University of Southern California Dissertations and Theses

Using VHSR multispectral imagery and object-based extraction to discover vernal meadows through vegetative persistence at Fort Ord, California

(USC Thesis Other)

Using VHSR multispectral imagery and object-based extraction to discover vernal meadows through vegetative persistence at Fort Ord, California

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content USING VHSR MULTISPECTRAL IMAGERY AND OBJECT-BASED EXTRACTION TO
DISCOVER VERNAL MEADOWS THROUGH VEGETATIVE PERSISTENCE AT FORT
ORD, CALIFORNIA

by

Charles Joseph Hanley III

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)

August 2015

Copyright 2015 Charles Joseph Hanley III

i

DEDICATION
I dedicate this document to my beautiful and amazingly patient wife, Shannon, without whom I
could never have completed this project. You are my true one and only, my soulmate, and I am
profoundly lucky to call you my best friend. And to my crazy little rugrat Connor, who is the
constant source of inspiration and joy in my life. I do this all for the two of you!

ii

ACKNOWLEDGMENTS
There are too many people to which I owe a debt of gratitude to mention them all here. On a
personal note, I must first thank my wife for putting up with the two-year disappearance of her
husband into a computer screen. I am so thankful to have returned to you from my long digital
sojourn. Your love and support are the cornerstones of my life, and I am profoundly grateful to
have you. To Connor, thank you for being the driving force that impels me on in life. Dad and
Gina, what can I even begin to say? I could not have gotten anywhere without you. You both
know very well the road I have traveled, and deserve thanks and praise for paving it ahead of me.
To my Mom, who has always lent an ear and had a kind supportive word, you have my eternal
appreciation. Thank you. To Doug and Jackie, you have been so supportive throughout this
process. I owe you a great debt of gratitude. To Dan O’Rourke, the generous donation of your
aircraft and piloting were an important part of this research project, and were greatly appreciated.
Professionally, I would like to thank Professor John Wilson. For agreeing to advise my
project, I am forever grateful. Not often in life does one receive tutelage from someone whom is
truly at the top of their respective field. I have always felt encouraged by your feedback. All the
backing I required to complete this work was graciously provided to me, at the behest of Dr.
Wilson. Thank you also to Professors Travis Longcore and Su Jin Lee, both for sitting on my
committee and for providing invaluable feedback. Professor Robert Vos deserves many thanks
as well, for serving as my de facto advisor while I conceived of this project. To Richard Tsung,
thanks for the amazing all-hours tech support. Thanks also go to Cary Steibel and Lyle Shurtleff
from the U.S. Army BRAC office, for being amenable and providing invaluable GIS data. And
last, but certainly not least, many sincere thanks to Devon Libby and the Digital Globe
Foundation, for graciously providing an imagery grant for this project.
iii

CONTENTS
DEDICATION ................................................................................................................................. i
ACKNOWLEDGMENTS .............................................................................................................. ii
LIST OF TABLES ......................................................................................................................... vi
LIST OF FIGURES ...................................................................................................................... vii
LIST OF ABBREVIATIONS ........................................................................................................ xi
ABSTRACT ................................................................................................................................. xiii
CHAPTER 1: INTRODUCTION ................................................................................................... 1
1.1 Aerial Wetland Surveillance ................................................................................................. 3
1.2 Wetlands and Vernal Pools at Fort Ord ................................................................................ 4
1.3 Scope ..................................................................................................................................... 8
CHAPTER 2: BACKGROUND AND LITERATURE REVIEW ............................................... 13
2.1 Vernal Pools ........................................................................................................................ 14
2.2 Vernal Pools at Fort Ord ..................................................................................................... 16
2.3 Remote Sensing of Wetlands and Vernal Pools ................................................................. 19
2.3.1 Surveying Wetlands Using Image Analysis 19
CHAPTER 3: METHODOLOGY ................................................................................................ 22
3.1 Selection of Study Area and Time-Period .......................................................................... 23
3.1.1 Criteria for Selecting Analytical Datasets 25
3.2 Study Design ....................................................................................................................... 27
3.3 Data Exploration ................................................................................................................. 32
3.4 Pre-Processing..................................................................................................................... 33
iv

3.5 Segmentation and Object-Based Classification .................................................................. 37
3.5.1 Fort Ord OBIA 37
3.5.2 Semantic Considerations 40
3.6 Analyzing Change ............................................................................................................... 42
CHAPTER 4: RESULTS .............................................................................................................. 47
4.1 Classification Results .......................................................................................................... 47
4.1.1 1941 Panchromatic Classification and Accuracy 48
4.1.2 2001 Ikonos Classification and Accuracy 50
4.1.3 2011 WorldView 2 Classification and Accuracy 52
4.2 Statistical Results ................................................................................................................ 55
4.2.1 70-Year CVP/PVP Comparison 56
4.2.2 Short- and Long-Term Changes in Self-Same CVP Sites 59
4.2.3 Performance of PI and WIPI 62
4.3 Flyover Observations .......................................................................................................... 63
CHAPTER 5: DISCUSSION AND CONCLUSIONS ................................................................. 69
5.1 On OBIA Accuracy............................................................................................................. 70
5.2 Comparing 70-Year Changes in PVPs and CVPs .............................................................. 72
5.3 Comparing Short- and Long-Term Patterns of Change in Self-Same CVPs ...................... 75
5.4 Sources of Error .................................................................................................................. 77
5.5 Final Thoughts .................................................................................................................... 81
REFERENCES ............................................................................................................................. 85
v

APPENDIX A: SUPPORTIVE IMAGERY ................................................................................. 95
APPENDIX B: ENLARGEMENTS OF CLASSIFICATION RESULTS ................................... 99
APPENDIX C: MAPS OF MEASURED METRICS ................................................................. 102
APPENDIX D: EXPLORATORY STATISTICS RESULTS .................................................... 109

vi

LIST OF TABLES
Table 1: Summary of Study Meadow Counts by Type and Year 33
Table 2: Object Class Names, Acronyms, and Descriptions 42
Table 3: Error Matrix (Contingency Table) for 1941 Panchromatic Classification 49
Table 4: Error Matrix (Contingency Table) for 2001 Ikonos Classification 51
Table 5: Error Matrix (Contingency Table) for 2011 WV2 Classification 53
Table 6: Group Statistics for 70 Year CVP/PVP Comparison 57
Table 7: Independent T-test Results for 70 Year CVP/PVP Comparison 57
Table 8: Group Statistics for Short- and Long-term Self-Same CVP Comparisons 61
Table 9: Dependent Samples T-test Results for Short- and Long-Term CVP Comparisons 61

vii

LIST OF FIGURES
Figure 1: Fort Ord MRS-BLM Restricted Site ............................................................................... 5
Figure 2: Previously Confirmed (ID # 80C02) Versus Previously Unconfirmed (ID # 80P03)
Vernal Pool Meadows (Source Imagery: USGS EarthExplorer, Google Earth,
8/25/2014) ...................................................................................................................... 6
Figure 3: The Fort Ord Core Area (USFWS 2005) ........................................................................ 9
Figure 4: PVP Site 70P14 Vegetative Persistence and Hydrology Indicators (Photo Credits:
Top Center, WorldView 2 2011; Bottom Right, C. Hanley 2015; All others,
Google Earth) ............................................................................................................... 11
Figure 5: Flyover Aerial of a Large Fort Ord Vernal Pool, 21 February, 2015 (Photo Credits:
C. Hanley) .................................................................................................................... 14
Figure 6: Mixed Chaparral and Oak Landscape inside the Restricted Area (Photo Credits: C.
Hanley) ......................................................................................................................... 17
Figure 7: Study Area Extent ......................................................................................................... 24
Figure 8: Exploratory Aerials Used to Confirm Site ID # 80P03 as Vernal (Photo credits,
clockwise from top left: Google Earth, USGS EarthExplorer, C. Hanley,
WorldView 2) ............................................................................................................... 26
Figure 9: Oblique Camera Angle of CVP (ID # 80P07) from Google Earth Pro ......................... 29
Figure 10: Analytical Rasters, Side by Side After Pre-processing ............................................... 30
Figure 11: Site 80C03, the Single Study Area Vernal Pool Visible from Public Roads (Photo
Credits: C. Hanley) ....................................................................................................... 31
Figure 12: 1941 B&W Raster Reassembled mosaic, Post-geometric corrections ........................ 34
Figure 13: NDVI Comparison, 2001 Ikonos (left) versus 2011 WV2 (right) .............................. 35
Figure 14: Novel False Color Composite of Ikonos 2001 Imagery .............................................. 36
Figure 15: Typical Linear FNEA and NN Workflow ................................................................... 38
Figure 16: Iceplant and Pampas Grass Comprised the "Invasive" Class (Photo Credits: C.
Hanley) ......................................................................................................................... 41
Figure 17: Lowland Complex Site ID Numbers ........................................................................... 44
Figure 18: 1941 Panchromatic Image Classification Results ....................................................... 49
Figure 19: 2001 Ikonos Image Classification Results .................................................................. 51
viii

Figure 20: 2011 WV2 Image Classification Results..................................................................... 53
Figure 21: Landscape Conversion to Disturbance, from 2001 to 2011 ........................................ 54
Figure 22: Box and Whisker Plots with Outliers, Comparing PVP to CVP, for Each
Analyzed Metric ........................................................................................................... 58
Figure 23: 70-Year WIPI for Self-Same CVP Sites ..................................................................... 64
Figure 24: 10-Year WIPI for Self-Same CVP Sites ..................................................................... 64
Figure 23: Erosion Scars Assisted in the Confirmation of Site ID# 80P07 as a CVP (Photo
Credits: C. Hanley) ....................................................................................................... 65
Figure 24: Inundation Frequencies, Compiled from a Review of 21 Historic Aerial Images ...... 66
Figure 25: CVP Site 80C01 Timeline Showing Persistent Dead Transition (DT) Vegetation
Band (Data Sources: Google Earth 2015; DigitalGlobe Foundation 2014; US
Army BRAC 2014) ...................................................................................................... 67
Figure 26: Major Erosion Scars Divert Surface Waters within the Study Area (Photo Credits:
C. Hanley) .................................................................................................................... 68
Figure 27: PVP and CVP Complexes in 1941 (top) and 2001 (bottom) ...................................... 73
Figure 28: PVP and CVP Complexes in 2001 (top) and 2011 (bottom) ...................................... 76
Figure 29: Five Previously Unconfirmed Vernal Sites within MRS-BLM .................................. 82
Figure 30: The First Ever Comprehensive Map of MRS-BLM Vernal Sites ............................... 83
Figure 31: PVP Site 70P05 Displays Floral Persistence Similar to that of Study Area CVP
(Source Imagery: Google Earth, C. Hanley—bottom right) ........................................ 95
Figure 32: Previously Unconfirmed CVP 80P03 (bottom row) Was Flooded During the
Flyover While Known CVP 80C02 Was Not (Source Imagery: USGS
EarthExplorer, Google Earth, C. Hanley—top and bottom) ........................................ 96
Figure 33: Comparative Shoreline Shapes for Known and Previously Unconfirmed CVP
(Photo Credits: C. Hanley) ........................................................................................... 97
Figure 34: Inundated CVP Site 11P01 (Previously Unknown), Likely Created by
Disturbance (Photo Credits: C. Hanley) ....................................................................... 97
Figure 35: Enlargement of Erosion Drainage Scar in Site ID# 80P07 (Photo Credits: C.
Hanley) ......................................................................................................................... 98
Figure 36: Enlargement of 1941 Classification Results ............................................................... 99
ix

Figure 37: Enlargement of 2001 Classification Results ............................................................. 100
Figure 38: Enlargement of 2011 Classification Results ............................................................. 101
Figure 39: Lowland Complex FD Changes by Type, 1941-2001 ............................................... 102
Figure 40: Lowland Complex FD Changes by Type, 2001-2011 ............................................... 102
Figure 41: Lowland Complex SI Changes by Type, 1941-2001 ................................................ 103
Figure 42: Lowland Complex SI Changes by Type, 2001-2011 ................................................ 103
Figure 43: Lowland Complex PAR Changes by Type, 1941-2001 ............................................ 104
Figure 44: Lowland Complex PAR Changes by Type, 2001-2011 ............................................ 104
Figure 45: Lowland Complex PI Values by Type, 1941-2001 ................................................... 105
Figure 46: Lowland Complex PI Values by Type, 2001-2011 ................................................... 105
Figure 47: 70-Year WIPI for Self-Same CVP ............................................................................ 106
Figure 48: 10-Year WIPI for Self-Same CVP ............................................................................ 106
Figure 49: Lowland Complex Percent Area Changes by Type, 1941-2001 ............................... 107
Figure 50: Lowland Complex Percent Area Changes by Type, 2001-2011 ............................... 107
Figure 51: Lowland Complex Percent Perimeter Changes by Type, 1941-2001 ....................... 108
Figure 52: Lowland Complex Percent Perimeter Changes by Type, 2001-2011 ....................... 108
Figure 53: Exploratory Statistics for Change in FD, 1941-2001 ................................................ 110
Figure 54: Exploratory Statistics for Change in SI, 1941-2001 .................................................. 111
Figure 55: Exploratory Statistics for Change in PAR, 1941-2001 ............................................. 113
Figure 56: Exploratory Statistics for Log(PI), 1941-2001 .......................................................... 115
Figure 57: Exploratory Statistics for % Area Change, 1941-2001 ............................................. 116
Figure 58: Exploratory Statistics for % Perimeter Change, 1941-2001 ..................................... 118
Figure 59: Exploratory Statistics Self-Same CVP FD Differences ............................................ 119
Figure 60: Exploratory Statistics Self-Same CVP SI Differences .............................................. 119
Figure 61: Exploratory Statistics Self-Same CVP PAR Differences.......................................... 120
x

Figure 62: Exploratory Statistics Self-Same CVP WIPI Differences ......................................... 120
Figure 63: Exploratory Statistics Self-Same CVP % Area Differences ..................................... 121
Figure 64: Exploratory Statistics Self-Same CVP % Perimeter Differences ............................. 121

xi

LIST OF ABBREVIATIONS
B&W Black And White (Panchromatic)
BLM Bureau of Land Management
BG Bare Ground/Dirt Track
BRAC Base Realignment and Closure
CERCLA Comprehensive Environmental Response, Compensation and Liability Act
CIR Colorized Infrared
CVP Confirmed Vernal Pool
DT Dead Transition
ETM/ETM+ Enhanced Thematic Mapper/Enhanced Thematic Mapper Plus
FB Fire Break/Burn
FD Fractal Dimension
FNEA Fractal Net Evolution Approach
GIS Geographic Information Systems
HD High Definition
HSR High Spatial Resolution
IV Invasive
KML Keyhole Markup Language
LL Lowland
LT Live Transition
MO Model Object
MRS-BLM Munitions Response Site-Bureau of Land Management
MSR Medium Spatial Resolution
xii

NDVI Normalized Difference Vegetation Index
NIR Near Infrared
NN Nearest Neighbor
OBIA Object-Based Image Analysis
OT Oak Tree
P or PERI Perimeter
PAR Perimeter to Area Ratio
PI Persistence Index
PR Paved Road
PVP Potential Vernal Pool
SI Shape Index
SPOT-5 Satellite Pour l’Observation de la Terre 5
TTA Mask Test and Training Area Mask
UL Upland
USGS United States Geological Survey
VHSR Very High Spatial Resolution
WIPI Weighted Intervals Persistence Index
WV2 WorldView 2

xiii

ABSTRACT

Vernal pools are rare, seasonal pools that form in landscape depressions and create temporary
habitat for many floral and faunal taxa. In California, as much as 90% of historic vernal pool
area has been displaced by agriculture and urbanization. Pools are commonly inhabited by
endemic, threatened, and endangered plants and animals, and are critical breeding areas for
California tiger salamanders (Ambystoma californiense) and fairy shrimp (Linderiella
occidentalis). Seasonal inundation and desiccation are driving factors behind the biotic
community structure around pools, both spatially and compositionally.
At Fort Ord, California, a rare subset of vernal pools occur perched atop relict sand dunes
in an arid chaparral environment. Fifteen vernal pools have been previously identified within the
base’s historic firing range impact area. At least 45 other lowland meadows within the impact
area meet pool topographic requirements and were evaluated for their potential to be vernal
habitats. This thesis proposes an object-based method of extracting vegetative patterns from
VHSR Ikonos and WorldView 2 satellite imagery, to compare persistence in vegetative patterns
over time. Classification results from three aerials collected over an 80-year interval were
subjected to a geospatial change analysis, and used to make short- and long-term comparisons of
known vernal meadows to themselves and other meadows in the study area. Two new metrics,
the Persistence Index and Weighted Intervals Persistence Index, were created for this study.
These indices normalize changes in geometric properties, enabling comparisons between known
vernal areas and study sites, and between self-same sites sampled at different times. PI and WIPI
results were consistent with the results from other analyzed metrics.
xiv

Strong persistence in several study sites, comparable to that of the known vernal areas,
likely indicates latent presence of a seasonal hydric regime and an elevation-based hydrological
gradient. The results of this study show that there is no statistically significant difference
between the way that vernal and other meadows change shape and size over time. This result
means that a number of lowland meadows in the impact area may have active or dormant vernal
pools because the two groups cannot be empirically differentiated from one another. This study
also positively confirmed the presence of at least five previously unrecognized vernal areas
through the detection of water in multiple aerial images. These findings merit further on-the-
ground investigation, as well as a geographical reconsideration of current conservation efforts.

