.

u. .. I

2a.“.
2

_. .L
ELL
. a

. L . ..
. $3312
3.3:. 31::
.333. Law...
, . L .4
. . . L

.19

. 3.2.3.‘

. , 5.95:1:
.. L.
rain
5.53.21
LU.

aganuahw
37....)
a thug-34.. 1:. ,
. .7 v.r>¢.4..::.:b_
, uyz

”Mmfdé
when
.
”1 m4?
1.. [021.1
”Mm

. .i
1. 537151.)...
3 - imrfl.

 

.v v i

.9.un

l «a .
2 ‘

 

.LLL

41531;...

LLLE:
3L

 

”LL.
._ A
bk 7....

 

This is to certify that the
dissertation entitled

ESTIMATION OF TROPICAL FOREST BIOPHYSICAL
ATTRIBUTES WITH SYNERGISTIC USE OF OPTICAL AND
MICROWAVE REMOTE SENSING TECHNIQUES

presented by

CUIZHEN WANG

has been accepted towards fulfillment
of the requirements for the

PhD. degree in Geography

 

 

o‘-

  

I

 

/ MabTProfessor’s Signature

5/;“/c»§/

I

Date

MSU is an Affirmative Action/Equal Opportunity Institution

 

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

4 AL.“

- IcI

 

6/01 c:/ClFIC/DateDue.p65-p.15

 

ESTIIVIATION OF TROPICAL FOREST BIOPHYSICAL ATTRIBUTES
WITH SYNERGISTIC USE OF OPTICAL AND MICROWAVE REMOTE
SENSING TECHNIQUES

By

Cuizhen Wang

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

Depaitment of Geography

2004

 

 

ABSTRACT

ESTIMATION OF TROPICAL FOREST BIOPHYSICAL ATTRIBUTES
WITH SYNERGISTIC USE OF OPTICAL AND MICROWAVE REMOTE

SENSING TECHNIQUES
By

Cuizhen Wang

Accurate estimates of tropical forest biophysical attributes provide quantitative
information in the assessment of human disturbances and the carbon sequestration in
global climate change studies. The research of my dissertation is thus to develop new
algorithms to estimate tropical forest biophysical parameters (forest fractional cover, leaf
area index, structures, and aboveground biomass) using synergy of optical and radar
remote sensing imagery and radiative transfer models. The case study is in Mae Chaem
Watershed, ChiangMai, Thailand. Ground data were collected during the field trips

sponsored by research projects.

A linear unmixing model in the vegetation index (MSA V1) domain was built to estimate
forest fractional cover with Landsat ETM+ image. The forest fractional cover map was
validated using both ground measurements and high-resolution IKONOS images. The
estimated fractional cover correlated with the ground-measured fractional cover
(R2=O.76) at 32 study sites and correlated with the IKONOS—calculated fractional cover
(R2=0.70) at 400 randomly selected locations. The leaf area index was estimated using a

modified Gaussian regression model with forest fractional cover results. The model was

examined with a x2 goodness-of-fit test. The correlation coefficient between the modeled

and ground-measured leaf area index values is 0.90.

A microwave/optical synergistic radiative transfer mode] was built to simulate the radar
scattering from the forest components. The leaf scattering and its attenuation to the
woody components (branches, trunks) were quantified with the leaf area index derived
from optical remote sensing data. The forest structural parameters, such as tree height and
stand density, were estimated through model inversion with JERS-l SAR and VNIR data.
The total root—mean—square error (RMSE) of tree height estimation was 4.6 meter and that
of stand density estimation was 300 trees/ha. The stand density estimation did not work
in tropical evergreen forests because of its saturation at around 500 trees/ha. In
accordance with ground measurements, tree height is negatively correlated with stand
density in the study area. The model inversion becomes questionable at mountainous
areas with high relief and steep slopes. The aboveground forest biomass is also calculated
with allometric equations and the modeled forest structural parameters. The total RMSE

is 88 ton/ha.

The methods developed in this study could be applied to estimate forest biophysical
attributes at regional or global scales. With optical remote sensing imagery, only forest
fractional cover and leaf area index could be estimated. When both optical and SAR data
are acquired, the forest structural parameters and aboveground biomass can be estimated.
These results could provide quantitative information in full carbon accounting in global

climate change studies.

 

To
my husband Lanwu Zhao and my baby girl Jessie with all my love

And to
my family in china for their infinite support and encouragement

iv

 

ACKNOWLEDGMENTS

First of all, I would like to thank Dr. J iaguo Qi, my faculty advisor, for his constant help
and encouragements during my graduate studies. Without his academic guidance, it is
impossible for me to complete this dissertation and move forward to my academic career.
I would also like to thank my committee members, Dr. David Skole, Dr. David Lusch,
Dr. Richard Kobe, and Dr. William Salas for their support and help even since the

dissertation proposal was emerged.

I am also grateful to all the faculty and staff members in Department of Geography,
Michigan State University, who help me to possess the knowledge as a geographer as
quick as I can. Special thanks to Dr. Randall Schaetzl and Ms. Sharon Ruggles who gave

me the strongest support when I was struggling in US as a foreign graduate student.

Also, I would like to thank Dr. Mark Cochrine, Walter Chomentowski, Jay Samek,
Eraldo Matricardi, Oscar Castaneda and Cameron Williams in the Center of Global
Change and Earth Observations (CGCEO) for image processing, technical support and
helpful comments on my research. Many thanks go to Ms. Diane Cox and Deana Haner,
very wonderful ladies for their assistance in my study and research. I would also express
my appreciation to Narumon (Nok) Wiangwang, Pari (Perry) Varnakovida, Chetphong
Butthep in CGCEO, and Dr. Charlie Navanujraha, Ms. Siam Lawavirojwong and Woody
in Mahidol University, Bangkok, Thailand. Their warrnhearted help and assistance in

field works in Thailand will never be forgotten.

Finally, my deepest thank to my husband Lanwu Zhao who gave me endless support and
love during my graduate study. Also to my little sweetheart Jessie who was born when I
began to write my dissertation. Her arrival motivated me to be a successful mother and

scientist. Thanks to my father, sisters, and brother in China for their love and many years’
support. This dissertation is also a very special memorial to my loved mother. I am sure

she will receive this message in heaven.

vi

TABLE OF CONTENTS

Chapter 1
Introduction ......................................................................................................................
1.1 Literature Review ...............................................................................................
1.2 Research Objectives ...........................................................................................
1 .3 References ..........................................................................................................
Chapter 2
Study Area and Field Measurements ............................................................................
2.1 Study Area ........................................................................................................
2.2 Field Measurements ..........................................................................................
2.2.1 First field trip (August 10—18, 2001) .........................................................
2.2.2 Second field trip (January 20-27, 2002) ...................................................
2.2.3 Field data processing .................................................................................
2.2.3.1 Forest fractional cover ..........................................................................
2.2.3.2 Stand density .........................................................................................
2.2.3.3 Aboveground biomass ..........................................................................
2.2.4 Ground data analysis ................................................................................. .
2.3 References ......................................................................................................... 1
Chapter 3
Estimation of Tropical Forest Fractional Cover with Landsat ETM+ and IKONOS
Imagery ............................................................................................................................ .
3. 1 Introduction ....................................................................................................... .
3.2 Remotely Sensed Data ...................................................................................... .
3.2.1 Landsat ETM+ image ............................................................................... .
3.2.2 High—resolution IKONOS Images ............................................................. I
3.3 Methods ............................................................................................................. I
3.3.1 A linear unmixing model .......................................................................... I
3.3.2 Optimal vegetation index (V1) .................................................................. I
3.4 Canopy Fractional Cover Analysis ................................................................... I
3.5 Validation .......................................................................................................... I
3.5.1 Validation with ground measurements ..................................................... ;
3.5.2 Validation with l-m IKONOS data .......................................................... ;
3.5.3 Seasonal adjustment of fiactional cover ................................................... .
3.6 Conclusion and Discussion ............................................................................... I
3.7 References ......................................................................................................... I
Chapter 4
Estimation of Leaf Area Index with Fractional Cover Data in Tropical Forests ..... ‘
4. 1 Introduction ....................................................................................................... ‘
4.2 Experimental Design and Field Measurements ................................................ I
4.2.1 LAI-2000 and fisheye photographs .......................................................... I
4.2.2 LA] ground measurements ........................................................................ I
4.2.3 LAI ~ forest fractional cover relationship ................................................. I
4.3 Model Development and Results ...................................................................... I
4.3.1 A modified Gaussian regression model .................................................... I
4.3.2 LA] estimation ........................................................................................... I

vii

4.4 Validation .......................................................................................................... E

4.5 Conclusion and Discussion ............................................................................... E
4.6 Reference .......................................................................................................... S
Chapter 5
A Microwave/Optical Synergistic Canopy Scattering Model and its Inversion to
Estimate Forest Structure ............................................................................................ 1t
5. 1 Introduction ..................................................................................................... 1 (
5 .2 Model Development ........................................................................................ 1 (
5.2.1 Modified Karam-IEM model .................................................................. 1(
5.2.1.1 Soil surface scattering ......................................................................... 1(
5.2.1.2 Leaf scattering ..................................................................................... 1(
5.2.1.3 Branch scattering ................................................................................ 1(
5.2.1.4 Trtmk scattering .................................................................................. 1(
5.2.1.5 Leaf-soil interaction ............................................................................ 1(
5.2.1.6 Branch-soil interaction ........................................................................ 1(
5.2.1.7 Trunk-soil interaction .......................................................................... 1(
5.2.2 Linkage to optical remotely sensed variables ......................................... 11
5.2.3 PDF functions of forest components ....................................................... 1]
5.3 Model Simulation and Validation............................. ...................................... 11
5.3.1 Remotely sensed data .............................................................................. 11
5 .3 .2 Model simulation .................................................................................... l 1
5.3.2.1 Contribution of leaves ......................................................................... 11
5.3.2.2 Contribution of branches ..................................................................... 1]
5.3.2.3 Contribution of trunks ......................................................................... 11
5.3.3 Model validation ..................................................................................... 12
5.4 Model Inversion and Forest Parameters Estimation ....................................... 12
5.4.1 Model inversion ...................................................................................... 12
5.4.2 Forest structural parameters by model inversion .................................... 12
5.4.3 Uncertainty analysis ................................................................................ 12
5.5 Conclusions and Discussion ........................................................................... 12
5.6 References ....................................................................................................... 13
Chapter 6
Aboveground Woody Biomass Estimation with Microwave and Optical Remotely
Sensed Data .................................................................................................................... 14
6.1 Introduction ..................................................................................................... 14
6.2 Biomass Estimation with a Simple Regression Method ................................. 14
6.3 Biomass Estimation with Compensation of leaf Attenuation ......................... 15
6.4 Biomass Estimation with Synergistic Model and Allometric Equations ........ 15
6.5 Conclusions and Discussions .......................................................................... 15
6.6 References ....................................................................................................... 16
Chapter 7
Conclusions and Future Envisions .............................................................................. 17
7.1 Concluding Remarks ....................................................................................... 17
7.2 Challenges ....................................................................................................... 17
7.3 Potential Applications for Other Studies ........................................................ 17

Viiii

Table 2-1
Table 2-2

Table 4-1
Table 5-1
Table 5-2
Table 5-3

Table 5-4
Table 5—5

Table 6—1

LIST OF TABLES

Average values of ground measurements in different forest types. ................ 27
Average values of ground measurements in different forest types when the
outlier was deleted. ......................................................................................... 27
Study sites in northern Michigan. ................................................................... 95
System parameters of J ERS—l SAR and NVIR sensors ................................ 132
Input parameters for model simulation. ........................................................ 132
Input parameters for model validation (in addition to ground measurements).
......................................................................................... 133
A set of forest structural parameters for each forest type in the model. ....... 133
Average and standard deviation of modeled tree height, stand density, and the
error of model inversion for each forest types. ............................................. 134
Coefficients of 0'0 ~biomass logarithmic curve fitting and their statistic tests.
......................................................................................... 164
ix

 

LIST OF FIGURES

Figure 2-1 The study area (Mae Chaem Watershed). ....................................................... 2
Figure 2—2 The DEM data (with hill shade effect) in the study area. ............................... 2
Figure 2-3 Vertical distribution of forest types in the study area. .................................... 2
Figure 2-4 Forest type map in the study area. .................................................................. 2
Figure 2-5 Study sites and ground control points (GCPs) in the study area ..................... 3
Figure 2-6 fisheye picture and its transformation in angular coordinates. ....................... 3
Figure 2-7 Correlation scatterplot matrix of ground—measured data. ............................... 3
Figure 2-8 Normalized average ground measurements in different forest types .............. 3
Figure 2-9 Modified normalized average ground measurements in different forest types
(the outlier of dry evergreen site was removed). ............................................ 3

Figure 3-1 ETM+(band4+3+2, 02/02/2000) (a) and IKONS images (band4+3+2,
02/27/2000 and 10/09/2002) (b) in Mae Chaem Watershed. The dots in (b)

represent the study sites in two field trips ....................................................... 7
Figure 3-2 DEM data with hillshade effect (a) and topographic correction with Rahman’.
BRDF model (b). ............................................................................................ 7
Figure 3-3 Dynamic ranges of the six vegetation indices calculated with simulated
spectral data in SAIL model. .......................................................................... 7
Figure 3-4 MSAVI image in Mae Chaem Watershed. ..................................................... 7
Figure 3-5 Forest fractional cover map in Mae Chaem Watershed. ................................. 7
Figure 3—6 Comparison of ground-measured and ETM+ estimated fc: scatterplot (a) and
clustered column (b). ...................................................................................... 7

Figure 3-8 Scatter plots of ground-measured and the H<ONOS estimated fc in 6 sites (a),
and the ETM+ and IKONOS estimated fc in 400 randomly selected points (b)

75
Figure 3-9 IKONOS and ETM+ subset images and their fc maps. .................................. 7
Figure 3-10 The ground-measured, IKONOS and ETM+ estimated fc scatterplot in dr:
and wet seasons. .............................................................................................. 7
Figure 3-11 The curve fitting model to adjust fc in dry season to wet season. ............ 7
Figure 3-12 Forest fractional cover map adjusted from dry season to wet season. ...... 7
Figure 4-1 Relationships between LAI (5-ring and 4-ring) with forest fractional cover.
The fractional cover is in the range of [0,1] .................................................... 9
Figure 4-2 Comparison of LAI measurements from LAI-2000 and fisheye Photos. The
1:1 line is also drawn in the plot. .................................................................... 9
Figure 4-3 LAI ~ forest fractional cover relationships measured with LAI-2000 and
fisheye photos in northern Michigan. ............................................................. 9
Figure 4-4 A modified Gaussian curve fitting (a) and its residual plot (b). ..................... 9
Figure 4-5 LAI distribution in the study area in dry (a) and wet (b) seasons. .................. 9
Figure 4-6 Comparison of modeled and adjusted measured LAI over the study sites in th«
watershed (second trip). .................................................................................. 9
Figure 4-7 Correlation between modeled and adjusted measured LAI over the study sites
with LAI <6. ................................................................................................... 9
Figure 5-1 Geometry of a forest canopy. ........................................................................ 13
Figure 5-2 PDF functions of leaves, branches, and trunks. ............................................ 13

Figure 5-3 JERS-l SAR (a) and VNIR data (NIR+Red+Green) (b) in the study area... 13

Figure 5-4 DEM data (a) and geometrically corrected JERS-l SAR image (b). ........... 13
Figure 5-5 Local incidence angle image (a) and the topographically corrected image (b).
The layovers and shadows are the white areas in (a). ................................... 13

Figure 5—6 Modeled backscattering of different forest components: leaf size (a), LAI (b),
branch radius (0), branch density (d), DBH (e), stand density (0, and tree

height(g) ............................................................................... 13‘
Figure 5-7 Backscattering simulation based on ground-measured forest parameters: LAI
(a), tree height (b), trunk height (0), DBH (d), and stand density (e). .......... 13

Figure 5-8 Comparison of J ERS-l SAR observed and modeled backscattering
coefficients: scatterplot of backscattering coefficients (a) and residuals (b).

The 1:1 line is also drawn in (a) .................................................................... 14
Figure 5-9 Scatterplot of stem height and DBH to tree height measured at study sites. 14
Figure 5-10 Modeled tree height distribution in the watershed .................................. 14
Figure 5-11 Modeled stand density distribution in the watershed. ............................. 14
Figure 5 -12 Error distribution of model inversion. .................................................... 14
Figure 5-13 Scatterplots of modeled and measured tree height (a) and stand density (b:

at all study sites. ............................................................................................ 14
Figure 5-14 Comparison of modeled and measured tree height (a) and stand density (b

in different forest types. ................................................................................ 14-
Figure 6-1 Scatterplot and curve fitting of 035R5_1~biomass at all study sites. .............. 16

Figure 6-2 Biomass distribution estimated with a simple 035RS_1~biomass regression

model ............................................................................................................. 16
Figure 6-3 The LA] distribution derived from J ERS-l VNIR data in the study area. 16
Figure 6-4 Relationships between ground-measured biomass and modeled backscattering

coefficients of forests (a) and their components (b). .................................... 16
Figure 6-5 Logarithmic curve fitting of SAR observed and modeled backscattering

coefficients with biomass .............................................................................. 16
Figure 6-6 Leaf scattering (a) and its attenuation to the woody forests (b) .................... 16
Figure 6-7 Woody scattering in the study area. .............................................................. 16‘
Figure 6-8 Biomass distribution with woody forest scattering in a regression model. .. 16‘.
Figure 6-9 Modeled DBH distribution in the watershed. ............................................... 17I
Figure 6-10 Forest biomass distribution from model inversion in the watershed. ..... 17I
Figure 6-11 Scatterplot of modeled and measured biomass at study sites. ................ 17
Figure 6-12 Average values of measured and modeled biomass (without outlier) and

the standard deviation of the modeled biomass in each forest type. ............. 17

xi

Chapter 1

Introduction

Forests are a major natural resource in the Earth biosphere and control a wide range of
environmental processes. Forest ecosystems play a vital role in energy, gas and water
exchanges between the land surface and the atmosphere. Tropical forests hold the largest
proportion of the Earth’s biodiversity, and account for approximately 40% of terrestrial
biomass (Foody et al. 1997). They play an important role in water, gas, and energy cycle
due to the geosphere, biosphere, and atmosphere interactions. Traditional forest mapping
with remote sensing imagery provide important information for the assessment of
deforestation, reforestation, and afforestation in tropical forests as a result of both human
activities and natural processes. But this forest/non-forest classification is a more

qualitative technique which introduces high uncertainty in full carbon accounting.

Therefore, we need continuous fields of biophysical distribution in tropical forests
(DeFries et al. 1999, 2000). Quantitative measure of forest area, density, biomass and
other biophysical attributes of tropical forests is necessary to fully assess the intensity and
extent of forest degradation, fragmentation, and recovery from human and natural
disturbances. These biophysical attributes are also important input parameters of the
carbon and climate models in global climate change studies. (Skole and Tucker 1993;
Myneni et al. 1998). This study thus aims to develop new algorithms to retrieve these
biophysical attributes in tropical forests with both optical and microwave remote sensing

techniques.

1.1 Literature Review

The quantification of tropical forest variables is not an easy task. Due to labor and cost
constraints and the physical difficulties to access the study areas, field in-situ
measurements of biophysical attributes are very limited in dense tropical forests. The
limited ground measurements by destructive methods, litter collections, and sawmill
census data may not be representative when applied to large areas of tropical forests. As
an alternative, remote sensing techniques provide timely and spatially continuous
measurements over large areas and, therefore, have the potential to offer a timely and cost

effective estimate of forest biophysical attributes (Uhl 1987).

Optical remote sensing records the energy of sunlight reflected from the Earth surface in
the wavelength range of ~0.4um (visible) to ~14.0um (thermal infra—red), comparable to
the size of atmospheric components such as aerosol and water vapor. These components
absorb and scatter sunlight and therefore reduce the quality of optical imagery. Tropical
forests are often covered by clouds which precludes information extraction from optical
remote sensing imagery. Aside from atmosphere effects, the spectral responses of optical
images are strongly dependent on the green leaf characteristics of the forest canopy. The
reflected energy from a forest canopy is determined by the physical properties of the
foliage such as chlorophyll concentration, internal leaf structure, the lignin component of
the cell walls, and the water content. Since the spectral responses of branches and trunks
are similar to those of bare soil surfaces, reflectance is not very sensitive to these woody

components in forests.

Microwave remote sensors use a different range of the electromagnetic spectrum from
optical remote sensors. The spectral wavelengths of the Synthetic Aperture Radar (SAR)
most commonly used in forest studies range from 3cm (X-band) to 30cm (P-band). These
wavelengths are much longer than the sizes of the atmospheric constituents so the
atmosphere is transparent to microwave remote sensing. The capability of cloud
penetration is a great advantage of microwave remote sensing in tropical forest studies.
At certain system parameters, the intensity and phase of a radar backscatter signal is
mostly determined by dielectric constants and forest structures. The radar backscatter is
also dependent on the wavelength of the signal. Signals with shorter wavelength (e. g. C—
or X-band) mainly scatter from crown layer. Longer wavelength signals (e. g. L- and P-
bands) can penetrate the crown layer and reach the soil surface underneath. Therefore, the
backscattered SAR signals contain the information of woody components like branches

and trunks in forests.

The biophysical attributes of green canopies have been a primary interest of the optical
remote sensing community. Various approaches have been developed to estimate these
attributes like forest fractional cover (Price 1992, Wittich 1997, Gutman and Ignatov
1998, Qi et al. 2000) and leaf area index (Peterson et a1. 1987, Pinty et a1. 1990, Qi et al.
1995). The high temporal (daily through yearly) and coarse spatial (250m to 1km)
resolution products of these biophysical attributes on a global scale were also developed

with the MODIS and MISR data aboard the Earth Observation System satellites

(Townshend et al. 1999; DeFries et al. 2000).

 

 

The attempt of using microwave remote sensing to retrieve biophysical attributes began
in early 19603. In May of 1960, the Terrain Handbook by Professor Bill Peake was
published which denoted a major transformation of radar remote sensing from qualitative
image interpretation to quantitative information retrieval (Ulaby 1998). During the past
decade, the application of SAR imagery in monitoring ecosystem processes has grown
significantly (Kasischke et al. 1997). Various types of SAR data, including single-
frequency, single-polarization, multi-frequency, multi-polarization, multi-temporal,
polarimetric and interferometric data, have been acquired and subsequently applied to
map forest areas and estimate biophysical attributes such as biomass (Dobson et al. 1992,
Le Toan et al. 1992, Luckman et al. 1998), forest structure (McDonald and Ulaby 1993,
Imhoff 1995, Sun and Ranson 1998), and soil moisture (Oh et al. 1992, Dubois et al.
1995, Shi et al. 1997). However, applications of microwave imagery are limited because
of the intrinsic speckle effects in SAR images and the topographic influence in
mountainous areas (Henderson and Lewis, 1998). Salas et al. (2002a, b) concluded that
although quantitative estimates of biomass are not stable due to intrinsic texture, system
noise, and environmental effects, multi-temporal analysis significantly improves biomass

estimates in secondary tropical forests.

Among the studies mentioned above, most of the forest biophysical estimation was
carried out using either optical or microwave remote sensing. For example, leaf area
index and forest fractional cover are estimated with optical remote sensing imagery,

while aboveground biomass, structure, and soil moisture are estimated with microwave

remote sensing imagery. Due to the different wave-target interaction mechanisms, each
type of imagery has its own advantages and disadvantages. Optical remote sensing is very
sensitive to vegetation chlorophyll, wetness, and leaf structures, but is less sensitive to
branches or trunks. Microwave remote sensing is sensitive to biomass and woody

structures, but highly affected by the attenuation from leaf canopies.

A synergistic use of optical and microwave remote sensing could fill the application gaps
so as to improve the estimation of forest biophysical attributes. A simple way of
optical/SAR synergism is image fusion that can be performed at the pixel level in which
physical measurements of sensors are merged (N ezry et al. 1993); at the feature level in
which characteristics of the original images are merged (Solberg et al. 1994); and at the
interpretation level (also called decision level) in which the labeled features are merged
(Rignot et al. 1996). At higher level of synergism, the canopy reflectance and scattering
models could be investigated and combined with remotely sensed data to simulate surface
conditions. However, it is not an easy task because of the very different mechanisms of

signal-surface interactions.

Only limited effort has been made on the synergy of optical and microwave remote
sensing. Moran et al. (2002) used Landsat TM imagery to eliminate vegetated and/or
moist fields when investigating the sensitivity of ERS-2 data to surface roughness of bare
soil. Rignot et a1. (1996) combined the classifications from TM and L-band SIR-C
images to achieve a higher accuracy of land cover and deforestation/regrowth mapping.

Salas (2002a) used both TM and J ERS-l data to map land use and land cover change, and

 

to assess deforestation and secondary vegetation in tropical forests. Moghaddam and
Dungan (2000) attempted to estimate foliage biomass by fusing SAR and TM data at the
pixel level. All of these studies are based on optical and SAR image fusions and their
results are mostly qualitative. These methods are far too simple to be applied in

quantification of forest biophysical attributes.

1.2 Research Objectives

This study aims to develop new algorithms to estimate specific biophysical attributes
using a microwave/optical radiative transfer model and both optical and SAR remotely

sensed data. More specifically, the following detailed objectives will be addressed:

1) Develop improved approaches to estimate forest canopy fractional cover using
optical remote sensing at medium to coarse resolutions and validate these
results with high resolution IKONOS imagery and ground data.

2) Relate leaf area index distribution to the forest fractional cover product.

3) Develop a microwave/optical synergistic canopy scattering model to simulate
scattering from all forest components and to estimate forest structural
parameters by model inversion using JERS-l SAR/VNIR data.

4) Map aboveground woody biomass distribution with the microwave/optical

synergistic model and JERS-l SAR/VNIR data.

