 

 

 

"v
rm; .

thing...

¥.:..,.. ,
.
a.
We vs...»

£15..
..u%

1:! guru”?
h uh.”- ﬁlﬂ*5u : h

.. v.1. 0L;wo|t5rnisﬂ..«
a...“

.r

1:"!!! .
bx ,. ‘V:

1:}. .1. .ivnnnvér

.1........a...|l.1......”n

H‘HFQIMT...”

3.: Intrﬁri. ”ca“:

 

””531:

Z O )3

 

This is to certify that the

dissertation entitled

Measuring and Modeling the Spatial Pattern
of Understory Bamboo Across Landscapes:
Implications for Giant Panda Habitat

presented by

Marc Alan Linderman

has been accepted towards fulﬁllment
of the requirements for

Ph.D. Fish. & Wildl.

degree in

 

 

./ ’

 

Major professor

’J 4.44 a

Date October 25, 2002

 

MS U is an Afﬁrmative Action/Equal Opportunity Institution 0-12771

 

 

LIBRARY
Michigan State
University

 

 

 

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

 

JUN, 0 3 “200

U1
‘4

 

 

 

 

 

 

 

 

 

 

 

 

 

6/01 cJCIRC/DateDuo.p65~p.15

 

MEASURING AND MODELING THE SPATIAL PATTERN
OF UNDERSTORY BAMBOO ACROSS LANDSCAPES:
IMPLICATIONS FOR GIANT PANDA HABITAT

by

Marc Alan Linderman

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

2002

ABSTRACT
MEASURING AND MODELING THE SPATIAL PATTERN
OF UNDERSTORY BAMBOO ACROSS LANDSCAPES:
IMPLICATIONS FOR GIANT PANDA HABITAT
by

Marc Alan Linderman

Understory vegetation is a critical component of biodiversity and an
essential aspect of habitat for countless wildlife species. However, compared to
overstory, information about understory vegetation distribution is scant, available
mainly over small areas or through imprecise large area maps from tedious and
time-consuming ﬁeld surveys. As a result, measuring and monitoring wildlife
habitat at the landscape level rarely include understory vegetation and has likely led
to serious misestimates of wildlife habitat and, consequently, biodiversity. As a
case study, we examined an approach to classifying understory bamboo, the staple
food of the giant panda (A iluropoda melanoleuca), from remote sensing imagery in
the Wolong Nature Reserve, China. We also used these data to estimate the
landscape-scale distribution of giant panda habitat, and model the human effects on
forest cover and the spatio-temporal dynamics of bamboo and the resulting
implications for giant panda habitat.

The spatial distribution of understory bamboo was mapped using an
artiﬁcial neural network and leaf-on remote sensing data. Training on a limited set
of ground truth data and using widely available Landsat TM data as input, a non-
linear artiﬁcial neural network achieved a classiﬁcation accuracy of 80% despite the

presence of co—occurring mid-story and understory vegetation.

Using information on the spatial distribution of bamboo in Wolong, we compared
the results of giant panda habitat analyses with and without bamboo information. Total
amount of habitat decreased by 29 — 56% and overall habitat patch size decreased by 16 —
48% after bamboo information was incorporated into the analyses. The decreases in the
quantity of panda habitat and increases in habitat fragmentation resulted in decreases of
41 — 60% in carrying capacity.

Using a spatio—temporal model of bamboo dynamics and human activities, we found that
local fuelwood collection and household creation will likely reduce secondary habitat
relied upon by pandas. Human impacts would likely contribute up to an additional 16%
loss of habitat. Furthermore, these impacts primarily occur in the habitat relied upon by
giant pandas during past bamboo die-offs. Decreased total area of habitat and increased
fragmentation from human activities will likely make giant pandas increasingly sensitive
to natural disturbances such as cyclical bamboo die-offs.

Our studies suggest that it is necessary to further examine approaches to monitor
understory vegetation and incorporate understory information into wildlife habitat
research and management. The success here to map bamboo has important implications
for giant panda conservation and provides a good foundation for developing methods to
map the spatial distributions of understory plant species. Knowledge of the spatial
distribution of bamboo is necessary to accurately measure the quantity and landscape
characteristics of giant panda habitat. And integrating more speciﬁc habitat information
into models of the combined effects of land cover change and natural disturbances allows

more accurate assessments of habitat.

Copyrighted by
Marc Alan Linderman
2002

To my family

ACKNOWLEDGEMENTS

I would like to extend sincere appreciation to the people and

organizations that made this research possible. Foremost, I would like to offer a
special thanks to my friends and family: My parents who made my education
possible, my grandparents for their unconditional love and support, B00 and Andy
who took it upon themselves to keep me sane, Amber and Anna for the inspiration,
Wayne for the encouragement and support to stay in school, Thomas (aka. The
Hand of God) for providing a ﬁtting nemesis, and everyone else for their support
and tolerance.

I would like to thank my committee, Dr. Jiaguo Qi, Dr. David Skole, and
Dr. Pat Soranno. Thanks to all the members of the lab: Chris Lepczyk (thanks for
all your input), Li An, Scott Bearer, Guangming He, Ed Laurent, and Anita
Morzillo. I really appreciate all the help from Aron Thomas and Aaron
Kortenhoven (thanks for chugging up the hills with me and the intellectual banter).
Thanks to all those for the support and help at Michigan State, Dr. Angela Merti g,
Dr. Kelly Millenbah, Dr. Scott Winterstein, Dr. Ricque Campa, Dr. Ashton
Shortridge, Dr. Tom Coon, and Dr. Bill Taylor. I would also like to thank everyone
for putting up with me during times spent studying and writing. And ﬁnally, Jack, I
sincerely appreciate all your help. You, of course, made all this possible and I very
much thank you for your patience and support.

I would also like to thank the coauthors of each chapter: 1) Jianguo Liu,

Jiaguo Qi, Li An, Zhiyun Ouyang, Jian Yang, and Yingchun Tan. 2) Scott Bearer,

vi

Li An, Yingchun Tan, Zhiyun Ouyang, and Jianguo Liu. 3) Li An, Scott Bearer.
Guangming He, Zhiyun Ouyang, and Jianguo Liu.

And, I would like to thank the Wolong Nature Reserve Administration for
their logistic support of our ﬁeldwork. Particularly, I would like to thank Yang Jian
for his amazing assistance, support, and technical insight. I would also like to thank
Tan Yingchun, Huang Jinyan, Zhou Shiqiang and Zhang Hemin. Ouyang Zhiyun
provided assistance in logistics. And Helen Liu, thank you for showing me around
Beijing.

In addition, I would like to gratefully acknowledge the ﬁnancial support
from the National Aeronautics and Space Administration (NASA Earth Science
Fellowship Award), the National Science Foundation (CAREER Award, formerly
Presidential Young Investigator Award), National Institutes of Health (National
Institute of Child Health and Human Development, R01 HD39789), National
Science Foundation (Biocomplexity Grant), American Association for the
Advancement of Sciences/T he John D. and Catherine T. MacArthur Foundation.
Michigan State University (All-University Research Initiation Grant, Global
Competence Fund, Institute for International Agriculture), the National Natural
Science Foundation of China, the Chinese Academy of Sciences, Ministry of
Science and Technology of China (G2000046807), and China Bridges International

(North America).

vii

TABLE OF CONTENTS

LIST OF TABLES ................................................................................................. x

LIST OF FIGURES .............................................................................................. . xi

CHAPTER ’1 — BACKGROUND AND RESEARCH IMPETUS ....................... . 1
1.1 Background ............................................................................................. 1
1.2 Research Topics ...................................................................................... 8

CHAPTER 2 - USING OPTICAL REMOTE SENSING DATA TO MAP THE
SPATIAL DISTRIBUTION OF UNDERSTORY BAMBOO IN A GIANT
PANDA

RESERVE ............................................................................................................... 13
2.1 15
2.2 Methods ................................................................................................... 18

2.2.1 Study Area ................................................................................. . 19
2.2.2 Data ............................................................................................. 21
2.2.3 Application of Artiﬁcial Neural Networks ................................ . 23
2.2.4 Comparison to Traditional Techniques ....................................... 26
2.3 Results ..................................................................................................... 27
2.3.] Evaluation Results ...................................................................... 27
2.3.2 Supervised Classiﬁcation ............................................................ 29
2.3.3 Full Reserve Results .................................................................. . 30
2.4 Conclusions ............................................................................................. 31

CHAPTER 3 — THE EFFECTS OF UNDERSTORY VEGETATION

ON WILDLIFE HABITAT ACROSS FRAGMENTED LANDSCAPES ........... . 47
3.1 Introduction ........................................................................................... . 49

3.2 Methods ................................................................................................... 51

3.2.1 Study Area ................................................................................. . 52

3.2.2 Comparison Study ....................................................................... 53

3.2.3 Bamboo Classiﬁcation ................................................................ 55

3.2.4 Incorporating Bamboo Information into Habitat Analysis ......... 56

3.2.5 Habitat Measures ....................................................................... . 59

3.3 Results ..................................................................................................... 60

3.4 Conclusions and Discussion .................................................................. . 63

viii

CHAPTER 4 — IMPACTS OF NATURAL AND HUMAN DISTURBANCES

ON FORESTS AND GIANT PANDA HABITAT ............................................... . 77
4.1 Introduction ............................................................................................. 79
4.2 Methods ................................................................................................... 84
4.2.1 Study Area .............................................................................. . 84
4.2.2 Data and Parameterization ....................................................... 87
4.2.3 Model Description .................................................................. . 92
4.2.4 Model Validation and Sensitivity Analyses .............................. 98
4.2.5 Household Impacts ................................................................. 101
4.2.6 Integrated effects of household activities and bamboo
dynamics .................................................................................. 103
4.3 Results ................................................................................................... 104
4.3.1 Model Validation and Sensitivity ........................................... 104
4.3.2 Household Impacts ................................................................. 108
4.3.3 Combined effects of household activities and bamboo
dynamics ................................................................................ 109
4.4 Conclusions and Discussion .................................................................. 1 1 1
CHAPTER 5 — SYNTHESIS AND CONCLUSIONS .......................................... 140
5.1 Summary ................................................................................................ 140
5.2 Conclusions ............................................................................................ 145
BIBLIOGRAPHY ................................................................................................. l 48

ix

Table 2.1
Table 2.2
Table 2.3
Table 3.1

Table 3.2
Table 3.3
Table 3.4
Table 4.1

Table 4.2
Table 4.3

Table 4.4

Table 4.5

Table 4.6

Table 4.7

Table 4.8

Table 4.9

LIST OF TABLES

Relationship between overstory species and abiotic factors and
understory bamboo .................................................................................... 37

Inﬂuences of early stopping and learning algorithms on validation and
veriﬁcation accuracy (percent) .................................................................. 38

Confusion matrix showing ground truthed values compared to
predicted presence/absence from full-reserve analysis ............................. 39

Inﬂuences of bamboo ﬁlter methods and ﬁlter sizes on habitat

quantity ...................................................................................................... 68
Inﬂuences of ﬁlters on patch number and size .......................................... 69
Inﬂuences of ﬁlter methods and ﬁlter size on total habitat ....................... 70
Inﬂuence of ﬁlter method on core habitat and carrying capacity ............. 71

Landscape characteristics of the measured households in 1997
compared to model scenarios .................................................................. l 18

Accuracy of the predicted household locations for the model scenarios
relative to the measured households in 1997 .......................................... l 19

Comparison of the inﬂuence of variations in individual parameters for
each multiplicative factor used within the fuelwood sub-model ............ 120

The inﬂuence of household characteristics on habitat over 65 years
(1965 — 2030) relative to a baseline scenario of 0 new households per
year and 0 m3/yr of fuelwood consumed after 1997 ............................... 121

The inﬂuences of household characteristics on habitat for models run
from 1997 to 2030 based on digital classifications and measured
household locations ................................................................................. 122

The inﬂuences of household characteristics on habitat below 2600
meters of elevation for model runs from 1997 to 2030 based on digital
classiﬁcations and measured household locations .................................. 123

The combined inﬂuences of bamboo die-offs and household impacts... 124

The combined inﬂuences of bamboo die-offs and household impacts
on habitat below 2600 meters of elevation .............................................. 125

The combined inﬂuences of bamboo die-offs and household impacts
on habitat fragmentation .......................................................................... 126

Figure 2.1

Figure 2.2

Figure 2.3

Figure 2.4

Figure 2.5

Figure 3.1

Figure 3.2

Figure 3.3

Figure 3.4

Figure 4.1

Figure 4.1

Figure 4.2

Figure 4.3
Figure 4.4

Figure 4.5

Figure 4.6

LIST OF FIGURES

Examples of past bamboo distribution maps for Wolong Nature
Reserve .................................................................................................. 42

Examples of Bashania fangiana beneath a typical overstory and
F argesia Robusta with co-occurring vegetation .................................... 43

Representative schematic of an artiﬁcial neural network ...................... 44

Effects of algorithm structure with variations in learning algorithm
and error goal ......................................................................................... 45

Full-reserve output maps from optimal algorithm input and structure. 46

Classiﬁcations of bamboo suitability based on proportion and
majority ﬁlters ........................................................................................ 69

The effects of ﬁlter methods and ﬁlter sizes on highly suitable habitat
class ........................................................................................................ 73

Habitat distribution throughout Wolong based on classiﬁcations
without bamboo, with bamboo (proportion ﬁlter), and with bamboo
(majority ﬁlter) ....................................................................................... 74

Maps of core habitat patches based on classiﬁcations without
bamboo, with bamboo (proportion ﬁlter), and with bamboo

(majority ﬁlter) ....................................................................................... 75
A conceptual depiction of bamboo and land-use effects on panda
habitat... ........................................................................................... . 130
Wolong Nature Reserve lies in the Qionglai Mountains between the
Tibetan plateau and Sichuan basin. ..................................................... 131
A conceptual ﬂow schematic diagram of the model ............................ 132
Examples of abiotic and biotic data used for the model ...................... 133
Examples of spatial data used for post-1997 simulations .................... 134

Comparisons between visual classiﬁcations of satellite data from
1965, 1974, 1987, and 1997 and predicted forest cover due to
household activities of corresponding years ........................................ 135

Comparisons of the inﬂuence of the three multiplicative factors
contained within the household sub-model .......................................... 136

xi

Figure 4.7

Figure 4.8

Figure 4.9

Predicted forest cover due to fuelwood collection compared to the
digital classiﬁcation at various starting volumes for low-elevation
forests ................................................................................................... 137

Impacts from household activities and bamboo dynamics ................... 138

The impacts of household activities and bamboo die-offs to habitat
below 2600 meters. ............................................................................. 139

xii

CHAPTER 1

BACKGROUND AND RESEARCH TOPICS

1.1 Background

There is a long history of studying wildlife/habitat relationships (Leopold
1933; Morrison et a1. 1998). While some economically and socially relevant
species have been studied in extreme detail, surprisingly little is known of the exact
habitat requirements for many species, how their habitat is distributed, how
continued human activities will inﬂuence the current distribution of habitat, and
how the distribution of habitat inﬂuences the local and global distribution of
species. This is particularly problematic at scales larger than the typical plot or
patch scale at which most of the past ecological studies have operated. Broad-scale
analyses such as landscape and regional studies have almost always necessarily
relied on habitat information available from remote sensing data (Roughgarden et
al. 1991). If useful theories are to be developed for a better understanding of
wildlife/habitat interactions and to offer viable management options, more detailed
information is needed over larger areas. In addition, these data need to be examined
relative to species needs. And, the processes and degree to which humans are
impacting wildlife habitat need to be better understood.

Habitat loss is a principal threat to the local and global persistence of

terrestrial fauna] and ﬂoral species (Sala et a1. 2000). There has been much

discussion on the degree of humans have impacted natural areas around the world.
Estimates have ranged from 40% of the terrestrial surface having been directly co-
opted to 80% directly and indirectly inﬂuenced (Vitousek et a1. 1997). This large
degree of inﬂuence on natural areas, and consequently wildlife habitat, necessarily
effects the distribution and persistence of almost every species on this planet.
Knowing where these impacts are occurring, how they relate to wildlife habitat, and
understanding the process related to impacts are needed to mitigate further losses
and develop a more sustainable human-environment relationship.

Landscape ecology ﬁrst emphasized the importance of the spatial
arrangement of resources relative to wildlife species and the need of examining
habitat at scales relevant to how species interact with the environment (Forman and
Godron 1986; Turner 1989). This was one of the ﬁrst efforts to extend wildlife-
habitat relationships to include the inﬂuence of the surrounding environment on
habitat quality, movement, and population persistence. Expanding ecological
analyses from the traditional patch relationships broadened ecological theory, but
was very data, and potentially, computationally intensive. Three advancements
allowed landscape ecology to include analysis of real heterogeneity found in nature:
accessible computational power, remote sensing, and Geographical Information
Systems (GIS) (Roughgarden et a1. 1991). Increased computational availability
allowed increased modeling and quantitative analyses of large amounts of data,
remote sensing provided data corresponding to the arrangement of natural features,
and GIS provided a system to relate the separate data sources to one another for

more intensive and realistic analyses.

Modeling wildlife~habitat relationships is a mainstay of ecology (Turner
1990). Models have ranged from simple linear regressions to complex nonlinear
systems analyses. However, intensive modeling has only become increasingly
popular with the development of personal computers. Particularly, spatially explicit
models have become more commonplace with the availability of computational
power and geographically registered data (Dunning et a1. 1995). While
geographical analyses of socio-economic relationships with the environment have a
long history, it has only been the recent availability of detailed geo-referenced data
that has allowed coupling theoretical developments with observed distributions of
natural resources. Remote sensing data and hand-held Global Positioning System
(GPS) devices have led this boom in geo-referenced data. However, many
environmental variables found necessary to relate wildlife to habitat at the patch
level have not been available from remote sensing sources.

One example of this is past attempts to map giant panda (Ailuropoda
melanoleuca) habitat. Past analyses have been restricted in their ability to
accurately measure panda habitat over large areas by the lack of information on the
spatial distribution of bamboo, the principal food source for pandas (Schaller et a1.
1985; Morain 1986). Throughout most of the panda range, bamboo is found in the
understory. Efforts to map the distribution of bamboo through ﬁeld surveys have
necessitated coarse estimates over large areas or detailed estimates restricted to
small areas. Attempts to map bamboo using aerial photography or remote sensing
have been unsuccessful in that conventional photo interpretation and digital image

analyses have difﬁculty categorizing understory vegetation (Morain 1986). This

lack of understory information, then, is likely leading to overestimates of panda
habitat. In addition, as bamboo is not correlated to forest cover or abiotic factors
(Linderman, unpublished data), the actual landscape distribution of panda habitat is
likely much different than if habitat is only calculated using traditional approaches.

Giant pandas historically extended across much of current China (Schaller et
a1. 1985). Fossils dating back to the Pleistocene era (approximately 3 million years
ago) have been found across China and in neighboring Burma. In the mid-1980’s.
MacKinnon and DeWulf (1985) estimated that the total available habitat has been
reduced to approximately 32,000 kmz. Drastic human population growth and
subsequent development and deforestation have led to the loss of much of the
panda’s former range. It is this extensive loss of habitat and isolation that has
resulted in the fragile nature of the current panda populations.

While rugged topography and isolation from humans protected the current
available panda habitat and populations, continued growth in human population and
development pose increasing risks to the remaining panda habitat. Currently,
pandas occupy only a portion of the remaining habitat, found only at the fringe of
their former range in Sichuan and Shaanxi provinces. Within the Min, Qionglai and
Min mountains, it is estimated that less than 1.000 pandas remain in the wild,
existing in isolated populations separated by topography and human development
(Zhang et a1. 1997). The isolation of distinct populations severely threatens the
genetic stability of the species (Lu et a1. 2001). However, other threats exist in the

form of incidental poaching, decreasing accessibility to quality habitat, and

continued development and human impacts of the remaining habitat (Schaller et al.
1985).

For example, between 1974 and 1989 it is estimated that the panda
population in Wolong Nature Reserve (one of the largest reserves dedicated to
protecting giant panda habitat) plummeted from 140 to approximately 74
individuals (Zhang et al., 1997). Schaller et al. (1985) suggest the loss of habitat
and incidental poaching most likely contributed to the attrition of the reserve’s
population. A more publicized explanation of the decrease of pandas in the reserve
was a mass-ﬂowering of bamboo that occurred in 1982 (Schaller et al. 1985).
While pandas are obligate grazers of bamboo and the species (Bashaniafangiana)
which mass-ﬂowered was the primary bamboo species for pandas in the reserve, it
is unlikely the loss of this species contributed signiﬁcantly to the loss of any
pandas. Schaller et a1. (1985) noted that no pandas died of starvation and very few
pandas suffered malnutrition. This is in stark contrast to a similar situation in the
Min mountains where in 1982 a sudden die-off of several bamboo species resulted
in drastic number of pandas dying of starvation. The die-off of the principal
bamboo species in Wolong Nature Reserve did not result in similar impacts on
pandas because a contiguous secondary bamboo species (F argesia robusta) was
accessible. However, continued human impacts threaten habitat throughout
Wolong with the most likely areas of development and use located in areas of
secondary bamboo. The loss of this habitat could result in drastic consequences for

the panda population in Wolong and poses a problem throughout much of the panda

range. The current fragmentation of habitat and inaccessibility to alternative
bamboo sources seriously threaten the remaining panda populations.

Bamboo is fairly unique in that it is a semelparous grass with a gregarious
ﬂowering and subsequent mass die-off (J anzen 1979). Few environmental
associates or triggers have been found and few conclusive hypotheses on the cause
for ﬂowering are available (Keeley and Bond 1999). The time between ﬂowering
(or intermast time) varies considerably depending on the species. Interrnast times
for bamboo species range from 10 years to over 100 years, but seem to show
extremely consistent intermast times within species (Janzen 1979). Taylor et al.
(1991) found that environmental variables contributing to stress of the bamboo
limits areas of ﬂowering. For example, large tracts of Bashania fangiana (common
name: Arrow bamboo) ﬂowered in 1983 in Wolong Nature Reserve. Patches of
Bashania at the extreme upper elevation range, on steep north-facing slopes, or in
large canopy gaps (and consequently different local moisture conditions) were less
likely to ﬂower than patches not stressed. Reestablishment after mass-ﬂowering is
difﬁcult to predict based on environmental conditions. However, slight increases in
recruitment were noted in small canopy gaps (Taylor and Qin 1993). Growth in
canopied areas was slightly lower than in gaps with an average of 15 years for
Bashania to return to full cover in most areas (Taylor and Qin 1993). However,
even after 15 years, visible signs of the ﬂowering and large tracts devoid of
Bashania can still be found throughout much of the range of Bashania (Linderman,

personal observation).

Bamboo’s unique life-history poses a signiﬁcant threat to the current
distribution of giant pandas. Where pandas previously were able to move to areas
of alternative bamboo species, current habitat fragmentation limits movement and
restricts panda’s access to areas containing other bamboo species. The dependence
of pandas on secondary bamboo species was apparent in Wolong during the recent
Bashania die-off and the consequences due to the lack of secondary bamboo during
the simultaneous die-off of several bamboo species in the Min Mountains. This
problem has become increasingly acute as humans appropriate increasing amounts
of habitat, but more relevant is that habitat loss is occurring in areas that may have a
considerably higher impact on panda populations than total quantity of habitat loss
predicts alone. If pandas are excluded from an alternative bamboo resource during
the subsequent ﬂowering of Bashania, the current panda population could be
seriously impacted. Furthermore, this points to wider problems and needs for
effective management of panda habitat and mitigation of human impacts. Current
knowledge on the spatial distribution of bamboo is inadequate relative to panda
needs, the spatial distribution of bamboo has a signiﬁcant impact on panda habitat,
and human activities are increasingly occurring in areas vital to panda survival.