1

CHAPTER 1: INTRODUCTION

Vernal pools are ephemeral wetlands that are geographically isolated from other water bodies
and inundated only on a seasonal basis. They provide habitat for a rich variety of endemic,
threatened, and endangered taxa of flora and fauna (Bauder 2005; Burne 2001; Carpenter, Stone,
and Griffin 2011; Keeler-Wolf et al. 1998; Lathrop et al. 2005; Reed and Amundson 2007; Stone
1992; Van Meter, Bailey, and Grant 2008; Van Thomme 2011; Zedler 1987). Though variations
of vernal pools occur worldwide, in most cases these environments have been drastically reduced
in number (Bauder 2005; Burne 2001; Carpenter, Stone, and Griffin 2011; Gibbes et al. 2010;
Keeler-Wolf et al. 1998; Kneitel 2014; Mattoni and Longcore 1997).
The primary threats to all temporary wetlands are urban and agricultural development
(Bauder 2005; Cormier 2001; Cutler 2006; Keeler-Wolf et al. 1998). In California alone, as
much as 90% of vernal habitat has been lost to development and farmland (Ferren et al. 1996;
Keeler-Wolf et al. 1998; Tannourji 2009). Of these remaining few, a rare subset exist in isolated
depressions amongst thickets of cismontane central maritime chaparral, a narrowly distributed
plant community with drought- and fire-resistant characteristics. An even more uncommon
subset of chaparral vernal pools, many of which may now be dormant or extinct, can be found
within the relict sand dunes of Monterey Bay’s Fort Ord. Because they are unusually unique and
occur inside an area of heavily restricted access, it is critical that these rare but diverse habitats
be identified for study and conservation.
Fort Ord’s vernal pools occur in both oak savanna and chaparral landscapes. Vernal
pools in the base’s northern zones typically develop in large, flat plains. These plains alternate
2

between grasslands and oak groves, and are densely dotted with large gopher mounds (mimas)
(Cox 1986; Reed and Amundson 2007). By contrast, vernal pools at the southern end of the base
are found isolated in grassy lowland depressions amongst dense stands of manzanita
(Arctostaphylos spp.) chaparral. In such circumstances, competition for space is intense and in
many locations there is a sharp break between upland chaparral and grassy lowland meadow
habitats. Vernal pools in particular develop strong, lasting vegetative patterns in and around
their margins because of the intermittent flooding to which their plant communities are
subjected. Likewise, non-vernal meadows in the area appear to resist the encroachment of
chaparral recruits as well, but the explanation as to why is less-readily apparent.
Patterning characteristic of vernal vegetation is also evident in some non-vernal lowland
meadows and raises the possibility that these meadows may have some latent potential to act as
vernal pools. It is possible that these areas are, or once were, ephemeral wetlands that have not
been previously recognized. The simplest explanation for the similarity in patterning is that the
same underlying process is happening in both types of location. That is, soil moisture gradients
are dictating the local floral distributions. This project primarily asked whether persistent
vegetation patterns could be measured and correlated to an inundation regime. Specifically, have
the shapes and sizes of vernal meadows changed over time in a way that is significantly different
than that of non-vernal meadows? How then would it be possible to quantify and compare such
changes? Moreover, are there remaining vernal meadows that have gone unnoticed or been
missed in previous site surveys that are discoverable using the methods demonstrated here?
One objective of this study was to create a single index number, henceforth referred to as
the Persistence Index (PI), to encapsulate the gross degree of size and shape change an ecological
patch undergoes over time. To make this evaluation, which required very high resolution aerial
3

photography and fine-scale classification, a method of object-based image analysis was
employed. The resulting classification outputs were analyzed at the patch level for a variety of
shape and size metrics. Changes in metric values were computed over short-term (10-year) and
long-term (70-year) difference intervals. The degrees of change in several key metrics were then
combined and normalized in a novel equation to create the Pl value for each study site. A
comparison of PI distributions for known vernal and non-vernal meadows was used to determine
if any differences between the two populations existed. If water is the factor controlling the
vegetation patterns in lowland meadows, as hypothesized, the relative magnitudes and variances
in PI values were expected to be significantly smaller for those areas that pool vernally, as
compared to those that do not.
1.1 Aerial Wetland Surveillance
Wetlands are frequently surveyed using flyovers and aerial image data. Panchromatic (B&W),
colorized infrared (CIR), and true-color aerial imagery have all been employed to this end
(Burne 2001; Carpenter, Stone, and Griffin 2011; Cutler 2006; Stone 1992; Tiner 2003). Manual
image interpretation via a stereoscope is a well-established but laborious method (Burne 2001;
Drägut and Blaschke 2006; Fallon 2013; Lathrop et al. 2005; Stone 1992). More recently,
supervised and unsupervised algorithms for automatic image classification have been developed,
and in the past several decades used routinely for wetlands surveying (Cutler 2006; Cormier
2001; Dissanska, Bernier, and Payette 2009; Fallon 2013; Frohn et al. 2009, 2012; Gibbes et al.
2010; Lichvar et al. 2006). Most of these perform pixel-based classifications on images that offer
no better than 15-30 m resolution (e.g. Landsat TM and ETM+ datasets). Because some vernal
pools may be smaller than three meters across (Zedler 1987), an alternative approach amenable
to use with higher resolution datasets, such as OBIA, is necessary.
4

Medium (15–5 m), high (5–1 m), and very high spatial resolution (<1 m) (MSR, HSR,
VHSR, respectively) satellite imagery in many spectral channels is now routinely available
through missions like Quickbird, Ikonos, WorldView-2, and SPOT-5. These higher resolution
images present highly-complex spectral components that confound typical pixel-based
classification processes (Blaschke 2010; Burnett and Blaschke 2003; Corcoran, Winstanley, and
Mooney 2010; Mallinis, Pleniou, and Koutsias 2010; Thompson and Gergel 2008; Yu and Zhang
2008). By comparison, object-based classification has been demonstrated to be highly accurate,
but is still an area of ongoing research and development (Corcoran, Winstanley, and Mooney
2010; Drägut and Blaschke 2006). OBIA utilizes texture, color, and spatial orientation to first
segment (outline), then classify objects in an image (Blaschke 2010; Burnett and Blaschke 2003;
Corcoran, Winstanley, and Mooney 2010; Drägut and Blaschke 2006; Hsu, Chua, and Pung
2000; Johansen et al 2007; Mallinis, Pleniou. and Koutsias 2010; Yu and Zhang 2008). OBIA
has been used to classify wetland areas, as well as to map vegetation patterns in landscapes
(Blaschke 2010; Cutler 2006; Dissanska, Bernier, and Payette 2009; Fallon 2013; Frohn et al.
2009, 2012; Gibbes et al. 2010; Johansen et al. 2007; Lichvar 2006; Mayer et al. 1997). This
thesis demonstrates a technique for integrating OBIA results with the analytical power of a
Geographic Information System (GIS) to extract vegetation patterns around potential and
confirmed vernal lowland meadows, and measure changes in those patterns over time. Historic
and contemporary aerial and satellite imagery were used to measure spatial persistence and
compare vernal lowlands with other lowland meadows.
1.2 Wetlands and Vernal Pools at Fort Ord
Fort Ord (Figure 1) occupies a plateau of relict Pleistocene and Holocene sand dunes on the
southeastern crescent of California's Monterey Bay (Smith et al. 2002). The former U.S. Army
5

Figure 1: Fort Ord MRS-BLM Restricted Site

base is a 28,000 acre (~45 square mile) CERCLA Superfund site with legacy pollution left over
from nearly 100 years of intense military activity (EMC Planning Group 2012; EMC Planning
Group and EDAW, Inc. 2001a; LFR, Weston Solutions, and Westcliffe Engineering, Inc. 2008).
While under Army control, over 8,000 ha (approximately 20,000 acres) of undeveloped
wilderness were restricted and used for historic firing ranges and training areas. Much of that
wilderness is in the impact area (Figure 1), and remains entirely off limits because it is still
polluted with dangerous munitions and unexploded ordnance. Ironically, the restricted area is
also habitat for several important ecosystems with significantly reduced ranges, including central
6

Figure 2: Previously Confirmed (ID # 80C02) Versus Previously Unconfirmed (ID # 80P03) Vernal
Pool Meadows (Source Imagery: USGS EarthExplorer, Google Earth, 8/25/2014)

maritime chaparral and collocated vernal pools (Burleson Consulting, Inc. 2006; Keeler-Wolf et
al. 1998). An unintentional consequence of the Army’s 75 year occupation of the land is that
many of the Fort Ord vernal pools have been protected from development (Bauder 2006; Keeler-
Wolf et al. 1998; Van Meter, Bailey, and Grant 2008), as can often happen with natural
landscapes near to noxious land uses (Longcore and Rich 2008). Yet, the Army’s treatment of
the land may have also led to a shift from natural wetlands and vernal pools to stock ponds and
artificial pools formed in artillery craters (Keeler-Wolf et al. 1998; Smith et al. 2002).
As visible from medium- and high-resolution imagery, there remain many depressions in
the restricted area’s landscape that are marked only as meadows or grasslands, but which bear a
strong visual resemblance to already-mapped vernal areas. Some of these grasslands (bottom
row, Figure 2) were confirmed as vernal pools during this analysis, through multiple water
sightings and extreme visual similarities. Other non-pooling meadows with similar appearances,
if they exist in lowland depressions, may themselves be dormant vernal areas. As such, they
should be examined thoroughly to assess their hydrologic potential.
7

Vernal pools, as well as much of their constituent biota, are protected by both federal and
state statutes, and they are recognized as highly valuable habitat areas for freshwater crustaceans
and amphibians (Keeler-Wolf et al. 1998; Kneitel 2014; TetraTech and Ecosystems West
Consulting Group 2014; USFWS 2005; Zedler 1987, 2003). At present, military installations
with de facto wilderness refuges, like that at Fort Ord, offer the best opportunity to reclaim and
restore vernal areas (Cooper and Perlman 1997; Keeler-Wolf et al. 1998; USFWS 2005).
Because vernal pools at Fort Ord are protected sites, knowing the locations of them is vital to any
conservation efforts. They are critical habitat and breeding areas for endangered, threatened, and
special status species like the California tiger salamander (Ambystoma californiense), fairy
shrimp (Linderiella occidentalis), and Contra Costa goldfields wildflowers (Lasthenia
conjugens) (Burleson Consulting, Inc. 2006; Keeler-Wolf et al. 1998; Lathrop et al. 2005; LFR,
Weston Solutions, and Westcliffe Engineering, Inc. 2008; Smith et al. 2002; USFWS 2005;
Wang et al. 2011). Some other vernal pool species may be largely endemic to Fort Ord.
Identifying new vernal pool habitat areas would thus be a critical boon to local and statewide
conservation efforts.
The specific definition of a vernal pool varies regionally, since they occur in a variety of
landscapes all over the globe (Bauder 2005; Burne 2001; Carpenter, Stone, and Griffin 2011;
Lathrop et al. 2005; Keeler-Wolf et al. 1998; Mattoni and Longcore 1997; Reed and Amundson
2007; Stone 2002; Van Meter, Bailey, and Grant 2008; Van Thomme 2011; Zedler 1987). In
general, vernal pools are ephemeral waterbodies that collect in a depression in the landscape
during late winter and early spring rains. They persist for at least two months, providing critical
habitat for local flora and fauna, and then slowly desiccate during the hot months of summer
(Burne 2001; Cormier 2001; Stone 1992). Importantly, because they are not permanent, they
8

cannot support fish populations that would otherwise predate upon the invertebrates and
amphibians that breed in vernal waters (Burne 2001; Lathrop et al. 2005; Munoz et al. 2009;
USFWS 2005; Zedler 2003). Pools can range in size from 1 m
2
to 36 ha, but are generally less
than 1 m deep (Keeler-Wolf et al. 1998; USFWS 2005). They lack inlets and outlets, and either
trap water with an impermeable horizon of clay or bedrock, or sit low enough to be below the
surface level of a perched water table (Bauder 2005; Burne 2001; Lathrop et al. 2005, Stone
1992). Often, a characteristic array of wildflowers will colonize the margins of a pool, making
them detectable even after surface waters are gone (Burne 2001; Burleson Consulting, Inc. 2006;
USFWS 2005; Van Thomme 2001; Zedler 1987).
1.3 Scope
Fort Ord vernal pools are part of the Fort Ord Core Area (Figure 3), a subzone of the larger
Central Coast vernal pool region (USFWS 2005). In the Fort Ord Core Area, vernal pools occur
in three basic environments: (1) in steep valleys supporting oak woodlands and annual grasslands
to the southeast; (2) in oak woodland savannas in the north; and (3) amongst the xeric (arid)
sandy soils and central maritime chaparral of the southwest relict sand dunes (Smith et al. 2002).
It is the last variety of pools that concern this study, as they reside in an uncommon vernal
environment that has been altered by human actions. For this research, vegetative alliances
described in the literature assisted in the identification of potential pool locations (Burleson
Consulting, Inc. 2006; Jones & Stokes Associates 1992; Keeler-Wolf et al. 1998; USFWS 2005;
Zedler 1987). The identification framework created here for characterizing dry or dormant pools
in coastal chaparral will be useful in informing future attempts to use remote sensing similarly in
other locations.

9

Figure 3: The Fort Ord Core Area (USFWS 2005)

A literature review uncovered no attempts to quantify Fort Ord’s extinct or dormant
vernal pool systems with aerial or remote sensing data. Although aerial imagery has been used
elsewhere for vernal pool detection several times in the past, these studies have focused
primarily on the identification of temporary waterbodies (Burne 2001; Carpenter, Stone, and
Griffin 2011; Lathrop et al. 2005; Stone 2002; Van Meter, Bailey, and Grant 2008). Because
Fort Ord has Mediterranean weather patterns (mild, moist winters and prolonged dry summers),
however, its vernal pools exhibit variable hydrological cycles. In some years they may not fill
with surface water at all, which makes water an unreliable indicator. Many organisms that use
10

vernal habitats in California are drought and fire adapted, and it is possible for pools to be
dormant for decades before re-establishing themselves when appropriate conditions return
(USFWS 2005; Van Dyke and Holl 2003; Van Dyke, Holl, and Griffin 2001).
This ability for vernal pools to lie dormant is particularly relevant at Fort Ord, where
military activities have altered the landscape. Many of the natural vernal pools may have been
degraded while, at the same time, artificial ones were being inadvertently created. Historical
images therefore provide a baseline for calculating changes in confirmed and potential vernal
pool sites, and OBIA offers an alternative method for examining the lasting effects of water and
disturbance on plant communities in and around vernal pools.
Several ground studies have mapped portions of the base’s extant vernal pools, but they
do not agree entirely on the locations of vernal pool meadows (Burleson Consulting, Inc. 2006;
Jones & Stokes Associates 1992; LFR, Weston Solutions, and Westcliffe Engineering, Inc. 2008;
Tannourji 2009). Aerial imagery provides a different perspective not afforded to samplers
working in situ. Landscape features such as the boundary between terrestrial and edge flora,
depressions and swales, seasonal wildflowers, and mima mounds are strong indicators of the
presence of latent hydric soil (Bauder 2000; Cormier 2001; Cutler 2006; Keeler-Wolf et al. 1998;
Reed and Amundson 2007; TetraTech and Ecosystems West Consulting Group 2014; USFWS
2005, Wang et al. 2011; Zedler 1987, 2003), but may not always be discerned from the ground.
While each of these indicators can be extracted using OBIA; this thesis primarily focused on
mapping locations and characteristics of the boundaries of terrestrial-edge species (Burleson
Consulting, Inc. 2006; Burne 2001; Keeler-Wolf et al. 1998; Montrone 2013; Reed and
Amundson 2007; Zedler 1987). This boundary zone is typically represented by a band of
constituent plants that undergo the most environmental stress of any in the localized vernal pool
11

Figure 4: PVP Site 70P14 Vegetative Persistence and Hydrology Indicators (Photo Credits: Top
Center, WorldView 2 2011; Bottom Right, C. Hanley 2015; All others, Google Earth)

community (Bauder 2000; Montrone 2013; Tannourji 2009; Zedler 1987). The boundary zone is
an extreme environment that is highly vulnerable to alterations in hydrologic regime, making it
an excellent indicator of change-over-time.
The underlying premise for this study was that vernal pools create a lasting imprint on the
local environment. This imprint is difficult to erase, even through burning, bulldozing, or
bombing. For example, a lowland meadow persists over an eight year time-series of photos, in
spite of surrounding disturbance and encroaching chaparral (Figure 4). The center and right-
hand panes show verdant springtime conditions that, upon close inspection, demonstrate vernal
indicators like damp soil and wildflowers. The site shown (ID# 70P14) is not a confirmed vernal
pool; whether or not it has vernal hydrology is unknown. It is, however, representative of other
PVP in appearance and floral composition. It may be a relict or dormant pool site that is still
imprinted on the land because of an ongoing subterranean hydrologic regime.
12

When a vernal pool disappears permanently due to a change in hydrology, it is possible
for its physical signature to persist in the landscape for years or decades. This project advanced
the hypothesis that vegetation patterns that persist in spite of repeated disturbances can be
combined with topographical data to determine the historic extent of vernal pool systems at Fort
Ord. The methodology adapted previously developed object-based image extraction techniques
to identify and analyze potential vernal pool sites. The process whereby relict vernal pool areas
were detected through OBIA is extensible to future studies of a similar nature. Discovering new
sites in this manner can provide an avenue for better understanding California’s coastal vernal
pool ecosystems. Using the evidence produced by this study, researchers will be able focus
survey effort on sites with the highest probability of having a vernal pool hydrology to collect
soil and plant samples, which can then be used to verify the predictions. Finally, this thesis has
produced the first comprehensive database of Fort Ord impact area vernal pool locations, which
will be immediately useful in instructing continued conservation and future restoration efforts.
13

CHAPTER 2: BACKGROUND AND LITERATURE REVIEW

Wetlands are important national resources with high ecological, economic, and social value
(Gibbes et al. 2010; Haas, Bartholomé, and Combal 2009). Remote sensing data has a long
history of use in the inventory and delineation of wetland resources, at both continental and
regional scales. Virtually every type of remote sensing data has been employed to study wetland
areas. Because wetlands are often in remote, inaccessible places covering vast stretches, the
acquisition of ground evidence might be impractical, or even impossible in many instances
(Cormier 2001; Haas, Bartholomé, and Combal 2009; Munoz et al. 2009; Stone 1992). Remote
sensing provides a solution to this problem and has proven to be a valuable tool for studying and
mapping wetlands worldwide.
In resolutions ranging from low to very high, aerial and satellite imagery provide wide
coverage, time-series data, and the ability to study remote sites ex situ. In particular, the
availability of historic aerial imagery and satellite data collected at regular return intervals makes
possible change assessments, and the detection of geographically isolated temporary waterbodies
(e.g. vernal pools). It is common practice in this field to use the presence of water in an image as
a wetland indicator (Burne 2001; Carpenter, Stone, and Griffin 2011; Dissanska, Bernier, and
Payette 2009; Frohn et al. 2009; Yu and Zhang 2008). However, because temporary isolated
waterbodies are ephemeral, they may not be flooded at the time of imaging (Andrew and Ustin
2008; Bauder 2000, 2005). Fort Ord is a case in point, as the majority of photos reviewed for
this project were taken while the landscape did not contain water features. Hence, other
indicators were needed to remotely detect vernal pools.
14

Figure 5: Flyover Aerial of a Large Fort Ord Vernal Pool, 21 February, 2015 (Photo Credits: C.
Hanley)

2.1 Vernal Pools
Vernal pools, like those found in woodlands (Figure 5), are a rare type of isolated ephemeral
wetland. They are characterized physically as topographical depressions that flood seasonally
and lack inlets or outlets (Bauder 2005; Burne 2001; Carpenter, Stone, and Griffin 2011;
Cormier 2001; Kneitel 2014; Stone 1992). Inundation typically lasts two to six months and is
followed by a lengthy desiccation period (Burne 2001; Cormier 2001; Stone 1992). Most pools
trap water because they are either located on a shallow or exposed sill of bedrock, or underlain
with a shallow clay aquitard (Carpenter, Stone, and Griffin 2011; Cormier 2001; Keeler-Wolf et
al. 1998; Mayer et al. 1997; Montrone 2013; Smith et al. 2002; USFWS 2005, Zedler 1987,
2003). In some cases, pool bottoms may extend below the surface elevation of a locally perched
water table (Ferren et al. 1996; Keeler-Wolf et al. 1998; Smith et al. 2002; Zedler 1987, 2003).
15