The dissertation is outlined as follows:

 

Chapter 2 describes the study area and ground data measured during the field trips. The
biophysical attributes measured at the study sites include forest fractional cover, tree
height, stem height, diameter at breast height (DBH), tree age, stand density and stand

biomass.

Chapter 3 focuses on the estimation of forest fractional cover with optical remote sensing
data. A linear unmixing model in vegetation index domain was built and the forest

fractional cover distribution was mapped in the study area. The fractional cover map was
validated with both ground measurements and high resolution (1-m) IKONOS data. This

chapter fulfills Objective 1.

Chapter 4 describes the estimation of leaf area index with the results of forest fractional
cover retrieved in Chapter 2. A third field trip was carried out to simultaneously measure
leaf area index with a LAI-2000 instrument, and forest fractional cover with fisheye
photos using GLA software. A modified Gaussian regression model was built to relate
leaf area index to forest fractional cover. The model was examined with a x2 goodness-
of-fit test. Then the leaf area index distribution was mapped with the forest fractional

cover product in the study area. This chapter fulfills Objective 2.

Chapter 5 focuses on the estimation of forest structural parameters in the study area. A
microwave/optical synergistic canopy scattering model was built to simulate the
scattering from all of the forest components. The leaf scattering and its attenuation to

other components were quantified with the leaf area index retrieved in Chapter 3, so that

only the woody structures, such as tree height, stand density, and diameter at breast
height (DBH) ‘were the unknown variables in the model. With J ERS—l SAR and VNIR
data, the model was inversed and the forest structures estimated. This chapter fulfills

Objective 3.

Chapter 6 describes three different methods in the estimation of aboveground woody
biomass in the study area. First, a simple regression model was built to estimate biomass
from JERS—l SAR backscattering coefficients. Second, another regression model was
built to estimate biomass from woody scattering when the leaf scattering and its
attenuation to woody components (branches and trunks) was quantified with JERS-l
VNIR data. Finally, according to allometric equations, the woody biomass was also
calculated using the forest structure parameters estimated in Chapter 4. This chapter

fulfils Objective 4.

Chapter 7 contains the conclusions and challenges of this study. Its potential applications
in land use and land cover change and carbon and global climate change studies are also

addressed.

1.3 References

DeFries, R. S., Townshend, J. R. G. and Hansen, M. C. (1999). Continuous fields of
vegetation characteristics at the global scale at 1 -km resolution. Journal of
Geophysical Research, 104, 16,911—16,923.

DeFries, R. S., Hansen, M. C., Townshend, J. R. G., Janetos, A. C. and Lovelands, T. T.
(2000). A new global l-km dataset of percentage tree cover derived from remote
sensing. Global change biology: 6, 247-254.

Dobson, M. C., Ulaby, F. T., Le Toan, T., Beaudoin, A. and Kasischke, E. S. (1992).
Dependence of radar backscatter on coniferous forest biomass. IEEE Transactions
on Geoscience and Remote Sensing: 30, 412-415.

Dubois, P. C., Van Zyl, J ., and Engrnan, E. T. (1995), Measuring soil moisture with
imaging radar. IEEE Transactions on Geoscience and Remote Sensing: 33(4):
915-926.

Foody, G. M., Lucas, R. M., Curran, P. C., and Honzak, M. (1997). Mapping tropical
forest fractional cover from coarse resolution remote sensing imagery. Plant
Ecology, Vol.13 l , 143-154.

Gutman G. and Ignatov, A. (1998). The derivation of the green vegetation fraction from
NOAA/AVHRR data for use in numerical weather prediction models.
lntemational Joumal of Remote Sensing: 19(8), 1533—1543.

Henderson, F. M. and Lewis, A. J. (1998). Principles and applications of imaging
RADAR, Manual of remote sensing, vol.2, third edition. John Wiley &Sons, New
York.

Imhoff, M. L. (1995). A theoretical analysis of the effect of forest structure on Synthetic
Aperture Radar backscatter and the remote sensing of biomass. IEEE
Transactions on Geoscience and Remote Sensing: 33(2), 341-352.

Kasischke, E. S., Melak, J. M. and Dobson, C. (1997). The use of imaging radars for
ecological applications — a review. Remote Sensing of Environment: 59, 141—156.

Le Toan, T., Beaudoin, A. and Guyon, D. ( 1992). Relating forest biomass to SAR data.
IEEE Transgtions on Geoscience and Remote Sensing: 30, 403-411.

Luckman, A., Baker, J ., Honzak, M. and Lucas, R. (1998). Tropical forest biomass
density estimation using JERS-l SAR: seasonal variation, confidence limits, and
application to image mosaics. Remote Sensing of Environment: 63, 126-139.

McDonald, KC. and Ulaby, F. T. (1993). Radiative transfer modeling of discontinuous
tree canopies at microwave frequencies. International Journal of Remote Sensing:
14(11), 2097-2128.

Moghaddam, M., and Dungan, J. (2000). Fusion of SAR and TM data for Quantitative
estimation of forest variables over an extended range of validity. Proceedings of
IEEE International Geoscience and Remote Sensing Symposium, 24-28, July,
Honolulu, Hawaii.

Moran, M. S., Hyrner, D. C., Qi, J. and Kerr, Y. (2002). Comparison of ERS-2 SAR and
Landsat TM imagery for monitoring agricultural crop and soil conditions. Remote
Sensing of Environment: 79, 243-252.

Myneni, R. B., Tucker, C. J ., Asrar, G. and Keeling, C. D. (1998). Interannual variations
in satellite-sensed vegetation index data from 1981 to 1991. Journal of
Geophysical Research: 103(D6), 6145-6160.

Nezry, E., Mougin, E., Lopes, A. and Gastellu—Etchegorry, J. P. (1993). Tropical
vegetation mapping with combined visible and SAR spacebome data.
International Journal of remote sensing: 14 (11), 2165-2184.

Oh, Y., Sarabandi, K., and Ulaby, F. T. (1992), An empirical model and an inversion
technique for radar scattering from bare soil surfaces. IEEE Transactions on
Geoscience and Remote Sensing: 30(2):370-381.

Peterson, D. L., Spanner, M. A., Running, S. W. and Teuber, K. B. (1987). Relationship
of thematic mapper simulator data to leaf area index of temperate coniferous
forest. Remote Sensing of Environment: 22, 323-341.

Pinty, B., Verstraete, M. M. and Dickinson, R. E. (1990). A physical model of the
bidirectional reflectance of vegetation canopies, 2: Inversion and validation.
Journalof Geophysical Research: 95(D8), 11,767-11,775.

Price, J .C. (1992). Estimating vegetation amount form visible and near infrared
reflectances. Remote Sensing of Environment: 41, 29-34.

Qi, J ., Cabot, F., Moran, M. S. and Dedieu, G. (1995). Biophysical parameter estimations
using multidirectional spectral measurements. Remote Sensing of Environment:
54, 71-83.

Qi, J., Marsett, R. C., Moran, M. S., Goodrich, D. C., Heilman, P., Kerr, Y. H., Dedieu,
G., Chehbouni, A. and Zhang, X. X. (2000). Spatial and temporal dynamics of
vegetation in the San Pedro River basin area. Agricultural and Forest
Meteorology: 105, 55-68.

Rignot, E., Salas, W. A. and Skole, D. L. (1996). Mapping deforestation and secondary
growth in Rondonia, Brazil using imaging radar and Thematic Mapper data.
Remote Sensing of Environment: 59, 167-179.

Salas, W.A., Ducey M. J ., Eignot, E. and Skole, D. (2002a). Assessment of JERS-l SAR
for monitoring secondary vegetation in Amazonia: 1. Spatial and temporal

10

variability in backscatter across a chrono-sequence of secondary vegetation stands
in Rondonia. International Journal of Remote Sensing: 23 (7), 1357-1379.

Salas, W.A., Ducey M. J ., Eignot, E. and Skole, D. (2002b). Assessment of JERS—l SAR
for monitoring secondary vegetation in Amazonia: II. Spatial, temporal, and
radiometric considerations for operational monitoring. International Journal of
Remote Sensing: 23(7), 1381-1399.

Shi, J. C., Wang, J ., Hsu, A. Y., O’Neil, P.E., Engman, E. T. (1997), Estimation of bare
surface soil moisture and roughness parameters using L-band SAR image data.
IEEE Transactions on Geoscience and Remote Sensing: 35:1254-1265.

Skole, D., and Tucker, C. J. (1993). Tropical deforestation and habitat fragmentation in
the Amazon: satellite data from 1978 to 1988. Science: 260, 1905-1910.

 

Solberg, A. H., Jain, A. K. and Taxt, T. (1994). Multisource classification of remotely
sensed data: fusion of Landsat TM and SAR images. IEEE Transactions on
Geoscience and Remote sensing: 32, 768-777.

Sun, G. and Ranson, K. J. (1998). Radar modeling of forest spatial patterns. lntemational
Joumil of Remote Sensing: l9 (9), 1769-1791.

Townshend, J. R. G. (1999). MODIS enhanced land cover and land cover change
product: algorithm theoretical basis documents (ATBD), version 2.0.

Uhl, C. (1987). Factors controlling succession following slash-and-bum agriculture in
Amazonia. Journal of Ecology: 75, 377-407.

Ulaby, F. T. (1998). SAR biophysical retrievals: lessons learned and challenges to
overcome. In Proceedings of Retrieval of Bio- and geophysical parameters from
SAR data for land applications, ESTEC, NL, October.

Wittich, K. P. (1997). Some simple relationships between land-surface emissivity,
greenness and the plant cover fiaction for use in satellite remote sensing.
lntemational Journal of Biometeorology: 41, 58-64.

11

Chapter 2

Study Area and Field Measurements

2.1 Study Area

The study area is in the topical forests in north Thailand, located at 98°00’ - 99°00’E and
18°00’ - 19°00’N. Chiang Mai, the second largest city in Thailand, is located in the lower
flatlands in the east of the watershed. The study area consists of four sub-watersheds:
Mae Chaem, Mae Wang, Mae Samoeng, and Mae Klang (Figure 2-1), covering an
approximate area of 6,692 sq. km. It is called the greater Mae Chaem Watershed in this
study. The watershed encompasses a wide range of topography, climate, vegetation types
and density, and human disturbance. Rich historical field survey data, such as land use /
land cover (LULC), soil texture, streams, and elevation, have been made available

through past research activities.

The topography varies greatly within the study area. The altitude ranges from 250 to
2,500 meters above the sea level. Mount Doi Inthanon (25 65.33m) at the center of the
study area is the highest point in Thailand. The digital elevation data (DEM at 30—m
resolution) covering the whole watershed (Figure 2-2) was acquired during the past
research projects in cooperation with Dr. Charlie Navanuj raha and Mr. Siam
Lawavirojwong in Mahidol University, Bangkok, Thailand. The low-elevation flat area in
the southeast is connected to the city of Chiang Mai, and the one in the middle is Mae
Chaem town. With limited urban settlements, the focus study area is generally above 500

meters in elevation.

12

 

The climate in the study area is tropical monsoon that consists of southwest (SW) and
northeast (NE) monsoons. The SW monsoon starts in mid-May, ends in mid-November,
and is characterized by moist winds from the Andaman Sea as the main source of rainfall
during this period. The SW monsoon season is also called the wet season in the study
area. During the wet season, there are two peaks of rainfall periods: J une-August and
September—October (Dr. Charlie Navanujraha, Mahidol University, personal
communication). The NE monsoon starts in mid-November, ends in mid-February, and is
characterized by cold dry winds from Siberia. The NE monsoon season is also called the
cold season in the study area. Both the cold season and the period between NE and SW

monsoons (March-June) compose the dry season as little rainfall occurs.

The general land cover types in the study area include forest, agriculture, grass
land/savanna, open land/bare soil, urban/settlement, and water bodies. Tropical forests
cover 80% of the area (Thailand LUCC case study, 1997). As outlined in Figure 2-3,
tropical forests can be categorized into five types based on the elevation: dry dipterocarps
(lower and higher), mixed deciduous, Pine transition, Tropical dry (lower) evergreen, and
tropical moist (upper) evergreen. Moist evergreen forest is also called cloud forest
because it is often covered by clouds throughout the year. Except for the Teak plantation
around the Ecotourism Forest Station, Royal Department of Forestry, Thailand, in the
southwest of the study area, large-size plantations were not observed. Aside from these
five major forests, the evergreen gallery is composed of rain evergreen forest. Although

at the same elevation with dry dipterocarps (Figure 2-3), rain evergreen forest is confined

l3

to narrow damp valleys mixed with dense bamboo. The spatial distribution of the
evergreen gallery is very limited in the study area and, therefore, was not surveyed. The
buffer zone in Figure 2-3 is the area of human settlement such as villages, open land, and
agricultural fields at lower elevations. The pictures of different forest types are also

shown in Figure 2-3.

A forest type map was made by visual interpretation of a pan-sharpened ETM+ image
acquired on February 2, 2000 (Figure 2-4). It was examined with ground surveys,
historical land cover maps, and high-resolution IKONOS imagery. Only dry dipterocarps,
mixed deciduous, and tropical evergreen forests were categorized in the forest type map.
Pine transition is a narrow zone distributed between the mixed deciduous and tropical
evergreen forests and was not mapped in Figure 2-4. The spectral signature of dry
evergreen is very similar with that of moist evergreen so that they cannot be separated

from ETM+ imagery.

Dry dipterocarps are deciduous forests generally located at low elevations. They can also
be divided into lower and higher dry dipterocarps that are subtly different in senescent
and leaf-off time. Common species of dry dipterocarps primarily include Rokfar, Okfar,
Dang, Makarrn, Teak, and Theng (personal communication with foresters at local areas).
In areas close to the human settlements, forests suffer frequently from short-term clear
cutting and burning for agriculture. After intense cultivation for several years, these
agriculture fields are often abandoned. As a consequence, dry dipterocarps re-grow in

these areas. Most of the dry dipterocarp forests in the Mae Chaem Watershed are

14

secondary regrowth forests disturbed by fire and logging. The stand density and

fractional cover of dry dipterocarps are low.

Mixed deciduous forests occur at higher elevations. The species of both deciduous and
evergreen forests primarily include Plung, Mako, Tueng, and Sompee (personal
communication with foresters at local areas). Mixed deciduous forests also suffer from
logging disturbance. However, due to the topographical difficulty to access these

elevations, the degree of disturbance is lower than that of dry dipterocarp forests.

Pine transition is a narrow zone distributed at an approximate elevation of 600-700m. The
occurrence of pine trees is an indicator that the forest distribution shifts from mixed
deciduous to tropical evergreen forests. The species are a mixture of Pine, Ko, Sompee,
and Sarapee (personal communication with foresters at local areas). Pine transition zone
is narrow and discontinuous, and its spectral signature is very similar to that of tropical
evergreen forests. It is difficult to identify pine transition from evergreen forests in ETM+

images. There is no natural pine forest in the Mae Chaem Watershed.

Tropical evergreen forests occur at higher elevations. They can be further divided into
tropical dry evergreen and moist evergreen. Tropical dry evergreen forests are distributed
just above the pine transition zone. The primary species include Ko, Taklan, Meng,
Sarapee, and Yaung (personal communication with foresters at local areas). Tropical
moist evergreen forests are found on the very top of mountains where the forests are

often covered by cloud and the microclimate makes it moist year round. Sor and K0 are

15

the primary species of tropical moist evergreen forests in the study area. Due to the
extreme difficulty to access these areas, tropical evergreen forests are generally dense

natural forests, almost undisturbed by human activities.

2.2 Field Measurements

2.2.1 First field trip (August 10-18, 2001)

The first field trip was made in the wet season of the study area. This field trip was a
partially follow-up program of the Southeast Asia Regional Research and Information
Network (SEARRIN) science meeting & workshop held in August 2001 in Chiang Mai,
Thailand. The trip included one day of training, three days of field survey, and five days
of field measurements. There were three ground measurements on the training day and
twelve measurements on the following days. The study sites were distributed from
Chiang Mai to Mae Chaem town along the Doi Inthanon 1009 Royal Road. These study
sites covered the following forest types: dry dipterocarps, mixed deciduous, pine
transition, and moist evergreen forests. No dry evergreen forests were measured during

this trip.

On the training day, scientists from the countries in Southeast Asia, who are the
participants of the SEARRIAN workshop, made a field tour from Chiang Mai to Mount
Doi Inthanon. Dr. David Skole, Dr. J iaguo Qi, and Mr. Jay Samek of the Center of Global
Change and Earth Observations (CGCEO) at Michigan State University, which is the
organizer of the SEARRIAN workshop, demonstrated the technique of using a digital

camera with a fisheye lens to measure forest fractional cover. Three sites of dry

16

 

 

dipterocarps, transitional forest, and upper tropical evergreen forest (moist evergreen)

were visited and fisheye pictures were taken on that day.

In the next three days, a field survey was made guided by Dr. Charlie Navanujraha. Two
transects were surveyed. One extended from Chiang Mai to the Ecotourism Forest Station
of the Royal Department of Forestry, Thailand (RDFT) located in the southwest of the
study area. The other extended from the Ecotourism Forest Station to the north of the
study area. During the field survey, forest types and their distributions were identified
and ground control points (GCPs) were selected for geometric image processing. Since it
was at the peak of raining season, only paved roads were accessible and there were
landslides all over the study area. The lengths of the two survey transects were thus
limited. The study sites were also selected during the field survey for intense

measurements later.

Based on the experiences during the field survey, detailed ground measurements at
selected sites were made in the following five days as a joint effort of Ms. Nok
Wiangwang, Mr. Perry Namakovida and myself (all associated with CGCEO, Michigan
State University), and forester Pluem from the Ecotourism Forest Station (RDFT). The
measured study sites were all located not far from the paved roads because of
accessibility issues. At each study site, one transect (150m long) perpendicular to the
paved road was made. To reduce the possibility of human disturbance, the first point was

at least 100m off the road.

17

At each point, a fixed-radius plot method was used to measure tree biophysical
parameters. Centered at the point, a circular area with a radius of 5 meters was
established and the tree height, stem height, and diameter at breast height (DBH) of each
tree in the circle were measured. The measurements were limited to the trees higher than
2 meters and DBH larger than 10 centimeters. The fixed—radius plot method is very time—

consuming, but accurate for stand density and biomass measurements.

2.2.2 Second field trip (January 20-27, 2002)

To collect more ground data and examine the temporal change in the study area, a second
field trip was made in the dry season in January 2002 by Mr. Siam Lawavirojwong,
Wood from Mahidol University and myself. The leaves of broadleaf forests, such as dry
dipterocarps and some mixed deciduous forests, had turned reddish/brownish and were
partially off. The leaves of the evergreen forests did not change much. Since there was
little precipitation in the dry season, and the grasses/shrubs underneath the forests were
senescent, more sites in the study area were accessible even from the dirt roads or trials in

the study area.

During this trip, twelve study sites from the first trip were re-visited and twenty new
study sites all over the watershed were measured. At the study sites where the biophysical
attributes had been measured during the first trip, only fisheye pictures were taken. At
each of the new sites, a 300m transect perpendicular to the road or trial was made
consisting of 10 points. The first point was at least 100m off the road to reduce the effect

of human disturbance. At each point, one fisheye picture was taken and a point-quadrant

18

method was used to measure stand density. Centered at this point, four quadrants were
divided along and perpendicular to the transect direction. The nearest tree in each
quadrant was chosen and the tree height, stem height, DBH, and its distance to the center
were measured. Based on the forester’s expertise, the tree age and species were also
recorded. Again, only those trees higher than 2 meters with a DBH larger than 10
centimeters were measured. The point-quadrant method was more efficient than the
fixed-radius method used in the first trip. However, the accuracy is lower than the fixed-
radius method because only four trees are used to represent the forest conditions at each

point. It may introduce some bias to the ground measured biophysical attributes.

During the two field trips, a total of 32 study sites were measured. Fifteen measurements
(12 with complete measurements, 3 with fisheye pictures only) were made in the wet
season and thirty-two measurements (20 with complete measurements, and 12 with
fisheye pictures only) were made in the dry season. These study sites were distributed

across the watershed (Figure 2-5).

At each study site, the GPS reading at the nearest open area was recorded (in UTM
coordinates) and the offset to the first point of measurement was estimated. The exact
locations of study sites were adjusted based on the GPS reading and its offset. In
addition, during the field surveys in both trips, GPS readings of approximately 60 ground
control points (GCPs) were collected at distinct locations, such as the curve of a road, the

middle of a bridge, and the comer of a reservoir. These GCP points were also distributed

l9

across the watershed (Figure 2—5) and were used for the geometric correction of the

remote sensing imagery.

2.2.3 Field data processing

During both field trips, tree species, age, tree height, stem height (the height between
ground and first branch), and DBH (diameter at breast height) were measured and fisheye
pictures were taken at the selected plots at each study site. The canopy height is the
difference of tree height and stem height. The canopy fractional cover, stand density and

stand biomass of each study site were also calculated from the measured data.

2.2.3.1 Forest fractional cover

Forest fi'actional cover (fc) is the percentage of tree canopy in unit area on the ground. At
each point along the transect of the study site, forest fractional cover was calculated from
the fisheye pictures taken Skyward from the forest floor with a digital camera and 180°
hemispherical (fisheye) lens. The pictures were circular images that recorded the size,

shape, and location of gaps in the forest overstory (Figure 2-6a).

When the fisheye pictures were taken, the camera was held higher than the
photographer’s head to avoid the non-vegetation effect on the ground from people and
understory vegetation or rocks. It was also kept horizontal with a two-dimensional level.
Topographic shading is an inevitable effect in mountainous areas where incident solar
radiation is influenced by not only surface orientation, but also a reduced view of the sky

hemisphere. In this study, most of the study sites were selected in the relatively flat or

20

shallow-slope areas to reduce the topographic shading effect. For the few sites with steep
slopes, only zenith angles of 0°-80° were selected so that the areas affected by

topography were blocked out.

The hemispherical digital pictures were analyzed and fc calculated using the Gap Light
Analyzer (GLA) software (Frazer et al. 1999). The pixel positions in circular images
were transformed into angular coordinates (10° interval in both azimuth and zenith
directions) (Figure 2—6b). The pixel intensities were classified into sky and non—sky
classes. The gap fraction was the percentage of sky brightness throughout the whole
image. The forest fractional cover was equal to (100.0- gap fraction) in the range of 0-
100%. The fc value at each study site was the statistical average of total points in the

transect.

2.2.3.2 Stand density

The fixed-radius plot method was used during the first trip. At each study site, the plot

density at one point, S, , was the total number of trees divided by the area of the plot (r =

5m). The stand density at the study site, S, was the statistical average of the five plot

densities along the transect:

S. = # of trees

2
' 727’

is.
i=1

S...

(2-1)

n

where n is the number of points along each transect (n=5 in the first trip), and r is the

radius of the plot.

21

The point-quadrant method was used during the second trip. At each study site during

this trip, the plot density at one point, S I. , and the stand density S were calculated as:

4
Si: 2 7 2 2
d1+dg+d3 +d4

n

 

(2-2)

S

 

where n is the number of points in each transect (n = 10 for most of the study sites in the

second trip), and d], d 2 , d 3 , d 4 are the distances of the nearest tree to the origin in the

quadrants of upper right, lower right, upper left, and lower left, respectively.

2.2.3.3 Aboveground biomass

The aboveground biomass, also called biomass density, is the total biomass of trees with
diameter 2 100m, including leaves, twigs, branches, bole, and bark (Brown et al., 1989).
For this study, the aboveground biomass included only the woody biomass. The leaves

are represented by the leaf area index and are not considered in the biomass calculation.

The stand biomass could be deduced from the total volume of the trees. When the trading

volume timber (i.e., trunk including wood and bark) is considered, the trunk volume of

one tree, V, , can be expressed in a first approximation as (Le Toan et al., 1992):

2
K=T.fl.[_D_i:£) -H (2-3)

22

where H is the tree height, and T is the shape factor. T = 0.45 was applied in tropical

forests (Le Toan et al. 1992). Similarly, the stand volume at each study site (VS) was

calculated using:

V =T-7z-[P—l2ili] -H-S (2-4)

 

S

where DBH and Hare the statistical means of DBH and tree height. S is the stand
density as defined in section 2.2.3.2. Defining the mean dry wood density (including

bark) as p, the stand biomass at each study site (B, ) was then calculated by:

B, = pIV (2-5)

5

In tropical forests, the dry wood. density is around 610 kg/m3 (Luckman et al., 1997).

2.2.4 Ground data analysis

The ground based geophysical and biophysical parameters at the 32 study sites are
analyzed in this section. The histograms of each parameter and the relationships among
them are shown in the correlation scatterplot matrix (Figure 2-7). The 95% confidence

ellipses are also drawn in the scatterplots.

For tree age, tree height, DBH, canopy height and biomass, there is a positive correlation
with elevation. There is no correlation between tree density and elevation. The
biophysical parameters have high correlations with tree age. Tree height, DBH, canopy
height, and biomass are highly positively correlated with tree age. Tree density is also
correlated with tree age, but the trend is negative, indicating that young trees in secondary

forests have higher density, while old trees in primary forests have lower density. The

23

tree height, canopy height, and DBH parameters are strongly positively correlated with
each other. Tree density has very weak correlations with the other biophysical parameters

except tree age as discussed above.

As shown in the histograms in Figure 2-7, none of the biophysical parameters at the study
sites are normally distributed. Except canopy height and tree age, all other parameters are
left-skewed, indicating more young trees and fewer old trees at the study sites. This bias
is especially obvious for stand biomass calculated from the measured biophysical

parameters using allometric equations.

Among the 32 study sites with complete measurements, there was 1 teak plantation, 8 dry
dipterocarps, 8 mixed deciduous, 1 rain evergreen (evergreen gallery), 5 pine transition, 6
dry evergreen, and 3 moist evergreen forest types. The teaks are also deciduous species
and belong to the dry dipterocarp forest type. The rain evergreen forest belongs to mixed
deciduous forests. The average of the biophysical parameters of the different forest types
is listed in Table 2-1. To make it comparable, the values in Table 2-1 were normalized to
[0,1] based on the maximum and minimum values for each parameter. This comparison is

shown in Figure 2-8.