This situation is not conﬁned to the giant pandas of China. Countless fauna,
as well as ﬂora, depend on environmental conditions that exist below canopies.
Therefore, past attempts reliant upon remotely sensed data are likely misestimating
species distributions and landscape arrangements of habitat. Efforts to derive
species-relevant data from remotely sensed images at a scale that species are

interacting with the environment are needed to more accurately reﬂect the spatial

distribution of habitat, allow more robust analyses of wildlife/habitat relationships
over large areas, and permit the development of more viable policy alternatives.
This dissertation was developed with the principal goals of deriving
pertinent giant panda habitat information from remotely sensed data, studying the
impact of detailed bamboo information on the distribution of habitat, and projecting
future impacts from human activities on giant panda habitat. It was hoped this
study would allow a more accurate representation of panda habitat over large areas,
increase our understanding of the current habitat conditions, and provide additional
information to management ofﬁcials to insure the protection of this endangered
species. To this end, I developed three primary areas of analysis: (1) I developed
methods to map the spatial distribution of bamboo from remote sensing data, (2)
analyzed the inﬂuence of the spatial distribution of bamboo on the landscape
distribution of giant panda habitat at the landscape scale, and (3) modeled the
spatial trends in human impacts relative to the spatial distribution of bamboo. A
review of the principal topics of this dissertation and the three research chapters

follows.

1.2 Research Topics

The ﬁrst goal was to develop a viable method to classify the
presence/absence of bamboo from remote sensing. The impetus for this was the
lack of detailed information on the spatial distribution of bamboo. Past efforts to

map the distribution of bamboo have relied on ground surveys. Due to the cost and

effort required, this approach resulted in either coarse maps over large areas (e. g.
1000’s of square kilometers) or detailed maps constrained to small areas (e. g. 10’s
of square kilometers). The detail of the coarse maps, while sufﬁcient for broad
estimates of giant panda habitat, were not nearly at the resolution at which pandas
interact with bamboo nor sufﬁcient to predict habitat use and resulting inﬂuence of
bamboo on the spatial distribution of habitat. To achieve the detail as in the small
area maps would be very costly and time consuming endeavors. Therefore, I
examined the utility of remote sensing data to map the distribution of bamboo. A
signiﬁcant obstacle to the successful classiﬁcations of bamboo from remote sensing
in the past is that throughout most of panda habitat bamboo is found as an
understory species (Morain 1986; Porwall and Roy 1991). Overstory cover reduces
the amount of irradiance at the understory, modiﬁes irradiant downwelling light,
intercepts reﬂected light from the understory, and modiﬁes upwelling reﬂected light
from the understory (Lillesand and Kiefer 1994). Therefore, the radiance measured
at a distant sensor is a complex integration of reﬂected energy with the main
contribution coming from the overstory features.

Irradiant energy obviously does reach the understory. Direct light from
canopy gaps, reﬂected light from overstory features, and transmitted light
(particularly near and shortwave infrared energy) interacts with the understory. The
ability of remote sensing then would be a matter of sufﬁcient signal from the
understory reaching the remote sensor relative to noise (or a sufﬁcient signal-to-
noise ratio). To examine this complex problem, we compared the utility of an

artiﬁcial neural network to parse the signal and classify the presence or absence of

understory bamboo. Using Landsat TM data as input and hundreds of ground
samples as training and validation data, we were able to achieve over 80% accuracy
in classifying the presence/absence of understory bamboo at a resolution of 30
meters for the entire Wolong Nature Reserve.

A subsequent study used the bamboo presence/absence data to examine the
effects of bamboo distribution on panda habitat throughout the reserve. Traditional
landscape studies have relied on remotely sensed data to map habitat. However,
information on the vertical and horizontal distribution of vegetation beneath
canopies is needed to accurately relate wildlife use and environmental conditions
(MacArthur and MacArthur 1961; James 1971). Theoretical studies and practical
application of landscape ecology then require more detailed environmental
information to more accurately predict the inﬂuence of the spatial arrangement of
habitat on wildlife species. A comparison was made between a reserve-wide
analysis on the distribution of habitat that did not use bamboo information (Liu et
al. 2001) and reserve-wide analyses including the spatial distribution of bamboo.
This has signiﬁcant implications for future panda habitat monitoring and
management. In addition, the comparison suggests data requirements for other
landscape-level analyses.

Finally, it has been long recognized that wildlife species require access to
speciﬁc resources (e. g. watering holes, forage, trees of sufﬁcient diameter, etc.) In
more general terms, landscape ecology is based on the principal that the

arrangement of resources inﬂuences ecosystem processes whether they are ﬁre

disturbances, wildlife movement, or human impacts. In particular, it has been noted

10

that certain resources can be vital for wildlife species. The spotted owl controversy
centered on the owls’ requirement of large areas of old-growth forest. Low-
elevation graze areas are critical to elk and large mammals in Yellowstone (Turner
et al. 1994). And, the availability of alternative bamboo sources is vital to the
survival of giant pandas. Even limited impacts to habitat, then, can have drastic
implications for a wildlife species. This is occurring in Wolong Nature Reserve.
Local farmers, while only harvesting a small proportion of the total forest area. are
concentrating their resource use (e. g., fuelwood collection) on easily accessible
forests at lower altitudes corresponding to the current distribution of the secondary
bamboo utilized by pandas. Continued impacts in these areas could have drastic
consequences during the next die-off of the principal bamboo species in the reserve.
In an attempt to understand the driving factors for the spatial distribution of
fuelwood collection, predict impacts under different scenarios in the future, and
offer possible mitigation strategies, we have modeled human activities in Wolong
Nature Reserve. The study suggested that the spatial arrangement of human
activities is as important as the sheer quantity of activity when considering possible
impacts on native species.

Wolong Nature Reserve is unique in that it is one of only a handful of
reserves dedicated to the protection of an extremely high proﬁle species, the giant
panda. However, the ecological problems occurring in Wolong are not unique at
all. Humans will continue to degrade habitat, more detailed information is needed
on speciﬁc habitat requirements such as understory vegetation, and more accurate

measurements of actual habitat area and spatial distribution are urgently needed for

ll

species around the world. The methods outlined in this dissertation hopefully
represent possible alternative approaches to developing these data, examining

ecological questions, and developing viable alternatives.

CHAPTER 2

USING OPTICAL REMOTE SENSING DATA
TO CLASSIFY TI-IE PRESENCE/ABSENCE OF UNDERSTORY
BAMBOO IN A GIANT PANDA RESERVE

In collaboration with

Jianguo Liu, Jiaguo Qi, Li An, Zhiyun Ouyang,
Jian Yang, and Yingchun Tan

13

Abstract

Understory vegetation is a critical component of biodiversity and an
essential habitat component for many wildlife species. However, compared to
overstory, information about understory vegetation distribution is scant, available
mainly over small areas or through imprecise large area maps from tedious and
time-consuming ﬁeld surveys. A practical approach to classifying understory
vegetation from remote sensing data is needed for more accurate habitat analyses
and biodiversity estimates. As a case study, we mapped the spatial distribution of
understory bamboo in Wolong Nature Reserve (southwestern China) using leaf-on
remote sensing data. Training on a limited set of ground truth data and using widely
available Landsat TM data as input, a non-linear artiﬁcial neural network achieved
a classiﬁcation accuracy of 80% despite the presence of co-occurring mid-story and
understory vegetation. These results suggest that understory vegetation inﬂuences
remote sensing data and classifying understory vegetation is feasible. The success
here to map bamboo has important implications for giant panda conservation and
provides a good foundation for developing methods to map the spatial distributions

of understory plant species.

14

2.1. Introduction

Understory vegetation is a signiﬁcant component of biological diversity and
critical habitat for countless wildlife species (MacArthur and MacArthur 1961,
Odum 1971, Schaller et al. 1985). However, monitoring of understory conditions
has been restricted to tedious and time-consuming ground surveys due to a lack of
alternative methods such as remote sensing. More speciﬁcally, whereas signiﬁcant
advances in regional ecology have been made from overstory mapping
(Roughgarden et a1. 1991), extensive spatial distribution information of understory
vegetation has remained unavailable due to the limitations of traditional remote
sensing classiﬁcation techniques. Thus, the development of a practical approach to
classify understory vegetation irrespective of canopy cover would signiﬁcantly
increase knowledge of the spatial distribution of understory vegetation and assist in
modeling wildlife habitat for species-speciﬁc studies and biodiversity estimates.

Classiﬁcation of vegetation from remote sensing data has typically relied on
multispectral optical data as variations in chemical, water and structural
composition between vegetation types all inﬂuence incident optical radiation
allowing differentiation between classes of vegetation (Asrar et a1. 1989). In sparse
forests or regions with signiﬁcant canopy gaps, direct reﬂectance from the
understory certainly inﬂuences the optical reﬂectance response as measured by a
distant sensor. In fact, in the near-infrared region, understory can dominate the
overall reﬂectance from open-canopy stands (Nemani et al. 1993). Furthermore,

several studies have shown that canopy background features have a signiﬁcant

15

inﬂuence on optical radiance measured by remote sensors even under considerable
cover (Huete et a1. 1985, Ranson et al. 1986, Guyot and Riom 1989, Bausch 1993).
In particular, studies on the ability to retrieve Leaf Area Index (LAI) values from
optical data have shown that, even in canopies up to 89% closed, understory
vegetation characteristics have a measurable effect on the radiance response
recorded by a satellite sensor (Nemani et al. 1993, Law and Waring 1994, Spanner
et a1. 1994, Qi et a1. 2000). The inﬂuence from background features is due not only
to direct reﬂectance from gaps in the canopy, but also multiscattering between
understory and the overstory as well as transmission through the canopy and is
therefore a combination of linear and non-linear contributions (Borel and Gerstl
1994). The degrees of linear and non-linear contributions vary relative to the
understory and overstory cover, structure, and other variables. Therefore, while
understory vegetation and background features inﬂuence the response measured at
distant sensors, the understory contributions can often result in a complex signal
that has not been amendable to traditional classiﬁcation techniques.

Attempts to use traditional methods to map understory vegetation, even a
coniferous understory with a leaf-off deciduous overstory, have had, at best, mixed
results (Stenback and Congalton 1990, Porwall and Roy 1991 , Ghitter et al. 1995).
Traditional approaches to classifying understory vegetation land cover from
multispectral data have been based on modeled categorizations of a discrete
thematic space using the statistical separability of pixel signatures (Jensen 1996).
These approaches are often restricted in their applications and accuracy for

classifying complex scenes. Topography, mixed classes and spectral similarities

l6

have all been shown to inﬂuence traditional classification accuracy (Lillesand and
Kiefer 1994). This becomes particularly relevant when considering the complex
contributions of understory vegetation. In addition, changes in canopy cover and
species composition, and understory species composition and cover from pixel to
pixel results in an extremely variable spectral signature, often with varying degrees
of non-linear contributions, that is difﬁcult to classify using standard methods.

In order to utilize remote sensing data for classifying understory vegetation,
a method to parse the highly variable and complex understory inﬂuences (canopy
gap and sub-canopy) from the integrated radiance received at the sensor then must
be used. Artiﬁcial neural networks are essentially non-parametric data
transformations that can account for non-linear effects given a sufﬁciently complex
partitioning of the classiﬁcation space (Atkinson and Tatnall 1997). To classify the
presence/absence of the variable spectral inﬂuence of understory vegetation, a non-
linear approach is almost certainly required as well. Understory contributions to the
radiance at the sensor are a mix of decreasing linear contributions from canopy gaps
as the overstory becomes increasingly dominant and increasingly non-linear
components due to multi-scattering and transmission. Furthermore, whereas
traditional classiﬁcation techniques make assumptions on the underlying model of
the data, neural networks are a non-parametric method and are capable of learning
the complex trends in data necessary to parse the variable and subtle trends in areas
of more canopy (Foody and Arora 1997, Atkinson and Tatnall 1997). Neural
networks should therefore be able to more accurately classify sub-canopy gap

vegetation while considering non-linear components of understory contributions. In

17

addition, this technique should be more capable of adapting to the variability in the
signature due to varying canopy and understory conditions. Therefore, to test the
practicality of classifying understory vegetation, we compared the ability of an
artiﬁcial neural network to predict the presence or absence understory vegetation
from optical remote sensing data based on limited ground-truth data relative to

traditional techniques.

2.2. Methods

The impetus for this study was the need for a practical method to classify
the spatial distribution of understory bamboo irrespective of the canopy conditions
over large regions. Bamboo plays a vital role in the survival of the endangered giant
panda (Ailuropoda melanoleuca) (Schaller et a1. 1985, Liu et al. 1999a). The impact
of the spatial distribution of bamboo on panda populations has been well-
documented (Johnson et a1. 1988, Taylor and Qin 1997). However, past panda
habitat analyses have been limited in their ability to conduct accurate habitat
assessments over large areas due to the lack of bamboo distribution maps with
sufﬁcient detail or extent as shown in Figure 2.1. Remote sensing would be a
preferable method for data acquisition at larger scales. However, methods to map
the extent of bamboo, even employing aerial photography, have not been successful
(Morain 1986, De Wulf et a1. 1988, Porwall and Roy 1991). The main problem in
classifying bamboo in much of the panda range is that it is typically found as an
understory species under variable canopy species, percent cover and densities. We

required an approach that would be able to classify the presence/absence of

18

presence/absence of understory bamboo irrespective of the overstory. Therefore, to
test the applicability of an artiﬁcial neural network to this problem, classifications
from a neural network and supervised classifications of the bamboo distribution

using limited training data were compared.

2.2.1 Study Area

Located between 102°52’ and 103°24’ E and 30°45’ and 31°25’ N in the
Qionglai Mountains of southwestern China, the reserve is approximately 200,000
ha in size and one of the largest parks dedicated to giant panda preservation (Liu et
al. 1999b, Liu et a1. 2001). It is estimated that over 10% of the remaining
worldwide panda population can be found within this reserve. The topographic
nature of Wolong directly relates to habitat preferred by panda. Within Wolong,
elevations range from 1200m to 6525m creating several climatic zones and
consequently high habitat diversity. This also provides cooler climatic conditions
that panda prefer coupled with precipitation levels necessary for bamboo of which
panda are obligate grazers. From the lowest elevation to approximately 1600
meters, the canopy consists mainly of evergreen broadleaf. From 1600 to 2000 m
there is an increasing mixture of the broadleaf and deciduous. The canopy is
dominated by deciduous vegetation from 2000 m to about 3600 m with an
increasing mix from conifer at the higher altitudes. Above this elevation, subalpine

conifers give way to alpine meadow at an elevation of approximately 4400 m.

19

Throughout the reserve forest canopies averaged 56% closure and rarely exceed
90%, maintaining signiﬁcant gap area for light penetration.

Within the reserve, bamboo occurs up to an elevation of 4500 m and is
typically an understory species. Up to eleven species of bamboo are found in
Wolong. Two species, Bashania fangiana and F argesia Robusta (Figure 2.2),
predominate. In forested areas, bamboo is distributed in patches ranging in size
from single plants to hundreds of meters across. Where bamboo does occur, it
averaged 41% of the ground cover. However, the spatial distribution of bamboo
seemingly does not follow any trends relative to overstory or abiotic factors (Reid et
al. 1989). Regression relationships based on ground samples of overstory cover,
slope, altitude and bamboo ground cover in forested areas are shown in table 2.1
and suggest that the prediction of bamboo presence/absence is independent of
within stand characteristics. The lack of correlation of the spatial distribution of
bamboo and environmental factors may be partly due to the unique episodic
synchronized die-offs of large areas of bamboo (Reid et a1. 1989, Keeley and Bond
1999). However, it is precisely this distribution that affects pandas.

The spatial distribution of bamboo is also inﬂuenced by human activities
and restricted environmentally (e. g. hi gh-altitude permanent rock, alpine meadow).
Grazing and agricultural use have effectively removed bamboo from some areas.
Other regions have been clear-cut leaving a mixed midstory shrub layer and a lower
occurrence of bamboo. Finally, selective logging has changed the species
composition of the overstory and reduced canopy cover in some areas. Few forested

areas contained large gap areas. Selective logging has been characterized by

20

random thinning. Clear-cuts and other land cover changes have typically been
larger than 60 x 60 m. While relatively limited in impact, these human activities are
transforming the landscape and introducing additional complexity in classifying the
land cover. The complex relationships between land cover and the lack of
substantial correlation between the presence of bamboo and canopy conditions
made prediction or classiﬁcation using traditional methods difﬁcult and required a
new approach for accurate classiﬁcation. Knowledge of the distribution of bamboo
irrespective of land cover is needed to allow better classiﬁcations of the suitability
of panda habitat, test the implications of the spatial distribution of habitat on panda
populations, and provide recommendations for future restoration and conservation

management.

2.2.2 Data

Field data were collected throughout the reserve during the summers of
1998 and 1999 and included vegetation ground sampling plots for algorithm
training and validation and Ground Control Points (GCPs) to allow registration of
the remote sensing data. Landsat Thematic Mapper (TM) data acquired over
Wolong Nature Reserve in September 1997 were used for this study. The Landsat
TM scene was registered to UTM WGS-84 coordinates to allow co-registration of
the ground truth data. The remote sensing data were registered using the GCPs
collected to an RMSE of less than one pixel. To gather representative ground data,

stratiﬁed sampling of the land covers and understory conditions was conducted.

21

Ground sample plots were located where access was possible and registered to the
remote sensing data through differential Global Positioning System (GPS) using
Trimble Pro XRS and Community Base Station receivers. Plots were selected
where the vegetation was relatively homogeneous over a 60m x 60m area. A
sample area was considered relatively homogeneous where similar percentages of
vegetation were distributed evenly throughout the entire sample plot. Therefore,
any 30m x 30m subplot would contain the same vegetation percentages regardless
where it was situated within the 60m x 60m plot. The 60m x 60m dimensions were
chosen as prior knowledge of the GPS position in relation to the remote sensing
grid was unknown and ensured one pixel of the Landsat TM data (30—meter
resolution) would be fully contained within a sample plot. For each plot,
information on the biota such as vegetation types for the overstory, midstory and
understory, as well as the corresponding percent cover, were recorded. Percentages
of vegetation cover in the overstory, > 5m, midstory, 2 — 5m. and understory, < 2m
were estimated visually for the 60m plots.

Training data for the artificial neural network were selected from the
vegetation ground truth data where positions were known at a suitable accuracy
(ﬁeld data were ﬁltered for GPS measurements with standard deviations greater
than 10 meters) and stratiﬁed to include a representative sample of vegetation
conditions. On this basis 189 sample plots were chosen. Approximately, two-thirds
of these data were used to train the neural network and the remaining third reserved
for validation. The data were categorized into presence/absence categories. If

bamboo cover was greater than 10%, the training value was assigned a 1 (presence),

22

else if the bamboo cover was less than or equal to 10% within the sample area, the
training value was assigned a 0 (absence). Such categorization was done for three
reasons. First, sampling methods were limited in assigning absolute ground cover
percentages over large sample plots. Therefore, data were binned into 0, 25, 50, 75
and 100% categories. For example, a sample of between 0% and 10% was
categorized as 0%. If a sample fell between 15% and 35% it was categorized as
25%, etc. Therefore, samples estimated to have 25, 50, 75 or 100% bamboo ground
cover were considered presence. Second, at less than 10%, the cover was extremely
insigniﬁcant, did not provide any useable biomass for pandas, and was considered
to have a very limited inﬂuence on the spectral response. And third, it was
anticipated that a binary categorization would reduce data transformation

complexity.

2.2.3 Application of Artiﬁcial Neural Networks

The multi-layer perceptron (MLP) is the most commonly used neural
network structure in remote sensing and was used in this study (Atkinson and
Tatnall 1997). The basic structure of an MLP is the existence of distinct data layers
through which the data are transformed (Figure 2.3). For example, the ﬁrst layer
(input layer) could be represented by a series of nodes consisting of the separate
bands of remote sensing data, ancillary data or any combinations of these data. The
next layer(s) are referred to as hidden layer(s) as the user does not have access to

the values within the nodes of these layers. Each hidden node is a weighted sum of

23

a function of the input vector. The ﬁnal output layer is the weighted-sum of a
function of all hidden nodes. The learning process of the neural network occurs by
back-propagation and error minimization according to the training data.

The training data are composed of input data and corresponding known
output values. The input data for this study consisted of the single TM pixels fully
contained within the sixty-meter ground sample plots. The output data consisted of
the corresponding ground sampled bamboo absence/presence values of those plots.
Therefore, the output layer consisted of a single node with the expected values of 0
(absence of bamboo) or 1 (presence or bamboo). Typically when training neural
networks, the hidden functions are initially set with a random weight value. The
input values are then passed through the hidden layers for the ﬁrst time and the
output from the hidden layers (the output layer) is compared to the expected
outcomes according to the ground sample training data. Learning algorithms then
modify the hidden layer weights to reduce the error between the training values and
algorithm outputs. This process is repeated as the individual hidden node weights
and biases are modiﬁed in an iterative process until acceptable error levels
compared to the training data are achieved. This repeated error minimization
essentially allows the algorithm to learn the nature of the training data (Atkinson
and Tatnall 1997). If the training data are perfectly representative, the algorithm
theoretically learns the nature of the full data set. However, it is often very difﬁcult
to collect a training set that is completely representative of the whole data set.
Consequently, image data not seen by the algorithm or consistent with the training

data can result in variations from the expected output values.

24

For this study, the neural networks were simulated in the Neural Network
module of Mathworks MATLAB (The MathWorks Inc. 1999). Several variations
of internal network structure, input data, and learning algorithms were tested to
determine optimal algorithm characteristics. Different combinations of the Landsat
TM (6 bands excluding thermal band) were examined as potential data input. As a
result, the input layer consisted of three to six input nodes depending on the layers
used. The structure of the hidden layers was also tested to determine the necessary
number of hidden layers and number of nodes per layer required. In addition,
generalization techniques, methods to reduce overﬁtting, were analyzed including
techniques such as early stopping by adjusting the training Mean Square Error
(MSE) goal and automated regularization utilizing the Bayesian Regularization
(BR) learning algorithm. Early stopping, as the term suggests, establishes a higher
error convergence threshold and thus stops the training process before full
convergence occurs. This can be done automatically by comparing an additional
reserved data set to determine when error minimization reduces generalization of
the full data set. However, since training data were limited, preset early stopping
error levels of 1x10'05 and 1x10’08 were tested relative to a full convergence level,
typically around lxlO'20 (known from preliminary trials).

Evaluation of individual algorithms was conducted by examining algorithm
performance through each of four stages: training, simulation, veriﬁcation and
validation. The algorithms were ﬁrst presented with training data (i.e. remote
sensing pixel values and corresponding ground-truth data presence/absence values).

If the algorithm was able to converge on the preset error goal, a 15 x 15km

25

evaluation subset of the remote sensing data was fed through the trained data, or
simulated, to output a predicted map of bamboo presence/absence. This output
subset was then examined to determine how well the predicted values conformed to
the expected output values of O or 1. As in many practical applications, it was not
possible to collect a completely representative training sample. Therefore, the
output values of pixel values not seen in training are expected to vary relative to the
expected outputs and were categorized if they did not conform exactly to the
expected presence/absence values of 1 or 0. The prediction maps were then veriﬁed
with the training data and validated using independent data. The optimal algorithm
characteristics were chosen based on the validation accuracy and conformance to
expected output levels for the entire image. This algorithm was then used to

simulate the remaining 17 subsets to produce a full map of the reserve.

22.4 Comparison to Traditional Techniques

It was hypothesized that the underlying restrictions and assumptions of
traditional classiﬁcation methodologies would not allow accurate classiﬁcation of
understory features. To allow comparison to traditional techniques, supervised
classiﬁcations of bamboo distribution were conducted on the 15 x 15km validation
subset using Erdas Imagine v. 8.3.1 (Erdas 1999). The same training data used to
train the neural network within the validation subset were used to gather supervised
signatures. These signatures were categorized to allow Maximum Likelihood

classiﬁcations and retained as single signatures for a Minimum Distance

26

classiﬁcation. This allowed testing of the effects of classiﬁcation algorithms and
merging of spectral signatures. Each method is a standard supervised classiﬁcation
approach. However, it was anticipated that the extreme topography and overstory
variation would require more than one signature to obtain accurate results.
Utilizing different combinations of signatures allowed testing of these effects. The
resulting output classes from each method were recombined into presence and
absence categories. These binary classiﬁcations were then compared to the

independent validation data to determine classiﬁcation accuracy.

