.d .
£3...
3.4.. .
3. n .
5.9”... ...... ..
. a... .5. w...
in n 1 .
fiéfimfimfl
as. a u ,
v. 50“..” H’?

..
swan“ . .
.2... 3.119.”...
. finfit

. Jar"... an?
n, Ffiuw I‘l

. .. an...
. {gawk}.ka
b . 4 . 1!. If. '-
"9..um1 .hw 1.4.2....
. 3.2.34.2
it . 19'
J.

. l S}
mffifiwnfif
L...)
a...

3.5.

.........l...

t: )2.
r .3... .25..)
. 3.3 .
35.3...
.43....
31......

. .. 2.3 thflnnldnuhu

1.. 2...... ..3
a 4.2.4.523: .2."
i...

. tin"...

3)
.L n
3 iii... $1.3 .r... II...
:13": VI’Yolliyh :Iu
{F.0...vl.uv.l~.tt:l~ ..

get...
5.27....

 

 

 

3 LIBRARY
ZOB‘I Michigan State
University

 

 

 

This is to certify that the
dissertation entitled

OBJECT-BASED IMAGE ANALYSIS FOR SCALING
PROPERTIES OF RANGELAND ECOSYSTEMS: LINKING
FIELD AND IMAGE DATA FOR MANAGEMENT DECISION

MAKING

presented by

JASON WILLIAM KARL

has been accepted towards fulfillment
of the requirements for the

Ph.D. degree in FISHERIES AND WILDLIFE

 

 

MAI/Wm

Major Professor’s Signature

Wima

Date

 

MSU is an Affirmative Action/Equal Opportunity Employer

fl.--n-.---.-.-2-.-o-a-.---.---u-o--u-r--v-u---u-------o-.-.-o-----------.-c-..

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5/08 K1IProlecc8Pres/CIRC/DateDue.indd

 

OBJECT-BASED IMAGE ANALYSIS FOR SCALING PROPERTIES OF
RANGELAND ECOSYSTEMS: LINKING FIELD AND IMAGE DATA FOR
MANAGEMENT DECISION MAKING
By

Jason William Karl

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
DOCTOR OF PHILOSOPHY
Fisheries and Wildlife

2009

 

ABSTRACT
OBJECT-BASED IMAGE ANALYSIS FOR SCALING PROPERTIES OF
RANGELAN D ECOSYSTEMS: LINKING FIELD AND IMAGE DATA FOR
MANAGEMENT DECISION MAKING
By
Jason William Karl
Management of semi-arid shrub-steppe ecosystems (i.e., rangelands) requires
accurate information over large landscapes, and remote sensing is an attractive option for
collecting such data. To successfully use remotely-sensed data in landscape-level
rangeland management, questions as to the relevance of image data to landscape patterns
and optimal scales of analysis must be addressed. Object-based image analysis (OBIA),
which segments image pixels into homogeneous regions, or objects, has been suggested
as a way to increase accuracy of remotely-sensed products, but little research has gone
into how to determine sizes of image objects with regard to scaling of ecosystem
properties. The purpose of my dissertation was to determine if OBIA could be used to
generate observational scales to match ecological scales in rangelands and to explore the
potential for OBIA to generate accurate and repeatable remote-sensing products for
managers. The work presented here was conducted in southern Idaho’s Snake River Plain
region. By comparing OBIA segmentation of satellite imagery into successively coarser
objects to pixel-based aggregation methods, I found that canonical correlations between
field-collected and image data were similar at the finest scales, but higher for image
segmentation as scale increased. 1 also detected scaling thresholds with image
segmentation that were confirmed via semi-variograms of field data. This approach

proved useful for evaluating the overall utility of an image to address an objective, and

identifying scaling limits for analysis. I next used observations of percent bare-ground
cover from 346 field sites to consider how hierarchies of image objects created through
OBIA could be used to discover appropriate scales for analysis given a specific objective.
Using a regression-based approach, I found that segmentation levels whose predictions of
bare-ground cover had spatial dependence that most closely matched the spatial
dependence of the field samples had the highest predicted-to-observed correlations.
When combined with geostatistical predictors, these changes in spatial variance with
scale led to robust predictions across a range of scales. Third, I demonstrated an
application of OBIA with the technique of regression kriging (RK), a geostatistical
interpolator, to make spatial predictions for three aspects of rangeland condition (percent
cover of shrubs, bare ground, and cheatgrass [Bromus tectorum L.]). Comparing spatial
predictions from generalized least-squares (GLS) regression to RK, I found that RK
implemented with OBIA produced more accurate results than GLS regression alone for
all three variables measured by cross-validated root mean-squared error. Finally, I
considered why techniques like OBIA, and remote sensing in general, are not more
widely used in routine rangeland management. Bolstering decision-making through 1)
better information tools and data to support management and 2) adaptive management
has been proffered as a means for making sound management decisions, but two recent
lawsuits in southern Idaho suggest that neither of these solutions is likely to be effective
at managing rangelands at scales commensurate with their threats unless there are
changes to the underlying management paradigm governing how the public participates

in the management process.

ACKNOWLEDGMENTS

This dissertation would not have been possible without the contributions and
sacrifice of a great number of people. I would like to specially recognize the following
individuals and institutions for their help and support.

Three individuals stand out because their contributions have directly led to the
successes I have realized. Through many discussions of with Dr. Jeff Yeo about how
science could effect better conservation of Idaho’s rangelands the beginnings of this
dissertation were hatched. Jeff not only gave me encouragement to pursue my Ph.D., but
also arranged for me to continue working for The Nature Conservancy (TNC) throughout
my degree program. Dr. Bob Unnasch has also been a great advocate for me within TNC
and a stimulating colleague to work with. My discussions with Bob have helped
immensely as the ideas presented here have matured. Dr. Brian Maurer, my graduate
advisor at Michigan State University (MSU), challenged me to grow intellectually during
my time at MSU and helped me develop and refine not only the ideas presented herein
but also my professional direction. I am very grateful to Jeff, Bob, and Brian for their
guidance, support, and friendship throughout my doctoral program.

My graduate committee, Dr. Jiagou Qi, Dr. Shawn Riley, and Dr. Gary Roloff,
challenged me to develop a deeper understanding of landscape ecology and remote
sensing and to broaden my perspectives on the interplay between science and social
factors in management decision making. They also provided insightful discussion and
invaluable review of this manuscript.

The staff of the Idaho Chapter of TNC and the other graduate students in Dr.

Maurer’s lab have also been a great source of support and encouragement. In particular, I

iv

would like to recognize and thank Karen Colson and Alan Sands for their support and
work on the Landscape Toolbox project, of which my dissertation research was a part.
Anne Axel, Andrea Dechner, and Erica Mize patiently listened as I talked through
aspects of my project, offered valuable suggestions, and, most importantly, helped me
keep my sense of humor.

The field data used in this dissertation were collected through the tireless work of
N. Bentley, K. Colson, K. DiChristina, A. Lucas, M. McCarty, and P. Murphy. Data
entry was performed by A. Asada and M. Cook. J. Russel and G. Mann of the Bureau of
‘ Land Management provided us with the field data from the Wildhorse allotment. The
Shoshone district of BLM provided logistical assistance during field data collection.

The research presented here was funded by the Idaho Chapter of The Nature
Conservancy, the M. J. Murdock Charitable Trust, the Lava Lake Foundation for Science
and Conservation, and The Nature Conservancy’s Rodney Johnson and Katherine
Ordway Science Endowment. My doctoral work at Michigan State University was
generously supported by a fellowship from the University’s Environmental Science and
Policy Program.

Finally, I would like to acknowledge and express my deepest gratitude for the
support and encouragement of my family who made great sacrifices to enable me to

pursue this dream.

TABLE OF CONTENTS

LIST OF TABLES ........................................................................................................... viii
LIST OF FIGURES ........................................................................................................... ix
CHAPTER 1 - INTRODUCTION ...................................................................................... l
Remote-sensing to the Rescue? ....................................................................................... 2
The Importance of Scale for Rangeland Decision-making ............................................. 7
Object-based Image Analysis .......................................................................................... 9
Dissertation Objective .................................................................................................. 11
Study Region ................................................................................................................. 12
Layout of the Dissertation ............................................................................................. I 4
References ..................................................................................................................... I 7

CHAPTER 2 - MULTIVARIATE CORRELATIONS BETWEEN IMAGERY AND
FIELD MEASUREMENTS ACROSS SCALES: COMPARING PIXEL

AGGREGATION AND IMAGE SEGMENTATION ..................................................... 23
Abstract ......................................................................................................................... 23
Introduction .................................................................................................................. 24
Study Area ..................................................................................................................... 28
Methods ......................................................................................................................... 31

Field Data Collection ................................................................................................ 31
Image Acquisition and Pre-processing ..................................................................... 32
Image Segmentation and Aggregation ...................................................................... 33
Statistical Analysis .................................................................................................... 34
Field Sample Variogram Analysis ............................................................................ 36
Spatial Predictions of Site Attributes ........................................................................ 37
Results ........................................................................................................................... 3 7
Canonical Correlations between Field and Image Data ............................................ 37
Correlations between Canonical Variables and Original Field Variables ................ 38
Spatial Predictions of Vegetation Parameters at Different Scales ............................ 41
Discussion ..................................................................................................................... 41
Conclusion .................................................................................................................... 45
Acknowledgments .......................................................................................................... 46
References ..................................................................................................................... 58

CHAPTER 3 - SPATIAL DEPENDENCE OF PREDICTIONS FROM IMAGE
SEGMENTATION: A VARIOGRAM-BASED METHOD TO DETERMINE
APPROPRIATE SCALES FOR PRODUCING LAN D-MANAGEMENT

INFORMATION ............................................................................................................... 62
Abstract ......................................................................................................................... 62
Introduction .................................................................................................................. 63
Study Area ..................................................................................................................... 69
Methods ......................................................................................................................... 70

Field Data Collection ................................................................................................ 70

vi

Image Acquisition and Pre-processing ..................................................................... 71

Image Segmentation .................................................................................................. 72
Statistical Analysis .................................................................................................... 73
Results ........................................................................................................................... 79
Discussion ..................................................................................................................... 83
Conclusion .................................................................................................................... 88
Acknowledgments .......................................................................................................... 90
References ................................................................................................................... I 01
CHAPTER 4 - MAKING SPATIAL PREDICTIONS OF ATTRIBUTES OF

RANGELAND ECOSYSTEMS USING REGRESSION KRIGING ............................ 106
Abstract ....................................................................................................................... I 06
Introduction ................................................................................................................ 1 07
Regression Kriging ..................................................................................................... I 10
Study Area ................................................................................................................... 1 15
Methods ....................................................................................................................... I 1 6
Field Data Collection .............................................................................................. 118
Image Acquisition and Processing .......................................................................... 119
Statistical Analysis .................................................................................................. 121
Evaluating Performance Of Predictions .................................................................. 124
Results ......................................................................................................................... 125
Discussion ................................................................................................................... 127
Implications ................................................................................................................ 132
Acknowledgments ........................................................................................................ 133
References ................................................................................................................... 1 49

CHAPTER 5 - DEATH BY A THOUSAND PAPERCUTS: THE ROLE OF
INFORMATION AND ADAPTIVE MANAGEMENT IN THE MANAGEMENT OF

SAGEBRUSH ECOSYSTEMS ...................................................................................... 153
Abstract ....................................................................................................................... 153
Introduction ................................................................................................................ 153
Natural Resource Management Paradigms ................................................................ I 5 6
Conflict Escalation and the Flow of Information in Rangeland Management ........... I 60
Adaptive Management ................................................................................................ 164
Data for Better Management Decisions ..................................................................... 1 66
Will Better Information Improve Rangeland Management? ...................................... 1 6 7
Public Participation in Rangeland Management — Putting Command and Control at
One End of a Management Spectrum ......................................................................... 1 70
Deciding the F arm and Extent of Stakeholder Participation ..................................... I 75
Alternatives to Command-and-Control Management ................................................. 1 77
Conclusions ................................................................................................................. 1 79
Acknowledgements ...................................................................................................... 182
References ................................................................................................................... 183

vii

LIST OF TABLES

Table 2.1. Percent cover variables estimated for each site that were used in the study. .. 48
Table 2.2. Correlations between the eight field variables sampled for this study. ........... 49

Table 3.1. Summary information for the ten successively coarser segmentation levels
used in this study ............................................................................................................... 91

Table 3.2. Variogram characteristics of the predicted and residual values and cross-
validation results for the regression analysis at each segmentation level. ........................ 92

Table 4.1. Summary statistics from the 346 field observations for the three percent cover
variables being predicted across the Wildhorse area: shrub, bare ground, and cheatgrass.
......................................................................................................................................... 134

Table 4.2. Generalized least squares (GLS) models for percent bare ground and percent
cheatgrass (Bromus tectorum L.) predicted from satellite image pixel data summarized by
image objects. ................................................................................................................. 135

Table 4.3. Cross-validation results for the generalized-least-squares regression model
(GLS) and the regression kriging model (RK) predictions of percent shrub, bare ground,
and cheatgrass cover. ...................................................................................................... 136

Table 4.3. Cross-validation results for the generalized-least-squares regression model
(GLS) and the regression kriging model (RK) predictions of percent shrub, bare ground,
and cheatgrass cover. ...................................................................................................... 136

Table 4.4. Proportion of the study area in five percent cover categories for predictions of

percent cover of the three rangeland attributes make via generalized least-squares
regression (GLS) and regression kriging (RK). .............................................................. 137

viii

LIST OF FIGURES

Images in this dissertation are presented in color.
Figure 1.1. Example of image segmentation in object-based image analysis (OBIA). 17
Figure 1.2. Study regions of southern Idaho considered in the following chapters .......... 18
Figure 2.1. Study area in Craters of the Moon National Monument, Idaho, USA ........... 50

Figure 2.2. Image segmentation versus pixel aggregation as a way of filtering information
from remotely-sensed image data. .................................................................................... 51

Figure 2.3. Canonical correlations between field measurements of percent cover and the
tasseled-cap transformed images aggregated to coarser scales by image segmentation and
pixel-based methods .......................................................................................................... 52

Figure 2.4. Correlations across scale levels between the eight original field variables and
the first canonical variables for the object-segmented (a) and pixel-aggregated (b) Ikonos
image. ................................................................................................................................ 53

Figure 2.5. Correlations across scale levels between the eight original field variables and
the first canonical variables for the object-segmented (a) and pixel-aggregated (b)
Landsat image. .................................................................................................................. 54

Figure 2.6. Sample semi-variogram of four percent cover measures from the 159 field
sites in the study area (see Table 1 for key to variable codes) .......................................... 55

Figure 2.7. Maps of the first canonical variable for the image segmentation at different
scales. ................................................................................................................................ 56

Figure 2.8. Effects of native image resolution on the persistence of image objects as an
image is segmented into coarser-scale representations of the original data. .................... 5 7

Figure 3.1. Location and ownership of the Bureau of Land Management’s Wildhorse
allotment in southern Idaho. ............................................................................................. 93

Figure 3.2. Example of a sample (empirical) semivariogram (black points) and the
variogram model (heavy solid line). ................................................................................. 94

Figure 3.3. Empirical semivariogram (dots) for percent bare ground cover in the
Wildhorse Allotment from field observations. ................................................................. 95

Figure 3.4. Changes in correlation with segmentation scale between predictions of
percent bare ground from generalized least-squares (GLS) regression and regression
kriging (RK). ..................................................................................................................... 96

ix

Figure 3.5. Empirical semivariograms for percent bare ground predictions and residuals
from the generalized least squares (GLS) regression models at different levels of image
segmentation. .................................................................................................................... 97

Figure 3.6. Plot of variogram range versus nugget-to-sill ratio (N SR) for the field
measurements of bare ground cover and generalized least-squares (GLS) regression
predictions for the ten segmentation levels considered. ................................................... 98

Figure 3.7. Correlation between predicted and observed percent bare-ground cover for the
10 segmentation levels plotted against the difference between the range of the variogram
model of the predictions and the range of the variogram model derived from the field
observations. ..................................................................................................................... 99

Figure 3.8. Changes in between- and within-object variance for the third tasseled-cap
band as segmentation scale increased. ............................................................................ 100

Figure 4.1. The Wildhorse study area in southern Idaho. ............................................... 138

Figure 4.2. Example of a sample (empirical) variogram (black points) and the variogram
model (heavy solid line) .................................................................................................. 139

Figure 4.3. Empirical variogram and spherical variogram model for the residuals of the
percent shrub regression model. ..................................................................................... 140

Figure 4.4. Empirical variogram and spherical variogram model for the residuals of the
percent bare ground regression model. ........................................................................... 141

Figure 4.5. Empirical variogram and spherical variogram model for the residuals of the
percent cheatgrass regression model ............................................................................... 142

Figure 4.6. Plot of predicted versus observed values for (A) the GLS regression percent
cheatgrass cover model (RMSE 3.63%. Standardized RMSE = 0.272), and (B) the
regression kriged percent bare ground cover model (RMSE=3.42%, standardized RMSE
= 0.256). .......................................................................................................................... 143

Figure 4.7. Plot of predicted versus observed values for (A) the GLS regression percent
bare ground cover model (RMSE 6.78%. Standardized RMSE = 0.8128), and (B) the
regression kriged percent bare ground cover model (RMSE=6.34%, standardized RMSE
= 0.5413). ........................................................................................................................ 144

Figure 4.8. Plot of predicted versus observed values for (A) the GLS regression percent
cheatgrass cover model (RMSE 6.87%. Standardized RMSE = 0.544), and (B) the
regression kriged percent bare ground cover model (RMSE=6.42%, standardized RMSE
= 0.509). .......................................................................................................................... 145

Figure 4.9. Spatial predictions of percent shrub cover in the Wildhorse area. ............... 146

Figure 4.10. Spatial predictions of percent bare ground cover in the Wildhorse area. 147

Figure 4.11. Spatial predictions of percent cheatgrass cover in the Wildhorse area. ..... 148

xi

CHAPTER 1 - INTRODUCTION

Rangelands, ecosystems where the natural vegetation is predominantly grasses,
forbs, and shrubs (Bedell 1998), are undergoing rapid ecological change globally
(National Research Council 1994; Woinarski et al. 2003). Even in the United States
where the majority of rangelands are in public ownership, effective management and
conservation has not been achieved because pervasive threats such as exurban
development, inappropriate livestock grazing, altered fire regimes, and non-native
invasive plants have not been addressed.

One requirement for successful management of natural resources is reliable and
timely information on the status and trend of the ecosystems being managed and human-
ecological interactions that affect those systems (Dietz et al. 2003). Rangeland managers
need ways of integrating information to inform actions taken at site (e. g., ecological site,
stand), landscape (e.g., management unit), and regional (e.g., planning area) scales
(Herrick et al. 2006). To be useful for decision-making, information must not only fit the
needs of managers in terms of content and timing (Sarewitz et al. 2007) but also be
accurate and be accompanied by characterizations of uncertainty in order to hold up to
scrutiny (National Research Council 2008).

However, personnel, budget and time constraints often limit efforts to collect the
data needed for management decision making. Routine approaches used by rangeland
managers to assess rangelands are either based on making extrapolations from a small
number of field samples or from infrequent, cursory “drive-by” surveys. When applied to
larger areas like public grazing allotments, this leads to subjectivity and inconsistency in

results and accuracy that cannot be quantified. It also fuels conflict over land

management because it leads to differing perceptions of condition and trend of the land.
Moreover, much of the available field data is outdated, many areas are difficult to access,
and managers have limited or declining resources to gather new data.

Recognizing these challenges, rangeland managers have, in recent years, begun to
embrace a landscape approach where a focus is placed on measuring and monitoring
rangelands consistently over large areas. Many new methods have been developed to
provide information to planners and managers across landscapes such as using satellite
imagery to detect changes in plant cover or predictive models to forecast future
conditions. However, a significant challenge in implementing these methods in rangeland
landscape management has been appropriately identifying scales of observation that

measure the effects of ecological processes driving rangeland changes.

Remote-sensing t0 the Rescue?

Remote sensing, generally referring to the collection of imagery and synthesis of
information from aerial or satellite imagery, has been held as one of the most promising
ways of generating information for rangeland management decision-making. Hunt et al.
(2003) explored some of the myriad possibilities of using remote sensing in rangeland
management including: assessment and monitoring of rangeland health, estimation of
biomass production for grazing management, and detecting and tracking invasive plant
infestations. There are several appealing features of remote sensing as a means for
generating information for managers. First is that information completely covering large
landscapes can be obtained at a fraction of the cost of sending personnel out in the field

to collect similar data. Second, remote sensing data are seen as objective and repeatable

measurements of landscapes because the sensor is recording radiation reflected from the
earth’s surface. Third, as archives of satellite imagery have grown since the 1970’s,
analysis of imagery is seen as a promising way of monitoring changes in rangeland
ecosystems over time — even enabling retrospective analyses (e. g., Malmstrom et al.
2008). As a result, interest in remote sensing as a means of providing information for
rangeland management is high and much research and development is happening.

In practice, however, remote sensing methods have not yet been widely adopted
into routine decision-making for rangeland management. This is despite the fact that
managers are clamoring for more and better information on rangeland condition and trend
and researchers are continually developing and promoting more precise and accurate
remote-sensing methods and products. There may be several reasons for this.

The first reason is that there remains a lack of awareness among rangeland
managers and ecologists as to the possibilities for implementing remote sensing into
routine management. Most rangeland professionals view remote sensing only as a means
for producing maps of vegetation classes and are not fully aware of the range of other
products available or how they might be used. A report by the National Research Council
(Steering Committee on Space Applications and Commercilization 2001) documented
two “gaps” that have hindered more widespread adoption of remote sensing: 1) a
knowledge gap between remote-sensing researchers and application developers, and 2) a
communications gap between application developers and end users. These gaps resulted
from reliance upon a model of technology transfer which assumed that people who could
benefit from new methods and technologies would naturally see their benefits and apply

the methods and technologies (van Kerkhoff et al. 2006). While there have been some

cases where such a “trickle-down” method has worked, the consensus is that overall this
is an inefficient means for getting remote sensing technologies implemented (Gilruth et
al. 2006).

A second reason is the difficulty in obtaining remotely-sensed data and the level
of technical skill required to analyze imagery. Not counting costs associated with
acquiring imagery, users must navigate non-intuitive websites to discover and obtain
imagery and then deal with esoteric data formats that may require specialized or custom
software to read. Software specific to image analysis is not widely available and most
rangeland professionals get no more than cursory exposure to remote sensing through
their college education and subsequent career training. Compounding this is the fact that
most remotely-sensed data are available as raw images whose value to rangeland
managers is untapped until converted into derived products to meet their needs.

A third factor is that many derived remote-sensing products that have been
proposed for rangeland management are not direct measures of rangeland parameters but
correlated surrogates, and an understanding of the strength of the correlation and its
practical uses and limitations is generally low among rangeland professionals. Vegetation
indices have been widely promoted for use in rangeland management because they are
correlated to photosynthetic activity and can be used to show patterns in greenness over
time (e.g., Dymond et al. 1992; Anderson et al. 1993; Pickup et al. 1994; Hunt et al.
2003). However, what these indices are actually measuring - the amount of radiation (i.e.,
light) reflected from different portions of the electromagnetic spectrum — is often poorly
communicated while emphasis is placed on attributes the indices are correlated with such

as photosynthetic activity or vegetation greenness. Factors that influence the strength of

the correlation (e.g., climatic variability, plant stress, scale), while they have been well
researched in most cases, have not been translated to managers who see only variable
performance of remote sensing products and question their overall reliability and
usefulness. The lack of experience with remote sensing and historically variable
performance may have contributed to a distrust in the technology (Wright et al. 2005).
Recently, efforts have been made to close the knowledge and communications
gaps in remote-sensing adoption and to make information products that are more useful
to managers. F ederally-sponsored programs have fostered the creation of service
providers who transform raw remote sensing data into useful information for decision-
making. One such provider was the University of Arizona’s Rangeview project
(http://rangeview.arizona.edu) which made vegetation index maps and analysis tools
derived frequently fiom the Moderate Resolution Imaging Spectrometer (MODIS)
satellite available to public and private rangeland managers via a webpage interface.
Networks of experts who can translate arcane remote sensing technologies into practical
applications for specific disciplines have also evolved. The National Geospatial
Extension Network (http://geospatialextension.org), built on the more traditional concept
of state-based agricultural extension services and extension agents, is a network of
remote sensing specialists in 14 states whose goal is to encourage the adoption of remote
sensing and other geospatial technologies in agricultural and environmental management
through outreach, education and direct assistance. As the networks of service providers
and channels of communication mature, more efficient communication of the prospects
and possibilities of remote sensing may speed the adoption of these technologies into

rangeland management.

A fourth reason that remote sensing may not have seen wide adoption in routine
rangeland management is that the kinds of information generated through remote sensing
often do not mesh with existing systems and information requirements for rangeland
management — especially within govemment agencies. While there is consensus that
remote sensing should be useful for assessing and monitoring rangeland conditions and
trends over large landscapes, there has yet to be agreement on how this should be
accomplished and which aspects of rangelands can be reliably measured for this purpose.
Additionally, in many agencies regulations mandate the collection of field data according
to prescribed methods and may serve as a disincentive to using remotely-sensed
information. Remotely-sensed data may not qualify for making management decisions
under agency regulations. Additionally, the manner in which field data is required to be
collected may impair its usefulness for training and validating remote-sensing products

A final reason that remote-sensing may not have been widely adopted in
rangeland management has to do with distrust of the technology on the part of rangeland
managers stemming from historically variable performance land managers have
experienced when applying remote-sensing to management decision-making. This may,
on its face, appear to mean that many remote-sensing technologies are not ready for use
in management systems or to be a product of researchers and academics “overselling” the
capabilities of remote sensing technologies to provide reliable information. However,
many of these problems stem from effects of scale that have not been thoroughly

considered or controlled for.

The Importance of Scale for Rangeland Decision-making

Scale is widely recognized as a critical attribute of ecological inquiries that not
only defines what patterns and processes can be measured, but also influences observable
relationships and governs the inferences that can be made from a set of data (Allen et al.
1982; O'Neill et al. 1986b; O'Neill et al. 1989; Wiens 1989). In order for data to be useful
for management decision-making, it must be collected and analyzed at spatial and
temporal scales that are relevant to the processes of interest to managers (O'Neill et al.
1986a) because different patterns can emerge at different scales for almost any ecosystem
(Wiens 1989).

In general terms, scale refers to the grain and extent of observations made in a
study area where grain refers to the finest level of spatial and temporal detail observable
and extent refers to the maximum area under consideration (Turner et al. 1989). Scale is a
characteristic of the set of observations and the choice of scale constrains the patterns and
processes that are observable (Burnett et al. 2003). Grain and extent define the upper and
lower limits of inference because elements of patterns below the grain cannot be detected
and inferences beyond the extent cannot be made without assuming a uniformity of
patterns and processes that is scale independent (Wiens 1989). Information at scales finer
than the observation grain is filtered out and considered noise, and information at scales
larger than the observation extent is also filtered out and becomes context for the
information retained (Wu 1999; Burnett et al. 2003). Thus, observations made with a set
grain and extent (i.e., at a specific scale) have the effect of integrating or smoothing the

dataset to give an interpretable message or signal (Allen et al. 1982; Burnett et al. 2003).

The concept of scale also has ecological interpretations where grain and extent
can be defined in terms of how organisms or ecological processes respond to their
environment with grain being the smallest area to which an organism responds to
heterogeneity in its environment (Addicott et al. 1987; Kotlair et al. 1990) and extent
being the coarsest set of enviromnental patterns to which organisms or ecological
processes react (Farina 1998). From an ecological perspective, scale is relative to an
organism or process and is not an inherent property of the environment (Wiens et al.
1989). Because natural landscapes are a complex product of many different processes and
environmental factors (Peters et al. 2006) that are difficult to characterize when viewed
from arbitrarily-selected spatial and temporal scales (Levin 1992; Burnett et al. 2003), a
significant challenge in ecological studies has been defining scales of observation that
match the relevant ecological scales for organisms or processes of interest.

A crucial decision in producing information for rangeland management is
selecting the appropriate scales of analysis for a particular objective. At a fundamental
level, choices in scale affect the observed variance of the attribute being measured
(Wiens 1989; Home et al. 1995). This can have effects ranging from decreased accuracy
of predictions to complete reversals of relationships between sets of variables (Belsky
1987; F uhlendorf et al. 1999; Beever et al. 2003). Compounding this is the fact that what
may be the best scale for measuring a phenomenon in one location may not be
appropriate someplace else (Fuhlendorf et al. 1999), and the same scale may not be best
for analyzing different ecosystem attributes (Addink et al. 2007). This context and
objective dependence of scale may be one of the root causes of variable performance of

remote-sensing products that has fi'ustrated rangeland managers.

Most studies that have employed remote-sensing have used the de facto scale
determined by the image’s pixel ground dimensions (i.e., resolution). In light of the
ecological definition of scale discussed above, the default resolution of imagery is an
arbitrary scale of observation that may or may not correspond to the scale of ecological
patterns and processes. If observational and ecological scales do happen to match, high-
quality results are possible, but the same scale might not work as well in other landscapes
or at a different point in time where the ecological scales are different. A default image
scale also might be suitable for understanding the distribution of one ecosystem attribute,
but not perform well for other attributes. Clearly, what is needed is a method for easily
determining how to select scales for remote-sensing studies that correspond to the

ecological scales of interest to rangeland managers.

Object-based Image Analysis

Aggregating similar, adjacent pixels into polygons (i.e., objects) as a way of
scaling imagery has been proposed as an alternative method to pixel-based remote
sensing. This approach is called object-based image analysis (OBIA) and research has
shown it to be a robust tool for creating information useful to managers from remotely-
sensed imagery. Because image objects are constructed in reference to information
captured in an image, and not arbitrarily, they have the potential to be defined to match
ecological scales (Hay et al. 2004).

Image segmentation works by grouping similar, neighboring pixels into discrete
regions —- or image objects — such that some measure of variance within the object does

not exceed a specified threshold (Burnett et al. 2003). The intent with segmentation is to

produce image objects that match objects on the ground (Woodcock et al. 1992). Taken
together, a set of image objects that completely tessellate an area into discrete units
constitutes one possible observational scale. There are many different methods for
aggregating pixels into image objects (see Neubert et al. 2006), and within each method
are parameters for controlling the scale that will result. Most significantly, parameters
affecting the degree to which information is filtered from the image — which in turn
affects things like how large the output objects are on average — have a significant effect
on how the image information relates to ground observations and what the resulting
objects can be used for (Figure 1.1). Up to some segmentation level, the information that
is lost is expected to be noise that is irrelevant to the analysis objectives. Past some point,
however, important information is lost, the objects no longer reflect patterns of interest
on the ground, and the accuracy of analysis results may decline (Addink et al. 2007).
Studies have found that OBIA outperforms traditional pixel-based methods in
many situations (e.g., Burnett et al. 2003; Hay et al. 2004; Wang et al. 2004; Luscier et al.
2006; Navulur 2007). While much work has been done to link OBIA with ecological
hierarchy theory (Wu 1999; Hay et al. 2001; Wu et a1. 2002; Burnett et al. 2003; Hay et
al. 2004), there is little objective guidance on how to choose segmentation levels to arrive
at sets of image objects that are of a scale suitable for addressing specific management
information needs. To date, the approach to selecting a set of image objects for a
particular purpose relies on trial and error and the researcher’s knowledge of the system
being studied (Burnett et al. 2003; Navulur 2007). Some authors have started to look for
ways to objectively determine “optimal” levels of segmentation (e.g., Feitosa et al. 2006;

Kim et al. 2006; Addink et al. 2007), but a better understanding of how, in general,

10

OBIA-derived observational scales relate to ecological scaling, and in particular, how
image objects relate to patterns on the ground is necessary to make OBIA a robust and

reliable tool for providing information to rangeland managers.

Dissertation Objective

The purpose of my dissertation was to explore the underlying properties of OBIA that
influence how image objects are defined relative to observable properties of rangeland
ecosystems. My goal was to determine if OBIA could be used to generate observational
scales to match ecological scales in rangeland ecosystems and to explore the potential for
OBIA to generate accurate and repeatable remote-sensing products for rangeland
managers. The specific objectives of my dissertation were to:

I. understand how the relationships between field observations and image
information change as scale increases in OBIA,

2. determine how scaling image data through OBIA changes the spatial dependence
of predictions of ecosystem attributes and see if these changes could be used to
identify appropriate levels of segmentation,

3. demonstrate how OBIA could be used in conjunction with geostatistical
interpolators to produce spatial predictions of attributes of rangeland condition
with quantified uncertainty that would be useful to rangeland managers, and

4. explore some of the management and social factors that ultimately will influence
if and how new techniques like OBIA are implemented in routine management

decision making.