Biologically, vernal pools are richly diverse temporary habitats, and many pool taxa are
specially adapted to their use (Bagella, Caria, and Zuccarello 2010; Burleson Consulting, Inc.
2006; Burne 2001; EMC Planning Group and EDAW, Inc.2001b; Tannourji 2009; USFWS
2005; Van Thomme 2001; Zedler 1987). California’s vernal pools are known to be habitat for an
assortment of endangered or threatened species of endemic flora and fauna. Such organisms
have evolved to rely heavily upon specific ephemeral sites (Cormier 2001; Cutler 2006; Frohn et
al. 2012; Gordon et al. 2012; Keeler-Wolf et al. 1998; Kneitel 2014; Lane, D’Amico, and Autrey
2012; Tannourji 2009; USFWS 2005). For instance, amphibians and freshwater crustaceans use
vernal pools for breeding because their ephemeral nature precludes the presence of fish that
would prey upon them and otherwise compete for resources (Burne 2001; Cormier 2001;
Lathrop et al. 2005; Munoz et al. 2009; USFWS 2005; Zedler 2003). In addition, mammals use
vernal pools as watering holes (Zedler 1987), migratory birds for temporary stopovers and
breeding (Van Thomme 2001), and some plants are entirely restricted to the harsh environments
of pool basins or edges (Rhazi et al. 2006; Tannourji 2009; Zedler 1987).
The inundation regime drives the structure of pool vegetation such that it often results in
three distinct, concentric vegetative bands (i.e. aquatic, edge, and terrestrial plant taxa), although
in some cases the banding may be non-concentric or fragmented (Bagella, Caria, and Zuccarello
2010; Keeler-Wolf et al. 1998; Montrone 2013; Schlising and Sanders 1982; Tannourji 2009;
USFWS 2005). Because pool soils have a cyclical hydrological regime, their moisture content
varies greatly throughout the year. Soil moisture also varies along an elevation gradient. The
lowest local elevations flood first and for the longest, while higher elevations receive shorter-
duration flooding that begins later in the year and ends earlier (Bagella, Caria, and Zuccarello
2010; Bauder 2000, 2005; Cormier 2001; Fernández-Aláez, Fernández- Aláez, and Bécares
16

1999; Rhazi et al. 2006). Both the timing and the duration of the inundation affect the
development of pool vegetation, especially the location, size, and continuity of bands (Bagella,
Caria, and Zuccarello 2010; Keeler-Wolf et al. 1998; Kneitel 2014; Rhazi et al. 2006; Zedler
1987).
2.2 Vernal Pools at Fort Ord
It could be argued that certain of Fort Ord’s vernal pools are amongst the rarest of all varieties,
and not just because they are part of the remaining 10% of California’s historic cohort. These
pools exist perched on a plateau of relict sand dunes that is uncommon in itself (Mattoni and
Longcore 1997, Smith et al. 2002) and exhibits moderately to well-drained soil with little or no
natural surface runoff (Smith et al. 2002). The local climate type is Mediterranean, which has a
very limited worldwide distribution. Similarly, the dominant upland vegetation type, coastal
central maritime chaparral (visible in the foreground of Figure 6), is a threatened ecosystem
endemic to only a few small areas of California (Van Dyke and Holl 2003; Van Dyke, Holl, and
Griffin 2001). In this environment, eggs and seeds of pool taxa can lay dormant within the soil’s
seedbank, awaiting the return of water before reviving (Gordon et al. 2012; USFWS 2005; Van
Dyke and Holl 2003; Van Dyke, Holl, and Griffin 2001; Van Thomme 2001, Zedler 1987). This
dormancy can last years or decades, and is critical in ensuring survival of pool species since
hydric regimes may vary greatly from year to year and pool to pool (Gordon et al. 2012; USFWS
2005). In some years, as has been observed by the author, pools do not fill at all (Andrew and
Ustin 2008; Bauder 2000). Vernal pools in California have also been demonstrated to have high
within pool and pool to pool variation in terms of community structure (Bagella, Caria, and
Zuccarello 2010; Barbour et al. 2003; Gordon et al. 2012; Rhazi et al. 2006; Tannourji 2009).

17

Figure 6: Mixed Chaparral and Oak Landscape inside the Restricted Area (Photo Credits: C.
Hanley)

Central maritime chaparral is drought-resistant, woody vegetation dominated by
manzanita (Arctostaphylos spp.) and coyote brush (Baccharis pilularis) (Smith et al. 2002;
TetraTech and Ecosystems West Consulting Group 2014; Van Dyke and Holl 2003; Van Dyke,
Holl, and Griffin 2001). At Fort Ord, the chaparral is also interspersed with Coast Live Oak
(Quercus agrifolia), Monterey Pine (Pinus radiata), coastal sage scrub, and small lowland
meadows (TetraTech and Ecosystems West Consulting Group 2014). Although this region of
California is both drought- and fire-prone and local plant taxa are well-adapted for it, current
drought conditions are severe and the burn regime has been unnatural for 100 years or more
(EMC Planning Group and EDAW, Inc.2001a; LFR, Weston Solutions, and Westcliffe
18

Engineering, Inc. 2008; Smith et al. 2002). A shift towards either climax terrestrial communities
(e.g. mature chaparral or oak woodland) or disturbance species (e.g. invasive annuals grasses)
should be measurable in the landscape at Fort Ord, in light of the current circumstances. If the
size and shape of meadow areas exhibit measurable consistency over time, then water is likely
playing a role. Regularly flooded soils can only support certain plant species, at the exclusion of
all others (Bauder 2000, 2005; Tannourji 2009; Zedler 1987, 2003). Included on the exclusion
list are terrestrial floral species such as chaparral plants, coast live oaks, and invasive annual
grasses. In the absence of water and in the face of disturbance, vernal pool species should be
locally extirpated in favor of the aforementioned terrestrial species.
Several ground studies of Fort Ord vernal pools have been conducted, most pursuant to
the Fort Ord Reuse Plan (Burleson Consulting, Inc. 2006; Jones & Stokes Associates 1992; LFR,
Weston Solutions, and Westcliffe Engineering, Inc. 2008; Tannourji 2009). These studies have
variously characterized local flora and fauna, mapped the extents of known pools, and surveyed
for characteristic species such as California tiger salamanders and fairy shrimp. No two studies
are in agreement about the quantity or locations of vernal pool sites, however, and no remote
sensing work has been conducted in this regard. Moreover, although some studies have
discussed factors that have degraded vernal habitat at Fort Ord, there are few specific site
locations described in the literature. The quantity and locations of degraded and lost vernal pools
at Fort Ord are unknown. Sources of degradation include water diversions to holding ponds,
erosion due to road and trail cuts, cratering from large explosions, soil contamination from
remaining ordnance, and changes in climate (Keeler-Wolf et al. 1998; Smith et al. 2002). The
study area, as described in the methods section, has 15 previously confirmed vernal pool sites
and approximately 51 other lowland meadows that were viewed as having the potential to be
19

dormant or relict vernal pools. Henceforth, this document will refer to all previously known
vernal pools as confirmed vernal pools (CVP). All other lowland meadows in the study area will
interchangeably be referred to as either study sites or potential vernal pools (PVP) (sensu Burne
2001).
2.3 Remote Sensing of Wetlands and Vernal Pools
To classify vernal pools visually, images are usually selected based on optimal flyover time
(generally spring, leaf-off imagery taken in wet years) (Burne 2001; Carpenter, Stone, and
Griffin 2011; Cormier 2001; Lathrop et al. 2005; Stone 1992; Van Meter, Bailey, and Grant
2008). Time-series studies have either examined multiple images from the same year to verify
sites by measuring inundation time, or assessed change between different years (Burne 2001;
Frohn et al. 2012; Stone 1992). B&W and CIR images are preferred over true color photos
because water has a much more distinct spectral signature in B&W and CIR (Burne 2001;
Cormier 2001; Frohn et al. 2009; Lane, D’Amico, and Autrey 2012; Lathrop et al. 2005; Stone
1992; Van Dyke and Holl 2003; Van Meter, Bailey, and Grant 2008). Supplementary
topographic and soil data from a GIS are routinely utilized to verify that PVP meet the basic
physical requirements of a vernal pool (Burne 2001; Carpenter, Stone, and Griffin 2011; Cormier
2011; Haas, Bartholomé, and Combal 2009; Lichvar 2006; Munoz et al. 2009; Stone 1992; Van
Meter, Bailey, and Grant 2008; Yu and Zhang 2008). The typical scale range employed in this
type of manual classification varies form 1:30,000 to 1:4,800 (Burne 2001; Carpenter, Stone, and
Griffin 2011; Cormier 2001, Stone 1992).
2.3.1 Surveying Wetlands Using Image Analysis
More recently, automatic classification using pixel- and object-based extraction algorithms have
been shown to produce similarly accurate results with less subjectivity (Burnett and Blaschke
20

2003; Corcoran, Winstanley, and Mooney 2010; Drăgut and Blaschke 2006; Hsu, Chua, and
Pung 2000). Supervised and unsupervised pixel-based classifications are often used on low to
medium resolution data (Blaschke 2010; Haas, Bartholomé, and Combal 2009; Thompson and
Gergel 2008). They are most effective when areas to be classified are locally homogenous and
single-scale (Burnett and Blaschke 2003; Drăgut and Blaschke 2006; Johansen et al. 2007).
Pixel-based classification schemes are well-developed, and used in many applications beyond
wetlands research. Their limitation for use in land cover analysis stems from the advent of HSR
and VHSR data, which exhibits increased heterogeneity at finer scales (Blaschke 2010; Burnett
and Blaschke 2003; Corcoran, Winstanley, and Mooney 2010; Mallinis, Pleniou, and Koutsias
2010; Thompson and Gergel 2008; Yu and Zhang 2008). Frohn et al. (2011) attempted to
overcome this problem with sub-pixel classification of isolated wetlands and vernal pools, using
Landsat ETM+ hyperspectral data (30–15 m spatial resolution). They demonstrated better than
87% producer’s and 97% user’s accuracies. Conversely, Fallon (2013) used EO-1 Hyperion 60
m resolution hyperspectral data to perform sub-pixel classification of spectral endmembers in
channeled scablands, with only limited success.
High resolution imagery is more spectrally complex than low resolution because of the
increased detail (Burnett and Blaschke 2003; Coburn and Roberts 2004; Corcoran, Winstanley,
and Mooney 2010; Frohn et al. 2009; Johansen et al. 2007). Spectral information alone cannot
sufficiently describe images in which small objects and textures are resolvable. Object-based
classification algorithms combine spectral data with texture, spatial context, size, and shape to
first segment an image into homogenous polygons, then merge the polygons into objects and
classify them according to some set of knowledge-based semantics (Baatz and Schäpe 1999;
Blaschke 2010; Dissanska, Bernier, and Payette 2009; Forestier et al. 2012, Halounova 2004;
21

Hsu, Chua, and Pung 2000; Johansen et al. 2007; Mallinis, Pleniou, and Koutsias 2010;
Smeulders et al. 2000; Yu and Zhang 2008). The divide that separates spectral interpretation
from semantic interpretation is called the semantic gap; it represents the difference between
automatically and manually extracted information (Dissanska, Bernier, and Payette 2009;
Forestier et al. 2012; Smeulders et al. 2000). Object-based classification can be used to make the
semantic gap smaller (Forestier et al. 2012). OBIA processes are built to mimic the manner in
which the human eyes and brain interpret visual information, and have been demonstrated to
produce more accurate results than pixel-based spectral analyses (Baatz and Schäpe 1999;
Blaschke 2010; Hsu, Chua, and Pung 2000).
Frohn et al. (2009) demonstrated a technique for accurately distinguishing isolated
wetlands using OBIA on medium resolution Landsat-7 imagery. More typically, high resolution
data from satellite missions like Quickbird, Ikonos, WorldView 2, and SPOT-5 are used for
OBIA. In another approach, historical B&W aerial photographs were subjected to OBIA by
Dissanska, Bernier, and Payette (2009) to analyze changes in peatlands and patterned fens.
Cutler (1998) applied OBIA to distinguish homogenous landscape areas for potential inclusion in
a sampling pool. Random OBIA polygons were then chosen for comparison between field data
and high-resolution satellite imagery. OBIA has also been applied more broadly in
environmental science to classify topographic landform types (Drăgut and Blaschke 2006), forest
and riparian habitat (Johansen et al. 2007), heterogenous and fragmented temperate rainforests
(Thompson and Gergel 2008), type and age class of forest stands (Coburn and Roberts 2004),
and vegetation patterns in African savannas (Gibbes et al. 2010).
22

CHAPTER 3: METHODOLOGY

This chapter details the process by which this thesis was constructed and the research it describes
executed. Descriptions of the selection of the study area and period are provided first in Section
3.1. Section 3.2 provides an overview of the study design and details previous scientific studies
that were particularly influential to the research described here. The next three sections, in
sequence, discuss the system used to explore, pre-process, and segment/classify the datasets.
Finally, the change analysis approach that was used is documented in Section 3.6.
The methods described below draw heavily upon previous scientific studies of chaparral
ecosystems, remote wetlands surveillance techniques, and image analysis algorithms.
Nonetheless, the route from concept to execution required much trial-and-error, a trait common
to object-based image analyses (Fallon 2013; Frohn et al. 2009; Yu and Zhang 2008; Zhang and
Maxwell 2006). Many details of the development of the segmentation and classification process
trees have been omitted here for brevity’s sake, and because they are inconsequential to the
outcome of this analysis. The landscape patterns are easily enough discerned that similarly
accurate OBIA results could be achieved in multiple ways.
Real-world objects vary in size, shape, texture, spectral signature, and spatial context, so
OBIA must account for these factors (Baatz and Schäpe 2000; Dissanska, Bernier, and Payette
2009). However, humans are able to classify visual objects immediately and discretely, without
consciously considering such things. For example, we know an automobile when we see one;
our brains automatically classify motorcycles, sedans, and trucks into this category even though
they do not look similar. It is much more difficult for computer software to do the same equally
23

well. Thus, according to Baatz and Schäpe (2000), the human eye is the ultimate judge of image
classification results. OBIA attempts to mimic the human ability to classify objects, in imagery
and video (Blaschke 2010; Hsu, Chua, and Pung 2000). Whereas pixel-based analyses utilize
non-spatial histograms that suffer from lost spatial information, object-based systems are better-
suited to handle the inherent spatial complexity of VHSR imagery (Hsu, Chua, and Pung 2000).
3.1 Selection of Study Area and Time-Period
This project’s study extent includes the majority of the historic impact area, along with the
interim action ranges and the Seaside munitions response area. In total, the 2,578 ha (6,371
acres) area is bounded by Eucalyptus Road to the north, General Jim Moore Boulevard to the
west, South Boundary Road to the south, and Impossible Canyon to the East. This location, also
known as MRS-BLM, was chosen because its terrain and dominant vegetation are distinctive
compared to the majority of the base’s grounds. Vernal pools in the impact area exhibit more
visually obvious vegetation patterns than those found elsewhere on Ft. Ord. The primary upland
vegetation is central maritime chaparral occasionally interspersed with coastal sage scrub, coast
live oak trees, and mixed conifer stands (Burleson Consulting, Inc. 2006; Smith et al. 2002;
TetraTech and Ecosystems West Consulting Group 2014; Van Dyke and Holl 2003). Lowlands,
by contrast, are typically grassy meadows (TetraTech and Ecosystems West Consulting Group
2014). The climate is Mediterranean and prone to severe drought (Bauder 2000, 2005; Smith et
al. 2002; USFWS 2005; Van Dyke, Holl, and Griffin 2001). The two soil complexes comprising
the study area are Baywood and Arnold, both of which are varieties of well-drained relict sand
dunes approaching 100 m deep (EMC Planning Group and EDAW, Inc.2001b; TetraTech and
Ecosystems West Consulting Group 2014; Smith et al. 2002). Of note, however, is the fact that
24

Figure 7: Study Area Extent

vernal soil samples taken within the study area by Burleson Consulting Inc. (2006) showed a
predominant association with clay rich, low-permeability Antioch soils.
The study area (Figure 7) is about 2.4 km (1.50 mi) to 8.3 km (5.15 mi) from the
coastline. The vertical range covers only around 215 m, from 60 to 275 m above mean sea level,
but contains significant microtopographic relief. The area generally decreases in elevation from
north to south and from east to west. However, the microtopography is such that hills and
valleys appear to occur isotropically in localized areas, independently of elevation.
25

Image data were used for either exploratory or analytical purposes, depending on the
type. Exploratory data were any historical aerial or satellite images that provided context or
assisted in framing the study. There were three primary sources for exploratory images: the U.S.
Army Base Realignment and Closure (BRAC) office at Fort Ord, the USGS EarthExplorer tool
(earthexplorer.usgs.gov/, last accessed 25 September 2014), and Google Earth Pro
(https://www.google.com/earth/, last accessed 31 March 2015). Data types included MSR B&W
photos, low resolution CIR, and VHSR multispectral satellite imagery. Exploratory data were
used to count and log historic inundation events, provide research context, support decision-
making processes, and assess physical site characteristics. In most instances, exploratory images
lacked complete metadata and did not include sensor information, spatial or spectral resolutions,
or acquisition dates. In some cases they were not georeferenced. Yet they still provided a
wealth of information and the opportunity for limited analysis. Unexpectedly, the images were
utilized to confirm that certain PVP are in fact CVP (Figures 2 and 8).
3.1.1 Criteria for Selecting Analytical Datasets
Analytical data included two raw VHSR multispectral satellite datasets and one high resolution
panchromatic aerial photograph. The project received a student research grant through the
commercial vendor DigitalGlobe and the DigitalGlobe Foundation that provided the requisite
data free of charge. The images were selected based upon the following six criteria:
1. Complete coverage of the study area;
2. Minimum of 10 years separation between datasets;
3. Maximum cloud cover of 1%;
4. Image acquisition date in late winter or early spring;
26

Figure 8: Exploratory Aerials Used to Confirm Site ID # 80P03 as Vernal (Photo credits, clockwise
from top left: Google Earth, USGS EarthExplorer, C. Hanley, WorldView 2)