2.3. Results

2.3.1 Evaluation Results

Evaluation and comparison of the simulations were initially conducted on a
15km x 15km evaluation subset. For each algorithm where convergence was
achieved, conformance to expected values, veriﬁcation, and validation results were
examined. It was found that the full TM data set (all six bands) was required for
adequate convergence. Inputs using fewer bands converged more slowly or not at
all. Adequate convergence was also not possible using an algorithm structure
containing only one hidden layer. For comparison between algorithm structures
using more than one hidden layer, veriﬁcation and validation results were

examined.

27

Prior to veriﬁcation and validation analyses, the outputs were categorized
into presence and absence values. This was necessary since the training data were
not a fully representative sample of all land covers, combinations of land covers,
topographic effects, and understory conditions. Therefore, variations in the output
value from the expected 1 or 0 were seen. For example, comparison of output
values corresponding to training data plots showed value ranges of 0.99-1.00 as
classiﬁcation values for the presence of bamboo and only 0.00 for absence.
However, output values for pixels where ground samples were available but not
used for training showed a larger spread. In fact, for land covers (clouds, exposed
rock, snow) where training data were not available, output values were as high as
2.5. In addition, since complete control even within sampled vegetation types and
abiotic factors was not absolute, output values in vegetated areas ranged from 0 —
1.90. Therefore, when the entire image was simulated, output values from the
network less than 0.50 or greater than 1.50 were considered absence and values
greater than 0.50 and less than 1.50 were considered presence.

Algorithm verification was signiﬁcantly higher when using at least 24
hidden nodes in the ﬁrst layer and as high as 100% agreement for all convergence
levels tested when using at least 24 nodes in the ﬁrst layer and 48 nodes in the
second hidden layer. The validation data for six algorithm architectures given in
table 2.2 show the basic trend in network training. Validation results ranged from
less than 50% from networks failing to fully converge (not shown) to 82% for the

most optimal method shown. As shown in table 2.2, learning algorithms (e. g.

28

BFGS quasi-newton (BFG) and Levenberg-Marquardt (LM)) had less inﬂuence on
the overall accuracy compared to early stopping levels.

Examples of the outputs from the neural network algorithm tests on the
15km x 15km study area are shown in Figure 2.4. Consistent areas of agreement
among the outputs can be seen (Figure 2.4 letterboxes) and represent trends in
landcover such as major clearings and agriculture areas. The differences are less
obvious. In tests run to convergence levels of 1x10’8 and 1x 10'20 (Figures 4b and
c), landcovers not represented in the training data are not classiﬁed within the
expected 0 — 1 range of values and are noted as very bright features. Compared to
the 1x 10'5 trials (Figure 2.4a), the latter image shows more consistent bamboo
classiﬁcation and better conformance to expected trends in output values (0 - 1). In
addition, compared to overstory classiﬁcations and knowledge of the region, the
1x10"5 method seems to retain landuse features (i.e. human appropriated areas,
permanent rock, etc.) better. Based on these visual assessments and the validation
results shown in table 2.1, the algorithm with 2 hidden layers, 24 and 48 hidden
nodes, and convergence level of 1x10'5 with classiﬁcation accuracy for the 15 x

15km study area of 80 - 82% was selected for full reserve analysis.
2.3.2 Supervised Classification
Performing supervised classiﬁcations of the validation subset yielded

consistently lower classification accuracy than the optimized neural network

methods. Merging the individual signatures into two, presence and absence,

29

signatures and then performing a Maximum Likelihood (ML) classiﬁcation yielded
a 71% classiﬁcation accuracy based on the validation data set. Using each of the
individual signatures to perform a Minimum Distance classiﬁcation resulted in a
69% classiﬁcation accuracy of the bamboo. Subdividing the ML signatures into
spectrally similar categories (i.e. similar aspect, slope, overstory vegetation)

produced similar results of 71% accuracy.

2.3.3 Full Reserve Results

Full reserve images of outputs from the neural net stopped at 1x10‘5 are
presented in Figure 2.5. Figure 2.5 shows the mosaick of the 18 outputs. A
comparison between the bamboo predictions from the neural network to the ﬁeld
survey map shows an excellent correspondence between the predicted and surveyed
distributions (Schaller et a1. 1985). Temporal differences of actual distribution
between the surveyed and predicted bamboo maps are expected from natural
dynamics and human disturbance, but since potential habitat is controlled by abiotic
factors such as altitude, precipitation, and slope, overall distribution should be
consistent through time. A confusion matrix of the full reserve output validation is
presented in Table 2.3. Overall accuracy based on all validation data was 80%.
The matrix shows more clearly the correspondence between predicted and ground
truth data and those pixels not being correctly categorized. Of particular note is the

technique’s inaccurate prediction of pixels containing bamboo as having an absence

3O

of bamboo, or false negatives. This category represents 75% of the incorrectly
assigned pixels.

Further analyses examining factors inﬂuencing the errors and these false
negatives in particular showed very interesting trends. For example, no relation
between the percent of canopy closure and prediction ability was found. In other
words, the distribution of canopy closure for the mis-categorized pixels was the
same as all the data. It should be noted, however, that complete canopy closure was
rare and bamboo was not found under 100% canopy closure. Nor was any relation
of mis—categorized pixels to the percent bamboo found. Again, the distribution of
mis-categorized pixels was similar to that of all the data. However, signiﬁcant
trends were found in regards to co-occurring midstory and understory vegetation
within mis-categorized pixels. In the case of 90% of the false negatives, a midstory,
typically 2 — 5m deciduous sub-canopy trees and shrubs, either partially covered
and/or intermingled with the bamboo. In every case of the false positives, a grass
understory, with similar characteristics as Bashania, covered the forest ﬂoor.
Ground truth plots with co-occurring shrub in the case of false negatives and grass

in the case of false positives in all represent 92% of the mis—categorized pixels.

2.4. Conclusions and Discussion

Data and algorithm requirements were found through testing various

combinations of input, training data format, and algorithm architecture for

successful neural network prediction of understory bamboo presence/absence. The

31

best results were obtained when using all six bands of the TM data as inputs.
Infrared wavelengths have a greater canopy penetration compared to shorter
wavelengths (Lillesand and Kiefer 1994). The TM sensor records information in
three bands in the infrared, one in the near and two in the shortwave infrared. In
addition, general application of the trained network was most accurate when
stopped from reaching full convergence. Using this information, we were able to
derive spatial distributions with signiﬁcant correspondence to independent data.
The results showed as high as 82% correspondence between predicted bamboo
distribution and ground truth validation data.

In comparison to the supervised classiﬁcations, clear gains were made using
the neural network for bamboo classiﬁcation. Relative to the observed distribution
of bamboo, the supervised classiﬁcations seem to be more closely related to the
general trends in the dominant vegetation. It is possible that the gains in the
bamboo classiﬁcation using the neural network are due to the ability of the neural
net to more precisely learn trends in the dominant vegetation to that of the co-
existence of bamboo. However, based on ﬁeld observations (Table 2.1) and
classiﬁcation errors we do not believe this is the case. While not providing
conclusive evidence, trends in the classiﬁcation and the lack of any discemable
correlation Of overstory vegetation in forested regions to the existence of bamboo
lead us to believe that the neural network is more capable of utilizing canopy gap
and sub-canopy inﬂuences to more accurately classify understory bamboo. The
neural network is probably more capable of classifying minority features, adapting

to the variable inﬂuences of changing canopy conditions, and accounting for the

32

non-linear effects of sub-canopy vegetation. The misclassification trends discussed
in the results and shown in table 2.3 lead us to believe that the neural network
method is basing the classiﬁcation, at least in part, on understory vegetation to
increase classification accuracy of understory bamboo. We believe, for example, in
the case of the false positives the neural network is falsely classifying other
understory grasses as bamboo. The grasses may be spectrally similar enough to the
bamboo as to cause false positive classiﬁcations. In the case of the false negatives,
it is possible that the co-occum'ng‘shrubs are simply masking understory features.
These pixels of co-occurring shrub and bamboo represent about 89% of the false
negatives. The neural net may also be training against shrub containing plots as
typically shrub dominated areas are devoid of bamboo. To test this, further data are
being collected with emphasis on samples where there is co-occurring vegetation.

Canopy cover rarely exceeded 90% in Wolong and was on average around
56%. Neural network classiﬁcation of canopy gap vegetation is, therefore, very
feasible. The increased classiﬁcation accuracy and correlation between mis-
categorized pixels and understory vegetation suggest this is probably occurring
here. The degree to which the neural net is capable of incorporating non-linear
effects such as multi-scattering and IR transmission due to sub-canopy vegetation is
unknown. Further studies are necessary to test this and utilize this information if it
is available. For example, to accurately classify the percent cover and density of
understory vegetation, important data for many habitat models and mapping studies,
the non-linear inﬂuences are most likely required. Much of the understory

vegetation occurs below the canopy. Determining the fraction of cover will require

33

that the contribution of the sub-canopy vegetation be distinguishable and
proportional to the fraction of sub-canopy vegetation. To this end, we are
examining the relationship of the percent cover of bamboo to optical remote sensing
data using neural networks. However, further theoretical studies and controlled
experiments are needed as well.

Data from other types of sensors may also contribute to these analyses and
make parameterization of understory conditions more applicable. Signiﬁcant
structural information (e. g. biomass and vertical distribution) can be inferred (and
consequently some differentiation between structurally distinct vegetation types)
from Synthetic Aperture Radar (SAR) (Luckman et a1. 1997, Treuhaft and Siqueira
2000) and lidar (light detection and ranging) (Lefsky et al. 1999). Classiﬁcation of
sub-canopy vegetation is typically not possible from canopy penetrating SAR or
lidar data (e. g. confusion between understory shrub and bamboo). However, we
anticipate that the fusion of the increased biomass and structural information with
the signature information available from optical sensor data may allow enhanced
classiﬁcations and biophysical parameterization.

It is widely recognized that the understory contains signiﬁcant biomass and
diversity of vegetation. However, it typically remains unclassiﬁed using traditional
remote sensing techniques. The use of artiﬁcial neural networks to extract the
complex information available from optical remote sensing data seems promising as
a method to accurately classify understory features. In addition, the neural
network’s ability to learn the complex trends in the data and to generalize across

land covers make this method broadly applicable. Practical approaches to

34

classifying understory vegetation are needed for studies requiring more accurate
information of biomass, biodiversity, and habitat conditions, as in the case for the
endangered giant panda. We believe the results from this study at least point to a
need for further analyses on the inﬂuence of understory vegetation on remote
sensing data, information available from other data sets, and practical methods to
use these data to classify understory vegetation. Research in this area has the
potential to provide a practical approach to classifying understory vegetation and
developing information on the quantity and spatial distribution of understory

vegetation species on a scale previously prohibitive.

35

Acknowledgements

We thank the Wolong Nature Reserve Administration, especially Zhang Hemin, for
their kind assistance and support of our ﬁeldwork. In addition, we thank Daniel
Rutledge and James Brown (Dept. of Fisheries & Wildlife, Michigan State
University) for their technical support. We gratefully acknowledge ﬁnancial
support from the National Aeronautics and Space Administration (NASA Earth
Science Fellowship Award), the National Science Foundation (CAREER Award,
formerly Presidential Young Investigator Award), American Association for
Advancement of Sciences/T he John D. and Katherine F. MacArthur Foundation,
Michigan State University (All-University Research Initiation Grant, Global
Competence Fund, Institute for International Agriculture), the National Natural
Science Foundation of China, the Chinese Academy of Sciences, and China Bridges

International (North America).

36

Table 2.1: Relationship between overstory species and abiotic factors and

understory bamboo.

 

 

 

 

 

 

 

 

 

Bamboo (%) Canopy (%) Sloie (deg) Altitude (m)

Mean 41.2 i 36.2 55.6 i 20.8 23.0 i 10.7 2583 i 285
R 0.1 l 0.30 0.14
R2 0.01 0.09 0.02

 

37

 

Table 2.2: Influences of early stopping and learning algorithms on validation

and verification accuracy (percent).

 

 

 

 

 

 

 

 

 

 

 

 

Learning Accuracy Accuracy
Algorithm MSE Goal (Veriﬁcation) (Validation)
LM 111105 100.0 80.0
LM 1x108 100.0 62.2
LM 1x1020 1000 55.6
BFG 1x105 100.0 82.2
BFG 111108 100.0 75.6

Bro 1x102" 100.0 68.9

 

38

ANN Prediction

Table 2.3. Confusion matrix showing ground truthed values compared to
predicted presence/absence from full-reserve analysis based on BFG 1x10'5
algorithm output using Landsat TM data. Numbers in parentheses represent

those absence and presence validation points containing co-occurring *grass

and Tshrubs, respectively.

 

 

 

 

(iround Truth Data
Absence Presence
of bamboo of bamboo Accuracy
Absence
of bamboo 31 9(8l) 78%
Presence
of bamboo 3(3*) 17 85%
Overall
Accuracy 91 % 65 % 80%

 

 

 

 

 

 

39

Figure Legends. (Images in this dissertation are presented in color)

Figure 2.1. Examples of past bamboo distribution maps for Wolong Nature
Reserve. Figure 2.1a shows the full-reserve map derived from ﬁeld survey work
conducted from 1979-1983 (Schaller et al., 1985). The inset shows the approximate
area of the map shown in Figure 2.1b, a higher detail, smaller extent map derived
from work done by Johnson et al. (Johnson et al., 1988) in response to a mass die-

off ofBasham'afangiana within the reserve.

Figure 2.2. Examples of (a) Bashaniafangiana beneath a typical overstory and (b)

F argesia Robusta with co-occurring vegetation.

Figure 2.3. Representative schematic of an artiﬁcial neural network. The arrows
represent a feed-forward process of transforming input data, such as remote sensing
imagery, to an output space (e. g. bamboo existence/absence). Networks are trained
through a priori knowledge of output and input relations (ground-truth data and
corresponding remote sensing pixel values) and a reiterative back-propagation of

training errors to update the hidden layer weights.

Figure 2.4. Effects of algorithm structure are shown with variations in learning
algorithm and error goal. Figure 2.4a is the output map ﬁom BFG training
algorithm with convergence threshold of MSE = 1x10'5. Figures 4b and c Show

output maps from training algorithms with convergence thresholds of MSE = 1x10'

40

8, and IX 1020, respectively. Effects of generalization are apparent between methods
using a 1x10'5 threshold (a) and LM 1x10‘8 and 1x10‘20 outputs (b and c). Better
delineation of an agriculture area (inset A) using a 1x10'5 threshold and decreased
generalization at 1x10"8 and 1x10’20 shown as brighter areas in 4b and c (inset B)

not conforming to expected output values.

Figure 2.5. Full-reserve output maps from optimal algorithm input and structure.
Input data were Landsat TM data excluding the thermal band and algorithm
characteristics included 2 hidden layers with 24 and 48 nodes. The BFG learning
algorithm with a convergence threshold of MSE = 1x 10'5 were used. Figure 2.5a
shows only the bamboo distribution with green areas representing bamboo and gray
areas the absence of bamboo. Figure 2.5b shows the good correspondence of the
neural network output as it compares to the full-reserve output. Neural network
bamboo prediction shown as green overlaying ground survey distribution (legend

corresponds to Figure 2.1b).

41

Figure 2.1

 

 

42

Figure 2.2

 

 

.3 .3
\E

 

 

 

43

Figure 2.3

Input Nodes (Ni) Hidden Nodes (HNi) Output Nodes

 

Figure 2.4

 

 

 

 

 

Figure 2.5

 

 

 

 

46

 

CHAPTER 3

THE EFFECTS OF UNDERSTORY VEGETATION ON WILDLIFE HABITAT
ACROSS FRAGMENTED LANDSCAPES

In collaboration with

Scott Bearer, Li An, Yingchun Tan,
Zhiyun Ouyang and Jianguo Liu

47

Abstract

Understory vegetation is an important component of biodiversity and habitat
for many wildlife species, but previous large-scale studies on biodiversity and
wildlife habitat have suffered from lack of detailed information about understory
distribution. Consequently, it is unclear how understory distribution inﬂuences the
analysis of habitat quantity and fragmentation. To ﬁll this gap, we have recently
developed a method of mapping understory bamboo distribution using remote
sensing data (Landsat TM) and ﬁeld data in Wolong Nature Reserve (China), one of
the largest reserves designated for conserving giant pandas (Ailuropoda
melanoleuca). In this paper, we compared the results from panda habitat analyses
with and without bamboo information. The results show that the spatial distribution
of bamboo has a substantial effect on the quantity and spatial arrangement of panda
habitat. Total amount of habitat decreased by 29 - 56% and overall habitat patch
size decreased by 16 — 48% after bamboo information was incorporated into the
analyses. The decreases in the quantity of panda habitat and increases in habitat
fragmentation resulted in decreases of 41 - 60% in carrying capacity. Our study
suggests that it is necessary to incorporate understory vegetation into wildlife

habitat research and management to avoid overestimation of habitat.

4s

3.1. Introduction

Understory vegetation is a signiﬁcant component of wildlife habitat in
forested landscapes. Understory vegetation provides forage and cover for ground
birds (Estades and Temple, 1999), small and large mammals (Lindzey and Meslow,
1977; Dueser and Shugart, 1978), and invertebrate species (Ernest, 1989). For
example, lack of understory forage availability can limit white-tailed deer
(Odocoileus virginianus) ranges (McCaffery et a1. 1974), and berry-producing
understory shrubs are an essential component of the North American black bear’s
(Ursus americanus) diet (Powell et a1. 1997). As a result, it is widely recognized
that detailed knowledge of vertical structure and horizontal distribution of
understory ﬂora is often necessary to predict wildlife-habitat relationships
accurately (MacArthur and MacArthur 1961, James 1971).

While the importance of understory vegetation has long been recognized
and often incorporated in ﬁne-scale analyses, the inﬂuence of understory vegetation
on broad-scale analyses of habitat has not been quantiﬁed. We believe
incorporating understory information will have considerable implications for many
broad-scale theoretical ecological issues and practical applications. For example,
behavioral studies have shown that giant panda (Ailuropoda melanoleuca) habitat is
a function of forest cover, slope, altitude, and understory bamboo (Schaller et a1.
1985, Liu et a1. 2001). Previous efforts to collect information on the availability
and spatial distribution of understory bamboo. however, have been conﬁned to

coarse estimates over large areas or more reﬁned maps at ﬁne scales (Schaller et a1.

49

1985, Johnson et a1. 1989). Therefore, detailed information on the spatial
distribution of bamboo over large areas has not been available and past studies on
giant pandas have not been able to incorporate understory vegetation information in
analyses of the spatial distribution of habitat (De Wulf et al. 1989, Liu et a1. 2001).
As a result, Liu et a1. (2001) suggest it is probable these studies are leading to
conservative estimates of the loss and fragmentation of giant panda habitat.

However, to what degree understory vegetation inﬂuences broad-scale
habitat analyses has not been examined. The lack of detailed understory
information has forced past studies to base landscape analyses of habitat only on
features that can be measured easily over large areas such as obvious land cover and
topography. If understory vegetation, vertical vegetation structure, and other
difﬁcult to measure factors are limiting or have a different spatial distribution
relative to measured features, estimates of total habitat and landscape metrics may
not accurately reﬂect the availability and spatial distribution of habitat. This is
likely the case for past analyses of giant panda habitat. Based on extensive ground
sampling and a recent classification of the distribution of understory bamboo using
remote sensing data and an artiﬁcial neural network, we have found that understory
bamboo is distributed independently of forest cover. Mapping giant panda habitat
based only on overstory and topography is, therefore, likely leading to an
overestimation of available habitat and underestimation of habitat fragmentation
and isolation.

To illustrate the importance of understory vegetation on the quantity and

spatial distribution of wildlife habitat, we examined the inﬂuence of understory

50

bamboo on giant panda habitat. Speciﬁcally, we compared panda habitat with and
without recently developed high-resolution (30-meter) classiﬁcation of the spatial
distribution of understory bamboo over a range of a couple hundred thousand
hectares. Given the importance of understory vegetation throughout forested
ecosystems around the world, our comparative approach offers an example of the
degree to which habitat may be over- or under-estimated and emphasizes the need

to incorporate understory vegetation in many broad-scale habitat analyses.

3.2. Methods

To assess the impact of understory vegetation information on the quantity
and spatial distribution of giant panda habitat, we compared the landscape
characteristics of habitat classifications with and without understory bamboo
information within a nature reserve in southwest China. Comparisons were made to
a previous multi-temporal analysis of landscape trends in giant panda habitat. Our
previous analysis (Liu et al. 2001) provided a good baseline for comparison as it
used common approaches to broad-scale habitat analyses, but due to the lack of
historical data was not able to incorporate information on the distribution of
understory bamboo. For the study presented here, the land cover, slope, and aspect
data used to derive the 1997 time-series habitat classiﬁcation in Liu et a]. were
combined with recently developed bamboo data to reclassify habitat suitability.

The resulting habitat classiﬁcations were then compared to the original habitat

51

classification. Comparisons were made based on habitat quantity, habitat

fragmentation, and panda carrying capacity.

3.2.1 Study Area

Our study was conducted in Wolong Nature Reserve, in the Qionglai
Mountains of Sichuan Province, China (located between 102°52’ and 103°24’ E,
and 30°45’ and 31°25’ N). Wolong is one of the largest reserves (approximately
200,000 ha) dedicated to giant panda conservation, and is estimated to contain
~10% of the remaining wild panda population (c. 1,000 individuals; Zhang et al.
1997). Within Wolong, elevations range from 1,200 m to 6,525 m, creating several
climatic zones and consequently high habitat diversity.

Throughout Wolon g and most of the panda range, bamboo is found
predominantly as understory species. The distribution of bamboo species present in
Wolong is restricted by elevation. The two predominant species, umbrella bamboo
(F argesia Robusta) and arrow bamboo (Bashania fangiana) (both utilized by
pandas), are typically found between 1,500 and 2,550 m and 2,550 to 3,500 m,
respectively. Bamboo is distributed in the understory in patches ranging in size
from a single plant to hundreds of meters across. However, predicting the spatial
distribution of bamboo within these elevation ranges has not been possible due to a
lack of any signiﬁcant relationship between overstory and abiotic variables and the

presence or absence of bamboo (Linderman et al. in review).

52

The distribution of overstory vegetation in the reserve is also related to the
elevation: from approximately 1,200 to 1,600 m, the canopy is composed primarily
of evergreen broadleaf (Quercus) with broadleaf shrubs and occasional grasses.
From 1,600 to 2,000 m there is an increasing mixture of deciduous broadleaf
(Betula, Acer) overstory and understory vegetation. From 2,000 m to about 2,700
m, the canopy is a mixture of deciduous (Acer, Betula) and evergreen (Larix, Tsuga)
broadleaf. Above 2,700 subalpine conifers (Abies) dominate until about 3,600 m
where subalpine conifers, fern, and grass understory grade into alpine thickets and
meadow (Reid et a1. 1989, Schaller et a1. 1985).

In the past several decades, human activity has been a major force behind
forest loss and degradation of panda habitat (Liu et al. 1999, 2001 ). Grazing and
agricultural use have effectively removed forest cover and bamboo from some
areas. Other areas have been clear-cut, leaving a mixed midstory shrub layer and,
consequently, less bamboo. Finally, selective logging in other areas has changed the

species composition in the overstory and reduced canopy cover.

3.2.2 Comparison Study

Classification of habitat suitability for the Liu et a1. study was based on
previous giant panda behavioral research. For example, pandas are rarely seen
outside of forested areas. Sick and hungry pandas have been found in non-forested
areas and panda droppings have been found in canopy openings, but the vast

majority of their activity is in areas containing forest cover (Schaller et a1. 1985). In

53

addition, studies have shown that the main altitude range of the panda is between
2,700 m and 3,200 m, but extends down to 2,000 m and occasionally up to 3,500 m
(Schaller et a1. 1985, Ouyang et a1. 1996, Liu et al. 2001). Pandas prefer gently
sloping regions, restricting their activity to slopes less than 45 degrees and
preferring areas with less than 15 degrees slope (Ouyang et al. 1996). Furthermore,
forest understory vegetation plays a particularly vital role. Bamboo comprises
~99% of the panda diet, and pandas spend up to 14 hours per day foraging due to
bamboo’s low nutrient and energy content (Schaller et a1. 1985). Because of
pandas’ obligate relationship with bamboo, panda habitat is strongly inﬂuenced by
bamboo availability and distribution (Johnson et a1. 1989, Reid et a1. 1989).