11

Study Region

The work presented in the following chapters was conducted in two study areas in
southern Idaho’s Snake River Plain region (Figure 1.2). The Snake River Plain was
formed as the North American continental plate moved over the Yellowstone Hotspot.
Frequent eruptions from shield volcanoes laid the basalt bedrock that characterizes the
geology of this region of Idaho. The defining features of this landscape are expanses of
unvegetated black rock created from a succession of volcanic eruptions that blanketed the
area with lava flows between 15,000 to 2,000 years ago (Owen 2008). Within the area of
the lava flows, some areas were not covered by the lava, forming kipukas — islands of
vegetation surrounded by lava. Two of these kipukas, Little Park (3,150 ha) and Laidlaw
Park (35,040 ha) were included in the study region. Soils within the areas not covered by
lava flows are mostly aridisols with low organic matter and subsurface accumulations of
either calcium carbonate or clay (Soil Survey Staff 2006a; Soil Survey Staff 2006b). This
region of southern Idaho sees annual precipitation ranging from approximately 25mm to
40mm. Generally, the topography of the region is flat to rolling hills and plateaus, and
elevations range from 1,272m to 1,712m.

The vegetation in the two study areas consists of several different sagebrush
communities occurring along gradients of moisture, elevation, and soil depth. The most
mesic sites with the deepest soils support basin big sagebrush (Artemisia tridentata Nutt.
ssp. tridentata) with an understory of Great Basin wild rye (Leymus cinereus (Scribn. &
Merr.) A. Love) and bluebunch wheatgrass (Pseudoroegeneria spicata (Pursh) A. Love).
Communities of Wyoming big sagebrush (Artemisia tridentata Nutt. ssp. wyomingensis

Beetle & Young) with an understory of bluebunch wheatgrass, Idaho fescue (F estuca

12

idahoensis Elmer), or bottlebrush squirreltail (Elymus elymoides (Ref) Swezey) typify
drier sites. In Little Park and the northern portions of Laidlaw Park and the Wildhorse
allotment, many areas that have previously burned have three-tip sage (Artemisia
tripartita Rydb.)-Idaho fescue communities. Finally, the driest sites with shallow soils
support only low-productivity low sagebrush (Artemisia arbuscula Nutt.) communities.

This 160,000—ha region of Idaho is actively managed on its public and private
lands. Approximately 98.5% of the study region is publically owned. Of that land, the _
Bureau of Land Management (BLM) is the largest public land management agency with
approximately 130,000 ha (82.6%). Lava flows, under the administration of the National
Park Service, account for about 20,000 ha (12.7% of the study region). In 2000, President
Clinton signed Proclamation 7373 expanding the boundary of the Craters of the Moon
National Monument to include the Laidlaw and Little Park kipukas. The State of Idaho
manages approximately 5,000 ha (3.2% of the study region) -— mostly trust lands managed
to produce revenues for Idaho schools through grazing leases. Private land amounts to
around 2,400 ha (1.5%).

The main land-use activities in the study areas are livestock grazing and
recreation. Grazing rights for BLM lands are leased to private land owners under terms
that set the number of livestock and timing and duration of grazing for defined areas (i.e.,
allotments). In the study region, the BLM currently leases grazing rights for the 97,300-
ha Wildhorse and 38,200-ha Laidlaw Park allotments.

Fire is a defining ecological process for this study region. Prior to the era of fire
suppression that began in the early 20‘'1 century, wildland fires (predominantly caused by

lightning strikes) exhibited approximately a 25- to 50-year return interval. Low-intensity

l3

fires maintained relatively open-canopy stands of sagebrush in a mosaic with grasslands.
As a result, fires tended to be smaller in size and less intense than what the region
currently experiences. Fire suppression and the invasion of highly competitive annual
grasses such as cheatgrass (Bromus tectorum L.) have changed the dynamics of fire in
sagebrush ecosystems. Historically intensive grazing and the exclusion of fire have led to
dense stands of sagebrush with depleted understories that are subject to intense fires and
invasion by cheatgrass. Fires are now larger and hotter and many burned areas no longer
regenerate into bunchgrass grasslands but instead transition into new
cheatgrass/ sagebrush states that promote more frequent and more intense fires.

It is in this complex array of natural and political systems that rangeland
management in Idaho takes place. More detailed descriptions of each study area are given

in the following chapters.

Layout of the Dissertation

In the following chapters 1 consider how OBIA of remotely-sensed images can be
used to identify appropriate scales for analysis and develop reliable information for
rangeland managers. Following this introduction chapter, the next two chapters deal with
the mechanics of the OBIA method, the implications of how OBIA can be used to scale
information related to rangeland ecosystems, and how OBIA can be used to determine
appropriate scales for analysis. I follow this with a practical application of OBIA for
rangeland management, and conclude with a look at some factors that control how and if
techniques like those I describe in this dissertation and the resulting data on landscape

condition will be used in public rangeland management decision-making.

14

Chapter two describes the OBIA method in detail and how it can be used to create
hierarchies of information at different scales. I present results showing that the
relationship between field and remotely-sensed data changes with scale, but that,
generally, OBIA is preferable to traditional pixel-based methods for up-scaling
information to look for landscape-level patterns. Finally, I explore how different scales of
information identified through OBIA can be used to generate information pertinent to
management of rangeland ecosystems

Chapter three considers how hierarchies of image objects created through OBIA
can be used to discover which scales are most appropriate for analysis of landscape
patterns. I use, as an example, data on percent bare-ground cover from the Wildhorse
allotment to show how scaling remote-sensing data through OBIA changes the spatial
dependence of predictions of bare ground cover. When combined with geostatistical
predictors, these changes in spatial variance with scale led to robust predictions across a
range of scales. This is a method that can be used in other situations to identify the
different scales of patterns recorded in image data and produce accurate products for use
in rangeland management.

Chapter four highlights an application of OBIA for generating information useful
for rangeland management. I again use the Wildhorse study area to demonstrate OBIA
used with the technique of regression kriging to make spatial predictions for three aspects
of rangeland condition. Regression kriging is a geostatistical predictor that improves on
regression predictions with remote sensing data by taking advantage of the spatial
autocorrelation of field samples. I give an overview of the regression kriging approach

and contrast it against standard linear regression for three rangeland attributes important

15

for assessing rangeland health and managing habitat for sagebrush obligate species like
Sage Grouse (Centrocercus urophasianus): percent shrub, bare ground, and cheatgrass
cover. I conclude with recommendations for interpreting OBIA-based regression kriging
results and implementing it in rangeland management.

The final chapter revisits the issue posed in this introduction of why remote
sensing information is not more widely used in rangeland management decision—making,
but considers it from a broader perspective of how management paradigms influence the
successful incorporation of many kinds of information. I briefly describe the command-
and-control management paradigm used by most public land management agencies and
review concepts of conflict escalation in natural resource management. Using recent
lawsuits from areas near my study region, I illustrate how command-and-control
management, lack of stakeholder engagement and a fixation on regulatory process have
all contributed to the management paralysis currently being experienced. Even though
both sides in the lawsuits argued that better information was needed to make management
decisions, it is uncertain if such information would have been used if available, and it is
also unlikely that its availability would have prevented conflict. Ultimately, for high-
quality information on the status and trend of rangeland landscapes to have an impact on
management decision-making, the current paradigms that control how decisions are made
for public rangelands and who is involved in the process need to change to become more

flexible and inclusive.

l6

 

Figure 1.1. Example of image segmentation in object-based image analysis (OBIA).
Through modifying image segmentation parameters, objects of different scales can be
created. Objective methods are needed to determine which of these scales would be most
appropriate for a given objective. Panel A is a portion of a 4-m resolution Ikonos image
shown as a false-color composite. Black areas are lava flows; the pink region at the top of
the panels is an aspen stand. In panels B, C, and D, the image has been segmented at
successively coarser scales. Note how distinct landscape features such as the landing
strip, recent burn, and aspen stand retain their definition as scale coarsens even though
objects around them have become much larger. All four panels are at a map scale of
1:70,000.

 

Boise
O

0
Twin Idaho
Falls Falls

 

 

 

 

 

 

 

 

 

 

D Study Area Boundaries
Land Ownership
I 7 Other Ownerships
BLM ; 7 _ 7 1. Wildhorse Allotment
Private 2. Laidlaw Park
r7‘ State of Idaho 7 N “I Km
- Lava Flows , . 0 10 20

 

 

 

Figure 1.2. Study regions of southern Idaho considered in the following chapters.

References

Addicott J. F., J. M. Aho, M. F. Antolin, D. K. Padilla, J. S. Richardson, and D. A. Soluk.
1987. Ecological neighborhoods: scaling environmental patterns. Oikos, 49:340-346.

Addink E. A., S. M. de long, and E. J. Pebesma. 2007. The importance of scale in object-
based mapping of vegetation parameters with hyperspectral imagery.
Photogrammetric Engineering and Remote Sensing, 73:905-912.

Allen T. F. H. and T. B. Starr 1982. Hierarchy: perspectives for ecological complexity,
University of Chicago Press, Chicago, IL.

Anderson G. L., J. D. Hanson, and R. H. Haas. 1993. Evaluating landsat thematic mapper
derived vegetation indices for estimating above-ground biomass on semiarid
rangelands. Remote Sensing of the Environment, 45:165-175.

Bedell T. E. 1998. Glossary of terms used in range management. 4th Edition, Society for
Range Management. Direct Press, Denver, Colorado.

Beever E. A., R. J. Tausch, and P. Brussard. 2003. Characterizing grazing disturbance in
semiarid ecosystems across broad scales, using diverse indices. Ecological
Applications, 13: 1 19-136.

Belsky A. J. 1987. The effects of grazing: confounding of ecosystem, community, and
organism scales. The American Naturalist, 129:777-783.

Burnett C. and T. Blaschke. 2003. A multi-scale segmentation/object relationship
modelling methodology for landscape analysis. Ecological Modeling, 168:233-249.

Dietz T., E. Ostrom, and P. Stern. 2003. The struggle to govern the commons. Science,
302:1907-1912.

Dymond J. R., P. R. Stephens, P. F. Newsome, and R. H. Wilde. 1992. Percentage
vegetation cover of a degrading rangeland from SPOT. Intemation Journal of Remote
Sensing, 13: 1999-2007.

Farina A. 1998. Principles and methods in landscape ecology, Chapman & Hall, London,
UK.

Feitosa R. Q., G. A. O. P. Costa, T. B. Cazes, and B. Feijo. 2006. A genetic approach for
the automatic adaption of segmentation parameters. Salzburg University, Salzburg,
Austria.

Fuhlendorf S. D. and F. E. Smeins. 1999. Scaling effects of grazing in a semi-arid
grassland. Journal of Vegetation Science, 10:731-738.

19

Gilruth P., S. Kalluri, J. W. Robinson, J. Townshend, F. Lindsay, P. Davis, and B. J. Orr.
2006. Measuring performance: moving NASA earth science products into the
mainstream user community. Space Policy, 22:165-175.

Hay G. J .and D. J. Marceau. 2004. Multiscale object-specific analysis (MOSA): an
integrative approach for multiscale landscape analysis.in SM de Jong and FD van der
Meer, editors. Remote Sensing and Digital Image Analysis: including the spatial
domain. Kluwer Academic Publishers, Dordecht, Germany.

Hay G. J ., D. J. Marceau, P. Dube, and A. Bouchard. 2001. A multiscale framework for
landscape analysis: object-specific analysis and upscaling. Landscape Ecology,
16:471-490.

Herrick J. E., B. T. Bestelmeyer, S. Archer, A. J. Tugel, and J. R. Brown. 2006. An
integrated framework for science-based arid land management. Journal of Arid
Environments, 65:3 19-335.

Home J. K. and D. C. Schneider. 1995. Spatial variance in ecology. Oikos, 74: 1 8-26.

Hunt E. R. Jr., J. H. Everitt, J. C. Ritchie, M. S. Moran, T. D. Booth, G. L. Anderson, P.
E. Clark, and M. S. Seyfried. 2003. Applications and research using remote sensing
for rangeland management. Photogrammetric Engineering and Remote Sensing,
69:675-693.

Kim M. and M. Madden. 2006. Determination of optimal scale parameter for alliance-
level forest classification of multispectral IKONOS images. International Archives of
Photograrnmetry, Remote Sensing and Spatial Information Sciences, Salzburg,
Austria.

Kotlair N. B. and J. A. Wiens. 1990. Multiple scales of patchiness and patch structure: a
hierarchical framework for the study of heterogeneity. Oikos, 59:253-260.

Levin S. A. 1992. The problem of pattern and scale in ecology. Ecology, 73:1943-1967.

Luscier J. D., W. L. Thompson, J. M. Wilson, B. E. Gorham, and L. D. Dragut. 2006.
Using digital photographs and object-based image analysis to estimate percent ground
cover in vegetation plots. Frontiers of Ecology and the Environment, 4:408-413.

Malmstrom C. M., H. S. Butterfield, C. Barber, B. Dieter, R. Harrison, J. Qi, D. Riano,
A. Schrotenboer, S. Stone, C. J. Stoner, and J. Wirka. 2008. Using remote sensing to
evaluate the influence of grassland restoration activities on ecosystem forage
provisioning services. Restoration Ecology, in press. -

National Research Council 2008. Public participation in environmental assessment and
decision making, National Academies Press, Washington, DC.

National Research Council 1994. Rangeland health: new methods to classify, inventory,
and monitor rangelands, National Academy Press, Washington, DC.

20

Navulur K. 2007. Multi-spectral image analysis using the object-oriented paradigm, CRC
Press, Taylor and Francis Group, LLC, Boca Raton, FL.

Neubert M., H. Herold, and G. Meinel. 2006. Evaluation of remote sensing image
segmentation quality - further results and concepts. International Archives of
Photograrnmetry, Remote Sensing and Spatial Information Sciences, Salzburg,
Austria.

O'Neill R. V., D. L. Deangelis, J. B. Waide, and T. F. H. Allen 1986a. A hierarchical
concept of ecosystems, Princeton University Press, Princeton, New Jersey.

O'Neill R. V., A. R. Johnson, and A. W. King. 1989. A hierarchical framework for the
analysis of scale. Landscape Ecology, 3:193-205.

O'Neill R. V., A. R. Johnson, and R. V. King 1986b. A hierarchical concept of
ecosystems, Princeton University Press, Princeton, New Jersey.

Owen, D. E. Geology of Craters of the Moon. 2008. Arco, ID, Craters of the Moon
National Monument, National Park Service.

Peters D. P. C., J. Yao, L. F. Huenneke et al. 2006. A framework and methods for
simplifying complex landscapes to reduce uncertainty in predictions. Pages 131-146
in J Wu, KB Jones, and OL Loucks, editors. Scaling and uncertainty analysis in
ecology: methods and applications. Springer, Dordrecht, The Netherlands.

Pickup G., G. N. Bastin, and V. H. Chewings. 1994. Remote-sensing-based condition
assessment for nonequilibrium rangelands under large-scale commercial grazing.
Ecological Applications, 4:497 -5 l 7.

Sarewitz D. and R. A. Pielke. 2007. The neglected heart of science policy: reconciling
supply of and demand for science. Environmental Science and Policy, 10:5-16.

Soil Survey Staff. Keys to soil taxonomy, 10th ed. 2006a. Washington, D.C., USDA
Natural Resource Conservation Service.

Soil Survey Staff. U.S. general soil map (STATSGOZ) for Idaho. USDA Natural
Resource Conservation Service . 7-5-2006b. 3-3-2009b.

Steering Committee on Space Applications and Commercilization, National Research
Council. Transforming remote sensing data into information and applications. 2001.
Washington, D.C., National Academy Press.

Turner M. G., R. V. O'Neill, R. H. Gardner, and B. T. Milne. 1989. Effects of changing
spatial scale on the analysis of landscape pattern. Landscape Ecology, 3:153-162.

van Kerkhoff L. and L. Lebel. 2006. Linking knowledge with action for sustainable
development. Annual Review of Environment and Resources, 31:445-477.

21

Wang L., W. P. Sousa, and P. Gong. 2004. Integration of object-based and pixel-based
classification for mapping mangroves with IKONOS imagery. Intemation Journal of
Remote Sensing, 25:5655-5668.

Wiens J. A. 1989. Spatial scaling in ecology. Functional Ecology, 3:385-397.

Wiens J. A. and B. T. Milne. 1989. Scaling of 'landscapes' in landscape ecology, or,
landscape ecology from a beetle's perspective. Landscape Ecology, 3:87-96.

Woinarski J. C. Z. and A. Fisher. 2003. Conservation and the maintenance of biodiversity
in the rangelands. The Rangelands Journal, 25: 157-171.

Woodcock C. and V. J. Harward. 1992. Nested-hierarchical scene models and image
segmentation. International Journal of Remote Sensing, 13:3167-3187.

Wright D. L. Jr., V. P. Rasmussen, Jr., and R. D. Ramsey. 2005. Comparing the use of
remote sensing with traditional techniques to detect nitrogen stress in wheat. Geocarto
International, 20:63-68.

Wu J. 1999. Hierarchy and scaling: extrapolating information along a scaling ladder.
Canadian Journal of Remote Sensing, 25:367-380.

Wu J. and J. L. David. 2002. A spatially explicit hierarchical approach to modelling
complex ecological systems: theory and applications. Ecological Modeling, 153:7-26.

22

CHAPTER 2 - MULTIVARIATE CORRELATIONS BETWEEN IMAGERY AND
FIELD MEASUREMENTS ACROSS SCALES: COMPARING PIXEL
AGGREGATION AND IMAGE SEGMENTATION

Abstract

To successfully use remotely-sensed data in landscape-level management, questions as to
the relevance of image data to landscape patterns and optimal scales of analysis must be
addressed. Object-based image analysis, which segments image pixels into homogeneous
regions, or objects, has been suggested as a way to increase accuracy of remotely-sensed
products, but little research has gone into how to determine sizes of image objects with
regard to scaling of ecosystem properties. We looked at how segmentation of a high-
resolution Ikonos and medium-resolution Landsat image into successively coarser objects
affected the multivariate correlation between image data and eight field measurements of
percent cover in a sagebrush ecosystem. For comparison we also looked at changes in
correlation as the image data were aggregated into larger square pixels. We found that
canonical correlations between field and image data were similar at the finest scales, but
higher for image segmentation than pixel aggregation for both images when scale
increased. For image segmentation, correlations between the canonical variables and the
original field variables were invariant with respect to size of the image objects,
suggesting linear scaling of vegetation cover in our study system. We detected a scaling
threshold with the Ikonos segmentation that was confirmed with a semi-variogram of the
sample data. Below the threshold interpretation of the canonical variables is consistent:
scale levels differ primarily in the amount of detail they portray. Above the threshold, the
meaning of the canonical variables changed. This approach proved useful for evaluating

the overall utility of an image to address an objective, and identified scaling limits for

23

analysis, but selection of appropriate scale for analysis will ultimately depend on the

objective being considered.

Introduction

Landscapes are composed of layers of patterns that occur from a variety of natural
and human-driven processes operating over vastly different spatial and temporal
frequencies (O'Neill et al. 1986; Wiens 1989). Remotely-sensed data (i.e., imagery)
capture information on the patterns related to many of these processes, but the challenge
comes in defining appropriate scales of analysis such that patterns relevant to an
objective can be: 1) separated out from the background noise generated at finer scales,
and 2) placed in context of processes operating at coarser scales (Wu 1999; Wu 2004).
The choice of a particular image carries implications for the types of patterns that can be
observed and also the scales over which the observations can occur. Thus, knowledge of
how the relationship between field and image data changes across scales and an
awareness of scaling thresholds (Wiens 1989) in a study landscape is necessary to select
both imagery and analysis scale for any particular application.

Scale is often operationally discussed in terms of the resolution and extent of
observations as the parameters of a given view of the natural world (Burnett et al. 2003).
Recognizing that scale is a product of the process of observation, a more useful definition
for scale in the context of ecological research is the unit of time and space over which a
phenomenon can be measured to provide meaningful results (Allen et al. 1982). When
observations of a natural system are made at a fixed resolution as with imagery, much

information about processes happening at below the resolution of a single pixel is lost

24

(Wu 1999; Burnett et al. 2003). Thus the scale at which the object or pattern of interest
occurs on a landscape must be considered when selecting appropriate image sources and
deciding on how to analyze the imagery.

Strahler et al. (1986) described two models for representing vegetation parameters
in image data based on the relationship between the size of a pixel and the size of the
pattern or object on the ground. A low-resolution (LRes) model occurs when the pixel is
the same size or larger than the ground object. Alternatively, a high-resolution (HRes)
model occurs when the pixels are smaller than the object or pattern of interest. By this
definition, any image can be both a high- and low-resolution representation of the area
within its extent depending on what patterns or objects are being observed. The choice of
whether to use a LRes versus HRes mode of analysis will heavily influence the type of
sensor and means of image analyses that are appropriate.

Both the LRes and HRes models present challenges with regard to selecting the
appropriate scale for analysis. With a LRes model, the most commonly recognized scale
problem is trying to detect objects or patterns that occur at a finer grain than the
resolution of the image data (Fisher 1997). This has prompted much research into
techniques to elucidate the composition of individual pixels (Atkinson 2004; Foody
2004). Less commonly acknowledged, though, is the problem with HRes models of
excessive heterogeneity — that is, the resolution of the image data is too high and the
relationships between field and image data are obscured by extraneous noise. Various
methods have been advanced to filter out signals recorded by image data from irrelevant

information including semivariance analysis (Meisel et al. 1998), wavelet analysis (Dong

25

et al. 2008), and image segmentation (Burnett et al. 2003; Hall et al. 2004; Woodcock et
al. 1992)..

Aggregating pixels into objects via image segmentation has been proposed as a
way of filtering out excessive heterogeneity in images to increase accuracy of image-
derived products. Wang et al. (2004) reported classification accuracies with image
objects were higher than the same classification method using a pixel-based approach.
Kim and Madden (2006) found that accuracy of forest-type classification increased as
Ikonos imagery was segmented into coarser objects until a threshold was crossed where
accuracy decreased again. Addink et al. (2007) found that an optimal level of image
segmentation could be identified that led to the highest accuracy of estimating leaf area
index and biomass, but that the optimal level of segmentation was different for each
parameter.

Image segmentation works by grouping similar, neighboring pixels into discrete
regions — or image objects -— such that some measure of variance within the object does
not exceed a specified threshold (Burnett et al. 2003). The intent with segmentation is to
produce image objects that match objects on the ground (Woodcock et al. 1992). There
are many different methods for aggregating pixels into image objects (see Neubert et a1.
2006), and within each method there are many adjustable parameters. Most significantly,
parameters affecting how much information is filtered from the image — which in turn
affects things like how large the output objects are on average - have a significant effect
on how the image information relates to ground observations and what the resulting
objects can be used for. Up to some segmentation level, the information that is lost is

expected to be noise that is irrelevant to the analysis objectives. Past a given threshold,

26

important information is lost and the accuracy of analysis results may decline (Addink et
al. 2007). The choice of parameters for how to aggregate pixels into objects, however, is
subjective (Wang et al. 2004) and usually involves a trial and error process for deciding
which set of objects best represents what the person doing the analysis interprets as
meaningful patches on the ground (Burnett et al. 2003; Feitosa et al. 2006; Navulur
2007)

Some authors have sought optimal levels of segmentation that maximize accuracy
of results with regard to a single response variable. Addink et al. (2007) found different
levels of segmentation produced the most accurate results for estimates of leaf-area index
and biomass. Feitosa et al. (2006) used a genetic algorithm to determine optimal
segmentation parameters for matching known landscape patterns for object classification.
Luscier et al. (2006), using a manual, iterative process, identified the segmentation level
that resulted in objects that gave the highest coefficients of agreement between field plot
photos and control data. Kim and Madden (2006) used variance of pixel spectral values
within an image object to determine the level of segmentation that would give the highest
accuracy for classifying forest types. Wang et al. (2004) used Battacharya Distance to
determine the segmentation level at which training data most closely matched image data
for mapping mangrove forests.

Addink et al. (2007), though, proposed that the optimal scale of observation (i.e.,
level of segmentation or size or image objects) for a given objective would depend on: 1)
the spatial heterogeneity of the landscape, and 2) the spatial and temporal extent and
frequency of the phenomenon of interest and the processes responsible for it. They also

observed that the optimal segmentation level may not be the same for all measures of a

27

system. F eitosa et al. (2006) suggested an automated approach for selecting segmentation
parameters for land cover classification by measuring fit against a reference set of objects
manually identified from imagery. To best identify the optimal scales of image
segmentation and to aid in judging the usefulness of a particular image for answering a
management question, a better understanding of how the relationship between field
measurements and image data changes as the image is aggregated by segmentation is
needed.

In this paper we looked at how segmentation of an image into successively
coarser objects affects the relationship between image and field data. Specifically, we
tested the multivariate correlation between eight field measurements of percent cover in
sagebrush ecosystems to a high-resolution Ikonos satellite image and a moderate-
resolution Landsat image both of which were aggregated into successively larger image
objects. For comparison we also looked at how the relationship between the field data
and images changed as the image data was aggregated into larger square pixels. In
particular, we were interested in whether: 1) the correlation between field and image data
would change with the scaling of image objects, 2) image segmentation would produce
higher correlations between field and image data than would pixel-based aggregation, and
3) changes in the multivariate correlation structure as scale changed would indicate

appropriate ranges of scale for further land cover classification work.

Study Area

Our study was conducted in sagebrush-steppe habitats within the Craters of the

Moon National Preserve in southern Idaho, USA (Figure 2.1). The Craters of the Moon

28

area was created during at least eight volcanic eruptive periods that blanketed portions of
the Snake River Plain with basaltic lava flows between 15 ,000 to 2,000 years ago (Owen
2008). Within the area of the lava flows, some areas were not covered by the lava and
formed kipukas — islands of vegetation surrounded by lava. Two of these kipukas, Little
Park (3,150 ha) and Laidlaw Park (3 5,040 ha) were sampled for this study.

In 2000, approximately 280,000 ha of the National Preserve, mostly under the
administration of the US Department of Interior Bureau of Land Management (BLM),
were added to the original Craters of the Moon National Monument. The Little and
Laidlaw Park areas are predominantly BLM lands with private inholdings. Private lands
were not considered in this study.

Within the parks, vegetation consisted of sagebrush-steppe communities in
various successional states. Precipitation in the study area ranges, on average, from
approximately 25cm to 40cm per year. In the absence of disturbance, plant communities
were historically dominated by basin big sagebrush (Artemisia tridentata Nutt. ssp.
tridentada), Wyoming big sagebrush (Artemisia tridentata Nutt. ssp. wyomingensis
Beetle & Young), or three-tip sagebrush (Artemisia tripartita Rydb.) with understories of
deep-rooting bunchgrasses such as Idaho fescue (F estuca idahoensis Elmer), bluebunch
wheatgrass (Pseudoroegneria spicata [Pursh] A. Love), needle and thread grass
(Hesperostipa comata [Trin. & Rupr.] Barkworth), and indian ricegrass (A chnatherum
hymenoides [Roem. & Schult] Barkworth).

Management practices (e.g., grazing) and wildland fire are the main processes that
cause change in plant community. Long-term, intensive grazing can lead to increasing

cover of shrubs, Sandberg’s bluegrass (Poa secunda J. Presl), and cheatgrass (Bromus

29

tectorum L.) and decreasing cover of deep-rooting native bunchgrasses and forbs.
Wildland fires reduce shrub communities to grasslands which initially are recolonized by
re-sprouting shrubs like three-tip sagebrush, green rabbit brush (Chrysothamnus
viscidiflorus [Hook.] Nutt.) and gray rabbit brush (Ericameria nauseosa [Pall. ex Pursh]
G.L. Nesom & Baird), and eventually transition back to climax sagebrush species.
Heavily disturbed sites are dominated by cheatgrass and annual forbs which also spread
into adjacent areas.

The Laidlaw and Little Park areas have seen intensive use, disturbance, and
management over the last 100 years. The rugged, sparsely vegetated lava flows make
good natural barriers to livestock, so the kipukas of this area have a long history of sheep
and cattle grazing. Today, private land owners lease the right to graze sheep and cattle in
the study area from BLM which determines the number of livestock that the area can
sustain and monitors compliance. While wildland fire has always been a part of this
ecosystem, fires have become more frequent in recent years. Over 10,250 ha of the two
parks burned between 2005 and when the sampling was completed for this study in 2007
— approximately 27% of the vegetated portion of the study area. BLM has carried out
extensive fire rehabilitation efforts throughout the study area including reseeding
bunchgrasses, forbs, and shrubs within many burns and deferring grazing for several

years on burned areas.

30

Methods

Field Data Collection

Between May 22 and August 22, 2007 we sampled vegetation on 159 sites in the
study area (21 in Little Park and 138 in Laidlaw Park) following the line-point intercept
method of Herrick et al. (2005). We recorded all top canopy and sub-canopy vegetation
and recorded soil surface type every meter along three 50m transects at each site. From
these data we estimated percent cover of each plant species and soil surface type. Cover
estimates were generalized to 12 measurements per site that were considered to be
representative of the character and condition of each site. From the 12 possible variables
per site, we removed those that were highly correlated (r2 > 0.9, n=159) and were left
with the eight variables used in this study (Table 2.1, Table 2.2).

Sample sites were selected with the aid of two scales of image objects derived
from a 2006 ASTER image. A set of coarse-scale image objects was used as sampling
strata. Fine-scale object polygons were created inside each coarse-scale object and a
number of the fine-scale objects randomly selected for sampling. Our assumptions in
using this approach were: 1) the plant community within a fine-scale object could be
considered homogeneous with regard to our sampling objectives, and 2) sampling within
an image object would ensure that we did not sample across ecological boundaries.

The location of sample sites was recorded with a GPS unit and differentially-
corrected. Average estimated error after differential correction was 4.13 m (2.39 m SD)

or approximately the size of one pixel on the imagery used in our analysis (see below).

31

Image Acquisition and Pre-processing

We acquired a GeoEye Ikonos 4m multi-spectral image for the study area on July
21, 2007, and a Landsat Thematic Mapper 5 image from August 10, 2007. We used the
atmospheric-correction method of Chavez (1996) to convert the image values to top-of-
atrnosphere reflectance for each image, and applied a tasseled-cap transformation using
the coefficients for the IKONOS sensor presented by Home (2003) and the Landsat
coefficients of Crist et al. (1986b).

The tasseled-cap transformation is a fixed linear combination of spectral bands
that ordinates reflectance data into new, orthogonal axes with specific interpretations
(Crist et al. 1986a; Jensen 1996). The first tasseled-cap band indexes overall brightness of
light reflected from the earth’s surface. The second band relates to the greenness of the
surface, and the third band has been interpreted as wetness or “yellowness” of the site.
Because it produces three uncorrelated bands of information that relate strongly to
vegetative site characteristics, the tasseled-cap transformation has been used successfully
in many vegetation mapping exercises. Also, because band four of the tasseled-cap
transformation contains much of an image’s “random noise” and generally does not
correlate strongly to surface features, the transformation has been used as a means to
improve image quality. Additionally, Addink et al. (2007) reported that the use of
correlated bands for image segmentation could inadvertently overweight the variability
that the bands represent on the ground. For all subsequent analyses we used tasseled-cap

bands one through three.

32

mire Segmentation; and Aggregation

We used Definiens Developer 7.0 (http://www.definiens.com/) to segment the
tasseled-cap images into image object polygons. Definiens implements the fractal net
evolution algorithm (FNEA) of Baatz and Schéipe (2000) to segment images. This
algorithm uses information recorded for each pixel (e. g., spectral reflectance or ancillary
data like elevation in either its original form or transformed as in the tasseled-cap) to first
group adjacent pixels into objects and then merge objects based on their similarity. The
FNEA evaluates the difference between two adjacent objects (pixels at first) and makes
the decision to merge them into a single region based on whether or not the increase in
local heterogeneity after the merge would exceed a limit imposed by a unitless scale
parameter (Baatz et al. 2000). The result is a minimization of the average object
heterogeneity across the image (Burnett et al. 2003). The unitless scale parameter is used
to set the average size of the image objects (Baatz et al. 2000), but also has the effect of
controlling the amount of heterogeneity allowable in the image objects. Increasing the
scale parameter yields, on average, larger objects with greater internal variability.
Although, it is important to note that the contrast between objects is an important
determinant of whether or not they will be merged as the scale parameter is increased.
With the FNEA, small, distinct objects can persist as image objects while surrounding
objects become large.

To create sets of image objects at different scales, we varied the scale parameter
of the segmentation algorithm. Each of the tasseled-cap bands was weighted equally in
the segmentation, and all other segmentation parameters were kept constant at default

values. We created 27 different segmentations of the Ikonos image ranging from the

33

finest-scale set with a median object size of 0.837 ha to the largest set with a median
object size of 1,034 ha (Figure 2.2). For the Landsat image, we created 12 segmentations
ranging in median object size from 9.61 ha to 1,089 ha.

For comparison to the image segmentation method, we also created simple
aggregates of the tasseled-cap pixels of the Ikonos and Landsat data to 13 different,
coarser-resolution images (Figure 2.2). Each aggregation was done from the original-

resolution tasseled-cap images.