Except tree density, the biophysical parameters values of dry dipterocarps, mostly young
secondary forests, were lowest. Moist evergreen forests are the primary and mostly
undisturbed forests and have the highest values for all biophysical parameters. The mixed

deciduous, pine transition and dry evergreen forests have biophysical values in between.

24

Among the 6 dry evergreen study sites, only one site (82-12) was young regrowth forest.
The other sites were degraded or undisturbed primary evergreen forests. However, from
Table 2-1 and Figure 2-8, the biophysical values of dry evergreen forests have a similar
range to those of mixed deciduous and pine transition. Their DBH values were even
lower than that of mixed deciduous forests that were mostly disturbed primary forests or
old secondary regrowth. If we treat Site 82-12 as an outlier and exclude it from the
statistical analysis, the normalized average (ranging from 0 to l) of the different forest
types is shown in Figure 2-9. The biophysical parameters of the dry evergreen sites have
slightly higher values than those in Figure 2—8. The values in Figure 2-9 are shown in

Table 2-2 and will be used in the data analysis in the following chapters.

The biophysical parameters of dry evergreen forests in Figure 2-9 are still
underestimated. The possible reason for this underestimation comes from the method
used for the measurements. All of the dry evergreen study sites were measured during the
second field trip in which a point-quadrant plot method was used, whereas a fixed-radius
plot method was used in the first field trip. The point-quadrant plot method provides
efficient ways to measure biophysical attributes in each plot, but the accuracy is lower
than the fixed-radius method. In forests in late succession with higher heterogeneity,
there are more young trees surrounded the old parent tree. Therefore, young trees have a
much higher probability of being picked in each quadrant which results in underestimated
measurements. Further data analysis and possible compensation for the point-quadrant

measurements need to be done in future research.

25

2.3 References

Brown, S., Gillespie, A. J. R. and Lugo, A. E. (1989). Biomass estimation methods for
tropical forests with applications to forest inventory data. Forest Science: 35(4),
881-902.

Frazer, G. W., Canham, C. D. and Lertzman, K. P. (1999). Gap Light Analyzer (GLA),
Version 2.0: Imaging software to extract canopy structure and gap light
transmission indices from true-color fisheye photographs, users manual and
program documentation. Copyright © 1999: Simon Fraser University, Burnaby,
British Columbia, Canada, and the Institute of Ecosystem Studies, Millbrook,
New York, USA.

Le Toan, T., Beaudoin, A. and Guyon, D. (1992). Relating forest biomass to SAR data.
IEEE Transactions on Geoscience and Remote Sensing: 30, 403-411.

Luckrnan, A., Baker, J ., Kuplich, T. M., Yanasse, C. C. F. and Frery, A. C. (1997). A
Study of the relationship between radar backscatter and regenerating tropical
forest biomass for spacebome SAR instruments. Remote Sensingof Environment:
60, 1-13.

26

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Table 2-1 Average values of ground measurements in different forest types.
tree H (m) stem H (m) DBH (cm) density (#lha) canopy H (m) biomass (ton/ha)
dipt 10.624 7.736 17.242 711.871 5.775 48.510
mixed 13.298 9.640 22.899 544.754 7.316 85.032
pine trans 14.265 10.212 24.564 754.859 8.105 170.128
dry ever 14.157 10.179 21 .704 605.443 7.956 105.493
moist ever 22.025 17.121 35.340 777.622 9.809 436.373
Table 2-2 Average values of ground measurements in different forest types when the

outlier was deleted.

 

 

 

 

 

 

 

 

 

 

 

 

 

ree H(m) [stern H (m) DBH cm ensity (#Iha) canopy H (m) biomass ton/ha)

dipt 10.624I 7.736 17.242 711.871 5.775 47.756
mixed 13.298 9.640 22.899 544.754 7.316 92.713
ine trans 14.265 10.212 24.564 754.859 8.105 95.786
dry ever 14.489 10.262 23.338 581.271 8.4541 200.409
moist ever 22.023 17.121 35.340 777.622 9.809] 426.682

 

 

27

 

Mae Chaem Watershed

  

Mae Chaem

Thaila map

*5

 

 

 

Figure 2-1 The study area (Mae Chaem Watershed).

 

El
- _
-
-
-
-
-
1:]
I:

 

 

 

 

Figure 2-2 The DEM data (with hill shade effect) in the study area.

28

 

dry dipterocarps
._ 3&2

 
 
     
    

V oist
Evergreen

Dry Evergreen

 

w,»- ‘ inc Transition
”.5, '

flli ‘
Mixed decrduous

A

   

«pf-$.37." I- -i f '

 

Evergreen Gallery

 

Buffer Zone Dry Dlpterocarps

 

 

Figure 2-3 Vertical distribution of forest types in the study area.

 

land cover type
‘ - Tropical evergreen
- Mixed deciduous
"T Dry dipterocarps
Agriculture/open

1o 5 0 1o 20 30 LWIOther
RKilometers

 

 

 

Figure 2-4 Forest type map in the study area.

29

 

      
    

   

Mae Chaem Mae Sam oeng

. Sites

A GCP points

 

 

 

Figure 2—5 Study sites and ground control points (GCPs) in the study area.

 

 

 

 

 

 

(a) (b)

Figure 2-6 fisheye picture and its transformation in angular coordinates.

30

AGE ELEVATION

TREE H

BIOMASS CANOPY H TREE DEN DBH CM

. "97' II.

 

 

 

 

E

 

 

3N

\\\\\\\\\\‘

 

 

2a..

 

 

\.
I

 

 

 

 

 

 

 

 

 

 

 

/

4

/

I I /

I I I ¢

I I I I I I I . . I I I /
.é- ' Q I” ' a“ ~55” ' " " " 8% 1717.1
ELEVATION AGE TREE_H DBH_CM TREE_DEN CANOPY_H BIOMASS

Figure 2-7 Correlation scatterplot matrix of ground-measured data.

31

 

0.9 - — i i ‘

 

 

3

c

“e’

2 0-7 “ _ ltree H (m)

:r

a [stem H (m)

“E’ 0-5 — DDBH (cm)

3 [jtree density (#/ha)
% 0.3 , .biomass (ton/ha)
é

o 0.1

:

 

 

 

 

 

 

 

 

 

 

 

dry dipt mixed pine dry moist
decid trans ever ever
forest types

Figure 2-8 Normalized average ground measurements in different forest types.

 

0 9 - E

0-7 7 .tree H (m)
mstem H (m)
[I DBH (cm)

ijtree density (#/ha)
[biomass (ton/ha)

 

 

.0
w

 

normalized measurements
0
U1

.0
_\
l

 

 

 

 

 

l

drydipt mixed pine dry moist
decid trans ever ever

 

 

 

 

 

 

 

forest types

Figure 2-9 Modified normalized average ground measurements in different forest

types (the outlier of dry evergreen site was removed).

32

Chapter 3
Estimation of Tropical Forest Fractional Cover with Landsat

ETM+ and IKONOS Imagery

3.1 Introduction

Tropical forests are being subjected to various extents of disturbances in addition to
outright deforestation. Selective logging, forest fire, and disturbance increases along
newly created forest edges and result in significant changes in forest structure and canopy
integrity and often promote the probability of other types of disturbance. Degradation, in
the form of decreased fractional cover, results from many direct and indirect human
impact including selective logging (Barros and U111 1995), forest fires (Cochrane et al.
1999), and biomass collapse in fragmented forests (Laurance et al. 1997). Forest cover is
an important indicator of forest degradation and forest thirming activities such as
selective logging and natural fire. An accurate assessment of the spatial extent of forest
cover is a crucial requirement for quantifying the sources and sinks of carbon from the

terrestrial biosphere (DeFries et al. 2000).

Forest fractional cover (fc) is the area covered by forest canopies per unit area on the
ground. At a larger or regional scale, it is also called forest canopy cover when the intra-
tree gaps are neglected. Field measurements generally provide good estimates of forest

fractional cover, but are often too expensive and time-consuming to be applied over large

33

areas. Although direct measurement of fc is theoretically possible with high spatial

resolution imagery, it is very expensive and sometimes impossible to cover large areas.

In general, natural tropical forests have dense fractional cover. In degraded areas, land
cover can be categorized into two types: trees and open areas. In the dry season, trees in
tropical forests can be characterized as green vegetation while open areas are often
senescent grass that has a similar spectral response to bare soil. Green vegetation has high
reflectance in the near infrared (NIR) and low reflectance in the red spectral regions,
while soil reflectance changes gradually with spectral wavelength. Therefore, vegetation
and bare soil have distinct spectral signatures that are separable in remote sensing images.
In many cases, however, a single pixel from satellite images does not contain pure
vegetation or open area. Instead, it often consists of a mixture of open area and
vegetation, especially in images with medium to coarse resolution. Therefore, the spectral
response of a pixel is a collective contribution from both green vegetation and the open
area. The spectral signature varies with the abundance of green vegetation. The more
green vegetation in a pixel, the higher the spectral response in NIR band and the lower in

red band.

Using data at medium to coarse resolutions, some effort has been made to estimate fc
with sub-pixel information in the reflectance domain (Quarmby, et al. 1992; Settle and
Drake 1993). Hanan and Prince (1991) combined red and near infrared reflectance in an
area-additive model to estimate the abundance of vegetation. Price (1992) estimated fc

and vegetation amount from visible and near infrared reflectance with AVHRR

34

observations in a “mixed pixel” case. J asinski and Eagleson (1990) used a pixel’s
distance to the soil line in the near infrared-red reflectance scattergram to estimate fc.
Radeloff et al. (1999) created fraction images in Jack Pine forests with spectral mixture
analysis of near infrared reflectance to detect budworm defoliation. Townshend (1999)
and DeFries et al. (2000) derived global continuous vegetation cover with coarse

resolution AVHRR (1km) and MODIS (250m, 500m, and 1km) data.

Among the studies above, linear spectral mixture analysis, or spectral unmixing theory,
was used to compute the abundance or percentage of vegetation cover in one pixel
(Wessman et al. 1997; Adams et al. 1993). These models operated in the spectral
reflectance domain. However, the use of reflectance values with linear unmixing models
is problematic. Although any reflectance band could be applied in the model, it is
difficult to choose the optimal one. Most studies applied the red or near infrared bands.
Although the surface reflectance in these two bands is primarily a function of the
dominance of vegetation, it is also influenced by moisture and the structure of both
vegetation and soil and other external factors such as atmosphere conditions and sun-
target-sensor geometry. When estimating the proportions of land cover within a pixel
with a linear mixture model in the reflectance domain, the standard error of estimates
between the percentages derived fiom mixture modeling and the actual land cover

percentages was as high as 11% (Townshend et al. 2000).

Vegetation indices (VIs), which are often mathematical combinations of different spectral

reflectances, suppress the spectral variations of reflectance, and make vegetation and

35

non-vegetation comparisons more stable (J asinski, 1990). In the past few years, more and
more attention has been made to estimate fc with the normalized difference vegetation
index (ND V1), one of the most commonly used vegetation indices. For example, Gutman
and Ignatov (1998) estimated regional green vegetation fraction distribution from
NOAA/AVHRR data using ND VI in a linear unmixing model. Zeng et al. (2000)
validated this method for each land cover type in the International Geosphere-Biosphere
Program (IGBP) land cover classification scheme and created global l-km fractional
vegetation cover maps with AVHRR data. Qi et al. (2000) reported that when using a
linear unmixing model with ND V1 instead of spectral reflectance, the atmospheric effect
was reduced. They also built vegetation fc maps in a semi-arid rangeland with l-km
SPOT VEGETATION, 30-m Landsat TM, and 3-m airborne TMS data. The results from

different datasets were highly comparable.

These studies assumed that a linear relationship between ND VI and canopy fractional
cover is adequate (Zeng et al. 2000; Wittich 1997; Gutman and Ignatov 1998). The
validity of the linear fc-NDVI relationship has been confirmed in some studies (Myneni
et a1. 1992; Carlson et al. 1990; Wittich and Hansing 1995). However, some other studies
(Myneni and Williams 1994; Carlson and Ripley 1997) reported that the relationship was

non-linear because ND VI was rapidly saturated when vegetation amount increased.

In this chapter, a linear unmixing model in the vegetation index domain was applied to

estimate canopy fractional cover in tropical forests within the study area. Six commonly

used vegetation indices were calculated and the one that was most linearly related with

36

vegetation amount was selected in the model. The forest fractional cover in the study area
was then estimated with an ETM+ image acquired in February 2000. The results were
validated with both ground measurement and l-m pan—sharpened H(ONOS data. The
seasonal variation of forest fractional cover distribution was also discussed and the

adjustment of the fractional cover map to wet-season conditions was examined.

3.2 Remotely Sensed Data

3.2.1 Landsat ETM+ image

A subset of ETM+ data (30-m resolution) was processed to estimate forest fractional
cover in the Mae Chaem Watershed (Figure 3-la). The image was acquired on February
2, 2000 during the dry season. The deciduous forests, such as dry dipterocarps and some
mixed deciduous species, were senescent and leaf-drop had begun. The leaves of
evergreen trees, however, did not have significant annual change. A small area of
evergreen forests was covered by cloud and was replaced by an ETM+ scene acquired

one month later (March 5, 2000).

The ETM+ image in this study was atmospherically, geometrically, and topographically

corrected before firrther processing.

Atmospheric Correction
The ETM+ image was radiometrically corrected using the information in its header file,
the ETM+ image was converted from digital number (DN) value to top-of—atmosphere

(TOA) radiance and then TOA reflectance in the range of [0,1]. To represent the true

37

surface conditions, the atmosphere effects need to be reduced so that the TOA reflectance

were converted to surface reflectance.

The ETM+ image was acquired in the early spring which is part of the dry monsoon
season. The atmosphere model was set as a typical Tropical Atmosphere, and the aerosol
model was Rural Extinction with visibility distance = 23km, a typical visibility in clear
air. The MODTRAN4.0 model was run to correct the atmosphere effect in the imagery.
The correction equations for bandl-5 were:

pmf, = 1.3923pmA, — 0.114

,omf2 ==1.3838,0m,,2 — 0.065

,osmf3 =1.3192,0T0A3 —0.0373 (3-1)
pm“ =1.3531pTOA4 —- 0.0201

pmfs =1.2419pm,,5 —0.0035

Here the ,0qu is the surface reflectance, and pm, is the reflectance on the top of

atmosphere. After atmospheric correction, the DN values were converted to surface

reflectance in the range of [0,1].

Geometric correction

According to the information in its header file, the ETM+ image was georeferenced to the
UTM projection, zone 47, spheroid and datum WGS84. However, due to system
distortions and the limited accuracy of the ground control, there was some deviation
between the UTM position in the image and its true position on the ground. In the areas
far from the image center, the deviation could be as much as several hundred meters.

Therefore, the georeferenced ETM+ image needed additional geometric correction.

38

The ground GPS recordings of around 60 positions were collected during field trips.
These ground control points (GCPs) were associated with unique ground features, such as
the curve of a road, the center of a bridge, and the corner of a reservoir. It was easy to
recognize these positions in the ETM+ image. With these GCPs, a second-order
polynomial model was selected to do the geometric correction. The model had a total

error of 23.6m, less than one pixel of the ETM+ image (30m).

Topographic correction

Mae Chaem Watershed is located in a mountainous area. Its elevation changes from
250m near Chiang Mai city to 2,550m at the peak of Mount Doi Inthanon. Slopes vary
from 0° to more than 50°. The topographic change in the study area makes the sun-
ground-satellite geometry complicated and the surface reflectance varies with its local
incidence angle. Therefore, the topographic effect in the surface reflectance has to be

corrected before further processing.

In this study, a bidirectional reflectance distribution function (BRDF) model (Rahman et
al. 1993) was applied to correct the topographic effect in the ETM+ surface reflectance
image. For each pixel, the local slope, azimuth aspect, and local incidence angle were
calculated from DEM data (Figure 3-2a). A nadir view was chosen as the standard view

direction, and the reflectance from its local incidence angle was adjusted to the one with

standard incidence angle (0°).

39

Let 05 , 6v , (p5 , (p, be sun zenith, sun azimuth, and satellite view zenith, view azimuth
angles which can be found in the image header file (gas =0° and to, can be any value for
ETM+ image), and a , ,6 be the slope and aspect which can be calculated from DEM
data, then local sun zenith 6?; , local view zenith 6;, and local relative azimuth angle p

can be calculated:

cos 0, = cosa cos 6, + srn a srn I93 cos(,6 — (0,)

cos 6:, = cosa cos 6,, + sin a sin 6,, cos(,B — gov) (3-2)

(DEW-<0.

 

Assuming the sun-satellite geometry as , 6,, , and to (relative azimuth angle) as standard

geometry, the surface reflectance in this geometry is modeled as (Rahman et al., 1993):

(cosélscosflv)"‘l . l—GI)2 (1+1—p0
0(<><ISI9.+<>0861,)“" [l+®2—2®cos(7r—§)]3/2 1+G

 

 

p(65’6v’¢):p ) (3'3)

where p0 and k are two empirical surface parameters. O is the function parameter

controlling the relative amount of forward (0 S O S 1) and backward scattering (-1 S O S

0). The phase angle 5 and geometric factor G can be calculated by

cos; = cos 6, cos I9, + sin 65 sin 6,, cosrp;

 

G = \/tan2 65 + tan2 6, — 2tan6$ tang, cosrp.
Similarly, the surface reflectance p'(6_; ,6; ,go') in local geometry can be calculated. The

correction coefficient p(¢95 ,6v,(p) / p'(6ls' ,6,',,,(p') was then applied to the ETM+ surface

reflectance image.

40

The topographically corrected image was much smoother even in areas with large
topographic variation. Figure 3-2b showed a subset of the ETM+ image (Band4+3+2)
near the peak of Mount Doi Inthanon before and after BRDF correction. It is obvious that
before BRDF correction, the topographic effect was dominant creating dark shadows in
the image. After correction however, the image was totally flattened and the shadows

were greatly reduced.

After atmospheric, geometric, and topographic correction, the ETM+ image was ready

for forest fractional cover estimation.

3.2.2 High-resolution IKON OS Images

For the purposes of spatial comparison and seasonal adjustment, ten IKONOS images
were acquired in the study area (Figure 3-1b), covering a total area of 673 km2 in Mae
Chaem Watershed. The study sites visited during the two field trips were also covered by
these IKONOS images. One IKONOS scene was acquired on February 27, 2000, 25 days
later than the ETM+ image acquisition, during the dry season. The remaining IKONOS
scenes were acquired on October 9, 2002, during the wet season. All of these IKONOS
images were geometrically corrected to the ETM+ image. No atmospheric and
topographic correction was done on these IKONOS images because only classification

and Visual interpretation were needed for the process.

The IKONOS images have four spectral bands at 4-m resolution and one panchromatic

band at l—m resolution. For each IKONOS image, the l-m panchromatic band was

41

merged with the 4-m spectral bands to produce a pan-sharpened multi—spectral image at
l-m resolution. Since the size of tropical trees is much larger than 1 meter, the pan-
sharpened IKONOS image reveals enough details for forest fractional cover calculation.
The pan-sharpened IKONOS images were then processed to calculate fc at a spatial unit
of 30x30 m, comparable with the ETM+ estimated fc. The IKONOS calculated fc values
were assumed as “ground truth” to validate the ETM+ estimation in continuous spatial

areas .

Since most of the IKONOS imagery was acquired in the wet season, the ETM+ estimated
fc from the dry season could be adjusted to the wet season conditions when the vegetation
was flourishing. After adjustment, the seasonal variation of fc in deciduous forests was

minimized and, therefore, the forest fractional cover map provided a better description of

forest integrity in the study area.

3.3 Methods

3.3.1 A linear unmixing model
For a surface with multiple components, the macroscopic linear mixing model used to

estimate the spectral response can be expressed as:
[D] = [RIC] (3-4)
where [D] is the data matrix, [R] is the response matrix, and [C] is the eigenvector matrix

consisting of the relative contributions of different components. It can also be written as:

d... = 2w... (3-5)
.i =1

42

where di’k is the measured response of pixel k in wavelength 1', n is the total number of

independent reflecting components in this pixel, ’1‘.)- is the response of component j in

wavelength 1', and CM is the relative contribution of component j in pixel k.

Since the ETM+ image in this study was acquired in the dry season, the open areas in the
forests were bare or covered by senescent grasses. Consequently, each pixel of the ETM+
image can be assumed to consist of two components: green tree canopy and open area.
The spectral response of open areas is characterized by bare soil and does not change
much all over the study area. The senescent grass has a similar spectral response to bare
soil and, therefore, is neglected. The reflected spectral values of these two components
are independent of each other. Let R be the total spectral response of the pixel at certain

spectral band, R the response of tree canopy, and R the response of open area,

canopy open
Eq.3-5 becomes:

R = Rama” fc + ROW (1 — fc) + 8 (3-6)
where fc is the forest fractional cover in the area corresponding to one pixel. An error
term 8 is introduced to account for some insignificant remaining components within the

pixel. In tropical forests during the dry season, the two components (green tree canopy

and open area) are dominant and the error term, a, is ignored.

Eq.3-6 suggests that at any wavelength, the total reflectance of a mixed pixel is the linear

combination of the reflectance from its vegetation and open area, weighted by their

percentage cover fc and (l-fc). With Eq.3-6, when R and R are known in an

canopy open

43

ETM+ image, the fc value can be easily calculated. However, the reflectance of surface
targets changes greatly at different wavelengths. For example, the reflectance of green
vegetation is high in NIR, but low in red wavelengths. Even at certain wavelengths, the

values of R and R are highly influenced by the vegetation wetness, structure, soil

canopy OpCII
moisture, and texture (J asinski, 1990). Therefore, it is difficult to choose an optimal

wavelength for R and R in Eq.3-6.

canopy open

A vegetation index (V1) is a mathematical combination of multiple spectral bands, which
can suppress the external effects mentioned above (Gutman et al. 1998; Qi et al. 2000).

Replacing the spectral response R with V], Eq.3-5 becomes:

VI = VI canopy fc + VIope" (1 — fc) (3-7)

where V1 is the vegetation index of green tree canopies in one pixel, and V1 is the

canopy open

vegetation index of open area in that pixel.

It should be noted, however, that although both vegetation index and spectral reflectance
are descriptions of vegetation properties, they are not linearly related. Eq.3-7 is not
directly deducted from Eq.3-6. Instead, it is an approximation by choosing the optimal VI

that is most linearly related to the vegetation amount in the study area.

3.3.2 Optimal vegetation index (V1)
Several vegetation indices (VIs) have been developed over the past decades. The most
commonly applied VIs include the Normalized Difference Vegetation Index (ND V1)

(Tucker et a1. 1979); the Soil Adjusted Vegetation Index (SA V1) (Huete et al. 1988); the

44

Modified Soil Adjusted Vegetation Index (MSA V1) (Qi et al. 1994); and some vegetation
indices for global applications such as the Enhanced Vegetation Index (E VI) (Huete et al.
1999); the Global Environmental Monitoring Index (GEM!) (Pinty and Verstraete 1992);
and MERIS Global Vegetation Index (MG V1) (Gobron et al. 1999). Among these
vegetation indices, ND VI is the earliest and the most widely applied vegetation index in
remote sensing applications. Most of other VIs are modifications ND V1 in an attempt to
depress the influence of atmosphere and surface conditions and to improve its accuracy in

extracting vegetation information from remotely sensed data.

To select the optimal VI in the linear unmixing model, all these VIs (ND V1, SA VI, MSA V],
E VI, GEM], and MG VI) were calculated using simulated surface reflectance in the
Scattering by Arbitrarily Inclined Leaves (SAIL) model (Verhoef 1984). The reflectance
was simulated as a function of leaf area index (LAI), an indicator of green vegetation
abundance. A forest of sparse to moderate density was modeled with LA] values from O
to 4.0 at an interval of 0.2, which is similar to most of the forests in the study area. The
conditions of dense evergreen forests were not considered because both VIs and forest fc
values became saturated when LA] was very high. The same sun-target—sensor geometry

as the Landsat satellite was used in the reflectance simulations.

As shown in Figure 3-3, the values of all six VIs increase with leaf area index. The E V],
SA VI, and GEM] quickly reach their saturation zone when LAI approaches 3.0. Although
ND VI has a higher value and increases almost linearly at lower LAI, it saturates at a

similar LAI threshold as the other three vegetation indices. MG VI has highest value and

45

does not saturate until LAI = 4.0. But, it increases rapidly when LA] is low, then quickly
slows down when LAI is higher than 2.0. The MG VI-LAI curve is polynomial. MSA VI has
a more linear relationship with LAI. It saturates only when vegetation is very dense (LAI
higher than 4.0). Since we are interested in an index that is most suitable for tropical
forests, MSA VI was chosen as the optimal vegetation index in this study because of its

sensitivity at higher forest densities.

The MSA VI reduces the effect of the soil background by using a soil adjustment function
that is determined from local soil reflectance. Unlike the empirical factors in other VIs,
the adjustment factor in the MSA VI changes as a function of canopy density and the slope

of the soil line (Qi et al. 1994):

MSA VI = pm —p,,, (1+L0) (3-8)
ION/R +pred +1‘0

 

where L0 is a soil adjustment function. It is expressed as a function of the soil line slope
and the reflectance properties on the ground:
L0 : [(pNIR _ pred ) X Slope +1+ ION/R “I” prod I2 _ 8'0 x Slope X (ION/R - pred) (3'9)

In our study area, the slope of the soil line is 1.2, calculated with the surface reflectance

values of soil surfaces in the atmospherically corrected ETM+ data.