To analyze the spatio-temporal trends of giant panda habitat in Wolon g
Nature Reserve over the past 32 years, Liu et a1. derived data corresponding to
pandas’ habitat preferences on data available from satellite images and topographic
maps. Images from 1965 (Corona satellite photographs), 1974 (Landsat MSS
imagery), and 1997 (Landsat TM imagery) were used to map the temporal
dynamics of the quantity and spatial distribution of forest cover in Wolong. The
photographs and imagery were used to derive classiﬁcations of forested/non-
forested for each time step. Slope and altitude values were derived from a digital
elevation model (see Liu et al. 2001 for details).

Detailed information on the spatial distribution of bamboo prior to 1997 was
not available. To be consistent over time and to provide a more conservative
estimate of habitat loss, bamboo information not was included in this previous

analysis. Therefore, habitat suitability was determined as a multiplicative

54

combination of the three factors (forest cover, altitude, and slope) available for the
32-year time span. Since non-forested areas are considered unsuitable habitat for
the panda, forest and non-forest classiﬁcations were multiplicative factors of l and
0, respectively. Slope and altitude multiplicative factors were proportional to the
observed use by pandas. The ﬁnal habitat classiﬁcation was a categorized suitability
measure of four classes termed highly suitable, suitable, marginally suitable, and

unsuitable (Liu et al. 2001).

3.2.3 Bamboo Classification

Remote sensing is a preferable method for mapping the distribution of forest
overstory over large extents. However, methods to map the extent of understory
vegetation like bamboo, even employing aerial photography, have not been
successful (Morain 1986, De Wulf et a1. 1988, Porwall and Roy 1991). The main
problem in classifying bamboo from remote sensing data is that bamboo occurs as
an understory species. The overstory typically limits spectral information from
understory bamboo, thus restricting traditional remote sensing classiﬁcation
approaches. To overcome this problem, we developed a new approach for
classifying the presence/absence of understory bamboo regardless of overstory
canopy (Linderman et al. in review). Using widely available Landsat TM imagery
as input and ground samples of understory and overstory conditions for training, a

non-linear artiﬁcial neural network achieved a bamboo classiﬁcation accuracy of

55

80% despite variations in canopy cover and the presence of co-occurring mid-story

and understory vegetation (Linderman et al. in review).

3.2.4 Incorporating Bamboo Information into Habitat Analysis

To examine the inﬂuence of bamboo information on landscape-level
analyses of panda habitat, we incorporated the above-mentioned bamboo data into
habitat analyses using the same procedures as those used in Liu et a1. (2001). For
consistency, we incorporated the bamboo data as an additional multiplicative factor
in addition to the three factors (forest cover, slope, and elevation) used in the
previous study.

The bamboo information was incorporated into habitat analyses in three
forms: unﬁltered data and two ﬁltered forms of these data. We incorporated the
unﬁltered data as a binary coverage (presence = l and absence = 0) at the 30-m
resolution derived from remote sensing imagery. Like many other species (e. g.,
Pearson et a1. 1996), however, pandas may perceive bamboo availability differently
from the patterns of bamboo distribution at the 30-m pixel level. For example,
previous behavioral studies (Schaller et a1. 1985) have found that pandas spend up
to 14 hours/day foraging, with signiﬁcant daily movement within and between
patches of bamboo. Typical distances traveled each day are around 300 m and
almost always <500 m. In addition, radio-tracking studies by Schaller et a1. (1985)
showed between-patch foraging distances were typically < 100 m. These studies

suggest that pandas perceive bamboo distribution at a scale of >30 m. Furthermore,

56

while most activity is largely restricted to regions with signiﬁcant amounts of
bamboo, pandas do use areas not containing bamboo, mainly for movement
between bamboo patches, territorial marking (primarily males) and travel to other
places (Schaller et a1. 1985). Thus, we used two ﬁlters (proportion filter and
majority ﬁlter) to reclassify the bamboo data to reﬂect pandas’ perceptions of
bamboo distribution, use of non-bamboo area, and the scale at which pandas
interact with bamboo. The bamboo data were ﬁltered by reclassifying the center
pixel of a square window based on the conditions of other pixels within the
window. The window sizes ranged from 3 x 3 pixels (90 x 90 m) to 21 x 21 pixels
(630 x 630 m), representing the lower and upper extremes in daily foraging distance
by a panda.

In the proportion filter method, the center pixel of a window was reassigned
a value of the proportion of pixels with bamboo within the window. This
represented an estimate of bamboo availability in the surrounding area as perceived
by a panda. The estimates were then categorized into three classes: highly suitable
(>50%), suitable (ZS-50%), and unsuitable (<25%). In other words, if more than
half of the pixels within a ﬁlter window contained bamboo, the center pixel was
classiﬁed as highly suitable. If less than one quarter of the pixels contained bamboo,
the center pixel was reclassiﬁed as unsuitable. Otherwise, the pixel was classiﬁed
as suitable. This classiﬁcation was based on bamboo availability as measured by
approximated panda use throughout the reserve and measured travel distance
between patches (Schaller et a1. 1985). Areas within Wolong known to historically

sustain high densities of pandas were used to represent high-quality bamboo habitat.

57

The majority ﬁlter adjusts the value of the center pixel to reﬂect the
conditions of the majority pixels within the window. For example, if at least 5 of the
9 pixels in a 3 x 3 window were classiﬁed as bamboo, the center pixel would be re-
classiﬁed as bamboo regardless of its original classiﬁcation. Like the proportion
ﬁlter, the majority ﬁlter maps the same areas of high-quality bamboo areas as
measured by the proportion ﬁlter. However, the proportion ﬁlter retains more
potential highly suitable, suitable, and marginally suitable habitat in areas of lower
levels of bamboo due to the relative inﬂuence of suitable bamboo habitat within the
proportion ﬁlter. The majority ﬁlter highlights only hi gh-quality bamboo areas and
provided an alternative estimate of the impact of bamboo on habitat availability and
fragmentation.

Classiﬁcations of overall habitat suitability were generated as in Liu et a1.
(2001) with the inclusion of the additional multiplicative factor of bamboo
suitability. This resulted in the same four categories of habitat suitability as defined
in Liu et a1. (2001): highly suitable, suitable, marginally suitable, and unsuitable.
Highly suitable bamboo areas (multiplicative factor of 1) resulted in no change to
the original habitat classiﬁcations. Unsuitable bamboo areas changed all previous
habitat classiﬁcations to unsuitable. Depending on the quality of the other three
factors used in the multiplicative index, suitable bamboo areas from the proportion
ﬁlter could degrade original habitat classiﬁcations by as much as one suitability

category.

58

3.2.5 Habitat Measures

Panda habitat under different methods of analysis was compared using three
measures: habitat quantity, fragmentation, and potential carrying capacity. Habitat
quantity included not only the total amount of all habitat classes (highly suitable,
suitable, marginally suitable), but also the amount of each suitability class.
Fragmentation measures the degree of discontinuity of habitat and is represented by
mean patch size. Potential carrying capacity was based on total quantity of core
habitat areas and density of pandas in the core habitat areas. A core habitat area was
designated as a habitat patch large enough to support at least one panda. Pandas’
home range varies from 3.0 to 6.0 km2 (Schaller et al. 1985). However, signiﬁcant
overlap occurs between home ranges. Schaller et al. (1985) suggest that prime
habitat (equivalent to highly suitable habitat in this study) has an average density of
l panda per 1.7 kmz. Therefore, to reduce the chance of underestimating total core
habitat, we deﬁned core habitat as any habitat (any combination of marginally
suitable, suitable, or highly suitable habitat) forming a contiguous patch of at least
1.7 kmz. Based on the frequency of observed use in different categories of habitat
(Ouyang et a1. 1996), we used density estimates of l panda per 3.4 and 5.1 km2 (2
and 3 times less population density than in highly suitable habitat) for suitable and
marginally suitable classes, respectively. Total area of each habitat suitability class
was then used to determine overall carrying capacity based on estimated population

densities for each habitat class.

59

3.3. Results

Incorporation of bamboo information into habitat analysis substantially
reduced the amounts of panda habitat and drastically changed the spatial
distribution of panda habitat (Figure 3.2, Tables 3.1 and 3.2). The impacts varied
with ﬁlter methods and ﬁlter sizes that resulted in different amounts and
distribution patterns of bamboo (see examples in Figure 3.1). For instance, using the
majority ﬁlter, highly suitable bamboo areas decreased by 11.8% to 32.8% for 3 x 3
and 21 x 21 windows, respectively, while mean bamboo patch sizes (3.97 to 17.9
ha) were about 8 to 34 times larger than those from the unﬁltered method (0.53 ha).
Total amounts of suitable and highly suitable bamboo areas derived from the
proportion ﬁlter were 59.9% to 92.6% more than those from the unﬁltered method,
respectively.

Unﬁltered bamboo data caused a considerable decrease in habitat quantity
and a marked increase in fragmentation (Table 3.1). While the amounts of all
habitat classes combined and each habitat class from the unﬁltered method were
similar to those from the majority ﬁlter (Table 3.1), mean patch size from the
unﬁltered method was considerably lower than the mean patch size based on the
two ﬁlter methods (Table 3.2). In other words, the ﬁlter methods resulted in larger
mean patch sizes and more contiguous habitat. As shown in Figure 3.2, the amount
of high-quality habitat (highly suitable habitat) was highest using the proportion
ﬁlter with ﬁlter sizes between 7 x 7 and 11 x 11 pixels. The quantity of high-quality

habitat using the majority ﬁlter was substantially lower for each ﬁlter size relative

60

to the proportion filter (Table 3.2). With the majority ﬁlter, the number of patches
of high-quality habitat decreased and patch size increased asymptotically relative to
ﬁlter size (Table 3.2). Differences in habitat class area relative to habitat
classifications not including bamboo ranged from a decrease of 40.1% for high-
quality habitat using the 11 x 1 1 proportion ﬁlter to a decrease of 54.2% for
marginally suitable habitat incorporating the unﬁltered bamboo data (Table 3.1).

Because results using different ﬁlter sizes varied considerably, determining
the appropriate ﬁlter size was necessary for comparisons of the results from
analysis without bamboo. For each ﬁltering method, we chose the ﬁlter size of 1 l x
l 1 pixels (330 x 330 m) according to behavioral studies of daily foraging activities
of pandas (typical daily travel distance was approximately 300 meters, Schaller et
a1. 1985). In addition, patch sizes and number of patches derived using a window of
1 1 x l 1 pixels were relatively consistent with those from neighboring ﬁlter sizes
(e.g. 7x 7 — 15 x 15 pixels).

Using the ﬁlter size of 11 x 11 pixels, habitat distribution patterns with and
without bamboo information differed considerably (Figure 3.3). Total habitat
quantity decreased by approximately 29% under the proportion ﬁlter (Table 3.3).
More signiﬁcantly, high-quality habitat was reduced by 34%. Patch size decreased
by 41% for suitable habitat and 18% for highly suitable habitat (Table 3.2). The
number of highly suitable patches decreased by 28%. Compared to the results from
analyses without including bamboo, bamboo information derived from the majority
ﬁlter (with ﬁlter size = l 1 x 11 pixels) also produced a smaller quantity of habitat

and a higher degree of fragmentation. Speciﬁcally, total habitat area decreased by

61

52% (Table 3.3), and mean patch size decreased by 48% and 16% for suitable and
highly suitable patches, respectively (Table 3.2).

Estimates of the carrying capacity based on analyses with and without
bamboo information offer further insight into the accuracy of each method. In the
19705, there were 145 pandas (Giant Panda Expedition 1974). Panda population
size declined to 72 in the early 19805 (China’s Ministry of Forestry and WF
1989). A recent population survey of Wolong has suggested that the current number
of pandas within Wolong is 74 individuals (Zhang et al. 1997). However, based on
habitat analyses without bamboo, carrying capacity (220 individuals) far exceeded
the survey estimates of the past and current panda numbers (Table 3.4). Including
bamboo in these analyses substantially lowered the estimate of the number of
possible pandas the current habitat in Wolong can support. Potential carrying
capacity derived from habitat analyses using unﬁltered bamboo data was only 27
pandas. However, the majority and proportion ﬁlter habitat classiﬁcations resulted
in estimates of 88 and 130 pandas, respectively (Table 3.4), falling within historical
and population range in Wolong (72 - 145). The fact that the carrying capacity
resulting from the majority ﬁlter method most closely matches current estimates of
panda numbers may suggest that pandas are concentrated in high-quality bamboo

are as .

62

3.4. Conclusions and Discussion

Incorporating information about understory vegetation into habitat analyses
of Wolong Nature Reserve resulted in substantially lower estimates of panda habitat
(at least 30%), lower estimates of habitat patch size (16 — 48%), and lower
estimates of carrying capacity (at least 40%). These new estimates have enormous
implications for panda conservation. For example, De Wulf et a1. (1988) reported
that total distribution area of giant pandas without bamboo information was
approximately 13,000 km2 (including Wolong, other reserves, and non-reserves). If
overestimation due to the lack of detailed information on the spatial distribution of
bamboo is consistent across the entire panda range, the total amount of panda
habitat is at least 3,900 km2 less than reported by De Wulf et a1. ( 1988). In addition,
notable isolation occurs between large habitat patches when bamboo information is
included (Figure 3.4). Pandas are extreme K-strategists, have low reproductive
success, and are currently at 'very low numbers (about 1,000). Sub-population
isolation and inbreeding are already a concern for this endangered species (Lu et a1.
2001).

To accurately estimate habitat distributions at the landscape scale, we have
found that data derived from remotely sensed imagery must relate to how species
perceive the environment. For this study, information from behavioral research
(Schaller et a1. 1985, Ouyang et a1. 1996) offered insight into pandas’ bamboo use
and relevant scales of interaction. We chose the most appropriate ﬁlter size to

reﬂect pandas’ daily movement patterns. The proportion ﬁlter offers the most

63

meaningful information relative to pandas’ perception of bamboo availability, as it
includes areas that have less than optimal bamboo still utilized by pandas. However,
the carrying capacity resulting from the majority ﬁlter most closely matches the
reserve’s current panda population size. Because the majority ﬁlter highlights
highly suitable bamboo area, this suggests current panda population size may be
below the potential carrying capacity as estimated by the proportion ﬁlter method.
Either habitat with less suitable bamboo is underused due to the low number of
pandas in the wild today, or some isolated patches of habitat are unavailable to
pandas.

The approach and results from this study have important implications not
only for panda conservation but also for biodiversity conservation and broad-scale
habitat analysis in general, as countless wildlife species depend on sub-canopy
structure and ﬂora, which are integral components of biodiversity and forested
ecosystems. Fortunately, information on understory structure and species
composition of forested ecosystems is becoming increasingly available (Linderman
et al. in review). Furthermore, biomass and structural information are being derived
from synthetic aperture radar and optical satellite data (Baret and Guyot 1991,
Beaudoin et a1. 1994). Recently, lidar (laser induced distance and ranging) has been
used to measure vertical structure below canopies (Lefsky et a1. 1999). As detailed
data of understory vegetation become available from remote sensing images, we
believe it is necessary and feasible to examine the impacts these data have on

landscape-scale habitat analyses, conservation strategies, and management policies.

Acknowledgements

We thank the Wolong Nature Reserve Administration for their logistic
support of our ﬁeldwork and the assistance from Jinyan Huang, J ian Yang, Hemin
Zhang, and Shiqiang Zhou. In addition, we gratefully acknowledge ﬁnancial
support from the National Aeronautics and Space Administration (NASA Earth
Science Fellowship Award), the National Science Foundation (CAREER Award,
formerly Presidential Young Investigator Award), National Institutes of Health
(National Institute of Child Health and Human Development, R01 HD39789),
American Association for the Advancement of Sciences/The John D. and Catherine
T. MacArthur Foundation, Michigan State University (All-University Research
Initiation Grant, Global Competence Fund, Institute for International Agriculture),
the National Natural Science Foundation of China, the Chinese Academy of
Sciences, Ministry of Science and Technology of China (G2000046807), and China

Bridges International (North America).

65

Table Legends

Table 3.1. A comparison between not including bamboo and three ﬁltered forms of
bamboo information in habitat classiﬁcations of giant panda habitat. The
differences in quantity of the four classes of habitat are shown based on habitat
classiﬁcations, including each of the four bamboo classiﬁcations (No Bamboo, N 0
Filter, Proportion Filter, and Majority Filter). The inﬂuence of ﬁlter window size is

shown for ten window sizes (3 x 3 - 21 x 21 pixels).

Table 3.2. The inﬂuences of bamboo data and classiﬁcation methodology (No
Bamboo, No Filter, Proportion Filter, and Majority Filter) and ﬁlter size (3 x 3 — 21
x 21 pixels) are shown for a) the number of habitat patches and b) the mean patch

size of habitat classes.

Table 3.3. The resulting quantity of total habitat is shown based on each of the four
bamboo classiﬁcations (No Bamboo, No Filter, Proportion Filter, and Majority
Filter) and ten ﬁlter sizes (3 x 3 — 21 x 21 pixels). Differences in total habitat
relative to the Liu et al. (2001) classiﬁcation (No Bamboo) show a loss of 29 —

56% habitat by including bamboo information.

Table 3.4. Inﬂuence of ﬁlter method on core habitat and carrying capacity. Core

habitat is deﬁned as contiguous areas of highly suitable, suitable, or marginally

66

suitable habitat forming patches > 1.7 kmz. Carrying capacity is a function of the

total area of each habitat class and panda densities.

67

Table 3.1. Inﬂuences of bamboo filter methods and filter sizes on habitat quantity.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Filter Habitat Class Area (ha)
M6221: and Highly Suitable Suitable “Efﬁgy Unsuitable
No Bamboo 1 1811 49329 991 1 135841
No Fm" 5764 23543 4542 173049
Proportion

Filter

3x3 6772 27411 12121 161318

5 x 5 6971 27884 14071 158696

7 x 7 7066 27908 14898 157750

9 x 9 7065 27713 15325 157518

11 x 11 7078 27447 15617 157480

13 x 13 7024 27140 15838 157619

15 x 15 6968 26808 15997 157848

17 x 17 6940 26479 16067 158136

19x19 6899 26159 16163 158401

21 x 21 6861 25818 16208 158735

Majority

Filter

3 x 3 6198 24404 4858 171429

5 x 5 6277 24209 4884 171522

7 x 7 6265 23931 4840 171856

9 x 9 6218 23608 4769 172296

11 x 11 6162 23234 4679 172817

13 x 13 6097 22825 4561 173409

15 x 15 6034 22384 4435 174039

17 x 17 6004 21982 4302 174604

19 x 19 5946 21585 4156 175205

21 x 21 5890 21206 4003 175792

 

68

 

Table 3.2. Inﬂuences of filters on patch number and size.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

w? :46 m3 E 88 :8 a x a
$4 m3 3a 26 £3 Sou e x 2
8.6 new 3m 08 S3 S: t x t
m? N: 8a 5: 8% $2 2 x 2
8.6 was 3% a : can 33 2 x 2 EE
QWM _N.D NMN ~mN~ 35M mmON : x — _ haw—Gnu:
Nam 8m 2 .N 5: £2 Sam 6 x 6
ﬁg 63 5.. S: can 3% h x N.
2 .N SN 2: 82 2% 8% m x m
3; m: 8.. 68m EM: 8% m x m
634 was 8a #3 8: £8 a x a
N: a: w? a? S: 88 2 x a
one 63 m2 :8 3% was t x t
86 Es as 8? £8 QR m. x 2
8m m2. mom em; 68m 2% 2 x 2 5:4.
Ra 8s 93 3mm 3% m2 .8. : x : 53.2.2.—
3». 6.6 N; 8S 3% 2mm 6 x a
63 can 5.. mm: 2% mm: b x h
2: VS 8.. 38 5: mm? m x m
m; ova 6: BEN 8m: 86.». m x m
was _ :5 _ NE 3% m8? 8:: _ _ 5:... A.z
com _ 6M: _ _ m2 E: mm; SQ _ $2.56.. A.z
655:5 1 6333 2.33 635.5 8% EE
2.2.362 635:5 EH: 2.23.52 633:5 2.3: Emcwmwwsoz
onmm :35— :32 8:35— .? uonE: Z . _

 

 

69

Table 3.3. Inﬂuences of ﬁlter methods and filter size on total habitat.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Filter T % Change
otal .1
Methods Habitat (ha) from No
and Slzes Bamboo”
No bamboo 71050
No ﬁlter 33848 -52.4
Proportion
Filter
3 x 3 46304 -34.8
5 x 5 48925 -3l.1
7 x 7 49872 -29.8
9 x 9 50104 -29.5
11x11 50141 -29.4
13 x 13 50003 -29.6
15 x 15 49773 -29.9
17 x 17 49485 -30.3
19 x 19 49221 -30.7
21 x 2| 48887 -31.2
Majority
Filter
3 x 3 35460 -50.1
5 x 5 35369 -50.2
7 x 7 35036 ~50.7
9 x 9 34596 -51.3
11 x 11 34075 -52.0
13 x 13 33483 -52.9
IS x 15 32853 -53.8
17 x 17 32288 -54.6
19 x19 31687 —55.4
21 x21 31100 -56.2

 

 

 

70

 

Table 3.4. Inﬂuence of filter method on core habitat and carrying capacity.

 

 

 

 

 

 

 

 

Core Habitat (ha)
Filter Highly Suitable Suitable NISalfiﬂlbalgy 52:33;
No Bamboo 11335.0 46253.4 8900.5 220
No Filter 1767.6 5258.5 786.6 27
Proportion 6636.1 23655 .7 10847 .4 130
Majority 5182.1 17512.3 3049.6 88

 

 

 

 

71

 

Figure Legends. (Images in this dissertation are presented in color)

Figure 3.1. Classiﬁcations of bamboo suitability based on proportion and majority
ﬁlters. Filter sizes include 3x3, 11x11, and 21x21 for the majority ﬁlter (3.1a, c, e,
respectively) and for the proportion ﬁlter (3.1b, d, f, respectively). White represents
areas of high quality bamboo areas and black represents unsuitable areas in each of
the images. The gray regions in the proportion ﬁlter maps (3.1b, d, 0 represent

areas of suitable bamboo areas.

Figure 3.2. The effects of ﬁlter methods and ﬁlter sizes on highly suitable habitat

class.

Figure 3.3. Habitat distribution throughout Wolong based on classiﬁcations without
bamboo (3.3a), with bamboo (proportion ﬁlter with ﬁlter size = 11 x 11 pixels,
3.3b), and with bamboo (majority ﬁlter with ﬁlter size = 11 x 11 pixels, 3.30).
Highly suitable, suitable, and marginally suitable habitats are shown in red, yellow,

and green, respectively. Unsuitable habitat is shown in gray.

Figure 3.4. Maps of core habitat patches (> 1.7 kmz) based on classiﬁcations
without bamboo (3.4a), with bamboo (proportion ﬁlter with ﬁlter size = 11 x 11
pixels, 3.4b), and with bamboo (majority ﬁlter with ﬁlter size = 11 x 11 pixels,
3.4c). Highly suitable, suitable, and marginally suitable habitats are shown in red,

yellow, and green, respectively. Unsuitable habitat is shown in gray.