S_tatistical Analysis

To measure the association between field and image data at each segmentation
and pixel-aggregation level, we used a multivariate canonical correlation analysis.
Canonical correlations establish the strength of association between a linear combination
of one set of variables and a linear combination of the variables of a second data set with
the goal of concentrating the correlation between the datasets into as few pairs of
canonical variables as possible (Johnson et al. 2002). For each segmentation and pixel
aggregation level, we used the first three tasseled-cap band values as our set of image
variables. In the case of segmentation aggregation, we used the average value of all pixels
within the object for each tasseled-cap band. For the pixel aggregation method, we used
the values for the first three tasseled-cap bands recorded in each pixel. Our field
measurement data set at each aggregation level consisted of the eight variables listed in
Table 2.1. When more than one sample point occurred in an object or image pixel, we
took the average of the field measurements. The canonical correlation analyses were

performed in R 2.4.1 (http://www.r-project.org) following Johnson and Wichem (2002).

34

In canonical correlation, there can be as many canonical correlates as the
minimum number of variables of either input dataset — three in our case. The first
canonical correlate is defined to contain the maximum amount of correlation between the
data sets. Additional canonical correlates are defined to represent additional association
between the variable sets but are uncorrelated with previous canonical correlates.
Assuming multivariate normality of both datasets, Johnson and Wichem (2002) describe
the following test for determining the significance of each canonical correlate from a

sample:

1 p
X(2p-k)(q—k)(a) < “(n ‘1‘”‘2’(P + q + 1)]111 HG “ 10:2) (1)

i=k +1
where n is the number of observations, p is the number of variables in the first data set, q

is the number of variables in the second dataset, k is the number of canonical correlates,

2 2
P,- is the square of the ith canonical correlation, and X( p_k)(q-k) (a) is the upper

(1000i)th percentile of a chi-square distribution with (p-k) and (q-k) degrees of freedom.
For each segmentation and pixel-aggregation level, we tested the significance of the three
canonical correlates at the a = 0.05 level.

To interpret the meaning of the canonical variables, each canonical variable can
be correlated back to the measurements of the original variables, giving a correlation
between the canonical variable and each original variable. While this approach does not
account for the joint effect of variables on the canonical variable or original variables that
are highly correlated, it provides a means of identifying which of the original variables
are contributing to the canonical correlation (Johnson et al. 2002). Additionally, these

correlations standardize the relationship between the canonical variable and the original

35

field variables, allowing us to compare the effect (albeit in only a univariate sense) of the

original variables across levels of segmentation and pixel aggregation.

Field Sample Vamgram An_al+ysis

To further explore whether patterns in the scaling of the Ikonos and Landsat
images reflected properties of the study area or just artifacts of the image segmentation
processes, we constructed sample semi-variograms fi'om the original field variables using
the coordinate values of our sample points. The sample semi-variance is a measurement
of the spatial autocorrelation of the field observations that are a specified distance (or
spatial lag) apart (Fortin et al. 2005). Plotting the semi-variance over a number of spatial
lags yields the semi-variogram which is used to determine how much of the observed
variation in the data are explainable by spatial autocorrelation. Semi-variance generally
increases with lag distance until a point is reached — called the sill - where the spatial
autocorrelation is zero (i.e., the distance at which samples are considered statistically
independent). Patterns of semi-variance against lag distance that do not increase
monotonically to the sill can indicate the presence of cyclic spatial structures or multi-

scale patterns in the data (Radeloff et al. 2000; Bellier et al. 2007). We calculated the

sample semi-variance 7901) at any spatial lag h following Fortin and Dale (2005) as:

7‘02 hmflze) z<x +h)l (2)

where z is the value of the field variable at sample point i, and n(h) is the number of pairs

of sampling points located a distance of h from each other. To compare the segmentation

36

results to the sample semi-variograrns, we used the mean distance between object

centroids for the segmentation levels being considered.

Spatial Predictions of Site Attributes

For any given segmentation or pixel aggregation level, the set of linear
coefficients that define how the sample tasseled-cap data were transformed into the
canonical variables can be obtained and applied to the original, full tasseled-cap datasets
to make map predictions of the canonical variables. We selected four Ikonos
segmentation levels representative of different scales to illustrate how spatial predictions
and interpretations of the canonical variable can change across scales. At each
segmentation level, we applied the coefficients for the first canonical variable to the mean
tasseled-cap values for each image object. This yielded a predicted first canonical
variable for each image object across the study area. We created maps of the first
canonical variable and interpreted the mapped canonical variable using the correlations

between the field variables and the canonical variable at those scale levels.

Results

C_anonicaL Correlations between Field and Image Data

Only the first two canonical correlates were significant for both the Ikonos and
Landsat segmentations at all scale levels. For pixel aggregation, though, all three
canonical correlates were significant at all aggregation levels. The third canonical

correlations for pixel aggregation, however, were generally weak across all scales

37

(average r2 of 0.099 and 0.109 for Ikonos and Landsat, respectively). Patterns of
canonical correlations across scale levels and patterns of the correlation between the
canonical variables and the original field variables were generally consistent across the
three canonical variables. For the purposes of comparing the segmentation and pixel
aggregation between Ikonos and Landsat imagery, we report here only the first canonical
correlate for each method.

At the finest scales, the first canonical correlations for Ikonos were similar
between the segmentation and pixel-aggregation methods, r2=0.763 and r2=0.779,
respectively (Figure 2.3). The first canonical correlations for the pixel aggregation
method decreased as the image was aggregated, while the first canonical correlations for
image segmentation remained high, gradually increasing with increasing size of the
Ikonos image objects.

Canonical correlations for the object-segmented Landsat image were higher than
the pixel-aggregated image at all but the finest scales (Figure 2.3), but the difference
between image segmentation and pixel aggregation was less than was observed for the
Ikonos image. Canonical correlations for the object-segmented images also increased as
median object sizes increased; whereas the canonical correlations for the pixel-
aggregated data generally decreased but also exhibited a high degree of variation as scale

increased.

Correlatior§ between Canonical Variables and Original Field Variables

Correlations of the Ikonos-segmented canonical variables with the field variables

were relatively invariant with increasing scale as median object size increased until a

38

scale of about 800ha where the correlations abruptly switched signs (Figure 2.4a). Below
the apparent scale threshold at 800ha, the first canonical variable had high positive
correlations with percent bare ground and forb cover and high negative correlations with
percent shrub and total canopy cover. The invariance of the correlation structure below
the threshold suggested that measures of percent cover for the communities in our study
area scaled linearly up until a median object size of approximately 800ha. Once the scale
threshold was crossed for the Ikonos segmentation, the signs of the canonical-to-field-
variable correlations reverted back to their original sign. Past the threshold, however, the
correlation structure was different than before, with the first canonical variable being
most strongly related to ground cover between plant canopies, litter, shrub, and total
canopy cover.

Pixel aggregation of the Ikonos image produced a canonical-to-field-variable
correlation structure that showed a high degree of variation as pixel size increased, but
little discemable pattern (Figure 2.4b). Correlation of the field variables with the image
data is very sensitive to the scale of the analysis, and the choice of scale has a large
influence on the interpretation of the canonical variable. The scaling threshold observed
for the segmentation method was not observable with the Ikonos pixel aggregation
method results.

Segmentation of the Landsat image produced a field-to-canonical-variable
correlation structure that was stable across all the scales we examined (Figure 2.5a) -
again suggestive of linear scaling. The structure of the correlations was similar to that
observed for the below-threshold Ikonos segmentation. The first canonical variable had

high positive correlations to percent bare ground and percent forb cover, and high

39

negative correlations to percent shrub cover and total canopy cover. In addition, however,
the Landsat first canonical variable also showed a stronger correlation with percent cover
of annual grasses (average across scales r2=-0.472 [range -0.680 to -0.348]) than did the
Ikonos analysis (average across scales r2=-0.296 [range -0.450 to 0323]).

For the Landsat image, however, pixel aggregation yielded more
consistent canonical-to-field-variable correlations than did the Ikonos pixel-aggregation
results (Figure 2.5b). However, the Landsat pixel aggregation method produced scale
regions where the signs of the correlations flipped in a manner similar to the scale
threshold identified in the Ikonos segmentations. However, because the correlation
structure remained the same, only backwards, in each of these instances, we concluded
that this was an artifact of the pixel aggregation method and did not signify scaling
thresholds for this system. The relative strength of the correlations between the field data
and the Landsat pixel canonical variables were similar to that of the Landsat
segmentations.

Variograrns constructed from the field observations also suggested the presence of
a scaling threshold. Each of the four variables (Shrub, TotCan, GCbPC, and Forb)
showed a linear increase in semi-variance until a lag distance of around 2,100m (Figure
2.6). The average distance between centroids of the Ikonos image objects at the threshold
where the canonical-to-field variable correlation structure changed was 2,160m —
corresponding to the local maxima in the semi-variograms. This suggests that the
discontinuity observed in the Ikonos segmentation results did represent a scaling

threshold for our study system that the Landsat segmentation results failed to detect.

40

Spatial Predictions of Vegetation Parameters at Different Scales

We selected four scales from the Ikonos image segmentation to illustrate spatial
predictions of the canonical variables and the effects of information filtering through
image segmentation and crossing scaling thresholds (Figure 2.7). Low values of the first
canonical variable were characterized by dense shrub cover, with little forb or bare
ground cover. High values of the canonical variable had high cover of forbs, relatively
more bare ground, and low shrub cover. Below the scale threshold, detail was lost as
coarser image objects were used to predict the canonical variable across the study area,
but the correlation structure between the canonical variable and the field variables -— and
ultimately the interpretation of the canonical variable — was consistent across scales.

Once the scale threshold was crossed, however, the correlation structure relating
the field variables to the canonical variable changed. At these scales, low values of the
first canonical variable had dense canopies of shrubs and high total plant cover. High
values of the first canonical variable had high ground cover between plant canopies and

overall high percent cover of litter.

Discussion

In our findings, image segmentation outperformed pixel aggregation at all but the
finest scales. The similar performance of the two methods at fine scales is not surprising
given that, at very fine scales, image objects are collections of only a few pixels. As
image resolution coarsens, however, pixels become a gross representation of patch
boundaries (Fisher 1997); whereas the image objects retain patch shape and landscape

spatial patterns. At large pixel sizes, observations made near boundaries may fall into

41

pixels that contain information from both sides of the boundary, obscuring the correlation
between the field observations and remotely-sensed data. In contrast, segmentation
maintains distinct boundaries and provides a data-driven approach for aggregating data.
Additionally, spatial aggregations based on fixed grids or arbitrary units may
mask the relationship between image and field data. Pixels are an imposition of a grid
with arbitrary origin and cell size relative to the natural system. Estimates or predictions
from pixel-based approaches may be variable due to shifts in the location of the grid
origin or changes in resolution (Jelinski et al. 1996). This phenomenon can occur even if
a landscape is not tessellated by a regular grid but into irregular units if the units are not
defined in reference to the landscape patterns (see Svancara et al. 2002). This is the
modifiable areal unit problem (MAUP) described by Oppenshaw (1984). Image
segmentation, because the tessellation is derived from the image data, can help avoid
MAUP problems if patterns relevant to a specific objective can be defined as image
objects. By itself, however, image segmentation does not solve the MAUP (Burnett et al.
2003) because parameters set arbitrarily can result in just another imposed tessellation of
a landscape — albeit one that is based on features as recorded in the image. Selecting
segmentation levels based on the strength of the relationship between field and image
data and avoiding scale thresholds would help avoid the MAUP for a particular objective.
The Ikonos and Landsat field-to-canonical-variable correlations both showed
scale invariance suggestive of linear scaling of percent cover in our study area. The
presence of a scale threshold detected by the Ikonos results was supported by the sample
semi-variograms, but the question remains of why the Ikonos segmentation results

detected the scale threshold while the Landsat results did. not. This may be related to the

42

difference in native resolution between Ikonos and Landsat and the effect this has on
detecting boundaries. The 4m (0.0016ha) resolution of the Ikonos data provides good
delineation of boundaries. For distinct patches, these boundaries will persist even for
small, distinct patches as the image is segmented into larger image objects (Figure 2.8).
The 30m (0.09ha) pixels of Landsat, on the other hand, do not represent irregular
boundaries or boundaries of small patches as clearly, and these objects tended to be
absorbed into surrounding objects at finer scales when segmenting the Landsat image. As
a result, the Ikonos segmentation was better at maintaining the spatial structure of the
landscape, and detection of the scale threshold may have been a result of finally
dissolving object boundaries preserved in the hi gher-resolution imagery. If this is the
case, then we might expect that in systems with more distinct patch boundaries or as the
extent of the landscape increased, that detection of larger scale thresholds might be less
sensitive to the base imagery resolution. Larger scale thresholds could not be identified
via image segmentation in our study area because there would be too few image objects.
Thus, it remains to be answered whether segmentation of the Ikonos and Landsat images
would be able to detect some of the additional scaling non-linearities suggested at longer
lag distances by the semi-variogram.

Objects resulting from image segmentation that are highly correlated with one or
more field variables may represent patterns that are the result of ecological or
management processes. The fact that different field variables may have unique optimal
levels of segmentation supports the theory that they result from different processes
operating at different rates and scales. Correlating segmentation levels to field

observations to determine scale thresholds and maximize the relationship between image

43

and field data helps avoid the trap of erroneously identifying segmentation levels as
significant due to inaccuracies or artifacts in the segmentation process or data (Hay et al.
2001; Wu 2004).

Correlations between the canonical variables and the field variables can give
insight into whether or not image information could be used as a surrogate for or to map a
measured field variable. The canonical correlations state the strength of overall
relationship between the field and image data and have the advantage that they can
consider many field and image variables at the same time. Poor overall canonical
correlations would mean that a particular image is not suitable for estimating any of the
measured response variables. If a strong overall canonical correlation is observed but
univariate estimates are desired, it would be expected that predictions of field variables
that are strongly correlated to the canonical variables would be more accurate than
predictions of variables with weak correlations. Comparing correlations between images
could indicate which image was the best to use for predictions or mapping. For example,
at fine segmentation levels where the canonical correlations between the different image
types (Ikonos and Landsat) and the field observations are similar, figures 2.4 and 2.5
suggest that for predicting annual grass cover, the Landsat image may provide marginally
better predictions than the Ikonos image because of its stronger correlation with the
Landsat canonical variables. Even if a particular sensor is capable of recording the type
of information necessary to detect what a researcher is interested in, though,
environmental factors may obscure the signal. The canonical correlation analysis
described above provides a means for assessing the suitability of an image for a particular

application. For instance, detection of individual plant species may be highly dependent

44

on things like phonological stage (e.g., Lass et al. 1997), site moisture (Hunt et al. 2005),
or plant density (Bulman 2000), and canonical correlation analysis at multiple
segmentation levels can help determine not only appropriate scales, but also identify the

best sensor and time of year for imagery.

Conclusion

The method we presented above provides valuable insight into how landscape
patterns scale and can inform the process of selecting imagery and determining the scales
of analysis. Image segmentation is a powerful method for identifying patterns in data at
different scales, but a lack of guidance on which segmentation levels to use may have
hindered its broader application Our results have shown that much information can be
gleaned from looking across a range of segmentation levels, and we anticipate that our
approach will help define a more rigorous way of setting segmentation parameters to
meet project objectives.

As higher-resolution imagery becomes more widely available, methods for
filtering out unnecessary information from the relevant signals will become increasingly
important. Pixel aggregation will likely continue to be a commonly used method for
upscaling image data because it is easy and fast to implement. Image segmentation is a
more involved approach, but our results clearly demonstrate that image segmentation is
superior to pixel aggregation. While image segmentation requires specialized software,
remote sensing and GIS programs that include segmentation algorithms are becoming
more prevalent (see Neubert et al. 2006). Vegetation mapping and classification based on

image segmentation (N avulur 2007) is the most common application of object-based

45

image analysis thus far, but image segmentation has applicability for other aspects of
ecological research like field sample design or exploring the effects of landscape-scale
patterns and context into species’ habitat associations. All of these applications, though,
will require objective ways to set segmentation parameters and determine the right scales
of analysis.

For a system like ours that scales linearly between thresholds, optimal scales of
segmentation should be defined first by determining the domains between scaling
thresholds and the strength of relationship between image and field variables in those
domains, and second to maximize the amount of information relevant to a specific
objective. Within domains where systems scale linearly, information is filtered from the
image data as it is segmented into larger and larger objects, but the overall interpretation
of the objects remains the same. Thus the goal becomes identifying the level of detail
needed to address the objective. If the inquiry is concerned with predictions over a small
area, then smaller image objects would be appropriate. If aggregate information is sought
or predictions made over large areas, then the detail lost with larger image objects may be

acceptable.

Acknowledgments

This research was supported in part by funding from the M.J. Murdock Charitable
Trust, The Nature Conservancy’s Rodney J ohnson/Katherine Ordway Science
Endowment, and the Lava Lake Institute. P. Murphy and A. Lucas worked tirelessly to

collect the field data used in this study. A. Asada digitized the field forms and performed

46

quality-control on the data that was collected. J. Qi, S. Riley, G. Roloff, and R. Unnasch

provided valuable comments and edits to this manuscript.

47

Table 2.1. Percent cover variables estimated for each site that were used in the study.

 

 

Code Variable Description

TotCan Total Canopy Cover Percentage of the site that has some live
vegetation over the soil surface.

Bare Bare Ground Cover Percentage of the site having no vegetation
over the ground and an exposed soil surface.

GCbPC Ground Cover between Percentage of the site that has plant litter in

Plant Canopy areas with no plant canopy.

Shrub Shrub Cover Percentage of site with shrubs in any canopy
layer.

BchGr Bunchgrass Cover Percent of site with bunchgrass in any canopy
layer.

AnGr Annual Grass Cover Percentage of site with annual grasses in any
canopy layer.

Forb Forb Cover Percentage of site with forbs in any canopy
layer.

Litter Litter Cover Percentage of site with dead plant material

laying on the ground.

 

48

Table 2.2. Correlations between the eight field variables sampled for this study. See
Table 2.1 for descriptions of the field variable codes.

 

 

TotCan Bare GCbPC Shrub BchGr AnGr Forb Litter
TotCan 1
Bare -0.5103 1
GCbPC -0.7697 -0.0264 1
Shrub 0.4786 -0.2893 -0.3317 1
BchGr 0.3605 -0.0277 -0.4065 0.0360 1
AnGr 0.4088 -0.3884 -0.1926 0.1004 -0.5083 1
Forb -0.1329 0.2571 -0.0240 -0.5289 0.0411 -0.1595 1
Litter -0.6772 0.0048 0.8418 -0.2824 -0.2724 -0.2315 0.0796 1

 

49

   
 
 
 
 
 
   

 

Expanded Craters of the Moon
S National Monument Boundary

- Sparser-vegetated lava

- Unvegetated lava

Land Ownership
m BLM
E Private

 

 

Figure 2.1. Study area in Craters of the Moon National Monument, Idaho, USA

50

 

Pixel
Aggregation

Image
Segmentation

     

     

-» tea: w:-
TCap "Brig tness" Band
Original IKONOS Image

 

 

400-ha CeITSize

 

Median 452ha

Figure 2.2. Image segmentation versus pixel aggregation as a way of filtering information
from remotely-sensed image data.

51

IKONOS - First Canonical Correlates Across Scales

 

 

 

-°- Image Aggregation
~r- Pixel Aggregation

l
o 200 460 600 800 who
Object/Pixel Area (ha)

 

 

 

Correlation Coefficient - R 2
0'0 012 0'4 0'6 0'8 110
7’
l
\\
s.
\
/

 

q

LANDSAT - First Canonical Correlates Across Scales

 

 

 

+ Image Aggregation

-r- Pixel Aggregation

r 7 l l r

0 200 400 600 800
Object/Pixel Area (ha)

 

 

 

Correlation Coefficient - R 2
010 0,2 0'4 016 0'8 1,0

 

Figure 2.3. Canonical correlations between field measurements of percent cover and the
tasseled-cap transformed images aggregated to coarser scales by image segmentation and
pixel-based methods. Segmentation of the Ikonos image (a) yielded higher and more
consistent correlations between field and image data than pixel aggregation. For the
Landsat image (b), there was a less distinct difference between segmentation and pixel
aggregation at all scales, but segmentation generally produced higher correlations.

52

A. Ikonos - image segmentation

 

 

 

 

 

 

 

 

 

 

N
tr
' “L
E o
.9
O
E O
C
8
9 LO
“6 9"
O
0
Q-
‘T I I I fii I I
0 200 400 600 800 1000
Obiecthixel Area I ha)
B. Ikonos - pixel aggregation
0.-
N
c:
' “2.
E o
.9
8
8 Q-
o o
I:
8
E It)
2 :5-
a I
0
°-_
‘7 I I I I I 7*
O 200 400 600 800 1000
Obiecthixel Area I ha)
—— —-- --o-- -—¢-— —0— --&-- --v-- —-o——

TotCan Shrub Litter AnGr BchGr Forb GCbPC Bare

Figure 2.4. Correlations across scale levels between the eight original field variables and
the first canonical variables for the object-segmented (a) and pixel-aggregated (b) Ikonos
image.

53

A. Landsat - image segmentation

 

    

 

 

 

 

 

 

 

 

0.
N
0510 '33".—"-*-—.“-‘::f==”“:<::::3:;.—--"
E O" / /\ .- — I: ..... 7 .7.
./ xv»— ~—‘---..—-r. -------
E I. 77177,: 7’ My. ' \N‘
807- C~
00
C
O
13
Em
895‘
O
0
°-_
6 260 460 600 800 1600
Obiecthixel Area I ha)
8. Landsat - pixel aggregation
0.-
N
a:
' '0-
E0
.9
9%
8°-
00
C
_o
5.0
“£9"
0
0
°-_
‘7 I ‘li l I I I
0 200 400 600 800 1000
Obiect/Pixel Area (ha)
—-— —-- --o-- --¢-— —0— --A-- --v-- —-¢-—

TotCan Shrub Litter AnGr BchGr Forb GCbPC Bare

Figure 2.5. Correlations across scale levels between the eight original field variables and
the first canonical variables for the object-segmented (a) and pixel-aggregated (b)
Landsat image.

54

% Shrub °/o TotCan

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

In
F— I °
ml—
8 3 3.
c C
E O... t“ I
.‘z' " g l
.E. .6. «a. '
«n m_ «n o |
o l
I
Q . o— l
o I I I I I I o I r I I I r
0 2000 6000 10000 0 2000 6000 10000
Lag Distance (m) Lag Distance (m)
% GCbPC % Forb
Nu—
‘— l0__
é“,— §
#30" +£2-
.2 fl .2
E ' §
In ‘14 l a)
o | 3'4
4 l
o__ ' O...
C I T1 I I l l O I I I I I I
0 2000 6000 10000 O 2000 6000 10000
Lag Distance (m) Lag Distance (m)

Figure 2.6. Sample semi-variogram of four percent cover measures from the 159 field
sites in the study area (see Table 2.1 for key to variable codes). Points on the graph were

smoothed with a cubic spline (gray line) for display purposes. Semi-variance increased
linearly until a lag distance of approximately 2,100m for all four measures. The dashed,

vertical line in each graph represents the Ikonos-segmented image objects’ average
separation distance of 2,160m (measured from the object centroids) where the scale

nonlinearity was observed.

55

..

Median Object 5.47 ha Median Object 21.3 ha
. E.
Median Object 452 ha Median Object 1030 ha

Figure 2.7. Maps of the first canonical variable for the image segmentation at different
scales. Up until a median object size of approximately 800ha (A, B, C), the canonical
variable represents mostly a gradient from high shrub and low forb and bare-ground
cover (green) to low shrub and high forb cover and bare ground (blue). Once the scale
threshold is passed (D), the canonical variable then becomes an axis of high shrub and
total canopy cover with low litter and ground cover between plant canopies (green)
versus low shrub and total canopy cover with high litter and ground cover between plant
canopies (blue).

56

 

&

Ir:

Ikonos 4m Image Medlan Object 7. 93ha “I

 

 

l;

m .a
Medlan Object 21.3ha
to

WI

 

Medlan Object 293. 8ha.

 

%

Landsat 30m Image A Medlan Object 9. 61 ha

 

 

I p
A?

' ' :1 J1
Medlan Object 36.3ha Medlan Object 267.5ha
‘1 L. ‘1

Figure 2.8. Effects of native image resolution on the persistence of image objects as an
image is segmented into coarser-scale representations of the original data. The region
highlighted in red is a landing strip in the study area. The higher resolution of the Ikonos
data (A) gives a more distinct edge to the landing strip and results in this feature
perpetuating through very coarse scales of segmentation despite the fact that it is a very
small feature (4.01ha). The larger pixel size of the Landsat image (B) does not define the
landing strip as distinctively, and as a result it is merged into the surrounding objects
sooner. This phenomena happens for small, distinct patches and may account for why the
Ikonos segmentation detects nonlinearities in scaling when the Landsat segmentation
does not.

 

57

References

Addink E. A., S. M. de Jong, and E. J. Pebesma. 2007. The importance of scale in object-
based mapping of vegetation parameters with hyperspectral imagery.
Photogrammetric Engineering and Remote Sensing, 73:905-912.

Allen T. F. H. and T. B. Starr 1982. Hierarchy: perspectives for ecological complexity,
University of Chicago Press, Chicago, IL.

Atkinson P. M. 2004. Resolution manipulation and sub-pixel mapping. Pages 51-70 in
SM de Jong and FD van der Meer, editors. Remote sensing image analysis: including
the spatial domain. Kluwer Academic Publishers, Dordrecht, The Netherlands.

Baatz M.and A. Schape. 2000. Multiresolution segmentation - an optimization approach
for high quality multi-scale image segmentation. Pages 12-23 in J Strobl, T Blaschke,
and G Griesebner, editors. Angewandte Geographische Informationsverarbeitung XII.
Wichmann-Verlag, Heidelberg.

Bellier E., P. Monestiez, J. Durbec, and J. Candau. 2007. Identifying spatial relationships
at multiple scales: principal coordinates of neighbor matrices (PCNM) and
geostatistical approaches. Ecography, 30:385-399.

Bulman D. 2000. Is the application of remote sensing to weed mapping just "S-pie in the
sky"? (December 20, 4 A.D.).

Burnett C. and T. Blaschke. 2003. A multi-scale segmentation/object relationship
modelling methodology for landscape analysis. Ecological Modeling, 168:233-249.

Chavez P. 8., Jr. 1996. Image-based atmospheric corrections - revisited and improved.
Photogrammetric Engineering and Remote Sensing, 62:1025-1036.

Crist E. P. and R. J. Kauth. 1986a. The tasseled cap de-mystified. Photogrammetric
Engineering and Remote Sensing, 52:81-86.

Crist E. P., R. Laurin, and R. C. Cicone. 1986b. Vegetation and soils information
contained in transformed Thematic Mapper data. Pages 1465-1470 European Space
Agency, Paris, France.

Dong X., P. Nyren, B. Patton, A. Nyren, J. Richardson, and T. Maresca. 2008. Wavelets
for agriculture and biology: a tutorial with applications and outlook. BioScience,
58:445-453.

Feitosa R. Q., G. A. O. P. Costa, T. B. Cazes, and B. Feijo. 2006. A genetic approach for

the automatic adaption of segmentation parameters. Salzburg University, Salzburg,
Austria.

58

Fisher P. 1997. The pixel: a snare and a delusion. Intemation Journal of Remote Sensing,
18:679-685.

Foody G. M. 2004. Sub-pixel methods in remote sensing. Pages 71-92 in SM de Jong and
FD van der Meer, editors. Remote sensing image analysis: including the spatial
domain. Kluwer Academic Publishers, Dordrecht, The Netherlands.

Fortin M. J. and M. Dale 2005. Spatial analysis: a guide for ecologists, Cambridge
University Press, Cambridge, UK.

Hall 0., G. J. Hay, A. Bouchard, and D. J. Marceau. 2004. Detecting dominant landscape
objects through multiple scales: an integration of obj ect-specific methods and
watershed segmentation. Landscape Ecology, 19:59-76.

Hay G. J., D. J. Marceau, P. Dube, and A. Bouchard. 2001. A multiscale framework for
landscape analysis: object-specific analysis and upscaling. Landscape Ecology,
16:471-490.

Herrick J. E., J. W. Van Zee, K. M. Havstad, L. M. Burkett, and W. G. Whitford 2005.
Monitoring manual for grassland, shrubland, and savanna ecosystems, USDA-ARS
Jomada Experimental Range, Las Cruces, New Mexico.

Home J. H. 2003. A tasseled cap transformation for IKONOS images. (May 5, 2003).
ASPRS.

Hunt, E. R. Jr., Hamilton, R., and Everitt, J. H. What weeds can be remotely sensed?
USDA Forest Service, Remote Sensing Applications Center A weed manager's guide
to remote sensing and G18. 2005.

Jelinski D. E. and J. Wu. 1996. The modifiable areal unit problem and implications for
landscape ecology. Landscape Ecology, 11:129-140.

Jensen J. R. 1996. Introductory digital image processing, Second Edition edition.
Prentice-Hall, Inc., Upper Saddle River, NJ.

Johnson R. A. and D. W. Wichem 2002. Applied multivariate statistical analysis, 5th
edition. Prentice Hall, Upper Saddle River, New Jersey.

Kim M. and M. Madden. 2006. Determination of optimal scale parameter for alliance-
level forest classification of multispectral IKONOS images. International Archives of
Photograrnmetry, Remote Sensing and Spatial Information Sciences, Salzburg,
Austria.

Lass L. W. and R. H. Callihan. 1997. Effects of phenological stage on detectability of
yellow hawkweed (hieracium pratense) and oxeye daisy (Chrysanthemum
leucanthemum) with remote multispectral digital imagery. Weed Technology,

11:248-256.

59

Luscier J. D., W. L. Thompson, J. M. Wilson, B. E. Gorham, and L. D. Dragut. 2006.
Using digital photographs and object-based image analysis to estimate percent ground
cover in vegetation plots. Frontiers of Ecology and the Environment, 4:408-413.

Meisel J. E. and M. G. Turner. 1998. Scale detection in real and artificial landscapes
using semivariance analysis. Landscape Ecology, 13:347-362.

Navulur K. 2007. Multi-spectral image analysis using the object-oriented paradigm, CRC
Press, Taylor and Francis Group, LLC, Boca Raton, FL.

Neubert M., H. Herold, and G. Meinel. 2006. Evaluation of remote sensing image
segmentation quality - further results and concepts. International Archives of
Photogrammetry, Remote Sensing and Spatial Information Sciences, Salzburg,
Austria.

O'Neill R. V., D. L. Deangelis, J. B. Waide, and T. F. H. Allen 1986. A hierarchical
concept of ecosystems, Princeton University Press, Princeton, New Jersey.

Oppenshaw S. 1984. Ecological fallacies and the analysis of areal census data.
Environment and Planning A, 16:17-31.

Owen, D. E. Geology of Craters of the Moon. 2008. Arco, ID, Craters of the Moon
National Monument, National Park Service.

Radeloff V. C., T. F. Miller, H. S. He, and D. J. Mladenoff. 2000. Periodicity in spatial
data and geostatistical models: autocorrelation between patches. Ecography, 23:81-
91.

Strahler A. H., C. Woodcock, and J. A. Smith. 1986. On the nature of models in remote
sensing. Remote Sensing of the Environment, 20:121-139.

Svancara L. K., E. O. Garton, K. Chang, J. M. Scott, P. Zager, and M. Gratson. 2002. The
inherit aggravation of aggregation: an example with elk aerial survey data. Journal of
Wildlife Management, 66:776-787.

Wang L., W. P. Sousa, and P. Gong. 2004. Integration of object-based and pixel-based
classification for mapping mangroves with IKONOS imagery. Intemation Journal of
Remote Sensing, 25:5655-5668.

Wiens J. A. 1989. Spatial scaling in ecology. Functional Ecology, 3:385-397.

Woodcock C. and V. J. Harward. 1992. Nested-hierarchical scene models and image
segmentation. International Journal of Remote Sensing, 13:3167-3187.

Wu J. 2004. Effects of changing scale on landscape pattern analysis: scaling relations.
Landscape Ecology, 19: 125-138.

60

Wu J. 1999. Hierarchy and scaling: extrapolating information along a scaling ladder.
Canadian Journal of Remote Sensing, 25:367-380.

61

CHAPTER 3 - SPATIAL DEPENDENCE OF PREDICTIONS FROM IMAGE
SEGMENTATION: A VARIOGRAM-BASED METHOD TO DETERMINE
APPROPRIATE SCALES FOR PRODUCING LAND-MANAGEMENT
INFORMATION

Abstract

A significant challenge in ecological studies has been defining scales of observation that
correspond to the relevant ecological scales for organisms or processes of interest.
Remote sensing has become commonplace in ecological studies and management, but the
default resolution of imagery ofien used in studies is an arbitrary scale of observation.
Segmentation of images into objects has been proposed as an alternative method for
sealing remotely-sensed data into units having ecological meaning. However, to date
selection of image object sets to represent landscape patterns has been largely subjective.
Changes in observation scale affects the variance and spatial dependence of measured
variables, and may be useful in determining which levels of image segmentation are most
appropriate for a given purpose. We used observationsof percent bare ground cover from
346 field sites in a semi-arid shrub-steppe ecosystem of southern Idaho to look at the
changes in spatial dependence of regression predictions and residuals for 10 different
levels of image segmentation. We found that segmentation levels whose regression
predictions had spatial dependence that most closely matched the spatial dependence of
the field samples also had the highest predicted-to-observed correlations. With the
incorporation of a geostatistical interpolator to predict the value of residuals at unsampled
locations, however, consistently high correlations could be achieved across many
segmentation levels suggesting that an appropriate range or domain of scales existed for
percent bare ground rather than a single best scale. More work is needed to develop

methods that consider a wide range of different ways to segment images into different

62

scales and select sets of scales that perform best for answering specific management
questions. The robustness of ecological landscape analyses will increase as methods are
devised that remove the subjectivity with which observational scales are defined and

selected.