In the MSA VI image (Figure 3-4), the tropical evergreen forests have high values (light
gray to white). The clear-cut areas, bare soil, and fallowed agricultural fields have low
values (dark gray or black), while the dry dipterocarps and mixed deciduous forests are in

between.

46

From Eq.3-7,fc can be expressed in the form of the MSA VI:

MSA VI — MSA V1, ,,
c = p (3-10)
MSA VI —MSA VI

canopy open

 

Here MSA VI and MSA VI are the endmembers in the linear unmixing model.

canopy open

MSA V1 is the vegetation index of full-cover green forests with fc = 1.0. MSA VI

canopy open

is the vegetation index of open area with fc = 0. The grasses and shrubs in open areas are

senescent in the dry season and, therefore, MSA V1 is close to the MSA VI value of bare

open

soil.

The two endmembers MSA VI and MSA VI in Eq.3-10 can be identified in the

canopy open
ETM+ image. A sand mining spot with a smooth light gray tone in the middle left of the
image was chosen to represent the open area in the study area. It had been observed from
the field trips that the top of Mount Don Inthanon was covered by the tropical moist
evergreen forests with very high fractional cover values. A small area with a smooth
bright red tone in the middle of the ETM+ image (band 4+3+2) was chosen to represent
full-cover green canopy. The mean of the MSA VI values in each area was calculated, and

the two endmembers in the linear unmixing model were thus determined: MSA VI =

canopy

0.71, and MSA VI = 0.16. These values were used in Eq.3-10 to compute the forest

open

fractional cover in the study area.

47

 

3.4 Canopy Fractional Cover Analysis

Figure 3-5 is the fc map estimated from the ETM+ image in the study area, binned into 10
groups with an interval of 10%. All pixels with fc value less than 0 or higher than 100%
(reaching saturation) were truncated into 0 or 100%, respectively. Non-forest surfaces,
such as villages, water bodies, fallowed agriculture fields, and clear-cuts, have a
fractional cover lower than 20%. Forests in the watershed, with fractional covers ranging
from 20% to 100%, are assigned a gradual color scheme changing from light grayish

green to dark green, indicating the increasing forest density.

The forest fractional cover map in Figure 3-5 shows a similar pattern with the land cover
and forest type map (Figure 1—4). Along with the altitude, the forest types change in a
sequence of dry dipterocarps, mixed deciduous, dry evergreen and moist evergreen. In
accordance, the fractional cover changes from low values in dry dipterocarps to

saturation (100%) in moist evergreen forests.

Even at each forest type, the fractional cover changes greatly. The fractional cover for
each forest type is also not smoothly distributed. Rather, it is scattered because of various
disturbances. Since the ETM+ image was acquired during the dry season, dry
dipterocarps, especially at lower elevations, were mostly senescent or leaf-off. The
vegetation index values are thus much lower than other forests and, therefore, the
fractional cover is lower. Some dry dipterocarp forests even have fractional cover lower
than 20%. Aside from the seasonal effect, dry dipterocarp forests suffer from intense

human disturbance of burning for agriculture or cutting for firewood (personal

48

communication with local foresters). Selective logging for valuable trees such as teak
also used to be a common human activity in the past decades. As a result, most of the dry
dipterocarp forests are young second-growth forests with a fractional cover ranges from
10% to 40%. The fc values of the forests are lower in areas close to villages and higher in

the mountains.

The mixed deciduous forests are found at higher elevations and are less affected by
seasonal variation and human activities. Clear-cuts for small area agricultural fields and
natural fires are thus the common types of deforestation. The fractional cover of mixed

deciduous forests ranges from 40% to 80%.

Evergreen forests occupy the highest elevations in the watershed. Human disturbance is
much lower than for other forests. Natural fire is thus the major disturbance in these
forests. Also, as the result of government policy decisions, some large areas, sometimes
the whole slopes, were cleared long time ago and evergreen forests regenerated slowly.
The fractional cover of the evergreen forests ranges from 70% to 100%. The fractional

cover values of most moist evergreen forests on the top of the mountains are saturated.

3.5 Validation

The forest fractional cover estimated with the ETM+ image was validated with both
ground measurements and hi gh-resolution IKONOS images. The IKONOS images
acquired in the wet season can also be processed to evaluate the seasonal change of the

forest fractional cover distribution in the study area.

49

3.5.1 Validation with ground measurements

The ground-measuredfc values were assumed to be correct (i.e. ground truth). At each
site, ten fisheye pictures were taken along a 150m transect (300-m in second field trip)
which was assumed to represent the area of 150x150m (or 300x300m). The fc values
calculated with these pictures were averaged to represent the fc in an area of 150x150 m
(300x300 m in second field trip). The fc estimated from ETM+ image at each site was
averaged using a 5x5 window (150x150 m) to match the area of the ground
measurements and to reduce the autocorrelations between pixels. Figure 3-6 compares the
ground-measured and ETM+ estimated fc values. To be seasonally matched, only

measurements from the dry season were compared.

Figure 3-6a shows that the ETM+ estimated fc is correlated with ground measurements
(R2 = 0.757). The regression line deviates from the 1:1 line, indicating that the ETM+
estimated fc is different from ground-measuredfc. In sparse forests Where the fractional
cover is low, the ETM+ estimation is lower than the ground measurements. Contrarily,
the ETM+ estimation is higher than ground measurement in dense forests. The ETM+
estimation matches well with ground measurements in forests with moderately high
density, in which the fc is around 70% - 85%. In very dense forests, the ETM+fc

estimation saturates while the ground measurements are less than 95%.

The difference is also shown in Figure 3-6b. Both the ETM+ estimated and ground-

measured fc values increases along the elevation gradient, in accordance with the forest

50

types changing in a sequence of Dry dipterocarps, Mixed deciduous, Pine transition, Dry
evergreen, and Moist evergreen. For Dry dipterocarps, the ETM+ estimated fc is much
lower than ground measurement. For Mixed deciduous and Pine transition forests, the
ETM+ estimated fc is lower than the ground measurements, but the difference is much
less than that of Dry dipterocarps. For evergreen forests, the ETM+ estimated fc is a little
higher than the ground measurements. The ETM+ estimated fc is saturated in the moist

evergreen forests (reaching 100%).

The difference between the ETM+ estimated and ground-measuredfc comes from the
different processing mechanisms. The ETM+ estimated fc in each pixel is calculated with
a linear unmixing model in which the vegetation index is a combined contribution of
green canopy and open area. Only green leaves in forests are considered in the model.
Therefore, the ETM+ estimated fc is actually the “green” fractional cover. The ground—
measured fc, however, is calculated from a binary classification of a hemispherical
photograph taken on the ground. All elements except open sky, including green leaves,
senescent leaves, stems, branches, and trunks, contribute to the value. In this meaning,

the ground estimated fc is the total cover of forest projected area.

The measurements shown in Figure 3-6 were made in the dry season when deciduous
forests are senescent or leaf-off. Therefore, in dry dipterocarps, the ETM+ estimated fc is
low due to a lack of green leaves. The ground-measuredfc, however, is much higher
because of the significant contribution from the woody components (Figure 3-6b). In the

mixed deciduous and pine transition forests, some species were partially leaf-off and

51

others remained evergreen. Consequently, both the ETM+ estimated and ground-
measured fc values are higher. The ETM+ estimated fc is still lower than the ground
measurements, but the difference is much smaller. Moreover, the vegetation index of pine
trees is generally lower than that of broadleaf trees due to the form of the needle leaves.
This also makes the ETM+ estimated fc in the pine transition zones lower than the ground
measurements. Both the dry evergreen and moist evergreen forests are much denser than
the other forest types of the area. The leaves remain green in. the dry season and,
therefore, both the ETM+ estimated and the ground-measuredfc values are high (>90%).
In dense evergreen forests, the in-tree gaps become dominant when calculating fc with
circular hemispherical photographs using the GLA software. However, these small-size
gaps are unrecognizable in the ETM+ image with 30-m resolution. As a result, the ETM+
estimated fc is higher than the ground measurements. In very dense, moist evergreen
forests, the ETM+ estimated fc reaches saturation while the ground measurement is less

than 95%.

Figure 3-6 explains the differences between ETM+ estimated and ground—measuredfc
values in each forest type. For fractional cover distribution in a large area with all forest
types, Figure 3-6a shows that the ETM+ estimated fc values are highly related to the
ground truth. This relationship indicates that the forest fractional cover distribution in the
study area can be estimated with ETM+ image with a reasonably high accuracy,
especially given the leaf-off issue. This validation, however, was based on ground

measurements in only 32 isolated sites. To make a spatial validation in the study area, the

52

high resolution IKONOS images were assumed as ground truth to compare with the

ETM+ estimated fc distribution.

3.5.2 Validation with l-m IKONOS data

The pan-sharpened IKONOS images have four spectral bands at l-m resolution in which
the openings between the trees are visible. In this study, the fc derived from the IKONOS
data was assumed to be correct (i.e. ground truth) wherever intensive ground
measurements are not available. Each IKONOS image was processed in six steps to

calculate fc in a unit area of 30x30 1n:

Step 1: Unsupervised classification

 

Based on the gray values in the four spectral bands, the pixels in the IKONOS image
were grouped into 50 clusters using the ISODATA unsupervised classification technique.
A maximum iteration of 50 in 95% convergence level was applied in the classification.
The signature file was thus created for next step.

Step 2: Supervised classification

With the experience of the field trips, the 50 signatures in the signature file were merged
into several signatures based on the land cover types and seasonal variation. With the
new signature file, a maximum likelihood supervised classification was made of the
IKONOS image and a land cover image was generated. There were 5 classes in the dry
season: forest, shaded forest, bare soil, shaded bare soil, and water body. For the wet
season, there were 8 classes: forest, shaded/dark forest, bright agriculture/open area,

fallowed agriculture/open area, bare soil, shaded bare soil, water body, and cloud/shadow

53

(if any). Figure 3-7 shows the feature spaces of each cluster between band 3 (red) and 4
(NIR). The bright forest and dark forest are combined into one cluster because they are
the major concern in this classification procedure. All classes are very separable.

Step 3: Majority filtering

In the land cover image created in step 2, there were often some isolated classes
embedded in large classes. For example, there were agriculture fields with one or two
pixels in forests. These pixels were not real classes and should be removed. Using a 3x3
moving window and a majority decision rule, the pixels in the small classes were merged
into the larger surrounding classes.

Stg; 4: Forest/non-forest class

The land cover image was recoded into a 0/1 binary image, assigning forest and shaded
(or dark) forest as l and all others as 0. The resulting forest/non-forest image maintained
a pixel size of 1 meter.

Steg 5: 30x30 matrix convolution

The forest/non-forest image was convolved with a 30x30 moving window to match the
pixel size of the ETM+ image. The resultant pixel value was a float point value between
0 and 1, which was equal to the forest fractional cover in a 30x30m area on the ground.
Step 6: Forest fractional cover

Applying the nearest neighbor technique, the image from step 5 was resampled to 30-m
pixel size. The result was the IKONOS calculated fractional cover distribution with a unit

area of 30x30 m.

54

The IKONOS estimated fc distribution in the study area is an alternative to “ground
measurements” to spatially validate the ETM+ estimation. It should be noted that, the
IKONOS fc values were calculated with unsupervised and supervised classification
techniques and were very different from both hemispherical photographs for ground fc

calculation and the linear unmixing model for ETM+fc estimation.

The ground-measured and the IKONOS and ETM+ estimated fc values were compared in
Figure 3-8. To avoid the seasonal effect, only fc data in the dry season were compared.
Only 7 study sites were covered by the IKONOS image in the dry season. All forest types
were included: dry dipterocarps (1), mixed deciduous (2), pine transition (1), and
evergreen (3). Among these 7 sites, only 6 sites were measured during the second field
trip. Therefore, in Figure 3-8a, there are 7 points of ETM+fc, 7 points of IKONOSfc,

and 6 points of ground-measuredfc.

When compared with ground fc values from fisheye photos, the IKONOS fc was
underestimated in sparse forests and overestimated in dense forests (Figure 3-8a). This
could be explained by their different processing mechanisms. In dry dipterocarp forests
that have low fractional cover in the dry season, the light-colored stems/branches and
senescent leaves are more easily mis-classified as bare soil in an IKONOS image.
However, they are important non-sky components in the hemispherical photos. In dense
evergreen forests, the in-tree gaps less than 1x1 1n were not observable in the IKONOS

image, whereas they played an important role in hemispherical photos. The small-area

55

 

gaps in the IKONOS classification images were also smoothed out during the majority

filter processing.

These differences aside, Figure 3-8a shows that the IKONOS fc is highly correlated to
ground measurements (R2 =0.97). This indicates that where intensive ground
measurements are lacking, the IKONOS fc could serve as ground truth for the purpose of
validation. The corresponding ETM+fc values at the limited study sites were also plotted
in Figure 3-8a. They too are highly correlated with the IKONOS fc (R2 =0.96). The
correlation line is close to the 1:1 line, indicating that the ETM+ image can be processed

to estimate fc with high accuracy.

Instead of the limited ground measurements, the IKONOS estimated fc values could be
compared with the ETM+ estimation in any corresponding areas. In this study, four
subset areas in different forest types were chosen for the comparison: dry dipterocarps,
mixed deciduous, transition zone, and evergreen forests. Since it is impossible to identify
pine transition forests in the ETM+ image, the transition zone covers mixed deciduous,
pine transition, and dry evergreen forests. The subset of evergreen forests covers both dry
evergreen and moist evergreen forests because it is difficult to identify them in the ETM+
image. The subset of evergreen forests was chosen nearby the peak of Mount Doi
Inthanon. In each subset, 100 positions were randomly selected to avoid the
autocorrelation effect. The fc value at each position was an average of a 3x3 window to

reduce the possibility of geometric mismatch between the ETM+ and IKONOS images.

56

As shown in the scatterplot (Figure 3-8b), the accuracy of the ETM+fc estimation varies
with different forest types. For dry dipterocarps that have a low forest density, the fc
values of both the IKONOS and ETM+ estimation are more scattered than other forest
types. Most of the ETM+fc values are clustered in a range of 20% to 40% while the
IKONOSfc values ranges from 10% to 60%. The distribution of the ETM+fc in dry
dipterocarp forests is smoother than that of the IKONOS, but the values are
underestimated. The ETM+ estimation is better for the mixed deciduous forests whose
density is higher. However, the ETM+fc is still underestimated, ranging between 30%
and 60%, whereas the IKONOS fc could be as high as 80%. For the transition zone that
has a much higher forest density, the ETM+ and IKONOS fc match well. Both the ETM+
and IKONOSfc values range from 60% - 90%, and the data points are closer to the 1:1
line. For the moist evergreen forests where the density is very high, both ETM+ and

IKONOS estimated fc values are saturated.

All points in Figure 3-8b are more or less scattered along the 1:1 line. The ETM+fc is
correlated with the IKONOS fc R2: 0.70). Similar to Figure 3-6a, it indicates that the
ETM+ image is a good source of forest fractional cover estimation over large areas.
Figure 3-8b also shows that when forest cover is not saturated, the higher the forest

density, the higher the accuracy of the fc estimation.

Figure 3-9 presents the H(ONOS image, the corresponding subset of the ETM+ image
and their fc maps. To avoid the seasonal variation, only the IKONOS image acquired in
the dry season (February 27, 2000) was compared. The IKONOS and ETM+ estimated fc

distributions show both similarities and differences in each forest type. In the fallowed

57

agriculture fields in the lower left of the subset, the IKONOS estimated fc values are
around 0, and the ETM+ estimated fc values are less than 20%. In dry dipterocarp forests
in the upper left of the subset, both the IKONOS fc and ETM+fc have a wide range of
distribution. However, the IKONOS fc is more scattered with a range of 0-60%, while the
ETM+fc has a smaller range of 20-40%. Conversely, in the evergreen forests in the right
of the subset, the IKONOS fc distribution is less variable than that of the ETM+. In the
mixed deciduous forests in the middle of the subscene, the fc distributions from the two

images are very similar.

The secondary forests that are visually interpreted in both images can also be identified in
the fc maps. In the middle of the subset, there is large area of regeneration of evergreen
forests in their early stage. Both ETM+ and IKONOS fc values are in a range of 20-60%,
much lower than the natural, mature evergreen forests nearby. In the late-stage secondary
forests such as the isolated small areas in the lower right of the subset, the ETM+fc map
has fractional cover values around 60-80%. However, in the IKONOS fc map, the values

exceed 90% and cannot be separated from the dense, mature evergreen forest nearby.

The differences between the IKONOS and ETM+ estimated fc values are mostly at local
scales. When all forest types are considered in the whole subset, the ETM+ estimated fc

distribution shows high similarity with the one from the IKONOS image. In Figure 3-9,

from west to east of the subset, the forest types change from dry dipterocarps, mixed

deciduous, and evergreen forests as visually interpreted in the ETM+ image. The fc

58

values in both the IKONOS and ETM+fc maps also show a gradual increase from 0 to

100%.

As seen in Figure 3—8 and 3-9, the accuracy of ETM+fc estimation is high when
compared with the IKONOS fc that is presumed correct. The accuracy also varies with
forest types in different fractional cover densities. For dry dipterocarps that have low
fractional cover in the dry season, the fc is underestimated with the ETM+ image,
although its distribution is less variable than the IKONOS fc. For mixed. deciduous and
transition zone forests, the correlation between the ETM+ and IKONOS estimated fc
values are the highest among all forest types. For evergreen forests, especially moist
evergreen forests, both the ETM+ and IKONOS estimated fc tends to saturate. The
ETM+ estimated fc map also reveals the different stages of forest clear-cut and regrowth
in evergreen forests. In general, the higher the forest density, the higher the accuracy of
ETM+fc estimation. Most of the tropical forests have high density and, therefore, the
method presented in this study is promising for large-area fractional cover estimation,

tropical forest evaluation, and ecosystem management.

3.5.3 Seasonal adjustment of fractional cover

The green fractional cover in the wet season is a more realistic representation of the
fractional cover distribution when all types of forests are leaf-on. However, sky
conditions during the wet season in the tropics are often cloudy or hazy. For satellite
sensors with coarse temporal resolution and wide swath such as ETM+ (16 days, 180km

swath width), it is often difficult to acquire clear images during the wet season. It is,

59

however, possible for the IKONOS satellite who has a sensor with a more flexible repeat
period (an average of 1.5 days). The IKONOS images acquired in the wet season can be
used to adjust the ETM+fc from the dry season to the “true” green cover of the peak

growing season.

Nine IKONOS images were acquired on October 9th, 2002 (wet season). These images
covered 28 study sites, among which 7 sites were measured in August 2001 (wet season).
One IKONOS scene was acquired on February 27th, 2000 (dry season), covering 7 study
sites, among which 6 sites were measured in January 2002 (dry season). Figure 3-10
shows the relationships between the IKONOS and ETM+fc in wet and dry seasons.
Three series of data points are displayed: IKONOS and ETM+fc values in the dry season
(7 points), IKONOSfc in the wet season, but ETM+fc in the dry season (28 points), and

ground-measuredfc in the wet season, but ETM+fc in the dry season (12 points).

The 7 points of IKONOS vs. ETM+fc in the dry season are closely distributed along the
1:1 line, indicating again that the ETM+fc during the dry season is estimated with high
accuracy. The relationships of IKONOS + ground fc in the wet season vs. ETM+fc in the
dry season, however, is not linear. Due to the seasonal variation of forest greenness, the
fc in the wet season, especially in dry dipterocarps and mixed deciduous forests, is much
higher than the fc calculated during the dry season. The fc in the pine transition and
evergreen forests does not change much across the seasons. Both IKONOS and ground-
measured fc values in the wet season follow a similar trend and, therefore, can be

combined as ground truth.

60

As shown in Figure 3-10, the ground truth (H(ONOS and ground-measuredfc) from the
wet season is logarithmically related to the ETM+fc calculated from the dry season. A
simple logarithmic regression was processed and the regression line is shown in Figure 3-
10. With the obvious three outliers in the upper left of the figure, the expected value is
higher than the observed when fc is less than 50%. The squared correlation coefficient
(R2) of the regression is only 0.64. When the three outliers were removed, a
logarithmically curve fitting was also made with the following equation:

fcwe, = 0.30 * ln(fcd0, /100 + 0.09) + 0.91 when fc < 91%

3-11
few, 2 fem when fc Z 91% ( )

Here, fc is the expected fc in the wet season, and fem, is the observed fc from the

we!

ETM+ image in the dry season. Using Eq.3-1 l, the expected fc is only 94% when

we!

ETM+fc is saturated (100%), which is less than the ground truth. Therefore, fcwe, is

underestimated in very dense forests. Since there is not much difference between fc

we!

and fc.»). in dense evergreen forests, we set few, = fem, when the ETM+ estimated fc 2

91%.

The curve fitting is demonstrated in Figure 3-11. The measured data points were closer to
the curve fitting line when the outliers were removed. The squared correlation coefficient
(R2) for this curve fitting was 0.87, much higher than the simple logarithmic regression in

Figure 3-10. The errors of the three coefficients in Eq.3-11 were much lower than the
expected values. A )(2 goodness-of—fit was also done to this curve fitting and the 12 =

0.15. For a 3-parameter curve fitting with 37 samples (40 data points —— 3 outliers), the

61

degree of freedom equals 34 (37 — 3 model coefficients). At 95% confidence level, the

standard 1305,34 = 48.60. Since 12 is much less than 15053, , we accept the curve fitting

described by Eq.3-l 1.

With Eq.3-11, the ETM+ estimated fc in the dry season could be adjusted to the “true”
green forest fractional cover of the wet season, a more realistic biophysical attribute in
tropical forest ecosystems. Figure 3-12 is the adjusted fractional cover map. It is obvious
that the distribution is much less spatially variable than the one in the dry season (Figure
3-5). For dry dipterocarps and mixed deciduous forests, the green cover is much higher
than the cover calculated from the dry season. The fractional cover of evergreen forests
does not change much. Except for the non-forest areas like agriculture fields and
settlements whose cover is less than 20%, the green forest fractional cover in the
watershed ranges from 20% to 100%. The fractional cover of most forest areas in the

watershed is higher than 50% during the growing season.

3.6 Conclusion and Discussion

Forest fractional cover is a good indicator of forest integrity. Quantifying the fractional
cover in tropical forests is very important for evaluating deforestation and recovery. In
this study, a linear unmixing model in a vegetation index domain was built to estimate the
forest fractional cover distribution in the tropical forests of the Mae Chaem Watershed in
northern Thailand. The Modified Soil Adjusted Vegetation Index (MSA V1), which was
most linearly related with vegetation abundance, was chosen as the independent variable

in the model. Two endmembers, full-cover forest canopy and open area with bare soil or

62

senescent grass, were calculated through statistical analysis. With the model and the
ETM+ image acquired in the dry season, the forest fractional cover distribution in the

watershed was mapped.

The forest fractional cover distribution in the study area results from the different extent
of both human disturbance and seasonal variation. Along the elevation gradient in the
mountainous watershed, the forest types change from dry dipterocarps at low elevation,
to mixed deciduous, pine transition, dry evergreen, and finally moist evergreen forests on
the top of the Mount Doi Inthanon. The forest fractional cover increases along this
gradient (downhill to uphill) in a similar trend to the forest type change. The dry
dipterocarp forests are frequently burned for agriculture and regrow after the agriculture
fields are abandoned. Moreover, as deciduous forests, the dry dipterocarps vary
seasonally. During the dry season, the leaves of dry dipterocarp forests become brown
and fall off. Therefore, as estimated with the ETM+ image, the dry dipterocarps have a
lower fractional cover between 10-40% in the dry season. The mixed deciduous forests
are composed of both deciduous and evergreen species. Also, because of the relative
difficulty to access these types, the mixed deciduous forests are less disturbed by human
activities. As a result, their fractional cover is between 40-80% in the dry season. The
pine transition and evergreen forests are composed of only evergreen species. Similarly
inaccessible due to the high topography, they are not degraded much as the dry
dipterocarps and mixed deciduous forests. The fractional cover of these forests is

generally higher than 70% in the dry season, and the seasonal variation is minimal. For

63

the moist evergreen forests on the top of Mount Doi Inthanon, the fractional cover is

often saturated (i.e. 100%).

In each forest type, the fc distribution reveals different stages of forest degradation and
recovery in the study area. For example, there were isolated areas of clear-cutting in the
evergreen forests in the early 19905. The regrowth of these areas shows in the fractional
cover map with the cover varying between 20 to 80%, lower than the surrounding mature

evergreen forests.

The forest fractional cover estimated with the ETM+ image was validated by isolated
ground measurements in the watershed. The ground-measuredfc was calculated using the
GLA software to process hemispherical photographs taken on the ground with a fisheye
lens. The ETM+fc values were correlated with those at the 32 study sites measured in the
dry season with R2: 0.76. Similar to the trend of ETM+fc, the ground-measuredfc
increased with elevation as the forest types changing from dry dipterocarps to mixed
deciduous, pine transition, and finally dry and moist evergreen forests. Due to the
methodological difference, the ETM+fc was lower than the ground—measuredfc in low-
density forests such as dry dipterocarps, and higher than ground-measuredfc in dense
forests such as the tropical evergreen. The ETM+fc was saturated in moist evergreen

forests at high elevations.

The forest fractional cover estimated with the ETM+ image was further validated in a

large spatial area by one hi gh-resolution IKONOS image that was also acquired in the dry

64

season, 25 days after the acquisition of the ETM+ image. The IKONOS fc was calculated
by image classification and visual interpretation. With limited study sites covered by this
scene (7 sites only), the squared correlation coefficient between the ground-measuredfc
and the IKONOS fc was 0.97, indicating the high probability of the IKONOS image
serving as an accurate validation source in this study. Also at these study sites, the ETM+
fc was highly correlated with the IKONOSfc with R2: 0.96. The correlation line was
very close to the 1:1 line in the scatterplot. The ETM+fc was also compared with the
IKONOS derived fc in a larger dataset, in which 100 points in each forest type were
randomly selected for the purpose of comparison (in total of 400 points). In the ETM+ vs.
IKONOSfc scatter plot, these 400 points were loosely clustered along the 1:1 line. The

squared correlation coefficient was 0.70.