72

Figure 3.1

 

 

 

 

E Highly Suitable
e . f Suitable
- No bamboo

 

 

 

 

73

Figure 3.2

 

Amount ot Highly Suitable Habitat

7200

gﬂ
88

 

6600 + Proportion

+ Majority

 

Area (ha)
2
8

 

 

6200
6000

 

5800

3 5 7 9111315171921

 

Number of Highly Suitable Habitat Patches

'8
._\

E

 

é

 

i

 

—0— Proportion
+ Majority

 

 

 

 

Number ot Patchea
8 8

3

 

 

 

 

O

f T

3 5 7 9 11 13 15 17 19 21

 

Mean Patch Size of Highly Suitable Habitat

4.00

 

 

3.50

'8

 

 

-0— Proportion
—I— Majority

 

2.50
J

 

Area (ha)

 

 

 

8

1.50

 

 

 

1m004
3 5 7 9111315171921

FitterShaManberotPixelaonOnOSIaota
Filterwooar)

 

 

 

 

74

Figure 3.3

 

 

 

 

 

 

0 10 20Kms

 

 

75

Highly Suitable
D Suitable

- Marginal
Unsuitable

Figure 3.4

 

 

 

 

 

010 20Kms

 

 

76

- Highly Suitable
E Suitable

- Mammal
m Unsuitable

CHAPTER 4

IMPACTS OF NATURAL AND HUMAN DISTURBANCES
ON FORESTS AND GIANT PANDA HABITAT

In collaboration with

Li An, Scott Bearer, Guangming He,
Zhiyun Ouyang, J ianguo Liu

77

Abstract

Human modiﬁcation of land cover has been a leading cause of ﬂoral and
fauna] species extirpation and loss of local and global biodiversity. As natural areas
become increasingly scarce, habitat becomes fragmented and animal and plant
populations become isolated. Thus, the spatial arrangement of human impacts is of
increasing importance. This is particularly evident in systems experiencing regular
natural disturbances such as in the mountainous areas of southwest China
supporting the remaining population of giant pandas (Ailuropoda melanoleuca).
Giant panda populations have been restricted to remnants of habitat ﬁ'om extensive
past land-cover change. We developed a spatiotemporal model of bamboo
dynamics and human activities and their impacts on panda habitat. We found that
local fuelwood collection and household creation will likely reduce critical habitat
relied upon by pandas during times of bamboo die-offs. Human impacts are
projected to result in up to 16% loss of habitat over the next 30 years. In addition,
we found that not only was it the accumulation of the loss of habitat, but also the
spatial location of the effects of human activities that will have a signiﬁcant impact
on giant pandas. The impacts primarily occur in habitat relied upon by giant pandas
during past bamboo die-offs. Therefore, human impacts are leading to decreased
total area of habitat and increased fragmentation and are likely making giant pandas

increasingly sensitive to natural disturbances such as the cyclical bamboo die-offs.

78

4.1. Introduction

It is widely recognized that habitat loss is one of the leading causes of local
and global species extinction (Ceballos and Ehrlich 2002). Natural (e. g. ﬁre,
drought, ﬂooding) and human (e. g. logging, urbanization) disturbances alter the
quantity, quality, and spatial arrangement of habitat, and, consequently, affect
species distribution and persistence. While many species have adapted to natural
disturbance regimes and are often resilient to natural changes in land cover or even
rely on them (Romme et al. 1995), human inﬂuences on land cover have reduced
the total quantity of natural areas, decreased the connectivity between species
populations, and destroyed or restricted access to secondary habitats that can be
crucial in times of natural disturbances. Increased understanding of the
implications of the spatial arrangement of human activities on landscapes relative to
the spatial arrangement of habitat are necessary to better predict habitat required by
species and allow mitigation of impacts on the remaining ecologically intact areas
at the earliest time possible.

Past studies of the impact of natural and human disturbances have
examined the total quantity of habitat lost (Sala et al. 2000), the resulting overall
fragmentation of habitat (Collins and Barrett, 1997), restrictions in habitat
connectivity (Diffendork et a1. 1995), and potential inﬂuences on meta-population
dynamics (Hess 1996). Models have also explored the implications of the spatial
arrangement of natural (Berjak and Heame 2002, Perry and Enright 2002) and

human (Liu 1993, Stéphenne and Lambin 2001) disturbances to natural systems.

79

However, the Spatial arrangement of human impacts relative to natural disturbance
regimes has received less attention. As a result, species area requirements and the
impact of human activities may be considerably underestimated or overestimated.
By not considering natural disturbances, past estimates may have misestimated not
only the total area necessary for many species, but also the relative importance of
habitat until now regarded as marginal or secondary habitat (F ahri g 2001). Further
studies on the arrangement of human activities and estimates of the quantity and
spatial distribution of critical habitat for population maintenance are needed.

The idea that the spatial arrangement of habitat and/or components of
habitat play a crucial role in ecosystem function, habitat suitability, and biodiversity
is certainly not new (Leopold 1933, Forman and Godron 1986). Recent studies
have expanded on this to also Show that the spatial arrangement of disturbances can
have a disproportional level of impact when they are concentrated in locations of
seasonal, breeding, or other crucial habitats. For example, Turner et a1. (1994)
found that elk winter survival in Yellowstone National Park after the 1988 ﬁres
were sensitive to the precise location and pattern of the burned areas. High-
elevation disturbance was less important than the loss of lower-elevation grasslands.
In addition, the removal of small wetlands surrounding larger wetlands has been
shown to critically impact populations in times of drought and even in non-drought
conditions (Takekawa and Beissinger 1989, Naugle et a1. 2001)

A similar situation exists for giant pandas in China. Once ranging across
most of present-day China, giant pandas are now extremely endangered and

restricted to only the western ﬁinge of their former habitat (Schaller et al. 1985).

80

Loss of habitat and poaching have reduced the panda population to its present level
of probably less than 1000 individuals (Zhang et al. 1997). While intentional and
incidental poaching have declined, land cover change pressures continue to threaten
vital habitat, even in designated protected areas (Liu et al. 2001). Decreasing
quantities of habitat are resulting in isolation of sub-populations of giant pandas and
may be increasingly restricting access to secondary habitats.

Past behavioral (Schaller et al. 1985) and more recent landscape studies
(Linderman et al. in review) have examined how pandas relate to the environment.
There is now an increased understanding of the local habitat requirements and the
scale at which pandas interact with the landscape. For example, Liu et al. (2001)
measured the aggregated landscape effects humans have had in the Wolong Nature
Reserve over time. However, little is known of the potential impacts local activities
will have in the future in terms of total area or spatial pattern of giant panda habitat
(Liu et al. 1999). In particular, little is known of the potential impact humans will
have on crucial panda habitat during periodic bamboo die-offs. Current policies
have mainly focused on protecting core habitat and have not incorporated the need
for secondary habitat in times of bamboo die-offs. Extensive human activity in
these areas following a die-off of the main bamboo might limit geographical groups
of pandas from required secondary habitat.

Fragmentation of panda habitat has resulted in isolation of sub-populations
and decreased genetic ﬁtness of the overall population that will have significant
implications for long-term conservation of the panda (Lu et al. 2001). However,

continued fragmentation and degradation of habitat may also have more short-term

81

implications. Similar to natural and human disturbances that have resulted in loss
of crucial habitat for other Species, the combined loss of habitat and the episodic
die-offs of bamboo increasingly threaten the panda. For example, in the mid-19708
several species of bamboo in the Min mountains mass-ﬂowered and subsequently
died—off. Over 138 pandas died of starvation since secondary bamboo species were
not available (Johnson et al. 1988). In 1983, a mass die-off of the primary bamboo
species (Basham'a fangiana) in the Wolong Nature Reserve also resulted in the loss
of the main food source for the local populations (Schaller et al. 1985). However,
at the time, pandas had access to secondary bamboo sources (F argesia robusta)
adjacent to primary habitat and there were no reported starvation deaths of pandas
(Schaller 1987). Access to secondary bamboo resources, therefore, appears to be
essential to the persistence of panda populations as a result of these large, cyclical
die-offs of bamboo species.

Liu et al. (1999) found, however, that human activities continue to have
considerable impacts to panda habitat and may have profound consequences in the
future. Commercial logging has likely been the main cause of measured losses of
giant panda habitat. Household activities have had limited impact on forested areas
in areas of preferred giant panda habitat. However, increasing numbers of
households, expanding agriculture activity, and increased fuelwood needs have
resulted in more extensive impacts of low-elevation forests. These are the same
forests relied on by pandas for Spring forage of new bamboo shoots and,

particularly, during times of massive die-offs of preferred high-altitude bamboo.

82

Giant pandas have co-existed with bamboo, their primary food source, for at
least thousands of years. Bamboo naturally mass-ﬂowers and synchronously dies-
off over extremely large areas every 3 — 120 years, depending on the species. In the
past, these die-off events were followed by a brief dormancy before the bamboo
regenerated to pre-ﬂowering biomass and distribution (Figure 4.1). Pandas
typically responded to die-off events by moving to other species of bamboo while
the die-off species regenerated. However, human activities are likely increasingly
interfering with the natural bamboo cycles and pandas interaction with bamboo.

For example, human activities are decreasing the total area of habitat. In addition, it
is likely that changes in forest structure and composition are also affecting the
ability of bamboo to regenerate to pre-ﬂower conditions (Figure 4.1). This leads to
two problems. First, humans are likely causing more loss of habitat than even
previously measured. Second, the spatial arrangements of the impacts may be
restricting pandas from secondary habitat relied upon during past bamboo die-offs.
In this study, we examined the dynamics of human impacts on forest cover
and bamboo distribution within Wolong Nature Reserve in southwest China (Figure
4.2). Similar to many reserves worldwide, local residents continue to rely on
natural resources in Wolong. For example, the population of local farmers residing
within the reserve has increased from approximately 2500 people in 1975 when the
reserve was established at its current size to about 4500 people in 2001. The rapid
rise in the number of local residents and an increased tourist industry within
Wolong have resulted in increased development and demand for natural resources

such as fuelwood, timber, and traditional medicinal plants. The combination of past

83

commercial logging and fuelwood collection has resulted in decreased quantity and
quality of panda habitat. Speciﬁcally, Liu et al. (2001) found that fuelwood
collection occurs primarily at low elevation areas in the general proximity of
households resulting in signiﬁcant loss of habitat pandas relied on in the 19808
during the last die-off of the primary bamboo Species. In addition, as fuelwood is
exhausted at low elevations, impacts are increasingly occurring at higher elevations
and core habitat areas.

Our model combines the spatiotemporal effects of household fuelwood
collection, new household creation, forest regrowth, and bamboo on giant panda
habitat. The objectives of this research were to (l) examine the relationship of
households to the environment, (2) examine the landscape effects of ﬁne-scale
household activities on giant panda habitat, and (3) predict future impacts of
different scenarios of household activities, combined with the dynamics of bamboo.
on giant panda habitat.

Using this model we examined the driving factors inﬂuencing the spatial
distribution of human activities including fuelwood collection and household
creation. Sub-models of household activities were then coupled with the
spatiotemporal distribution of forest cover and understory bamboo die-off models to
determine the potential impacts to panda habitat. Speciﬁcally, we focused on the
loss of habitat from bamboo die-offs, the spatial arrangement of human impacts,

and the potential implications for giant panda habitat.

84

4.2. Methods

4.2.1 Study Area

The study was conducted in Wolong Nature Reserve, which is located
between 102°52’ and 103°24’ E and 30°45’ and 31°25’ N covering approximately
200,000 ha (Figure 4.2). Wolong is one of the largest reserves dedicated to giant
panda conservation and is estimated to contain about 10% of the remaining wild
panda population (Zhang et al. 1997).

Approximately 40% of the reserve is currently forested (Linderman,
unpublished data). Elevations range from 1200 m to 6525 m creating several
climatic zones and consequently high biological diversity. The distribution of
overstory vegetation in the reserve is related to the elevation: from approximately
1,200 to 1,600 m, the canopy is composed primarily of evergreen broadleaf
(Quercus) with broadleaf Shrubs and occasional grasses. From 1,600 to 2,000 m
there is an increasing mixture of deciduous broadleaf (Betula. Acer) overstory and
understory vegetation. From 2,000 m to about 2,700 m, the canopy is a mixture of
deciduous (Acer, Betula) and evergreen (larr’xfsuga) broadleaf. Above 2,700
subalpine conifers (Abies) dominate until about 3,600 m where subalpine conifers,
fern, and grass understory grade into alpine thickets and meadow (Schaller et al.
1985, Reid et al.1989). Most forests in the reserve were logged (either clear cut or
selectively-cut) from 1916 until the reserve was established in 1975, reaching peak

intensity between 1961 and 1975 (Schaller et al. 1985).

85

Commercial logging typically resulted in relatively large clearcuts
distributed throughout the reserve. Logging has been ofﬁcially banned in the
reserve since 1975; however, to varying degrees illicit logging does continue (M.
Linderman, personal observation). Other human activities have also been a major
contribution to forest loss and, consequently, the Spatial distribution of habitat (Liu
et al. 1999, 2001). In 2001, 4,440 local residents in 1098 households resided within
the reserve. The majority of these residents are farmers with the primary economic
activities consisting of farming maize and vegetables, raising livestock such as pigs
and yaks, and collecting wild herbs. A household usually relies on fuelwood for
heating, cooking, and livestock feed preparation using 8 — 30 m3 of wood each year
(An et al. 2001). Selective logging for fuelwood collection typically changes the
species composition in the overstory and reduces canopy cover until all overstory
vegetation is removed. In some areas, forest cover is also removed by grazing and
agricultural use. Land tenure is through assignment by the local government and
typically limits agriculture to 0.5 — 1.0 hectares per household. Grazing areas are
seemingly local commons and are maintained by heavy grazing activity.

Wolong also contains up to 4000 plant Species including up to 11 species of
bamboo, two of which constitute the signiﬁcant portion of the total biomass,

F argesia robusta and Bashania fangiana. F. robusta grows to a height of 4 — 6
meters often in dense clumps spread through the elevational range of 1500 — 2600
meters. Pandas often migrate down the mountains in late spring to eat the new
shoots and were shown to rely on F. robusta during times of mast-seeding of the

primary bamboo, B. fangiana. B. fangiana is a smaller, thinner bamboo growing to

86

heights of 1 —- 1.5 meters. However, it often grows as large carpets of dense growth
as large as several hectares area at elevations of 2600 — 3600 m.

Most bamboo species in China are semelparous and are characterized by a
synchronized mast-seeding occuning every 3 — 120 years, depending on the species
(J anzen 1976). Mast-seedings of B. fangiana are estimated to occur every 45 years.
In 1983, a B. fangiana mast-seeding and corresponding mass die-off occurred over
an extensive area within the reserve. Pandas in Wolong continued to utilize small
patches that did not ﬂower and old stems, but were forced to migrate to lower
elevations in the winter of 1986 to forage on F. robusta. It is estimated that another
mast-seeding of B. fangiana will occur between 2025 — 2030. The intervals
between mast-seedings of F. robusta are unknown, but estimated at 60 — 80 years
(Schaller et al. 1985). The last mast-seeding of F. robusta is believed to have

occurred in 1949 (Wolong Administration, personal communication).

4. 2.2 Model Description

The model was developed in the SELES (Spatially Explicit Landscape
Event Simulator) framework (Fall and Fall 2001, Fall et al. 2001). SELES is a high-
level programming language that facilitates modeling of the temporal and spatial
dynamics of gridded landscapes. SELES also allows the incorporation of
georeferenced raster data, the design of sub-systems that interact on the landscape,

and temporal and spatial dynamics.

87

This model is intended to provide insight into the landscape-level interaction
between forest, humans, and understory bamboo. It seeks to balance parsimony and
realistic landscape driving factors for each aspect of the model and to examine the
inﬂuence of each of the model processes. The model included four sub-models:
household fuelwood activities, household creation, forest regrowth, and bamboo
growth and die-offs. The resulting impacts of the distribution of human activities
and bamboo spatiotemporal dynamics are integrated directly into giant panda
habitat suitability models and allow model predictions to be measured in terms of
changes to landscape indices of panda habitat. The four sub-models and their
interactions are shown in Figure 4.3. Each of the human activities inﬂuences the
spatial distribution of forest cover. Forest regrowth allows for forest re-
establishment and annual growth of non-climax forests. The spatial distribution of
forest cover inﬂuences the spatiotemporal dynamics of bamboo regrowth and die-
off. The suitability of giant panda habitat is then determined from forest cover
along with abiotic factors. Bamboo is incorporated into habitat models as
presence/absence of bamboo and, consequently, giant panda habitat.

The probability of a landscape event (e. g. fuelwood collection) occurring at
each cell is determined by the current state of the pixel (e.g. slope, aspect, forest
cover) and proximity to cells of a speciﬁc state (e.g. pixel containing a household).
The number of landscape events is determined by the sub-model characteristics
(e. g. total number of households) with the location of the event stochastically
determined relative to the cell probability. The event probabilities for each cell

(depending on the speciﬁc landscape event) could be determined by forest cover (%

88

pixel), slope and aspect (degrees), elevation (meters), agriculture cover (% pixel),
distance to households or roads. The modeling framework also allows for
landscape events to spread to neighboring cells (e. g. if a cell does not contain
sufﬁcient fuelwood for the annual collection of a household fuelwood needs,
fuelwood collection can spread to a neighboring cell to meet the annual needs of the
household).

The four sub-models and habitat suitability model are described below along
with examples of probability functions:

0 Fuelwood Collection - It was assumed that households collect fuelwood
based on availability, accessibility, and previous fuelwood collection
activity. Typically, fuelwood is collected around the household. AS these
areas are diminished, foraging extends to the neighboring areas
characterized by easy accessibility (Liu et al. 2001). Many households have
been forced to travel several kilometers to collect annual stocks of fuelwood
(An et al. 2001). Accessibility is characterized in this model by the distance
to collection site, slope, and elevation and is summarized as a cost function
relative to the distance to roads and main paths and topographic variability
(i.e. slope and elevation difference along path to the pixel). The probability

function was a linearly decreasing function of increasing cost:

P(fuelwood 1 cost) = (l — (cost/maximum cost))

89

Pixel forest cover and average yield per hectare determined availability.
Households are also more likely to return to the same pixel, if sufﬁcient
forest volume exists (An, unpublished data), or neighboring cells of
previous fuelwood extraction. Therefore, a higher probability of collection
was assigned to cells previously harvested and neighboring cells. The
overall probability of fuelwood extraction for each forested pixel is then a
multiplicative combination of these factors.

Household Creation — Since 1965 new households were established as
children moved out of their parents’ home or through immigration into the
reserve. Immigration is now restricted to new spouses of current residents.
Each new household is presumed to also establish its own agriculture land,
clearing the forest area or occupying previously deforested area. New
household establishment is based on suitable agriculture land and proximity
to transportation routes and other households (An, unpublished data). The
precise X,Y location of the actual residence is not included in this model.
Rather, households including the physical residence, agriculture land,
garden area, and surrounding edge are presumed to occupy cells of the
landscape. Suitable agriculture areas were based on abiotic factors: slope,
aspect and elevation. While agriculture activity occurs on Slopes up to 40
degrees, low slope areas are preferred. Preference for low elevation areas
was also noted. For example, relative probabilities for household placement

based on elevation were empirically measured as:

90

{ 0.00 (h > 2500) }
P(household l h) = { 0.08 (2250 < h S 2500) }
{ 0.82 (1750 < h S 2250) }

{ l.OO(hS 1750)}

In areas of higher elevation, preference is given to south-facing slopes to
maximize sunlight. Households were also more likely to develop land
adjacent to previously established houses within short distances (typically
less than 2 kms) to major transportation routes.

Forest Regrowth — Four forest types (Figure 4.4) were identiﬁed throughout
the reserve based on elevation and species distribution (Schaller et al. 1985).
Regrowth models were derived for each of the predominant species within
each elevational zone from published and empirical data (Yang and Li
1992). Initial stand volume was estimated for each elevational zone based
on approximate time and intensity of commercial logging activity. Regrowth
is calculated based on succession age and logistic regrowth of total volume.
Bamboo — The spatial distribution of bamboo was derived from 1997
remote sensing data and extensive ﬁeld surveys conducted from 1998 —
2000 (Linderman et al. in review). Regrowth and spatial and temporal die-
off characteristics were obtained from the literature (Janzen 1976, Schaller
et al. 1985, Reid et al. 1991, Taylor et al. 1991, Taylor and Qin, 1993). For
example, Taylor and Qin (1993) noted that while small forest gaps have

higher recruitment and growth rates, it is unlikely that bamboo will recover

91

from a mast-seeding in clearcuts (likely due to microclimate differences).
Reid et al. (1991) found that bamboo in the middle of its altitudinal range
had faster growth rates. This was included as a normal probability
distribution function deﬁned by the middle of the known altitudinal range
and known altitudinal range. And, Taylor et a1. (1991) found correlations
between the spatial distribution of B. fangiana die-off and environmental
conditions such as elevation, aspect and forest cover. For example, due to
the additional stress of steep, north-facing slopes bamboo was ten times less

likely to ﬂower during mass die-offs:

{0.1 (aspect _>_ 330 & slope > 40)}
P(dieoff | aspect,slope) = {0.1 (aspect S 30 & Slope > 40)}

{ 1.0 (30 < aspect < 330 & slope S 40)}

Habitat Suitability -The ﬁnal habitat classiﬁcation was a categorized
suitability measure of four classes termed highly suitable, suitable,
marginally suitable, and unsuitable (Liu et a1. 2001). The impacts from
human activities are reﬂected in the habitat suitability model as impacts
from fuelwood activity and agriculture development. Bamboo was
incorporated by measuring habitat containing bamboo. Measures of habitat
quantity and suitability allow analysis of the temporal and spatial dynamics

in terms of giant pandas, the inﬂuence of household characteristics, and

92

possible future scenarios on the likely landscape implications for giant
panda habitat.

Landscape events occurred on an annual time frame. The ﬁrst landscape event
each year is the establishment of new households and associated agricultural
development. Each household then collects its annual fuelwood volume. At the
end of the year, forest regrowth occurs for each forested cell and the suitability of

panda habitat updated.

4.2.3 Data and Parameterization

Several sources of data were used to parameterize, for input, and to validate
the model. Satellite data and topographic maps were used to describe the landscape
distribution of human activities, vegetation, and abiotic features throughout the
reserve. Socio—economic and demographic data were collected from local
government agencies and a household survey that was conducted from 1998 — 2001
(An et al. 2001). Literature on panda habitat and bamboo (particularly relating to
the 1983 B. fangiana die-off) were used for further parameterization of sub-models.

Georeferenced data used for this model included remote sensing
measurements of forest cover, topographic maps, and remote and ground surveyed
household distributions within Wolong. Additional layers were created by the
model relative to the initial layers such as distance to households. Each of these
layers was updated as the model progressed (e. g. forest cover was modiﬁed) where

applicable based on landscape events. Information on the landscape-level

93

distribution of forest was obtained from classiﬁcation of four dates of remote
sensing data, 1965, 1974, 1987, and 1997. The 1965 data are Corona stereo-pair
photographs acquired as part of the Corona photo-reconnaissance satellite project
(USGS Eros Data Center, Sioux Falls, South Dakota). The 1974 data are Landsat
MSS images and the 1987 and 1997 data are Landsat TM images. To account for
the spectral and spatial differences between the data, each image was visually
interpreted into forest and non-forest areas (for classiﬁcation details see Liu et a1.
2001). Topographic data were derived from a Digital Elevation Model (DEM).
Examples of these data are shown in Figure 4.4. The 1997 digital classiﬁcation of
the forest cover is a supervised classiﬁcation of the Landsat TM data based on over
300 ground samples. The bamboo distribution was derived from ﬁeld surveys
conducted from 1998 — 2000 and 1997 Landsat TM data (Linderman et al., in
review).

Location of each household was measured through the use of ﬁeld
measurements and Ikonos l-m resolution satellite imagery. Ikonos imagery
acquired in 2000 by Spacelmaging was georeferenced using ground control points
measured using a Global Positioning System with sub-meter accuracy (Trimble
Pathﬁnder Pro XRS receiver and Community Base Station). Households were then
identiﬁed in the images and the location recorded. All households created on or
before 1965 were used to create the initial distribution of households corresponding
to the initial forest cover information. New households have been added to the
reserve at a steadily decreasing rate each year between 1965 and 1997.

Approximately twenty-four new households were created each year.