Introduction

Scale is widely recognized as a critical attribute of ecological inquiries that not
only defines what patterns and processes can be measured, but also influences observable
relationships and governs the inferences that can be made from a set of data (Allen et al.
1982; O'Neill et al. 1986b; O'Neill et al. 1989; Wiens 1989). In order for data to be usefill
for management decision-making, it must be collected and analyzed at spatial and
temporal scales relevant to the processes of interest to managers (O'Neill et al. 1986a)
because different patterns can emerge at different scales for almost any ecosystem (Wiens
1989)

In general terms, scale refers to the grain and extent of observations made in a
study area where grain refers to the finest level of spatial and temporal detail observable
and extent refers to the maximum area under consideration (Turner et al. 1989). Scale is a
characteristic of the set of observations and the choice of scale constrains the patterns and
processes that are observable (Burnett et al. 2003). Grain and extent define the upper and
lower limits of inference because elements of patterns below the grain cannot be detected
and inferences beyond the extent cannot be made without assuming scale-independent
uniformity of patterns and processes (Wiens 1989). Information at scales finer than the

observation grain is filtered out and considered noise, and information at scales larger

63

than the observation extent is also filtered out and becomes context for the information
retained (Wu 1999; Burnett et al. 2003). Thus, observations made with a set grain and
extent (i.e., at a specific scale) have the effect of integrating or smoothing the dataset to
give an interpretable message or signal (Allen et al. 1982; Burnett et al. 2003).

The concept of scale also has ecological interpretations where grain and extent
can be defined in terms of how organisms or ecological processes respond to their
environment with grain being the smallest area to which an organism responds to
heterogeneity in its environment (Addicott et al. 1987; Kotlair et al. 1990) and extent
being the coarsest set of environmental patterns to which organisms or ecological
processes react (Farina 1998). From an ecological perspective, scale is relative to an
organism or process and is not an inherent property of the environment (Wiens et al.
1989). Because natural landscapes are a complex product of many different processes and
environmental factors (Peters et al. 2006) that are difficult to characterize when viewed
from arbitrarily-selected spatial and temporal scales (Levin 1992; Burnett et al. 2003), a
significant challenge in ecological studies has been defining scales of observation that
match the relevant ecological scales for organisms or processes of interest.

The use of remotely-sensed data from satellite imagery or aerial photography has
become commonplace in ecological studies and management. Most studies that have
employed remotely-sensed data have used the de facto scale determined by the image’s
pixel ground dimensions (i.e., resolution). In the context of the ecological definition of
scale discussed above, the default resolution of imagery is an arbitrary scale of
observation that may or may not correspond to the scale of ecological patterns and

processes. Three potential problems may arise when, rather than basing scale selection on

64

ecological processes or organisms of interest, scales are selected in an arbitrary manner or
for convenience (Addicott et al. 1987): 1) it becomes difficult to compare results between
studies because the units represent different scales, 2) a given observational unit may not
correspond to an ecological patch appropriate for the variable of interest which could lead
to misinterpreting relationships between events (see also Wiens 1989), and 3) because
different processes in the same system may occur at different scales, it can be difficult to
know the ones to which observed patterns are related.

In the cases where scales different from the image resolution are selected, either
from convenience or through an analysis of scaling relationships, scale is usually adjusted
by aggregating pixels into regularly shaped units (i.e., grids of square cells). Karl and
Maurer (in prep) found that correlations between images aggregated by regular, square
grid cells and field measurements of semi-arid shrub-steppe ecosystems were
unpredictable as the size of the cells increased. Also, as pixels (i.e., grain) become large
relative to the extent, analysis results can be influenced by small shifts in the reference
grid that defines the pixel boundaries (Jelinski et al. 1996).

Segmentation of image pixels into polygons has been proposed as an alternative
to pixel-based methods for sealing remotely-sensed data (Burnett et al. 2003; Wu et al.
2006a). Image segmentation works by grouping similar, neighboring pixels into discrete
regions called objects in such a manner that the pixels within an object are more alike
than they are to other pixels around the object (Burnett et al. 2003). Segmentation is a
rescaling of the original image data with the intent to produce a set of objects that match
conditions on the ground (Woodcock et al. 1992). Taken together, a set of image objects

that completely tessellate an area into discrete units constitutes one possible observational

65

scale. There are many different methods for aggregating pixels into image objects (see
Neubert et al. 2006), and within each method are parameters for controlling the scale that
will result. Most significantly, parameters affecting the degree to which information that
is filtered fi'om the image — which in turn affects things like how large the output objects
are on average — have a significant effect on how the image information relates to ground
observations and what the resulting objects can be used for. Up to some segmentation
level, the information that is lost is expected to be noise that is irrelevant to the analysis
objectives. Past a given threshold, important information is lost and the accuracy of
analysis results may decline (Addink et al. 2007).

The choice of parameters for how to aggregate pixels into objects, however, is
subjective (Wang et al. 2004) and usually involves a trial and error process for
subjectively deciding which set of objects best represents what the person doing the
analysis interprets as meaningful patches on the ground (Burnett et al. 2003; Feitosa et al.
2006; Navulur 2007). However, ecologists may be more likely to select intuitive,
anthropocentric scales for studying systems rather than scales at which processes are
actually happening (Wiens 1989)

The optimal scale of segmentation is defined as the set of objects that results in
the lowest prediction error of modeled parameters. Methods for determining optimal
scales of analysis have used one of two basic approaches: analysis of changes in image
object parameters with changes in scale, and comparison of products derived from
segmentation at different levels to some sort of “ground tru ” data.

Using changes in object parameters with scale as a means for identifying optimal

scales is based on the theory that when the scale of observation changes, the variance of

66

the variables measured also changes because as grain increases, a greater proportion of
environmental heterogeneity becomes contained in each grain unit and is lost to
observation (i.e., within-grain variability increased and between-grain variability
decreases, sensu Wiens 1989). Plots of between-grain variance against grain-size, called
scalograms (sensu Wu 2004), have been used to look for patterns that deviate from what
would be expected from homogeneous systems (O'Neill 1989; Wiens 1989). Wu (2004)
used changes in landscape metrics (e. g., within patch heterogeneity, fragmentation) as
scale increases to look for evidence of scaling non-linearities in pixel-based image
analysis. For image segmentation, Kim and Madden (2006) used variances of pixels
within image objects at different segmentation levels to determine the best scale for
mapping forest types.

Comparisons to ground-truth data for determining optimal segmentation levels is
also common. F eitosa et al. (2006) used a genetic algorithm to iteratively find the
combination of segmentation parameters yielding image objects that most closely
matched pre-defined patches. Wang et al. (2004) used a multivariate distance measure,
Battacharya distance, to determine the segmentation level that gave a mangrove forest
classification that most closely matched training data. Addink et al. (2007), using a
regression technique, found optimal levels of segmentation for estimating leaf-area index
and biomass, and showed that segmentation levels that produced the best results were not
the same for each variable.

Changes in scale can also affect the degree of spatial autocorrelation between
observations, but little research has been done on how spatial dependence (i.e., how

autocorrelation changes as a function of distance between observations) changes with

67

increasing levels of image segmentation. Geostatistical techniques such as semivariance
analysis have been used to explore spatial dependence in order to characterize effects of
scale on observational datasets and to look for evidence of ecological scales (Cullinan et
al. 1997; F ortin et al. 2005; Garrigues et al. 2006). Geostatistical techniques may be
useful for understanding how segmentation changes spatial dependence of remotely-
sensed data and identifying optimal scales of image segmentation for a given task.

The objective of this study was to determine if, as observational scale increases,
changes in spatial dependence could be used to identify segmentation levels that most
closely match field measurements and could be considered relevant scales for mapping
ecosystem features. As an example, we used observations and predictions of percent bare
ground cover, an important attribute for assessing the condition of rangelands (National
Research Council 1994; Pyke et a1. 2003), in a semi-arid shrub-steppe ecosystem of
southern Idaho. Specifically, we looked at how the semivariance of linear regression
predictions and residuals changed as successively coarser levels of image segmentation
we employed. Through comparison to semivariance calculated from field samples, we
detennined which sets of image objects best represent observed patterns of bare-ground
cover. We also compared the semivariance results to scalograms of within- and between-
object variance to determine if these easier-to-implement methods could also produce
similar results. Finally, we discuss the benefits and limitations of our approach to
defining scales through image segmentation and the implications of scale selection for

providing data for land management planning.

68

Study Area

For this study we considered the Bureau of Land Management’s (BLM) 97,308-
ha Wildhorse grazing allotment in southern Idaho (Figure 3.1, 43.028°N, 113.864°W).
The defining features of this landscape are expanses of unvegetated black rock created
from a succession of volcanic eruptions that blanketed this area with basaltic lava flows
15,000 to 2,000 years ago (Owen 2008). Soils in the study area are mostly aridisols with
low organic matter and subsurface accumulations of either calcium carbonate or clay
(Soil Survey Staff 2006b; Soil Survey Staff 2006a). The terrain in the Wildhorse
allotment is flat to gently rolling with elevations ranging from 1,272 m to 1,557 m.
Average annual precipitation ranges from 24.9 cm to 32.6 cm based on the PRISM map
of average annual precipitation from 1971 to 2000 (PRISM Group, Oregon State
University, http://www.prismclimate.org, created June 16, 2006).

Current vegetation communities consist of a mosaic of mountain big sagebrush
(Artemisia tridentata Nutt. ssp. vaseyana (Rydb.) Beetle), three-tip sage (Artemisia
tripartita Rydb.), and Basin big sagebrush (Artemisia, tridentata Nutt. ssp. tridentata)
types. Native understory composition is principally bunchgrasses: bluebunch wheatgrass
(Pseudoroegneria spicata (Pursh) A. Love), Idaho fescue (F estuca idahoensis Elmer), and
bottlebrush squirreltail (Elymus elymoides (Raf) Swezey). Cheatgrass (Bromus tectorum
L.) abundance is variable within the study area, reaching high densities in areas that have
fi'equently burned and sites disturbed by intensive livestock use (e. g., near watering
troughs, corrals, and loading areas).

This study area has seen an active fire history with 18 wildfires within the last 20

years, and eight of those being greater than 200 ha. Over the last 20 years, 80.04% of the

69

Wildhorse Allotment has burned. The frequent, large fires in this area have contributed to
the spread of cheatgrass and other invasive species in the allotment.

The majority of the study area is in public ownership with the Bureau of Land
Management (BLM) being the largest single land steward — managing approximately
93,317 ha (95.8%) of the study area. Approximately 1,305 ha (1.3%) of the study area is
in private ownership, and 2,843 ha (2.9%) managed by the state of Idaho. The main land

use in the study area is cattle and sheep grazing.

Methods

Field Dga Collection

For this project we used data collected by the Shoshone District BLM (G. Mann
and J. Russel, unpublished data) from 468 field observations of percent cover collected
between June 21, 2006 and August 6, 2008. Percent cover of vegetation and soil surfaces
was estimated for each site from a single 15.15m transect using the line-point intercept
method as part of the Fire Effects Monitoring and Inventory Protocol (Keane et al. 2005).
The starting location of each transect was recorded with a GPS and differentially
corrected.

We excluded 122 of the field sites due to missing data or fires that occurred
between when the field measurements were taken and when the satellite imagery was
acquired. Percent bare ground cover was calculated on the remaining 346 sites. For the
purposes of rangeland assessment and monitoring, bare ground is considered land surface
not covered by vegetation, rock, or litter (Bedell 1998; Pellant et al. 2005). Percent bare

ground cover was calculated as the proportion of the points along each transect where no

70

plant canopy was intercepted and the soil surface was recorded as exposed soil (Herrick
et al. 2005). Because regression analyses and kriging methods are sensitive to data
normality (Cressie 1993; Bailey et al. 1995), we used a square-root transformation to

achieve normality of the percent bare ground measurements.

Image Acquisition and Pre-processing

We acquired a Landsat Thematic Mapper (TM) 5 scene from July 11, 2008. We
ortho-rectified and geo-registered it using publically-available 1/3 arc-second digital
elevation models and l-m resolution color aerial photography to have a nominal ground
resolution of 30111. We also atrnospherically corrected the Landsat image using Chavez’s
(1996) dark-object-subtraction method and converted the 8-bit image values to estimated
at-sensor reflectance. For this study we considered the six TM bands in the visible, near-
and shortwave infrared regions that had a ground resolution of 30m and omitted the 60m
thermal infrared band.

Because image segmentation is sensitive to correlations between bands (N avulur
2007), we used the tasseled-cap transformation to convert the original six TM bands into
four nearly orthogonal bands. The tasseled-cap transformation, originally defined by Crist
(1986), uses a linear combination of the input bands to create new composite bands that
have a consistent interpretation of brightness, greenness, and wetness for composite
bands one, two, and three, respectively. To achieve the same interpretation of the
tasseled-cap bands between image sources, tasseled-cap coefficients must be calculated
specifically for each sensor. For the Landsat image we used Crist’s (1986) Landsat 5

coefficients to generate the tasseled-cap bands for the Wildhorse allotment from the six

71

original Landsat bands. The tasseled-cap transformation produces the same number of
output bands as input bands, but certain bands (e. g., tasseled-cap bands four and six for
Landsat TM) contain too much noise to be a useful source of information. Tasseled-cap
bands four and six were not used in image segmentation or subsequent analyses.

We additionally calculated a modified soil-adjusted vegetation index (MSAVI) —
an index of vegetation greenness that is designed to be sensitive to changes in vegetation
cover in areas with high amounts of exposed soil or rock (Qi et al. 1994; Gilabert et al.
2002). The MSAVI was not used in the segmentation process because it correlates highly
with several of the tasseled-cap bands, but was used in subsequent regression analyses
because previous research has shown that vegetation indices correlate well with percent
cover of vegetation (e.g., Anderson et al. 1993; Hunt et al. 2003; Buono et al. 2005;
Marsett et al. 2006) and bare ground (e.g., Bertiller et a1. 2002; Harris et al. 2003; Buono

et al. 2005).

Image Segmentation

All image segmentation was done with Definiens Developer 7.0
(http://www.definiens.com) using the multi-resolution segmentation algorithm described
by Baatz and Schape (2000; Burnett et al. 2003). Multi-resolution segmentation works by
merging adjacent pixels in the first iteration and objects in later iterations and evaluating
the increase in local heterogeneity. If after merging the local heterogeneity is below a set
threshold, the merged objects are retained, otherwise the merge is not kept and a different
combination of objects is tried. This process continues until all possible merges below the

threshold are made. The multi-resolution segmentation method requires the user to set the

72

threshold value through a unitless scale parameter (p5). By increasing ps, the acceptable
level of heterogeneity of pixels within objects is increased and, on average, segmentation
produces larger objects. It is important to note that multi-resolution segmentation does
not specify a minimum object size and groups of pixels that are very distinct will be
maintained as separate objects while the surrounding objects increase in size. For our
analyses, we varied only p, in multi-resolution segmentation and used default values for
all other parameters.

We segmented the tasseled-cap-transformed image for the Wildhorse allotment
multiple times to create a set of image segmentations that became progressively coarser
in scale. For each successive segmentation run, we incremented p, by a small amount
(Table 3.1). The image objects for each segmentation level were attributed with the mean
and standard deviation of the pixels within the object for all the tasseled-cap bands and
the MSAVI and then exported for statistical analysis. For each segmentation level, the
output variables were assessed for normality and transformed as necessary using either

square-root or log transformations.

S_tatistical Analysis

To assess the spatial dependence of the field measurements of percent bare
ground cover, we constructed a semi-variogram between all pairs of observations using

the following equation:

r‘( h)=2—(—,52[z(s) z(s +h)]’ (1)

73

where 70‘) is the estimate of semivariance for all observations separated by a distance h,

n is the total number of observations, z(s,-) is the measured percent bare ground cover at

point 3,, z(s,-+h) is the measured percent bare ground cover at a point a distance of h

away, and nflt) is the number of points separated by a distance of h (Fortin et al. 2005).
Distances were collapsed into discrete distance classes (i.e., lags), and the estimated
semivariance plotted over all lags considered to produce an empirical variogram (Figure
3.2). We fit a spherical semivariogram model to thepercent bare ground variogram using
the ordinary least-squares (OLS) regression method in R’s gstat package (Pebesma 2004).
We checked for anisotropy (i.e., directional differences in spatial dependence) and
determined that it was not significant for percent bare ground cover. Accordingly, we
used a single omni-directional variogram model for the spatial dependence of bare-
ground cover. The semivariogram model was characterized by (Bailey et al. 1995): the
nugget (i.e., variability at distances smaller than the shortest distance between sample
points — includes measurement error), sill (i.e., total observed variation of the variable),
and the range (i.e., distance at which two observations could be considered independent).
The nugget-to-sill ratio (N SR) of a variogram is commonly used as a measure of the
proportion of the total observed variation that cannot be explained by spatial dependence
of the variable (i.e., by the variogram model). The NSR is commonly used in geostatistics
to characterize the proportion of data variability that cannot be described by the
variogram model (Kravchenko 2003).

We then used a combination of generalized least-squares regression (GLS;
Pinheiro et al. 2000) and regression kriging (RK; Hengl et al. 2004) to predict percent

bare ground cover from the tasseled-cap band and MSAVI values summarized in the

74

image object polygons at the different segmentation levels. For each segmentation level,
we used the following basic steps:

1) selection of objects containing one or more field sites,

2) GLS regression between average percent bare ground within each object to the

tasseled-cap and MSAVI values for the objects,

3) variogram modeling of GLS bare-ground predictions and residuals,

4) RK using the GLS model and the residuals variogram, and

5) cross-validation of the GLS and RK results.

All statistical analyses were performed in R version 2.4.2.

The first step for each segmentation level was to select the image objects that
contained one or more field sample locations. These objects were extracted to a new
dataset and each polygon was attributed with the mean and standard deviation of the
pixels occurring in it for the four tasseled-cap bands and MSAVI. Coordinate values were
also included for the geometric center of each polygon. When more than one sample
location fell within an image object polygon, the field-measured percent bare ground for
each selected image object was determined by averaging the measurements of all sites
within the polygon. We used this method to avoid artificially inflating sample sizes of
objects as segmentation levels increased and because it was akin to taking multiple
samples within an object. The image values summarized by object polygon, coordinate
values, and average percent bare ground cover were exported in a single table for
analysis.

The second step at each segmentation level was to use GLS regression to establish

the relationship between the field measurements of percent bare ground and the image

75

values summarized by image object and predict percent bare-ground cover. We included
all 10 irnage-band measures (i.e., mean and standard deviation of pixels per object for the
four tasseled-cap bands and MSAVI), and the coordinate values in an initial regression
model. The only interaction term we considered was between the X and Y coordinate
values. Models were created separately for each segmentation level, and we used a
backward stepwise procedure to select the most parsimonious model at each level. The
GLS regression method allows for specification of the covariance between samples. In
the context of this study, we used GLS in order to incorporate the spatial covariance
between samples (Bailey et al. 1995). Spatial covariance is incorporated into GLS in an
iterative fashion via a semivariogram model found from the residuals of an initial OLS
regression model (Hengl et al. 2004). The regression is then rerun with GLS using the
semivariogram model to specify covariance of the samples. We used R’s nlme package
(Pinheiro et al. 2000) for all GLS regression modeling.

The predictions of the GLS regression model at each segmentation level were
assessed using leave-one-out cross validation. In this process, one observation, selected at
random, was withheld from the regression and the predicted value for that location
compared to the observed value. The omitted point was then replaced and another
observation randomly selected. This procedure was repeated 100 times and the results
used to estimate a correlation between predicted and observed measures of percent cover.
We employed this method for assessing the GLS regression to have a measure that was
comparable to what was available for RK (below).

We hypothesized that segmentation levels yielding the best GLS regression

results did so because the image objects were defined in such a way as to minimize

76

spatial dependence between objects. If this were the case, then the predictions of the GLS
model for the sample locations would have a similar spatial dependence to the field
measurements of percent bare ground cover and the residuals of the GLS model would
show little spatial dependence. To test this, we constructed variograms for the predicted
percent bare ground and the residuals of the bare ground prediction at each segmentation
level and compared them to the variogram of the field sample locations by way of their
NSR and range. Because anisotropy for the bare-ground predictions and residuals was
negligible for each scale, we used omni-directional variogram models for the spatial
dependence of the GLS predictions and residuals.

If substantial spatial dependence exists in regression model residuals, then
geostatistical techniques can be employed to improve the predictions of the model. The
RK method was developed as a way to exploit spatial autocorrelation in datasets to
increase the accuracy of predictions (Odeh et al. 1994; Odeh et al. 1995; Hengl et al.
2004; Karl and Maurer in prep). Regression kriging uses a GLS regression model to
predict a response variable and then uses the geostatistical technique of kriging (Krige
1966; Goovaerts 1997; Hengl et al. 2004) to predict the value of the GLS residuals at

unsampled locations. Thus the RK predictor for a variable at an unmeasured location,

2(S0 ) , is the sum of the GLS regression prediction and the predicted residual:
p A n
Z(S0)=Zflk °qk(S0)+Z/1i°e(si) (2)
k=0 i=1

where the flk and qk(s0) are the regression coefficient and the value of the kth predictor

variable at the unknown location, respectively, the A,- are the kriging weights determined

from the i known locations using the variogram model of the GLS-model residuals, and

77

e(s,-) is the GLS-model residual value at point i (Hengl et al. 2007). All RK predictions

and cross-validation was accomplished with the gstat package in R (Pebesma 2004).

The objective of implementing RK in addition to the GLS modeling was to
determine if added predictive capability could be gained through explicitly incorporating
spatial dependence into the predictors. At each segmentation level, we used the GLS
model and the residuals variogram model developed above to create RK predictions of
percent bare ground cover. We assessed the RK predictions using the same cross-
validation procedure described above to derive a correlation between predicted and
observed percent bare ground cover. For each segmentation level, we compared the
predicted-vs-observed correlations for the GLS and RK predictions.

Kim and Madden (2006) suggested that changes in pixel variance within and
between objects could be used to identify appropriate scales of segmentation for analyses.
This follows, in part, from Wiens’ (1989; see also O'Neill 1989) prediction that the
variance of a variable would decrease as scale coarsened because more of the system’s
variability would be contained within the observations (i.e., objects). Such measures, if
they proved to be reliable in identifying appropriate levels of segmentation, would be
idea] because they are much easier to generate than regression-based or geostatistical
measures and would not require the use of any empirical field data. To test if best-
perforrning segmentation levels for predicting percent bare ground cover could be
predicted using only information on changes in variance of image objects, we constructed
scalograms of changes in within- and between-obj ect variance with respect to scale. At

each segmentation level average within-object variance was calculated as:

78

n
2
20,.

i=1

“w '(T—B <3)
where n is the total number of image objects and 0,-2 is the variance of pixel values within
the ith image object at that scale. Between-obj ect variance was calculated as the variance
of the object means across all image objects in a segmentation level:

2
n 1 n
2 ’9'“ij
nj=l

1:1

0' =
b (n—l) (4)

 

where 37,- and 55]- are the mean pixel values for each image object in a segmentation level.

By these equations, within- and between-object variances can be calculated for only one
image band at a time. We calculated within- and between-obj ect variance for each
tasseled-cap band and the MSAVI at each segmentation level to construct our

scalograms.

Results

The empirical variogram for percent bare-ground cover field measurements
showed strong spatial dependence (Figure 3.3). The variogram model we fit had spatial
dependence of bare-ground cover extending to a range of 14,719m. The nugget and sill of
the variogram model were 0.0036 and 0.0147, respectively, with the NSR being 0.2474 —
meaning that about one-fourth of the observed variation in percent bare ground cover
could be considered short-range or within plot variability that could not be explained by

the model of spatial dependence.

79

Segmentation of the TM image into successively coarser scales resulted in an
exponentially-decreasing number of image objects with scale (Table 3.1). As p,
increased, the size of the objects also increased. Within any segmentation level, the
distribution of object sizes was not normal, and the typical area was best characterized by

the median object size (3mm). Median object size increased roughly with the square of p,

(5mm = 0305* $1.85, R2=0.999). Minimum object area also increased with the square of p,

2.11

(smm = 0.0034* s , R2=O.979) and ranged from a single Landsat pixel at the smallest p,

up to 116 pixels at ps=50. Maximum area increased with scale parameter until ps=20

where it remained relatively constant. At p5: 40, maximum object area jumped abruptly.

Cross-validated correlations between the GLS regression predictions and the field
measurements at different scales ranged from R=0.5553 to R=0.6677 (Table 3.2). For the
RK predictions, correlations were higher at each segmentation level except ps=50 and
ranged from 0.5792 to 0.7223. There was some evidence that GLS predictions increased
slightly in correlation with field measurements as segmentation level increased (Figure
3.4, slope = 0.0013, R2=0.3145, p-value=.0917 for test of zero slope). The RK

predictions, however, showed no increasing trend with segmentation level increases

(slope = 0.0001 , R2=0.0109, p-value=0.7744 for test of zero slope). Correlations reached

local maximums at ps=20 and 35 for GLS and ps=20, 35, and 45 for RK predictions.

Nugget-to-sill ratios for GLS predictions ranged from 0.1000 to 0.5435 (Table
3.2). This meant that depending on the segmentation level, anywhere between 10% to
approximately half of the variation in predicted percent bare-ground cover could not be

attributed to spatial dependence. For the GLS residuals, NSR ranged from 0.000 to

80

0.5451, and, in general, as NSR for the residuals decreased the GLS model correlations
increased (R2 = 0.6927). Ranges of the variogram models at different segmentation levels
were from 15,189m to 22,041m for the predicted values and 3,035m to 12,507m for the
prediction residuals (Table 3.2). Ranges for the residual variogram models were always
shorter than those of the predicted values and tended to decrease with increasing
segmentation level as more of the spatial dependence in bare-ground cover was
accounted for by coarser image segmentation. Segmentation levels whose residual

variograms had the longest ranges also tended to have the largest differences between the
GLS regression and RK predictions of bare-ground cover (R2=0.5502).
At fine levels of segmentation (e.g., ps=5), the form of spatial dependence for the

GLS residuals more closely matched the field empirical variogram than did the spatial

dependence from the GLS predictions (Figure 3.5). As segmentation level increased to
ps=20, the variograms for the predicted values became more similar in form to the field-
derived variogram, and the spatial dependence of the regression residuals decreased (i.e.,

range decreased, Figure 3.5). Beyond ps=20, the variogram range of GLS predictions

values increased again until ps=35 where the predicted-values variogram again closely

matched the field-derived variogram. When measured by the predicted-values variogram

range and NSR, the segmentation levels that most closely matched the field-derived bare-
ground cover variogram (i.e., ps=20 and 35) corresponded to the locally maximum
correlations for the GLS-only predictions (Figure 3.6). This also held for the RK results

with the exception of ps=45 which had the highest predicted-to-observed value

81

correlation but whose spatial dependence of prediction values was very different from the
field-derived variogram (Figure 3.6).

When the correlation between predicted and observed bare-ground cover was
plotted against the difference in range between the predicted-values variogram and the
field-derived variogram for the GLS regression predictions there was some evidence for a

general decrease in correlation as the difference in ranges increases (Figure 3.7,

R2=0.2188, p-value=0. 1728 for test of zero slope). For RK, however, there was no trend

in correlation with changes in difference between the variogram ranges (R2=O.0050, p-

value=0.8463 for test of zero slope). These results are consistent with results above that
segmentation levels producing predictions with spatial dependence matching that
observed in the field yielded the best predictions when geospatial statistical techniques
were not employed. The use of geostatistical techniques, however, could achieve similar
results regardless of scale because they accounted for spatial dependence not captured by
the regression prediction.

Within- and between-obj ect variances increased with segmentation level for all
four tasseled-cap bands and the MSAVI. Because the form of increase for within- and
between-obj ect variances was similar for each band, we report only those values for
tasseled-cap band three (Table 3.1, Figure 3.8). Tasseled-cap band three was used
because it was highly significant in the GLS regression models at all segmentation levels.
Both within- and between-object variances increased with median object size along a

power curve (i.e., a straight line on a log-log plot). The increase in within-obj ect variance

0.1287

(0.0038*smm ) followed Wiens (1989) prediction for within-grain variance increasing

0.2407

with scale. The increase in between-object variance (0.0095*s,,,,-,, , R2=0.972) that we

82

observed was contradictory to Wiens (1989) and O’Neill (1989) who proposed that
between-grain variances should decrease with scale. Neither within- nor between-object
variance, however, showed any marked departures from the power trend line that
corresponded to differences in correlation between predicted and observed bare-ground

cover at different segmentation levels.

Discussion

When considering only GLS regression, the scales where spatial dependence of
prediction most closely matched that of the field measurements in terms of NSR and
range performed the best. The likely reason for this lies in the fact that image
segmentation is a non-arbitrary aggregation of pixels into units that have implicit
meaning with reference to the landscape being studied (Hay et al. 2004). By grouping
together similar adjacent pixels into objects in a manner that minimizes variance within
each object (Baatz et al. 2000; Burnett et al. 2003), image segmentation preserves the
variance between the objects. When the variance between objects is maximized, the
objects are better able to model the large-scale variation in the observations. Hence there
is less unexplained variation in the residuals. Scales that match the field variogram
perform better because the objects at that scale are defined in such a way that the variance
between the objects is similar to the variance of the field measurements not only in
magnitude, but also in covariance between samples. By contrast, aggregation of pixels by
regular grids of arbitrary shapes tends to decrease variance as scale coarsens (Wiens

1989). This not only obscures the spatial dependence of a variable but also decreases the

83

chance of defining a set of arbitrary units with similar variance as the field
measurements.

The slight increase in correlations between GLS-regression predictions and field
observations of bare ground is consistent with the findings of Karl and Maurer (in prep).
Looking at canonical correlations between field observations and both Landsat and
Ikonos imagery, Karl and Maurer reported that correlations between field and segmented-
image data increased with scale when image pixels were aggregated by image
segmentation. On the other hand, field-to-image correlations decreased with scale as
pixels were aggregated by square grid cells of increasing size.

The high correlation between the RK prediction and field-measured bare-ground
cover at ps=45 did not follow the patterns observed for RK at other scales. The likely

reason for this result has to do with how the Landsat image was segmented. First, we
used the original Landsat pixels to create the image objects at each scale rather than
deriving coarser scales through merging of objects in the next finest scale. This was
because merging objects at one scale to produce a coarser scale enforces a patch
hierarchy on the study area that may not be appropriate, and segmentation that does not

enforce strict hierarchies can show ephemeral objects that appear at one scale and are
later replaced by different sets of objects (Hay et al. 2003). Segmentation up to ps=40
grouped pixels similarly at each scale, steadily produced a set of image objects that

gradually grew larger by, in effect, dissolving boundaries between objects. At ps=45,

while most small objects were similar to finer scales, the largest image objects had

different boundaries and configurations. This was evident in Table 3.1 where the

maximum object size for ps=45 was much smaller than for ps=35 or 40. Because this

84

event happened at one of our coarsest scales, and because the number of image objects at

scales coarser than ps=50 gave too few objects for reliable regression or variogram

estimates, we cannot tell if the high correlation at ps=45 was a unique event or signified a
scaling transition. Owing to the fact that the jump in correlation appeared to have come
from a reconfiguration of the object boundaries, the high correlation at ps=45 did strongly
suggest that other ways of combining pixels into meaningful objects (i.e., by exploring
other parameter options in addition to ps) may provide better results than what we

reported here. Feitosa et al. (2006) used a genetic algorithm to vary three multi-resolution
segmentation parameters to find the combination that best matched pre-defined patches —
a similar approach could be adapted to identifying scales that maximize regression-based
prediction accuracy.

For making RK predictions, the scales which GLS-residual semivariograms had
low NSR and long ranges showed the largest gain in correlation to field observations over
GLS. This suggested that the potential to improve overall predictions by predicting
residual values at unsampled locations could be assessed by considering the NSR and
range of a regression-residuals variogram. Low NSR and long ranges usually indicate
that higher accuracies can be achieved in making predictions to unsampled locations
(Kravchenko 2003). Using kriging and inverse-distance weighting (another geostatistical
interpolator) on soil properties, Kravchenko (2003) found that variables with strong
spatial structure could be mapped more accurately than those that had weak spatial
structure regardless of the variance of the soil property.