The ETM+fc estimated in the dry season can be adjusted to the “true” green cover with
ground measurements and IKONOS images from the wet season. A logarithmical
relationship was observed and a curve fitting model was built using the combined
ground-measured and IKONOS fc in the wet season and the ETM+fc in the dry season.
With this relationship, the ETM+fc map in the dry season was adjusted to wet season

conditions, a better representation of tropical forest canopy cover.
The successful estimation of forest fractional cover is an obvious improvement over

simple forest/non—forest binary classifications. Knowing the fractional cover distribution

in tropical forests provides a better way to measure both forest degradation and its

65

recovery. It is also very helpful for estimating forest biomass in tropical forests and

understanding their role in carbon sequestration and global change.

66

3.7 References

Adams, J. B., Smith, M. O. and Gillespie, A. R. (1993). Imaging spectroscopy:
interpretation based on spectral mixture analysis. 145-166. In: Remote
Geochemical Analysis: Elemental and Mineralogical Composition. Pieters, C. M.
and Englert, P.A. (Eds), Cambridge University Press, Cambridge.

Barros, A. C. and Uhl, C. (1995). Logging along the Amazon River and estuary: patterns,
problems, and potential. Forest ecology apd Management: 77, 87-105.

Carlson, T. N., Perry, E. M. and Schmugge, T. J. (1990). Remote estimates of soil
moisture availability and fractional vegetation cover for agricultural fields.
Agriculture, Forestry, and Meteorology: 52, 45-70.

Carlson, T. N. and Ripley, DA. (1997). On the relation between NDVI, fractional
vegetation cover, and leaf area index. Remote Sensing of Environment: 62, 241-
252.

Cochrane, M. A., Alencar, A., Schulze, M. D., Souza Jr., C. M., Nepstad, D. C.,
Lefebvre, P. and Davidon, E. A. (1999). Positive feedbacks in the fire dynamic of
closed canopy tropical forests. Science: 284,1833-1835.

DeFries, R. S., Hansen, M. C., Townshend, J. R. G., J anetos, A. C. and Lovelands, T. T.
(2000). A new global l-km dataset of percentage tree cover derived from remote
sensing. Global change biology: 6, 247-254.

Frazer, G. W., Canham, C. D. and Lertzman, K. P. (1999). Gap Light Analyzer (GLA),
Version 2.0: Imaging software to extract canopy structure and gap light
transmission indices from true-color fisheye photographs, users manual and
program documentation. Copyright © 1999: Simon Fraser University, Burnaby,
British Columbia, Canada, and the Institute of Ecosystem Studies, Millbrook,
New York, USA.

Gobron, N., Pinty, B., Verstraete, M. and Govaerts Y. (1999). The MERIS Global
Vegetation Index (MGVI): description and preliminary application. lntemational
Journal of Remote Sensing: 20 (9), 1917-1927.

Gutman G. and Ignatov, A. (1998). The derivation of the green vegetation fraction from
NOAA/AVHRR data for use in ntunerical weather prediction models.
International Joumal of Remote Sensing: 19(8), 1533-1543.

Hanan, N. P. and Prince, S. D. (1991). Spectral modeling of multicomponent landscapes
in the Sahel. lntemational Journal of Remote Sensing: 12 (6), 1243-1258.

Huete AR. (1988). A soil-adjusted vegetation index (SAVI). Remote Sensing of
Environment: 25, 295-309.

67

Huete, A.R., Justice, C. and Van Leeuwen, W. (1999). MODIS vegetation index, MODIS
algorithm theoretical basis document, NASA Goddard Space Flight Center,
Greenbelt.

Laurance, W. F ., Laurance, S. G., F erreira, L. V., Rankin-de Merona, J ., Gascon, C. and
Lovejoy, T. E. (1997). Biomass collapse in Amazonian forest fragments. Science:
278,1117-1118.

J asinski, M. F. (1990). Sensitivity of the normalized difference vegetation index to
subpixel fractional cover, soil albedo, and pixel scale. Remote Sensing of
Environment: 32, 169-187.

J asinski, M. and Eagleson, P. S. (1990). Estimation of subpixel vegetation cover using
red-infrared scattergrams. IEEE Transactions on Geoscience and Remote Sensing:
28 (2), 253-267.

Myneni, R. B., Asrar, G., Tanre, D. and Choudhury, B. J. (1992). Remote sensing of solar
radiation absorbed and reflected by vegetated land surfaces. IEEE Transactions on
Geoscience and Remote Sensing: 30, 303-314.

Myneni, R. B. and Williams, D. L. (1994). On the relationship between FPAR and NDVI.
Remote Sensing of Environment: 49, 200-211.

Pinty, B. and Verstraete, M. M. (1992). GEMI: a non-linear index to monitor global
vegetation from satellites. Vegetatio: 101, 15-20.

Price, J.C. ( 1992). Estimating vegetation amount form visible and near infrared
reflectances. Remote Sensing of Environment: 41, 29-34.

Qi, J ., Chehbouni, A, Huete, A. R. and Kerr, Y. (1994). A modified soil adjusted
vegetation index (MSAVI). Remote Sensinggf Environment: 48, 119-126.

Qi, J ., Marsett, R. C., Moran, M. S., Goodrich, D. C., Heilman, P., Kerr, Y. H., Dedieu,
G., Chehbouni, A. and Zhang, X. X. (2000). Spatial and temporal dynamics of
vegetation in the San Pedro River basin area. Agricultural and Forest
Meteorology: 105, 55-68.

Quarmby, N. A, Townshend, J. R. G., Settle, J. J ., Milnes, M., Hindle, T. L., and Silleos,
N. (1992). Linear mixture modeling applied to AVHRR data for crop area
estimation. International Journal of Remote Sensing: 13, 415-425.

Radeloff V. C., Mladenoff, D. J. and Boyce, M. S. (1999). Detecting Jack Pine budworm
defoliation using spectral mixture analysis: separating effects from determinants.
Remote Sensing of Environment: 69, 156-169.

Rahman, H., Pinty, B. and Verstraete, M. M. (1993). Coupled surface-atmosphere
reflectance (CSAR) model, 2, semi-empirical surface model usable with NOAA

68

advanced very high resolution radiometer data, Journal of Geophysical Research:
98(20), 781-801.

Settle, J. J. and Drake, N. A. (1993). Linear mixing and estimation of ground cover
proportions. lntemational Journal of Remote Sensing: 14, 1159-1177.

Souza Jr. C. and Barreto, P. (2000). An alternative approach for detecting and monitoring
selectively logged forests in the Amazon. lntemational Journal of Remote
Sensing: 21, 173-179.

Stone, T. and Lefebvre P. (1998). Using multi-temporal satellite data to evaluate selective
logging in Para, Brazil. International Journal of Remote Sensing: 19, 2517-2526.

Thailand Land Use and Land Cover Change Case Study (1997). Southeast Asia, IHDP,
IGBP, WCRP program.

Townshend, J. R. G. (1999). MODIS enhanced land cover and land cover change
product: algorithm theoretical basis documents (ATBD), version 2.0.

Tucker, J. J. (1979). Red and photographic infrared linear combinations for monitoring
vegetation. Remote Sensing of Environment: 8, 127-150.

Verhoef, W. (1984). Light scattering by leaf layers with application to canopy reflectance
modeling: the SAIL model. Remote Sensing of Environment: 16, 125-141.

Wessman, C. A, Bateson, C. A. and Benning, T. L. (1997). Detecting fire and grazing
patterns in tallgrass prairie using spectral mixture analysis. Ecological
Applications: 7, 493-511.

Wittich, K. P. and Hansing, O. (1995). Area-averaged vegetative cover fraction estimated
from satellite data. lntemational J oumgl of Biometeorology: 38, 209-215.

Wittich, K. P. (1997). Some simple relationships between land-surface emissivity,
greenness and the plant cover fraction for use in satellite remote sensing.
International Journal of Biometeorology: 41, 58-64.

Zeng, X., Dickinson, R.E., Walker, A. and Shaikh M. (2000). Derivation and evaluation
of global l-km fractional vegetation cover data for land modeling. Journal of
Applied Meteorology: 39(6), 826-839.

69

 

 

Mae Chaem

l
02/27/2000

 

 

 

 

(a) (b)
Figure 3-1 ETM+(band4+3+2, 02/02/2000) (a) and IKONS images (band4+3+2,
02/27/2000 and 10/09/2002) (b) in Mae Chaem Watershed. The dots in (b) represent the

study sites in two field trips.

70

 

 

 

 

- 1707-1991
[:1 1992— 275
[:1 2276— 2500

 

 

 

 

 

 

 

(a) (b)

Figure 3-2 DEM data with hillshade effect (a) and topographic correction with

 

 

 

 

Rahman’s BRDF model (b).

1.0 -
>< _
g 0'8 _._ GEMI
E 0 6 + NDVI
g ‘ + SAVI
"g —— MSAVI
‘5 0'4 ‘ + EVI
9 + MGVI
> 0.2 I

0.0 I l I l

o 1 2 3 4 5
. Leaf Area Index

Figure 3 -3 Dynamic ranges of the six vegetation indices calculated with simulated

spectral data in SAIL model.

71

 

 

 

 

 

Figure 3-4 MSAVI image in Mac Chaem Watershed.

 

  
 

-I
-

 

 

 

Figure 3-5 Forest fiactional cover map in Mac Chaem Watershed.

72

 

ETM+ estimated fc (%)
(02l02/2000)

 

 

 

 

 

  
  

   

0 I I I I
0 20 40 60 80 100
ground measured fc (%)
(01/I20-2712002)
(a)
100-» -
80 I

.3 aground
.2 I, a n ETM+

  

   

.
I59.

   

fine

          

 

 

 

 

 

   

 

dry dipter. nixed pine dry ever. rroist
decid. trans. ever.

0))
Figure 3-6 Comparison of ground-measured and ETM+ estimated fc: all forest types

combined (a) and by forest type (b).

73

200 a

 

 

 

 

 

 

 

150 ~
2
E
wt
g 100 -
.o
q.
o
g o forest(bright+dark)
a bare
50 ~ A brightopen
+ dark open
x shaded bare
0 water
0 l T l l
O 50 100 150 200

DN of band 3 (Red)

Figure 3-7 Feature space of different classes in the supervised classification of

IKONOS images.

74

 

 

100 [ I I l . l ' /$ 1 00
Z i ' /
—e—fc_ground_Jan : j j CG, / Q

  
 
 
 
 

90 “ Y=M0+M1‘X

 

 

 

 

 

 

 

’(‘cu‘ M0 32.38| ; [1‘1
3 80 M1 osozoelmgm _ g 80 Z
S R 0.98309' +
I I I - —h
In 5 f ____________ i ...... ; __ O
8 7° .‘ s 2 2 , 5 7° 5'
In . . : s ' o.
.0 60 , 60 m
.E I I C / I . . g
s i <1: F . . ’
.9 50 — --------- ------- . ---------- / -------- g ~~~~~~~ g IIIIIIIIII i ----------- 1 ---------- — 50 ‘8
1:3 3 . ,’ 3 3 a = 9 3
3 a ,/ 3 I .3 -I{}>—fc_ETM_Feb 11
U) f ‘ 3 ; 5 Y=M0+M1'X S

 

 

 

 

 

 

 

 

; ‘3" I . ; || M0 0.37416]

30 7' "' 3' " ' ‘ g‘ "l M1 1.o1o1|[ 30
I i II R 0.9804611

20 l .__ l l l l I 20

2o 30 40 50 60 70 80 90 100
IKONOS fc in dry season (Feb)

     
  

 

 

 

 

 

 

 

 

 

 

 

(a)
100
0 dry dipterocarps 00
a: mixed deciduous
’3‘ 8O - otransition zone g) Q
E X o 0%
A I
r: evergreen q) o
3 O x o
8 60 ~ 0 g
'” 6o
5 x
8 4O -
+
E *°
In 20 -
R2: 0.70
O I I I l
O 20 40 60 80 100
IKONOS fc (dry season) (%)
(b)

Figure 3-8 Scatter plots of ground-measured and IKONOS estimated fc in 6 sites (a),

and the ETM+ and H(ONOS estimated fc in 400 randomly selected points (b).

75

 

, ETM+ image (4+3+2), 02/02/00
IKONOS 1mage(4+3+2), 02/27/00

 

   

IKONOS estimated fc ETM+ estimated fc

Figure 3-9 IKONOS and ETM+ subset images and their fc maps.

76

100

 

 

 

 

 

 

 

 

g 80 —
.2
8
z 60 -
O
E
'2
a 40 —
1:
§
2 20 ,x” Ofc_ground_wet
m I
x” 2 X fc__ikonos_wet
// R ”0'64 Ifc_ikonos_dry
O I I I I I
O 20 4O 60 80 1 00
ETM+ to in dry season (%)
Figure 3-10 The ground-measured, H(ONOS and ETM+ estimated fc scatterplot in dry

and the wet seasons.

 

 

 

 

c: 0.8 r
O
U)
(U
Q)
m .
5 0.6 -
3
.E
8
E 0.4 - . ‘ ~ -
8 y = m1 + m2 * In(M0+m3)
§3 Value Error
m m1 0.907 0.017
C2) 0.2 m2 0.298 0.049 —
0 m3 0.086 0.053
5 Chisq 0.145 NA
R 0934 NA
0 4 l
0 0.2 0.4 0.6 0.8 1
ETM+ fc in dry season
Figure 3-11 The curve fitting model to adjust fc in the dry season to the wet season.

 

 

 

 

 

 

 

 

 

 

 

 

77

 

   

. ._-.I
A
‘5
u

‘II-IUIM

 

 

 

Figure 3-12 Forest fractional cover map adjusted from the dry season to the wet

season.

78

Chapter 4
Estimation of Leaf Area Index with Fractional Cover Data in

Tropical Forests

4.1 Introduction

Leaf area index (LAI), the total single-sided leaf area per unit ground area, is another
important biophysical attribute of forest ecosystems. It controls the partitioning of
incoming solar energy into sensible and latent heat fluxes and is a good indicator of
energy, gas and water exchanges between land surface and atmosphere. Because of its
important rule in the physiological processes such as photosynthesis, transpiration and
evapotranspiration, LAI has been widely used to parameterize and validate models of
ecosystem functioning, biosphere-atmosphere interaction, vegetation growth, net primary
production and other environmental processes at landscape to global scales (Sellers and

Schimel 1993).

The direct measurement of LA] involves cutting and measuring the leaves in a small area.
It provides the most accurate measurements. However, this method is very time
consuming and destructive. As a result, this method is not recommended for long-term
large scale measurements. An alternative way is to indirectly measure LA] with
commercial instruments like the Li-Cor LAI-2000 and fisheye photography (Welles
1990). The accuracy of these indirect methods highly depends on the sunlight radiation

model that these instruments use and the atmosphere conditions at the time of

79

measurement. Both direct and indirect LAI measurements provide ground data for the

validation of remote sensing methods at large scales.

One of the primary approaches to estimate LAI using remote sensing imagery is the
polynomial regression between LAI and vegetation indices (VIs) from radiometric
measurements (Qi et al. 2000; Baret 1995; Best and Harlan 1985; Peterson et al. 1987).
This relationship is non-linear (Tucker et al. 1981) and saturates at LAI 2 3 depending on
the type of vegetation index used, the canopy type studied, and the experimental
conditions (Spanner et al. 1994; Baret and Guyot 1991). This empirical LAI— VI
relationship is also affected by leaf pigments, internal structure, orientation distribution,
leaf clumping, woody reflectance, and heterogeneity in tree height and tree gaps (Turner

et al. 1999).

Another commonly applied approach is to estimate LA] by the inversion of bi-directional
reflectance distribution function (BRDF) models (Pinty et al. 1990; Goel and Thompson
l984a,b; Qi et al. 1995). So far, more than a dozen BRDF models have been developed
for various types of surface such as cropland, grassland, and bare soil surfaces (Strahler
1994). Some of these models are mathematically invertible so that biophysical attributes
can be calculated. However, a BRDF model often has multiple input parameters. For
large-scale model inversion, some of the essential input forest parameters are often
unavailable and have to be simplified. This simplification often results in unreasonable

LAI values for forests with high heterogeneity of distribution and topography.

80

Because of the topographical difficulty to access, it is impossible to measure LA] with
LAI-2000 instrument at study sites in the watershed. However, as discussed in Chapter 3,
the forest fractional cover was measured with a fisheye camera at each study site, and the
forest fractional cover product from remotely sensed data was validated. In this chapter, a
simple regression model was developed to retrieve forest LAI distribution in the study
area from forest fractional cover product in Chapter 2. The adjustment of LA] product

from the dry season to the wet season was also performed.

4.2 Experimental Design and Field Measurements

4.2.1 LAI-2000 and fisheye photographs

A variety of instruments have been developed which apply radiation models to indirectly
measure the vegetative canopy attributes. When passing through the canopy, solar
radiation is mainly alternated by foliage amount and orientation. The spatial distribution
of the foliage is assumed to be random in the radiation model. The common strategy of
the radiation model is to describe how radiation is affected as it passes through a canopy
with some well-defined, geometrical canopy attributes, then make appropriate radiation
measurements and invert the model to estimate the value of these attributes (Welles and
Cohen 1996). The success of the radiation model is thus determined by how closely the

real canopy conforms to the idealized one.

It should be noted that, when passing through a canopy, the radiation is not only affected

by leaves, but any other opaque object like stems, fruits, and branches. As a consequence,

81

 

 

although LA] is defined as leaf area index, the LA] estimates with indirect methods should

be better described as plant projected area index (Welles and Norman 1991).

The LAI-2000 Plant Canopy Analyzer developed by LI-COR company is the most
commonly used commercial instrument to measure leaf area index and some other
vegetative canopy attributes. It uses hemispherical optics and a ringed detector that
simultaneously measures diffiise radiation in five distinct angular bands about the zenith
(Welles and Cohen 1996). For a single-sensor LAI-2000 equipment, a reference reading
is made at the open area above the canopy, instantly followed by one or more below
canopy readings. Sometimes another open reading is made after the below canopy
readings. For each ring, the canopy’s gap fraction at the view angle of that ring is
assumed to be the ratio of the below and above canopy readings (Welles and Norman,
1991). With a sunlight radiative transfer model, the LAI, foliage amount density, and
foliage orientation can be calculated. The LAI-2000 has proven to be very useful in short

canopies like agricultural crops and grasses.

However, for tall canopies like forests, the application of LAI-2000 is limited. For each
LAI measurement, a simultaneous reference reading is needed in an open area that is
large enough to avoid the effects of forest scattering. In this study, all of the study sites
are located in the deep forests in Mae Chaem Watershed and, therefore, it was impossible

to measure LA] with the LAI-2000 instrument during the two field trips.

82

 

As discussed in Chapter 3, the digital photos with fisheye lens were taken at the study
sites and processed with the Gap Light Analyzer (GLA) software to calculate the gap
fraction. A similar radiation model was also applied in GLA to calculate LAI over certain
zenith angles. For example, LAI-4ring is the effective leaf area index integrated over the
zenith angles 0° to 60°, and LAI-5ring is the effective leaf area index integrated over the
zenith angle 0° to 75° (Frazer et al. 1999). The forest fraction (in 0-1 scale) is 1.0 — gap
fraction. As shown in Figure 4-1, LA] values at all study sites from the two field trips are
exponentially related to the forest fractional cover. In sparse forests, there is more open
space at larger zenith angle (60°-75°) and, therefore, the values of LAI-5ring are lower
than LAI-4ring. Contrarily, in dense forests, the fractional cover is higher than gap

fraction and therefore, the values of LAI-5ring are higher than LAI-4ring.

Similar to the LAI-vegetation index relationships, the LAI-forest fractional cover curve
begins to saturate at LAI around 3.0. However, the LA] values measured with fisheye
photos are unreasonably low in Figure 4-1. Even for the dense moist evergreen forests
with fractional cover higher than 95%, the LAI values are less than 4.0. The radiation
model in GLA software assumes that the atmosphere conditions above the forest canopies
are constant at the time of measurements at all of the study sites, which in fact varies
greatly depending on different seasons, amount of cloud cover or cloud properties. It

could be the possible reasons for the low LAI values in Figure 4-1.

83

4.2.2 LAI ground measurements

From the radiation model in GLA software, the LAI and fractional cover are
logarithmically correlated (Figure 4-1). However, the LA] values in GLA software with
fisheye photos are underestimated. While the forest fractional cover can be measured
with fisheye photos, LAI-2000 provides a good measurement of LA]. Since it is
impossible to use LAI-2000 in Mae Chaem Watershed, I made additional measurements
during the third field trip in the hardwood forests in northern Michigan in October and

November 2002.

Between October 1 and November 12, 2002, four forested areas were visited and totally
42 study sites were measured (Table 4-1). These sites were close to river, lakes, or large
open areas so that the LAI-2000 requirements were met. At each study site, two open
readings and five under-canopy readings were recorded with LAI-2000 to measure LAI,
following the TIIIII sequence. Three digital photos were taken with the fisheye lens
and were processed using GLA software to calculate fractional cover and LAI. As shown
in Figure 4-1, for forests with fractional cover lower than 95%, the LAI-4ring and LAI-
5ring values do not change much when calculated in GLA software. Therefore, only LAI-

4ring readings are compared in this analysis.

The study sites visited in early October simulates the canopy conditions of tropical
forests with higher cover, such as mixed deciduous and tropical evergreen in Mae Chaem
Watershed (in the dry season). The forests in northern Michigan are typical deciduous

broadleaf forests. In late October and early November, the leaves become yellowish and

84

begin to drop off which is analogous for canopy conditions of dry dipterocarps and sparse
mixed deciduous forests (in the dry season) in Mae Chaem Watershed. Therefore, all
these study sites at different dates could approximately represent the ground
measurements in our study area in Mae Chaem Watershed and may provide ground data
for the LA] estimation. It should be noted that, the structure of tropical forests and
hardwood temperate forests are different. Some errors may be introduced with this

approximation.

The LA] values from LAI-2000 and fisheye photos over study sites in northern Michigan
were compared in Figure 4-2. The 1:1 line was also drawn in the figure to demonstrate
the variation of the two LAI measurements. Although LAI values from fisheye photos are
a little higher than the ones from LAI-2000 for forests at low fractional cover, they are
much lower for forests with higher fractional cover. The LA] values from fisheye photos
increase very slowly when forest cover is higher. Among all the study sites with different
dates (leaf-on and leaf-off), the LAI-2000 measurements range from 0.88 to 5.04, while
the LAI values from fisheye photos (LAI 4-ring) only change from 1.39 to 2.66. Figure 4-

2 confirms that fisheye photos are not a good source of LAI measurements.

4.2.3 LAI ~ forest fractional cover relationship

As discussed in Chapter 2 and 3, fisheye photos provide reasonable forest fractional
cover measurements. Replacing LAI values in Figure 4-1 with the LAI-2000
measurements, the relationship between LA] and forest fractional cover was shown in

Figure 4-3. Most forests in Michigan are old secondary or natural forests and there was

85

no sparse forests found at the study sites. Therefore, the forest fractional cover values at
all study sites in Figure 4-3 are higher than 0.6. The LAI and forest fractional cover

values fit an exponential relationship.

4.3 Model Development and Results

4.3.1 A modified Gaussian regression model

A modified Gaussian curve fitting model was developed with the ground measurements
in northern Michigan. In Figure 4-4a, both the LAI and forest fractional cover values are
extended to the origin to represent the sparse forests with fractional cover from 0 to 0.6, a
simulation of dry dipterocarp forests in the dry season in Mae Chaem Watershed in
Thailand. The residuals of the regression are plotted in Figure 4-4b. Among the 42 study
sites, there are only four sites with absolute residual values larger than 1.0, and all these 4
sites have fractional cover around 0.85 or higher. Most of the residuals are distributed
between i0.5. Figure 4-4 also indicates that the higher the forest fractional cover, the
higher the residuals. It is reasonable because the LAI-2000 measured LA] is a measure of
plant projected area instead of green foliage area. As the fractional cover becomes higher,

there introduces more error in the LAI—2000 readings.

The regression equations are expressed as:

LAI = 0.217 + 0.058 * e‘5’578‘If"°-°33”> 0.6 < fc $1.0
LAI = 0.85 * fc 0.0 < fc s 0.6

(4-1)

Here fc is the forest fractional cover at the scale of [0,1]. There was no ground data when

fc is lower than 0.6. However, from Figure 4-2, the relationship between LAI and

86

fractional cover is almost linear when the fractional cover is low. Therefore, a linear

relationship in Eq.4-1 was used for 0.0 < fc S 0.6 .

A x2 goodness-of-fit test was examined to the modified Gaussian regression. The x2 was
calculated from:

N (LAIO, —LA1,, )2
12 =2 bLAI p (4-2)

i=1 exp

 

Here N is the total number of ground measurements, N=42 in the study sites in the

northern Michigan. LAlob is LAI-2000 measurements, LAI,Xp is the modeled LAI values

S

from Eq.4-1. At the degree of freedom of 38 (N-4 coefficients in the model) and

confidence level of 95%, the critical value, 130,38 , is 55.76. As shown in Figure 4-4a,

the x2 value in this test is 9.712, far less than 130538 . Therefore, the null hypothesis

cannot be rejected and the modified Gaussian regression model is valid. Moreover, the
correlation coefficient (R) between modeled and measured LAI values is 0.895, also

indicating that the model could characterize the LA] ~ forest fractional cover relationship

very well.

4.3.2 LAI estimation

The forest fractional cover distribution in Mae Chaem Watershed in Thailand has been
mapped in chapter 3. With the modified Gaussian regression model, the LAI distribution
could be mapped. Figure 4-5 is the LA] distribution in the study area in dry and the wet
seasons, resulting from the forest fractional cover distribution in both seasons. The open

areas that are non-forests are masked out.