5. Panchromatic resolution finer than 1.0 m; and
6. Multispectral channels available.
Of the datasets from which to choose, one Ikonos (0.82 m panchromatic, 3.28 m blue,
green, red, red edge, yellow, coastal, and near infrared) and one WorldView 2 (0.50 m
panchromatic, 2.00 m blue, green, red, red edge, yellow, coastal, near infrared 1, and near
infrared 2) dataset were selected. The Ikonos data was acquired 07 February, 2001, while that
from WorldView 2 was obtained ten years later, on 15 April, 2011. Both presented 0% cloud
cover and 100% aerial coverage of the study area. The analysis was restricted to the red, blue,
green, and near infrared channels (NIR) for each multispectral dataset. The panchromatic aerial
27

mosaic was taken in 1941, and is of unknown but sufficiently high resolution to warrant
inclusion as a baseline image for this study; it appears to have been resampled to a pixel
resolution of 0.5 m.
Although the datasets did not have identical spatial resolutions, the OBIA was performed
separately on each and the change analysis was based on vector polygons. Therefore all data
were assumed to be functionally comparable in spite of the minor discrepancies in pixel scale.
Positional differences that could not be resolved through georeferencing and geometric
corrections were ignored because the final change analysis was based upon the geometry of
polygons, not their absolute real-world positions. To wit, a change in polygon size or shape
implies a change in boundary location, even if the precise coordinates of the new and old
locations are not accurately known.
3.2 Study Design
This study began as an investigation of historical aerial imagery. It was rapidly evident that
while the combination of anthropogenic disturbances and natural processes lead to short-term
dynamism, at the landscape level the processes are largely deterministic. This study is based on
the hypothesis that water explains the apparent tendency for localized climax communities to
remain geographically bound, rather than expanding when disturbance presents an opportunity to
do so.
An initial assumption that water would rarely be visible in historical aerials of the study
area proved to be incorrect. Although looking for visible water surfaces directly is simpler and
in line with approaches that have already been established in previous studies, this seemed
poorly suited for use at the Fort Ord study area because the region is so arid. As fact would have
it, numerous photographs show water surfaces in at least a few of the known vernal pool sites.
28

Surprisingly, water features were confirmed at five additional locations more than once (see
Figures 2 and 8). These PVP locations also exhibited some of the same vegetative patterns as
those visible around CVPs.
Such vegetation patterns were quantified in each analytical image, using Trimble
eCognition Developer 9.0 for segmentation and classification, and ArcMap 10.3 for pre- and
post-processing and analysis. The final datasets were also parametrically analyzed in IBM SPSS
Statistics 21. These patterns were used as a proxy for water in the environment because plants
must be specially adapted to survive in flooded soil. In theory, the way plant communities
develop in areas with steep hydrologic gradients should be notably different than the way the
same constituent species come together in the absence of standing water.
The postulate for this thesis is that relict or dormant vernal pool sites are still
discoverable between historic and contemporary imagery, because the vegetation patterns in and
immediately surrounding them appear strongly persistent. In this view, isolated lowland
meadows with margins that are fixed in time and space are that way because of seasonal
flooding. The persistence of vegetation patterns is not typical of a highly disturbed environment,
such as the study area. The history of continual disturbance (e.g. fire, mastication, explosions,
mowing, road and trail cuts, water diversions, and bulldozing) should be enough to shift the
whole landscape toward a more homogenous state, via both primary and secondary forms of
succession (Bauder 2000, 2005; Gopal 1986; TetraTech and Ecosystems West Consulting Group
2014; Van Dyke and Holl 2003; Van Dyke, Holl, and Griffin 2003). If meadow islands are
permanent fixtures, it could be because some underlying process or processes are helping
maintain the status quo. The presence of water at or just below the soil’s surface could surely
29

Figure 9: Oblique Camera Angle of CVP (ID # 80P07) from Google Earth Pro

account for such patterning, because it would restrict the nearby vegetation types to a much
narrower cohort.
CVP site locations (n = 15) were compiled using previously published studies and maps
as source information. An additional five meadows regularly developing water features were
also noted, bringing the CVP total to 20. PVPs (n < 51) were identified through a manual
examination of satellite imagery, aerial photographs, Google Earth Pro, and the digital USGS
topographical base layer available through ArcGIS Online. Data exploration was primarily
conducted with a dual approach that included the simultaneous use of ArcMap 10.3 and Google
Earth Pro. Preliminary utilization of Google Earth included visual scrutiny of the study area,
logging of historic inundation events using the timeline feature, and 2.5D visualization of the
study area. Google Earth is excellent for studying spatial relationships when used in a decision-
support role (Haas, Bartholomé, and Combal 2009; Jiang et al. 2012; Wang et al. 2011). For
instance, the camera can be panned, tilted, and rotated (Figure 9) to view different perspectives
30

Figure 10: Analytical Rasters, Side by Side After Pre-processing

of image rasters draped over a 2.5D fine-scale model of the Earth’s surface topography, and the
vertical axis can be stretched by up to a factor of three.
Raster images were pan-sharpened and clipped to the boundaries of the study area before
being uploaded into eCognition and subjected to OBIA. Figure 10 shows the three analytical
rasters in B&W (1941, far left) and true color (2001 and 2011, from center to far right), as they
appeared after the completion of post-processing. In a manner conceptually similar to that of
Zhang and Maxwell (2006), the primary Model Object (MO) was determined to be the outermost
edge of a study meadow, wherever chaparral and meadow habitats converged. Convergent
regions typically constituted the boundary between upland and lowland vegetation types.
This study demonstrates a method for tracking changes in upland/lowland boundaries to
allow a comparison between vernal pool-containing and other meadows. An ad hoc
segmentation algorithm based on the Fractal Net Evolution Approach (FNEA), developed by
Baatz and Schäpe (2000), was created for each set of rasters, after which supervised and visual
methods were used to classify the image segments (Baatz and Schäpe 2000; Blaschke 2010;
Halounova 2004; Mallinis, Pleniou, and Koutsias 2010). Objects were classified with input from
31

Figure 11: Site 80C03, the Single Study Area Vernal Pool Visible from Public Roads (Photo
Credits: C. Hanley)

various components of the available feature space (discussed later in Section 3.5). Finally, the
fully classified set of objects were exported to shapefiles and uploaded back into ArcMap. The
patch analyst extension (Rempel, Kaukinen, and Carr 2012) was used to calculate area, perimeter
(PERI), perimeter-to- area ratio (PAR), shape index (SI), and fractal dimension (FD). Patch
metrics were then used to compare short- and long-term changes in CVP and PVP boundaries.
Attempts were made to collect ground evidence to verify OBIA results, but only one
vernal meadow, 80C03 (Figure 11), is visible from a road or access point open to the public. A
request to access the restricted zone was denied due to unsafe conditions relating to unexploded
ordnance and munitions. Instead, a low-altitude flyover in a Piper PA-28 Cherokee was
conducted above the study area on 21 February, 2015. The route traversed was an approximate
north-south sawtooth pattern spanning the full extent of the study area. In flight, full HD video
32

of the ground was recorded from a GoPro HERO 3 digital camera mounted to the underside of
the right wing, as photographs were taken from the cockpit. Conditions were clear and bright,
with a flight time of 42 minutes at an altitude between 250 and 400 m above sea level.
3.3 Data Exploration
Persistent spatial patterns at the interface between meadow areas and chaparral habitats, mima
mounds, and the regular occurrence of oak stands nearby, are noticeable landscape features
common to many time-series images of the study area. Any location with a lasting meadow was
initially considered for inclusion as a study site. Unlike other areas of Fort Ord, where oak trees
mix with invasive annual grasses to create savanna ecosystems, the grassy lowland meadows in
the study area invariably have no growth of oak trees within their margins. Interestingly, many
meadows appear to coincide with small oak groves within the upland vegetation at their southern
and western peripheries. By contrast, oaks tend to appear as isolated individuals throughout
much of the remainder of the upland ecosystem.
A digital USGS topographic 10 m contour map, marked with depressions and drainage
areas, was used to locate and log the locations of low spots in the landscape (n = 69) in an
ArcGIS point feature class. The point file was converted to KML format, uploaded into Google
Earth Pro, and used to confirm the physical feasibility of indicated sites. From the combination
of topographical data and historical visual information, 54 PVPs were included in the first
iteration.
Upon review of all the available datasets, it was determined through water sightings that
the number of CVPs should be increased to 19, while the number of PVPs varied from image to
image. The discrepancy in PVP count was due to disqualifications mediated by extreme
disturbance. Namely, prescribed burns and mastications just prior to the 2001 and 2011 image
33

Table 1: Summary of Study Meadow Counts by Type and Year
Type 1941 2001 2011

CVP sites

18

17

18
PVP sites 41 38 7
Unique sites* 4 2 2
*Unique sites appear only in a single analytical image.

acquisition dates left more than 50% of the study area too denuded to perform accurate OBIA.
For the purposes of this research, a minimum regrowth time of at least eight years was needed to
see the re-establishment of a distinct upland/lowland boundary.
Recent disturbances were equal cause for jettisoning both CVP and PVP sites, except
where a visible water feature existed in the analytical image. Sites were also eliminated when
they did not persist from 1941 to 2001, or were not present in the 1941 image. The 2011 data
included two unique sites, both of which are vernal pools. However, only one of these is a
previously confirmed CVP. The study sites are summarized by year and type in Table 1.
3.4 Pre-Processing
Raw analytical data were projected and georeferenceed immediately upon acquisition. This
study utilized the projected coordinate system WGS_1984_UTM_Zone_10N. The multispectral
analytical images were pan-sharpened in ArcMap using a combination of their multispectral and
panchromatic bands, with Gramm-Schmidt transformations. The default channel weights were
chosen with respect to the sensor of acquisition.
34

`
Figure 12: 1941 B&W Raster Reassembled mosaic, Post-geometric corrections

Prior to export, attempts were made to correct alignment issues between the three
datasets. Since the panchromatic picture contained clear mosaic errors and the Ikonos data was
acquired off-nadir, the WorldView 2 imagery was selected as the base layer to which the others
were aligned. It was impossible to eliminate every alignment error, but most were significantly
reduced. The Ikonos data was corrected using georeferenced control points. Control points were
also used to align the 1941 raster, but the image contains some minor registration issues that are
unresolvable due to slight mosaic misalignments. All attempts were made to correct the errors
by first clipping the geotiff raster along the original mosaic lines, which are clearly visible in
Figure 12, then georeferencing the new pieces and reassembling the mosaic. However, some
distortion remained, in spite of these efforts.
35

Figure 13: NDVI Comparison, 2001 Ikonos (left) versus 2011 WV2 (right)

The final pre-processing step performed in ArcMap was the creation of a Normalized Difference
Vegetation Index (NDVI) band, which combines the Red and near-infrared (NIR) channels to
create a new channel that is useful in assessing the health and state of vegetation (Fallon 2013;
Highfield, Ward, and Laffan 2008). Red light is absorbed by chlorophyll during photosynthesis,
while near infrared wavelengths are reflected by healthy plant cells. Most commonly, the
brighter regions in Figure 13 coincide with areas of healthy flora. The NDVI band was
calculated according to the standard NDVI equation:

Image layers were mixed in eCognition using trial-and-error. The approach focused on
the creation of clear visual distinctions between the most important image objects, like meadow
margins, oak trees, and chaparral stands. The final weighted layer mixture was a variant of a
weighted bands CIR mix. The mix included red, green, blue, NIR, NDVI, and panchromatic
(1)
36

Figure 14: Novel False Color Composite of Ikonos 2001 Imagery

channels in order to enhance the separation between the feature spaces of different object classes,
with each color gun having up to four layers used in combination (Figure 14).
Halounova (2004) recognized the need to perform signature space enhancement, doing so
by using Gauss and median filters to reduce the amount of overlap in the spectral signatures of
classes. Similar separation was achieved here manually, via noise-reducing contrast stretches
produced with either inverse or negative gamma equalizations. An example of the novel false
color composite mix, as it appeared when applied to the 2001 imagery, is visible in Figure 14. In
the center of the inset is a CVP with a band of Dead Transition (orange) around the meadow
37

(purple). Oak trees are visible as blue objects, while the chaparral appears teal. Roads, bare
ground, and fire break are the darkest objects visible.
3.5 Segmentation and Object-Based Classification
Typical OBIA workflows employ the Fractal Net Evolution Approach, which Baatz and Schäpe
(2000) introduced commercially in eCognition software, followed by a Nearest Neighbor (NN)
classification (Figure 15). FNEA has since been referenced numerous times in the literature,
particularly in conjunction with multiresolution segmentation (Blaschke 2010; Burnett and
Blaschke 2003; Halounova 2004; Mallinis, Pleniou, and Koutsias 2010, Smeulders et al. 2000).
In segmentation, grayscale texture is used to create polygons known as segments that
represent homogenous regions. Multiresolution segmentation involves segmenting an image at
multiple resolutions, in order to fuse homogenous lower level sub-objects into higher level model
object classes. The result is a tessellation of polygons spanning the image extent, with each
texture-based polygon ideally containing adjacent pixel groupings that represent cohesive real-
world objects. Subsequently, classification of the segmented objects is performed. A NN
algorithm (Blaschke 2010; Halounova 2004; Smeulders et al. 2000; Yu and Zhang 2008) is the
most commonly applied classification technique. NN classifies horizontally across an image,
assigning segments as semantic objects according to their spectral signatures, densities, and
spatial contexts. The classification parameters one chooses to employ can be optimized in
eCognition, but it is often easier to make the choices on a trial-and-error basis instead. The
results of a simple linear FNEA and NN classification are depicted in Figure 15.
3.5.1 Fort Ord OBIA
Segmentation usually requires some supervision through the input of texture training areas,
manual adjustments to segment boundaries, or the creation of additional dissolves not triggered
38

Figure 15: Typical Linear FNEA and NN Workflow

by the chosen algorithm (Dissanska, Bernier, and Payette 2009; Halounova 2004). Though
object-based classification cannot be performed without first segmenting an image (Baatz and
Schäpe 2000), it is possible to iterate through cycles of merging, re-segmenting, and re-
classifying portions of an image to improve result accuracy. Several iterations may be necessary
because noise in the image data can negatively impact segmentation if it occurs on a similar
spatial scale as an object’s texture (Baatz and Schäpe 2000). Per Zhang and Maxwell (2006), the
segmentation of sub-objects can ultimately effect the drawn accuracy of the model objects. Thus
it is preferable to over-segment the image with smaller polygons and merge upwards, stopping
before the MO is merged with other objects.
The Fort Ord imagery was analyzed using a derivative of the FNEA approach that
included a top-down, bottom-up scheme. Model objects were segmented and classified
iteratively, going both forward and backwards as necessary. This multi-directional tactic
allowed the small number of PVP and CVP sub-objects to be selected and merged into model
39

objects and classified manually, ensuring that lowland features and their boundaries were
extracted with the highest precision possible. There was a conscious tendency to err on the side
of commission and occasionally include adjacent sub-objects whenever the original segmentation
was incorrect, as lowlands were later unclassified and re-segmented at a finer scale to improve
the accuracy of the results. Spectral- and shape-related thresholds were then utilized to classify
the remainder of the image because feature space optimization and the NN algorithms were
computationally intensive and time-consuming. The cost in time was doubly important since the
entire OBIA was constructed from scratch for each image, using trial-and-error.
To illustrate how thresholds were used, consider two equivalent road segments. Due to
any number of chromatic aberrations or variations in the objects themselves, each road segment
may attenuate differing proportions of a given wavelength of light. The reflected portion of said
wavelength gives the object brightness in that color. Thus one road segment may have a
brightness in the blue band, for instance, that overlaps the signature space of another object class
while the other road segment does not. Adding a second threshold for another band, such as red,
allows the first threshold to be adjusted while still excluding another object class that overlaps in
blue (provided there is signature space separation in the new band).
While NN classification would have been easier to execute on a conceptual level, and the
method described herein required more effort and direct input from the user, it also achieved
much more accurate classification results. The scheme was refined using incremental changes in
threshold values. Moreover, a shapefile of burn area polygons was created in ArcMap, exported
to eCognition, and used to classify burn areas in the 2011 image. Integration of the two data
streams significantly reduced the time required to perform the OBIA, as well as improved the
40

accuracy. Finally, the accuracies were assessed using error matrices created from Test and
Training Area (TTA) masks, which were constructed from a carefully selected group of samples.
3.5.2 Semantic Considerations
Prudence dictates that segmentation and OBIA results must be visually inspected to ensure that
classifications match the semantic knowledge base (Dissanska, Bernier, and Payette 2009).
However, experience has shown that the semantic structure of the study is equally, if not more,
important in producing accurate results. Choosing the optimal number of classes and naming
them aptly was not easy to accomplish. The imagery was exclusively comprised of wilderness or
natural landscapes, which increased the challenge because clear physical, as well as semantic,
distinctions between object classes were often hard to achieve. For example, it was necessary to
ask where the boundaries between transitional vegetation and lowland grasses were when a
gradient existed rather than a break. Likewise, how many oak trees constituted an oak grove, or
what was the appropriate classification of a study site that had water features but was not a
previously confirmed vernal pool? What if object classes existed in one image but not others?
These were some of the questions that needed answers.
Clearly, object class semantics are paramount to OBIA success. In this case, a structure
was required that could be applied equally to every image in the time-series sequence. The Fort
Ord impact area time-series presented many challenges because there was a high-degree of
variation inherent in several object classes. For example, the “Invasive” class (Figure 16)
included both iceplant (Carpobrotus edulis) and pampas grass (Cortaderia jubata), which differ
considerably in texture and appearance. Another class, “Upland,” likely encompassed anywhere
from one to several species of Arctostaphylos manzanita, coyote brush, Monterey ceanothus
(Ceanothus cuneatus var. rigidus), and chamise (Adenostoma fasciculatum). “Upland” probably
41

Figure 16: Iceplant and Pampas Grass Comprised the "Invasive" Class (Photo Credits: C. Hanley)

also included some unsegmented oak trees, small patches of bare ground, invasives, and the
occasional coastal sage scrub community. Overall, there are more than 1,000 plant species that
have been identified on the base, meaning that most categories had a catch-all component similar
to the invasive and upland classes (BLM 2010). Table 2 details the overarching semantic
structure employed during the analysis, as it was run in eCognition. Classes that related to
lowland structures, such as Live Transition, Dead Transition, Lowland, and Water were later
reclassified in ArcMap into a supra class, termed “Lowland Complex” (not described in Table 2).
Although it was at times difficult to differentiate between within-meadow classes, it was rarely
difficult to establish the lowland-upland boundary. Consequently, it was deemed appropriate to
analyze lowlands as complexes, rather than at the constituent level. However, OBIA and
classification required a more granular approach that necessitated the additional classes to reduce
the semantic gap.
42

Table 2: Object Class Names, Acronyms, and Descriptions
Class Name Description

Bare Ground/Dirt Track (BG)

Unpaved roads, tracks, paths, trails, and bare ground
Building Buildings or similar human structures
Dead Transition (DT) Band of grey (dead or senescent) upland vegetation
surrounding vernal meadows
Fire Break/Burn (FB) Any area that had been recently cleared of vegetation,
but was not bare ground (e.g. recent burns, fuel breaks,
youngest chaparral age classes)
Invasive (IV) Typically iceplant or pampas grass
Live Transition (LT) Areas where dead transition was reviving or chaparral
was recruiting into meadow areas
Lowland (LL) Local depression, grassy zone inside or adjacent to
Dead/Live transition where water features may or may
not develop
Oak Tree (OT) Individual oak trees and groves
Paved Road (PR) Any pavement, asphalt, or gravel surface
Upland (UP) Chaparral vegetation that had reached minimum age
class of 8 years
Upland Prairie Any grassland not located in a lowland depression
Water (W) Visible water features