The success of the technique we have outlined in this paper for identifying

appropriate levels of image segmentation for making predictions of ecosystem attributes

85

depends on two factors. First, there must be enough spatial dependence in the attribute of
interest to detect and use in conjunction with available remote-sensing data. Percent bare-
ground cover had a well-defined spatial dependence in the Wildhorse allotment, making
it a good candidate for this study. While some other variables for the same area had
similar degrees of spatial dependence (e.g., percent shrub cover) a number variables of
interest to managers had either poor overall spatial dependence that (e.g., percent cover of
forbs) or spatial dependence exhibiting such a short range to not be useful with our
moderate-resolution Landsat imagery (e.g., cheatgrass [Bromus tectorum L.]). Precision
of the field measurements is an important consideration when assessing spatial
dependence because imprecise measurements can inflate the variogram nugget (Webster
et al. 2007) and decrease the observed spatial dependence. Second, enough field samples
must be collected to accurately characterize the spatial dependence that is present, and the
samples must be collected in the right manner. Webster and Oliver (1991) found through
simulations that variograms created using fewer than 50 samples did a poor job of
estimating actual spatial dependence. They recommended at least 100 samples to
construct a reliable variogram. In addition, the distribution of the sample locations is
important. If samples are taken on a regular grid, the shortest lag distance possible will be
the grid spacing and it will be impossible to evaluate short-range spatial dependence if
samples are spaced far apart. One strategy when selecting sample locations for use in
geostatistics is use multi-level or nested sampling (Webster et al. 2007) to achieve a
variety of distances between samples.

Our results suggest that identification of “best” scales is important when using

standard regression-based predictions. However, the incorporation of geostatistical

86

techniques, like RK, that account for spatial dependence in residual values made
predictions less sensitive to scale selection. While we looked only at a single dependent
variable in this study, percent bare-ground cover, we expect that predictions from
methods employing geostatistical techniques would be relatively invariant to the choice
of scale as long as the conditions outlined in the previous paragraph are met. The
significance of this is twofold. First, fine levels of segmentation can be used for making
predictions without needing to assess many different scales to find the one that best
matches the variable of interest. Second, because optimal levels of segmentation may not
be the same between different variables (Wiens et al. 1989; Addink et al. 2007), the
ability of predictors like RK to achieve high correlations across a range of scales means
that a single scale may be useful for predicting multiple variables across a landscape.
Our results support Addink et al.’s (2007) observation that specific segmentation
levels may be defined to maximize the accuracy of a prediction. However, our results
also suggest that there may not be a single “characteristic” scale that best describes the
distribution of a variable, rather there is a range of scales over which good predictions are
possible (i.e., in the case of bare-ground cover, RK produced high correlations over most
scales considered). Horne and Schneider (1995) argued against the notion of
characteristic scales on the grounds that apparent characteristic scales (i.e., those that
maximize between-grain variance) often change over short time pans and can only be
reliably detected through long-term study because landscape patterns are often the result
of complex interactions of many physical and biological processes. Instead, our results
support the concept of scaling domains (Wiens 1989) which are ranges of scales within

which the relationship between variance and scale is consistent due to a similar set of

87

underlying processes that control landscape patterns (Wu et al. 2006b). Karl and Maurer
(in prep) found that scaling through image segmentation was a better method than square
pixel-aggregation techniques for characterizing scaling domains and thresholds from
remotely-sensed data and field observations.

While we did find evidence of optimal scales for predicting the distribution of
bare ground cover in the Wildhorse area through our regression/semivariogram analysis,
the graphs of between- and within- object variance change were not informative. This
also had to do with the fact that the segmentation algorithm, by minimizing local variance
around each object, preserved or increased between-object variance. Aggregation of data
by regular grids of arbitrary units incorporates disparate observations into a single grain
as scale coarsens. At grain sizes smaller than dominant landscape patterns, the likelihood
of aggregating dissimilar pixels may be low and variance will decrease slowly with
coarsening scale. Once grain size exceeds the patch size, however, variance will decrease
rapidly with scale as more and more dissimilar pixels are grouped within the same grain.
In contrast, image segmentation, because it is a data-driven aggregation maintains the
distinction between dissimilar regions and shows only incremental changes in within- and

between—obj ect variance.

Conclusion

The fact that the ability to predict ecosystem attributes from remotely-sensed data
is closely tied to the scale of analysis is not new. It is a product of the well-known
relationship between measurements of an attribute’s variance and the scale of observation

that occurs because the attribute is non-randomly distributed across a landscape (Wiens

88

1989; Home et al. 1995; Fuhlendorf et al. 1999; Wu et al. 2006a). The contribution of our
results is that the spatial dependence of predictions also changes with scale and this
information can be used to select scales that maximize predictive ability. Scales yielding
predictions whose spatial dependence most closely matched spatial dependence of
percent bare ground measured from field samples also tended to yield the most accurate
predictions.

Our results also support the idea that there is not a single optimal scale for
understanding spatial variance of an ecosystem variable, but rather a range of appropriate
scales. When considering only regression-based predictors, discrete scales could be
identified that produced the most accurate predictions based on comparing the spatial
dependence of the predictions to the spatial dependence of the field measurements.
However, when spatial dependence was accounted for through incorporating a
geostatistical interpolator, high accuracy could be achieved across a range of scales.

The process of reducing information on landscape patterns to discrete units is a
means of reducing the complexity of ecological systems so that we can understand the
relationships between patterns and processes (Burnett et al. 2003). Scaling theory
supports the notion that there may be many appropriate ways for aggregating landscape
data depending on the objective. In our results, we found evidence of this. Image
segmentation shows much promise for sealing image data for ecological analyses, but
more work is needed to develop methods that consider a wide range of different ways to
segment images into coarser scales and select sets of scales that perform best for

answering specific management questions. The robustness of ecological analysis of

89

landscapes will increase as methods are developed to remove the subjectivity with which

these discrete units are drawn.

Acknowledgments

Field data was collected through the tireless work of N. Bentley, K. Colson, K.
DiChristina, A. Lucas, M. McCarty, and P. Murphy. A. Asada and M. Cook assisted with
data entry. J. Russel and G. Mann of the Bureau of Land Management provided us with
the field data from the Wildhorse allotment. J. Qi, S. Riley, G. Roloff, and R. Unnasch
provided insightful discussion during this study and invaluable review of this manuscript.
This work was supported, in part, by the Idaho Chapter of The Nature Conservancy, the
M. J. Murdock Charitable Trust, the Lava Lake Foundation for Science and
Conservation, and The Nature Conservancy’s Rodney Johnson and Katherine Ordway
Science Endowment. The Shoshone district of BLM provided logistical assistance during

field data collection.

90

Table 3.1. Summary information for the ten successively coarser segmentation levels

used in this study. The segmentation scale parameter (p3), a unitless number specified
in the multi-resolution segmentation algorithm described by Baatz and Schape
(2000), controls the size of the image objects and the within-object variance. Some
authors have suggested using changes in between- and within-object variance as a
way to identify appropriate levels of image segmentation.

 

 

Variance of
Object
Segmentation Number Means Mean
Scale of Median (Between- Within
Par arneter Image Area Minimum Maximum object Object
(Ps) Objects (ha) Area (ha) Area (ha) Variance) Variance
5 10452 5.77 0.09 491 0.0159 0.004583
10 2660 21.53 0.63 1401 0.018545 0.005676
15 1 185 46.7 0.72 1618 0.021825 0.006299
20 694 74.09 2.07 2344 0.026782 0.006743
25 491 1 17.24 2.97 2129 0.028742 0.007078
30 351 170.29 3.69 2204 0.032919 0.007329
35 267 228.62 7.20 2219 0.035934 0.007505
40 226 282.25 10.44 7640 0.038275 0.00772
45 186 324.92 10.44 3976 0.039849 0.0079
50 162 393.66 10.44 6230 0.04 0.007997

 

91

 

 

 

 

 

92

menmd Ewmd Sim wmemd Seed m _eed eemdfi ode—d emeed meeed em

mmmnd Seed 3&6 eeeed :ceed eeeed deg: mmed meeed emeed me

eened emeed an _ .v 3m _ .e Seed weeed mmw.e_ emwmd emeed mmeed ev

emwed Reed m med. eeeed wveed eeeed 93.2 ememd Seed e Bed w m

domed Smed Edd ehmmd mmeed m Heed mat“: evmmd mmeed N _eed em

Seed eemed Rad Seed meeed mmeed 3m. _ N meevd eveed w . ee.e mm

_ Sod emmed m d Nd ewe _ .e neeed fl Sed em d .2 wen _ .e mveed eeeed em

emved mneed emmd emvmd Seed Seed afie— NM: fie nmeed neeed 2

enved mmmmd 3&2 Gumd eeeed Nmeed Sedm dead emeed eeeed e—

dged emwmd neflfl mwtcd 5 .ed eveed din—m Em _ .e mmeed meeed m

me So and mmz Em awwzz d5 $2 :5 5392 as
owned amend $68me

cote—2.80 _§Emom 38%on 23m
:ofivaM 5:858me

BEE—3-380

 

.638 053 wEwE commmocwoc 05 use .258 $qu mosecmémmo— Bum—805w 05 com comm—oz? mmob “8-28-262 sweep:
355.8% mg» 252w 23 8082. 32030 98 38:65 5953 mco_8_otoo 628m assented 8m. 35 38.—wot? 5953
gauges é 5% 8:8 am 9 somwsz .83.:le .Ewouamze owe Emééwwsz .quuam .emooduammsz
62:30.5 wage—.8 05 was @2320on 20m 05 82% 238w 83 H522— 8.“ Gamers,» .8395 2E. .35. couflcofiwom
some 3 mains“ commmouwoc 05 com 3:62 acumezgémob 98 82? _§Em8 use 38:65 05 no 8356830 EEwotm> .N.m 035.

 

  
 
 
 
 
  
    
 

 

D Study Area Boundary
Land Ownershi
1T3 BLM

D Private

717?. State of Idaho

- Lava Flows

 

 

 

 

:
5.1 ,

 

 

 

 

 

N d
A fl 0 :P g F .30 Idaho F.1d]:
f rse u

.Twin Falls

 

 

 

 

0 10 20

Figure 3.1. Location and ownership of the Bureau of Land Management’s Wildhorse
allotment in southern Idaho.

 

93

 

 

emrvarrance 'Y

S

 

 

Distance between points (i.e., lag)

 

Figure 3.2. Example of a sample (empirical) semivariogram (black points) and the
variogram model (heavy solid line).

94

Wildhorse Bare-ground Variogram model

I l l l

 

0.015 —

0.010 ‘

Semivariance of % Bare Ground

0.005 "

 

 

 

 

T V l l
5000 10000 15000 20000
Distance (m)

Figure 3.3. Empirical semivariogram (dots) for percent bare ground cover in the
Wildhorse Allotment from field observations. The variogram model (line) has a nugget of
0.0036, sill of 0.0147, and range of 14,719m. The nugget-to-sill ratio (N SR), a measure
of the proportion of variability in the observations not explainable by distance between
locations, was 0.2474.

95

 

0.75

P
\1
r

 
  

Correlation Coefficient (R)
.0
0)

.°

01

or
O

'0 GLS Regression Correlation
if Bigfisl‘flfirgnfloflfi'atfl‘i
0.5 I

5 1 0 15 20 25 30 35 40 45 50

Segmentation Scale Parameter

 

 

 

Figure 3.4. Changes in correlation with segmentation scale between predictions of
percent bare ground from generalized least-squares (GLS) regression and regression
kriging (RK). For GLS regression, there is some evidence for a slight increase in
correlation as scale coarsens (solid line, slope = 0.0013, p=0.0917 for test of slope = 0).
For RK, there is no trend for increasing slope (dashed line, slope=0.0001, p=0.7744 for
test of slope = 0). The R2 values reported on the graph are for the fit of the points to the
trend lines. Correlations reached local maximums at segmentation levels of 20 and 35 for
GLS and 20, 35, and 45 for RK. Segmentation levels of 20 and 35 had variogram model
parameters that most closely matched the field-derived empirical variogram for percent
bare ground cover.

96

.32.? 38:55 2: do 85203—5 33% 38.55 98 mfisemmfi commmocwou 2: 82m oocoecoqoe Ream 2688

8 328 03% 9:322: demand :03 $2 2589 mESwotaZEom .038 $5 285m dmnum 3 a: 35233 33% engined—om
05 8 3:83 29: 2:83 mos—g 36:55 Be mEEwothEom douanEwom owns: do £26— Eocotme See “a 2368 commmocwoc
$qu 33:3 ammo. dam—88% 05 80¢ $3632 23 283235 95ch 9:3 8082. How mESonZEom Butfifim .m.m 059m

 

 

 

 

   

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Ase Scams E 8:905 E magma E5 eczema
eeeeN eeemv eeeer deem eeeeN eeemp eeeeF deem eeeem eeemw eeeeF deem eeeem eeemp eeeew deem
_ _ r _ p _ _ _ _ _ _ _ may
m.
A
w
I I . mood e.
. Mu...
o o O
r r o o .. Sod n...
o 9
o . o s.
m
d _ _ q W
b _ _ _ a
m.
L r «8.0 m
W.
1 I veed U
m
r I wood 0.
w...
.. r r I L 1 wood A
m
n
_ _ _ _ _ _ fl d _ q q q %

mm 250$ mu 250w ON 030w m 250%

97

0.6

0.5

,3 0.4 7, v i, 7 _ F. _
$2
a? 03 fiepL _ _ ‘—
fi Obs 35
0
8 ‘ 3Q 10.
3
z 0.2 ------- —m v i ____ _. -07 __ 7
5
2°. 15 50 '

0.1 H A — h sw—g a , _ -

0 ,

14,000 16,000 18,000 22,000

 

 

45
O

 

Variogram Range (m)

Figure 3.6. Plot of variogram range versus nugget-to-sill ratio (N SR) for the field
measurements of bare ground cover and generalized least-squares (GLS) regression

predictions for the ten segmentation levels considered. Segmentation levels closest to the
field-observation variogram parameters had the highest correlation between predicted and

observed values for percent bare ground.

98

 

0.8

 

 

 

 

 

0 75 jeLmsion f _0
' Alissmsion Kriging-
= A
'3 07
s . ’ A *W _
g _ ‘ R2=0.005 A AA
° 1'— —————-l-——-—_
° 065 ——— -0 h A * _i ,
C ' A I A
.2
3
2
D
O
0:
0.55 r a a of .___j _
0.5 e

 

0 1000 2000 3000 4000 5000 6000 7000 8000
Difference in Variogram Model Range from Field Variogram (m)

Figure 3.7. Correlation between predicted and observed percent bare-ground cover for the
10 segmentation levels plotted against the difference between the range of the variogram
model of the predictions and the range of the variogram model derived from the field
observations. For generalized-least-squares (GLS) regression only, there is some
evidence that accuracy of regression predictions decreases as the spatial dependence
(measured by the variogram model range) of the image objects becomes more different
than the spatial dependence of the field observations (solid line, p=0.1728 for test of
slope = O). For regression kriging (RK) there is no trend apparent (dashed line, p=0.8463
for test of slope = 0). The R2 values reported on the graph are for fit of the points to the
trend lines.

99

0.045

 

 

0044? ~ ~ ~ 0

0.035 ~~ ——fi

   
 

0.03 * r

o Between-object Variance F

 

0.025 --

 

Tasseled-Cap Band 3 Variance

 

 

 

0.02 ~- -— ——-— i I Within-object Variance
0.015 <+ * * ~
0.01 -«~- ~ * ~~~ w a i f
------------------ I

_, -_._’___... ————— .-_-— I -' I

0.005 til—u’L- ~- ~ a ~ ~~ V f k» «e _ i a
o . . .
0 50 100 150 200 250 300 350 400
Median Object Size (ha)

Figure 3.8. Changes in between- and within-object variance for the third tasseled-cap
band as segmentation scale increased. Because all bands showed similar patterns and the
third tasseled cap band was a highly significant variable in regression models at all
segmentation levels, we have shown only results from that band. Wiens (1989) proposed
that plots of variance against scale on a log-log plot would have slopes between 0.0 and -
1.0 because as scale increased there would be a decrease in between-obj ect variance as
more of the landscape heterogeneity was contained within each grain. Contrary to this
prediction, we found that between-object variance increased with segmentation. This was
likely due to the fact that objects were defined in a manner to minimize within-object
variability and therefore preserved between-object variance. Lines are power functions fit
to the variance measures and would be straight lines on a log-log plot.

100

References

Addicott J. F., J. M. Aho, M. F. Antolin, D. K. Padilla, J. S. Richardson, and D. A. Soluk.
1987. Ecological neighborhoods: scaling environmental patterns. Oikos, 49:340-
346.

Addink E. A., S. M. de Jong, and E. J. Pebesma. 2007. The importance of scale in object-
based mapping of vegetation parameters with hyperspectral imagery.
Photogrammetric Engineering and Remote Sensing, 73:905-912.

Allen T. F. H. and T. B. Starr 1982. Hierarchy: perspectives for ecological complexity,
University of Chicago Press, Chicago, IL.

Anderson G. L., J. D. Hanson, and R. H. Haas. 1993. Evaluating landsat thematic mapper
derived vegetation indices for estimating above-ground biomass on semiarid
rangelands. Remote Sensing of the Environment, 45:165-175.

Baatz M.and A. Schéipe. 2000. Multiresolution segmentation - an optimization approach
for high quality multi-scale image segmentation. Pages 12-23 in J Strobl, T
Blaschke, and G Griesebner, editors. Angewandte Geographische
Informationsverarbeitung XII. Wichmann-Verlag, Heidelberg.

Bailey T. C. and A. C. Gatrell 1995. Interactive spatial data analysis, Addison-Wesley.

Bedell T. E. 1998. Glossary of terms used in range management. 4th Edition, Society for
Range Management. Direct Press, Denver, Colorado.

Bertiller M. B., J. O. Ares, and A. J. Bisigato. 2002. Multiscale indicators of land
degradation in the Patagonian Monte, Argentina. Environmental Management,
30:704-715.

Buono, J ., Heilman, P., Williams, D., and Guertin, P. Assessing indicators of rangeland
health with remote sensing in southeast Arizona. RMRS-P-36, 508-510. 2005.
Fort Collins, CO, USDA Forest Service, Rocky Mountain Research Station.
USDA Forest Service Proceedings.

Burnett C. and T. Blaschke. 2003. A multi-scale segmentation/object relationship
modelling methodology for landscape analysis. Ecological Modeling, 168:233-
249.

Chavez P. 8., Jr. 1996. Image-based atmospheric corrections - revisited and improved.
Photogrammetric Engineering and Remote Sensing, 62: 1025-1036.

Cressie N. 1993. Statistics for spatial data, revised edition, Wiley and Sons, New York,
NY.

101

Crist E. P. and R. J. Kauth. 1986. The tasseled cap de-mystified. Photogrammetric
Engineering and Remote Sensing, 52:81-86.

Cullinan V. I., M. A. Simmons, and J. M. Thomas. 1997. A Bayesian test of hierarchy
theory: scaling up variability in plant cover from field to remotely sensed data.
Landscape Ecology, 12:273-285.

Farina A. 1998. Principles and methods in landscape ecology, Chapman & Hall, London,
UK.

Feitosa R. Q., G. A. O. P. Costa, T. B. Cazes, and B. F eijo. 2006. A genetic approach for
the automatic adaption of segmentation parameters. Salzburg University,
Salzburg, Austria.

Fortin M. J. and M. Dale 2005. Spatial analysis: a guide for ecologists, Cambridge
University Press, Cambridge, UK.

F uhlendorf S. D. and F. E. Smeins. 1999. Scaling effects of grazing in a semi-arid
grassland. Journal of Vegetation Science, 10:731-738.

Garrigues S., D. Allard, F. Baret, and M. Weiss. 2006. Quantifying spatial heterogeneity
at the landscape scale using variogram models. Remote Sensing of the
Environment, 103:81-96.

Gilabert M. A., J. Gonzalez-Piqueras, F. J. Garcia-Haro, and J. Melia. 2002. A
generalized soil-adjusted vegetation index. Remote Sensing of the Environment,
82:303-310.

Goovaerts P. 1997. Geostatistics for natural resource evaluation, Oxford University Press,
New York, NY.

Harris A. T. and G. P. Asner. 2003. Grazing gradient detection with airborne imaging
spectroscopy on a semi-arid rangeland. Journal of Arid Environments, 55:391-
404.

Hay G. J ., T. Blaschke, D. J. Marceau, and A. Bouchard. 2003. A comparison of three
image-object methods for the multiscale analysis of landscape structure. Journal
of Photogrammetry and Remote Sensing, 57:327-345.

Hay G. J .and D. J. Marceau. 2004. Multiscale object—specific analysis (MOSA): an
integrative approach for multiscale landscape analysis.in SM de Jong and FD van
der Meer, editors. Remote Sensing and Digital Image Analysis: including the
spatial domain. Kluwer Academic Publishers, Dordecht, Germany.

Hengl T., G. B. M. Heuvelink, and D. G. Rossiter. 2007. About regression-kriging: from
equations to case studies. Computers & Geosciences, 33: 1301-1315.

102

Hengl T., G. B. M. Heuvelink, and A. Stein. 2004. A generic fiamework for spatial
prediction of soil variables based on regression-kriging. Geoderma, 120:75-93.

Herrick J. E., J. W. Van Zee, K. M. Havstad, L. M. Burkett, and W. G. Whitford 2005.
Monitoring manual for grassland, shrubland, and savanna ecosystems, USDA-
ARS J omada Experimental Range, Las Cruces, New Mexico.

Home J. K. and D. C. Schneider. 1995. Spatial variance in ecology. Oikos, 74:18-26.

Hunt E. R. Jr., J. H. Everitt, J. C. Ritchie, M. S. Moran, T. D. Booth, G. L. Anderson, P.
E. Clark, and M. S. Seyfried. 2003. Applications and research using remote
sensing for rangeland management. Photogrammetric Engineering and Remote
Sensing, 69:675-693.

Jelinski D. E. and J. Wu. 1996. The modifiable areal unit problem and implications for
landscape ecology. Landscape Ecology, 11:129-140.

Keane, R., Lutes, D., Key, C., and Benson, N. FIREMON - sampling methods
documents. 2005. Moscow, ID, Fire Research and Management Exchance
System, National Biological Information Infrastructure.

Kim M. and M. Madden. 2006. Determination of optimal scale parameter for alliance-
level forest classification of multispectral IKONOS images. International
Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences,
Salzburg, Austria.

Kotlair N. B. and J. A. Wiens. 1990. Multiple scales of patchiness and patch structure: a
hierarchical framework for the study of heterogeneity. Oikos, 59:253-260.

Kravchenko A. N. 2003. Influence of spatial structure on accuracy of interpolation
methods. Soil Science Society of America Journal, 67: 1564-1571.

Krige D. G. 1966. Two-dimensional weighted moving average trend surfaces for ore-
evaluation. Journal of the South Africa Institute of Mining and Metallurgy, 66:13-
38.

Levin S. A. 1992. The problem of pattern and scale in ecology. Ecology, 73: 1943-1967.

Marsett R. C., J. Qi, P. Heilman, S. H. Beidenbender, M. C. Watson, S. Amer, M. Weltz,
D. Goodrich, and R. Marsett. 2006. Remote sensing for grassland management in
the arid southwest. Rangeland Ecology and Management, 59:530-540.

National Research Council 1994. Rangeland health: new methods to classify, inventory,
and monitor rangelands, National Academy Press, Washington, DC.

Navulur K. 2007. Multi-spectral image analysis using the object-oriented paradigm, CRC
Press, Taylor and Francis Group, LLC, Boca Raton, FL.

103

Neubert M., H. Herold, and G. Meinel. 2006. Evaluation of remote sensing image
segmentation quality - further results and concepts. International Archives of
Photogrammetry, Remote Sensing and Spatial Information Sciences, Salzburg,
Austria.

O'Neill R. V. 1989. Transmutations across hierarchical levels. Pages 59-78 in GS Innis
and RV O'Neill, editors. Systems Analysis of Ecosystems. International
Cooperative Publishing, Fairland, MD.

O'Neill R. V., D. L. Deangelis, J. B. Waide, and T. F. H. Allen 1986a. A hierarchical
concept of ecosystems, Princeton University Press, Princeton, New Jersey.

O'Neill R. V., A. R. Johnson, and A. W. King. 1989. A hierarchical framework for the
analysis of scale. Landscape Ecology, 3:193-205.

O'Neill R. V., A. R. Johnson, and R. V. King 1986b. A hierarchical concept of
ecosystems, Princeton University Press, Princeton, New Jersey.

Odeh I. and A. B. McBratney. 1994. Spatial prediction of soil properties from landform
attributes derived from a digital elevation model. Geoderma, 63:197-214.

Odeh I., A. B. McBratney, and D. Chittleborough. 1995. Further results on preditiction of
soil properties from terrain attributes: heterotopic cokriging and regression
kriging. Geoderma, 67 :2 1 5-226.

Owen, D. E. Geology of Craters of the Moon. 2008. Arco, ID, Craters of the Moon
National Monument, National Park Service.

Pebesma E. J. 2004. Multivariable geostatistics in S: the gstat package. Computers &
Geosciences, 30:683-691.

Pellant, M, Shaver, P., Pyke, D. A., and Herrick, J. E. Interpreting indicators of rangeland
health, version 4. Technical Reference 1734-6. 2005. Denver, CO, US.
Department of the Interior, Bureau of Land Management, National Science and
Technology Center.

Peters D. P. C., J. Yao, L. F. Huenneke et al. 2006. A framework and methods for
simplifying complex landscapes to reduce uncertainty in predictions. Pages 131-
146 in J Wu, KB Jones, and OL Loucks, editors. Scaling and uncertainty analysis
in ecology: methods and applications. Springer, Dordrecht, The Netherlands.

Pinheiro J. C. and D. M. Bates 2000. Mixed-effects models in S and S-PLUS, Springer-
Verlag, New York.

Pyke D. A., J. E. Herrick, P. Shaver, and M. Pellant. 2003. What is the standard for
rangeland health assessments? Pages 764-766 Durban, South Africa.

104

Qi J ., A. Chehbouni, A. R. Huete, Y. H. Kerr, and S. Sorooshian. 1994. A modified soil
adjusted vegetation index. Remote Sensing of the Environment, 30:1311-1319.

Soil Survey Staff. U.S. general soil map (STATSGO2) for Idaho. USDA Natural
Resource Conservation Service . 7-5-2006a. 3-3-2009a.

Soil Survey Staff. Keys to soil taxonomy, 10th ed. 2006b. Washington, D.C., USDA
Natural Resource Conservation Service.

Turner M. G., R. V. O'Neill, R. H. Gardner, and B. T. Milne. 1989. Effects of changing
spatial scale on the analysis of landscape pattern. Landscape Ecology, 3:153-162.

Wang L., W. P. Sousa, and P. Gong. 2004. Integration of object-based and pixel-based
classification for mapping mangroves with IKONOS imagery. Intemation Journal
of Remote Sensing, 25:5655-5668.

Webster R. and M. A. Oliver. 1991. Sample adequately to estimate variograms of soil
properties. European Journal of Soil Science, 43:177-192.

Webster R. and M. A. Oliver 2007. Geostatistics for environmental scientists, John Wiley
& Sons Inc., Hoboken, NJ.

Wiens J. A. 1989. Spatial scaling in ecology. Functional Ecology, 3:385-397.

Wiens J. A. and B. T. Milne. 1989. Scaling of 'landscapes' in landscape ecology, or,
landscape ecology from a beetle's perspective. Landscape Ecology, 3:87-96.

Woodcock C. and V. J. Harward. 1992. Nested-hierarchical scene models and image
segmentation. International Journal of Remote Sensing, 13:3167-3187.

Wu J. 2004. Effects of changing scale on landscape pattern analysis: scaling relations.
Landscape Ecology, 19: 125-1 38.

Wu J. 1999. Hierarchy and scaling: extrapolating information along a scaling ladder.
Canadian Journal of Remote Sensing, 25:367-380.

Wu J ., B. K. Jones, H. Li, and O. L. Loucks 2006a. Scaling and uncertainty analysis in
ecology, Springer, Dordrecht, The Netherlands.

Wu J .and H. Li. 2006b. Concepts of scale and scaling. Pages 3-15 in J Wu, KB Jones, H
Li, and OL Loucks, editors. Scaling and Uncertainty Analysis in Ecology:
Methods and Applications. Springer, Dordrecht, the Netherlands.

105

CHAPTER 4 - MAKING SPATIAL PREDICTIONS OF ATTRIBUTES OF
RANGELAND ECOSYSTEMS USING REGRESSION KRIGING

Abstract

Sound rangeland management requires accurate information on rangeland condition over
large landscapes. The typical approach to making spatial predictions of attributes related
to rangeland condition (e.g., shrub or bare ground cover) is via regression between field
and remotely-sensed data. This has worked well in some situations but has limited utility
when correlations between field and image data are low and does not take advantage of
all information contained in the field data. I compare, for three rangeland attributes
(percent cover of shrubs, bare ground, and cheatgrass [Bromus tectorum L.]) in a
southern Idaho study area, spatial predictions from generalized least-squares (GLS)
regression to a geostatistical interpolator, regression kriging (RK) that combines GLS
regression with spatial interpolation of the residuals to improve predictions of rangeland
condition attributes over large landscapes. I employed a remote-sensing technique,
object-based image analysis (OBIA), to segment Landsat TM 5 image pixels into
polygons (i.e., objects) because previous research has shown that OBIA yields higher
image-to-field data correlations and can be used to select appropriate scales for analysis.
Spatial dependence, the decrease in autocorrelation with increasing distance, was
strongest for percent bare ground (samples autocorrelated up to a distance [i.e., range] of
12,646m), but present in all three variables (range of 3,653m and 768m for shrub and
cheatgrass cover, respectively). As a result, RK produced more accurate results than GLS
regression alone for all three variables measured by cross-validated root mean-squared

error. The ability to create maps quantifying how prediction confidence changes with

106

distance from field samples is a significant benefit of regression kriging and makes this
approach suitable for landscape—level management planning. The results of RK could be
used in assessments of rangeland conditions over large landscapes and, if repeated over

time, could be used for landscape-level rangeland monitoring.

Introduction

Sound rangeland management and decision-making requires accurate information
on the condition of rangelands over large landscapes. Definitions of rangeland condition
may vary depending on management objectives, but they usually include factors related
to the integrity of soil, hydrologic, and/or vegetation processes (National Research
Council 1994). Many of these factors are measures of continuous attributes like percent
shrub cover, bare ground, or biomass. For example, when managing Sage Grouse
(Centrocercus urophasianus) habitat, percent cover and height of sagebrush (Artemisia
spp.) and grasses are important (Connelly et al. 2000; Crawford et al. 2004). Grazing
impacts are frequently measured by utilization (Holechek et al. 2001) or amount of bare
ground (Pickup et al. 1994). Stocking levels are determined, in part, by the amount of
available forage biomass (Holechek 1988; Hunt et a1. 2006; Karl et al. in review).

The use of remote sensing to map or make spatial predictions of rangeland
attributes has been widely explored, in part, because field sampling is usually not an
effective means for collecting data over large areas (Hunt et al. 2003). Interpolation
between field sampling locations is also difficult in heterogeneous landscapes. The hope

with remote-sensing is that sensor technologies and processing methods will yield an

107

efficient method for sampling large landscapes and making reliable spatial predictions of
rangeland attributes.

Typically, the approach to making spatial predictions of rangeland attributes from
remotely-sensed imagery has been to develop a statistical model that relates the field
measurements to the image band values at those points (Forster 1980; Dymond et al.
1992; McKenzie et al. 1999; Qi et al. 2002). The resulting statistical model is then
applied to all the pixels in the image. While this works well in some situations, it has
limited utility when the correlation between field and image data is low, and it does not
take full advantage of all of the information that can be obtained from the field samples.

Geostatistical interpolators can also be used to make spatial predictions of
continuous variables using field samples. Predicted values for unsampled locations are
determined by a combination of the surrounding samples weighted by their distance to
the unknown location (Fortin et al. 2005). Kriging, originally developed by Krige (1966)
for mineral exploration, is one of the most widely used geostatistical interpolators
because of its fidelity to the sample data (i.e., prediction errors are minimized close to
sample points) and ability to generate spatial predictions of standard error (see discussion
below). In short, kriging uses a model of the spatial dependence between samples (i.e.,
how autocorrelation changes as a function of distance between samples) as well as the
distance to neighboring sample points to create estimates at unknown locations (Bailey et
al. 1995). The number and distribution of sample points affect the scale of predictions
that can be made via statistical interpolators. This is especially true for kriging as sample
points of varying distances apart are needed to estimate spatial autocorrelation. Given the

cost and time required to collect field samples, it is often difficult to obtain enough

108

samples to make spatial predictions at a scale that is useful to rangeland managers using
statistical interpolators alone.

Hybridized approaches have been developed that combine regression models with
geostatistical predictors. Such approaches yield tighter integration of field and remote-
sensing data — making better use of high-cost field observations and improving the
accuracy of remote-sensing-derived map products. Several studies have found that these
hybridized methods perform better than either approach used separately (Odeh et al.
1994; Odeh et al. 1995; Goovaerts 1997; Hengl et al. 2007).