94

A household survey was conducted from 1998 — 2001 and included each of
the households within Wolong (An et al. 2001). Households were queried on
fuelwood use, fuelwood collection, agricultural activity, household creation, and
other associated socioeconomic and demographic information. Additional
socioeconomic and demographic information was obtained from local government
records. Census information was obtained from each township within the reserve.
Local governments also maintain information on land allocated to each household.
From these data it was found that each household maintains on average 0.7 ha of
agricultural land. Including the area of the physical house, garden area, and
surrounding edge, the typical total area is approximately 0.8 hectares. Therefore,
the scale of the model was chosen to be 90 x 90 meters (0.81 hectares).

Fuelwood use was determined by surveying over 50 households and
physical measurements of daily use (All et al. 2001). The volume of wood varied
between 8 and 30 m3 and averaged 15 m3. An annual volume of wood used by each
household in the model was then 15 m3. Preference for fuelwood collection and
household creation sites was derived from comparing DEM and slope coverages,
and house locations and fuelwood sites. Distance to fuelwood collection sites varied
from 50 meters to over 5 kilometers. The average distance for 100 households
surveyed was 500 meters (Li An, unpublished data). Households preferred to
collect fuelwood in ﬂat areas (< 20° slope) and had a decreasing probability relative
to elevation.

Uncertainty in the 1965 stand volume of the various forest types posed the

most difﬁcult parameterization problem. While basic coverage information was

95

available from satellite photographs, data on the average volume throughout the
reserve were scant. Quantitative information dating back 35 years is difﬁcult to
obtain or non-existent. Schaller et al. (1985) suggest that much of the reserve was
commercially logged from the 1916 until 1975. Measurements taken in the late
1990’s found much of the lower altitude forests to be well below old-growth
volumes. Average volumes for broadleaf forests below 2600 m were approximately
80 m3/ha (Ouyang et al., unpublished data). It is likely that these forests were the
ﬁrst to be harvested in the ﬁrst half of the century and have regrown to current
volume levels. Based on regrowth data for the broadleaf forests in Wolong, we
estimated the average volume for 1965 to be approximately 45 m3/ha. Stand
volume for subalpine conifers was on average approximately 300 m3/ha (Ouyang et
al., unpublished data). Subalpine stand volume was high enough such that
variations in estimates would not signiﬁcantly inﬂuence the model.

Forest regrowth was included in the model to allow for previously logged
regions to regenerate and the addition of biomass and regrowth in selectively
logged cells. Separate logistic regrowth models were developed for each forest type
based on species composition, stand age, and altitudinal zone (Figure 4.4b). Model
parameters were derived from over 30 plots distributed throughout the reserve (Liu
et al. 1999) and Species regrowth and yield models derived by the Sichuan
Department of Forestry (Yang and Li 1992).

The spatiotemporal trends of the two dominant bamboo Species in Wolong,
B. fangiana and F. robusta, were included in models running from 1997 to 2030.

Bamboo was not included in models begun in 1965 as data on the distribution of

96

bamboo are not available prior to 1997. The initial spatial distribution of the
presence or absence of bamboo in 1997 was derived from remote sensing data and
extensive ground surveys (Linderman et al. in review). Species delineation was
based on altitudinal range and ﬁeld surveys (Schaller et al. 1985). B. fangiana and
F. robusta make up approximately 80% of the total bamboo within the reserve and
are the primary food sources for pandas in the reserve. Therefore, three classes of
bamboo are included in the model: B. fangiana, F. robusta, and all others (Figure
4.5b). However, because detailed life-history information of other species is
unavailable and to provide a conservative estimate of habitat impacts from bamboo
dynamics, die-offs for only the two main bamboo species were included in the
model.

Growth functions of each bamboo species and environmental inﬂuences on
regrowth and die-off characteristics were gathered from the literature (e. g. Taylor et
al. 1994). In 1983, approximately 85% of B. fangiana throughout the study area
died after ﬂowering. The areas of die-off were non-randomly distributed across the
landscape with bamboo patches experiencing environmental stress (altitudinal
limits, steep north facing slopes, and in large clearings) were more likely to not
ﬂower. Analyses of B. fangiana recovery after a mass-ﬂowering showed that
typical patches require 15-20 years to regenerate with an initial dormancy of 0 — 5
years. In addition, Taylor and Qin (1993) found lower seedling establishment and
poor overall regeneration in large forest clearcuts.

Behavioral studies have shown that panda habitat is a function of forest

cover, slope, altitude, and understory bamboo (Schaller et al. 1985, Ouyang et al.

97

1996). Ultimately, then, pandas are constrained by canopy cover while understory
bamboo strongly inﬂuences the quantity, quality, and distribution of habitat within
forested areas (Johnson et al. 1989, Reid et al. 1989). While both overstory and
understory forest components are known constraints, past estimates of the
availability and distribution of panda habitat over large extents have been based
solely on overstory features and abiotic factors, not on understory bamboo
distribution (De Wulf et al. 1988, Liu et al. 2001). Therefore, we determined habitat
suitability using a multiplicative combination of the three factors (forest cover,
altitude, and slope) available for the 30-year time span per Liu et al. (2001).
Because non-forested areas are considered unsuitable habitat for the giant panda,
forest/non-forest classiﬁcations were multiplicative factors of l or 0, respectively.
Slope and altitude multiplicative factors were proportional to the observed use by
pandas. For model scenarios run from 1997 — 2030, bamboo presence or absence

was included as an additional multiplicative factor of l or 0, respectively.

4.2.4 Model Validation and Sensitivity Analyses

Variations in individual parameters such as the rate of new household
creation, fuelwood use, and forest characteristics were made to test the validity of
each parameter and the sensitivity of impacts to each individual parameter. Exact
correspondence between the model predictions and the measured distributions was
not expected as the model is stochastic with the probability of a landscape event

(e. g. fuelwood collection, bamboo die-off) occurring in a cell based on the

98

biophysical parameters of the cell. Based on the variability of initial tests, the
model was run 20 times for each scenario to account for the stochastic processes.

To examine the inﬂuence of various parameters on the model predictions,
parameters from each of the sub-models were tested to examine the relative
inﬂuence on the related landscape event for model runs from 1965 - 1997. The
overall inﬂuence of each component of household submodel (abiotic, proximity,
and cost function) was examined by comparing various model scenarios (e. g. model
predictions excluding multiplicative factors) and the actual household distribution.
This was done to show the overall inﬂuence each function had on the model
selection of new households and as some functions could not be varied
systematically (e. g. abiotic inﬂuences were based on conditional probabilities).
Accuracy and landscape metrics were calculated based on the average accuracy and
indices of the final output from each of the 20 simulations. Systematic analyses of
parameter sensitivity of individual parameters were also conducted for the fuelwood
model such as the propensity to return to previous fuelwood collection sites and
distance to fuelwood collection sites. Since parameterization of stand volumes for
broadleaf forests below 2600 meters contained relatively large uncertainty, several
average stand volumes for the broadleaf forests were tested including 30, 45, 60, 75,
and 90 m3/ha. Accuracy assessments and landscape measures of fuelwood
activities were based on forest cover within the 5 km buffer.

The model was validated based on Simulations started in 1965 with the
initial distribution of forest based on the classiﬁcation of forest/nonforest categories

from these photos. The original distribution of households was based on all

99

households established prior to or in 1965. Validation Simulations were run for 32
years to correspond to the latest remote sensing data available. Two validation
methods were used. First, predicted fuelwood impacts on forest cover were
compared to visual delineations of forest cover from 1974, 1987, and 1997 satellite
imagery (Liu et al. 2001). Visual delineation of the four time steps provided a
consistent methodology for forest classiﬁcation relative to the forest classiﬁcation
used for the 1965 Corona photographs. The 1974, 1987, and 1997 data were used
for visual comparison of the multitemporal trends in predicted versus the measured
forest changes. Second, measurements of the distribution of households and digital
classifications of forest cover as measured in 1997 were compared to ﬁnal outputs
from the model. Digital classiﬁcation of the 1997 forest cover was possible with
extensive ground sample data and provided a more detailed snapshot of the current
distribution of forest cover. Household locations as measured in 1997 were
compared to locations of predicted households at the end of the 32-year simulations.
Comparisons between the simulated landscapes and measured landscapes
included a direct comparison between predicted change areas and measured change.
In addition, comparisons were made in the differences in the quantity of forest area
and disturbed areas and landscape metrics such as patch size, Shape, and
complexity. Timber logging activities continued after the establishment of the
reserve in 1975 and there is little information available on the total quantity or area
affected. To minimize the effect of natural and external inﬂuences on the accuracy
assessments, analyses were constrained to regions within 5 kms of the current

household distribution. This distance corresponds to the maximum distance

100

households travel to collect fuelwood. The accuracy of the model is reported as the
percentage of predicted pixels that correspond to measured pixels (e. g. predicted
non-forest vs. measured non-forest pixels). This ignores possible omission errors
and was used because of the difﬁculty in distinguishing natural variability in and
external impacts on forest cover from household activities even within 5 kms of the
households. Visual comparisons of model predictions and measured forest cover
change are Shown for comparison between commission and omission errors.

Given the difﬁculty distinguishing between timber logging, fuelwood
collection, and natural variability in forest cover, accuracy assessments of the model
predictions relative to measured landscape trends (particularly those from the
detailed classiﬁcation) do not provide a complete picture. Additional measures and
indices are used to compare the outputs of the model to measured impacts. The
impacts measured from projection simulations are reported as the landscape indices
relative to the impact of interest (e. g. household distribution and forest cover).
Indices were chosen relative to Riitters et al. (1995) and included total number of
patches, mean patch Size, corrected perimeter to area (p/a) ratio (Baker and Cai
1992), connectivity between patch centroids (Forman and Godron, 1986), and the

aggregation index (He et al. 2000).

4.2.5 Household Impacts

To examine the relative inﬂuences of different household scenarios on the

landscape, a variety of scenarios were run from 1965 until 2030 with 1965 baseline

10]

parameters and 1965 land-cover and household data used until 1997. Model
parameters were then varied for 1997 to 2030 to examine possible changes in
household characteristics. These scenarios represent situations where new policies
are introduced after 1997. Parameters examined included fuelwood consumption
per household and the household growth rate (or immigration/emigration rate).
Varying household characteristics was done to examine the effects of household
parameters on model outputs in the absence of natural and external (e. g.
commercial logging) variability. This allowed relative comparisons between
models and to highlight scenarios of interest to be run in the 1997 - 2030 scenarios
described below. Model scenarios were compared based on impacts to giant panda
habitat as deforestation for fuelwood and household construction removed habitat
and forest regrowth returned land to potential habitat.

Several levels of fuelwood consumption (30, 15, 10, 5, and 0 m3/yr) and
household growth (36, 24, 12, 0, -12, and —24 new houses created or removed each
year) were examined. Several combinations of these parameters were also
examined (Table 4.3). These levels were chosen to reﬂect possible future
household characteristics resulting from new policies and management efforts such
as subsidies, forest restrictions, and/or increased accessibility to electricity. For
example, efforts to eliminate fuelwood collection and reclaim agriculture land were
initiated in 2000. Subsidies have been offered in exchange for managing forests.
However, it is unlikely that fuelwood collection will completely cease. As
economic opportunities increase and demand for meat products increases, farmers

may increase the production of livestock and, consequently, actually increase

102

fuelwood consumption. Electricity prices are also currently unaffordable for most
local farmers, particularly for energy consumptive needs such as heating and
cooking. In addition, efforts to encourage emigration out of the reserve are being
instituted. However, an increasing preference by younger adults to establish new
households and in response to subsidy opportunities, new households have
increased at higher rates than in past years. To reﬂect the possible range of values,
fuelwood consumption levels were chosen based on the current maximum known
household consumption (to reﬂect increased livestock production) to no fuelwood
use. Household creation rates ranging from increases in household establishment to
a net emigration were examined to reﬂect the possible policy inﬂuences on

household activities.

4.2.6 Integrated effects of household activities and bamboo dynamics

Projection scenarios were run for 33 more years (1997 — 2030) to examine
the combined effects of human activities and potential bamboo die-offs. The model
was run for 33 years starting from 1997 with the latest available household and
forest cover data. These models were coupled with the spatiotemporal trends of
understory bamboo (information on the spatial distribution of bamboo is available
only from 1997). This coupled model was used to predict actual impacts on forest
cover, possible bamboo dynamics and associated interactions with forest cover, and

possible implications for giant panda habitat.

103

The latest household and forest data as measured from satellite imagery and
collected in the ﬁeld were used to develop the ﬁnal model. Figure 4.5a shows the
measured forest cover based on digital classiﬁcations of Landsat TM data acquired
in 1997. Figure 4.5b shows the measured distribution of bamboo in 1997. These
data, as well as measured household locations, were used in the scenario runs from
1997 — 2030 to examine potential impacts on giant panda habitat from the coupled
human activities and bamboo dynamics.

Four scenarios of household creation and fuelwood consumption were
chosen from the scenarios tested above. These included household creation rates of
24, 12, -24 and 0 houses per year. Fuelwood consumption levels of 15, 10, and 0
mg/year were also used. The scenario of 0 houses/year and fuelwood consumption
of 0 m3/year was used as a baseline for comparison. The other scenarios were
chosen to represent the responses found in section 3.2 above and represent ranges in
predicted future possibilities under business as usual and management actions to
reduce fuelwood use and household creation. The parameters of 12 houses/year and
10 m3/year, representing efforts to encourage emigration and decreased reliance on
fuelwood, were used as the most likely conditions for the next 30 years.

Three bamboo scenarios are presented. First, habitat availability is
measured if neither B. fangiana nor F. robusta die-off during the model run. B.
fangiana has a fairly well known intermast period of 45 years with the next die-off
predicted to occur approximately in 2027. Therefore, two die-off times of F.
robusta in addition to a B. fangiana die-off occurring in 2027 were coupled with the

above mentioned household parameters, a B. fangiana die-off occurring twenty

104

years prior to the B. fangiana and a simultaneous die-off. Twenty years was chosen
as recovery time as studies of past die-offs in the reserve suggest 15 — 20 years until
complete reestablishment. However, as forest cover affects the probability of
reestablishment of bamboo, the dynamic interaction between human activities and
bamboo ﬂowering may affect the reestablishment of bamboo to its former

distribution.

4.3. Results

4.3.] Model Validation and Sensitivity

Figure 4.6 shows a multitemporal comparison of the predicted 32-year
simulation of household activity and the measured forest cover within 5 km of all
households. There appears to be a good correspondence between the model outputs
and measured forest distribution. The basic trends in forest cover are comparable
between measured and predicted distribution of forest cover though some
differences from extraneous activities are apparent. In addition, the 1997 modeled
household distribution, based only on the initial 1965 distribution of households, is
very similar to that of the measured distribution from ground surveys and satellite
imagery for 1998 — 2000 (Figure 4.7a).

Various model scenarios resulted in distinct differences in the distribution of
household locations. Considering all household factors resulted in a slight increase

in the number of patches and p/a ratio, a 44% larger mean patch Size, and slightly

105

higher connectivity and aggregation indices (Table 4.1). Not including topographic
preferences resulted in 71% more patches of households (Table 4.1) and in house
placement in regions of distinct topographic relief (e. g. areas of extreme slope)
(Figure 4.7b). Excluding the distance and topographic variation from main
transportation routes resulted in a wide distribution of households (Figure 4.7c).
The number of patches was more than three times the measured distribution. Mean
patch size and p/a ratio were both considerably lower (Table 4.1). And, the lack of
a proximity factor results in decreased clumping of households (low connectivity
and aggregation indices), smaller patch size and an increase in the number of
patches relative (Table 4.1) to the measured distribution of households (Figure
4.7d).

Accuracy in terms of predicted household locations agreeing precisely with
measured locations of household distribution varied from 20 — 27% (Table 4.2).
Precise prediction of the 90 x 90 meter plot was not the intention of this model.
Foremost, this is not saturated system. This leads to quality areas not being
occupied and complex edges that are difﬁcult to predict empirically. In addition,
the model has a stochastic decision process and also does not include household-
level inﬂuences on household placement. However, as the spatial arrangement of
households may have an impact on habitat, particularly crucial secondary habitat,
we examined the percent of predicted households falling in close proximity (l, 2,
and 3 pixels) of measured households. Over 88% of the predicted households from
the complete model were within three pixels of the measured locations (Table 4.2).

This suggests that the model was predicting households essentially within the same

106

areas as those measured to also contain households. Not including the distance
function resulted in the lowest accuracy of 63% for predicted households within 3
pixels of measured households. Not including a preference to create new
households next to existing households resulted in an accuracy of 80%. Excluding
the selection based on abiotic factors (i.e. slope and elevation) resulted in an
accuracy of 81%. The complete sub-model had an accuracy of 27, 68, 82, and 88%
for predicted households within 0, l, 2, and 3 pixels from measured households
(Table 4.2).

Analyses conducted for each of the fuelwood parameters Showed expected
inﬂuences from variations in the distance and proximity factors (Table 4.3).
Relaxing the tendency for households to collect fuelwood from previously cleared
areas resulted in more fragmented impacts and is reﬂected in the landscape metrics.
Variation of the proximity factor three times more or less likely to return to
previous sites resulted in 35% fewer and 52% more patches and 54% larger and .
34% smaller patch sizes, respectively (Table 4.3). In addition, perimeter and
connectivity indices Show increasing clustering as the proximity factor is increased.
Accuracy of predicting impact areas, however, only varies by less than two percent.
Varying the cost factor by 20% resulted in similarly expected results. Easing the
inﬂuence of the distance factor resulted in more dispersed impact occurring in
smaller patches. This is seen in the patch characteristics with more and smaller
patches and decreased p/a ratios and decreased connectivity (Table 4.3). Increased
probability of using near areas conversely increased patch size, decreased patch

number, and increased connectivity between patches. Patch Size varied by 17.9% to

107

33.7% and patch number varied by 24.1% and 20.5% for a twenty percent decrease
and increase in the cost factor, respectively (Table 4.3).

Expected trends in deforestation relative to initial stand volume were
decreasing area of impact and decreased fragmentation Since more volume was
available in preferred collection areas (Table 4.3). While the outputs using each of
the above mentioned initial volumes shown in Figure 4.8 do seemingly conform
largely to these expectations, increased peripheral impacts occur at both increased
initial volumes and decreased volumes. Landscape metrics and overall model
accuracy also Show this trend in the model (Table 4.3). The lowest number of
patches was at initial forest stand volumes of 45 m3/ha. Decreasing stand volume
resulted in larger overall impact patch size, particularly the core area nearest to
households; however, smaller peripheral impacts were more common. As initial
stand volume was increased, the overall impact decreased, however small pockets
of impact emerged where more continuous impacts previously existed. These
trends are clearly shown in the decreasing patch perimeter and mean patch size.
The accuracy of predicted impact sites relative to measured impact also reﬂects
more concentrated impacts as initial volume is increased (Figure 4.8). AS fuelwood
activity is focused on core areas near households, model accuracy increases. At an
initial stand volume of 30 m3/ha, the overall prediction accuracy is approximately
55%. As the volume is increased, model accuracy increased to 64% when using
90m3/ha (Table 4.3). The increase in accuracy is largely a result of smaller areas
being affected only near households and decreased inﬂuence of stochasticity in

choosing distant fuelwood Sites.

108

4.3.2 Household Impacts

Projected household impacts on panda habitat are shown in Table 4.4.
Current levels of household creation and fuelwood consumption resulted in nearly
10% more habitat loss below 2600 meters. An increase in fuelwood consumption
(30 m3/yr) could impact up to 16% of the low-elevation forest. Reducing fuelwood
consumption by two-thirds reduced habitat loss by 59% compared to current trends.
However, to maintain levels of habitat as measured in 1997, the removal of
households at a rate of at least 50% the current level of creation or complete
cessation in fuelwood collection is required. However, as seen from a 50% increase
in household creation with no fuelwood collection, increased population and
resulting household creation contributed little to habitat loss because considerable
areas around households are already cleared of forest cover. Most likely estimates
(due to recent policy and management changes) for future new household creation
and fuelwood consumption (12 households/year and 10 m3/year) resulted in
approximately 30% less forest loss than current levels of new household creation

and fuelwood consumption.

4.3.3 Combined effects of household activities and bamboo dynamics

Human impacts on available giant panda habitat are shown in Figure 4.9a.

Increased impacts around households and near main transportation routes are

evident. The magnitude of these impacts for the four scenarios is shown in Table

109

4.5. Decreased area of impact relative to the 1965 to 1997 time period is expected
as forests recover biomass. However, up to 6.5% of all habitat is still predicted to
be lost relative to the current forest conditions in the near future from fuelwood
collection. And, these losses take place in a relatively small area of the reserve
with over 95% of the impacts occurring below 2600 meters of elevation. A
disproportion percent of habitat below 2600 meters will be impacted with up to
17% of low elevation habitat being lost (Table 4.6). Particularly worrisome is that
up to 15% of highly suitable habitat is impacted.

The total area of habitat containing bamboo if no die-offs occur is
approximately 30,000 ha. Up to 52% of this habitat may be lost during a
simultaneous die-off of B. fangiana and F. robusta (Table 4.7, Figure 4.9b).
Human impacts on habitat are relatively small compared to the loss from bamboo
die-offs. However, the total habitat available when considering possible bamboo
die-offs is only 14,000 — 16,500 ha. Human impacts are predicted to result in up to
1500 ha of loss of bamboo habitat representing approximately 10% of the
remaining habitat below 2600 meters (Table 4.8). In addition, mean patch sizes
were considerably lower after bamboo die-offs compared to the landscape without
bamboo die-offs (Table 4.9). In addition, the cumulative long-term effects of
human activities on habitat (the relatively permanent land-cover changes and spatial
location of these changes) now make pandas increasingly vulnerable to bamboo
dynamics. Human impacts have further reduced the size of the mean patches of
habitat suggesting that timing of bamboo die-off and human activities will inﬂuence

access to habitat (Table 4.9). AS human impacts are focused in areas of F. robusta

110

habitat, areas relied on by pandas during the 1983 mass-ﬂowering of B. fangiana,
loss of access to these habitats may have considerable impact on local panda
populations. The model predicts that up to over half of the initial bamboo habitat
will be lost given the coupled impacts of human activities and a simultaneous
bamboo ﬂowering. Even if F. robusta ﬂowers with a sufﬁcient time for recovery
(~20 years) prior to B. fangiana ﬂowering, total habitat remaining is only about half
the original habitat. Figure 4.10 shows the distribution of these impacts. Even with
conservative estimates of ﬂowering rate of F. robusta, total habitat in certain areas
is severely reduced. This is particularly noticeable in the north and west edges of

the extent of habitat in the reserve (Figure 4.10).

4.4. Conclusions and Discussion

The landscape approach to modeling the human and biological dynamics of
Wolong Nature Reserve provided considerable insight into the historical trends and
conditions of the reserve, the driving factors of human causes of land cover change,
and the potential consequences of human alterations of land cover and bamboo
dynamics on panda habitat. In addition, this particular model can be used to predict
conﬂicts between human activities and species’ habitat requirements and more
speciﬁcally suggests human modiﬁcation of secondary habitat can have
considerable ecological consequences.

A lack of detailed historical information for Wolong posed problems for

accurately modeling the human interface with the environment. Except for the

111

Corona photographs used for this study, very little information on the state of the
forest in 1965 was available. However, by comparing projections of household
activities from 1965 to a time where we have more detailed information permitted a
more detailed estimate of forest conditions and insight into factors contributing to
habitat loss. Comparisons of predicted forest loss from 1965 to 1997 to measured
forest conditions in 1997 for several scenarios of the average starting volume of
low-elevation forest conﬁrmed that these forests were already at relatively low
volumes and magniﬁed human impacts since 1965. It is possible that large-scale
logging ocurred concurrently with household fuelwood collection from 1965 until
1975 or later. While timber activities almost certainly continued after 1975, no
large-scale logging was noted by researchers present in the reserve from 1983 to the
1990’s. Forest loss after 1975 until 1997 are likely due to a combination of
households and ﬁner-scale timber activities and exacerbated by already low average
stand volumes from previous large-scale activity. As these forests are increasingly
lost, fuelwood activities move to higher elevation forests with increasing losses of
core habitat.