87

Similar to that of forest fractional cover in Chapter 3, the seasonal change of LA]
distribution is significant. In the dry season, the leaves on deciduous trees become
senescent and were partially off and, therefore, most of the dry dipterocarps and mixed
deciduous forests have very low LAI values (less than 1.0). The tropical evergreen forests
have LAI values 4.0 or higher. In moist evergreen forests, the LA] values saturate at
around 10.9. In Figure 4-5a, there is a very narrow zone between mixed deciduous and
tropical evergreen forests where the LA] values are around 1.0 (corresponding to the
fractional cover 70-80%). It could be the pine zone transition where the seasonal change
is less significant than deciduous or mixed deciduous forests, but more significant than

tropical evergreen forests.

In the wet season, the tree leaves are green and healthy, and most of the dry dipterocarps
and mixed deciduous forests have LAI higher than 1.0 (Figure 4-5b). The narrow zone in
Figure 4-5 a merges to the tropical evergreen forests with LAI greater than 4.0. The LA]
increment in tropical evergreen forests is not as significant as other forest types, but there
are more saturated areas than those in the dry season. The open areas in Figure 4-5b are
human settlements (town, villages) and agricultural fields. Surrounding these open areas,
there are some forests that have LAI values far less than 1.0 in the wet season. Most of
these forests are sparse young regrowth of dry dipterocarps and are highly disturbed by
human activities. Therefore, the LAI values in these forests are very low even in the wet

season.

88

4.4 Validation

As discussed earlier, because of the physical difficulty to access the study sites and the
lack of open areas in deep forests, it is difficult to measure the leaf area index over the
study sites in tropical forests with LAI-2000. As a result, it is impossible to directly

validate the modeled LAI distribution mapped in this study.

As an alternative, an indirect validation was made in this section. The leaf area index
(LAI-4ring) can be calculated from the fisheye photos over the study sites in Mae Chaem
Watershed in northern Thailand. As shown in Figure 4-2, although the LAI from fisheye
photos were not as sensitive as that of LAI-2000, there was a linear relationship between
these two LAI values in the LA] range between 0 and 6. With this linear relationship, the
LAI-4ring values at the study sites in the watershed in Thailand could be adjusted to the
equivalent values of LAI-2000 measurements. These adjusted LAI measurements were
assumed as ground “truth” to validate the modeled LAI distribution in this study. To be
consistent with the modeled LAI distribution in the dry season, only the fisheye LAI-4ring

values at the 32 study sites during the second field trip were adjusted.

The comparison of the modeled and adjusted measured LAI values at the study sites in
the watershed was shown in Figure 4-6. In the LAI range between 0 and 4, the modeled
LAI fitted with them well and the data were scattered along the 1:1 line. The adjusted LAI
was underestimated when the ground LAI was higher. The adjusted LAI was still less than

4 when the modeled LAI reached 6. In moist evergreen forests, the modeled LAI values

89

were saturated at 10.9 as shown in the LA] map, while the adjusted measured LAI values

were only around 5.

The adjusted measured LAI from fisheye photos was not the real ground truth. Since the
fisheye measured LAI was adjusted from that of LAI-2000 measurement when LAI was
less than 6.0, it did not provide good validation when the LA] was higher than 6.0 in the
watershed. When LAI was less than 6.0, the adjusted measured LAI could be applied for
the purpose of validation. The linear regression line between modeled and adjusted
measured LAI was very close to the 1:1 line and the square correlation coefficient R2 =

0.67 (Figure 4-7).

4.5 Conclusion and Discussion

In this chapter, a modified Gaussian regression model was built to simulate the
relationship between LAI and forest fractional cover in the study area. The LAI values
were measured from the LAI-2000, and forest fractional cover values were from fisheye

photos as discussed in Chapter 3. In the x2 goodness-of-fit test, the actual x2 from the
model was only 9.71 , far less than the critical 16.05.33 (55.76) at the 95% confidence level.
The correlation coefficient between the modeled and ground-measured LAI values was as

high as 0.90. Most of the residuals distributed between i0.5 and only 4 out of the 42

ground measurements had residuals higher than 1.0. All these statistical analyses
supported that the model was valid and could be applied to estimate LAI distribution with

forest fractional cover product in Chapter 3.

90

The LA] distributions in both dry and wet seasons were mapped in this study. The
seasonal change of the LAI distributions was significant. Most of the dry dipterocarps and
mixed deciduous forests have LAI values less than 1.0 in the dry season while higher than
1.0 in the wet season. The LAI values of the tropical evergreen forests were always
around 4.0 and higher in both seasons, but there were more areas that were saturated
because of the forest fractional cover saturation. The LAI values in saturated areas were
higher than 10.0. A narrow zone with LAI around 1.0 was shown in the LA] map in the
dry season, possibly the pine transition zone whose seasonal variation was lower than
deciduous forests but higher than tropical evergreen. On the other hand, some forests
around the open areas (villages, agricultural areas) have very low LAI values even in the

wet season, indicating intense human disturbances in these forests.

The method in this study provided a new approach to map LAI with forest fractional
cover which could be estimated from remotely sensed data. It should be noted, however,
the ground data were not measured in the study area in Thailand. Because of the
topographical difficulty, as well as the limited open area in deep tropical forests, it is
impossible to make LAI—2000 measurements in mountainous tropical forests in Mae
Chaem Watershed. Instead, some deciduous forests in northern Michigan in different
seasons (leaf—on and leaf-off) were selected to approximately represent the tropical
forests in the study area. Since the gap fraction is the major key to calculate LAI and
forest fractional cover values, this alternative is acceptable if the fractional cover has

large range of variation, e. g. from sparse to dense forests at the study sites in northern

91

Michigan. The ground measurements in this study lacked of forest samples with sparse

forests that should be considered in the future research.

In this study, it was found that although the GLA software also gave the readings of LAI
(4-ring or 5-ring), these LAI values are significantly underestimated. LAI-2000
measurements are more realistic than LAI values from GLA with fisheye photos.
Theoretically, the GLA/fisheye photo and LAI-2000 apply the same radiation model. The
primary difference is that fisheye photo assumes that the atmospheric conditions are
constant at all seasons and under all cloud cover properties, whereas the LAI-2000 makes
one or two open readings simultaneously with the canopy readings. The gap fraction of a
fisheye photo is calculated with binary image clustering techniques. The gap fraction of
LAI-2000, however, is the ratio of the canopy readings to the open readings. If this
difference can be compensated for the fisheye photo and GLA software, it is possible to
make reasonable LAI as well as fractional cover measurements with fisheye photos. Then
digital cameras with fisheye lens will become a more economic and efficient tool for
ground measurements in tall canopies, especially the dense mountainous forests where

LAI-2000 is in limited usage.

92

4.6 Reference

Baret, F. (1995). Use of spectral reflectance variation to retrieve canopy biophysical
characteristics. In Advances in Environmental Remote Sensing (M. Darson and S.
Plummer, Eds. ), John Wiley and Sons, Inc., 34-51.

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

Best, R. G. and Harlan, J. C. (1985). Spectral estimation of green leaf area index of oats.
Remote Sensing of Environment: 17, 27-36.

Frazer, G. W., Canham, C. D. and Lertzman, K. P. (1999). Gap Light Analyzer (GLA),
Version 2.0: Imaging software to extract canopy structure and gap light
transmission indices from true—color fisheye photographs, users manual and
program documentation. Copyright © 1999: Simon Fraser University, Burnaby,
British Columbia, Canada, and the Institute of Ecosystem Studies, Millbrook,
New York, USA.

Goel, N. S. and Thompson, R. L. (1984a). Inversion of vegetation canopy reflectance
models for estimating agronomic variables. IV: Total inversion of the SAIL
model. Remote Sensing of Environment: 15, 237-253.

Goel, N. S. and Thompson, R. L. (1984b). Inversion of vegetation canopy reflectance
models for estimating agronomic variables. V: Estimation of LAI and average leaf
angle using measured canopy reflectances. Remote Sensing of Environment: 16,
69-85.

Peterson, D. L., Spanner, M. A, Running, S. W. and Teuber, K. B. (1987). Relationship
of thematic mapper simulator data to leaf area index of temperate coniferous
forest. Remote Sensing of Environment: 22, 324-341.

Pinty, B., Verstraete, M. M. and Dickinson, R. E. (1990). A physical model of the
bidirectional reflectance of vegetation canopies, 2: Inversion and validation.
Joumalof Geophysical Research: 95(D8), 11,767-11,775.

Qi, J ., Cabot, F., Moran, M. S. and Dedieu, G. (1995). Biophysical parameter estimations
using multidirectional spectral measurements. Remote Sensing of Environment:
54, 71-83.

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

Sellers, P.J. and Schimel, D. (1993). Remote sensing of the land biosphere and
biochemistry in the EOS era: science priorities, methods and implementation -

93

EOS land biosphere and biogeochemical panels. Global and Planetary Change: 7,
279-297.

Spanner, M., Johnson, L. and Miller, J. (1994). Remote sensing of seasonal leaf area
index across the Oregon transect. Ecological Applications: 4(2), 258-281.

Strahler, A. H. (1994). Vegetation canopy reflectance modeling - recent developments
and remote sensing perspectives. Remote Sensing Reviews: 15, 179-194.

Tucker, C.J., Holben, B. N., Elgin, J. H. and McMurtrey, E. (1981). Remote sensing of
total dry matter accumulation in winter wheat. Remote Sensing of Environment:
1 1, 171 -l 90.

Turner, D. R, Cohen, W. B., Kennedy, R. E., F assnacht, K. S. and Briggs, J. M. (1999).
Relationships between leaf area index and Landsat TM spectral vegetation indices
across three temperate zone sites. Remote Sensingof Environment: 70, 52-68.

Welles, J. M. (1990). Some indirect methods of estimating canopy structure. Remote
Sensing Reviews: 5(1), 31-43.

Welles, J. M. and Norman, J. M. (1991). Instrument for indirect measurement of canopy
architecture. Agronomy Journal: 83, 818-825.

Welles, J. M. and Cohen, S. (1996). Canopy structure measurement by gap fraction
analysis using commercial instrumentation. Journal of Experimental Botany:
47(302), 1335-1342.

94

Table 4-1 Study sites in northern Michigan.

 

 

 

 

 

 

GPS Location .
Study area Date Study srtes

(UT M, WG884)
Muskegon River
Watershed (552741, 4789298) 10/1/2002 5
Burchfield Park (697576, 4720093) 10/2/2002 5

' ' 10/5/2002 16

Rose Lake erdlrfe (716891, 4742766)
Comm/anon 1 1/12/2002 10
MSU baker (706798,4732483) 1 1/12/2002 6
woodlot

 

 

 

 

 

 

95

 

x LAI-4ring

 

 

o LAI-5ring , i
I
I

Leaf Area Index
N
823%

a. g
0 I . a I

0.2 0.4 0.6 0.8 1 .0
forest canopy cover

 

 

Figure 4-1 Relationships between LAI (5-ring and 4-ring) with forest fractional

cover. The fractional cover is in the range of [0,1].

 

LAI (Fisheye)
OD

 

0 1 2 3 4 5 6
LAI (LAI-2000)

Figure 4-2 Comparison of LAI measurements from LAI-2000 and fisheye photos.

The 1:1 line is also drawn in the plot.

96

LAI (LAI2000)

 

 

 

 

5 ............................... 5 -
I g
’8" 4 _ .................................................. .00 “0:10. H
O @Oog
a 10 %>9
< 3 _ ...................................................... j. ............. A
:L O ; O .
_ ; OIO O I
3 2 _ , 06m@@QL.—no a
8 C?
1 __.. o .o ........ 9...: ........ 'I
O I I I I I
0.65 0.7 0.75 0.8 0.85 0.9 0.95

forest canopy cover (Fisheye)

Figure 4-3

LAI ~ forest fractional cover relationships measured with LAI-

2000 and fisheye photos in northern Michigan.

 

0 0.2 0.4 0.6
forest canopy cover
(a)
Figure 4-4

0.8

residual

97

 

 

 

 

1.5 I I f I I
I O
1 ... ,.. . 0.. NJ
‘ o
o
05 _ ................. I. Q .0 .Q a
i 00 o
.' O O
0 _ O 8 O . .. .. BQQ . ‘I
O 0 c0) 5 I
. O 0 @(g): 0 aogo
-0.5 - , -- ~ :~ - I
-1 s .. ..
O o
.15 i I I I i
0.65 O 7 0.75 0.8 0.85 0.9 0.95

forest canopy cover

(19)

A modified Gaussian curve fitting (a) and its residual plot (b).

 

 

 

 

 

 

 

 

 

(b)

Figure 4-5 LAI distribution in the study area in dry (a) and wet (b) seasons.

 

modeled LAI

 

 

 

O I I l I I

 

0 2 4 6 8 10 12
adjusted fisheye LAI

Figure 4-6 Comparison of modeled and adjusted measured LAI over the study sites in

the watershed (second trip).

 

 

 

 

6 O
5 ~ 0
7t o
4 -
_l
g 3 O .
E ° ’
E 2 - . o. o
1 _
o 3: o o R2=0.67
0 1 2 3 4 5 6
adjusted fisheye LAI

Figure 4-7 Correlation between modeled and adjusted measured LAI over the study

sites with LAI <6.

99

Chapter 5
A Microwave/Optical Synergistic Canopy Scattering Model

and its Inversion to Estimate Forest Structure

5.1 Introduction

Forest canopy structure is the combination of forest texture (the qualitative and
quantitative composition of the vegetation as to different morphological elements) and
forest structure (the spatial arrangement of these elements) (Barkman 1979). Forest
structural parameters are important input variables in ecosystem functioning and forest
harvest models. The understanding of forest structure is also necessary in accurate
estimation of forest biomass. Depending on the study interest, the forest canopy structure
can be described at several levels of integration: forest/non-forest at large scales, forest
types in different communities, within-forest patches (e.g., successional developments
phases, leaf area index), individuals (e. g., species and tree size, height), plant parts (e.g.,
crown and stem), and plant organs aboveground (leaves, branches, flowers, etc)
(Bongers, 2001). From this definition, both forest fractional cover and leaf area index are
structural parameters. They are described in Chapter 3 and Chapter 4 because they are all
green leaf properties and have been intensively studied using optical remote'sensing. The
forest structural parameters described in this chapter are mainly woody structural
properties such as tree height, stem height, diameter at breast height (DBH), and forest

stand density.

100

Optical remote sensing techniques have been applied in the retrieval of forest structural
parameters. Tall vegetation species create larger shadows, but they also expose a larger
portion of their vertical structure to the sensing systems when viewed from off-nadir
directions. Consequently, optical sensors with off-nadir view angles can be used to study
the vegetation biophysical structures (Qi et al. 1995). Li and Strahler (1992) built a
geometric-optical model in which a tree canopy was a collection of individual geometric
objects that cast shadows on a contrasting background: sunlit crown, shaded crown, sunlit
background and shaded background. In this model, the three-dimensional structure of the
canopy is the primary factor influencing the reflectance from the canopy (Strahler and

J upp 1990).

A good way of forest structure retrieval with optical remote sensing data is the inversion
of the geometric-optical model (Woodcock et al. 1994; Shimabukuro and Huemmrich
1995). Scarth and Phinn (2000) defined forest structure as the horizontal and vertical
distribution of components within a plant community. They applied TM images, DEM,
and field measurements in an inverted Geometric-Optical model (Li and Strahler 1992) to
estimate crown diameter, tree density, crown cover projection, and the “treeness”
parameter. Recent developments in high spatial resolution remote sensing techniques
leads to small scale characterization of canopy structure such as small forest gaps and
individual crown characteristics. These structural properties are important in monitoring

forest changes resulting from selective logging activities (Bongers 2001).

101

Microwave remote sensing techniques also have been applied to retrieve structural
information in temperate forests. Previous studies have demonstrated that forest structural
variation may have a substantial effect on L-band backscatter of forest stands with same
biomass (Imhoff 1995). The spatial pattern such as inter-tree gaps and tree clumping in
uneven forests may create significant differences in radar backscattering at finer

resolution (Sun and Ranson 1998).

The forest structural information retrieved from SAR backscattering is wavelength-
dependent. Signals with shorter wavelength (e. g. X- and C-bands) mainly contain
information in crown layers. More information about woody structures and biomass
could be retrieved from longer wavelengths (e. g. L- and P-band) because the signals can
penetrate the crown layer (Le Toan et a1. 1992). The backscattering is also polarization-
dependent. The backscattering in each polarization contains different structure
information. HH-polarized backscattering is dominated by the ground-canopy interaction
mechanism while direct crown backscattering contributes more significantly to the V\/-
and HV-polarized backscattering (McDonald and Ulaby 1993). Karam et al. (1995) also
showed that the HV backscattering coefficient was due almost solely to the long branches
at all ages. Therefore, the dynamic range observed in HV polarization reflected the

physiological changes of long branches as the canopy age increased.
The phase difference (the phase of the complex polarimetric coherence) in polarimetric

SAR data, resulting from the crown attenuation and stalk-ground double bounce, contains

much forest structural information (Ulaby et al. 1987). Phase difference could be

102

calculated in a polarimetric backscattering model (Ulaby et al. 1990). Hoekman and
Quifiones (2002) proposed a new method of polarimetric coherence decomposition

instead of power to relate polarimetric signal and forest structures.

Structural information can also be described by image texture, the spatial variability of
image tones that describes the relationship between elements of surface cover (Wulder et
al. 1998). For images over forests, the variation in texture is related to changes in the
spatial distribution of forest canopies. The medium to coarse resolution optical remote
sensing imagery often has a smooth texture so that the information in texture variation is
limited. However, the high resolution imagery such as IKONOS is able to reveal detailed
variation of image texture. These data can be used to detect internal stand shade
conditions as mutual shadowing that is more dominant in mature forest than secondary
regrowth. SAR signal penetrates forest canopies much deeper than sunlight and therefore,
contains more structural information than optical imagery. Luckman et al. (1997) showed
that coefficient of variation (CV) of SAR backscattering intensity exhibited the least
variation within each regrowth age, and was the best method to measure texture. Aside
from the deficiency of speckle noise, the standard deviation of SAR intensity and SAR

image texture is useful in forest structure retrieval (Manninen and Ulandder 2001).

In this chapter, a microwave/optical synergistic radiative transfer model was developed
and the backscattering from forest components was simulated while the leaf properties
were measured from optical remotely sensed data. The forest structures were the input

parameters of the model. The model was validated with J ERS-l SAR and VNIR data and

103

 

ground measurements. Then, with a nonlinear least square (LS) optimization method, the
model was inverted and the forest structures estimated. Finally, an error analysis was

performed to examine the accuracy of the model inversion.

5.2 Model Development

The forest can be simplified as a three-layer vegetation composition: leaf canopy,
branch+leaf canopy, and trunk atop of a rough soil surface (Figure 5-1). For different
forest types, the components in each layer and the values of their biophysical parameters
may vary. The saplings and seedlings and the grass/shrub underneath the forest are not
considered in the simulation. The backscatter intensity is the result of additive

contribution from the components in all layers:

0' = O-layerl + O-IayerZ + O-Iayer3 + 050i! (5-1)

total
In a 2nd-order solution of radiative transfer equations, the total backscatter consists of
surface scattering from soil underneath, volume scattering from leaves, branches, trunks,
interaction between forest components and soil surface, and 2nd-order non-coherent
scattering:

0' : 050i] + Gleaf + abranch + 0'

total trunk

(5-2)

+ aleaf—soil + O-branclz-soil + O- + U

trunk —soil noncoherenr

The microwave canopy scattering model developed by Karam et al. (1995) had a
detailed explanation of the scattering properties of the canopy-related components in
, in Eq.5-2 is a very simple lst-order

forests. However, in Karam’s model, the 0'

solution of radiative transfer equations, assuming the soil as a continuous and slightly

rough dielectric surface. It is thus insufficient to simulate radar backscatter under various

104

soil roughness and moisture conditions in different forest types, particularly in sparse,
young, secondary forests in which the soil surface plays an important role in determining

the total backscatter.

Three modifications to Karam’s model were made in this study. The first was to
introduce an integral equations model (IEM) bare surface scattering model (Fung et al.
1992) into the canopy scattering model. The second was to link the model to optical
remote sensing variables. Since Karam’s model was originally developed for conifer
forests, the probability distribution function (PDF) of each of the forest components was

also modified to fit the characteristics of tropical forests in the study area.

5.2.1 Modified Karam-[EM model

The first-order solutions of radiative transfer equations for backscattering coefficients (in
power units) of the components in Eq.5-2 are described in Eq.5-3 to Eq.5-12. The
second-order solutions of the non-coherent backscattering are much smaller than the

first-order ones and, therefore, the equations are not listed in this study.

5.2.1.1 Soil surface scattering

Soil surface backscattering in forests depends on its dielectric constant, surface
roughness, and attenuation from the forest canopies above the soil surface. Surface
roughness can be described with a root-mean-square (rms) of the surface roughness
height and an auto-correlation function. The scattering coefficient of soil ground

scattering (in power units) can be expressed as:

105

 

_ S
0'30” — tlprzpr3p0'pqr3qz'2quq (5-3)

where p and q are polarizations (H or V), r, is the polarized attenuation factor from the

ith layer (leaf canopy, branch+leaf, and trunk), and of”, is the pq element of the surface

scattering matrix given in the [EM model (Fung et al. 1992).

The IBM model extends its application to a wide range of soil surface conditions. It
reduces to the small perturbation model when the surface is slightly rough and to the
Kirchhoff model when surface roughness is large. Based on an approximate 2nd-order
solution of a pair of integral equations for the tangential surface fields, the soil surface
scattering is composed of single and multiple scattering when the surface rms slope is
large:

5 _ single multiple _
0m _ 0m +0120 (5 4)

Since most natural terrain has a small rms slope, single scattering dominates over the
multiple scattering in most situations (Shi et al. 1997). The single scattering coefficient

(in power units) can be described as (Fung et al. 1992):

2 W" (‘ka ,0)

n!

n
PC]

 

sin le k2 —2/r252 00 Zn
0' g :78 2 ES 1

M (5’5)

 

 

n=l

where k is the wave number, k, = k cos 7] , k, = k sin 77 , 77 is incident angle, 3 is the rms

height of the soil surface, I is the scattering intensity, and W" is the Fourier transfomi of

the nth power of the surface auto-correlation function.

106

 

Multiple scattering is significant only when the surface rms slope is large. It dominates
the cross-polarization backscattering. The multiple scattering coefficient (in power unit)

can be described as (Fung et al. 1992):

k252)n+m

 

Umultiple _ k2 6—3/(2252 Z Z (2
pq _ 47:6

I I
n: lm= l n’m'

- IRCIfgqu (u,v)]W" (u - k,.,v) - W'" (kx + u,v)dudv

) [1+]?!

3: -21.,22 (kis
+167: 22

n_1 m_ 1 n! my
. [[le, (u,v)|2 + 17;, 01,”ng (—u,-—v)] . W’" (u — k,.,v)W" (u + kx,v)dudv
(5-6)
where f M and F p q are the coefficients of the Kirchhoff component and its complementary

component in the electromagnetic field on soil surface. They are described in detail in

Fung et al. (1992).

5.2.1.2 Leaf scattering

In Eq.5-2, the leaf volume scattering coefficient (in power unit), 0,6,0] , is an additive

contribution from all of the leaves in the first (JIM) and second (01:4) layers in forests.

It is a lSt order solution of the radiative transfer equation (Karam et al. 1995):

_ l 2
Glen] _ Clea] + aleaf
_4flQIeafl . 1_Tlprlq +4 lea/2 . 1_T2Pr2(1 -T 2'
_ ,,., k+k) ,,., k+k) ‘P'
( lp lq 86C]? ( 2p 2q seen

(5-7)

‘I
where k0. (i=1,2,j=p, q) is the extinction coefficient in the ith layer and jth polarization.

Qi,”! and Qi,?” represent scattering from one leaf within layer 1 and layer2,

107

 

2 . . . .
Q55] = nl -<Fl::“f >, where F 11:” IS the element of the scattermg amplitude matrix for a

' leaf, and n 1 is the number of leaves in this layer. The ensemble average <> is used in the

equation to represent the statistical average over all leaves in this layer.

5.2.1.3 Branch scattering

Similarly, the branch volume scattering in Eq.5—2, abram , is the 1St order solution of

radiative transfer equation describing the scattering from branches in the second layer:

l—z'zprzq

 

_ branch
Ubranch _ 472.qu . . z.1p2-1q (5-8)

(k,p+k2q)sec77

 

branch _ branch 2 branch - . - -
Here Q m _ n2 2<Iqu >, F M 1s the element of the scatterlng amplitude matrlx for

a branch in layer2, and n2 is the number of branches in this layer.

5.2.1.4 Trunk scattering

The description of trunk volume scattering in Eq.5-2, 0mm, , is similar with Eq.5-7 and

Eq.5-8 except the attenuation factors:

1 — 2' 2'
_ trunk 3P 3‘}
Ummk _ 47Zqu I k k . 1'”)qu . zI2p’z-2q (5-9)
( 3p + 3g ) sec 77

 

Here Qtrunk = n3 '<Fpt:;mk

P4

 

2 . . . .
>, F p?” 18 the element of the scattermg amplltude matrlx for a

branch in layer3, and n 3 is the number of trunks in this layer. The volume backscattering

of trunks suffer from the attenuation from both layer 1 (71) and layer2 (12).

