS VALIDATION POINTS
Figure D1 The SWIR/NIR Results Validation Points histogram for serious tree death.
Figure D2 The SWIR/NIR Results Validation Points histogram for healthy trees. The main range for the SD points in Figure D1 is
shifted slightly to the right in comparison to the HLH points.
103
Figure D3 The NBR Results Validation Points histogram for serious tree death.
Figure D4 The NBR Results Validation Points histogram for healthy trees. The main curve for SD points, in Figure D3, is shifted
slightly to the left in comparison to the HLH points.
104
Figure D5 The NDMI Results Validation Points histogram for serious tree death.
Figure D6 The NDMI Results Validation Points histogram for healthy trees. The main range of the SD points in Figure D5 is shifted
slightly to the left in comparison to the HLH graph.
105
Figure D7 The NDVI Results Validation Points histogram for serious tree death.
Figure D8 The NDVI Results Validation Points histogram for healthy trees. The main body of SD points in Figure D7 is shifted
slightly to the right in comparison to HLH.
106
Figure D9 The RGI Results Validation Points histogram for serious tree death.
Figure D10 The RGI Results Validation Points histogram for healthy trees. There is correlation between the main sets of points in
these ranges so this index is not useful for differentiating the classes.
107
Figure D11 The Tasseled Cap Bright Results Validation Points histogram for serious tree death.
Figure D12 The Tasseled Cap Bright Results Validation Points histogram for healthy trees. The values are spread over too wide a
range to be able to use this index to identify a change in spectral signature.
108
Figure D13 The Tasseled Cap Green Results Validation Points histogram for serious tree death.
Figure D14 The Tasseled Cap Green Results Validation Points histogram for healthy trees. The values are spread over too wide a
range to be able to use this index to identify a change in spectral signature.
109
Figure D15 The Tasseled Cap Wet Results Validation Points histogram for serious tree death.
Figure D16 The Tasseled Cap Wet Results Validation Points histogram for healthy trees. The values are spread over too wide a range
to be able to use this index to identify a change in spectral signature.
Abstract (if available)
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Using Landsat and a Bayesian hard classifier to study forest change in the Salmon Creek Watershed area from 1972–2013
PDF
Using census data, urban-land cover classification, and dasymetric mapping to measure urban growth of the lower Rio Grande Valley, Texas
PDF
Integrating Landsat and California pesticide exposure estimation at aggregated analysis scales: accuracy assessment of rurality
PDF
Questioning the cause of calamity: using remotely sensed data to assess successive fire events
PDF
Finding the green in greenspace: an examination of geospatial measures of greenspace for use in exposure studies
PDF
Using GIS to identify potential dynamic marine protected areas: a case study using shortfin mako shark tagging data in New Zealand
PDF
Evaluating glacier movement fluctuations using remote sensing: a case study of the Baird, Patterson, LeConte, and Shakes Glaciers in central southeastern Alaska
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
Classifying the alpines: developing a methodology to track environmental changes in the alpines utilizing remote sensing and ecozone vegetation patterns
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
PDF
Investigation and analysis of land use/land cover to determine the impact of policy developments in cities
PDF
Remote analysis of avalanche terrain features: identifying routes, avoiding hazards
PDF
The role of amenities in measuring park accessibility: a case study of Downey, California
PDF
A comparative study of ground and satellite evapotranspiration models for southern California
PDF
Does the Bay Area have a social center? Delimiting the postmodern urban center of the San Francisco Bay Area
PDF
Assessing the transferability of a species distribution model for predicting the distribution of invasive cogongrass in Alabama
PDF
Surface representations of rainfall at small extents: a study of rainfall mapping based on volunteered geographic information in Kona, Hawaii
PDF
Exploring urban change using historical maps: the industrialization of Long Island City (LIC), New York
PDF
Mixed forest image classification of paper birch: using AVIRIS bandwidths ranging from 530 to 745 nm
PDF
A Maxent-based model for identifying local-scale tree species richness patch boundaries in the Lake Tahoe Basin of California and Nevada
Asset Metadata
Creator
Gillis, Trinka
(author)
Core Title
Use of remotely sensed imagery to map sudden oak death (Phytophthora ramorum) in the Santa Cruz Mountains
School
College of Letters, Arts and Sciences
Degree
Master of Science
Degree Program
Geographic Information Science and Technology
Publication Date
07/01/2014
Defense Date
05/29/2014
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
landsat,OAI-PMH Harvest,Phytophthora ramorum,ramorum blight,Santa Cruz,sudden oak death
Format
application/pdf
(imt)
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Wilson, John P. (
committee chair
), Lee, Su Jin (
committee member
), Rashed, Tarek (
committee member
)
Creator Email
tgillis@usc.edu,trinkagillis@gmail.com
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c3-428759
Unique identifier
UC11286567
Identifier
etd-GillisTrin-2603.pdf (filename),usctheses-c3-428759 (legacy record id)
Legacy Identifier
etd-GillisTrin-2603.pdf
Dmrecord
428759
Document Type
Thesis
Format
application/pdf (imt)
Rights
Gillis, Trinka
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
Tags
landsat
Phytophthora ramorum
ramorum blight
sudden oak death