One hybridized approach is a derivation of kriging called regression kriging (RK).
Regression kriging first seeks to establish a linear relationship between the field samples
and a secondary dataset, typically a remotely—sensed image, and then uses kriging to
predict the value of the residuals of the regression model to improve the prediction
(Hengl et al. 2004; Hengl et al. 2007; Odeh et al. 1995; Odeh et al. 1994). In RK, the
predicted value at an unsampled location is the sum of the regression prediction and the
kriged residuals. The technique of RK has been used successfully in many applications
including mapping precipitation (Lloyd 2005), leaf-area index of boreal forests
(Berterretche et al. 2005) and predicting soil properties (Odeh et al. 1995; Lopez-
Granados et al. 2005; Yemefack et al. 2005) but has not yet seen wide application in
rangeland management.

Regression kriging has the potential to be a broadly useful technique for making
spatial predictions of aspects of rangeland condition because it provides an approach to
tightly integrating field and remote sensing data to get the most information out of each

source. My objective was to demonstrate the utility of the RK approach to provide

109

information on attributes of rangeland ecosystems in a southern Idaho landscape. I
compare the RK method to linear regression between field and satellite image data for
percent shrub, bare ground, and cheatgrass (Bromus tectorum L.) cover in a southern
Idaho landscape (Figure 4.1). Finally, I provide discussion on the uses and limitations of
the RK technique to providing information for rangeland management and getting the

most information out of available data sources.

Regression Kriging

A brief introduction of the concepts of geostatistical predictors as they apply to
kriging is offered prior to discussing RK. For more detailed discussion of geostatistical
prediction, see Bailey and Gatrell (1995), Goovaerts (1997), and Fortin and Dale (2005).

Kriging, like other geospatial predictors, has at its core one of the fundamental
rules of geography that“... near things are more related than distant things” (Tobler
1970, p. 236). In practical terms, this means that observations of a variable at nearby
locations are not independent (i.e., they are spatially autocorrelated). In most cases, the
spatial autocorrelation is highest when two observations are very close and decreases as
distance between observations grows until some point when observations can be
considered to be independent. In the context of the classical statistics that most ecologists
are taught, autocorrelation is problematic and efforts are made to ensure independence of
sample data. In geostatistics, however, spatial autocorrelation is useful for improving
predictions if the autocorrelation can be characterized.

Kriging makes predictions of a variable at unknown locations by taking a

weighted average of the variable’s value from sample locations where it was measured.

110

The weights are selected to minimize prediction error variance using: 1) the distance that
each sample point is to the unknown location and 2) the spatial autocorrelation of the
sample locations (Hengl et al. 2007).

Considering a sample of n observations of a rangeland variable (e. g., percent
cover of shrubs) taken at known locations S,- that can be written as 2(51), 2(32), ..., 2(S,J,

the predicted value of the variable at an unknown location is:
n
2(50) = 2 11'2“.) (1)
i=1

where 2 (S 0) is the predicted value at the unknown location so and the 7», are the weights

applied to the i observations. The optimal weights for any unknown location are found by
way of a variance/covariance matrix of the observations and the spatial covariance of the
observations to the unknown point. The spatial covariances between known and unknown
locations, of course, are not known and must be estimated via a model of the spatial
dependence (i.e., the change in covariance with distance between observations) of the
target variable. Spatial dependence is quantified using semivariance, a measure of
dissimilarity between two locations that is a function of the distance, h, between them
(Fortin et a1. 2005). Semivariance is calculated between all pairs of observations as:

N?) =2” -—(-h)ZIZ(S--) 2(S- +10] (2)

where ”701) is the estimate of semivariance for all observations separated by a distance h,
and n(h) is the number of points separated by a distance of h (Fortin et al. 2005). In
practice, distances are combined into distance groups (i.e., lags) to make calculation

easier. A plot of estimated semivariance over a range of lag distances is called an

111

empirical semivariogram and is an expression of the spatial dependence of the target
variable (Figure 4.2). A model that is fit to the empirical variogram (see Bailey et al.
1995; F ortin et al. 2005) estimates the spatial covariances needed to find the kriging

weights. The semivariogram model specifies: l) the portion of the total observed variance
(0'2) that cannot be explained by lag distance (i.e., the nugget), 2) the proportion of the

total observed variability (i.e., the sill) that can be explained by distance (i.e., nugget to
sill ratio), and 3) the maximum distance over which spatial autocorrelation is present (i.e.,
the range).

Kriging has a number of assumptions that must be met before it can be
successfirlly used. One assumption of kriging is that the predictor variables be normally
distributed (for discussion of normality assumptions in kriging, see Bailey et al. 1995;
Hengl et al. 2004). Often, data must be transformed to achieve a linear relationship
between the predictor and response variables. Predicted values from kriging can be easily
back-transformed, but problems exist with back-transforming estimates of variance.
Hengl et al. (2004) provide an example for finding (1 -a)% confidence intervals for
kriging predictions made with logistic-transformed data.

A second assumption to which the results of kriging are very sensitive is that the
data must be spatially stationary, or in other words, lack any kind of first-order trend with
respect to the geography of the study area. The presence of a significant trend in the data
may result in misspecification of the spatial dependence necessary to perform kriging and
biased predictions.

Several methods have been proposed for “dc-trending” data prior to kriging. An

understanding of a few of these methods is helpful to avoid confusion and clarify how

112

RK works. Kriging performed on data that have no geographic trend is called ordinary
kriging (Bailey et al. 1995). Universal kriging incorporates estimation of trend along with
the spatial prediction. Many authors agree that the term universal kriging should be
reserved for cases where the trend is modeled as a function of only the positions of the
sample points (i.e., using latitude and longitude coordinate values) (Hengl et al. 2004). If
other variables are used to estimate trend, the procedure is known as kriging with external
drift (KED). With KED, the equations for making predictions from sample data are
solved all at once via generalized least squares (Hengl et al. 2004). However, the trend
and residuals can be fitted separately and the results summed together to obtain the
spatial prediction. This procedure was called RK by Odeh et al. (1994; see also Odeh et
al. 1995) and has the advantage that it is more stable than KED and can be used with a
variety of different regression methods (Hengl et al. 2004).

Regression kriging starts with a standard multiple regression predictor for
determining the relationship between a measurements of a dependent variable, 2, and a

set of p predictor variables - designated with q rather than the standard notation of x to

avoid confusion with coordinate values (Hengl et al. 2007) - taken at known locations S ,~:

Z(Si)=:flk *qk(Si)+ei (3)

where flk are the regression coefficients and e,- is the residual, accounting for the

difference between the predicted and observed value. In RK, the regression coefficients
are usually found using GLS in order to account for spatial autocorrelation of the
residuals (Cressie 1993; Hengl et al. 2004). Typically, reflectance values or metrics

derived from ratios of image bands (e. g., NDVI) from satellite imagery are used as the

113

predictor variables. Hengl et al. (2004) recommended using a linear combination of the
satellite imagery such as principal components to improve the regression results.

In the context of standard regression, the residuals over all predictions are
assumed to be independent with mean zero. Residuals are considered to be variation in
the data that cannot be explained by the predictor variables, but if the residuals exhibit
spatial dependence then an estimate of the value of the regression residual at unknown
locations can be made using ordinary kriging. To accomplish this, semi-variances are
calculated over a range of lag distances from the GLS model residuals using Equation 2
above and an empirical semi-variogram constructed. A semi-variogram model is fit and

used to derive the kriging weights for estimating the residual at unknown locations.

The RK predictor for a rangeland variable at an unmeasured location, 2 (S 0) ,

then becomes the sum of the GLS regression prediction and the predicted residual:
P A n
z(s0)=Zflk ’qk(so)+Z’1i°e(Si) (4)
k=0 i=1

where qk(so) is the value of the kth predictor variable at the unknown location, the Xi are
the kriging weights determined from the 1' known locations using the variogram model of
the GLS-model residuals, and e(S,-) are the GLS-model residual value at point i (Hengl et

a1. 2007).

The variance of the RK prediction consists of the variance of the GLS regression
model plus the variance of the kriged residuals (see Cressie 1993; Hengl et al. 2004).
This composite variance reflects the increase in prediction uncertainty as the location gets
farther away from the observation point and further away from the regression mean.

When using transformed data, variance estimates of RK cannot be easily back-

114

transformed because they are not symmetric around the regression plane (Hengl et al.
2004). Hengl et al. (2004), however, provided an example of how (1-a)% confidence
intervals could be constructed from RK predictions with logistic-transformed data. This

example can be adapted for other transformations.

Study Area

The regression kriging approach to making spatial predictions of rangeland
attributes was applied to the 97,308 ha Bureau of Land Management (BLM) Wildhorse
Allotment in southern Idaho (Figure 4.1, 43.028°N, 113.864°W). The Wildhorse
Allotment lies within the Snake River plain and is partially within the Craters of the
Moon National Monument. Physiographically, the study area is mostly flat plateaus or
gently rolling hills, ranging in elevation from 1,272m in the southwest corner to 1,557m.
Precipitation ranges from 24.9cm to 32.6cm based on the PRISM map of average annual
precipitation from 1971 to 2000 (PRISM Group, Oregon State University,
http://www.prismclimate.org, created June 16, 2006).

Current vegetation communities are dominated by a mosaic of mountain big
sagebrush (Artemisia tridentata Nutt. ssp. vaseyana (Rydb.) Beetle), three-tip sage
(Artemisia tripartite Rydb.), and Basin big sagebrush (Artemisia, tridentata Nutt. ssp.
tridentata). Principal understory grasses are bluebunch wheatgrass (Pseudoroegneria
spicata (Pursh) A. Léve) and Idaho fescue (F estuca idahoensis Elmer). Cheatgrass
(Bromus tectorum L.) abundance is highly variable within the study area, reaching high
densities in disturbed sites and sites that have frequently burned. Dominant ecological

sites within the study area are loamy Basin big sagebrush/bluebunch wheatgrass,

115

Wyoming big sagebrush (Artemisia tridentata Nutt. ssp. wyomingensis Beetle &
Young)/bluebunch wheatgrass, and three-tip sage/Idaho fescue, and sandy Basin big
sagebrush/needle-and-thread grass (Hesperostipa comata (Trin. & Rupr.) Barkworth) —
indian ricegrass (Achnatherum hymenoides (Roem. & Schult.) Barkworth) (USDA
Natural Resources Conservation Service 2003)

This study area has seen an active fire history with 18 wildfires within the last 20
years, and 8 of those greater than 200 ha. Over the last 20 years, 80.04% of the Wildhorse
Allotment has burned. The frequent, large fires in this area have contributed to the spread
of cheatgrass and other invasive species in the allotment.

The majority of the Wildhorse Allotment is in public ownership with the Bureau
of Land Management (BLM) being the largest single land steward — managing
approximately 93,317 ha (95.8%) of the study area. Approximately 1,305 ha (1.3%) of
the study area is in private ownership, and 2,843 ha (2.9%) managed by the state of

Idaho. The main land use in the study area is cattle and sheep grazing.

Methods

For this study, I employed a remote sensing technique called object-based image
analysis (OBIA). In OBIA, the pixels in an image are first segmented into objects (i.e.,
polygons) in such a way that the pixels within an object are more similar than the pixels
in neighboring objects (Burnett et al. 2003; Baatz et al. 2000). I used the multi-resolution
segmentation (MRS) method developed by Baatz and Schape (2000) as implemented in
the Definiens Developer 7.0 program (http://www.definiens.com). In the MRS method,

contiguous pixels are initially grouped together to form an object. Subsequently,

116

neighboring objects that are similar (according to the parameters set) are merged into
larger objects until a threshold of heterogeneity is reached within the object. The MRS
method accepts a number of parameters that control how pixels and objects are merged
into larger objects. The color and shape parameters control the degree to which the
objects are defined by spectral versus textural information and the compactness of the
objects, respectively. A unitless scale parameter controls the size (i.e., scale) of the
objects by specifying the degree of similarity that will result in neighboring objects being
merged.

A unique feature of OBIA as opposed to other methods of scaling-up satellite
imagery is that highly distinct objects (e.g., disturbed areas, water bodies), even though
they may be small, can persist as image objects while the overall scale of the surrounding
objects become larger. This mimics the way that humans perceive the surface of the
earth when interpreting aerial imagery. By selection of the appropriate segmentation
parameter sets, objects defined through OBIA can approximate habitat patch structures
and be formed into scale hierarchies that correspond to different levels of ecosystem
organization (Blaschke et al. 2002; Wu et al. 2002; Wu 1999). This offers many
advantages to rangeland ecology.

For this study I created a set of image objects from Landsat Thematic Mapper
(TM) 5 imagery and derived mean and standard deviation of pixel values for each object
(see below). Regressions were run against these values for each object. Finally, the
kriging predictions were made for the centroids of the Landsat image objects and the

predicted values assigned to the entire image object.

117

Field Dela Collection

I used data that were collected by the Shoshone District BLM (G. Mann and J.
Russel, BLM, unpublished data). The original purpose of the field data collection was to
look at the effects of restoration activities and wildland fire in the Wildhorse Allotment.
Sampling sites were located randomly within the study area and 468 locations were
sampled between June 21, 2006 and August 6, 2008. A single 15.15m (50ft) transect was
set and percent cover of plant species was recorded using the line-point-intercept method
described by Herrick et al. (2005). The location of each sample point was recorded with a
GPS and differentially corrected.

Of the 468 original sites, I excluded 122 because the sites had burned between the
date they were sampled and when the satellite imagery was collected (see below). For the
remaining 346 observation sites, I calculated percent shrub cover as the number of
sample points where any shrub species was encountered in any of the canopy layers or as
a basal hit divided by 50. For the purposes of rangeland assessment and monitoring, bare
ground is considered land surface not covered by vegetation, rock, or litter (Bedell 1998;
Pellant et al. 2005). Percent bare ground cover was calculated as the proportion of the 50
points where no plant canopy was intercepted and the soil surface was recorded as
exposed soil (Herrick et al. 2005). Cheatgrass cover was calculated as the number of
points along the transect where cheatgrass was encountered in any of the canopy layers or

as a basal hit divided by 50.

118

Image Acquisition and Processing

I acquired an ortho-rectified Landsat Thematic Mapper (TM) 5 satellite image for
the study area from July 11, 2008. Landsat TM 5 is a multispectral sensor with pixel
dimensions (i.e., resolution) of 30m on a side and six bands (not counting the thermal
infrared band which I did not use): blue (0.45 to 0.52 pm), green (0.53 to 0.61 pm), red
(0.63 to 0.69 pm), near infrared (0.78 to 0.90 pm), and two short-wave infrared bands
(1.55 to 1.75 pm and 2.09 to 2.35 pm). I applied a dark-object subtraction method to
correct for the effects of atmosphere (Chavez, Jr. 1996), and converted the image values
to reflectance (i.e., percent of incident light that is reflected for each band).

The original multispectral bands of an image are highly correlated, and this can be
undesirable when segmenting the image in OBIA (N avulur 2007). I used the tasseled-cap
transformation (Jensen 1996) to obtain a set of bands having low correlations with each
other. The tasseled-cap transformation is a linear combination of the original image bands
that is defined such that each of the output bands has a specific interpretation (Crist et al.
1986). Tasseled-cap coefficients must be defined for each sensor, and the number of
tasseled-cap bands possible equals the number of original bands from the sensor. The first
tasseled-cap band is interpreted as brightness of the land surface. The second band is
interpreted as “greenness” and correlates highly with plant photosynthetic activity. The
third band is interpreted as wetness and correlates with vegetation and soil moisture.
Band four, in most instances, captures much of the noise in the image and generally is
discarded. Band five contains useful information (i.e., is not a noise band) but usually

does not have a clear interpretation.

119

I used bands one, two, three, and five of the tasseled-cap transformed image in
Definiens Developer’s (version 7.0) MRS algorithm to create a set of image objects for
the study area. The scale parameter of the segmentation algorithm was set so that the
median size of the image objects consisted of at least 20 pixels. All other segmentation
parameters were left at default values. Segmentation resulted in 13,646 image object
polygons with a median size of 5.22 ha (range 0.09 to 459.81 ha). To each image object
polygon, I assigned the mean and standard deviation of the pixels within the object for
each band. The image object polygons and their attribute values were exported to a GIS
layer. I intersected the image object polygon layer with the sample points to obtain a table
that had the field measurements, tasseled-cap image values, and coordinates for each
sample point. This table became the input for the statistical analysis. Additionally, I
created a point layer of geometric centroids from the image object polygons that was used
in the kriging analyses to predict the value of each image object.

Karl and Maurer (in prep) showed that there may be an optimal scale of
segmentation of an image for a variable being mapped. At scales below this optimal level
(including pixel-level analysis not using OBIA), regression will not account for all of the
spatial variability expressed in a semi-variogram of the original field observations, and
the model residuals will show spatial autocorrelation. In this case, RK may yield better
results than standard regression. At the optimal scale, the semi-variance of the model
predictions matches that of the original field data and the residuals show little or no
spatial autocorrelation. At this point, RK performs no better than standard regression for
making predictions. Optimal segmentation levels may vary for different rangeland

attributes depending on the nature of their spatial autocorrelation. Karl and Maurer found

120

near-optimal segmentation levels through analysis of successive levels of segmentation;
however, in practice, this method is currently time-consuming and cumbersome to
implement. Regression kriging, because it can extract more information from spatially
autocorrelated regression residuals, offers a way to achieve similar results using the same
set of image objects for different rangeland attributes and without having to worry about
finding a near-optimal segmentation level. For this reason, I chose a small scale
parameter for segmentation that was likely to be below the optimal level for any of the

three attributes I was considering.

 

StatisticaflAnirlysis

For the sake of exploring the benefits of RK for making spatial predictions, I
compared the results of the RK technique to a standard GLS regression approach. Each
method was repeated for percent shrub, bare ground cover, and cheatgrass cover. All
statistical analyses were done in R 2.7.2 (http://www.r-proj ect.org) using the CAR
package for data transformations (Fox 2000), GSTAT package for variogram modeling
and kriging (Pebesma 2004), and the NLME package for GLS modeling (Pinheiro et al.
2000).

The regression analysis started with 13 predictor variables: mean pixel values and
standard deviation of pixel values by image object for tasseled cap bands 1, 2, 3, and 5
and a soil-adjusted total vegetation index (Wallace et al. 2003). In addition, I included X
and Y coordinate values relative to the center of the study area and the product of the
relative X and Y coordinates. Through exploratory data analysis, I checked each variable

for normality and transformed them as necessary. I identified outliers (i.e., those points

121

 

with a value more than two standard deviations from the mean) for each variable, and
investigated each one individually using the study datasets and l-m color aerial
photography. If the investigation revealed that the outlier was atypical of the range of
variability found in the study area or another problem was identified, the point was
excluded; otherwise it was retained.

I used GLS to establish the linear relationship between the independent and
dependent variables. Percent shrub, bare ground, and cheatgrass cover were modeled
separately. Initial models included all variables. The only between-variable interaction
considered was between the X and Y coordinate values. I used a backward-stepwise
process to find a parsimonious model for each variable (Table 4.2). At this point, the
regression coefficients for the final model were taken and applied to the full set of image
objects to obtain the GLS-only prediction for the Wildhorse area.

The GLS method supports specification of the covariance and spatial
autocorrelation between the sample points (Cressie 1993). To employ this option,
however, requires a reiteration of the regression modeling because a variogram needs to
be constructed fi'om a set of model residuals to estimate spatial autocorrelation. The GLS-
only results were done without specifying any covariance structure. For RK, I constructed
an empirical variogram from this initial model’s residuals and defined a variogram model
consisting of a nugget, range, and sill. All variogram models used in this project were of
a spherical form. This variogram model was used to create a variance/covariance matrix
for a second GLS nm. The updated regression coefficients from this final model were

used to make the spatial predictions on the full image object dataset (Table 4.2). I used

122

the residuals from the final GLS model to construct a second variogram and define a
variogram model that would be used in the kriging portion of the spatial predictions.

The GSTAT package in R uses the GLS model and the variogram model defined
from the GLS residuals to predict the value of each of the three percent cover attributes at
the centroids of all the image objects. The kriging routine in GSTAT also produces a
variance estimate for each regression-kriged location that is the sum of the variance of the
regression estimate and the kriging variance.

Because I used a square-root transformation on my response variables, I squared
the model predictions to back-transform my GLS regression and RK predictions.
Goovaerts (1997) noted that the use of transformations with regression kriging can lead
to values in the results that are outside of the physical range of the response variable (e. g.,
negative values, percentages greater than 100). Hengl et al. (2004) suggested masking-out
or manually correcting such values. Accordingly, 1 limited percent cover predictions to
between zero and 100%. I also calculated and mapped 95% confidence intervals for the
regression kriging results for both percent cover variables using the following formula

adapted from Hengl et al. (2004):

2.0.9.09.)=(Z‘<S.)i1.96-6(S.>)2 (s)

where Z (S 0) is the predicted value at location So, and 6'(So) is the standard deviation

of the estimated value at location So.

123

Evaluating Performance Of Predictions

Root mean-squared error (RMSE) was a convenient measure of performance that
could be easily derived and directly compared for each method. I used a leave-one-out
cross validation method to calculate RMSE for the GLS regression model and RK
predictions. In leave-one—out cross validation, a sample point is omitted and a value
predicted for that point’s location. The omitted point is then replaced and another point is
omitted and predictions made again. This process is repeated until all points have been
omitted. The differences between the predicted and observed values when each point is

omitted are collected and used to calculate RMSE as:

 

n

Z(z<S.>—2<S.))2

_ i=1
RMSE _ n (6)

 

where 2(SJ was the observed value for location S,- and 2(5) was the predicted value for

the same location over all n points. I also used the cross-validation results to construct
plots of predicted versus observed values for each prediction method on each variable.
The RMSE can be standardized by total standard deviation of the observed
samples as a way to compare results between variables of different types (Hengl et a1.
2004). An advantage to using standardized RMSE is that there is no need to back-
transform the data. The more accurate a model is, the lower its standardized RMSE.
Hengl et al. (2004) proposed that a standardized RMSE of around 40% is a satisfactory
accuracy for predicting landscape attributes with regression kriging. I calculated

standardized RMSE for each prediction method on each variable to compare results.

124

Results

The GLS models for each of the three rangeland attributes had adjusted R2 values
ranging from 0.299 for percent bare ground cover to 0.531 for percent shrub cover (Table
4.2). Spectral (i.e., average pixel values per image object) and textural (i.e., standard-
deviation of pixel values per image object) were important in predicting cover of all three
variables. For percent shrub cover, there was a significant trend from east to west. For
percent cheatgrass cover, there was a significant trend from north to south, and for bare
ground cover, there was a second-order trend in both latitude and longitude. Factoring in
the geographic trends in addition to the variables derived from the Landsat imagery and
image objects is necessary to meet the stationarity requirements for kriging the model
residuals.

The empirical variograms of the GLS model residuals showed differing amounts
of spatial autocorrelation in the residuals for each rangeland attribute. For percent shrub
cover, the nugget-to-sill ratio (an expression of the proportion of the variability in the
residuals cannot be explained by distance) was 0.377 (Figure 4.3). The distance at which
residuals of the shrub cover GLS model were no longer spatially autocorrelated was
3,653m. Percent bare ground cover had a higher nugget-to-sill ratio, 0.448, meaning that
less of the variability in bare ground cover could be explained by distance (Figure 4.4).
The range of the bare-ground residuals, however, was much higher at 12,646m,
indicating that the spatial autocorrelation was present over longer distances. The residuals
for the percent cheatgrass cover model, had the lowest nugget-to-sill ratio at 0.023,

suggesting that values of residuals could be projected to unknown locations nearby the

125

field points (Figure 4.5). However, this spatial autocorrelation was operating over a
relatively short distance of 768m.

Cross validation run on the GLS model and the regression kriging showed that in
all cases regression kriging led to better predictions for the three rangeland attributes
(Table 4.3). For percent shrub cover, correlation between the original observations and
the cross-validated RK predictions was 0.7353 where the correlation for the cross-
validated GLS predictions was 0.7018 (Figure 4.6). Calculated RMSE decreased from
3.63% for the GLS model to 3.42% for the RK predictions. For percent bare ground
cover, cross-validation correlations were 0.5800 and 0.6475 for the GLS and RK
predictions, respectively (Figure 4.7). The RMSE for bare ground cover went from 6.78%
for GLS to 6.35% for RK. Finally, cross-validation correlations for percent cheatgrass
cover were 0.5092 and 0.5900 for the GLS and RK predictions, respectively (Figure 4.8).
Estimated RMSE decreased from 6.87% for the GLS model predictions to 6.42% for the
RK predictions.

Standardized RMSE gives a way to compare between different models (Hengl et
al. 2004), and a standardized RMSE of 0.4 or less is generally indicative of a good
prediction. Standardized RMSE values were less that 0.4 only for the percent shrub cover
predictions (Table 4.3). In all cases, though, RK produced smaller standardized RMSE
than the GLS model alone, and in the case of percent bare ground cover, the difference in
standardized RMSE between GLS and RK was large (81.28 for GLS vs. 54.23 for RK).

While the overall model fit and accuracy levels of the GLS and RK predictions
may be similar for a rangeland attribute, the spatial distribution of the predictions may

vary considerably. For percent shrub cover, most of the study area was predicted to have

126

less then 5% shrub cover (Table 4.4). The mean of the field shrub cover measurements
was 3.53% (SD 5.06, Table 4.1). RK predicted a somewhat lower proportion of the
landscape as in the less than 5% cover category than did GLS (Table 4.4). The spatial
distribution of percent shrub cover from RK showed a more aggregated pattern than the
GLS predictions (Figure 4.9). The higher degree of aggregation in the RK predictions is a
product of the spatial autocorrelation of the regression residuals. For bare ground the
majority of the study area was predicted to have between 5 and 15% cover of bare ground
(Table 4.4). The mean bare ground measurement across the field observations was 10.74
(SD 8.34, Table 4.1). For bare ground cover, the RK method yielded slightly higher
predictions on average over much of the landscape (Table 4.4). The spatial distribution
for percent bare ground cover was also more aggregated for the RK than for the GLS
predictions (Figure 4.10). The long range associated with the percent bare ground model
residuals led to the aggregation that is apparent in the RK predictions. Finally, the GLS
regression and RK methods both predicted the majority of the study area to be between 5
and 15% cover of cheatgrass (Table 4.4). The mean cover of cheatgrass found in the field
samples was 12.22% (SD 7.93, Table 4.1). In the case of cheatgrass, the RK predictions
were generally lower than the GLS predictions. The spatial distribution of cheatgrass was
largely similar between the GLS regression and RK methods (Figure 4.11). This is due to

the short range of spatial autocorrelation in the cheatgrass regression residuals.

Discussion
The results above demonstrate that regression kriging can increase the accuracy of

spatial predictions of rangeland variables over regression-based predictors alone. In some

127

cases, like percent bare ground, the difference between a standard regression approach
and RK can be large. However, this may not always be the case. Percent cover
measurements from a different study area showed no significant spatial autocorrelation
and RK gave no better predictions than GLS (J.W.K. and K. Colson, unpublished data).
This may be because there was little spatial autocorrelation in the rangeland attributes to
begin with or, more likely, that there weren’t enough field samples to detect and model
the spatial autocorrelation.

This speaks to an important point to consider, namely that geostatistical
predictors, including regression kriging, often carry a higher implementation cost than
standard regression-based methods. Sufficient samples must be collected to estimate the
variable’s spatial autocorrelation with a variogram model. A general rule of thumb used
among geostatisticians is that a minimum of 100 sample points is needed to construct a
reliable variogram model (Webster et al. 1991). In addition to the number of points, the
distribution of the points in the study area is also important. Because variance of the
regression-kriged predictor increases away from known locations, one would ideally
want good coverage of sample points across the study area. However, a regular point
spacing prevents estimation of spatial autocorrelation at lag distances shorter than the
distance between observation points, so it is also desirable to have some points located
close to each other in order to characterize short-range spatial autocorrelation. Lark
(2002), through simulation studies, found that the nature of the spatial autocorrelation of
the variable being mapped influenced the optimal configuration of sample points for
estimating the variogram. Variables showing spatial dependence over long ranges can be

sampled effectively with regularly-spaced sampling schemes. Variables exhibiting short-

128

range spatial dependence are best sampled by scattered clusters of points. In the case
where there is no a priori knowledge on the nature of the spatial dependence of a
variable, sampling along linear transects was deemed the most efficient for estimating
spatial dependence from the field observations. Webster and Oliver (2007) advocated that
when data are to be collected to estimate spatial dependence of a variable, a multi-stage
or nested hierarchical sampling design that produces a distribution of sample locations
that have a variety of distances between points (i.e., some sites are close together, some
are far apart) would be efficient at collecting data to estimate semi-variance.

While RK yielded improvements in the standardized RMSE for all three
variables, the predictions for percent bare ground and cheatgrass cover were still higher
than the 40% rule proposed by Hengl et al. (2004). Given the RMSE associated with the
bare ground and cheatgrass maps, these predictions should still be useful in assessing
large changes in cover across the study area. In large part, the high RMSE for these
variables is a result of the low correlation between the sample locations and the satellite
imagery, and several factors could be contributing to this. First, field-to-image
correlations generally decline as the pixel size of the native imagery increases (Karl and
Maurer, in prep). Repeating this study using higher resolution satellite imagery
segmented by OBIA may yield higher correlations. Second, the single 15.15m transect
measured at each field site may not be a good approximation of the conditions at that site.
More intensive sampling at each site may increase the correlation between field and
image data.

Knowing a priori whether or not RK will yield significantly better results than

GLS regression is difficult to determine. However, variograms of GLS model residuals

129

are relatively easy to construct as long as coordinate values are associated with the
measurements at each field point. If significant spatial structure exists in the residuals,
manifest as a low nugget-to-sill ratio and a long semi-variance range (Kravchenko 2003),
then RK should provide better predictions than GLS regression. Empirical variograms
with scattered points that have no discemable pattern indicate a lack of spatial structure,
and in these cases, RK generally will not perform any better than GLS and may actually
lower cross-validation results due to over-fitting of the data.

With any statistical technique, understanding the limits of the inference space is
important. Inference space refers to the limits of inferences, or generalizations, that can
be made from a set of observations or data (Dixon et al. 1994) and includes the range of
parameter values as well as the spatial extent of the inference. In terms of parameter (i.e.,
dependent variable) values, predictions made beyond the range of values that were
sampled in the field may be suspect because assumptions must be made that the
relationship between the dependent variable and the independent variables is the same
outside of the range of values sampled. For the variables considered in this study, limits
of inference for the variables are set by the minimum and maximum values observed for
each of these variables in the field (Table 4.1). This highlights the importance of
sampling across the full range of conditions if the objective is to produce predictive maps
across landscapes.

The spatial limits of inference are determined by both the distribution of the
sample points and the spatial autocorrelation of the variable. For a variable with a high
degree of spatial autocorrelation (i.e., would have a long range in a variogram model), the

spatial inference space would be larger than for a variable having low spatial

130

autocorrelation. Thus, the empirical variograms and variance estimates from RK are
useful in determining the spatial limits of inference from regression modeling.

The nugget of the variogram model represents the portion of the observed
variance in the field observations that cannot be explained by spatial autocorrelation (i.e.,
by distance). One determinant of the magnitude of the nugget is the variability in the
samples (Bailey et al. 1995). Sample variability can have two causes: 1) natural
variability in the system, and 2) imprecision in the methods used to sample the system.
The first cause cannot be directly controlled, but can be addressed in part through
measures like defining system bounds to minimize variability or employing stratification.
This could have been done in modeling the percent shrub cover attribute by first
classifying the satellite image to discriminate between sites with shrub cover and those
without, and then performing the RK only for those areas with shrub cover.

The second cause is under the control of the investigator, and methods involving
kriging argue for the use of precise methods and strong measures to minimize inter-
observer variability. The field measurements used in this study were made with a single
15.15m transect with only 50 measurements taken to calculate percent cover. The finest
cover measurement discemable using this method is 2%, and the single transect may not
have been located such that it was representative of the site. The more closely that
measurements reflect the true condition of the sample site, the smaller the variogram
nugget, and the more accurate the results of the kriged predictions can be. In practice,
though, the precision and accuracy of measurements at each site must be weighed against

the need to have a lot of sites sampled to estimate empirical variograms.

131

The spatial maps of confidence generated from RK are different from those of
GLS regression because the confidence decreases with distance from the field
observation points until the range of the variogram is reached. For locations beyond the
range of the residuals variogram, the confidence of RK is the same as that of the GLS
regression predictions. The ability to create maps depicting how confidence changes with
distance to observation points is a significant benefit of the regression kriging approach.
Maps of the RK variances show more directly the effects of proximity to the sample
locations and can be useful in determining how adequate sampling was for the kriging
aspect of regression kriging, or where additional samples might be located to improve or
increase consistency of predictions by suggesting additional places where sampling might

occur.

Implications

Regression kriging is a potentially powerful tool for making spatial predictions of
rangeland condition attributes. Like standard regression techniques, RK leverages the
correlation between field observations and remotely-sensed data to make predictions over
large landscapes. Its advantage over standard regression techniques, however, comes in
that it exploits the spatial autocorrelation of field observations to improve predictions —
thereby making better use of expensive and difficult-to-collect field data. Predictions
from RK are especially suitable for landscape—level planning efforts because the
uncertainty in the predictions and how it changes through space can be quantified. The
results of RK could be used in assessments of rangeland conditions over large landscapes,

and if repeated over time, could be used for monitoring purposes as well.