The model is relatively parsimonious. The placement of new households is
explained by only four factors: distance to roads, proximity to other households,
slope, and elevation. Using only these four factors, however, the model accurately
predicts household creation nearly 90% of the time within 3 pixels of the measured
distribution of households. Fuelwood collection also is only based on a few
landscape variables: distance to roads, previous fuelwood collection locations,

slope, and elevation. Again, the model captures the trend in household reductions

112

in forest cover. This simplicity and success of the model suggests basic landscape-
level factors have considerable inﬂuences on the spatial distribution of human
activities. Households were present in the reserve prior to the establishment of the
current transportation routes. These roads and the introduction of mechanized
transportation have likely led to increased agricultural activity along these routes
and increased access to forests near the roads. In addition, as the reserve is situated
in a relatively mountainous area, topography plays a signiﬁcant role in shaping the
spatial distribution of household activities. Agricultural activity requires relatively
ﬂat land and easy access to transportation. While fuelwood collection is less
dependent on the collection site qualities than the cost factor of the distance to
roads, the slope, elevation change, and overall accessibility inﬂuence the location of
collection sites.

This model also provided means to examine the role of household
characteristics on possible future impacts to giant panda habitat. Most notably,
considerable changes in fuelwood consumption and/or household creation rates
were required to maintain the current area of forest. While an increase in household
creation itself led to only small decreases in forest area, even limited fuelwood
consumption resulted in relatively large losses in habitat. This is likely because
areas around existing households have already been deforested. Therefore, the
placement of new households is likely not to result in further loss of forest.
However, even small amounts of fuelwood required for the large number of
households already in the reserve results in increased impacts on forest cover. A

decrease in the number of current households such as through emigration does limit

113

the loss of forest from agricultural activity. However, very large decreases in the
number of households or simultaneous decreases in average household consumption
rates are needed to retain the current forests. These results are similar to estimates
as measured by Liu et al. (1999) who showed that relatively high rates of
emigration were necessary to restore habitat and suggest that most efforts should
focus on reducing fuelwood collection and providing alternative energy sources for
the current households while providing viable means and incentives to encourage
emigration.

In this study, we further examined the landscape-level factors governing the
spatial distribution of human impacts on forests and its implications for giant panda
habitat in light of episodic die-offs of bamboo. The continued removal of low-
elevation forests considered less important than core habitat areas may seriously
restrict current sub-populations of giant pandas within Wolong from necessary
secondary habitat when the next B. fangiana die-off occurs. Simultaneous or near-
simultaneous ﬂowerin gs of B. fangiana and F. robusta will be detrimental to panda
populations. Even with a sufﬁcient interval between ﬂowerings to allow biomass
recovery (15 — 20 years for both species (Taylor and Qin, 1993)), decreasing access
to F. robusta in the future may have similar consequences for the remaining pandas
in Wolong. In 1983, pandas had direct access to signiﬁcant amounts of F. robusta.
Pandas eventually shifted to secondary food sources after the last B. fangiana die-
off. The availability of F. robusta during the mast-seeding of B. fangiana in 1983
likely buffered the studied populations from the mass-Starvations faced by pandas in

the Min Mountains. Landscape models are useful in describing the basic driving

114

factors control events occurring at the landscape level. Abiotic, transportation, and
biophysical variables can strongly control the location and distribution of human
activities. However, by coupling the multitemporal spatial dynamics of bamboo
with human activities, the implications of even limited impacts to critical habitat
can be seen.

The success of the model to predict similar distributions of households and
local land-cover change suggest basic landscape features such as topography and
accessibility can play a large role in Shaping the spatial distribution of human
activities. The limitations of the model point to areas where further study is needed.
Most decisions such as consumption level, propensity to use alternative energy
sources, emigration rates and new household formations are made at the household-
level and are not included in this landscape model. Furthermore, more detailed
information on the biophysical characteristics such as total available biomass,
growth rates, and efﬁciency of conversion of biomass to fuelwood are needed.

Even with these limitations, however, the model does provide insight into
the causes and driving factors leading to past habitat loss, future impacts, and the
spatial arrangement of these impacts. Past studies have identiﬁed the impacts of
broad-scale logging and household fuelwood collection on giant panda habitat
(Schaller et al. 1985. Mackinnon and DeWulf 1989), however, few studies have
noted the combined inﬂuences of these activities. We found that broad-scale
logging is likely providing conditions that exacerbated impacts from households to
low-elevation forests. Regrowth following clearcutting of these forests provided

trees preferred for fuelwood consumption.

115

In addition, lower volumes resulting from logging activity led to larger areas
required to meet household demands. Similar to past studies (Liu et al. 2001), this
study found that access to giant panda habitat increases the likelihood and quantity
of losses of habitat. However this study also points out that access clearly plays a
role in determining the Spatial distribution of household activities. Households
require access to roads to transport cash crops to local cities and the ease of access
to preferred fuelwood also seemingly shapes fuelwood collection decisions. The
distribution of roads had a considerable impact on the past distribution of
households and associated activities. Future development of and providing
additional access to Wolong and other reserves will likely make these areas
vulnerable to similar impacts. Lastly, found as previous studies (Liu et al. 1999)
that extensive changes in household characteristics are needed to maintain habitat.
The model suggests that the number of new households and the average volume of
fuelwood consumed by each household has a strong inﬂuence on the quantity of
future impacts. Decreases on the order of 150% in new households or a complete
cessation of fuelwood collection are needed to preserve habitat at current levels.
Providing viable energy alternatives to the local population will likely considerably
decrease impacts, particularly the fringe habitat adjacent to primary habitat.

Bamboo die-offs will occur in Wolong every 40 — 60 years. These die-offs
reduce total habitat available to pandas until regrowth. We found that even though
human impacts on forests in Wolong are relatively small compared to losses of
habitat from bamboo die-offs, losses of habitat from bamboo are temporary and

have not historically Signiﬁcantly affected the overall panda population. The

116

current extent and spatial distribution of human impacts may, however, have
considerable consequences on panda populations, particularly during these
temporary losses of bamboo. In addition, human reductions in forest cover reduce
the likelihood of bamboo regeneration to previous levels. This model offers insight
into the importance the Spatial arrangement of even limited human impacts can
have on giant panda habitat. Worldwide, human impacts may be underestimated as
the combined effects of human and natural disturbances have not frequently been
considered. Simultaneous examination of natural and human disturbances may be
necessary to assess total habitat area requirements for species persistence.
Examining the spatial distribution of human activities in conjunction with natural
disturbances further highlights the relevance of examining the consequences of

human activities in other areas and the implications for species persistence.

117

Acknowledgements

We thank the Wolong Nature Reserve Administration for their logistic support of
our ﬁeldwork and the assistance from J inyan Huang, Jian Yang, Hemin Zhang, and
Shiqiang Zhou. In addition, we gratefully acknowledge ﬁnancial support from the National
Aeronautics and Space Administration (NASA Earth Science Fellowship Award), the
National Science Foundation (CAREER Award, formerly Presidential Young Investigator
Award and Biocomplexity in the Environment Grant), National Institutes of Health
(National Institute of Child Health and Human Development, R01 HD39789), American
Association for the Advancement of Sciences/The John D. and Catherine T. MacArthur
Foundation, Michigan State University (All-University Research Initiation Grant, Global
Competence Fund, Institute for International Agriculture), the National Natural Science
Foundation of China, the Chinese Academy of Sciences, Ministry of Science and

Technology of China (G2000046807), and China Bridges International (North America).

118

Table 4.1. Landscape characteristics of the measured households in 1997
(Households 1997) compared to model scenarios. Landscape indices include the
perimeter to area ratio (P/A Ratio), centroid connectivity index (Connectivity) and
aggregation index (Aggregation).

 

 

 

 

 

 

Mean
No. of . PIA .. .
Patches Patch Size Ratio Connectrvrty Aggregation
(ha)
H°“Seh°'ds 94.00 40931 1.49 0.046 0.423
1997
All Parameters 110.35 59101 1.50 0.053 0.513
N0 memmy 261.00 24905 1.41 0.015 0.271
Factor
N0 Ab‘°"° 161.90 40229 1.46 0.034 0.426
Factor
No Cost Factor 280.60 23152 1.29 0.009 0.337

 

 

 

 

 

 

 

 

119

Table 4.2. Accuracy of the predicted household locations for the model scenarios
relative to the measured households in 1997 (Households 1997). Accuracy as
predicted household locations occurring at measured household locations (titled 0)
and predicted locations within 1, 2, and 3 pixels (labeled 1, 2, and 3, respectively)
of measured household locations. Uncertainties represent one standard error of the
accuracies of the 20 simulations conducted for each scenario.

 

 

 

 

 

 

Accuracy (%)
0 1 2 3

N0 COS‘ 20.6 :1: 1.3 47.3 a 2.4 57.1 :1: 2.6 63.0 :1: 2.4
Factor

N0 Prox‘m‘ty 21.2 :1; 1.1 54.3 :1: 1.7 70.6 :1: 1.3 79.8 a 1.4
Factor

N0 Abmm 22.4 1 1.4 55.8 a 2.4 71.8 :1: 2.0 81.2 :1: 2.3
Factor

A" 27.4 a 0.7 67.9 5; 1.5 82.5 1 1.9 88.3 :1: 1.9

Parameters

 

 

 

 

 

120

 

Table 4.3. Comparison of the inﬂuence of variations in individual parameters for
each multiplicative factor used within the fuelwood sub-model. Landscape
characteristics of the model results are also given for various levels of initial
volume per hectare for low-elevation forests.

 

 

 

 

 

 

 

 

 

Parameter Number of 1:51:23 P/A Connectivity Aggregation

Patches . Ratio Index Index
Size (ha)

Proximity 0.33 125.2 75.8 1.668 0.719 0.858
Factor 1 192.2 49.2 1.606 0.336 0.829
3 291.5 32.7 1.538 0.170 0.794
Distance 0.8 145.8 65.8 1.630 0.546 0.848
Factor 1 192.2 49.2 1.606 0.336 0.829
1.2 231.6 40.4 1.587 0.277 0.816
30 21 1.4 51.9 1.567 0.365 0.830
Initial 45 192.2 49.2 1.606 0.336 0.829
Volume 60 258.7 33.6 1.540 0.212 0.812
(ml/ha) 75 265.9 30.3 1.502 0.161 0.805
90 246.3 30.5 1.502 0.167 0.802

 

 

 

 

 

 

 

 

 

 

 

121

 

Table 4.4. The inﬂuence of household characteristics on habitat over 65 years
(1965 — 2030) relative to a baseline scenario of 0 new households per year and 0
m3/yr of fuelwood consumed after 1997.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Home F“e’w°°.d Total Habitat Habitat <
Growth Rate Consumptlon
(houses/yr) (m3/yr) (%) 2600 m (%)
24 0 -0.06 -0.18
24 5 -l .34 ~3.79
24 10 -2.61 -7.36
24 15 -3.32 -9.33
24 30 -6.06 - 1 5 .84
0 15 -2.77 -7.74
12 15 -3.21 -8.99
24 15 -3.32 -9.33
36 15 -4.31 -l 1.74
0 0 0.00 0.00
-12 15 0.01 0.03
-24 15 0.05 0.13
12 10 -2.26 ~6.41
36 0 -0.11 -0.45

 

 

 

 

 

 

122

Table 4.5. The inﬂuences of household characteristics on habitat for models run
from 1997 to 2030 based on digital classiﬁcations and measured household
locations.

 

 

 

 

 

 

Fueliwmd 0 10 15 15
(m /yr)
Parameter
Household
(households/yr) O 12 24 '24
Marginal 7669. 9 7662.5 7657.1 7663.9
Total Suitable 41 863.2 40227.8 38881.8 40606.4
Habitat .
(ha) Highly 9659.3 9199.1 8780.3 9309.9
Total 59192.4 57089.4 55319.2 57580.2

 

 

 

 

 

 

 

 

123

Table 4.6. The inﬂuences of household characteristics on habitat below 2600
meters of elevation for model runs from 1997 to 2030 based on digital
classifications and measured household locations.

 

 

 

 

 

 

 

 

 

 

 

 

Fuel3vvood 0 10 1 5 15
(m lyr)
Parameter
Households

(hou seholds/yr) 0 l 2 24 -24

Marginal 72.1 65.0 59.3 66.1
Habitat Suitable 16523.2 14929.4 13636. 7 15296.3
(ha) Highly 5391.4 4958.3 4580.2 5063.9
Total 21986.6 19952.8 18276.1 20426.3

 

 

124

 

 

 

 

 

 

 

 

 

@302 wdSc— ﬁnomm— wan—m— oms: 0852 mdnwv— Oncmc— odvoam .88.

0.3mm 0.23 034% e. _ new #3me N. _ mum wdccm QmSN m. _ wav bswi
85
.85.":

_NNS Gnome 562 w adw _m _dmcw N33 @403 «632 Ramona oBSSm

odmmm od—mm ﬁommm mdvmm m.mo~m mdmmm Qmmmm admmm wébmm _aEwSE

Geo»

RON Rom Rom hmom Son Eon Rom hmom SON $933

scﬂwcﬂd

C3»

53 boom Rom Son hmom 53.. Ken boom 32 toéE
8398 m uoaoEEam

. m»

cm. vm- em 3 N _ Q a o o B23885

38:88:

2 2 m. m. S 2 o o o @8335

 

 

 

 

 

 

 

 

 

 

 

Table 4.7. The combined inﬂuences of bamboo die-offs and household impacts. B.
fangiana die-offs are held constant at 2027 while F. robusta die-offs are varied

between 2007 and 2027.

125

 

Table 4.8. The combined inﬂuences of bamboo die-offs and household impacts on

habitat below 2600 meters of elevation. B. fangiana die-offs are held constant at

2027 while F. robusta die-offs are varied between 2007 and 2027.

 

 

 

 

 

 

 

 

 

mev0 93%. ﬁmmom m._~_ N. ”Same 2o: Wmmho ~6me 5.33 .88.
”com— mhomm 053: wdoom @30— _.2 mm wswom wénmm @5me >233
3.:
233m:
_dnvv _.N .3 ”ﬂow advow 5.3.94 092% @394 macaw mdoac oBSEm
o.m Gm QM Gm m.m Ym NM ~.m med _mEmE—z
Cw...»
bmom Rom snow Row Row Rom hmom Rom Rom :923
SEES». .m
Coo»
Rom boom Rom 58 Row Bow Rom boom «.2 $923
8358 .n— BEEF—ma.
A;
em- 3. vm em 2 N. o o o Exosomaoé
8:53:03
23::
m. 2 2 2 o. o_ o o o coca—oak

 

 

 

 

 

 

 

 

 

 

 

 

126

Table 4.9. The combined inﬂuences of bamboo die-offs and household impacts on
habitat fragmentation.

 

 

 

 

 

 

 

 

Fuelwood
(“ﬂ/yr) 0 0 0 24 24
Household 0 0 0 15 15
P (households/yr)
arameter F I) l
.' ’0. “S 3 N/A 2007 2027 2007 2027
(dlC-OIUCaI')
B: fang'ana N/A 2027 2027 2027 2027
(dle-off year)
Marginal 2.85 2.47 2.49 2.48 2.47
M63“. Suitable 4.16 1.96 1.80 1.89 1.73
Patch 8126
(ha) Highly 2.33 1.88 1.81 1.80 1.72
All 6.88 3.06 2.86 2.93 2.71

 

 

 

 

 

 

 

 

 

127

Figure Legends. (Images in this dissertation are presented in color)

Figure 4.1. A conceptual depiction of previous bamboo dynamics effects on habitat
(solid line) and the combined bamboo and land-use effects on habitat (dashed line)

based on model predictions and projected continued effects.

Figure 4.2. Wolong Nature Reserve lies in the Qionglai Mountains between the

Tibetan plateau and Sichuan basin.

Figure 4.3. A conceptual ﬂow schematic diagram of the model.

Figure 4.4. Figure 4.4a shows the digital elevation model (DEM) and Figure 4.4b

shows the distribution of forest types.

Figure 4.5. Examples of spatial data used for post-1997 simulations. Figure 4.5a is
the digital classiﬁcation of forest cover from 1997 Landsat TM data and Figure 4.5b
shows the classiﬁcation of presence/absence of bamboo throughout the reserve.
Species delineation of bamboo is based on elevational differences and ground

surveys.

Figure 4.6. Comparisons between visual classiﬁcations of satellite data from

1965, 1974, 1987, and 1997 and predicted forest cover due to household activities

of corresponding years.

128

Figure 4.7. Comparisons of the inﬂuence of the three multiplicative factors
contained within the household sub-model. Accuracy of each scenario is shown
relative to the measured households with corresponding predicted households and
measured households shown in black, incorrectly predicted households are shown
in dark gray, actual households where no households were predicted are shown in
white. Figure 4.7a Shows the predicted household distribution in 1997 relative to
the actual distribution. Figure 4.7b is without abiotic preferences, Figure 4.7c

without cost factors, and Figure 4.7d without proximity inﬂuences.

Figure 4.8. Predicted forest cover due to fuelwood collection compared to the
digital classiﬁcation at various starting volumes for low-elevation forests. Figure

4.8a — e are with starting volumes of 30,45, 60, 75, and 90 m3/ha, respectively.

Figure 4.9. Impacts from household activities and bamboo dynamics. Figure 4.8a
shows impacts from household growth of 24 houses per year and an annual
fuelwood consumption of 15 m3/yr. Figure 4.8b Shows the remaining habitat after
a simultaneous die-off of F. robusta and B. fangiana while other bamboo species

remain.
Figure 4.10. The impacts of household activities and bamboo die-offs to habitat

below 2600 meters are shown. Figure 4.10a shows the impacts from households

alone while Figure 4.10b includes loss of habitat from bamboo die-offs.

129

Figure 4.1

 

Bamboo and Land-Use Effects on Habitat

 

 

 

 

 

 

 

 

 

 

 

ﬂ 1
9. , .
< :
6666665666 66605665
x '5 b on e\o\w\r§\e\e\e\e¢¢b¢b9

Year

 

130

 

 

Figure 4.2

 

   

Wolong Nature Reserve

 

 

 

131

Figure 4.3

 

 

Fuelwood
Collection
Bamboo
Distribution

 

 

 

 

 

 

 

 

 

Habitat

 

 

Giant Panda

 

 

 

Household
Creation
Forest
Cover

 

 

 

 

 

 

 

Abiotic Factors
(e.g. Slope,
Elevatlon)

 

 

 

132

Figure 4.4

 

 

 

 

 

v?
.- Sub-Alpine Fir
Deciduous Broadleaf
Ever teen Bl‘oadleaf
Non crest

 

 
 

 

 

133

 

Figure 4.4

 

 

Nonforest

 

 

 

 

B. fangiana
F. robusta
other
-No bamboo

 

 

 

 

 

 

 

134

Figure 4.5

Measured Forest Loss Predicted Forest Loss

 

135

Figure 4.6

 

 

 

 

 

 

 

136

Figure 4.8

 

 

 

 

 

 

 

 

 

 

137

 

 

 

 

 

 

Modeled/Measured

D Forest/Forest

- Nonforest/Forest
Forest/Nonforest
1:1 Nonforest/Nonforest

Figure 4.9

 

 

 

 

 

 

 

 
  

- Highly Suitable
- Suitable

- Matguial

- Unsuitable

 

 

138

 

 

Figure 4.10

 

 

 

- Marginal
- Unsuitable

 

 

139

 

 

 

CHAPTER 5

SUMMARY AND CONCLUSIONS

5.1. Summary

Considerable research has been conducted on giant panda behavior and
habitat requirements. As with all other species, however, viable conservation
options will only come from detailed studies of habitat requirements, human
impacts, and spatio-temporal dynamics of the environment at the time and spatial
scales relevant to long-term conservation. For this dissertation, I examined three of
these issues: the spatial distribution of the primary food source of pandas, bamboo,
the implications of the distribution of bamboo on spatial characteristics of habitat at
the landscape scale, and the combined impacts of human activities and bamboo
spatiotemporal dynamics on giant panda habitat across a 200,000 ha nature reserve.

The lack of detailed information on the spatial distribution of bamboo has
been a problem for panda conservation. Pandas rely on bamboo for over 99% of
their diet, must feed for up to 14 hours a day, and require an average of 12.5 kg of
bamboo each day. Pandas are therefore completely restricted to habitat with easy
access to plentiful supplies of bamboo. Measuring and modeling panda habitat then

requires knowledge of the spatial distribution of understory bamboo.

140

The results from the neural network classiﬁcation are promising. We found
the artiﬁcial neural network improved classiﬁcation accuracy by 11 - 13% over
maximum-likelihood classiﬁcations and resulted in a reserve-wide classiﬁcation
accuracy of 80%. It is difﬁcult to assess the signiﬁcance of the improvements in
accuracy. The increases in accuracy may be in part to the improved classiﬁcation
performance of neural nets found in other classiﬁcation studies. However, a 12%
increase in accuracy of a binary classiﬁcation is relatively large. In addition, we
found evidence that the neural network classiﬁcation was being inﬂuenced by
understory vegetation. Co-occurring understory vegetation was strongly correlated
with the misclassiﬁcations. False positives were almost always associated with the
presence of other grass species. False negatives largely corresponded to signiﬁcant
mid-story cover. Further studies on this approach are needed. Studies on the
inﬂuencing factors and the ability of the neural net to distinguish between variable
vegetation conditions are needed. However, the study does point to a promising
approach to understory vegetation detection from remote sensing data.

The detailed information gained from these analyses improved our ability to
measure the distribution of panda habitat at a landscape scale. More accurate
estimates of the quantity of habitat are now possible. In addition, more detailed
assessments of the landscape characteristics of giant panda habitat in Wolong are
available. Compared to previous estimates of total habitat available to pandas in
Wolong, including bamboo in habitat estimates reduced the total available habitat
by 29 — 56%, patch size by 16 — 48%, and carrying capacity by 41 — 60%. These

results suggest that panda habitat has been overestimated. Not including bamboo

141

into habitat analyses not only overestimates total habitat, but also underestimates
fragmentation. Particularly notable is the increased isolation of larger patches. The
isolation and fragmentation of habitat my lead to increased isolation of and
decreased contact between groups of pandas in the reserve. However, more studies
are needed on the movement patterns and dispersal abilities of pandas. Ideally,
tracking devices capable of recording several positions each day to examine how
pandas are interacting with the environment and long-term monitoring to examine
range and dispersal tendencies of pandas are needed.

Finally, we combined predictions of the spatio-temporal dynamics of the
measured distribution of bamboo with human impacts in the reserve. Commercial
logging, household fuelwood collection, and agricultural activities have seriously
degraded habitat throughout Wolong. Using socioeconomic characteristics of
households and measured characteristics of the spatial distribution of household
activities from 1965 — 1997, we were able to project future human impacts on forest
cover through the year 2030. The immediate impacts from households are
relatively limited compared to past commercial logging. However, when these
impacts are coupled with the temporary loss of bamboo during times of die-off, the
Spatial distribution of household impacts may have considerable impact to panda
populations.

Giant panda populations have been restricted to remnants of habitat from
extensive past land-cover change and are now extremely sensitive to natural die-
offs of bamboo. Continued human activities may have considerable impacts on

giant panda populations in the near future. Using a spatio-temporal model of

142

bamboo dynamics and human activities, we found that local fuelwood collection
and agricultural activity will likely reduce critical habitat relied upon in past
bamboo die-offs. According to the model, a simultaneous die-off of the two
predominant bamboo species would result in up to 52% of the current area of
habitat to be lost. Human impacts would likely only contribute up to an additional
16% loss of habitat. However, these impacts primarily occur in habitat relied upon
by giant pandas during past bamboo die-offs. The spatial distribution, as well as the
total quantity, of human impacts on giant panda habitat must be taken into
consideration. Future policies need to take into consideration where households are
being placed and fuelwood collection is taking place. A ﬁrst step to protecting
panda habitat would be to absolutely restrict new households in areas in these
crucial areas, including secondary habitat needed by pandas in times of losses of
bamboo. Second, priority should be placed on relocating current households in
crucial habitat. Finally, fast growing forest plantations have been placed near
households and there are areas in the reserve where pandas have been extirpated
and which are unlikely to be used by pandas in the near future. Increased reliance
on plantations and areas not important to pandas for fuelwood is critical to decrease
pressure on panda habitat.