3.6 Analyzing Change
43

From eCognition, the user is able to access and export over 100 spectral, geometric, or class-
related metrics. However, these data were not used because the vernal pool analysis was not
spectral in nature. More importantly, post-hoc corrections performed in ArcMap would have
rendered eCognition’s geometric values incorrect. The spatial metrics described below in this
section were therefore calculated using ArcMap after post-hoc corrections had been made.
The change analysis approach that is presented in this section is novel in the sense that,
unlike the other components of this study, it was not guided by previously published research. In
essence, the ad hoc development of the change analysis arose from the problem of quantifying
the spatial differences between an object and later versions of itself. Since the registration
between the images was inexact, and the average offsets were far greater than the images’ spatial
resolutions, performing an overlay or examining changes in centroid positions were not
considered appropriate analytical pathways. Instead, all lowland complexes were analyzed using
the Patch Analyst extension, which computed values for several key geometric properties. In
total, five initial metrics—area, perimeter, PAR, SI, and FD—were calculated for every meadow
site (ID codes shown in Figure 17) in each analytical image, with PAR, SI, and FD being unitless
numbers. A similar set of metrics was used by Backoulou et al. (2011) to evaluate island patches
of aphid-stressed plants in wheat fields.
Changes in the geometric properties of polygons must be evaluated with caution because
area and perimeter do not scale equally. Complicating the issue is the fact both segmentation
granularity and landscape composition can lead to discrepancies in object smoothness. Changes
in smoothness have a greater impact on perimeter than area. One way to handle the problem is
to calculate a shape index. The SI equation normalizes the perimeter-to-area ratio of a polygon
44

Figure 17: Lowland Complex Site ID Numbers

by comparing it to a simple shape, like a square, of equal size (Backoulou et al. 2011). SI also
makes the PAR metric redundant. Instead another useful metric is the fractal dimension which,
in simplest terms, is a measure of the degree of complexity of a shape’s outline (Backoulou et al.
2011). The study meadows tend to have amorphous outlines that are relatively compact and
simple, but the smoothness can exhibit some variance over time.
To compare images, differences in metric values for self-same sites were derived for the
1941–2001 and 2001–2011 intervals. The percent change in each category was calculated as
well. Finally, a new metric introduced for the first time here, the Persistence Index, was
determined for each site as follows:

(2)
45

where A
i
= the initial area, ΔA
t
= the area change at time t, SI
i
= the initial shape index, ΔSI
t
=
the change in the SI at time t, FD
i
= the initial fractal dimension, and ΔFD
t
= the change in the
FD at time t. Given that A
i
, SI
i
, and FD
i
cannot equal or sum to zero, the range is -∞ < PI < ∞.
PI is also unitless, and can be interpreted as the mean of the change ratio across all three
metrics. In theory, it can range to infinity in either direction. Shapes with PI values very close to
zero have undergone very little change since the previous measurement. The index gives the
relative degree of combined shape and size change of an object, either positive (i.e. expansion) or
negative (i.e. contraction).
PI can also be weighted by interval (WIPI) by dividing by the number of intervals,
allowing for groups with a variety of temporal ranges to be compared. For instance, 2011 WIPI
value distributions were compared to 2001 WIPI values for self-same CVP locations, with an
interval length of 10 years:

where C = the number of intervals and C ≠ 0. Thus WIPI for the 7 decadal intervals from 1941–
2001 is given by:

Likewise, WIPI for the single decadal interval from 2001–2011 is given by:

WIPI normalizes the persistence index for the effects of time, as the absolute value of a given
patch’s PI would be expected to increase as time passes.
(3)
(4)
(5)
46

Patch and class level analytics were performed in both ArcMap and SPSS, including
exploratory regression, dependent, and independent t-tests. For comparative purposes, meadows
were placed into one of the two previously described nominal categories, PVP or CVP. The
CVP category comprised both previously confirmed vernal pool meadows and those discovered
during this research. PVP were defined as any other persistent lowland meadow study site
within the study area. For the 1941–2011 interval, it was possible to compare the groups of
CVPs and PVPs because sufficient samples existed. However, between 2003 and 2011,
approximately half of the study area was burned or masticated. As a consequence, PVP sites
were reduced from 38 to seven. CVP sites suffered less, being reduced from 16 to 15. Thus
WIPI was developed to facilitate a comparison between self-same CVP sites over time. Whereas
the 70-year interval allowed for the comparison of two different groups—CVP and PVP—using
the Student’s (independent samples) t-test, the 10-year interval only allowed for CVP to be
compared with themselves using the Paired (dependent) samples t-test.
Spatial analytics were attempted in addition to parametric statistics, but were ultimately
dropped when an adequate model could not be constructed using exploratory regression. The
datasets were not well-suited to the requirements of a regression analysis. Redundancy in some
of the metrics manifested itself as multi-collinearity, while spatial autocorrelation of the residuals
and small sample sizes further confounded the test. Moreover, a regression requires a continuous
dependent variable, which the datasets in question did not adequately possess. Conversely, the
data were amenable to parametric analysis, allowing the data collected from this OBIA to be
analyzed with only the occasional logarithmic transformation or removal of an outlier.

47

CHAPTER 4: RESULTS

The results presented here are broken up into three parts. In Section 4.1, the accuracy reports for
the classification procedures are given, while Section 4.2 contains the key statistical outputs from
SPSS. The last section focuses on the observations resulting from the low altitude flyover.
4.1 Classification Results
Classification accuracy was assessed in eCognition using error matrices generated from TTA
masks. Masks were created from MO level segments, using sample sets carefully chosen to
represent the within-class variation inherent in the data. Error, or confusion, matrices
(contingency tables) like the ones shown later in this section provide estimates of Type I and II
errors on a per class basis. Classification errors can manifest in two basic ways. The former,
Type I, happen when the correct action was not taken. The latter, Type II, arise when the
incorrect action was taken.
Type I errors are errors of omission that occur when a pixel should have been included in
a given class, but was mistakenly omitted instead. Omission errors are reflected in the columns
of a contingency table. The sum of each column is actually the total number of pixels that
should belong to that class in the real world, according to ground evidence. The Producer’s
accuracy is another way of describing Type I error, in that it gives the estimated proportion of the
true number of pixels that were classified correctly into a given class. For example, consider a
situation where an image contains 100 true pixels of grass; if 75 are classified as grass and 25
48

errantly classed into another category, the omission error would be 25% and the Producer’s
accuracy would be 0.75 (75%).
By comparison, Type II errors are related to User’s accuracy and commission errors, and
are interpreted from a table’s rows. Type II commission errors are those that result from a pixel
being classified into a class to which it does not belong. The way to distinguish between Type I
and II errors is to recall that, for a given class, some proportion of pixels were not included that
should have been (Type I) while other pixels were included that should not have been (Type II).
Type II errors are caused when pixels are falsely counted in the class being examined. In the
same manner that Producer’s accuracy relates to Type I error, the User’s accuracy is an alternate
expression of Type II error. In essence, the User’s accuracy gives the relative likelihood that
someone visiting a pixel’s real-world location would actually find the expected class of object.
4.1.1 1941 Panchromatic Classification and Accuracy
Classifying the 1941 B&W image required the fewest object classes, since its lower resolution
and lack of color made classification to a higher granularity impossible. Only classes that were
easily and unmistakably identifiable were included in the image semantics. The object classes
included were Upland (UL), Fire Break/Burn (FB), Lowland (LL), Bare Ground/Dirt Track
(BG), Oak Tree (OT), and Live Transition (LT). With this semantic scheme, it is possible that
undetectable errors may have occurred if Dead Transition (DT) and Invasive (IV) objects were
present in the image, as they were not incorporated in the class structure. Both classes would
have been counted in the LT class by mistake. Such a determination would be impossible to
make, however, without ground evidence—which does not exist for the 1941 dataset.
Figure 18 presents the classification results and Table 3 follows with the error matrix for
the 1941 panchromatic image. A visual examination of the output map shows that the
49

Figure 18: 1941 Panchromatic Image Classification Results
Table 3: Error Matrix (Contingency Table) for 1941 Panchromatic Classification
CLASS UL FB LL BG OT LT SUM
UL 8,904,365 0 0 0 0 0 8,904,365
FB 0 149,975 0 0 0 0 149,975
LL 0 0 598,688 0 0 0 598,688
BG 0 0 0 72,254 0 0 72,254
OT 0 0 0 0 37,731 0 37,731
LT 0 0 0 0 0 325,437 325,437
SUM 8,904,365 149,975 598,688 72,254 37,731 325,437 10,088,450

Prod. 1 1 1 1 1 1 1
User 1 1 1 1 1 1 1
Overall 1

50

classification proceeded with a high degree of accuracy. On the other hand, the error matrix
overestimated the accuracy in calculating 0% types I and II errors for each object class. Since
perfect classification performance is highly improbable, the accuracy results shown in Table 3
are dubious. The real accuracy could not have been 100%, but the reason for the overestimate is
unclear. Nonetheless, the vector output results easily pass inspection when scrutinized with the
human eye. It is reasonable to expect that the true accuracy may still be on par with the 2001
and 2011 results, which both exceeded 98% overall.
4.1.2 2001 Ikonos Classification and Accuracy
The 2001 Ikonos imagery was significantly more complex than the B&W from 1941. An
additional six classes were needed, doubling the total from the first OBIA (Figure 19, next page).
The combination of multispectral bands and increased spatial resolution led to an improved
ability to distinguish between classes. However, the gain in sensitivity is also what necessitated
the expansion of the semantic structure. The original MO classes were maintained, and six new
classes were added, including Building, Dead Transition, Invasive, Paved Road (PR), Upland
Prairie, and Water. Only three of the newly added classes were considered during the accuracy
assessment. Building, upland prairie, and water objects were not evaluated in the error matrix
because their occurrence rates were disproportionately low. In all, there were five classified
building objects and one each of upland prairie and water, out of a total of 10,643 object
segments. Since these three classes together occupied less than 0.1 percent of the total image
area, their bearings on the overall accuracy of the OBIA were negligible.
The accuracy estimates produced using the remaining nine classes were very high (Table
4, next page), equaling the accomplishments of Frohn et al. (2011). Visually, the classification

51

Figure 19: 2001 Ikonos Image Classification Results
Table 4: Error Matrix (Contingency Table) for 2001 Ikonos Classification
CLASS LL UL DT BG FB OT PR LT IV SUM
LL 68,186 0 0 0 0 0 0 0 0 68186
UL 0 1,515,886 0 1,928 14,210 1,145 0 0 0 1533169
DT 0 0 15,530 0 0 0 0 0 0 15530
BG 0 0 0 53,842 13,113 0 0 0 0 66955
FB 0 0 0 2,328 87,514 0 0 0 0 89842
OT 0 100 0 0 0 159,154 0 0 0 159254
PR 0 0 0 0 0 0 49,956 0 0 49956
LT 0 0 0 0 0 0 0 21,648 0 21648
IV 0 0 0 0 0 0 0 0 9,285 9285
SUM 68,186 1,515,986 15,530 58,098 114,837 160,299 49,956 21,648 9,285 201,3825

Producer 1 0.9999 1 0.9267 0.7621 0.9929 1 1 1
User 1 0.9887 1 0.8042 0.9741 0.9994 1 1 1
Overall 0.9837

52

results showed an exceptional degree of fidelity to the original image. Like the 1941 results,
there were some ostensibly error-free classes, such as Lowland, Paved Road, Live Transition,
and Invasive. However, these results are somewhat more believable because those classes were
either classified manually or, in the case of Paved Road, had a sizable separation in feature
space. The most confusion transpired amongst the Upland, Oak Tree, Bare Ground/Dirt Track,
and Fire Break/Burn classes. BG and FB pixels were most often confused for one another, with
each class lending to the majority of Type I and II errors in the other. Consequently, the BG
User’s accuracy of 80.42% and the FB Producer’s accuracy of 76.21% were significantly lower
than those of the other classes. Upland, for instance, demonstrated high User’s and Producer’s
accuracies (98.87% and 99.99% respectively), in spite of a slight propensity toward commission
errors involving BG, FB, and OT pixels. Oak tree objects and their shadows were known
confounding factors as they were too small and numerous to reliably segment as individuals.
Not surprisingly, it was not unheard of for the occasional oak tree to go entirely unsegmented.
Typically these oaks were encompassed within a larger upland object, which is reflected as Type
I error for the Oak Tree class and Type II error for the Upland class.
4.1.3 2011 WorldView 2 Classification and Accuracy
The 2011 semantic structure had an extra class included, that of Athletic Field, to accommodate
the addition of two small sod grass fields in an athletic compound at the southern tip of the study
area. As there were numerous water features visible in this image (Figure 20), the Water class
was incorporated into the classification accuracy assessment (Table 5). The Athletic Field,
Building, and Upland Prairie classes were ignored for reasons similar to that of the 2001
evaluation. Altogether, 10 classes were included in the error matrix for the 2011 dataset, with
nine being common to the 2001 image and six common to all three years.

53

Figure 20: 2011 WV2 Image Classification Results
Table 5: Error Matrix (Contingency Table) for 2011 WV2 Classification
CLASS LL DT BG UL FB PR OT W LT IV SUM
LL 235,207 0 0 0 0 0 0 0 0 0 235,207
DT 0 38,977 0 0 0 0 0 0 0 0 38,977
BG 0 0 170,946 0 25,099 0 0 0 0 0 196,045
UL 0 0 0 1,556,450 0 0 0 0 0 0 1,556,450
FB 0 0 0 0 3,209,234 0 0 0 0 0 3,209,234
PR 0 0 0 0 0 117,178 0 0 0 0 117,178
OT 0 0 0 0 0 0 168,586 0 0 0 168,586
W 0 0 0 0 0 0 0 106,129 0 0 106,129
LT 0 0 0 0 0 0 0 0 33,479 0 33,479
IV 0 0 0 0 0 0 0 0 0 44,999 44,999
SUM 23,5207 38,977 170,946 1,556,450 3,234,333 117,178 168,586 106,129 33,479 44,999 5,706,284

Prod. 1 1 1 1 0.9922 1 1 1 1 1
User 1 1 0.8720 1 1 1 1 1 1 1
Overall 0.9956

54

Figure 21: Landscape Conversion to Disturbance, from 2001 to 2011

The WV2 imagery presented a unique set of challenges for OBIA because widespread
disturbance effects were visible throughout the study area (Figure 21). The disturbances were
due to recent prescribed burns and manifested in a gradient of immature upland vegetation. The
most recent and severe disturbances were concentrated at the northeast corner of the study area.
Within the disturbed zone, immature upland chaparral graded from youngest to oldest in a
westerly direction, echoing the chronology of the previous decade’s land management practices.
The total area of Fire Break/Burn class grew from an estimated 155.62 ha (384.54 acres) in 2001
to 1,350.32 ha (3,336.73 acres) in 2011, nearly an order of magnitude increase. In 2011, Fire
Break/Burn coverage increased to about 52% of the study area, up from a modest 6% in 2001.
55

As so much of the study area had been mowed, masticated, or burned in the 10-year
interval between datasets, the classification scheme was necessarily arbitrary. The most
reasonable course of action was to set a criteria for minimum growth years, and place any areas
that failed to reach the age-restriction into the Fire Break/Burn category. Doing so meant that, in
2011, FB was a class that included upland, lowland, invasive, and transitional vegetation.
Segments not already classified as Oak Tree, Bare Ground/Dirt Track, Paved Road, or Water
were classified into the FB class wherever they intersected a polygon shapefile of relevant
controlled burn sites, or otherwise contained indicative visual cues (e.g. an obvious recent
vegetation disturbance).
While painstaking care was applied to ensure the most accurate OBIA possible, the
arbitrary nature of the FB class seems to have led to another overestimate of the procedure’s
performance. Like the results from 2001, the 2011 error matrix showed a similar, but less
extreme tendency toward confusing FB with BG. The other classes again exhibited
overestimated accuracies of 100%. The estimated overall accuracy of 99.56% is therefore
believed to be higher than the true value as well.
4.2 Statistical Results
The Patch Analyst extension was used to derive the base geometry metrics described in Chapter
3, Section 6 (Area, Perimeter, PAR, SI, and FD), which were then used calculate PI, WIPI,
percent area change, and percent perimeter change. Metrics were computed per lowland
complex, for each year of analysis.
Since so many PVP areas were eliminated from the 2011 classification, the full short- and
long-term analyses were not comparable. Therefore, the statistical analysis for each was
conducted differently. For the 70-year interval, sufficient PVPs (n = 26-31, depending on
56

outliers) and CVPs (n = 16) were available to be able to compare the two as separate groups
through independent samples t-tests. At the 10-year scale, there were not enough PVPs (n = 5)
to enable their inclusion. Instead, 14 self-same CVPs were compared using dependent samples t-
tests. Two sites, 70C29 and 80C11, presented water features but were eliminated because they
were within burn areas. In this manner, a short- and long-term comparison of geometric changes
in CVP sites was achieved. For PVP sites, only a comparison between these and CVP sites at the
70-year scale was possible. The results are summarized below, and a series of maps
documenting more complete results are included in the appendices at the end of this document.
4.2.1 70-Year CVP/PVP Comparison
To evaluate CVP and PVP at the 70-year time scale, the two populations were compared using
independent samples t-tests. Metric values measured in 1941 and again in 2001 were used to
calculate the change in each for every site. The analysis included the following six metrics:
1) Change in Fractal Dimension (Δ_FD)
2) Change in Shape Index (Δ_SI)
3) Change in Perimeter-to-Area Ratio (Δ_PAR)
4) Percent Area Change (%Δ_AREA)
5) Percent Perimeter Change (%Δ_PERI)
6) Log base 10 of Persistence Index (Log(PI))
Area change and perimeter change were expressed in terms of a percentage to normalize
them, since they do not scale equally. Persistence indices were transformed into normal
distributions by a base 10 logarithmic transformation with a positive shift of 0.25. Table 6
provides the full set of group statistics. Table 7 summarizes the t-test results, including Levene’s
Test for Equality of Variances. The remaining graphical outputs, such as the box and whisker
57

Table 6: Group Statistics for 70-Year CVP/PVP Comparison
Metric Type N Mean Std. Deviation Std. Error Mean
Δ_FD
PVP 31 .0342 .1101 .0198
CVP 17 .0546 .0886 .0215
Δ_SI
PVP 31 .9358 1.6836 .3024
CVP 17 1.0426 1.4661 .3556
Δ_PAR
PVP 28 .0150 .0792 .0150
CVP 16 .0168 .0418 .0105
%Δ_AREA
PVP 26 22.7111 50.4222 9.8886
CVP 17 33.5210 45.2974 10.9862
%Δ_PERI
PVP 27 40.1622 53.0898 10.2171
CVP 17 65.7299 67.5221 16.3765
Log(PI)
PVP 29 -.3612 .2871 .0533
CVP 16 -.2874 .1879 .0470