132

Acknowledgments

G. Mann and J. Russel from the Shoshone District BLM provided me with the field data
for this study. M. McCarthy, K. Colson, and K. DiChristina assisted in data collection
during 2008. B. Maurer, R. Unnasch, G. Roloff, S. Riley, and J. Qi provided thoughtful
reviews and editorial comments on this manuscript. This work was supported by grants
from the M.J. Murdock Charitable Trust, The Nature Conservancy’s Rodney Johnson and
Katherine Ordway Science Endowment, and the Lava Lake Foundation for Science and

Conservation.

133

Table 4.1. Summary statistics from the 346 field observations for the three percent cover
variables being predicted across the Wildhorse area: shrub, bare ground, and cheatgrass.

 

 

% Shrub % Bare Ground % Cheatgass
Minimum 0.00 0.00 0.00
Maximum 28.00 51.85 56.00
Median 1.41 8.66 11.57
Mean 3.53 10.74 12.22
Standard Deviation 5.06 8.34 7.93

 

134

Table 4.2. Generalized least squares (GLS) models for percent bare ground and percent
cheatgrass (Bromus tectorum L.) predicted from satellite image pixel data summarized by

image objects.

 

 

 

 

Variable % Shrub % Bare Ground % Cheatgrass
Intercept 0.6147*** 5.878*** 0233*”
Mean TCap 1 --- 0.272*** ---
TCap 3 0.715*** 0.238*** -0.270***
TCapS -0.937*** --- 0285*”
SATVI 1.0614* * * --- ---
Standard Deviation TCap 1 -0.2147*** --- ---
TCap 2 -0.1466* -0.195*** 0.216***
TCap 3 0.0983* 0.135*** ---
Positional RelX 0.2397*** -23.852*** ---
RelY --- -64.671*** -0.181***
RelX * RelY --- 77.117*** ---
R2 0.5311*** 0.299*** 0306*"
* p-value between 0.05 and 0.01

*** p-value less than 0.001

135

Table 4.3. Cross-validation results for the generalized-least-squares regression model
(GLS) and the regression kriging model (RK) predictions of percent shrub, bare ground,
and cheatgrass cover. R2 is the correlation between the predicted and observed values
from the cross validation. RMSE is the root-mean-squared error, and Standardized RMSE
is the RMSE divided by the standard deviation of the observed values after Hengl et al.
(2004)

 

 

Standardized
Model R2 RMSE RMSE
% Shrub Cover GLS 0.4925 3.63% 27.15
RK 0.5407 3.42% 25.55
% Bare Ground GLS 0.3356 6.78% 81.28
Cover RK 0.4173 6.35% 54.23
% Cheatgrass GLS 0.2593 6.87% 54.40
Cover RK 0.3481 6.42% 50.90

 

136

Table 4.4. Proportion of the study area in five percent cover categories for predictions of
percent cover of the three rangeland attributes make via generalized least-squares
regression (GLS) and regression kriging (RK).

 

Proportion of Study Area in % Cover Category
< 5% 5 -10% 10 -15% 15 - 20% 20 - 25% > 25%

 

Shrub Cover

RK Prediction 0.606 0.212 0.134 0.025 0.008 0.015
Lower 95% CI 0.630 0.224 0.108 0.019 0.005 0.013
Upper 95% CI 0.584 0.200 0.152 0.037 0.010 0.017

GLS Prediction 0.691 0.222 0.059 0.012 0.002 0.015
Lower 95% CI 0.722 0.209 0.045 0.009 0.001 0.015
Upper 95% CI 0.654 0.236 0.074 0.018 0.002 0.015

Bare Ground

RK Prediction 0.1 12 0.330 0.249 0.182 0.082 0.045
Lower 95% CI 0.157 0.334 0.236 0.179 0.061 0.034
Upper 95% CI 0.071 0.319 0.263 0.182 0.107 0.058

GLS Prediction 0.153 0.378 0.286 0.117 0.029 0.036
Lower 95% CI 0.192 0.397 0.263 0.093 0.021 0.033
Upper 95% CI 0.124 0.347 0.303 0.143 0.042 0.040

Cheatgrass Cover

RK Prediction 0.130 0.337 0.335 0.153 0.040 0.006
Lower 95% CI 0.180 0.364 0.315 0.1 19 0.018 0.003
Upper 95% CI 0.094 0.294 0.340 0.188 0.074 0.010

GLS Prediction 0.112 0.306 0.375 0.145 0.044 0.017
Lower 95% CI 0.143 0.339 0.358 0.123 0.025 0.013
Upper 95% CI 0.083 0.272 0.365 0.189 0.068 0.023

 

137

 

 

    
  

 

 

fl ET ' i ‘7 '. D Study Area Boundary
' Land Ownershi

A 11 BLM

L I” D Private

 

 

 

 

 

 

 

 

    
 
  

 

 

A if? :1: .30 Idaho Falls
_ '3 ise n .

Km 0
0 10 20 Thin Falls

Figure 4.1. The Wildhorse study area in southern Idaho.

 

 

138

 

 

emrvarrance y

 

S

 

V

Distance between points (i.e., lag)

Figure 4.2. Example of a sample (empirical) variogram (black points) and the variogram
model (heavy solid line).

139

% Shrub Residuals Variogram Model

1

 

0.010 ‘

0.008 ‘

0.006 "

0.004 ‘

Semivariance of transformed % cover

0.002 ‘

 

 

 

5000

1
10000

Distance (m)

15000

20000

Figure 4.3. Empirical variogram and spherical variogram model for the residuals of the
percent shrub regression model. The nugget, sill, and range were 0.0034, 0.0090, and

3,653m, respectively.

140

 

% Bare Ground Residuals Variogram Model

I l l J

 

0.012 ‘ °

 

 

Semivariance of transformed % cover

0.002 ‘

 

 

1 l 1
5000 10000 15000 20000

Distance (m)

Figure 4.4. Empirical variogram and spherical variogram model for the residuals of the
percent bare ground regression model. Semivariance is expressed in transformed units
squared. The nugget, or variance of the residuals unexplained by distance, was 0.0047.
The sill of the variogram model, or the total variance of the residuals was 0.0105 The
range of the variogram model, or distance at which locations are no longer spatially
autocorrelated, was 12,647m.

141

 

% Cheatgrass Residuals Variogram Model

 

 

 

 

 

 

l L 1 1
O
. O
o O
0.0124 ..0 ' .oo o .
. e . . e e . 0 e .
e e e .
O
. ' ~ 0. . ‘0 . .
g 0.010 "‘ k 9 . r.
O
8 ” . O
°\°
2 0.008 -
8
E
“5
0.006 -
8
C
(D
'C
(D
.2
g 0.004 -
a)
1’
0.002 -i
l l I l
5000 10000 15000 20000
Distance (m)

Figure 4.5. Empirical variogram and spherical variogram model for the residuals of the
percent cheatgrass regression model. The nugget, sill, and range, were 0.0023, 0.0100,

and 768m, respectively.

142

.330
n $32 808388 .smvaummzé 085 .88 258» 28 E88 wowed. someone 2: 50 as .385 u 5.32 802.08%
.o\om0.m mmEyc 3008 .560 389350 0:830 :23“:on 040 05 $0 80 32.9» 03.836 303 082020 .3 $20 .06 250E

.260 s .326 88620 .260 4. 2:5 caused

 

 

 

0.20 0V0 00.0 00.0 0N0 920 9.0 00.0 00.0
_ p _ _ .0 _ _ _ r _ .nv
.. ..u. 1w o
co 0 e 0.
$ 0 one o e I O 0
. . ...... .. ... .... m
0 O. .\O ‘. ‘ . 0 a a
e e \\ 000 co 0 . M M
O . . \\.\ O 0.. Im m- ”-
\\M O .00 s S
O \\\. O. O O . u- “-
xxxx 0 e e e O l W. m.
. \\ o o
. .0 w w
\\ l7u
.. . o... m
0 J J
o O mmmndnm m 0 0 w denm <

 

 

 

 

 

 

143

.232

u mass. Savages. .xxfmmsav .088 .98 258» as .523 89: 8.3%.: 05 80 2a £23 n mmzm BNsawcsm
£050 mmsav 302: 5.60 959% 9:3 68.50 "~232on mAO 05 A3 .80 83? 03.830 383 080053 .«o 305 .né oSwE

 

 

 

330 .x. 259.0 2am 330:5:— ._w>oo <9 2520 2am 0223.5
0N0 0N0 £20 0V0 00.0 V0 0.0 N0 _..0

_ _ _ _ _ O _ _ _ _ 0
IO 0

O O

m m

T0 8 IO 8

a 3

P P

..o m . ..o m

z a z m

9 9

T0 m. IO m

9 n 8 n

u u

P P

F0 al.. 10 %

.V O \ . .V O

o o

A A

O m r..O m
0 m5~00fl- 1.9 O 00m.0HM C..

 

 

 

 

 

 

0
<1

144

 

A000 .0

u 322 855%.... fiatonmmzé .25... .38 232w 23 .828 8...... Samoa... 2.. 80 Ba .32 u 3.5. 3.632%

$5.0 ”.525 3on .960 $83350 6082— :ommmohwou 01—0 05 A3 .00 mos—g 003030 25.3 033on 00 $20 .04. 2:me

.830 .\.. $893.5 33:3..."—

0N.0 0N_.0 0:.0 0J0 0W0 0&0

 

       

 

 

. \
O .0. f. t C.» .\
O. 0(r .3“ ;
.0
. ......
. C. C .
o to $ 0
0 o
.
.
a 82.1...

     

are

30

9‘0 v'o 9‘0 2'0

 

 

Janos % ssumeaug pauasqo

m

.050 .x. 380330 630:3...“
0N_.0 0...0 0_..0 0W0

 

 

 

O N0000NM

LTO 0'0

9'0 v'o 9'0 2'0

 

 

<1 1910:) % sswfiaeaug pauasqo

145

.08: BEBE 000000000 $00 0000: Q 05 :E: RE0E_-00:000=00 $3 0030— Am 0:3
000000000 050 0505— 03300000 5 .08: 32005000000000 $3 00003 046 AU .08: 32005000000000 $00 0030— 0.5 Am
60020000 020000000 A0400 8.803-600— 00N=8000w 9. .0000 8.8.0:? 0.0. E 00>00 0300 E00000 .00 000030000 .0030 .06 000E..—

    

         

        

00>00 .xbm A I
.300 ..03 8 2 I
.050 $2 8 S I

H .950 :2 o. m I

00.60 $0 8 0

.960 .x: v B

0:20 «:0000m

146

.08: 00300: 00002080 $3 0000: Q 000 :00: 32005000000000 $3 0030— Am 0:3 0000:0000
CDC wEwE 007.00qu a .000: _0>00000-00000w000 $3 00003 mAO AU .08: 03000000000580 $3 0030— mAO Am 5000:0000
0000803.— Amqov 00.8030000— 00N=00000w Q. .0000 00500:? 05 E 00>00 058% 0000 0000000 00 000006000 _0_000m .030 00=wE

 

           

00.50 $3 A I
00.50 $3 3 om I
00.50 $om 8 w— I
0900 $2 0. o. I
00.60 $3 8 n

00>00 $m v B
tfischmv unam— .X.

147

 

.000: 0020:: 0000000000 $3 0000: E 000 :00: 33807000000000 $3 0030— Am .005 000000000
33: @093 0000803.— 5 .08: 33900000020000 $3 0000 3 mam AD .000: _0>00.0_-00000_0000 $3 0030— wAO Am 00020000
00600030 Amqov 3.000.030.00— 00N=00000m 2 .0000 09500:? 05 0_ 00>00 30090000 0000000 00 300000000 _0=00m ._ :0 000wE

 

         

530 $3 A I
533 $3 8 ow I
00.60 $8 8 2 I
00.60 $2 8 S g
.650 $2 00 n

00060 $m v D
20.300000 $

148

References

Baatz M.and A. Schape. 2000. Multiresolution segmentation - an optimization approach
for high quality multi-scale image segmentation. Pages 12-23 in J Strobl, T Blaschke,
and G Griesebner, editors. Angewandte Geographische Informationsverarbeitung XII.
Wichmann—Verlag, Heidelberg.

Bailey T. C. and A. C. Gatrell 1995. Interactive spatial data analysis, Addison-Wesley.

Bedell T. E. 1998. Glossary of terms used in range management. 4th Edition, Society for
Range Management. Direct Press, Denver, Colorado.

Berterretche M., A. T. Hudak, W. B. Cohen, T. K. Maiersperger, S. T. Gower, and J. L.
Dungan. 2005. Comparison of regression and geostatistical methods for mapping leaf
area index (LAI) with Landsat ETM+ data over a boreal forest. Remote Sensing of
the Environment, 96:49-61.

Blaschke T., M. Conradi, and S. Lang. 2002. Multi-scale image analysis for ecological
monitoring of heterogeneous, small structured landscapes. Pages 35-44.

Burnett C. and T. Blaschke. 2003. A multi-scale segmentation/object relationship
modelling methodology for landscape analysis. Ecological Modeling, 168:233-249.

Chavez P. 8., Jr. 1996. Image-based atmospheric corrections - revisited and improved.
Photogrammetric Engineering and Remote Sensing, 62: 1025-1036.

Connelly J. W., M. A. Schroeder, A. R. Sands, and C. E. Braun. 2000. Guidelines to
manage sage grouse populations and their habitats. Wildlife Society Bulletin, 28:967-
985.

Crawford, J. A., Olson, R. A., West, N. E., Mosley, J. C., Schroeder, M. A., Whitson, T.
D., Miller, R. F., Gregg, M. A., and Boyd, C. S. Ecology and management of sage-

grouse and sage-grouse habitat. Journal of Range Management 57, 2-19. 2004. 12-15-
0004.

Cressie N. 1993. Statistics for spatial data, revised edition, Wiley and Sons, New York,
NY.

Crist E. P. and R. J. Kauth. 1986. The tasseled cap de-mystified. Photogrammetric
Engineering and Remote Sensing, 52:81-86.

Dixon P. M.and K. A. Garrett. 1994. Statistical issues for field experimenters. Pages 439-
450 in RJ Kendall and TE Lacher, Jr., editors. Wildlife toxicology and population
modeling: integrated studies of agroecosystems. CRC Press, Boca Raton, FL.

149

Dymond J. R., P. R. Stephens, P. F. Newsome, and R. H. Wilde. 1992. Percentage
vegetation cover of a degrading rangeland from SPOT. Intemation Journal of Remote
Sensing, 13:1999-2007.

Forster B. C. 1980. Urban residential ground cover using Landsat digital data.
Photogrammetric Engineering and Remote Sensing, 46:547-558.

Fortin M. J. and M. Dale 2005. Spatial analysis: a guide for ecologists, Cambridge
University Press, Cambridge, UK.

Fox J. 2000. An R and S-PLUS companion to applied regression, Sage Publications, Inc.,
Thousand Oaks, CA.

Goovaerts P. 1997. Geostatistics for natural resource evaluation, Oxford University Press,
New York, NY.

Hengl T., G. B. M. Heuvelink, and D. G. Rossiter. 2007. About regression-kriging: from
equations to case studies. Computers & Geosciences, 33:1301-1315.

Hengl T., G. B. M. Heuvelink, and A. Stein. 2004. A generic framework for spatial
prediction of soil variables based on regression-kriging. Geoderma, 120:75-93.

Herrick J. E., J. W. Van Zee, K. M. Havstad, L. M. Burkett, and W. G. Whitford 2005.
Monitoring manual for grassland, shrubland, and savanna ecosystems, USDA-ARS
J omada Experimental Range, Las Cruces, New Mexico.

Holechek J. L. 1988. An approach for setting stocking rate. Rangelands, 10:10-14.

Holechek J. L., C. H. Pieper, and C. H. Herbel 2001. Range management: principles and
practices, Prentice Hall, Upper Saddle River, New Jersey.

Hunt E. R. Jr., J. H. Everitt, J. C. Ritchie, M. S. Moran, T. D. Booth, G. L. Anderson, P.
E. Clark, and M. S. Seyfried. 2003. Applications and research using remote sensing

for rangeland management. Photogrammetric Engineering and Remote Sensing,
69:675-693.

Hunt E. R. Jr. and B. A. Miyake. 2006. Comparison of stocking rates from remote
sensing and geospatial data. Rangeland Ecology and Management, 59:11-18.

Jensen J. R. 1996. Introductory digital image processing, Second Edition edition.
Prentice-Hall, Inc., Upper Saddle River, NJ.

Kravchenko A. N. 2003. Influence of spatial structure on accuracy of interpolation
methods. Soil Science Society of America Journal, 67:1564-1571.

Krige D. G. 1966. Two-dimensional weighted moving average trend surfaces for ore-
evaluation. Journal of the South Afi'ica Institute of Mining and Metallurgy, 66:13-3 8.

150

Lark R. M. 2002. Optimized spatial sampling of soil for estimation of the variogram by
maximum likelihood. Geoderma, 105:49-80.

Lloyd C. D. 2005. Assessing the effect of integrating elevation data into the estimation of
monthly precipitation in Great Britain. Journal of Hydrology, 308: 128-150.

Lopez-Granados F ., M. Jurado-Exposito, J. Pena-Barragan, and L. Garcia-Torres. 2005.
Using geostatistical and remote sensing approaches for mapping soil properties.
European Journal of Agronomy, 23:279-289.

McKenzie N. and P. Ryan. 1999. Spatial prediction of soil properties using
environmental correlation. Geoderma, 89:67-94.

National Research Council 1994. Rangeland health: new methods to classify, inventory,
and monitor rangelands, National Academy Press, Washington, DC.

Navulur K. 2007. Multi-spectral image analysis using the object-oriented paradigm, CRC
Press, Taylor and Francis Group, LLC, Boca Raton, FL.

Odeh I. and A. B. McBratney. 1994. Spatial prediction of soil properties from landform
attributes derived from a digital elevation model. Geoderma, 63:197-214.

Odeh I., A. B. McBratney, and D. Chittleborough. 1995. Further results on preditiction of
soil properties from terrain attributes: heterotopic cokriging and regression kriging.
Geoderma, 67 :215-226.

Pebesma E. J. 2004. Multivariable geostatistics in S: the gstat package. Computers &
Geosciences, 30:683-691.

Pellant, M, Shaver, P., Pyke, D. A., and Herrick, J. E. Interpreting indicators of rangeland
health, version 4. Technical Reference 1734-6. 2005. Denver, CO, US. Department
of the Interior, Bureau of Land Management, National Science and Technology
Center.

Pickup G., G. N. Bastin, and V. H. Chewings. 1994. Remote-sensing-based condition
assessment for nonequilibrium rangelands under large-scale commercial grazing.
Ecological Applications, 4:497-517.

Pinheiro J. C. and D. M. Bates 2000. Mixed-effects models in S and S-PLUS, Springer-
Verlag, New York. .

Qi J ., R. Marsett, P. Heilman, S. Biedenbender, M. S. Moran, and D. Goodrich. 2002.
RANGES improves satellite-based information and land cover assessments in
southwest United States. EOS, Transactions of the American Geophysical Union,
83:601,605-606.

Tobler W. R. 1970. A computer movie simulating urban grth in the Detroit region.
Economic Geography, 46:234-240.

151

USDA Natural Resources Conservation Service. Soil survey geographic (SSURGO)
database for Blaine County area, Idaho. 2003. Ft. Worth, Texas, USDA Natural

Resources Conservation Service.

Wallace 0. C., J. Qi, P. Heilman, and R. Marsett. 2003. Remote sensing for cover change
assessment in southeast Arizona. Journal of Range Management, 56:402-409.

Webster R. and M. A. Oliver. 1991. Sample adequately to estimate variograms of soil
properties. European Journal of Soil Science, 43:177-192.

Webster R. and M. A. Oliver 2007. Geostatistics for environmental scientists, John Wiley
& Sons Inc., Hoboken, NJ.

Wu J. 1999. Hierarchy and scaling: extrapolating information along a scaling ladder.
Canadian Journal of Remote Sensing, 25:367-3 80.

Wu J. and J. L. David. 2002. A spatially explicit hierarchical approach to modelling
complex ecological systems: theory and applications. Ecological Modeling, 153:7-26.

Yemefack M., D. G. Rossiter, and R. Njomgang. 2005. Multi-scale characterization of
soil variability within an agricultural landscape mosaic system in southern Cameroon.
Geoderma, 125:117-143.

152

CHAPTER 5 - DEATH BY A THOUSAND PAPERCUTS: THE ROLE OF
INFORMATION AND ADAPTIVE MANAGEMENT IN THE MANAGEMENT OF
SAGEBRUSH ECOSYSTEMS

Abstract

Contentious, command-and-control (i.e., unilateral decision-making with little public
involvement) management of publically-owned sagebrush ecosystems in the western US.
has created barriers that have impeded management of sagebrush ecosystems that
continue to deteriorate due to threats unabated. Bolstering decision-making through 1)
better information tools and data to support management and 2) adaptive management
has been proffered as a means for making sound management decisions. Two recent
lawsuits involving rangeland management in southern Idaho suggest that neither of these
solutions is likely to be effective at managing rangelands at scales commensurate with
their threats unless there are changes to the underlying management paradigm governing
how the public participates in the management process. Alternatives to command-and-
control management exist, but no single management paradigm will be suitable for all
situations. Only with management approaches that provide a platform for discovering and
negotiating differences in values, knowledge, and assumptions between stakeholders can
adaptive management using the best available information work as intended for making

rangeland management decisions.

Introduction

Settlers of the American West called it the Sagebrush Sea — the over 100 million
acres in the Western United States covered by sagebrush-dominated ecosystems. Often

labeled as “barren wasteland,” this expansive rangeland habitat harbors a wide diversity

153

of plants and animals. It also provides essential ecosystem services such as erosion
control, regulation of hydrologic cycles, recreational opportunities, and rangelands for
livestock production. However, once considered endless, the Sagebrush Sea is being
increasingly degraded or converted to agricultural or residential landscapes. Across its
range, healthy sagebrush ecosystems are considered endangered, and in some locales,
unique sagebrush ecosystems are critically endangered (Noss et al. 1995). It is estimated
that less than 10% of sagebrush habitats remain intact (Dodge 2007), the rest degraded
fiom overgrazing and converted to agriculture or residential landscapes, and almost all
sagebrush habitats have been affected by livestock grazing (West 1995). Pellant (1990)
reported that over 2 million acres of sagebrush habitat in Idaho’s Snake River Basin have
been converted to sites dominated by invasive annual grasses that exclude native plants
and alter natural fire regimes, and other invasive species are spreading largely unchecked,
disrupting natural processes and out-competing native species (Asher et al. 1998).

The traditional view of managing for the health of sagebrush ecosystems was that
poor conditions were caused by some discrete disturbance — typically grazing. Removal
of that disturbance was presumed to be sufficient to restore an ecosystem to its original
state in all but the most egregious cases because ecosystems were seen as trending back
to a stable climax state (USDA Forest Service 2002; Briske et al. 2005). This
equilibrium-based theory of ecology has set human uses of rangelands such as grazing at
odds with rangeland health, and dichotomizes habitat degradation as either human caused
or not (Sayre 2005). It is clear now that these assumptions of linear progress back to a
climax community are too simplistic (see Briske et al. 2005) and that factors contributing

to declining conditions of sagebrush ecosystems have complex, synergistic causes that, in

154

some cases, will require active management even to maintain the status quo. The
interactive nature and evolution of new threats to rangeland health have led to a situation
where rangeland condition will continue to deteriorate unless the threats are directly
addressed through management action.

The majority of sagebrush habitat in the western US. is on publically-owned
lands for which the rights to graze are leased to private land owners, but regulated by
government agencies. Management of publically-owned sagebrush ecosystems in the
western US. has historically been contentious. Disagreement over management goals and
a general distrust of public land management agencies have created barriers to managing
sagebrush ecosystems (sensu Lachapelle et al. 2003), and the resulting frustration has led
to challenges of many proposed management actions either by ranchers or environmental
groups (Cawley 1993). The result has effectively been a management gridlock on public
rangelands where procedural requirements, risk-avoidance, lack of meaningful public
involvement spurs appeals, and litigation that have made management planning so
laborious that the plans that do make it to implementation stages rarely do so in a timely
fashion (USDA Forest Service 2002).

The objective of this paper is to explore the human processes that drive the
management of publically-owned sagebrush ecosystems and discuss options that have
been proposed for breaking the paralysis that is currently preventing more active
management to restore rangelands and address threats. I will examine the traditional
resource-management paradigm most commonly used in rangelands and how this can
lead to conflict escalation. Next, I will discuss two proffered solutions: 1) bolstering

decision-making by obtaining better data and developing information tools to support

155

management, and 2) adaptive management. I will illustrate these issues with case studies
from rangeland management in southern Idaho and discuss how, ultimately, rangeland
degradation is unlikely to stop unless the underlying management paradigm governing
how the public participates in the management process changes. Finally, I will explore
several alternatives involving differing degrees of public participation and discuss their
strengths and challenges in addressing evolving threats to the West’s sagebrush

ecosystems.

Natural Resource Management Paradigms

Environmental atrocities such as the harmful effects of pesticides documented by Rachel
Carson in Silent Spring (1962) and the burning of the Cuyahoga River, Ohio in 1969
fomented a movement in the 1960’s and 70’s that resulted in governmental regulations
intended to protect and improve the quality of the environment. The purpose of these
regulations was to prevent or ameliorate harmful impacts on the environment caused by
human activities. While they have clearly been effective at addressing many of the most
pressing environmental crises that led to their implementation, the institutionalization of
these regulations in government agencies has led to what has been described as a
“command-and-control” (sometimes referred to as “synoptic”) management paradigm
(Holling et al. 1996; Lachapelle et al. 2003).

Holling et al. (1996) described the command-and-control paradigm as one where
a problem is identified and a direct, feasible, and effective solution is implemented that
either controls the processes that cause the problem or addresses the effects. This method

works Well for easily definable phenomena that exhibit straightforward cause/effect

156

relationships and where there is a clear, universally-agreed-upon desired outcome. When
responding to gross enviromnean problems like those that prompted enactment of
existing environmental regulations, the initial use of command-and-control management
is usually successful (Holling et al. 1996). However, as the command-and-control
strategy becomes entrenched in bureaucracies, emphasis can shift from the original intent
of the regulations to a process-based system (USDA Forest Service 2002) whose focus
becomes cost efficiency and standardization. Holling et al. (1996) refer to this as the
”pathology of natural resource management.” As complex, nonlinear problems and
competing value systems are encountered, the command-and-control paradigm breaks-
down and conflict ensues.

The command-and-control management model says that all decisions should be
scientifically-based and made by experts. Under such a model, it is presumed that the
experts (i.e., the managers or others informing the managers) know the resource best and
therefore are most qualified to the decisions for their management (Riley et al. 2002;
Lachapelle et al. 2003). Also, under this model data are considered suspect unless
gathered, analyzed, and interpreted by experts, and public participation is primarily a
method of collecting or disseminating information (Lachapelle et al. 2003). Modern
natural resource management in US. government agencies has largely been built on a
command-and-control management paradigm (Holling et al. 1996).

A fundamental assumption of a command-and-control natural resource
management paradigm is that natural resource managers represent the public interest. It is
unrealistic, though, to expect that, in most circumstances, a single objective can be

defined that will please everybody. Lachapelle et al. (2003) concluded, “Assuming a

157

singular public interest in a pluralistic society may be fatal to any natural resource
planning process and results from implicit assumptions about goal agreement.”
According to Lachapelle et al., a command-and-control management paradigm does not
work for “wicked” and “messy” problems which they define as ones where: multiple
goals exist (the “wicked” part), there is a lack of clear information on cause-effect
relationships, limited time and resources to affect changes, and inequities exist in access
to information and in the distribution of power (the “messy” part).

Another problem with a command-and-control management paradigm applied to
natural resource management is that regulatory requirements aimed at increasing quality
and objectivity of management can stifle the decision-making process. Most government
agencies view procedural obligations as hurdles to overcome rather than opportunities to
improve the management process or involve stakeholders (Lachapelle et al. 2003). Such a
short-term risk-avoidance strategy has led agencies to tend toward creating legally
acceptable documents rather than seeking to resolve concerns or accommodate other
viewpoints (Wik et al. 2000).

Finally, public involvement in a command-and-control management framework
often fuels conflict by providing stakeholders with opportunities to voice their concerns
only as a formality Efforts to formalize and standardize public involvement in agency
management planning have left local communities feeling marginalized, decreasing their
trust in public agencies, and fueling opposition to planning efforts (Davenport et al.
2007). Public apathy toward natural resource management has been borne from cynicism,

lack of trust, and thinking that agencies don’t “acknowledge, hear, or honor what the

158

public has to say” (Lachapelle et al. 2003). As a result, the typical public response is to
forego any effort to engage the management agencies and go straight to court.

Ironically, if the response by the public (e. g., environmental groups or ranchers)
to failures of the command-and-control paradigm to address complex problems involving
different human-value systems is to demand more regulation and prohibition of actions,
as if often the case with lawsuits brought against resource management agencies, “then
the [command-and-control] pathology is deepened, because this applies a command-and-
control solution to a problem initiated by command-and-control” (Holling et al. 1996).
This situation may have occurred recently in a 2004 lawsuit brought against the Bureau
of Land Management (BLM) by the Western Watersheds Project (WWP).

The WWP is a non-profit organization headquartered in Idaho whose mission is to
reduce what it sees at the deleterious effects of livestock by eliminating grazing on public
lands. One of WWP’s main tools is litigation and the threat of litigation “. .. to enforce
our nation's environmental laws and improve management of our public lands”
(http://www.westemwatersheds.org/legal/legal.htrnl). Western Watersheds holds grazing
rights to approximately 4,000ac of Idaho state land and owns property in Idaho, giving it
standing to bring lawsuits against land-management agencies.

In their 2004 suit (WWP vs. K Bennett, BLM 392 F.Supp.2d 1217), WWP
alleged that the BLM did not follow National Environmental Policy Act (N EPA)
requirements in renewing grazing leases for 28 allotments in the J arbidge Resource Area,
Idaho by conducting separate ecological assessments for each allotment instead of a
comprehensive environmental impact statement (EIS). WWP also argued that the BLM

increased permissible grazing levels on the allotments in the face of decreasing sage

159

grouse populations and declining rangeland condition and despite a management plan for
the area that stated preference would be given to wildlife population and rangeland health
before grazing levels would be increased. The presiding US. District Court Judge found
in favor of WWP and ruled that BLM had violated the NEPA by failing to obtain public
comment on environmental assessments of proposed grazing changes and for not
adequately considering the cumulative effects of increasing grazing levels across the
majority of the allotments. The judge granted summary judgment to WWP and issued an
injunction prohibiting any grazing on the 28 allotments until a full environmental impact
statement, including public comment, could be conducted. Four years later, the
environmental impact statement has yet to be completed, although BLM was successful
in having the injunction partially lified.

This case is indicative of many other cases that are brought against government
land-management agencies each year. The fear of additional litigation has led to a higher
standard being applied to ongoing planning processes further delaying their progress and
adding to their cost — contributing to the “process predicament” (USDA Forest Service
2002). Sadly, cases like this one have also made the agencies more reticent to share
information and collaborate with the public. The end result is that fewer projects are
being executed that address threats facing sagebrush ecosystems that are, arguably, more

crucial than livestock grazing.

Conflict Escalation and the Flow of Information in Rangeland Management

Lachapelle et al. (2003) described five barriers to natural resource management

that may cause conflict in natural resource management:

160

. Inadequate goal definition — either poorly defined goals, well defined broad goals but
poorly defined specific objectives, or goals unacceptable to stakeholders.

. Lack of trust — both within management agencies as well as between stakeholders.
Constraints to building trust in public management efforts include: competing values
over the resource, knowledge gaps, limited stakeholder involvement, and agency staff
turnover (Davenport et al. 2007).

. Procedural obligations — requirements of NEPA and other regulations aligned toward
command-and-control management.

. Inflexibility - with regards to time, funding, personnel, and the ability to try new

approaches.

. Institutional design — entrenched values, culture and goals of management agencies

that exclude or isolate stakeholders. These fall under three main themes: power held

by agencies or staff, influence of special interests, and public apathy

These impediments are common in command-and-control management paradigm.

Conflict over natural resources is commonplace (Ayling et al. 1997), and arises

from differences in values given to natural resources and competition over them (Buckles

1999). Conflict can range from minor irritations and frustrations to outright violence, and

tends to escalate as engagement in the management process and opportunities for redress

become more difficult. As conflict escalates, stakeholders increasingly try to gain and

demonstrate power relative to other stakeholders (Yasmi et al. 2006). Any conflict may

go through a number of stages, but the goal of each stage is to gain advantage over the

other side and either secure or exclude access to the resource. Yasmi et al. (2006)

recognized eight stages of natural resource conflict, of which litigation was a later stage.

161

They found that while the exact pattern of escalation varied widely, there were common
escalation pathways. Half of the pathways they identified included litigation as either an
intermediate or terminal stage of conflict.