The coupled model of the spatiotemporal dynamics of bamboo and human
activities Showed that even limited impacts to critical habitat may have signiﬁcant
implications. Continued removal of forests considered less important as core
habitat areas will seriously restrict current sub-populations of giant pandas within

Wolong. The model predicts fuelwood collection alone will reduce access in many

143

areas of prime panda habitat. Relatively large areas of low-elevation highly suitable
and suitable habitat adjacent to prime habitat are predicted to be lost to household
activities. Any additional impacts, as seen from 1975 to 1997, will likely
considerably add to the impacts from households alone.

The model, however, also suggests other causes and driving factors leading
to past habitat loss and future impacts. Past broad-scale logging likely provided
conditions that exacerbated impacts from households by providing trees preferred
for fuelwood collection. In addition, lower volumes resulting from logging activity
led to larger areas required to meet household demands. Access also clearly plays a
role in determining the spatial distribution of household activities. Households
require access to roads to transport cash crops to local cities and the ease of access
to preferred fuelwood also seemingly Shapes fuelwood collection decisions.
Finally, the model suggests that the average volume of fuelwood consumed by each
household largely controls the quantity of future impacts. Providing viable energy
alternatives to the local residents will likely considerably decrease impacts,
particularly the fringe habitat adjacent to primary habitat.

Landscape models are useful in describing the basic driving factors of
events occurring at the landscape level. Abiotic and biophysical variables can
strongly control the location and distribution of human activities. However,
examining the spatial distribution of human activities in conjunction with natural
disturbances further highlights the relevance of these activities and potential
inﬂuence on the biotic system and permit more accurate estimates of the overall

habitat requirements for species persistence.

144

5.2 Conclusions

These studies point to the need for considerable information on
environmental conditions relevant to a species for more accurate estimates of
quantity, quality, and spatial arrangement of habitat. Combining these more
detailed habitat assessments to dynamic models of human activities allow improved
estimates of human impacts and potential consequences otherwise not available.
From these studies, it is clear that more research is required on remote sensing
techniques to gather relevant biophysical data and methods to incorporate these data
into habitat analyses. Furthermore, studies on the coupled effects of continuing
human appropriation of habitat and natural disturbance regimes would contribute to
the current literature on the minimum area concept, landscape dynamics, population
persistence, and the impacts socioeconomic activities on biodiversity.

Remote sensing has provided a wealth of information not available through
traditional approaches. Remote sensing data have provided global land cover
classiﬁcations, long-term monitoring of biophysical attributes, and even
socioeconomic data. As sensors become more appropriate for measuring
biophysical properties that are relevant to ﬂoral and fauna] species, ecosystem
processes, and human needs, more research is required on methods to extract these
necessary data. For this study it was clear we need detailed information on the
spatial distribution of understory bamboo, however, traditional techniques did not
lend themselves to these analyses. The application of an artiﬁcial neural network

provided increased classiﬁcation accuracy and seemingly relied on understory

145

vegetation. More intensive and controlled studies, however, are needed.
Particularly, controlled ﬁeld studies on understory contributions to the radiance
measured by a distant sensor are needed. Also, ﬁeld studies with well controlled
overstory and understory conditions would contribute to this approach. In addition,
this speciﬁc approach can be applied to the detection of other understory vegetation.
The general approach, however, needs to be further developed in the application of
remote sensing data to ecological studies. More accurate estimates on vegetation
characteristics are needed for quantitative assessments of habitat.

More detailed biophysical data would allow more accurate assessments of
habitat for particular species and biodiversity in general. Current estimates likely
overestimate available habitat. Categorizing habitat based on basic land-cover
information almost deﬁnitely overlooks necessary structural (cover, biomass, etc.)
information, less dominant vegetation species, and sub-canopy species. These
attributes have been Shown repeatedly by local studies to signiﬁcantly inﬂuence .
species presence and behavior. Certainly other variables may contribute to Species
distribution, persistence, or movement at broader scales (such as the landscape
scale), however, ignoring these variables, as shown in this study, can result in
serious misestimates of the quantity and spatial characteristics of habitat.

Viable management options will not be possible without accurate estimates
of habitat. Knowledge of the spatial distribution of habitat is necessary to gauge
and mitigate human impacts. For example, the spatial distribution of bamboo and
location of human activities in Wolong have particular importance to pandas. More

accurate assessments of habitat will have considerable implications for predicting

146

panda population trends, palticularly in light of metapopulation theory. It is now
clear that there is far less habitat available to giant pandas than previously
estimated. The fragmentation and isolation estimated in this study suggests pandas
are likely more vulnerable than previously believed. In addition, more accurate
estimates of habitat are needed to derive the full impact of human activities and
viable mitigation options. We now see that the cumulative human effects to habitat
in Wolong Nature Reserve may have considerable impacts on giant pandas.
Management must take into consideration not only the total quantity of the impacts,
but also the spatial distribution of the activities in developing viable policy
alternatives. This study shows the need for more detailed landscape-scale habitat
information to more accurately monitor habitat and biodiversity. In addition, these
data would allow more accurate estimates of human impacts on species and more

viable management alternatives.

147

BIBLIOGRAPHY

Adams, J. B., D. E. Sabol, V. Kapos, R. A. Filho, D. A. Roberts. M. O. Smith and
A. R. Gillespie (1995). "Classiﬁcation of multispectral images based on
fractions of endmembers: Application to land-cover change in the Brazilian
Amazon." Remote Sensing of the Environment 52: 137—154.

An, L., J. Liu, Z. Ouyang, M. Linderman, S. Zhou and H. Zhang (2001).
"Simulating demographics and socioeconomic process on household level

and their impacts on giant panda habitats." Ecological Modelling 140(1-2):
31-49.

Asner, G. P. and D. B. Lobell (2000). "A biogeophysical approach for automated
SWIR unmixing of soils and vegetation." Remote Sensing of Environment
74: 99-1 12.

Asrar, G., R. B. Myneni and E. T. Kanemasu (1989). Estimation of plant-canopy
attributes from spectral reﬂectance measurements. Theor_'y and Applications
of Optical Remote Sensing. G. Asrar. New York, John Wiley & Sons.

Atkinson, P. M. and A. R. L. Tatnall (1997). "Neural networks in remote sensing."
International Joumgtl of Remote Sensing 18(4): 699-709.

Baker, W. L. and Y. Cai (1992). "The r. le-programs for multiscale analysis of
landscape structure using the grass geographical information—system."
Landscape Ecology 4: 291-302.

Baret, F. and G. Guyot (1991). "Potentials and limits of vegetation indices for LAI
and APAR assessment." Remote Sensing of Environment 35: 161-173.

Bausch, W. C. (1993). "Soil background effects on reﬂectance-based crop
coeffecients for corn." Remote Sensing of the Environment 46: 213-222.

Beaudoin, A., T. Le Toan, S. Goze, E. Nezry, A. Lopes, E. Mougin, C. C. Hsu, H.
C. Han, J. A. Kong and R. T. Shin (1994). "Retrieval of forest biomass from
SAR data. " International Journal of Remote Sensing 15(14): 2777-2796.

Berjak, S. G. and J. W. Heame (2002). "An improved cellular automaton model for
simulating ﬁre in a spatially heterogeneous Savanna system." Ecological
Modelling 148(2): 133-151.

Carpenter, G. A., S. G0pal, S. Macomber, S. Martens and C. E. Woodcock (1999).
"A neural network method for mixture estimation for vegetation mapping."

Remote Sensing of Environment 70: 138-152.

Ceballos, G. and P. R. Ehrlich (2002). "Mammal population losses and the
extinction crisis." Science 296(5569): 904-907.

148

China's Ministry of Forestry and WF (1989). Conservation and Management Plan
for Giant Pandas and Their Habitat. Beijing, China.

Collins, R. L. and G. W. Barrett (1997). "Effects of habitat fragmentation on
meadow vole (Microtus pennsylvanicus) population dynamics in experiment
landscape patches." Landscape Ecology 12(2): 63-76.

Cross, A. M., J. J. Settle, N. A. Drake and R. T. M. Paivinen (1991). "Subpixel
measurement of tropical forest cover using AVHRR data." International
Journal of Remote Sensing 12(5): 1119-1129.

De Wulf, R. R., R. E. Goossens, J. R. MacKinnon and S. C. Wu (1988). "Remote
Sensing for Wildlife Management: Giant Panda Habitat Mapping from
LANDSAT MSS Images." Gwrto International 1: 41-50.

Diffendork, J. E., M. S. Gaines and R. D. Holt (1995). "Habitat fragmentation and
movements of three small mammals (sigmodon, microtus, and
peromyscus)." Ecology 76(3): 827-839.

Dueser, R. D. and H. H. Shugart (1978). "Microhabitats in a forest-ﬂoor small
mammal fauna." Ecology 59: 89-98.

Dunning, J. B., D. J. Stewart, B. J. Danielson, B. R. Noon, T. L. Root, R. H.
Lamberson and E. E. Stevens (1995). "Spatially explicit population models:
Current forms and future uses." Ecological Applications 5(1): 3-1 1.

Erdas (1999). Imagine. Atlanta, GA.

Ernest, K. A. (1989). "Insect herbivory on a tropical understory tree: effects of leaf
age and habitat." Biotropica 21: 194-199.

Estades, C. F. and S. A. Temple (1999). "Deciduous-forest bird communities in a
fragmented landscape dominated by exotic pine plantations." Ecological
Applications 9(573-585).

Fahrig, L. (2001). "How much habitat is enough?" Biological Conservation 100(1):
65-74.

Fall, A. and J. Fall (2001). "A domain-speciﬁc language for models of landscape
dynamics." Ecologicgl Modelling 141(1-3): 1-18.

Fall, A., D. G. Morgan and D. Daust (2001). "A framework and software tool to
support collaborative landscape analysis: Fitting square pegs into square
holes." Transactions in GIS 5(1): 67-86.

Foody, G. M. (1996). "Relating the land-cover composition of mixed pixels to

artiﬁcial neural network classiﬁcation output." Photogrammetric
Engineering & Remote Sensing 62(5): 491-499.

149

Foody, G. M. and M. K. Arora (1997). "An evaluation of some factors affecting the
accuracy of classiﬁcation by an artiﬁcial neural network." International
Journal of Remote Sensing 18(4): 799-810.

Forman, R. T. and M. Godron (1986). Landscape Ecology. New York, Chichester,
Brisbane, Toronto, Singapore, John Wiley & Sons.

Forman, R. T. T. (1995). Land mosaics : The ecology of Lindscapes and regions.
Cambridge; New York, Cambridge University Press.

Ghitter, G. S., R. J. Hall and S. E. Franklin (1995). "Variability of Landsat
Thematic Mapper data in boreal deciduous and mixed-wood stands with

conifer understory.” International Journal of Remote Sensing 16(16): 2989-
3002.

Giant Panda Expedition (1974). "A survey of the giant panda (Ailuropoda
melanoleuca) in the Wolong Natural Reserve, Pingwu, Northern Szechuan,
China." Acta Zoologica Sinica 20: 162-173.

Gilabert, M. A., F. J. Garcia-Haro and J. Melia (2000). "A mixture modeling
approach to estimate vegetation parameters for heterogeneous canopies in
remote sensing." Remote Sensing of Environment 72: 328-345.

Guyot, G. and D. G. J. Riom (1989). "Factors affecting the spectral response of
forest canopies: A review." Geocarto International 3: 3-18.

Haines-Young, R., D. R. Green and S. Cousins, Eds. (1993). Lgldscape Ecology
and Geographical Information Svstemp. Bristol, PA, Taylor & Francis.

He, H. S., B. E. DeZonia and D. J. Mladenoff (2000). "An aggregation index (A1) to
quantify Spatial patterns on landscapes." Landscape Ecology 15: 591-601.

Hess, G. R. (1996). "Linking extinction to connectivity and habitat destruction in
metapopulation models." American Naturalist 148(1): 226-236.

 

Huete, A. R., R. D. Jackson and D. F. Post (1985). "Spectral response of a plant

canopy with different soil backgrounds." Remote Sensing of Environment
17: 37-53.

James, F. C. (1971). "Ordinations of habitat relationships among breeding birds."
The Wilson Bulletin 83: 215-236.

Janzen, D. H. (1976). "Why bamboos wait so long to ﬂower." Annual Review of
Ecology Eﬂl Systematics 7: 347-391.

Jensen, J. R. (1996). Introductory Digital Image Processing - A Remote Sensing
Perspective. Upper Saddle River, New Jersey, Prentice Hall.

150

Johnson, K. G., G. B. Schaller and J. Hu (1988). "Responses of giant pandas to a
bamboo die-off." National Geographic Resem 4(2): 161-177.

Keeley, J. E. and W. J. Bond (1999). "Mast ﬂowering and semelparity in bamboos:
the bamboo ﬁre cycle hypothesis." The AmericapNaturm 154(3): 383-
391.

Law, B. E. and R. H. Waring (1994). "Remote sensing of leaf area index and
radiation intercepted by understory vegetation." Ecological Applications
4(2): 272-279.

Lefsky, M. A., G. G. Parker, T. A. Spies, D. Harding, W. B. Cohen and S. A. Acker
(1999). "Lidar remote sensing of the canopy structure and biophysical

properties of Douglas-ﬁr western hemlock forests." Remote Sensing of
Environment 70: 339-361.

Leopold, A. (1933). Game Management. Madison, Wisconsin, The University of
Wisconsin Press.

Lillesand, T. M. and R. W. Kiefer (1994). Remote Sensing and Image
Interpretation. New York, John Wiley & Sons, Inc.

Lindzey, F. G. and E. C. Meslow ( 1977). "Home range and habitat use by black
bears in Southwestern Washington." Journal of Wildlife Mﬂrgement 41:
413-425.

Liu, J. (1993). "ECOLECON: An ECOLogical-ECONomic model for species
conservation in complex forest landscapes." Ecological Modelling 70: 63-
87.

Liu, J ., M. Linderman, Z. Ouyang, L. An, J. Yang and H. Zhang (2001).
"Ecological degradation in protected areas: The case of Wolong Nature
Reserve for giant pandas." Science 292(5514): 98-101.

 

Liu, J ., Z. Ouyang, Y. Tan, J. Yang and H. Zhang (1999). "Changes in human
population structure: Implications for biodiversity conservation."
Population and Environment 21(1): 45-58.

Liu, J ., Z. Ouyang, Z. Yang, W. Taylor, R. Groop, Y. Tan and H. Zhang (1999). "A
framework for evaluating the effects of human factors on wildlife habitat:
The case of giant pandas." Conservation Biology 13(6): 1360-1370.

Lu, 2., J. S. Martenson, S. Mainka, H. Shi-Qiang, Z. Zhihe, G. Li, W. Pan, X. Mao,
S. J. O'Brien, Johnson W.E., M. Menotti-Raymond and N. Yuhki (2001).
"Patterns of genetic diversity in remaining giant panda populations."
Conservation Biology 15(6): 1596-1607.

151

Luckman, A., J. Baker, T. Mora Kuplich, C. da Costa Freitas Yanasse and A. C.
Frery (1997). "A study of the relationship between radar backscatter and
regenerating tropical forest biomass for spacebome SAR instruments."
Remote Sensing of Environment 60: 1-13.

MacArthur, R. H. and J. W. MacArthur (1961). "On bird species diversity."
Ecology 42(3): 594-598.

McCaffery, K. R., J. Tranetzki and J. Piechura, Jr. (1974). "Summer foods of deer
in Northern Wisconsin." Journal of Wildlife Management 38: 215 - 219.

Morain, S. A. (1986). "Surveying China's agricultural resources: Patterns and
progress from space." Geocarto Intemational 1: 15-24.

Morrison, M. L., B. G. Marcot and R. W. Mannan, Eds. (1998). Wildlife-Habitat
Relationships: Concepts & Applications. Madison, Wisconsin, The
University of Wisconsin Press.

Naugle, D. E., R. R. Johnson, M. E. Estey and K. F. Higgins (2001). "A landscape
approach to conserving wetland bird habitat in the prarie pothole region of
eastern South Dakota." Wetlands 21(1): 1-17.

Nemani, R., L. Pierce, S. Running and L. Band (1993). "Forest ecosystem processes
at the watershed scale: Sensitivity to remotely-sensed Leaf Area Index
estimates." International Journal of Remote Sensing 14(13): 2519-2534.

Nemani, R., L. Pierce, S. Running and L. Band (1993). "Forest ecosystem processes
at the watershed scale: sensitivity to remotely-sensed Leaf Area Index
estimates." Intemzaional Journal of Remote Sensing 14(13): 2519-2534.

Ouyang. Z., Y. Tang and H. Zhang (1996). "Biodiversity spatial pattern analysis in
Wolong." China Man and Biosphere 3.

Pearson, S. M., M. G. Turner, R. H. Gardner and R. V. O'Neill (1996). An
organism-based perspective of habitat fragmentation. Biodiversity in

Managed Landscapes: Theory and Practice. R. Szaro, Johnston, DW.
London, Oxford University Press: 77 - 95.

Perry, G. L. W. and N. J. Enright (2002). "Spatial modelling of landscape
composition and pattern in a maquis-forest complex, Mont Do, New
Caledonia." Ecological Modelling 152(2-3): 279-302.

Porwall, M. C. and P. S. Roy (1991). "Attempted understorey characterization using

aerial photography in Kanha National Park, Madhya Pradesh, India."
Environmental Conservation 18(1): 45-50.

152

Powell, R. A., J. w. Zimmerman and p. E. Seaman (1997). Ecology and Belmior

of North Amergan Black Bears: Home Ranges, Rabitat and Social
Organization. London, Chapman & Hall.

Qi. J., H. Kerr, M. S. Moran, M. Weltz, A. R. Huete, S. Sorooshian and R. Bryant
(2000). "Leaf area index estimates using remotely sensed data and BRDF
models in a semiarid region." Remote Sensing of Environment 73: 18-30.

Qi , J ., M. Weltz, A. R. Huete, S. Sorooshian, R. Bryant, Y. H. Kerr and M. S.
Moran (2000). "Leaf area index estimates using remotely sensed data and
BRDF models in a semiarid region." Remote Sensing of Environment 73(1):
18-30.

Ranson, K. J ., C. S. T. Daughtry and L. L. Biehl (1986). "Sun angle, view angle,
and background effects on Spectral response of simulated balsam ﬁr
canopies." Photogrammetric Engineering and Remote Sensing 52(5): 649-
658.

Ranson, K. J ., D. C. S. T. and B. L. L. (1986). "Sun angle, view angle, and

background effects on spectral response of simulated balsam ﬁr canopies."
Photogrammectric Engineering and Remote Sensing 52(5): 649-658.

Reid, D. G., J. Hu, S. Dong, W. Wang and Y. Huang (1989). "Giant panda

Ailuropoda melanoleuca behaviour and carrying capacity following a
bamboo die-off." Biological Conservation 49: 85-104.

Reid, D. G., A. H. Taylor, H. Jinchu and Q. Zinsheng (1991). "Environmental

inﬂuences on bamboo Bashania Fangiana growth and implications for giant
panda conservation." Journal of Applied Ecology 28: 855-868.

Riitters, K. H., R. V. O’Neill, C. T. Hunsaker, J. D. Wickham, D. H. Yankee, S. P.
Timmins, K. B. Jones and B. L. Jackson (1995). "A factor analysis of
landscape pattern and structure metrics." Landscape Ecology 10: 23-39.

Romme, W. H., M. G. Turner, L. L. Wallace and J. S. Walker (1995). "Aspen, elk,
and ﬁre in Northern Yellowstone National Park." Ecology 76(7): 2097-
2106. '

Roughgarden, J ., S. W. Runing and P. A. Matson (1991). "What does remote
sensing do for ecology?" Ecology 72(6): 1918-1922.

Sala, O. E., F. S. Chapin 111, J. J. Arrnesto, E. Berlow, J. Bloomﬁeld, R. Dirzo, E.
Huber-Sanwald, L. F. Huenneke, R. B. Jackson, A. Kinzig, R. Leemans, D.
M. Lodge, H. A. Mooney, M. Oesterheld, N. L. Poff, M. T. Sykes, B. H.

Walker, M. Walker and D. H. Wall (2000). "Global biodiversity scenarios
for the year 2100." Science 287(5459): 1770-1774.

 

153

Schaller, G. B. (1987). "Bamboo shortage not only cause of panda decline." Nature
327: 562.

Schaller, G. B., J. Hu, W. Pan and J. Zhu (1985). The Giant Pandas of Wolong.
Chicago and London, The University of Chicago Press.

Spanner, M. A., L. L. Pierce, D. L. Peterson and S. W. Running (1994). "Remote
sensing of temperate coniferous forest leaf area index: The inﬂuence of

canopy closure, understory vegetation and background reﬂectance."
Ecological Applications 4(2): 258-271.

Stenback, J. M. and R. G. Congalton (1990). "Using Thematic Mapper imagery to

examine forest understory." Photogrammetric Engineering and Remote
Sensing 56(9): 1285-1290.

Stéphenne, N. and E. F. Lambin (2001). "A dynamic simulation model of land-use
changes in Sudano-sahelian countries of Africa (SALU)." gp'culture,
Ecosystems and Environment 85(1-3): 145-161.

Takekawa, J. E. and S. R. Beissinger (1989). "Cyclic drought, dispersal, and the
conservation of the Snail Kite in Florida: Lessons in critical habitat."
Conservation Biology 3(3): 302 - 311.

Taylor, A. H. and Z. Qin (1993). "Bamboo regeneration after ﬂowering in the

Wolong Giant Panda Reserve, China." Biological Conservation 63(23]-
234).

Taylor, A. H. and Z. Qin (1997). The dynamics of temperate bamboo forests and
panda conservation in China. The Bamboos. G. P. Chapman. San Diego,
Harcourt Brace & Company: 189-21 1.

Taylor, A. H., D. G. Reid, Z. Qin and J. Hu (1991). "Spatial patterns and
environmental associates of bamboo (Bashaniafangiana Yi) after mass-

ﬂowering in southwestern China." Bulletin of the Torrey Botanical Club
1 18(3): 247-254.

Taylor, A. H. and Q. Zisheng (1993). "Bamboo regeneration after ﬂowering in the
Wolong Giant Panda Reserve, China." Biological Conservation 63: 231-
234.

The MathWorks Inc. (1999). Matlab. Natick, MA.

Treuhaft, R. N. and P. R. Siqueira (2000). "Vertical structure of vegetated land
surfaces from interferometric and polarimetric radar." Radio Science 35(1):

141-177.

Turner, M. G. (1989). "Landscape ecology: The effect of pattern on process."
Annual Review of Ecological Svstems 20: 171-197.

154

Turner, M. G. (1990). "Spatial and temporal analysis of landscape patterns."
Landscape Ecology 4(1): 21-30.

Turner, M. G., Y. Wu, L. L. Wallace, W. H. Romme and A. Brenkert (1994).
"Simulating winter interactions among ungulates, vegetation, and ﬁre in
northern Yellowstone Park." Ecological Applications 4(3): 472-496.

Vitousek, P. M., H. A. Mooney, J. Lubchenco and J. M. Melillo (1997). "Human
domination of Earth's ecosystems." Science 277(5325): 494-499.

 

Yang, Y. and C. Li, Eds. (1992). Sichuan Forests. Yunnan, China Forestry Press.
Zhang, H., D. Li, R. Wei, C. Tang and J. Tu (1997). "Advances in conservation and

studies on reproductivity of giant pandas in Wolong." Sichuan Journal of
Zoology 16: 31-33.

155

 

 

        
  

 

M IIIIIIIIIIIIIIIIIIIIIIIIIIIIII

1111111111111111111111111111111 11111.11

93 0233 820 _