Table 7: Independent T-test Results for 70-Year CVP/PVP Comparison

Equal
Variances
Levene's
Test for
Equality of
Variances
t-test for Equality of Means

95% Confidence
Interval of the
Difference
F Sig. t df
Sig. (2-
tailed)
Mean
Diff.
Std.
Error
Diff. Lower Upper
Δ_FD
Assumed .725 .399 -.655 46 .516 -.0204 .0311 -.0830 .0423
Not

-.698 39.477 .489 -.0204 .0292 -.0794 .0386
Δ_SI
Assumed .420 .520 -.219 46 .827 -.1067 .4863 -1.0856 .8721
Not

-.229 37.148 .820 -.1067 .4668 -1.0524 .8389
Δ_PAR
Assumed 2.298 .137 -.081 42 .936 -.0017 .0212 -.0449 .0414
Not

-.095 41.848 .925 -.0017 .0183 -.0386 .0351
%Δ_AREA
Assumed .039 .844 -.715 41 .479 -10.8099 15.1233 -41.3520 19.7322
Not

-.731 36.919 .469 -10.8099 14.7811 -40.7616 19.1418
%Δ_PERI
Assumed 1.524 .224 -1.400 42 .169 -25.5676 18.2689 -62.4358 11.3005
Not

-1.325 28.246 .196 -25.5676 19.3023 -65.0911 13.9559
Log(PI)
Assumed 3.214 .080 -.923 43 .361 -.0738 .0800 -.2351 .0875
Not

-1.039 41.572 .305 -.0738 .0711 -.2173 .0696

58

Figure 22: Box and Whisker Plots with Outliers, Comparing PVP to CVP, for Each Analyzed
Metric

plots shown in Figure 22, are included in the appendices and were used to help ensure that each
population met the assumptions of normality. The boxplots show the range of values for PVPs
(leftmost box in each plot) and CVPs (rightmost box), for each metric analyzed in the study. The
horizontal black bars represent, in order from lowest to highest, the minimum observed value,
the first quartile, the median, the third quartile, and the maximum observation. The box itself
highlights the interquartile range, while the small black circles depict outliers. Extreme outliers
were removed prior to the final analysis.
The most important statistics provided are the means, standard deviations, and standard
error means from Table 6, along with the p-values given by the Levene’s Significance and the t-
test Significance (2-tailed) in Table 7. A significant Levene’s statistic (p ≤ 0.05) would indicate
that the assumption of equal variances was violated, necessitating the use of lower rows of values
59

(labeled “Not”). However, equal variances were met across all categories, so the upper rows
(labeled “Assumed”) were used. For values of Significance (2-tailed), a p-value of 0.05 or less
would have indicated a significant difference in the PVP and CVP means.
There were no data categories with a statistically significant 2-tailed p-value at a 95%
confidence interval. In fact, no metric would have passed at a 90% confidence interval either.
The means, standard deviations, and standard error means appear to be comparable between site
types as well, particularly for the transformed PI and the changes in FD, SI, and PAR. The
percent changes in area and perimeter were non-significant as well, but did display greater
relative discrepancies. Nonetheless, by the metrics and methods applied here, the CVP and PVP
populations are indistinguishable from the vantage point of geometric change.
4.2.2 Short- and Long-Term Changes in Self-Same CVP Sites
CVP were compared to themselves in order to check for geometric instabilities that might have
been discoverable at a finer temporal scale. The reasoning was that environmental variability
might cause lowland boundaries to exhibit short-term spatial dynamism, with them eventually
converging to a mean location as a result of geographic restrictions. Water could be the cause of
such a restriction, but drought and climate instability might lead to variation only evident on
shorter time scales.
Since the 2011 image contained so few PVP sites, a comparison using these data would
have led to a dataset that was too small to be analyzed. To ensure a larger sample size, PVPs
were eliminated from the short- and long-term comparison. The remaining CVPs were matched
against their self-same sites, and compared at 10- and 70-year intervals. The change in approach
necessitated a shift from independent to dependent sample t-tests. Logarithmic transformations
similar to those previously described were used to create normal distributions, and the Weighted
60

Intervals Persistence Index was substituted for the non-weighted PI to normalize for interval
length:
1) Log base 10 of change in FD, 10-year minus 70-year (Log(ΔFD10) – Log(ΔFD70))
2) Log base 10 of change in SI, 10-year minus 70-year (Log(ΔSI10) – Log(ΔSI70))
3) Log base 10 of change in PAR, 10-year minus 70-year (Log(ΔPAR10) – Log(ΔPAR70))
4) Percent Area Change, 10-year minus 70-year (%ΔAREA10 – %ΔAREA70)
5) Percent Perimeter Change, 10-year minus 70-year (%ΔPERI10 – %ΔPERI70)
6) Weighted Intervals Persistence Index, 10-year minus 70-year (WIPI_10 – WIPI_70).
The results of this part of the analysis are harder to interpret, mainly because they seem to
conflict with one another. The group statistics given in Table 8 show that, on average, CVP area
decreased over the shorter but increased over the longer of the two intervals. Interestingly,
perimeter change increased over both intervals instead of similarly decreasing over the short-
term. Percent area and percent perimeter change are cumulative, so the absolute value of either
mean would be expected to greater over the longer interval. Area should change the most, as it
scales faster in both directions when a polygon changes size. Accordingly, it is unclear how
mean perimeter grew while mean area shrunk over the short interval. WIPI, on the other hand,
was stable across both time periods.
The primary values of interest for the results of the dependent sample t-test presented in
Table 9 are found in the significance (2-tailed) column. Like the previous t-tests, a 95%
confidence interval for detecting differences in means is indicated by p ≤ 0.05. Not surprisingly,
percent area and percent perimeter changes showed significant differences between the two time
intervals (p = 0.003 and p = 0.018, respectively). Even slow growth (positive or negative) would
eventually lead to the same kinds of results. By contrast, low, stable WIPI values over both
61

Table 8: Group Statistics for Short- and Long-term Self-Same CVP Comparisons
Pair Mean N Std. Deviation Std. Error Mean
Log(ΔFD10) – Log(ΔFD70)
-.0775 14 .0239 .0064
-.067 14 .0476 .0127
Log(ΔSI10) – Log(ΔSI70)
.2076 14 .6954 .1858
.4160 14 .2363 .0632
Log(ΔPAR10) – Log(ΔPAR70)
-1.0979 14 .2253 .0602
-1.2209 14 .5270 .1409
%ΔAREA10 – %ΔAREA70
-6.5806 14 15.6050 4.1706
41.3330 14 43.3527 11.5865
%ΔPERI10 – %ΔPERI70
15.6284 14 33.6389 8.9904
77.4321 14 65.7887 17.5828
WIPI_10 – WIPI_70
.0501 14 .1330 .0356
.0474 14 .0333 .0090

Table 9: Dependent Samples T-test Results for Short- and Long-Term CVP Comparisons

Paired Differences
t df
Sig. (2-
tailed) Mean
Std.
Deviation

95% Confidence
Interval of the
Difference
Std.
Error
Mean Lower Upper
Log(ΔFD10) –
Log(ΔFD70)
-.670 13 .514 -.0106 .0591 .0158 -.0447 .0235
Log(ΔSI10) –
Log(ΔSI70)
-.935 13 .367 -.2084 .8343 .2230 -.6901 .2733
Log(ΔPAR10) –
Log(ΔPAR70)
.760 13 .461 .1229 .6051 .1617 -.2265 .47233
%ΔAREA10 –
%ΔAREA70
-3.693 13 .003 -47.9136 48.5477 12.9749 -75.9442 -19.8830
%ΔPERI10 –
%ΔPERI70
-2.714 13 .018 -61.8037 85.2204 22.7761 -111.0085 -12.5989
WIPI_10 –
WIPI_70
.066 13 .949 .0027 .1532 .0410 -.0858 .0912

62

intervals are likely attributable to an overall stability in size and shape. Put another way, a small
WIPI means that the spatial appearance of a shape has changed little over time. When the 10-
year and 70-year WIPI means are so similar (p = 0.949), and so small (

10
= 0.0501,

70
=
0.0474), the interpretation is that size and shape are temporally fractal. That is, regardless of the
length of the time interval, the sizes and shapes of the polygons appear unchanged.
The most vexing results are the non-significant differences between intervals in the FD,
SI, and PAR change tests. Since the datasets were log transformed, their means are not useful in
understanding the directionality of the changes. However, since each is also a cumulative
metric, a longer interval should result in more change. Counterintuitively, similarity in FD, SI,
and PAR changes over both intervals means that change was happening more quickly during the
shorter time period. Such a result supports the idea that short-term dynamics are juxtaposed with
long-term stability.
4.2.3 Performance of PI and WIPI
The Persistence Index equation was created to provide a useful description of wholesale changes
in polygon dimensions, including area, shape index, and fractal dimension. The Weighted
Intervals Persistence Index makes it possible to compare PI values across unequal time intervals.
The Persistence Index equation minimizes the effect of minor changes so that major changes
stand out more. PI can be likened to the human eye, which is able to gloss over minor
differences in a polygon to see that the underlying shape is still the same.
In the first analysis, the performance of PI was in keeping with the other results. The
logarithmic transformation came at the price of lost directional knowledge because all the PI
values were first shifted to ensure none were less than zero. The shift was necessary to perform
the transformation, but it also filtered out information about whether change was in the form of
63

expansion or contraction. Still, the finding that the PVP and CVP groups were not different was
consistent with the findings of the analysis at large. Conversely, the results of the second
analysis were cryptic. With the other results pointing to an overall increase in change over the
short-term, the absolute value of WIPI would be expected to increase too. Yet the differences
between the two intervals was non-significant. The 10- and 70-year WIPI values are mapped in
Figures 23 and 24. The WIPI results can easily be reconciled with the area and perimeter results
because the latter two are cumulative while the former is not. Since WIPI is normalized for the
effects of time, it effectively tracks the change rate while area and perimeter changes are
considered in gross. Hence a constant rate of change would give both a constant WIPI and
changes in area and perimeter.
Reconciliation between WIPI and the other metrics is challenged by the fact that the
index describes a slow, constant rate of change. FD, SI, and PAR changes indicate a different
trend. They instead show that change appears to have sped up over the 10-year period. The
reasons for the discrepancy in the results are unknown. One possible explanation is that WIPI is
less sensitive to small changes in metric values that are more easily discerned by one of the more
focused metrics.
4.3 Flyover Observations
Even the highest resolution aerial photography cannot substitute for the understanding an in situ
visit can impart. Despite the fact that a request for ground access to the study sites was denied,
the airspace above Fort Ord is not restricted and low-altitude flyovers are permitted. The
primary goal in performing a flyover was to put real eyes on the study sites. Since the only
vernal pool visible from a public road was inundated at the time of this study, there seemed to be
64

Figure 23: 70-Year WIPI for Self-Same CVP Sites

Figure 24: 10-Year WIPI for Self-Same CVP Sites
65

Figure 25: Erosion Scars Assisted in the Confirmation of Site ID# 80P07 as a CVP (Photo Credits:
C. Hanley)

a reasonable possibility of other sites being inundated as well. The flyover was the only
opportunity to collect ground evidence to verify classification results.
During the flight, there were seven CVP sites noted with visible water surfaces, along
with an eighth site possessing a large erosion gully leading directly into its center. Of the seven
sites with water in them, two were CVP confirmed for the first time during this study. The
eighth site was a PVP that had a detectable water feature in 2011 to go along with the erosion
drainage evident during the flyover, the combination of which was compelling enough to cause
the type to be switched to CVP prior to the final SPSS analysis. Figure 25 shows the erosion
feature (circled in red) in the eighth site (ID# 80P07) from opposite angles, with the aircraft wing
visible in the foreground of the photo on the left.
Overall, the flyover confirmed several ideas that were considered during data exploration.
First, one of the justifications used for reassigning PVPs as CVPs was the frequency of water
sightings. The observation was made that certain PVPs had confirmed water features more often
66

Figure 26: Inundation Frequencies, Compiled from a Review of 21 Historic Aerial Images

than some known vernal pools. By extension, it was also obvious that inundation did not occur
equally across the landscape, nor predictably with time. In other words, there appeared to be no
relationship between the filling of one pool and the filling of another in a given photo. The
flyover substantiated the idea that some pools fill more often than others, or at least hold water
for longer. Notably, when not inundated, CVP look superficially similar to PVP. Total
inundation frequencies were calculated after the flyover for 21 historic and contemporary images
(including flyover data), the results of which are presented in Figure 26.
A few PVPs, particularly toward the western margins, appeared to be converting into
uplands. Most, however, were verdant grasslands that visibly stood out amongst the various age
classes of chaparral. By the time of the flyover, around 75% of the study area’s mature chaparral
67

Figure 27: CVP Site 80C01 Timeline Showing Persistent Dead Transition (DT) Vegetation Band
(Data Sources: Google Earth 2015; DigitalGlobe Foundation 2014; US Army BRAC 2014)

had been removed. After burning or mastication, the only vegetation left standing were oak
trees. Based on photographic evidence, it will take an estimated 8–10 years for the study area to
return to a measurable state again, and 2–3 times as long as that for the chaparral to reach full
maturity, when it can become decadent and exceed 5 m in height.
Unfortunately, the flyover did not settle matters of identifying invasive and dead
transition species. The expedition did nothing to refute the assertion that the invasive class
contained only pampas grass and ice plant, but neither did it provide confirmation. Dead
transition vegetation, on the other hand, was given this name for the lack of a better description.
The only new information provided by the flight was that dead transition appears to come from a
taller species than first anticipated. The consistent shape and size of the DT band around some
CVP is so striking that it almost looks artificial (Figure 27). At the risk of being speculative, it
seems at least plausible that DT is a remnant of very old chaparral that was never removed
during previous land management regimes. A mechanical clearing of the surrounding brush
68

Figure 28: Major Erosion Scars Divert Surface Waters within the Study Area (Photo Credits: C.
Hanley)

could account for the smooth shape and consistent size, with the cleared areas slowly maturing
into healthy chaparral again while the remaining plants buffering the lowland eventually senesce
and die.
Certainly, vegetation removal is not the only anthropogenic scar visible on the landscape.
The flyover revealed that wide, deep erosion gullies now crisscross the study area. These gullies,
visible as ragged gashes in Figure 28, clearly alter the way that water flows through the
environment, and could be responsible for changing or stopping inundation regimes in nearby
lowlands. There also exist many leftover vehicles, buildings, bunkers, and rifle pits spread
throughout study area that are sources of pollution, as well as physical disturbances.

69

CHAPTER 5: DISCUSSION AND CONCLUSIONS

In the Fort Ord study area, lowland complexes are akin to Burnett and Blaschke’s “island
patch[es] in a sea matrix” (2003, p. 238). For them to exist, they must combat the slow
encroachment of maritime chaparral. This thesis pursued the explanation as to why lowland
meadows, in the absence of a known vernal pool, are able to persist and not be swallowed up in
the sea of upland vegetation by which they are surrounded. The evidence put forth here indicates
that the upland-lowland border interactions are spatially stagnant. That is, when not artificially
disturbed, they change very little whether measured over short or long time intervals. Lowland
grasses do not invade manzanita or chamise chaparral, likely because the shrubs are allelopathic
and poison the surrounding soils (Van Dyke, Holl, and Griffin 2001). On the other hand,
persistent annual hydric cycles are the most plausible explanation as to why chaparral is repelled
from the vernal and non-vernal pool lowlands with equal vigor.
According to the data presented here, the border interactions that vernal and non-vernal
pool lowlands have with upland chaparral are similar, in terms of their spatial context. The
empirical numbers confirm the observation that CVPs and PVPs are more than just superficially
similar. PVPs respond in similar ways to the same environmental pressures as CVPs. That is
not to suggest that all PVPs in the study are really vernal pools. Rather, vernal pools are
probably common amongst the PVP group, in order for them to exhibit such holistic similarities
to CVPs. To be sure, directly spotting water or visiting a site in person are the best, if not the
only, ways to confirm that a PVP site is actually a vernal pool. Nonetheless, the evidence
suggests that most of the PVP sites must at least exhibit damp soil conditions that favor the
70

development of grasslands, while simultaneously repelling upland vegetation. A few may
actually display surface pools of water under the correct climatic conditions, although this is
speculation based on suggestive but ultimately inconclusive visual evidence. Since the region is
currently in the midst of a severe and prolonged drought, the likelihood of ascertaining the truth
in the near future is low. If disturbances have re-routed water away from areas that were once
vernal pools, they may now be relict or dormant instead. Dormancy is nevertheless reversible, so
it is worthwhile to examine the feasibility of restoring the ecological potential of such areas.
In the field of ecology, it has been said that when the same phenomenon occurs in two
places, the root cause is likely the same. Similarly in geography, the closer two instances of a
phenomenon are in space, the more similar they are likely to be in character. Bearing these
adages in mind, a collective of logic, observation, and empirical evidence stand in support of the
assertions made in the closing paragraphs of this document. What follows now is the final
discussion of the meaning behind the results this study produced. Though there are multiple
possible interpretations, the principle of parsimony (Occam’s razor) instructs that the simplest,
most succinct explanation is the best. In this case, the best explanation is that water is the
driving environmental force behind lowland vegetation patterns in Fort Ord’s MRS-BLM zone.
The final arguments as to why are presented below.
5.1 On OBIA Accuracy
Converting raw image data into useful intelligence requires considerable domain knowledge of
the subject matter being interpreted. Image interpretation demands much technical skill of the
analyst as well. Naturally, the results of an image analysis need to somehow be vetted post-hoc
to judge their quality. Where the object-based extraction algorithms used in this study differ
from those typically described in the literature is in the use, or lack thereof, of the nearest
71

neighbor approach. NN uses object class samples, select feature space components, and a
horizontally moving window to evaluate object pixels and their spatial relationships. To reduce
the semantic gap, the methods used to segment and classify the Fort Ord imagery employed a
higher-degree of supervision than is typical of NN. Consequently, the manual adjustments made
to segments and class assignments somewhat negated the efficacy of generating an error matrix,
which is better suited to estimate the performance of automatic classification algorithms.
In terms of the accuracy assessment itself, experience has proven that the cohort of
samples inputted into a TTA mask can have an effect on the subsequent error matrix. Accuracy
is also tied in with segmentation performance and semantic structure, and so is subjective in
several ways. In the cases where MO classes were infrequently represented, occupied a small
percentage of the total area, or incurred unusual amounts of manual intervention, the trend was to
exhibit 100% accuracy. In short, the overestimation of accuracy was likely to occur whenever
the sample size was small or the classification not fully-automated. Water features detected in
the 2011 data illustrate this principle well, since they were rare and isolated to a few known
areas. There was little likelihood of a Type I or II error occurring in the water class because
water features were classified manually, and no other objects had mixed water features in them.
The segmentation also fared well at determining shorelines—the areas that were most susceptible
to errors. It is easy to understand, then, why the Water class had an estimated 100% accuracy.
Between the ways the OBIA was conducted and the TTA samples selected, there was virtually
no chance of detecting an error, even if it did exist.
While it is clear that a few errors went unnoticed in the accuracy assessment, upon
looking at the results with the eye, it is equally clear that very reliable classifications were
achieved. Importantly, the boundaries of MO classes with lowland associations were thoroughly
72

examined and corrected to ensure accurate measurements would be taken. For these reasons, the
OBIA results were deemed appropriate for analytical use.
5.2 Comparing 70-Year Changes in PVPs and CVPs
The ability to observe landscape processes in action is limited to the moments when a sensor or
an observer are present to record a snapshot. It follows that the study of such processes must
rely heavily on whatever evidence is left behind. In this case, since inundation events are
inconsistent at best, vegetation patterns were seen to be a good proxy for the presence of water in
lowland meadows. As many of the meadows with known vernal pools exhibit long-term spatial
persistence in floral patterns, it was inferred that patterns in vernal pool meadows must be
different than patterns in non-vernal meadows. Both types of lowland complex are shown on the
following page in Figure 29. Suffice to say, if two things are not the same, there must be a way
to tell them apart. A belief that this difference exists is what compelled this analysis.
Although CVPs and PVPs are similar in many respects, they do differ greatly in the
singularly important category of confirmed water sightings. Notwithstanding the obvious
difference, many non-vernal lowlands bear the hallmarks of a hidden hydric soil regime. The
initial data exploration uncovered two different types of pattern that were associated with vernal
lowlands. One involved the way in which some vernal pools had wide margins of dead
vegetation that clearly defined the meadow’s perimeter and appeared to never move. This
patterning may not be natural since it was not evident in all CVPs, nor in most other lowland
study sites, but certain PVPs did exhibit the same conditions. Coincidently or not, these few
PVP were subsequently confirmed to be vernal pools after numerous logged water sightings
enabled their proper reassignments as CVPs. In so doing, the second type of pattern emerged as
the fixed nature of upland/lowland boundaries was observed.
73