Early stages of conflict consist of one party trying to bring the other party around
to their way of thinking or to negotiate an arrangement that is amenable to both parties. A
core component of these interactions is information (Yasmi et al. 2006). In a friendly
environment, information can be shared openly to arrive at agreeable solutions. However,
in a situation of conflict, information flows become restricted. Information favorable to
one parties’ view is used to attack the other, and contradicting information is suppressed.
With regard to conflicts over natural resource management, this can manifest itself as an
information arms race where agency land-managers race to find data to make and back
their decisions, and opposing groups (not always environmental groups) race to offer
contrary data that contradicting agency opinions to force their desired outcomes. This
often leads to agencies trying to “bulletproof” planning efforts through attention to
process minutiae and “incontrovertible” information. (Lachapelle et al. 2003).

The result has been a management paralysis where agency managers are so
occupied with attending to details of regulatory requirements, fighting litigation, or
avoiding threatened litigation that few actions actually get done (USDA Forest Service
2002). Managers have invested heavily in technologies and research to fill in the
information gaps hoping that it will help them fight off legal actions. Uncertainty will
always exist, and therefore it is not possible to plug all of the holes or close all avenues
for special—interest groups to try to thwart management actions through appeals and legal

action. This, however, is treating the proximate cause of challenges to management

162

actions, and not the root causes of marginalized stakeholders — inflexibility in the
management process, and an unwillingness or inability to incorporate alternative
objectives/ goals. In order for natural resource managers to proactively treat and restore
rangelands while still providing for sustainable uses, the barriers preventing management
and fueling conflict must be addressed.

The command-and-control paradigm upon which public natural resource
management in the United States is built invites conflict by reliance on flawed
assumptions and an over-emphasis on process. Far from encouraging parties to work
together in natural resource management, command-and-control management as it is
implemented today provides disincentives to cooperation (Grumbine 1997). In order for a
party to have their views seriously considered in management decision-making, it often
involves legal action. This fits the pattern of organizations like WWP who prefer legal
actions to trying to work cooperatively with inflexible or unwilling government agencies.
The entry point for lawsuits contesting land management in rangelands often involves
how information is or is not used in making rangeland management decisions and the
quality and quantity of that information. There have been many proposals of ways to
address information gaps or uncertainties that are often seen as the Achilles’ heel of the
land management decision-making. Two of these are: 1) adaptive management, and 2)
the development of new information or information technologies for management

decision-making.

163

 

Adaptive Management

Adaptive management has been widely promoted as a way to make progress in
natural resource management despite uncertainty in causal relationships of the resources
or success of the intended intervention (Walters et al. 1990; Lee 1999; NEPA Task Force
2003; Walters 2007). Adaptive management considers management actions as
“experiments” that allow managers to take actions that address pressing needs even
though the success of the action or even the exact outcomes are uncertain or unknown
(Walters et al. 1990). Typically, adaptive management consists of four steps that are
repeated cyclically (figure 1, Walters et al. 1990; The Nature Conservancy 2001; Ruhl
2007): 1) recognition of the problem and definition of management objectives, 2)
selecting targets for management and developing strategies, 3) implementing
management actions, and 4) monitoring and evaluating performance - leading back to
revised or new management objectives.

Management of complex systems like rangeland habitats is difficult because ofien
objectives cannot be distilled into clear-cut actions and there may be little understanding
of what the response of the system will be to any particular action. In theory, this kind of
situation lends itself well to an adaptive management approach (Ruhl 2007). However, to
date, adaptive management has been more influential as a concept rather than as an actual
management model (Lee 1999). Many projects never make it past the defining strategies
stage (USDA Forest Service 2002), and for those that are implemented, monitoring is
ofien either not implemented or cut short as management priorities change (Walters
2007). In looking at the success of implementing adaptive management, Walters (2007)

concluded that the majority of his case studies were failures in that no actions were taken

164

or monitoring was not implemented. Walters traced these failures back to three
institutional problems: 1) a lack of resources and will for implementing requisite
monitoring, 2) an unwillingness on the part of the management agencies to accept
uncertainty, and 3) a lack of leadership and willingness to engage in a new management
paradigm. Lee (1999) asserted, however, that adaptive management still has the potential
to be an efficient method for addressing ecosystem threats and providing answers to
uncertainties about natural resource management.

Implemented under a command-and-control paradigm, adaptive management is
unlikely to solve the management paralysis faced by public land-management agencies
because of a reliance on philosophies and methods that short-change not only the types of
knowledge that can be considered for management (Cortner et al. 1999) but also the
identification and inclusion of a diversity of human values in natural resource
management (Lee 1999; Riley et al. 2003). Adaptive management can be implemented in
a way that includes a wider group of stakeholders in “participatory management” where
identification of threats and formulation of solutions can be shared among public land
managers and private citizens (Riley et al. 2003). Atypical of how it is applied in a
command-and-control paradigm, Lee (1999) proposed that adaptive management only be
used after stakeholders have agreed to a common set of goals for managing natural
resources. Ruhl (2007), however, argued that there is good reason to doubt whether
adaptive management can ultimately succeed within public agencies without substantial

changes in the administrative laws that established the command-and-control paradigm.

165

Data for Better Management Decisions

In an effort to plug holes in management plans left by lack of knowledge of base
conditions or uncertainty in outcomes, much effort has gone into developing new
information gathering, analysis, and communication tools. The National Research
Council (1994) highlighted the paucity of baseline information on rangelands in the US.
and called for a national rangeland inventory and monitoring system. This effort spurred
many new efforts to develop assessment and monitoring protocols (e. g., Pellant et al.
2005; Pyke et al. 2002) and data tools (e.g., Rangeview Geospatial Tools for Natural
Resource Management, http://rangeview.arizona.edu) to fill the gap. The widespread
adoption of information technologies like geographic information systems (GIS) have
made it much easier to assemble and integrate disparate information to inform
management (Lee 1999). Given the expense involved with collecting and analyzing field
data, many managers begun to look toward remote sensing and modeling technologies as
a way to gather or create information over larger landscapes.

These efforts are an attempt to reduce uncertainties associated with land
management and document the decision-making process in order not only to improve
management, but also to reduce the risk of legal challenges (Lachapelle et al. 2003).
Rangeland management is in desperate need of good information upon which to
formulate defensible management options that are likely to succeed (National Research
Council 1994), but in the context of a command-and-control management paradigm, a
fixation on information (i.e., removing all uncertainty) and risk avoidance can lead to
excessive analysis and contribute to management paralysis (USDA Forest Service 2002).

Also, value differences between stakeholder groups, a lack of agreement on what

166

qualifies as knowledge or information on rangeland ecosystems, and an insistence that
data be perfect before they can be considered useful for decision-making can lead to
information not being used in resource management decision-making even when it may

be specifically requested (van Wyk et al. 2008).

Will Better Information Improve Rangeland Management?

Calls have been made for the amount and quality of information that goes into
rangeland management to be improved (National Research Council 1994), but at the
same time management decisions must be made. A combined emphasis on applying the
principles of adaptive management and improving access to high-quality information is
necessary to advance rangeland management. However, focus on these two things alone
is not likely to change the current management gridlock being experienced. Both the
adaptive-management and information systems approaches failed in an attempt to revise
grazing management in central Idaho in part because they were hamstrung by the legacy
of an inflexible command-and-control agency management system which prevented their
fully intended use.

In 2005, WWP sued the US. Forest Service (USPS) and livestock operators over
grazing on four allotments in the Sawtooth National Forest and the Sawtooth National
Recreation Area (WWP vs. USF S, CV-05-189-E-BLW). Again, WWP asserted that the
agency failed to follow NEPA by not conducting environmental assessments of grazing
impacts in the allotments and by failing to use all information available to it on capability
of the land to be grazed in development of their grazing plans. The USFS countered that

the best-available information was flawed, although it made no attempt to quantify how

167

inaccurate it was or to document that it was even considered. Further, they argued that
their plan was based on principles of adaptive management and management would be
modified from monitoring data as needed. The same judge that heard the previous WWP
vs. BLM case heard this case also and ruled that USFS had violated NEPA standards by
failing to use information that it had (and had used in other situations) and failing to
disclose that information thereby stifling public participation. Furthermore, he ruled that
USFS could not use adaptive management as a proxy for analyzing requisite and
obtainable information to formulate a “best” management option and that their failure to
thoroughly define their adaptive management process was in violation of NEPA. The
USFS was ordered to redo its environmental impact statement of grazing on the four
allotments and grazing was enjoined until that was completed.

In the end, the private ranchers who leased the grazing rights for these four
allotments presented a detailed adaptive management plan to the court based on
independent assessment of grazing capability (Karl et al., in review) and field-based
monitoring, and they were successful at restoring grazing to the allotments while the
USF S completed the environmental impact statement. It is worth noting, however, that
the USFS’ proposed grazing plan would have reduced grazing on the Sawtooth National
Forest Allotments, just not by a large enough margin to satisfy WWP, and when the
injunction was lifted, grazing was restored to its original, higher level while the USPS
completed its mandated revisions.

There are many examples in rangeland management where a lack of information
hampered decision-making and led to poor or contested decisions. A persistent theme of

the WWP lawsuits in Idaho is that decision-making is arbitrary and capricious because it

168

happens without needed data. While many ecologists and managers are quick to point out
that they don’t have all the information they would like to have, it is not entirely clear if
and how effectively additional information would be used. The WWP vs. USFS lawsuit
cited above illustrates this dilemma — additional information was available but not used
for grazing management planning.

Several factors may be contributing to why available information is not used in
land management decision-making. The first of these is a reluctance to use data that is
known to be imperfect. This was the case in the WWP vs. USFS case where the Forest
Service managers did not trust the analysis of grazing capability because of its limitations
and inaccuracies. Second is uncertainty about what the information actually . portrays and
appropriate ways to use it. This is a common predicament where extraordinary effort is
invested into collecting data (either in the field or via remote sensing) and it never gets
used because it is unclear how to process, analyze, and interpret the data. A third reason
has to do with lack of data or data consistency across an entire study area. Land owners
and managers primarily concerned with their own lands and interests often create datasets
that stop at ownership boundaries and are subsequently of limited usefulness for broader
management planning. Finally, the regulations and institutional structure of many land
management agencies can be rigid to the point of making it inordinately difficult to
acquire and use new or different data (USDA Forest Service 2002).

Even if the state of data collection and use for rangeland management were to
instantly improve, it would likely make a limited contribution to alleviating decision-
making headaches because contested decisions and litigation would still mire down the

process. This is because under the command-and-control paradigm it is easy for

169

managers to find themselves at odds with groups of stakeholders, and one
disenfranchised stakeholder can throw a wrench in the works of decision making. In
order for information systems and adaptive management to have a significant
contribution to rangeland management, management institutions need to explore different
management paradigms and come to view command-and-control management as one end

of a spectrum of public engagement in decision-making.

Public Participation in Rangeland Management — Putting Command and Control at One

End of a Management Spectrum

Public participation in decision making involves defining a group of stakeholders
and giving those stakeholders (or a subset of them) roles in the decision-making process.
Stakeholders have been defined as individuals or groups with an interest, or stake, in the
decision being made (Margoluis et al. 1998) and can either affect a decision or action, be
affected by it, or both (Grimble et al. 1997). Identifying the right group of stakeholders is
not a trivial task, and for contentious or high-profile decisions, the pool of potential
stakeholders may be quite large. Grimble (1998; see also Grimble et al. 1997) proposed a
four-step method for identifying stakeholders and analyzing their needs and values based
on project objectives and context (see Grimble et al. 1997). Prell et al. (2007) advocated
using social network analysis to identify key participants from a large pool of
stakeholders. Such a detailed method may be, at present, beyond the capacity and time
constraints of most agencies. However, understanding the stakeholders of a management
decision is critical to being able to understand the impacts of that decision (Grimble et al.

1997) and to ultimately decide whether or not to involve the public in decision-making.

170

Public involvement in decision-making can take many different forms.
Traditionally in command-and-control management, agency decision-makers have relied
on a prescribed and formulated series of public meetings and comment periods to inform
stakeholders of management alternatives and to solicit feedback. This approach, while it
has worked in some situations, has ultimately engendered a good deal of frustration and
public disillusionment with the decision-making process (USDA Forest Service 2002).
Informal public participation methods such as field tours with groups of stakeholders or
partner meetings can also be effective for discussing objectives and soliciting feedback
when used at the right scale and in the right situation. At the other end of the spectrum is
consensus-based management where a formal stakeholder group is organized and vested
with the authority to make binding management decisions (see Brower et al. 2001).

Irvin and Stansbury (2004) point out, that engaging the public in decision making
incurs costs to the decision-making process in time and money, and that not all
management decisions need the same level of public engagement. The extent to which
the public should be involved in natural-resource management, though, is also influenced
by other factors such as: the nature of the decision being made, flexibility to extend
decision-making authority to the public, the values of the stakeholders, trust among
stakeholder groups, uncertainty and risk in the outcomes, and the potential for conflict
over decision.

The nature of the decision being made and the context in which the decision is
made greatly influences the amount and kind of public participation needed (Lawrence et
al. 2001). Some decisions are of small consequence or are non-controversial (e. g.,

treating invasive species with an approved herbicide). Public participation in this decision

171

may not need to extend beyond informing stakeholders prior to the action. At the other
end of the spectrum are large decisions that affect the management goals and
fundamental objectives of large landscapes (e. g., resource management plans) or involve
controversial topics (e.g., grazing levels or management of large carnivores). In these
situations diverse groups of stakeholders may need to be assembled to try to achieve a
consensus set of goals and objectives. In some cases enlisting public participation may be
counter-productive such as when the decisions that a manager makes may be strongly
constrained by state or federal regulations, when emergency decisions must be made.

Where the decision-making authority lies and how flexible that authority is in
allowing stakeholders to share it with public agencies is a large determinant of the nature
and degree of public involvement. Extensive public participation in decisions where
either the decision-making authority cannot be shared or the decision will be made
irrespective of the public input may be construed as misleading (Davenport et al. 2007) or
at best be a low-benefit and costly use of public participation (Irvin et al. 2004). A
paradigm of public involvement often involves innovative and flexible methods (Shindler
et al. 1999). Many agencies that are attempting new models for decision-making have
bumped up against regulatory limitations (Ruhl 2007) or institutional disincentives for
managers to try new approaches (Grumbine 1997). Other agencies, because they are
swamped with challenges to decisions, have over-emphasized regulatory requirements —
becoming mired in a “process predicament” that relegates public participation to town-
hall meetings and comment periods (USDA Forest Service 2002). Lachapelle et al.

(2003) identified institutional design as a barrier to natural resource management because

172

 

entrenched values, cultures and goals of management agencies exclude or isolate
stakeholders.

Each stakeholder may value wildlife or natural resources differently and for
different reasons than other stakeholders. When operating as an employee of a public
agency, the manager is perceived as representing the public trust and the values of agency
stakeholders are defined by their regulatory mandates. The degree to which a
management decision relates to a stakeholder’s values will determine their level of
interest in taking part in the decision-making process. Likewise, the degree to which
values of different stakeholders diverge will impact the potential for conflict regarding
the management decision and influence the form of public participation. In situations
where values are consistent, actions may be taken involving a small group of
stakeholders. As values diverge, however, the scale of public participation in the
decision-making process must grow and the decision-making process slows down as a
result.

Trust among stakeholders has been identified as a crucial prerequisite to effective
decision-making (Shindler et al. 1999). A lack of trust between stakeholders and
management agencies can lead to apathy about decision-making processes or attempts to
circumvent decision-makers through legislation and litigation (Lachapelle et al. 2003;
USDA Forest Service 2002). Building trust takes commitment on the part of the land-
management agencies to devote resources to the process and honor the process of public
participation and decisions it produces (Davenport et al. 2007; Ruhl 2007). Constraints to
building trust in public management efforts include: competing values over the resource,

knowledge gaps, limited stakeholder involvement, and agency staff turnover (Davenport

173

et al. 2007). In situations where low trust exists, opportunities for public involvement in
decision-making will be compromised, and until some level of trust can be restored,
participation, even when solicited, may be low and challenges to decisions should be
expected.

Uncertainty and risk in the outcomes of management actions can also help
determine the level of public participation in decision-making processes. For actions that
are risky or with uncertain outcomes, having public support for decisions is valuable.
When public support is valuable for backing management decision-making, an elevated
level of public participation is warranted (Lawrence et al. 2001).

The appropriateness of public involvement and the style of that involvement is
also dependent on the state of and potential for conflict surrounding the topic (Yasmi et
al. 2006). When levels of conflict are low, problems can be addressed through open
communication and facilitated discussion and the opportunities for public participation in
decision-making are high. As conflict escalates, stakeholders become more hardened in
their standpoints and the conflict must be addressed before public participation can be
effective.

Schindler and Cheek (1999) gave some additional suggestions for successfully
engaging the public in land management decisions. First, interactions between
stakeholders should be open and inclusive. Second, public participation will require
skilled leadership and development of long-term relationships between the stakeholders.
Involvement of stakeholders in the management process should be early and continuous,
providing the best opportunity to define management objectives agreeable to all parties

and vest a sense of ownership in the process. Finally, efforts of the stakeholders must

174

result in action being taken; reinforcing the idea that stakeholder participation is valued.
Once the decision has been made to involve the public in a land or resource management

decision, the manager must be committed to that process.

Deciding the Form and Extent of Stakeholder Participation

Balancing the costs of stakeholder involvement with the perceived and real
benefits of public engagement is needed for effective and timely decision-making, but
managers have not been given many tools to help them weigh these tradeoffs. Lawrence
and Deagen (2001) proposed a decision tree, based on the management-decision making
model by Vroom and Yetton (1973), for selecting the appropriate form of public
participation in natural-resource management. Their process was based on six questions
related to the importance of public commitment to decisions, congruence of goals among
stakeholders, time/cost, and the influence of public opinion. Recommendations ranged
from the manager making a unilateral decision to group decision-making.

The decision tree proposed by Lawrence and Deagren (2001) has several
important benefits. The first is that it explicitly recognizes that there are situations in
which public participation is not warranted or where limited public involvement or
command-and-control management would be most efficient. These options more
closely reflect real-world situations where the cost and delay of public participation may
not be necessary or justifiable in every case. While undoubtedly a great simplification
of a complex process, there is merit in the approach of distilling questions of whether
and how to involve stakeholders in decision-making to a few basic considerations.

Lawrence and Deagren suggest that forcing their six questions to have yes or no

175

answers could be a shortcoming of their approach, but ultimately should not hinder the
applicability of their approach as this is necessary anytime complexity is reduced to a
series of guiding rules.

Lawrence and Deagren’s (2001) model, however, does have some shortcomings
that should be addressed to improve its usefillness for managers. They suggest that their
decision tree naturally favors the most efficient (and ultimately cost effective) options,
yet many management decisions are time sensitive, and time and cost should be explicitly
incorporated in the model. Second, there are many situations in natural resource
management where past history with management agencies has created an atmosphere of
distrust that makes many stakeholders wary of unilateral decisions (Davenport et al.
2007) and level of trust should be explicitly incorporated. Third, the model does not
account for regulatory limitations to public participation or the degree to which the public
has the authority to provide input or make decisions. Lawrence and Deagen write that
giving the public a venue for expressing their views increases their acceptance of a
decision despite the fact that they had no authority or influence in making that decision.
Many authors have found this to have just the opposite effect — an increase frustrations
and decrease trust in agencies (USDA Forest Service 2002; Lachapelle et al. 2003; Irvin
et al. 2004; Davenport et al. 2007) — therefore the model should reflect agency limitations
to public involvement. A fourth shortcoming of Lawrence and Deagren’s decision tree is
that it only implicitly considers the size or context of a decision in terms of whether or

not it would require public participation.

176

Alternatives to C ommand-and-C ontrol Management

One alternative to command-and-control management is adaptive impact
management. Riley et al (2003) proposed that by shifting the focus of adaptive
management from the resource itself to ameliorating impacts of stakeholder-identified
threats to the resource, adaptive management could better address management problems
that are important to the stakeholders and the general public. They termed this approach
adaptive impact management and conclude that by focusing on impacts of threats,
Adaptive impact management has several distinct advantages over adaptive management
as it has traditionally been applied. First, it increases the relevance of resource
management to values that society holds. Second, it engenders greater stakeholder
satisfaction, resulting in fewer conflicts. Third, stakeholders create a management
fiamework that is more capable of withstanding change and dealing with uncertainty, and
finally, learning and sharing information becomes a natural part of the management
process.

Another alternative, which is at the opposite end of the participation spectrum
from command-and-control management, and that has worked in some cases, is
consensus-based management. Consensus-based conservation is a collaborative
management process where multiple stakeholders come together to define common
values, establish management objectives, develop action plans, and implement and
monitor the plans in an adaptive-management-style framework (Paulson 1998). A
fundamental assumption of the consensus management paradigm is that human values are
a core component of the management process (Lachapelle et al. 2003; Riley et al. 2003).

Instead of public land-management agencies defining objectives and making decisions

177

 

with minimal stakeholder input, in a collective-management approach the agencies
become one of a group of partners (but may retain leadership roles for moving the
process along). Each stakeholder acts according their own interests, promoting their
values in the management process. In this light, the public land manager’s role is to be an
advocate for the ecological health of the resource and secure opportunities for access to
an array of users. Scientific information is a crucial contributor to the management
process but decision-making is based on collectively agreed upon goals and values.

Consensus-based management seeks to build trust among the stakeholders. Trust
is the lodestar of collective management, and the examples of where it has been
successfully applied highlight the effort that has gone into building trust among the
stakeholders (Davenport et al. 2007). This does not require that stakeholders forsake their
identities or agree on everything, only that they be able to define some common goals for
management. Trust must be manifest on two levels: interpersonal — trust between the
people representing the stakeholders that the other is honest and genuine; and institutional
— trust that the agencies and other organizations participating are committed to the
collective management process (Davenport et al. 2007).

Consensus-based conservation is not without its limitations, however. The Upper
Colorado River Basin Recovery Program was initiated in 1987 to address continuing
development of water resources in the Upper Colorado River while simultaneously
protecting endangered fish species in the river (Brower et al. 2001 ). Given the large, and
diverse group of stakeholders, a consensus-based decision-making process was adopted.
The group negotiated a set of goals, and subsequently water developments have been

approved and steps taken to protect endangered fish populations. However, this

178

consensus-based approach is not working and has led to a decline in endangered fish
populations, made possible by special dispensation given to the project to allow it to
operate outside of normal federal authority over endangered species (Brower et al. 2001).
The failings of consensus-based conservation in this case were attributed to the values
and objectives of the program not being closely enough aligned with protection of the
endangered species, stakeholders being preoccupied with political agendas, and
disparities in the power ascribed to different stakeholders rendering some stakeholders
with more say than others (Brower et al. 2001). This led to a management process that
was vulnerable to control by special interests and made to look successful by reliance on
procedural goals and measures of success (e. g., number of studies launched).

Brower et al.’s (2001) recommendations for successfully implementing
consensus-based management are a useful caution for management under the collective
paradigm. They suggested that first, success be judged by the status or recovery of the
resource being managed and not by process goals. Second, regulatory agencies should
retain authority over the resource and reserve the ability to intervene if the condition of
the resource deteriorates. Finally, funding for collective management should be provided,
where possible, through a source with an agenda to protect the resource being managed to

offset the disproportionate influence of special interest groups.

Conclusions

The legacies of command-and-control management and trivial public involvement
in decision-making have engendered contention in rangeland management to the point

that actual or threatened legal actions have rigidly enforced a “status quo” management

179

situation. Meanwhile, threats to rangeland condition are operating at scales beyond what
typical, non-controversial management activities can address. Breaking management
paralysis for sagebrush ecosystems will hinge more on how we define, empower, and
hold accountable stakeholder groups than on what new information is brought to bear or
tools developed to aid decision-making. If this is the case, the questions remain of how
we decide who participates as a stakeholder and what form of management is employed
for different situations. Formal stakeholder analysis may be appropriate for large, and
potentially contentious, planning efforts or to develop a stakeholder pool for initially
moving beyond command-and-control management, but in most cases managers will
already have a pretty good idea who the stakeholders are. Decision rules such as those
described by Lawrence and Deagen (2001) appear, at present, to be the most useful way
to decide what level of stakeholder participation is appropriate. Case studies highlighting
the use of different levels of public involvement in different situations and how this
impacted information use, adaptive management, and ultimately conflict would also be
helpful.

Regardless of which form of management is used or how stakeholders were
identified, it remains clear that regulatory and legal systems in the United States are set
up so that one disenfranchised group (e.g., WWP) can make management difficult
through appeals and litigation. In the two court cases discussed above, the agencies were
faulted for failures to follow established procedures, inadequately considering the effects
of proposed management, and a failure to make use of the best information available to
the agency regardless of the fact that the data had obvious shortcomings. It is impossible

to say that a stakeholder-inclusive management approach would have resulted in different

180

management proposals and prevented the lawsuits from WWP. But it is evident that the
command-and-control paradigm under which the agencies have operated has mired them
in the “process predicament” (USDA Forest Service 2002) and encouraged the legal
challenges.

Management approaches that include meaningful stakeholder participation
provide a platform for discovering and negotiating differences in values, knowledge, and
assumptions. From here, direction can be set for how to acquire and use information so
that it becomes useful for making decisions (Sarewitz et al. 2007) and not an entry-point
for attacking managers. This offers a chance to break the management gridlock and start
taking actions to address the threats that sagebrush ecosystems in the western US. are
facing. By focusing on threats in the context of adaptive impact management,
stakeholders who have historically treated each other with acrimony may be able to find
agreement on some pressing management needs.

Such an approach has been successful when the Idaho Chapter of The Nature
Conservancy has partnered with the disparate groups of public and private land managers
to promote aggressive actions and policies for the prevention and control of invasive
species. The focus on weeds, “the one thing that we all love to hate” (A. Talsma, The
Nature Conservancy, personal communication) allowed groups with sometimes very
different sets of values to put aside their differences in other areas and move ahead with a

more crucial management need.

181

Acknowledgements

M. Kaplowitz, B. Maurer, G. Roloff, S. Riley, J. Qi, R. Unnasch, and R. Walker
provided thoughtful reviews and editorial comments on this manuscript. This work was
supported by the Idaho chapter of The Nature Conservancy and a grant from the M.J.

Murdock Charitable Trust.

182

 

References

Asher J. and C. Spunier. 1998. The spread of invasive weeds in western wildlands: a
state of biological emergency. Bureau of Land Management.

Ayling R. D. and K. Kelly. 1997. Dealing with conflict: natural resources and dispute
resolution. Commonwealth Forestry Review, 76: 182-185.

Briske D. D., S. D. F uhlendorf, and F. E. Smeins. 2005. State-and-transition models,
thresholds, and rangeland health: a synthesis of ecological concepts and perspectives.
Rangeland Ecology and Management, 58: 1-10.

Brower A., C. Reedy, and J. Yelin-Kefer. 2001. Consensus versus Conservation in the
Upper Colorado River Basin Recovery Implementation Program. Conservation
Biology, 15:1001-1007.

Buckles, D. Cultivating peace: conflict and collaboration in natural resource
management. 1999. Ottawa, Canada, International Development Research Centre
(IDRC)/World Bank.

Cawley R. M. G. 1993. Federal land, western anger: The Sagebrush Rebellion and
environmental politic, University Press of Kansas, Lawrence, KS.

Cortner H. J. and M. A. Moote 1999. The politics of ecosystem management, Island
Press, Washington, DC.

Davenport M. A., J. E. Leahy, D. H. Anderson, and P. J. Jakes. 2007. Building trust in
natural resource management within local communities: a case study of the Midewin
National Tallgrass Prairie. Environmental Management, 39:353-368.

Dodge S. R. 2007. Sagebrush in Western North Americs: habitats and species in
jeopardy. Pacific Northwest Research Station Science Finfinds, March 2007:1-2.

Grimble, R. Stakeholder methodologies in natural resource management. 1998.
Chatham, UK, Natural Resources Institute. Socioeconomic Methodologies. Best
Practices Guidelines.

Grimble R. and K. Wellard. 1997. Stakeholder methodologies in natural resource
management: a review of principles, contextx, experiences and opportunities.
Agricultural Systems, 55:173-193.

Grumbine R. E. 1997. Reflections on "What is Ecosystem Management?". Conservation
Biology, 11:41-47.

Holling C. S. and G. K. Meffe. 1996. Command and control and the pathology of natural
resource management. Conservation Biology, 10:328-337.

183

Irvin R. A. and J. Stansbury. 2004. Citizen participation in decision making: is it worth
the effort? Public Administration Review, 64:55-65.

Lachapelle P. R., S. F. McCool, and M. E. Patterson. 2003. Barriers to effective natural
resource planning in a "messy" world. Society and Natural Resources, 16:473-490.

Lawrence R. L. and D. A. Deagen. 2001. Choosing public participation methods for
natural resources: a context-specific guide. Society and Natural Resources, 14:857-
872.

Lee K. N. 1999. Appraising adaptive management. Ecology and Society, 3:3.

Margoluis R. and N. Salafsky 1998. Measures of success: designing, managing, and
monitoring conservation and development projects, Island Press,
Washington DC.

National Research Council 1994. Rangeland health: new methods to classify, inventory,
and monitor rangelands, National Academy Press, Washington, DC.

NEPA Task Force. Modernizing NEPA Implementation: the NEPA task force report to
the council on environmental quality. 2003. Council on Environmental Quality.

Noss, R. F., LaRoe, E. T. 111, and Scott, J. M. Endangered ecosystems of the United
States: a preliminary assessment of loss and degradation. Biological Report 28. 1995.
Washington, D.C., US Fish and Wildlife Service. 12-7-2004.

Paulson D. 1998. Collaborative management of public rangeland in Wyoming: lessons in
co-management. Professional Geographer, 50:301-315.

Pellant, M, Shaver, P., Pyke, D. A., and Herrick, J. E. Interpreting indicators of rangeland
health, version 4. Technical Reference 1734-6. 2005. Denver, CO, US. Department
of the Interior, Bureau of Land Management, National Science and Technology
Center.

Prell, C., Hubacek, K., and Reed, M. Stakeholder analysis and social network analysis in
natural resource management. 2007. Leeds, UK, Sustainability Research Institute,
School of Earth and Environment, University of Leeds.

Pyke D. A., J. E. Herrick, P. Shaver, and M. Pellant. 2002. Rangeland health attributes
and indicators for qualitative assessment. Journal of Range Management, 55:584-597.

Riley S. J ., D. J. Decker, L. H. Carpenter, J. F. Organ, W. F. Siemer, G. F. Mattfeld, and
G. Parsons. 2002. The essence of wildlife management. Wildlife Society Bulletin,
30:585-593.

Riley S. J ., W. F. Siemer, D. J. Decker, L. H. Carpenter, J. F. Organ, and L. T. Berchielli.
2003. Adaptive impact management: an integrative approach to wildlife management.
Human Dimensions of Wildlife, 8:81-95.

184

 

Ruhl J. B. 2007. Regulation by adaptive management - is it possible? Minnesota Journal
of Law, Science & Technology, 7 :21-57.

Sarewitz D. and R. A. Pielke. 2007. The neglected heart of science policy: reconciling
supply of and demand for science. Environmental Science and Policy, 10:5-16.

Sayre N. F. 2005. Rangeland degradation and restoration in the "Desert Seasz" social and
economic drivers of ecological change between the sky islands. Pages 349-3 52 in GJ
Gottfried, BS Gebow, LG Eskew, and CB Edminster, editors. Connecting mountain
islands and desert seas: biodiversity and management of the Madrean Archipelago II.
Proc. RMRS-P-3 6. US. Department of Agriculture, Forest Service, Rocky Mountain
Research Station, Fort Collins, CO.

Shindler B. and K. A. Cheek. 1999. Integrating citizens in adaptive management: a
propositional analysis. Ecology and Society, 3:9.

The Nature Conservancy. Conservation by design: a framework for mission success.
2001. Arlington, Virginia, The Nature Conservancy.

USDA Forest Service. The process predicament: how statutory, regulatory, and
administrative factors affect National Forest management. 2002. Washington, DC.

van Wyk E., D. J. Roux, M. Drackner, and S. F. McCool. 2008. The impact of scientific
information on ecosystem management: making sense of the contextual gap between
information providers and decision makers. Environmental Management, 41:779-791.

Vroom V. H. and P. W. Yetton 1973. Leadership and decision-making, Pittsburgh
University Press, Pittsburgh, PA.

Walters C. 2007. Is adaptive management helping to solve fisheries problems? Ambio,
36:304-307.

Walters C. J. and C. S. Holling. 1990. Large-scale management experiments and learning
by doing. Ecology, 71:2060-2068.

West N. E. 1995. Strategies for maintenance and repair of biotic community diversity on
rangelands.in R Szaro, editor. Biodiversity in managed landscapes. Oxford University
Press, New York.

Wik, J ., Calswell, L., Clark, R., DuVarney, A., McElfish, J ., Hogan, A., Solomon, R., and
Sutton, J. Reclaiming NEPA's potential: can collaborative processes improve
environmental decision-making? 2000. Laramie, WY, Institute for Environment and
Natural Resources, University of Wyoming.

Yasmi Y., H. Schanz, and A. Salim. 2006. Manifestation of conflict escalation in natural
resource management. Environmental Science and Policy, 9:538-546.

185