108

5 .2. l .5 Leaf—soil interaction
When calculating the interactions between scatterers in different layers, the soil surface is
assumed to be a specular or a slightly rough surface to simplify the radiative transfer

equations. In Eq.5-2, the backscattering coefficient caused by the interaction between

leaves and soil surface, 0',eaf_30,, , is an additive contribution from the leaves in first (leaf

canopy) and second (leaf+branch) layers. It can be described as:

O-leaf—soil :. O-leafl-soil + 0160f 2—soil (5-10)

where

rlp—r

k

-4kzszcoszn . Qleafl . 14 . T
3qz’2q

(7,,”ka = 4acosn-rlprzpr3p 2(Rpp +qu)-e M k

lq- lp

_4kzs2 c0327] . Qleafz . TZP — T24

pq ' 3a la
k2q —k2p

0",,af2_m., = 47: cosn-z'lprzpr3p -(Rpp +qu)-e

Here R pp and R W are the Fresnel reflect1V1ty of the 5011 surface at p or q polarlzatlon.

5 .2. 1 .6 Branch-soil interaction

Similarly, abm,,c,,_soi, in Eq.5—2 can be described as:

—4kzszcoszq branch TZP—Tzq
6 -Q -——-r,,r,q (5-11)

Ubranch—soil = 47: COS 77 . Tlpz-ZpZ-iip . R Pq k

pp

2q — 2p
5.2.1.7 Trunk-soil interaction
Similarly, 0",,.unk_soi, in Eq.5-2 can be described as:
_ 4 . R . —4k252coszn . trunk . 2.3P _ T30 5 12)
Urrunk—soil — ”C0877 TlpTZpZ-3p . pp 6 qu .TZqZ-lq ( -
k3,] — k3,,

109

5.2.2 Linkage to optical remotely sensed variables

Optical remotely sensed data have been successfully applied to extract green forest
biophysical attributes such as leaf area index (LAI), an indicator of green leaf biomass.
The amount of green leaves controls the magnitude of attenuation from the leaf canopy
in the canopy scattering model. Therefore, linking optical remote sensing variables to the
model could provide quantitative information for the estimation of forest woody

biomass.

As described in Eq.5-3 te Eq.5-12, there is an attenuation factor in each layer of the
forest (two factors in the second layer coming from the branches and leaves,
respectively). Assuming the layer height is H, the attenuation factor in either forward or
backward direction could be described as:

T : e—k,H/cosr7 (5-13)

I

where t= p or q, the polarization status of the signal. The extinction coefficient (k,) is

the total extinction cross section of all leaves statistically averaged over the orientation
probability distribution:

k = 4”NIm<F,, > (5-14)

I
0

 

where F a is the ce—pelarizatien component of the scattering matrix, or scattering
amplitude tensor, of a single scatterer. N is the density (#/m3) of the scatterers in this

layer.

110

In the canopy scattering model, the leaf is modeled as an ellipse. Let a and b be the half-
length and half-width of an elliptic leaf respectively, the leaf density number N in Eq.5-
14 can be related to the leaf area index (LAI) by:

LA]

N: (5-15)
7rab(H1 +H2)

 

It was assumed the same leaf density in first and second layers. In tropical forests, LAI
can be estimated from green fractional cover (fc) that is derived from optical remotely

sensed data as described in chapter 4:

LA] = 0.217 + 0.058 - e(f‘-O'O33)2'5'578 0.4 < fc < 1.0
LAI = 0.85 - fc 0 < fc s 0.4

(5-16)

MSA VI — MSA VI

open

MSA VI — MSA VI

canopy open

in a linear unmixing model as described in

 

where fc =

Chapter 3. The value of fc is in the range of [0,1]. MSA V1 is the Modified Soil Adjusted

Vegetation Index and the detailed description could be found in Qi et al. (1994).

5.2.3 PDF functions of forest components

In the canopy scattering model, the total backscattering in one layer is the statistical
addition of the backscattering from all scatterers with a probability distribution function
(PDF). As shown in Figure 5-2, the distribution of leaves in tropical forests is assumed to
be approximately spherical, the branches are more plagiophile, and the trunks are almost

vertical. The PDF equations of each forest component are listed below:

111

PDFW = Isin 204/2 [07:]
3 7t 57:
PDF = 3 / — ——,— 5-17
branch ICOS al (4) [6 6 :1 ( )
PDmek = cos6 01K??? [0,3]

where 0t is the inclination angle of scatterers. For elliptic leaves, it is the angle between
vertical and the normal vector of the disc in clock-wise direction. Branches and trunks are
simulated as cylinders and CI. is the angle between vertical and the axis of the cylinder in
clock-wise direction. The denominators in Eq.5-17 are the normalization factors that are

the integration of each function in the limited range of (X values.

5.3 Model Simulation and Validation

By replacing the simple soil scattering model with the IBM model and linking some
biophysical parameters to optical data, a microwave/optical synergistic canopy scattering
model was developed to simulate scattering from tropical forests. The study area is Mae
Chaem Watershed as described in Chapter 2. The remotely sensed data in microwave
spectrum were JERS—l Synthetic Aperture Radar (SAR) data, and optical spectrum were
JERS-l VNIR (visible + near infrared) data. Only backscatter in L-band and HH

polarization with incidence angle of 35° was modeled in order to match the JERS-l SAR

configuration.

5.3.1 Remotely sensed data

The study area covers two scenes of JERS-l SAR data (path/row): 132/269 and 132/270.

Two scenes of J ERS-l VNIR data in the same orbit were also acquired. The JERS-l SAR

112

is an active sensor that transmits L-band microwave signals and detects the characteristics
on the Earth surface with the backscattered signals. JERS-l VNIR is a high-resolution
radiometer that observes targets by receiving solar radiation reflected from the earth
surface in three visible and near-infrared (NIR) bands. The system parameters of these
sensors are listed in Table 5-1. The SAR data was acquired 8 days earlier than the VNIR

data. The surface change during this period was thus neglected.

The two scenes of SAR or VNIR data are in the same orbit and therefore, it is easy to
mosaic them into a larger scene (Figure 5-3). Beth SAR and VINR mosaic image covered
most of the study area with gaps on the east and west edges. With the information in the
metadata, both SAR and VNIR images were geerefereced using UTM projection (zone
47) with spheroid and datum WGS84. There was no cloud cover in the VNIR images.
The atmospheric effect in VNIR images was corrected with a simple dark—object-

subtraction (DOS) technique (Chavez 1988).

SAR and VNIR images were geometrically corrected by setting the reference image as
the ETM+ image acquired on February 2, 2000, which has been geometrically corrected
with ground control points (GCP) collected during field trips. From a nadir view, the
VNIR images did not have much geometrical distortion. The total error was 6.14m in a
second-order polynomial geometric model. The JERS-l SAR is a side-looking sensor
with a shallow incident angle. As a result, its imagery is severely distorted in the

mountainous area.

113

SAR data was processed with a 3x3 Lee filter to reduce the speckle noise and increase the
signal/noise ratio with increased number of looks. During the geometric correction,
considering the large area of the coverage and the severe topographic variation, the study
area was divided into four parts with a narrow overlay zone. Each part of the data was
geometrically corrected by setting the ETM+ image as the reference. In a 3rd-order
polynomial geometric model, the total errors are 51.68, 25.23, 42.13, and 48.78 meters
for the lower right, upper left, upper right, and lower left parts, respectively. Since the
look direction of JERS-l SAR was from west to east, the errors in west-east direction
were much higher than north-south direction. The four parts were then mosaiced again

(F ig.5-4b). Even splitting the total area into 4 sub-areas, it is still difficult to correct the
geometric distortion in the watershed. In the areas with high elevation and slopes (Fig.5-
4a), the geometric correction has very high error. To constrain the total error in one pixel,
the pixel size of geometrically corrected SAR data was resampled to 60 by 60m with a

nearest neighbor method.

SAR backscattering was also affected by the topographic variation in mountainous areas.
There are several types of topographic distortions in the SAR image: foreshortening,
layover, and radar shadow (Hendenson and Lewis, 1998). F ereshortening occurs when
the ground range difference between two ground positions is reduced to the slant range
difference, which is shorter because of the side-looking characteristics of SAR and the
topography on the ground. When the shortening becomes greater, the position on the top

of the mountain will be closer to the antenna than the position on the bottom, then

114

layover occurs. In this case, the radar signal cannot access the opposite slope, and radar

shadow occurs.

The foreshortening effect can be corrected with DEM data. During topographic
correction, the backscattering amplitude, DN, of the SAR image can be corrected to a
reference surface on which the local incidence angle of each pixel is 0 (Van Zyl et al.
1993)

sin 77 cos 90

DN = DN* (5-18)

topoc0r

sin 770
where 770 is SAR incidence angle, 77 is local incidence angle, and «9,, is azimuth slope.
Setting 6 as local slope angle and Ago as the relative azimuth angle between local aspect
and SAR azimuth, local incidence angle?) is calculated:

c0377 = cochos 770 + sinI9sin 770 cos Ago (5-19)
The azimuth slope 60 can be solved from tri-angle relationships:

tha = tgbl sin Ago (5-20)

In the areas with layover and radar shadow, the SAR data were permanently lost and

cannot be corrected. The conditions when layover and radar shadow occur are:

Layover: 77 < 900 and 6 < 7;

Radar shadow: 77 > 900

Figure 5-5 is the local incidence angle (a) and the topographically corrected image (b).
The areas with layovers and shadows were shown as white color in the local incidence

angle image. The backscattering amplitude in these areas cannot be used in the following

115

analysis. In high mountainous areas with large local incidence angles, although the
layovers and shadows did not occur, the foreshortening effect was so severe that the
correction accuracy was very poor. For this reason, the application of JERS-l SAR data

in mountainous area is limited.

5.3.2 Model simulation

The model was simulated with a set of parameters listed in Table 5-2. The range of each
parameter was determined with the experience gained during field trips. The mean of the
parameter was selected approximately based on the frequency of occurrences. To
examine the contribution of these biophysical parameters on the backscattering, only the
values of one parameter was changed during each simulation, while other parameters
were represented by their mean values. With the configuration in Table 5-2, in L-band
HH polarization, branch volume scattering, trunk-surface double bounce, and leaf volume

scattering are the primary components in total backscattering (Figure 5-6).

5.3.2.1 Contribution of leaves

In the microwave/optical synergistic forest scattering model, the leaves are simulated as
elliptic discs of which a], a2 are the radii of long and short axis, and r=al/a2 is the elliptic
ratio of the leaf. Figure 5-6a shows the variation of backscattering in different leaf size
and elliptic ratio. When both the leaf size and elliptic ratio are small (a]=0.02m and

r=l .2), the backscattering coefficient is very low and then increases rapidly with higher
elliptic ratio. When the leaf size is larger, the backscattering decreases slowly with

elliptic ratio. With much larger leaves (a1=0.07), there is almost no difference in the

116

backscattering with the change of elliptic ratio. The backscattering difference in the range
of leaf sizes could be more than 2dB in small elliptic ratio (r=1.2) while only 0.3dB in
large ratio (r=1.5). Therefore, a larger leaf size (a1=0.05) and elliptic ratio (r=1.5) is used

in the following simulations.

The contribution of different components in the forest model is shown in Figure 5-6b
with the change of LAI, which is directly related to leaf size and canopy height (Eq.5-15).
When LAI values increase, both the leaf volume scattering and leaf-soil interactions
increase rapidly then slow down to reach the saturation point. The contributions of other
components decrease due to the attenuation from the leaves. As described in Eq.5-13 and
5-14, the attenuation loss in dB unit is more or less linearly related to the amount of
leaves. With the configuration listed in Table 5-2, the increased backscattering from
leaves are compensated with the backscattering loss from other components, and

therefore, the total backscattering increase is only ldB in Figure 5-6b.

5.3.2.2 Contribution of branches

The branch-related backscattering varies with both branch size and density. The branch
length is assumed to be same as the height of layer 2. As shown in Figure 5-6c, in L-band
HH polarization, the modeled branch volume scattering and branch-soil interaction are
very sensitive to branch radius. The branch volume scattering has the highest value at
branch radius = 0.018m and lowest value at radius = 0.028m, and the difference is as high

as 8dB. The branch-soil interaction has an obvious low value at radius = 0.033m, and the

117

difference between highest and lowest scattering is around 3dB. The backscattering of

other components changes with a much smaller scale.

With the increase of branch density (#/m2), both the branch volume scattering and
branch-soil interaction have higher values, and the increment is 5dB and 3dB,
respectively (Figure 5-6d). The branch-soil interaction is more quickly to reach
saturation. The scattering of other components decreases almost linearly because of the

attenuation in branch layer.

5.3.2.3 Contribution of trunks

The trunk-related backscattering varies with both trunk size and density. Trunk size is
determined by DBH and trunk height, and the trunk density is same as the stand density
in the forest. In contrast with the contribution of leaves and branches, the trunk-soil

double bounce is much stronger than trunk volume scattering.

The trunk-soil interaction and trunk volume scattering are sensitive to the variation of
DBH (Figure 5-6e). The difference of the highest and lowest values in trunk-soil
interaction is as high as 9dB. The difference in trunk volume scattering is only 3dB, but
the curve is more complicated. The leaf and branch volume scattering is not affected by
the variation of DBH. The scattering in other components decreases almost linearly. In
accordance with the curve of trunk-soil interaction, the total scattering increases slowly,

then decreases after DBH > 0.16m. The change in total scattering is 2.6dB.

118

As shown in Figure 5-6f, when stand density (# of trees / m2) increases, the trunk-soil
interaction increases, then decreases after stand density > 0.08. The possible reason is
that, with denser trees, there is less space for double bounce between trunks and soil
surface. The decrease is around 3dB in the range of stand density in Figure 5-6f. On the
other hand, the trunk volume backscattering increases up to 5dB. It is obvious that there
is more chance of trunk volume backscattering at higher stand density. The leaf and
branch volume scattering is not affected by the variation of stand density. The scattering
in other components decreases ahnost linearly. With the compensation of increased trunk

volume scattering, the total backscattering decreases only 1dB.

It was found in ground-measured data that trunk height was almost linearly correlated
with tree height (R2=0.81). Here the backscattering is simulated with the variation of tree
height, which affects the heights of three layers (leaf canopy, leaf+branch, and trunk).
These variations in turn result in different contribution from leaves and branches. In
Figure 5-6g, the scattering from all components (except leaf volume scattering) is
tremendous. When the tree height increases from 6m to 24m, the branch volume
scattering, the component that has highest contribution to total scattering, increases up to
6dB. The trunk volume scattering increases slowly, then drops 7dB. The trunk-soil
double bounce drops more than 40dB. The soil surface backscattering, leaf-soil and
branch-soil interaction even reaches infinity (-80 dB in the model) when the tree is higher
than 18m. With the compensation of decreased scattering from these components, the

total backscattering increases 2dB.

119

5.3.3 Model validation

Field measurements were made in January 20-27, 2002 during the second field trip, one
week earlier than the JERS-l SAR data acquisition. Among the 32 study sites visited,
there were 9 dry dipterocarp forests, 9 mixed deciduous, 5 pine transition, 6 dry
evergreen, and 3 moist evergreen. With a point-quadrant plot method, the biophysical
attributes at each site were measured: tree height, trunk height, DBH, stand density. The
forest fractional cover and LAI values were estimated with JERS-l VNIR imagery. The
microwave/optical synergistic canopy scattering model is validated with these ground

measurements and JERS-l SAR data.

In addition to the ground-measured biophysical attributes, other inputs of the model are
listed in Table 5-3. The values of some parameters be different in each forest type. The
leaf and branch radii and soil conditions are selected based on the experiences in
fieldwork. The dielectric constants (8) of forest components are calculated based on
vegetation moisture and local temperature with the model described in Ulaby et al.

(1990). Only backscatter in L-band with incidence angle of 35° and incidence azimuth

angle of 278° is modeled to be consistent with JERS SAR system.

As shown in Figure 5-7, there are logarithmic relationships between modeled
backscattering (LHH) and ground-measured biophysical parameters. The modeled
backscattering increases rapidly and then reaches the saturation point at LAI = 4 (Figure
5—7a). The increments of modeled backscattering with tree height (Figure 5-7b), trunk

height (Figure 5-7c), and DBH (Figure 5—7d) are also observable, but the residuals are

120

higher because of the attenuation from leaves. The relationship between modeled
backscattering and stand density is very weak (Figure 5-7e). One possible reason is that
stand density is not a good indicator of forest growth. A young secondary forest could
have similar stand density as natural undisturbed forest. In late succession stage, the stand
density of natural forest declines, a similar phenomenon as the degraded forests disturbed

by human and natural activities (Aber and Melillo 2001).

The modeled backscattering coefficients are compared with J ERS-l SAR measurements.
In Figure 5-8a, all points were scattered along the 1:1 line, indicating that the model fits
well with SAR data. The root mean square error (RMSE) of the model simulation can be

calculated as:

 

M 0 0 2
2(05/11? - O-mod e!)
RMSE = M M (5-21)

 

where M is the number of study sites that were used for validation. The RMSE in this
study was 0.94dB in comparison with the JERS-l SAR observed and modeled
backscattering coefficients at the 32 study sites. As shown in the residual plot in Figure 5-
8b, most of the residuals scatter between ildB. The only site with residual higher than
2dB is teak plantation in the study area, whose regular pattern results in very high
backscattering in JERS-l observation. Figure 5-8 indicates that the synergistic canopy
scattering model could be used to predict the backscattering in the study area at an

accuracy of around 1dB.

121

5.4 Model Inversion and Forest Parameters Estimation

5.4.1 Model inversion
In the microwave/optical forest scattering model, the backscattering coefficient is

determined by several forest structural parameters: leaf size, leaf density (#fha), dielectric

constants 8W , (9th , 8mm, , heights of each layer, radii of branches and trunks, and
density of branches and trunks (#/ha).

0 _ . .
atom] — f(leaf .. Slze’ leaf __ denSlty, gleaf. Sbranch.8rrunk. Hlaycrl, HlayerZ, Hlayer3 ’

branch _ radius, branch _ density, DBH, 5 tan d _ density)

(5-22)

The forest structures are unique in each forest type. In accordance with the forest type
map created from optical remotely sensed data, only three major forest types are
considered in this chapter: dry dipterocarps, mixed deciduous, and tropical evergreen
forests. Table 5-4 lists a series of parameters that are applied in the model for each forest
type. As described in Chapter 2, some of the forest structural parameters are highly
correlated to each other. Figure 5-9 is the scatterplots of ground-measured stem height vs.
tree height and DBH vs. tree height. The regression equations are:

stem_H = 0.559*tree_H —l.328

(5-23)
DBH = 1.332 *tree__H + 4.29

Here the units of stem_H and tree_H are meters while the DBH is in centimeters.

The leaf density can be calculated with Eq.5-15 and 5-16 from optical remotely sensed
data. With the relationships in Table 5-4 and Eq.5-23, the model only needs two input

parameters: tree height and forest stand density. Then Eq.5—22 becomes:

0'0 = f (tree_ H, s tan d _ density) (5-24)

mod e!

122

For each pixel in the JERS-l SAR data, a least-square-error optimization technique was
applied to estimate these two forest structural parameters with model inversion. The

standard criteria listed below:

- 0 0 2
mln(0mod el - O-SAR)

The initial values of tree_H and stand_density for each forest type were given at the first

iteration. At each iteration, a new 0210“, was generated and compared with GEAR (J ERS-

l) at each pixel. The model stopped when the derivative of 0(3),”, between two iterations

was less than 10'6 or the iteration numbers exceeded 104. Then the tree_H and

stand_density at that iteration were outputted.

5.4.2 Forest structural parameters by model inversion

Figure 5—10 is the modeled tree height distribution in Mae Chaem Watershed. The non-
forest areas as shown in the forest type map (Figure 2-4) are masked out in Figure 5-10
and the following figures. Most of the trees in the watershed are higher than 10 meters.
Most of the dry dipterocarps are lower than 15 meters. Only some isolated dry
dipterocarps are around 20 meters. The mixed deciduous forests have a wider range
between 15 to 30 meters and the distribution is more heterogeneous than dry dipterocarp
forests. There are no trees higher than 30 meters in dry dipterocarps and mixed deciduous
forests. The tropical evergreen forests have a very wide range changing from 15 meters to
higher than 30 meters. There are also some isolated areas with tree heights around 12

meters.

123

Figure 5-11 is the modeled forest stand density distribution in the watershed. In dry
dipterocarps and mixed deciduous forests, there is a wide variety of stand density ranging
from 100 to 600 trees/ha and higher. The stand density of dry dipterocarps is much higher
than mixed deciduous. In accordance with the ground-measured parameters, the stand
density is negatively related to tree height. When the tree is higher than 15 meter, the
stand density in this area is around 100-400 trees/ha. When the tree is lower than 15
meter, the stand density could reach 600 trees/ha or higher. Opposite with the tree height
map, there is a trend of decreasing stand density from east to west in mixed deciduous
forests. The model does not work for tropical evergreen forests. The modeled stand

density distribution saturates at around 500 trees/ha.

5.4.3 Uncertainty analysis
The error 8 of the model inversion is defined as the absolute difference between modeled

and JERS-l SAR observed backscattering coefficients. It is in the unit of dB:

0 — 0.2/1R (5-24)

‘9 : Umodel
In Figure 5-12, most of the areas in dry dipterocarp forests in the watershed have errors
less than 1dB, indicating that the forest structural parameters estimation with model
inversion works well in these forests. In the mixed deciduous forests where the
topographic relief is high, the error could reach 2-3 dB. In some areas on the tops of
mountains, mostly in tropical evergreen forests where the elevation is very high and slope

is very steep, the error is even higher than 4dB. In these areas, the tree height and stand

density are highly overestimated or underestimated. As shown in Figure 5-5b, the

124

 

topographic correction of the SAR image did not work well in these areas, which

introduces high errors from topographic effects.

The modeled tree height and stand density were also compared with ground-measured
values at the study sites. The average of the model inversion error at all study sites was
only 0.44dB although the individual values range from 0 to 3.26 dB. In Figure 5-13a, the
modeled and measured tree height values at most of the study sites scattered along the 1:1
line. There was obvious overestimation in young forests in which the trees were shorter.
These forests were mostly dry dipterocarps in which the measured trees were between 5-
10 meters whereas the modeled ones were 10-15 meters. For forests with higher trees, the

modeled and the measured values fit well.

There is one outlier in Figure 5-13a with low value of measured tree height (13.18 meter)
but high value of modeled tree height (28.0 meter). It is a degraded and recovered
tropical dry evergreen forest with high heterogeneity at the study site (82-14). The
ground measurements could be underestimated using point-quadrant plot method as
discussed in Chapter 2. On the other hand, the canopy scattering model did not account
the scattering contribution from the interaction between young (short) and old (tall) trees
at this site. As a result, a higher tree height was introduced to match the JERS-l SAR
observed scattering coefficients. The error of model inversion at this study site is only

0.19dB.

125

The measured and modeled stand density at the study sites also scattered along the 1:1
line (Figure 5-13b). The modeled stand density at most of the tropical evergreen sites,
including the outlier (S2-14) in Figure 5-12, saturate at about 500 trees/ha. For other

forests (mostly dry dipterocarps and mixed deciduous), the modeled and measured stand

density fit well.

For each forest type, the average and standard deviation of modeled tree height, stand
density and the error of model inversion at all study sites were listed in Table 5-5. The
modeled tree height increased from dry dipterocarps, mixed deciduous, pine transition, to
tropical evergreen forests. The pine transition had the highest standard deviation.
However, in the forest type map, the pine transition, dry evergreen, and moist evergreen
forests were combined as tropical evergreen forests. In this way, the standard deviation is
only 2.93. In accordance with ground measurements, there is a negative relationship
between tree height and stand density. Young secondary dry dipterocarps have higher
density, while the tropical evergreen forests with less human disturbance have lower
density. The standard deviation of moist evergreen forests is very low (0.58) because of
the saturation of model inversion in these forests. Since most study sites were chosen at
relatively flat areas, the topographic effect at these sites was minimal and the average
errors of model inversion at all forest types were less than 1dB. The pine transition had
the highest standard deviation that leaded to a model inversion error of around MB in
tropical evergreen forests. The model inversion in dry dipterocarps and mixed deciduous

forests worked very well.

126

In Figure 5-14, the average values of modeled tree height (Figure 5—14a) and stand
density (Figure 5-l4b) were compared with the ground-measured values in each forest
types. The modeled tree heights were generally overestimated. The overestimation was
more obvious in dry dipterocarps and moist evergreen forests, and less obvious in mixed
deciduous, pine transition, and dry evergreen forests. The total root-mean-square-error
(RMSE) in tree height estimation was 4.6 meter in all study sites, and 3.8 meter when S2-
14 was removed. The modeled stand densities in dry dipterocarps and pine transition fit
well with ground-measured values. But they were overestimated in mixed deciduous and
dry evergreen forests, and were underestimated in moist evergreen forests because of the
saturation in model inversion. The total RMSE in stand density estimation was 300 #/ha

in all study sites, and 299 trees/ha when 82-14 was removed.

5.5 Conclusions and Discussion

In this chapter, a microwave/optical synergistic radiative transfer model was built to
simulate the radar scattering from forests in the study area. The backscattering
coefficients of the forest components, such as branches and trunks, were modeled. The
leaf scattering and its attenuation to the woody components were quantified in the model
with leaf area index retrieved from optical remotely sensed data using the method in
Chapter 4. The root-mean-square error (RMSE) of the model was 0.94 dB when

compared with JERS—l SAR backscattering coefficients.

Two forest structural parameters, tree height and stand density, were estimated by model

inversion with least-square optimization techniques and JERS-l SAR and VNIR data.

127

The error of model inversion was less than 1 dB in most of the areas in the watershed.
However, in the areas with high relief and steep slopes, the model inversion error could
be higher than 4 dB, which introduced high error in the estimation of forest structural
parameters. When all study sites were considered, the total root-mean-square error of tree
height estimation was 4.6 meter. The total root-mean-square error of stand density
estimation was 300 trees/ha. The average model inversion error was 0.44 dB although the

individual values ranged from 0 to 3.26 dB.