Figure 29: PVP and CVP Complexes in 1941 (top) and 2001 (bottom)
74

CVPs and PVPs were compared at the population level across six size- and shape-related
metrics, using the independent samples t-test. The test compares the means of two populations
for a given parameter, provided that the datasets are normally distributed and display equal
variances. To accept the hypothesis that CVPs and PVPs in the study area are fundamentally
different is to reject the null hypothesis that they are the same. Statistically speaking, rejecting a
hull hypothesis with a 95% confidence interval (p ≤ 0.05) is making the statement that there is
less than a 5% probability that the pattern evident in the data is the result of chance. Accepting
the null hypothesis does not denote the same thing as does rejecting it, though. The former
simply acknowledges that any differences in the two populations are not significant enough to be
thought of as different with 95% confidence, while the latter states that, with 95% confidence,
the two populations are significantly different. Rejecting a null hypothesis is therefore a stronger
statistical statement. That said, accepting the null hypothesis for all of the metrics is a strong
result in itself. It is clear that in the context of shape and size changes, statistical differences
between the PVP and CVP sites are not evident in these data.
Indeed, there must be a difference in plant patterning around meadows containing and
not-containing vernal pools. The distributions of 70-year SI and FD differences for CVPs and
PVPs were not expected to be similar. That fact that they were not different still supports the
original assertion of this thesis though; namely that water is the ultimate arbiter of plant
distributions at the upland-lowland interface. The preponderance of evidence presented in the
scientific literature supports this notion as well. P-values for SI and FD of 0.514 and 0.397,
respectively, suggest these datasets are undoubtedly alike. The similarity is confirmed by the
results of the PI t-test. Although a log transformation was required, the p-value of 0.781
underscores the similarity of the two distributions. From these data, there is no reason to suggest
75

that boundary changes occur differently in PVP meadows than in CVP meadows. These results
are compelling, and difficult to attribute to anything other than flooded lowland soils that sharply
give way to xeric upland conditions at slightly higher elevations.
5.3 Comparing Short- and Long-Term Patterns of Change in Self-Same CVPs
The rate of change analysis for self-same CVP produced inconclusive results. The meaning
behind the results could be something as profound as an unforeseen ecological mechanism, or as
mundane as variation in the smoothness of drawn boundaries. Even occasional edge effects
caused by road cuts might have been enough to throw off the results (visible as long, thin,
whitish features adjacent to several lowlands illustrated in Figure 30, next page). The data
provides little insight into what is really happening.
Qualitatively, it is apparent that, year after year, CVPs have similar appearances as past
versions of themselves. Some images show that there are locations where stretches of a CVP
boundary may move slightly back and forth. There are also almost always well-defined portions
of the depression where upland or transitional vegetation break in long, smooth, stationary arcs.
The arc-like pool edges appear virtually unmoved over time, the imprint of a flood cycle that has
continued for millennia. Mima mounds within pool margins also create small depressions
between them that can harbor lesser vernal pools in years that lack enough rainfall to fill the
greater pool to capacity. Even though the distributions and characteristics of transitional
vegetation vary from pool to pool, there is always a central region that is the flood zone within
which only grass can survive. As the neighboring chaparral matures but is unable to encroach
upon the flood zone, the vernal meadow appears to be cut into the surrounding brush, as if it has
been carved into the land surface itself.
76

Figure 30: PVP and CVP Complexes in 2001 (top) and 2011 (bottom)
77

In a very real sense, pool boundaries are inscribed on the landscape. Yet, typical
meadows encompass more than just the flood zone. Usually there is a subtly elevated terrace
where grasslands eventually grade into transitional or upland vegetation. These elevated areas
are presumably affected by the flooding as well. During inundation, the water table rises enough
to be above ground in the flood zone, and should be close to the surface at slightly higher
elevations nearby. Thus the flooding effects more than just the areas that are visibly submerged.
Upland vegetation may suffer as much for being near a pool as it would for being in one. It is
logical to assume that, in drier years without surface pools, the water table may still rise enough
to create the damp soil conditions necessary to rebuff upland encroachment.
The observed consistency of vernal meadow margins is supported by the mean
differences in WIPI, but does not match the results from the other metrics. The performance of
PI in the 70-year analysis lends it credibility as a measurement. However, low standard errors
and standard error means across the board in the self-same comparison suggest that none of the
data is random. Since the visual appearance of the study sites denotes both short- and long-term
stability, as does the WIPI t-test result, the sensible conclusion is that CVPs are geometrically
stable over an indefinite time frame. The most likely explanation for the remaining results is that
they have been interpreted incorrectly, possibly because the relationships between the metrics is
more complicated than expected. The numbers are too self-consistent to be discounted, but they
also do not match the most obvious explanation. Therefore, there is likely a mistake in final
interpretation of the meaning.
5.4 Sources of Error
In any research endeavor, there are opportunities for errors to inject uncertainty into the results.
Even in a well-designed study there are avenues by which errors can arise. For instance, the
78

photographs that were used in this analysis were not perfectly co-registered. It is possible that
distortions, particularly in the base photograph from 1941, resulted in classifications that were
not true to life at the time. Yet, such an error would actually be all the more reason to accept the
results of this study, since unequal distortion across an image would lead to a more randomized
pattern of change in the data than was actually present. In regards to the 1941 photo specifically,
the individual mosaic tiles were not all of equal spatial resolution. Prior to its use here, however,
the entire mosaic had clearly been resampled to a 0.5 m pixel size.
The most likely place for error to have occurred would have been in the image analysis.
Both segmentation and classification require thorough examination to eliminate gross errors, but
for large rasters, optical revue is only practical on a cursory scale. Fortunately, since lowland
meadows and the immediately adjacent upland chaparral were the only targets of the analysis,
manually scanning for and correcting errors was feasible. The highest probability for error in the
entire study was in the classification semantics and their individual applications. The landscape
was much more mixed in the later images than first anticipated, owing greatly to disturbances
and successional vegetation. There were instances where determining the proper classification
for an object was difficult without ground evidence. Inputs from higher resolution imagery (0.25
m) added marginal clarification, but also led to a point of diminishing returns as the objects
became more and more spectrally mixed. In point of fact, a slight reduction in resolution of the
segmented image might have improved results more because objects would have appeared more
blurry at the edges. The blur would have had an anti-aliasing effect that would generalize the
appearance of objects and their boundaries. Increasing resolution does the opposite; objects turn
spectrally intricate and become fractal at their edges.
79

Resolution discrepancies may have also accounted for the unexplained growth in mean
perimeter that contrasted with the shrinkage of the mean area for CVPs at the 10-year interval.
The higher resolution image acquired in 2011 made it possible to more accurately segment the
irregular edges of many meadows, which would have had a much greater effect on measures of
perimeter than area. Moreover, this effect was probably exacerbated by minor changes in plant
composition at meadow boundaries.
Another problem related to boundary edges come from edge effects. Edge effect, in
essence, is the influence an environmental type has on its neighbors, which typically penetrates
some distance into the neighboring landscapes and affects their ecosystems. There are two
potential manifestations of edge effect that were probable in this study. The first, the edge effect
happening at lowland/upland boundaries, was the primary focus of this research. Where the
slope angle was the steepest, the edge effect was narrow. Simply put, the transitional band
between upland and lowland vegetation had a thin width when the soil moisture gradient (i.e.
slope angle) was steep. When the slope angle was lower, as was often the case in the upper
terrace of a meadow, the soil moisture gradient was obviously shallower and the transition band
was wider and visibly more diverse. The identities of transitional flora were not known for
certain at the time of this writing, making it difficult to create an adequate name in the
classification scheme. Suspected constituents included annual and native grasses, sandmat
manzanita (A. pumila), chamise, Monterey ceanothus, poison oak (Toxicodendron diversilobum),
ice plant, and pampas grass.
The other type of edge effect of concern here is evident in areas where a road has been
maintained close enough to the boundary of a lowland to become part of the edge itself. Most of
the time, the road is cut just inside or just outside of a boundary, as opposed to right up against it.
80

Either way, the result is sometimes a narrow band of upland sandwiched between a road and a
lowland, or vice versa; both grass and chaparral that are caught in the middle are subjected to
extra environmental pressure from the isolation and the nearby road. In the end roads were not
judged to be problematic enough to account for them in this study, but they easily could have
contributed some error, especially to falsely support the findings of geometric stability presented
herein.
The true identities of the lowland grasses might also be a source of error, as there is
undoubtedly some mixture of non-vernal, non-native species combined with natives. The native
grasses themselves may be of either vernal or non-vernal varieties, as well. Since this study
treated all grasses equally, there is a deeper level of granularity to be sought after. Imagery was
available to this study with sufficiently high accuracy to identify individual grass and possibly
even indicative wildflower species. Such an effort in the future could provide definitive
evidence for vernal pool presence in the absence of water, yet was unachievable here because
commensurate ground evidence did not exist and could not be acquired.
Finally, the ground evidence that was captured during the flyover brought to light a type
of classification error that was unanticipated. The severity of erosion gullies, and their ostensible
coincidence with roads and trails, made it clear that many of the linear structures classified as
Bare Ground/Dirt Track were better classified into a separate erosion category. Yet, since Bare
Ground was not considered in the statistical analysis, unclassified erosive features probably did
not skew the results. At the same time, there can be no question that the true classification
accuracies for the 2001 and 2011 analyses were marginally lower than estimated.
81

5.5 Final Thoughts
There exists a phenomenon that prevents chaparral and oak trees from encroaching upon the
bottoms of lowland depressions in the Fort Ord study area. For known vernal pools, the reason
is self-evident. The cyclical presence of water in the form of seasonal flooding prevents upland
vegetation from competing in the lowland areas. Thus the vernal lowlands are able to exist as
grassy meadow islands surrounded by a dense matrix of manzanita and coast live oak.
This study has provided reasonable argument that a similar hydric process is occurring in
many of the other lowland sites as well. Five previously unrecognized vernal pools have been
discovered and reported for the first time here (Figure 31). Each of these sites now requires
consideration as protected wetland habitat for California tiger salamander, fairy shrimp, and
Contra Costa goldfields wildflowers. Tiger salamanders in particular require a 2 km upland
habitat buffer, so the importance of these findings should not be overlooked.
In addition, a wide variety of empirical and observational evidence has been provided to
warrant a more extensive study of the PVP sites at Fort Ord. On average, CVP and PVP margins
change about the same amount over time. They both appear as lush and verdant grasslands
during the spring if no water features are present, and each repels the encroachment of upland
vegetation equally well. The similarities between PVPs and CVPs are noteworthy.
A strong case has been made that there remain vernal pools in the Fort Ord study area
that are as yet unrecognized. Other PVPs studied here could be the remnants of lost vernal
pools, having suffered too much nearby disturbance to become vernally active of late, but
maintaining a hydric soil cycle that keeps them still thriving as grasslands. This study was
unable to identify relict CVPs specifically without seeing an inundation event or collecting
ground evidence in situ. In the absence of direct confirmation, future studies should focus
82

Figure 31: Five Previously Unconfirmed Vernal Sites within MRS-BLM

efforts toward assessing the PVPs with similar vegetative persistence as CVP. Such studies
would be particularly impactful if they combined soil moisture testing to check for latent hydric
potential with observations of flora. Sites identified as having unusually damp soils may be good
candidates for restoration efforts aimed at returning the land to a more natural state. Soil
samples could also be collected for ground evidence, to be included in a spectral analysis that
uses remote sensing data to compare vernal and non-vernal soil compositions.
The remote sensing study of Fort Ord’s vernal and non-vernal lowlands presented in this
thesis was, to the author’s knowledge, the first of its kind. This work presented a method for
integrating historic and contemporary remote sensing data with GIS and parametric statistics,
83

Figure 32: The First Ever Comprehensive Map of MRS-BLM Vernal Sites

to achieve a novel, spatiotemporal perspective on the interactions between upland and lowland
plant communities. Additionally, a new metric, the Persistence Index, was created from patch
data derived in a GIS. PI was demonstrated to be a valid blanket descriptor of the degree of
change in patch geometry. Over a 70-year period at Fort Ord, vernal meadow patches changed
about the same amount as non-vernal patches, indicating an underlying similarity between them.
Finally, through incidental observation, five new vernal pools were discovered and identified
during the image exploration and analysis portion of the study. In Figure 32 these five, along
with all 15 other known vernal pools in the impact area, have now been compiled to reveal the
first complete map of vernal sites within the MRS-BLM boundaries. The work the map
84

represents was applied for the first time here, but it should not be the last effort. If the legacy of
this project is to simply be a link in the chain that advances the field of spatial ecology with
clarity, then it will have achieved its prime directive.

85

REFERENCES
Andrew, Margaret E. and Susan L. Ustin. 2008. The role of environmental context in mapping
invasive plants with hyperspectral image data. Remote Sensing of Environment, 112: 4301–
4317.
Baatz, Martin and Arno Schäpe. 2000. Multiresolution segmentation: an optimization approach
for high quality multi-scale image segmentation. In Angewandte Geographische
Informations-Verarbeitung, edited by Josef Strobl, Thomas Blaschke, and Gerald Griesebner,
XII. Wichmann Verlag, Karlsruhe: 12–23.
Backoulou, Georges F., Norman C. Elliott, Kristopher Giles, Mpho Phoofolo, and Vasile Catana.
2011. Development of a method using multispectral imagery and spatial pattern metrics to
quantify stress to wheat fields caused by Diuraphis noxia. Computers and Electronics in
Agriculture, 75: 64–70.
Bagella, Simonetta, Maria Carmela Caria, and Vincenzo Zuccarello. 2010. Patterns of
emblematic habitat types in Mediterranean temporary wetlands. Comptes Rendus Biologies,
333(9): 694–700.
Bauder, Ellen T. 2000. Inundation effects on small-scale plant distributions in San Diego,
California vernal pools. Aquatic Ecology, 34: 43–61.
----. 2005. The effects of an unpredictable precipitation regime on vernal pool hydrology.
Freshwater Biology, 50: 2129–2135.
Barbour, Michael, Ayzik Solomeshch, Carol Witham, Robert Holland, Rod MacDonald, Sarel
Cilliers, Jose A. Molina, Jennifer Buck, and Janelle Hillman. 2003. Vernal pool vegetation
of California: variation within pools. Madroño, 50(3): 129–146.
86

Blaschke, Thomas. 2010. Object-based image analysis for remote sensing. ISPRS Journal of
Photogrammetry and Remote Sensing, 65: 2–16.
Burnett, Charles and Thomas Blaschke. 2003. A multi-scale segmentation/object relationship
modelling methodology for landscape analysis. Ecological Modelling, 168: 2–249.
BLM (Bureau of Land Management). 2010. Vascular plants of Fort Ord. Last modified
10/12/2010.
http://www.blm.gov/style/medialib/blm/ca/pdf/hollister.Par.93735.File.dat/FtOrdPlantsAlph1
2Aug10_c.pdf.
Burleson Consulting, Inc. 2006. Wetland monitoring and restoration plan for munitions and
contaminated soil remedial activities at Former Fort Ord. Prepared for the U.S. Army Corps
of Engineers, Sacramento, CA.
Burne, Matthew R. 2001. Massachusetts aerial photo survey of potential vernal pools.
Prepared for the Massachusetts Division of Fisheries & Wildlife.
Carpenter, Laurel, Janice Stone, and Curtis R. Griffin. 2011. Accuracy of aerial photography
for locating seasonal (vernal) pools in Massachusetts. Wetlands, 31: 573–581.
Coburn, Craig and Arthur C.B. Roberts. 2004. A multiscale texture analysis procedure for
improved forest stand classification. International Journal of Remote Sensing, 25(20): 4281–
4308.
Cooper, Daniel S. and Dan L. Perlman. 1997. Habitat conservation on military installations.
Fremontia, 25(1): 3–8.
Cormier, Tina A. 2001. Statistical and cartographic modeling of vernal pool locations:
Incorporating the spatial component into ecological modeling. MS Thesis, University of
New Hampshire.
87

Corcoran, Padraig, Adam Winstanley, and Peter Mooney. 2010. Segmentation performance
evaluation for object-based remotely sensed image analysis. International Journal of Remote
Sensing, 31(3): 617–645.
Cox, George W. 1986. Mima mounds as an indicator of the presettlement grassland-chaparral
boundary in San Diego County, California. American Midland Naturalist, 116(1): 64–77.
Cutler, Justin Elliot. 2006. Accuracy assessment of high-resolution multispectral satellite
imagery for remote sensing identification of wetlands and classification of vernal pool in
Eastern Sacramento County, California. MS Thesis, California State University at
Sacramento.
Dissanska, Maria, Monique Bernier and Serge Payette. 2009. Object-based classification of
very high resolution and panchromatic images for evaluating recent change in the structure of
pattern peatlands. Canadian Journal of Remote Sensing, 35(2): 189–215.
Drăgut, Lucian, and Thomas Blaschke. 2006. Automated classification of landform elements
using object-based image analysis. Geomorphology, 81: 330–344.
EMC Planning Group. 2012. Final reassessment report: Fort Ord reuse plan final reassessment.
Prepared for Fort Ord Reuse Authority, 165 pp.
EMC Planning Group and EDAW, Inc. 2001a. Fort Ord reuse plan, volume I: context and
framework. Prepared for Fort Ord Reuse Authority, 211 pp.
----. 2001b. Fort Ord reuse plan, volume 4: final EIR. Prepared for Fort Ord Reuse Authority,
352 pp.
Fallon, Michael. 2013. Spectral mixture analysis of EO-1 Hyperion imagery in the channeled
scablands of Eastern Washington. MS Thesis, Northwest Missouri State University.
88

Ferren Jr., Wayne R., David M. Hubbard, Sheila Wiseman, Anuja K. Parickh, and Nathan Gale.
1998. Review of ten years of vernal pool restoration and creation in Santa Barbara,
California. Ecology, Conservation, and Management of Vernal Pool Ecosystems—
Proceedings from a 1996 Conference, California Native Plant Society: 206–216.
Fernández-Aláez, Camino, Margarita Fernández-Aláez, and Eloy Bécares. 1999. Influence of
water level fluctuation on the structure and composition of the macrophyte vegetation in two
small temporary lakes in the northwest of Spain. Hydrobiologica, 415: 155–162.
Forestier, Germaine, Anne Puissant, Cédric Wemmert, and Pierre Gançarski. 2012. Knowledge-
based region labeling for remote sensing image interpretation. Computers, Environment, and
Urban Systems, 36: 470–480.
Frohn, Robert C., Ellen D’Amico, Charles Lane, Brad Autrey, Justicia Rhodus, and Hongxing
Liu. 2012. Multi-temporal sub-pixel Landsat ETM+ classification of isolated wetlands in
Cuyahoga County, Ohio, USA. Wetlands¸32: 289–299.
Frohn, Robert C., Molly Reif, Charles Lane, and Brad Autrey. 2009. Satellite remote sensing of
isolated wetlands using object-oriented classification of Landsat-7 data. Wetlands, 29(3):
931–941.
Gibbes, Cerian, Sanchayeeta Adhikari, Luke Rostant, Jane Southworth, and Youliang Qiu.
2010. Application of object based classification and high-resolution satellite imagery for
savanna ecosystem analysis. Remote Sensing, 2: 2748–2772.
Gopal, B. 1986. Vegetation dynamics in temporary and shallow freshwater habitats. Aquatic
Botany, 23: 391–396.
89

Gordon, Sarah P., Christina M. Sloop, Heather G. Davis and J. Hall Cushman. 2012. Population
genetic diversity and structure of two rare vernal pool grasses in central California.
Conservation Genetics, 13: 117–130.
Grant, Evan H.C. 2005. Correlates of vernal pool occurrence in the Massachusetts, USA
landscape. Wetlands, 25(2): 480–487.
Haas, Eva M., Etienne Bartholomé, and Bruno Combal. 2009. Time series analysis of optical
remote sensing data for the mapping of temporary surface water bodies in sub-Saharan
western Africa. Journal of Hydrology, 370: 52–63.
Halounova, Lena. 2004. The automatic classification of B&W aerial photos. The International
Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences,
34(XXX): 455–460.
Highfield, Linda, Michael P. Ward, Shawn W. Laffan. 2008. Representation of animal
distributions in space: how geostatistical estimates impact simulation modeling of foot-and-
mouth disease spread. Veterinary Research, 39(17): 1–14.
Hsu, Wynne, Tat-Seng Chua, and Hung K. Pung. 2000. Approximating content-based object-
level image retrieval. Multimedia Tools and Applications, 12: 59–79.
Jiang, Zhenlan, Shiliang Su, Changwei Jing, Shengpan Lin, Xufeng Fei, and Jiaping Wu. 2012.
Spatiotemporal dynamics of soil erosion risk for Anji County, China. Stochastic
Environmental Research and Risk Assessment, 26: 751–763.
Johansen, Kasper, Nicholas C. Coops, Sarah E. Gergel, and Yulia Stange. 2007. Application of
high spatial resolution satellite imagery for riparian and forest ecosystem classification.
Remote Sensing of Environment, 110: 29–44.
90

Jones & Stokes Associates. 1992. Flora and fauna baseline study of Fort Ord, California.
Prepared for U.S. Army Corps of Engineers, Sacramento, CA: 290 pp.
Keeler-Wolf, Todd, Diane R. Elam, Kari Lewis, and Scott A. Flint. 1998. California vernal pool
assessment preliminary report. Prepared for the California Department of Fish and Game,
161 pp.
Kneitel, Jamie M. 2014. Inundation timing, more than duration, affects the community structure
of California vernal pool mesocosms. Hydrobiologica, 732: 71–83.
Lane, Charles R., Ellen D’Amico, and Brad Autrey. 2012. Isolated wetlands of the southeastern
United States: abundance and expected condition. Wetlands, 32: 753–767.
Lathrop, Richard G., Paul Montesano, Jason Tesauro, and Brian Zarate. 2005. Statewide
mapping of vernal pools: a New Jersey case study. Journal of Environmental Management,
76(2005): 230–238.
LFR, Weston Solutions, and Westcliffe Engineers, Inc. 2008. FORA ESCA remediation
program; final community involvement and outreach program plan. Prepared for the Fort
Ord Reuse Authority.
Lichvar, Robert W., David C. Finnegan, Stephen Newman, and Walter Ochs. 2006. Delineating
and evaluating vegetation conditions of vernal pools using spaceborne and airborne remote
sensing techniques, Beale Air Force Base, CA. Prepared for: U.S. Army Corps of Engineers,
1–18.
Longcore, Travis and Catherine Rich. 2008. Invertebrate conservation at the gates of hell.
Wings, Spring 2008: 9–13.
91

Mallinis, Giorgos, Magdalini Pleniou, and Nikos Koutsias. 2010. Object-based vs pixel-based
mapping of fire scars using multi-scale satellite data. The International Archives of the
Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXVIII-4(C7): 1–5.
Mattoni, Rudi and Travis R. Longcore. 1997. The Los Angeles coastal prairie, a vanished
community. Crossosoma, 23(2): 71–102.
Mayer, Alex S., Patrick P.E. Carriere, Claudio Gallo, Kurt D. Pennell, Tammy P. Taylor, Glenn
A. Williams, and Lirong Zhong. 1997. Groundwater Quality. Water Environment
Research, 69(4): 777–844.
Montrone, Ashton. 2013. Hydrologic and vegetative modeling of vernal pools in the Sierra
Nevada. MS Thesis, University of Nevada, Reno.
Moore, Craig, Monica Bastian, and Harold Hunt. 2001. Long-term vegetational and faunal
succession in an artificial Northern California vernal pool system. Prepared for California
Department of Transportation, FHWA/CA/TL-2001/36: 1–49.
Munoz, Breda, Virginia M. Lesser, John R. Dorney, and Rick Savage. 2009. A proposed
methodology to determine accuracy of location and extent of geographically isolated
wetlands. Environmental Monitoring and Assessment, 150: 53–64.
Rhazi, Laila, Mouhssine Rhazi, Patrick Grillas, and Driss El Khyari. 2006. Richness and
structure of plant communities in temporary pools from western Morocco: influence of
human activities. Hydrobiologica, 570: 197–203.
Reed, Sarah and Ronald Amundson. 2007. Sediment, gophers, and time: a model for the origin
and persistence of mima mound—vernal pool topography in the Great Central Valley.
Vernal Pool Landscapes. In Studies from the Herbarium, edited by R.A. Schlising, and D.G.
Alexander, 14: 15–27.
92

Rempel, Robert S., Daniel Kaukinen, and Angus P. Carr. 2012. Patch analyst and patch grid.
Ontario Ministry of Natural Resources. Centre for Northern Forest Ecosystem Research,
Thunder Bay, Ontario.
Schlising, Robert A. and Ellen L. Sanders. 1982. Quantitative analysis of vegetation at the
Richvale Vernal Pools, California. American Journal of Botany, 69(5): 734–742.
Smeulders, Arnold W.M., Marcel Worring, Simone Santini, Amarnath Gupta, Ramesh Jain.
2000. Content-based image retrieval at the end of the early years. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 22(12): 1349–1380.
Smith, Douglas, Bob Curry, Donald Koslowski, Regina Williams, Fred Watson, Leslie Turrini-
Smith, and Wendi Newman. 2002. Watershed and riparian assessment report: Bureau of
Land Management lands, former Fort Ord, Monterey County, California. Prepared for the
Watershed Institute, California State University at Monterey Bay, WI-2002-01: 151 pp.
Stone, Janice S. 1992. Vernal pools in Massachusetts: aerial photographic identification,
biological and physiographic characteristics, and state certification. MS Thesis, University
of Massachusetts Amherst.
Tannourji, Danielle N. 2009. Ecological factors suitable for the endangered Lasthenia
conjugens (Asteraceae). MS Thesis, San José State University.
Tetra Tech Inc. and EcoSystems West Consulting Group. 2014. 2013 biological monitoring
report for units 7, 5E, and 23E; units 15, 21, 32, and 34; units 18 and 22; and Ranges 43–48
Former Fort Ord. Prepared for U.S. Army Corps of Engineers, Sacramento, CA: 1–209.
Thompson, Shanley D. and Sarah E. Gergel. 2008. Conservative implications of mapping rare
ecosystems use high spatial resolution imagery: recommendations for heterogenous
fragmented landscapes. Landscape Ecology, 23: 1023–1037.
93

Tiner, Ralph W. 2003. Estimated extent of geographically isolated wetlands in selected areas of
the United States. Wetlands, 23(3): 636–652.
USGS. Earth Explorer Application. 2014. U.S. Geological Survey.
USFWS (U.S. Fish and Wildlife Service). 2005. Recovery plan for vernal pool ecosystems of
California and Southern Oregon. Portland, Oregon. xxvi + 606 pages.
Van Dyke, Eric. and Karen Holl. 2003. Mapping the distribution of maritime chaparral species
in the Monterey Bay area. Prepared for U.S. Fish and Wildlife Service, Ventura Fish and
Wildlife Office: 22 pp.
Van Dyke, Eric, Karen D. Holl, and James R. Griffin. 2001. Maritime chaparral community
transition in the absence of fire. Madroño, 48:221–229.
Van Meter, Robin, Larissa L. Bailey, and Evan H.C. Grant. 2008. Methods for estimating
vernal pool habitat in the northeastern United States. Wetlands, 28(3): 585–593.
Van Thomme, Eric. 2001. Restored vernal pools in California and post-creation monitoring and
management. Student On-Line Journal, Department of Horticultural Science, University of
Minnesota, 7(4): 1–6.
Wang, Ian J., Jarret R. Johnson, Benjamin B. Johnson, and H. Bradley Shaffer. 2011. Effective
population size is strongly correlated with breeding pond size in the endangered California
tiger salamander, Ambystoma californiense. Conservation Genetics, 12: 911–920.
Yu, Huan and Shu-Qing Zhang. 2008. Application of high resolution satellite imagery for
wetlands cover classification using object-oriented method. In The International Archives of
the Photogrammetry, Remote Sensing and Spatial Information Sciences, edited by Chen Jun,
Jiang Jie, and John van Genderen, Vol. XXXVII. Part B7: 521–526.
94

Zedler, Paul H. 1987. The ecology of Southern California vernal pools: a community profile.
Biological Report, Prepared for the U.S. Fish and Wildlife Service, 85(7.11): 137 pp.
----. 2003. Vernal pools and concept of “isolated wetlands.” Wetlands, 23(3): 597–607.
Zhang, Yun and Travis Maxwell. 2006. A fuzzy logic approach to supervised segmentation for
object-oriented classification. ASPRS Annual Conference, Reno, NV.

95

APPENDIX A: SUPPORTIVE IMAGERY

Figure 33: PVP Site 70P05 Displays Floral Persistence Similar to that of Study Area CVP (Source
Imagery: Google Earth, C. Hanley —bottom right)
96

Figure 34: Previously Unconfirmed CVP 80P03 (bottom row) Was Flooded During the Flyover
While Known CVP 80C02 Was Not (Source Imagery: USGS EarthExplorer, Google Earth, C.
Hanley —top and bottom)
97

Figure 35: Comparative Shoreline Shapes for Known and Previously Unconfirmed CVP (Photo
Credits: C. Hanley)

Figure 36: Inundated CVP Site 11P01 (Previously Unknown), Likely Created by Disturbance
(Photo Credits: C. Hanley)
98

Figure 37: Enlargement of Erosion Drainage Scar in Site ID# 80P07 (Photo Credits: C. Hanley)
99

APPENDIX B: ENLARGEMENTS OF CLASSIFICATION RESULTS

Figure 38: Enlargement of 1941 Classification Results
100

Figure 39: Enlargement of 2001 Classification Results
101

Figure 40: Enlargement of 2011 Classification Results
102

APPENDIX C: MAPS OF MEASURED METRICS

Figure 41: Lowland Complex FD Changes by Type, 1941-2001

Figure 42: Lowland Complex FD Changes by Type, 2001-2011
103

Figure 43: Lowland Complex SI Changes by Type, 1941-2001

Figure 44: Lowland Complex SI Changes by Type, 2001-2011
104

Figure 45: Lowland Complex PAR Changes by Type, 1941-2001

Figure 46: Lowland Complex PAR Changes by Type, 2001-2011
105

Figure 47: Lowland Complex PI Values by Type, 1941-2001

Figure 48: Lowland Complex PI Values by Type, 2001-2011
106

Figure 49: 70-Year WIPI for Self-Same CVP

Figure 50: 10-Year WIPI for Self-Same CVP
107

Figure 51: Lowland Complex Percent Area Changes by Type, 1941-2001

Figure 52: Lowland Complex Percent Area Changes by Type, 2001-2011
108

Figure 53: Lowland Complex Percent Perimeter Changes by Type, 1941-2001

Figure 54: Lowland Complex Percent Perimeter Changes by Type, 2001-2011
109

APPENDIX D: EXPLORATORY STATISTICS RESULTS
110

Figure 55: Exploratory Statistics for Change in FD, 1941-2001
111

Figure 56: Exploratory Statistics for Change in SI, 1941-2001
112

113

Figure 57: Exploratory Statistics for Change in PAR, 1941-2001
114

115

Figure 58: Exploratory Statistics for Log(PI), 1941-2001
116

Figure 59: Exploratory Statistics for % Area Change, 1941-2001
117

118

Figure 60: Exploratory Statistics for % Perimeter Change, 1941-2001
119

Figure 61: Exploratory Statistics Self-Same CVP FD Differences

Figure 62: Exploratory Statistics Self-Same CVP SI Differences
120

Figure 63: Exploratory Statistics Self-Same CVP PAR Differences

Figure 64: Exploratory Statistics Self-Same CVP WIPI Differences
121

Figure 65: Exploratory Statistics Self-Same CVP % Area Differences

Figure 66: Exploratory Statistics Self-Same CVP % Perimeter Differences

Abstract (if available)

Abstract Vernal pools are rare, seasonal pools that form in landscape depressions and create temporary habitat for many floral and faunal taxa. In California, as much as 90% of historic vernal pool area has been displaced by agriculture and urbanization. Pools are commonly inhabited by endemic, threatened, and endangered plants and animals, and are critical breeding areas for California tiger salamanders (Ambystoma californiense) and fairy shrimp (Linderiella occidentalis). Seasonal inundation and desiccation are driving factors behind the biotic community structure around pools, both spatially and compositionally. ❧ At Fort Ord, California, a rare subset of vernal pools occur perched atop relict sand dunes in an arid chaparral environment. Fifteen vernal pools have been previously identified within the base’s historic firing range impact area. At least 45 other lowland meadows within the impact area meet pool topographic requirements and were evaluated for their potential to be vernal habitats. This thesis proposes an object‐based method of extracting vegetative patterns from VHSR Ikonos and WorldView 2 satellite imagery, to compare persistence in vegetative patterns over time. Classification results from three aerials collected over an 80‐year interval were subjected to a geospatial change analysis, and used to make short‐ and long‐term comparisons of known vernal meadows to themselves and other meadows in the study area. Two new metrics, the Persistence Index and Weighted Intervals Persistence Index, were created for this study. These indices normalize changes in geometric properties, enabling comparisons between known vernal areas and study sites, and between self‐same sites sampled at different times. PI and WIPI results were consistent with the results from other analyzed metrics. ❧ Strong persistence in several study sites, comparable to that of the known vernal areas, likely indicates latent presence of a seasonal hydric regime and an elevation‐based hydrological gradient. The results of this study show that there is no statistically significant difference between the way that vernal and other meadows change shape and size over time. This result means that a number of lowland meadows in the impact area may have active or dormant vernal pools because the two groups cannot be empirically differentiated from one another. This study also positively confirmed the presence of at least five previously unrecognized vernal areas through the detection of water in multiple aerial images. These findings merit further on‐the‐ground investigation, as well as a geographical reconsideration of current conservation efforts.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

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

Questioning the cause of calamity: using remotely sensed data to assess successive fire events

PDF

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

PDF

A Maxent-based model for identifying local-scale tree species richness patch boundaries in the Lake Tahoe Basin of California and Nevada

PDF

Spatial and temporal patterns of long-term temperature change in Southern California from 1935 to 2014

PDF

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

PDF

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

PDF

Modeling burn probability: a Maxent approach to estimating California's wildfire potential

PDF

Using GIS to identify potential dynamic marine protected areas: a case study using shortfin mako shark tagging data in New Zealand

PDF

Investigation and analysis of land use/land cover to determine the impact of policy developments in cities

Asset Metadata

Creator Hanley, Charles Joseph, III (author)

Core Title Using VHSR multispectral imagery and object-based extraction to discover vernal meadows through vegetative persistence at Fort Ord, California

School College of Letters, Arts and Sciences

Degree Master of Science

Degree Program Geographic Information Science and Technology

Publication Date 06/24/2015

Defense Date 05/07/2015

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

Tag chaparral,classification,Ft. Ord,Ikonos,multispectral,OAI-PMH Harvest,object‐based image analysis,Persistence Index,vernal pool,Weighted Intervals Persistence Index,WorldView 2

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Wilson, John P. (committee chair), Lee, Su Jin (committee member), Longcore, Travis R. (committee member)

Creator Email cjhanley@usc.edu,hanley_c_j@hotmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-576672

Unique identifier UC11300946

Identifier etd-HanleyChar-3511.pdf (filename),usctheses-c3-576672 (legacy record id)

Legacy Identifier etd-HanleyChar-3511.pdf

Dmrecord 576672

Document Type Thesis

Format application/pdf (imt)

Rights Hanley, Charles Joseph, III

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