The model worked well for the estimation of structural parameters in dry dipterocarps
and mixed deciduous forests. Most of the dry dipterocarps were lower than 15 meters.
The mixed deciduous forests had a wider range between 15 to 30 meters and the
distribution was more heterogeneous. The stand density in these forests varied from 100
to 600 trees/ha or higher. In accordance with the ground measurements, the tree height
was negatively related to the stand density in these forests. The error of model inversion
was less than ldB in most of these forest areas. The modeled tree height in tropical
evergreen forests ranged from 15 to 30 meters and higher. The model inversion to

estimate stand density did not work well in these forests. Its distribution saturated at

approximately 500 trees/ha.

In mountainous areas, the estimation of forest structures with model inversion was highly
affected by the quality of SAR imagery and the accuracy of topographic correction. The
evergreen forests, especially moist evergreen forests, are distributed on the top of

mountains with high relief. It was difficult to correct the topographical effects to JERS-l

128

SAR data in these areas. The error of model inversion in tropical evergreen forests was
higher than that in other forests. In some areas when the relief was very high, the
topographic effect cannot be corrected. The errors of model inversion in these areas were
higher than 4dB, which introduced high uncertainty in the estimation of forest structural
parameters in these areas. A more accurate DEM data and a better correction algorithm

technique for the topographic correction to SAR imagery will be studied in the future.

The dense leaf cover in tropical forests is a big obstacle in SAR remote sensing. It highly
attenuates the scattering from woody components like trunks, branches, and their
interaction with soil surface. The model developed in this study took the advantage of
leaf properties quantified with optical remotely sensed data. As a result, the modeled
backscattering was more sensitive to woody parameters in tropical forests. Therefore, this
model can also be used for the woody biomass estimation in tropical forests which is

discussed in next chapter.

129

5.6 References

Aber, J. D., and A. Melillo (2001). Terrestrial Ecosystem. HP Harcourt Academic Press.

Barkman, J. J. (1979). The investigation of vegetation structure and texture. 123-160. In:
Werger, M. J. A. (ed.), The study of vegetation. Dr. W. Junk Publishers, Hague.

Bongers, F. (2001). Methods to assess tropical rain forest canopy structure: an overview.
Plant Ecology: 153, 263-277.

Chaves, P. S. Jr. (1988). An improved dark-object subtraction technique for atmospheric
scattering correction of multispectral data. Remote Sensing of Environment: 24,
459-47 9.

Fung, A.K., Lee, 2., and Chen, K. S. (1992), Backscatter from a randomly rough
dielectric surface. IEEE Transactions on Geoscience and Remote Sensing: 30 (2):
356-369.

Hoekman, D. H. and Quifiones, M. J. (2002). Biophysical forest type characterization in
the Colombian Amazon by airborne polarimetric SAR. IEEE Trans Transactions
on geoscience and Remote Sensing: (40)6, 1288-1300.

Imhoff, M. L. (1995). A theoretical analysis of the effect of forest structure on Synthetic
Aperture Radar backscatter and the remote sensing of biomass. IEEE
Transactions on Geoscience and Remote Sensing: 33(2), 341-352.

Karam, M. A., Amar, F., Fung, A. K., Mougin, E., Lopes, A., Le Vine, D. M. and
Beaudoin, A. (1995). A microwave polarimetric scattering model for forest
canopies based on vector radiative transfer theory. Remote Sensing of
Environment: 53, 16-30.

Le Toan, T., Beaudoin, A. and Guyon, D. (1992). Relating forest biomass to SAR data.
IEEE Transactions on Geoscience and Remote Sensing: 30, 403-411.

Li, X., and Strahler, A. H. (1992). Geometrical-optical bidirectional reflectance modeling
of the discrete-crown vegetation canopy: effect of crown shape and mutual
shadowing. IEEE Transactions on Geoscience and Remote Sensing: GE-30, 27 6-
292.

Luckman A. J. (1997). The effects of topography on mechanisms of radar backscatter
from a coniferous forest and upland pasture. IEEE Transactions on Geoscience
and Remote Sensing: 36(5), 1830-1834.

Manninen, A. T. and Ulander, L. M. H. (2001). Forestry parameter retrieval fiom texture
in CARABAS VHF-band SAR images. IEEE Transactions on Geoscience and
Remote Sensifiz: 39(12), 2622-2633.

130

McDonald, KC. and Ulaby, F. T. (1993). Radiative transfer modeling of discontinuous
tree canopies at microwave frequencies. International Journal of Remote Sensing:

14(11), 2097-2128.

Qi, J ., Cabot, F ., Moran, M. S. and Dedieu, G. (1995). Biophysical parameter estimations
using multidirectional spectral measurements. Remote Sensing of Environment:

54, 71-83.

Search, P. and Phinn, S. (2000). Determining forest structural attributes using an inverted
Geometric-Optical model in mixed Eucalypt forests, southeast Queensland,
Australia. Remote Sensing and Environment: 71, 141—157.

Shi, J. C., Wang, J., Hsu, A. Y., O’Neil, P.E., Engman, E. T. (1997), Estimation of bare
surface soil moisture and roughness parameters using L-band SAR image data.
IEEE Transactions on Geoscience and Remote Sensing: 35:1254-1265.

Shimabukuro, H. F. and Huemmrich, K. (1995). Remote sensing of forest biophysical
structure using mixture decomposition and geometric reflectance models.
Ecological Application: 5(4), 993-1013.

Strahler, A. H. and Jupp, D. L. V. (1990). Modeling bidirectional reflectance of forests
and woodlands using Boolean models and geometric optics. Remote Sens'mg and
Environment: 34, 153-178.

Sun, G. and Ranson, K. J. (1998). Radar modeling of forest spatial patterns. lntemational
Journal of Remote Sensing: 19 (9), 1769-1791.

Ulaby, F. T., Held, D., Dobson, M. C., McDonald, K. C. and Senior, T. B. A. (1987).
Relating polarization phase difference of SAR signals to scene properties. IEEE
Trans Transactions on geoscience and Remote Sensing: 25 , 83-92.

Ulaby, F .T., Sarabandi, K., Mcdonald, K., Whitt, M. and Dobson, M. C. (1990).
Michigan microwave canopy scattering model. International Journal of Remote

Sensing: 11(2), 1223—1253.

Van Zyl, J. J ., B. D. Chapman, B. D., P. Dubois, and J. Shi, (1993). The effect of
topography on SAR calibration. IEEE Transactions on Geoscience and Remote

Sensing: 31 (5), 1036-1043.

Woodcock, C. E., Collins, J. B., Gopal, S., Jakabhazy, V. D., Li, X., Macomber, S.,
Ryherd, S., Harward, V. J., Levitan, J., Wu, Y. and Warbington, R. (1994).
Mapping forest vegetation using Landsat TM imagery and a canopy reflectance
model. Remote Sensing of Environment: 50, 240-254.

Wulder, M. A., LeDrew, E. F ., Franklin, S. E. and Lavigne, M. (1998). Aerial image
texture information in the estimation of northern deciduous and mixed wood
forest leaf area index (LAI). Remote Sensing of Environment: 64, 65-76.

131

Table 5-1

System parameters of J ERS-l SAR and NVIR sensors.

 

J ERS-l satellite

Altitude:

Orbit:
Orbital direction: Descending

~5 70km

97.67°, sun-synchronous

 

SAR

VNIR

 

Spectral frequency

L-band (1.275 GHz)

Band]: 0.52 — 0.60 pm (green)
Band2: 0.63 — 0.69 pm (red)
Band3: 0.76 — 0.86 pm (NIR)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

View angle Incident angle (right-sided): IFOV:
Near-range; 35° 21.3 (nadir viewing)
Far-range: 42°
Polarization HI-I /
Pixel size 12.5m 18m
Swath width 75km 75km
# of Looks 3 /
Acquisition date 02/27/1998 03/ 06/ l 998
Table 5-2 Input parameters for model simulation.
parameter range mean
Leaf size (m) 1/2 long axis 0.02 ~ 0.07 0.05
elliptic ratio 1.2 ~ 1.5 1.5
Branch size (m) radius 0.016 ~ 0.047 0.031
Trunk size (111) radius 0.06 ~ 0.22 0.1
LAI 0.2 ~ 8.0 1.0
Branch density (#/m2) 0.5 ~ 3.0 1.5
Tree density (#/mT) 0.03 ~ 0.18 0.1
Tree height (m) 6 ~ 24 10

 

 

 

 

132

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Table 5-3 Input parameters for model validation (in addition to ground
measurements).
dry mixed pine dry moist
dipterocarps deciduous transition evergreen evergreen
Leaf radii (0.07,0.05) (0.05,0.033) (0.03,0.02)
(m) (31,32)
Branch 0.022 0.029 0.031 0.033
radius (m)
8133f (l6.94,-5.7) (22.29,-7.26) (28.23,-8.95) (35.23,-
10.78)
ebmnch (19.7,-6.53) (34.45,-
10.49)
gmmk (13.57,-4.66) (26.53,-
8.41)
Sson (5.3,-0.65) (10.23,-
1.95)
Soil Root-mean—square height (m): 0.015
roughness Correlation length (m): 0.15
Table 5-4 A set of forest structural parameters for each forest e in the model.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

dry dipterocarps mixed deciduous tropical
evergreen

Leaf3-D size (7, 5, 0.5) (5, 3.3, 0.5) (4, 3, 0.7)
(cm)
812a! (l6.94,-5.7) (22.29,-7.26) (28.23,-8.95)
gm” (5.3,-0.65) (5.3,-0.65) (10.23,-1.95)
3mm (19.7,-6.53) (19.7,-6.53) (26.53,-8.41)
grrunk (l3.57,-4.66) (13.57,-4.66) (23.26,-7.53)
Height of layer 1 (tree_H- (tree_H-stem_I—I)2*/ 3 (tree_H-

stem_H)*2/ 3 stem_H)/ 3
Height of layer 2 (tree_H- (tree_H-stem_H)* 1/ 3 (tree_H-

stem_H)* 1/ 3 stem_H)*2/ 3
Height of layer3 Stem_H Stem_H Stem_H
Branch radius DBH/4 DBH/4 DBH/4
Branch density Stand_density* 8 Stand_density* l 6 Stand_density*32

 

133

 

 

Table 5-5 Average and standard deviation of modeled tree height, stand density, and
the error of model inversion for each forest types.

 

 

 

 

 

 

tree H density error
(m) stddev (#/ha) stddev (dB) stdev
Dry dipt 14.34 2.35 701.11 328.05 0.13 0.40
Mixed dec 14.65 0.58 677.33 313.01 0.18 0.46
Pine trans 16.73 5.93 747.80 342.98 0.76 1.41
Dry ever 15.65 2.45 639.80 313.16 0.74 0.66
Moist ever 25.99 0.41 499.33 0.58 0.50 0.86

 

 

 

 

 

 

 

 

 

134

Figure 5-2

 

 

 

 

 

 

 

 

Figure 5-1 Geometry of a forest canopy.
1.4 _
2222222 trunk leaf 2 2 - - — branch
1.2 2
1 a l‘\. . lfi
\ /
O 8 / x, ,/ \.
B ' A ‘,‘ / \' ’I \.
D. . [I \‘ 'I \‘
0.6 — I \ I 1
‘~ I
.‘ .1 \ I \. ,
0'4 I 2/ I I \
‘2’ I I. ’1'
0.2 2 .’ 'I I ,' I
l -- x I.
0 I J, I I I I I I ‘ I
0 20 40 60 80 100 120 140 160 180
inclination angle

PDF functions of leaves, branches, and trunks applied in the synergistic

canopy scatter model.

135

 

 

 

 

 

 

 

 

Figure 5-3 JERS-l SAR (a) and VNIR data (NIR+Red+Green) (b) in the study area.

136

 

 

z. .
1-1 ‘
xiv-.314 ' ,I

 

 

 

 

 

 

Figure 5-4 DEM data (a) and geometrically corrected JERS-l SAR image (b).

 

 

 

 

 

 

 

Figure 5-5 Local incidence angle image (a) and the topographically corrected image

(b). The layovers and shadows are the white areas in (a).

137

-8.5 I I l r I
A '9 ‘ :-: : ‘ A
m F ‘ :':-~.
.3 ----------------------- -----‘--$.'.;_;~--.-
e .5?
0'1 -9 5 W
"In-
'0
2 - — - , -
g -10 - _/ - I
g ./ ——-a1=0.07
'/ ........ a1=0.06
40.5 r- ,/ -—-- a1=0.05
./ —-a1=0.04
,/ ----- a1=0.03
/ ._-— =
-11 2' I J J 1 31 0.02

modeled sing (dB)

modeled slgmo (dB)

modeled sigmo (dB)

 

 

 

 

 

 

 

 

 

1.2 1.25 1.3 1.35 1.4 1.45 1
leaf elliptic ratio (a1/a2)

 

.5

 

'5 l I l I I l
-10 _\M -
’ . ' ‘ ‘ ' ‘ - - - , ______ c
45 _ ,- .......... _ ( )
-20-fi_____.______~::"__“: ..
-25 — — —total
" '”’ ‘— " ‘— '— ~\ —\ _\ -— soll
- - leaf
'30 ‘ 232:“; -_v -_-_____,‘.: --------- ‘ ----- leaf-soil
“~-"’ "“‘"--—- ----- branch
-35 --_ ______________________ .. ---- branch-soil
-------- —-- --—~ trunk
-------- trunk-soil
_40 l l l l l l

 

 

 

 

0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05

branch radius (m)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

'5 I T u— r 1 r
-10 -fl— _
-15 _ —I (e)
-20 ________._, _________ _
'25 "' —.§_ -\
‘\ —\ ‘ ----------- total
-30 _ _ _,‘- f \ -\_ a» ‘___ ._ — -— soil
/“‘~\:_~/:"_" :~ * \ -2 — — :93; 'l
- _ ----- ., ‘\-“ _ ----- ea-sor
35 “ ------- y r - c ----- brancE I
"‘-~-_-,_ ---- branc -soi
-40 — ------- ‘ --—— trunk
~ ------ trunk-soil
-45 M 1 1 1 1 1
0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
DBH/2 (m)
0 I l r
'10 ‘ _____f_. ----------- "I (g)
-20 _ r -------------------- a
\
‘\
'30 r “::-:;=.:_\_ ,,,,,, a
‘ “\?*->: “"~ ~- --
'40 ' ‘~.“‘\\ “ 1 ——total
_ ‘\ _ -- sell
502 .\\ __leaf
‘. ‘ ----- leaf-soil
'60 5 ‘. *\ 4 ----- branch
'~ ---- branch-soil
-70 — z, \ - ----trunk
2 --------- trunk-soil
-80 I I ~ 1 I
5 10 15 20 25

tree height (m)

138

modeled slng (dB)

modelded sigmo (dB)

modeled slgmo (dB)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

1 1 I I I T I
-10 _ (b)
-15 _ , ’ ’ I , v - ---- ‘-
.{_ /
-20 - / """" d
/
-25 J —
__~ -_ ‘‘‘‘‘‘ total
-30 P - --------- _ T. :——-— v; 2‘; 7:: '* SOII
v“.\'“~r~;‘; "‘- _ . — _ 'eaf
‘35 _ x . runawa‘o“ ~~....- ----- leaf-soil
----- branch
, .—.- branch-soil
40 ~, ‘1 ---- trunk
-45 1 1 1 1 1 1 1 - ...... UUHK’SOII J
0 2 4 6 8
LAI
-5 I I I I
-10 W, ---------- T
-15 r —
_m_ ~~--:‘:ii:fV
-25 fl‘ .\ ‘ n —total
\ _\ _\ ‘ -— soil
5 -\ ,\ __§ — — leaf
-30 - ~‘ _--_2__ _______________________ f ----- leaf-SOil
I"’--"¢_:::_-—- -‘\ ~~- “- ----- branCh
-35 I __ _______ __, .......... ___ -—-- branch-S<
_______ ~__~__ --—-trunk
.40 1 1 1 1 ------- trunk's‘)",
0.5 1 1.5 2 2.5 3
branch density (#lm2)
-5 1 1 1 1 r 1
—10 -
-15 — d
,\ ‘\
-25 r ‘\ ‘\ - _ —I0t?'
‘ \ ‘\ -— sell
\ \ ‘ ‘ ‘\ ‘\ _ - laf
‘30 «Nun- \ ..‘,~ ,_ __-___-.:-7=-=';: """ leaf-soil
’I-‘Lh;-:':-' ‘ T ‘ - ..... branCh
-35 ~ I, “an“ 2 ‘ § ‘ q ----— branch-92¢
"“--\-_ - ~ ----trunk
-N‘ ‘ ........ uurk-SOII
40 L 1 1 _[ _L I |

0.05 0.075 0.1 0.125 0.15 0.17
stand density (#/m2)

Figure 5-6 Modeled
backscattering of different

5

forest components: leaf size

(a); LAI (b); Branch radius
(c); Branch density (d); DB
(e); Stand density (1); Tree
height (g).

H

modeled sigmo (dB)

modeled sing (dB)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

-2 -——-———~ _2 1‘ -___. - - _ -_ _- - WW. ,I
I
’4 d (a) A '4 _I (b) I
m I
g I
-5 — c -6 d I
E I
-8 - o o g 8 I
O O O . ' _ . . O I
o .6” . o o z .8 o 9. :. . I
4049.‘ O 3-10“ .’09 . I
‘.90 E 0.3 0
-12 d o -12 - . I
-14 I I I I -14 r I I _iI
0 2 4 6 8 12 o 5 1o 15 25 30
LAI tree height (m)
-2 ‘ -2 _ ......
(C) I (d)
‘4 ‘ I -4 '
I 3
i 'o
-6 I . g 5 -
I E
8 I '5 -8
- - , . . ’
: :°.:°:.-° ’2 r 2’ .2:
-102 ' ° 9 o, , -§-10- °° .0.
k . E ’9’ .
-12 “ O z -12 —I . £
'14 l l I I -14 1 I 1 1 I
0 5 10 15 20 0 10 20 30 4O 50
stem height (m) DBH (cm)
-2 ..... _
(e)
A -4 I
m
3
o ‘6 3
E
2"
to -8 r o .
E o .3 .0... o o o
g '10 “ O Q 0 .
CE) ° 0... ’
-12 _
O
-14 I I
O 500 1000 1500
stand density (#lha)
Figure 5-7 Backscattering simulation based on ground-measured forest parameters:

LAI (a), tree height (b), trunk height (c), DBH (d), and stand density (e).

139

Figure 5-8

 

 

 

 

 

 

 

 

 

(a)
A '8 d ’ o
m 0
3 ’ o ’ :
o '9 d ’ ’ 0 o
E ’ o o
m
'5 -10 — . ’
-o o ’ o
2 O
.8 -11— . o o
O o
E
-12 _
O
'13 l l I
-14 -12 -1O -8
JERS-1 sing (dB)
3 _w_, m W
O
2 _ 0))
. O
_ 1 . . t .
g Q .. I
E 0‘ """"""""""" . “‘“3'; “““ “.3“? """"" I
2 ° 0 o o o I
-1 O . O
.
-2 _ I
-3 l l fl I
-14 -12 -10 -8
modeled sing (dB)

Comparison of J ERS-l SAR observed and modeled backscattering

coefficients: scatterplot of backscattering coefficients (a) and residuals (b). The 1:1 line is

also drawn in (a).

140

 

40
x stem_H R2 = 0_72
El DBH R2 = 0.84

 
 

35-

  

 

30-

25‘

(m) and DBH (cm)

stem_H
8

 

 

 

0 5 1 0 1 5 20 25 30
tree_H (m)

Figure 5-9 Scatterplot of stem height and DBH to tree height measured at study sites.

 

 

 

 

 

Figure 5-10 Modeled tree height distribution in the watershed.

141

 

  
 

Density

600
600+

 

 

 

Figure 5-11 Modeled stand density distribution in the watershed.

 

 

 

 

 

Figure 5-12 Error distribution of model inversion.

142

modeled stand density (#lha)

Figure 5-13

-¥ N N
01 O 01

_x
O

l

modeled tree height (m)

01

O

2000

1500

1000

500

Scatterplots of modeled and measured tree height (a) and stand density (b)

 

 

l

 

.0

O.“

90.

RMSE=4.6 meter (with outlier)
RMSE=3.8 meter (without outlier)

 

 

I I I I I

5 10 15 20 25

measured tree height (m)

30

(a)

 

l

RMSE=320 #/ha (with outlier)
RMSE=319 #/ha (without outlier)

0.
O

‘:.. ”O
O.

 

 

 

I I I

500 1000 1500 2000

measured stand density (#lha)

(b)

at all study sites.

143

 

 

 

 

 

 

 

 

 

 

 

25 ~—— — ~ —— — 7~ ~— —
E, 20 -— — — — —— — ~— —
i” 15 _ h _ ___. _ _ ‘3; ._ lmeasured
g — _" T T ‘ “ . Dmodeled
3 . , ,
i:

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

l 7 I I I

dry mixed pine dry moist

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

dipt trans ever ever
(a)
1000
7'7 800
S I'—.—..
s a _ I
g 600 .. _- .. __ __ lmeasured
2 f I. __ Elmodeled
% 400 ~ __ is- _
U . ,
'5 200 " —j. "’7
O I If I I I4 I
dry mixed pine dry moist
dipt trans ever ever
(b)

Figure 5-14 Comparison of modeled and measured tree height (a) and stand density (b)

in different forest types.

144

Chapter 6
Aboveground Woody Biomass Estimation with Microwave and

Optical Remotely Sensed Data

6.1 Introduction

Most of the world’s terrestrial biomass (SO-92%) is contained in forest ecosystems
(Beaudoin et al. 1994). Aboveground forest biomass, or so-called biomass density, is of
great importance in understanding the state and dynamics of ecosystems and their
interaction with global carbon cycles/climate change studies because of its ability to lock
up carbon and its potential for carbon exchange with the atmosphere through forest
destruction and regrowth (Luckman et al. 1998). In this study, aboveground biomass
represents the total woody biomass of trees of diameter 10cm or larger, including twigs,
branches, trunks, and bark (Brown and Lugo, 1992). The amount of green leaves could be
represented by LAI, which has been deeply studied with optical remote sensing
techniques (Peterson et a1. 1987; Baret and Guyot 1991; Qi et al. 1995, 2000; Turner et

al. 1999).

In the past decades, forest biomass mapping from SAR data has become one of the most
promising applications of radar remote sensing to vegetation studies. A number of studies
have been done to estimate aboveground forest biomass using SAR remotely sensed data.
One of the most common approaches is to empirically relate SAR backscattering

intensity to biomass with ground measurements. Le Toan et al. (1992) found that there

145

was a good correlation between SAR intensity and forest biomass for lower frequency
such as P- and L-band data. C-band data primarily interacted with the crown layer
(Saatchi and McDonald 1997) and rapid saturation of the C-band backscattering
coefficient occurred with the increase in forest biomass (Dobson et al. 1992). Cross
polarized backscatter (HV) was found to have the highest correlation with forest biomass.
HH is related to both trunk and crown biomass, whereas W and particularly HV returns

are linked to crown biomass (Beaudoin et al. 1994).

Luckman et a1. (1997) found that lower frequency (L-band) SAR imagery could be used
to discriminate different levels of forest biomass up to 110 ton/ha. With the effect of
topography, the biomass threshold, above which there appears to be no further increase in
L—band radar backscatter, was around 60 tons per hectare (t/ha), and the error is estimated
to be around 30% of the biomass value (Luckman 1998). Other studies (Le Toan et al.
1992, Dobson et al. 1992) using P- and L—band data in temperate forests indicated that it
was possible to estimate biomass up to 200 ton/ha. With sensors at very long wavelengths
(>1m) such as VHF SAR data, the backscattering coefficient is strongly correlated to
characteristics of the tree trunk layer and the signal saturation is not observed up to 360

tons/ha (Melon et al. 2001).

However, the empirical models did not add any physical explanation of the processes that
drive the changes in backscattering. The biomass retrieval by means of relating biomass
to SAR backscattering is shown to be ill-posed (Dobson et al. 1995). These relationships

are highly affected by the structural properties of the forest such as species, age, height,

146

diameter, and tree component orientation distributions (Dobson et al. 1995), and other
factors as topography, soil moisture, history, and local disturbances. It is also problematic

when expanding the measured commercial biomass to the total aboveground biomass.

Several canopy scattering models were successfully developed to simulate radar
backscattering in temperate forests when these structural parameters were taken into
account (Karam et a1. 1995 ; Ulaby et al. 1990; Sun et al. 1991). These models are a
theoretical approximation of forest scattering based on the first or second resolution of
radiative transfer functions (RTF). Among these models, the soil surface was simply
assumed as a continuous and slightly rough dielectric surface. In forests with higher
biomass, the soil backscattering is low and this simplification is acceptable. However, in
sparse forests such as highly degraded or young secondary forests, the contribution of soil

backscattering is under-estimated which introduces high error in biomass estimation.

Due to the complicated natures of the input parameters, it is not a straightforward task to
develop an effective inversion algorithm of these physical models for the purpose of
biomass retrieval. Pierce et al. (1994) applied an artificial neural network to fulfill the
model inversion of forest biophysical attributes. Ranson et al. (1997) used a combination
of forest succession model and a canopy scattering model to develop the relationships
between biomass and backscattering coefficients. The approach produced similar results
with ground measurements and observed SAR data when aboveground biomass is less

than 150 ton/ha, in the same level of direct empirical biomass retrieval.

147

When applied to SAR data, both the empirical and microwave RTF model—based biomass
retrieval approaches only work on forest with sparse to medium biomass such as
temperate coniferous or broadleaf forests. In dense tropical forests, because of the strong
attenuation from the leaf layer, the radar backscattering is no longer sensitive to biomass
when the forest reaches the threshold biomass level. If the contribution of leaves to the
total backscattering can be quantified, the accuracy of woody biomass estimation could

be improved at a much higher level.

Optical remotely sensed data is mostly sensitive to the green leaf properties of the forest
canopy. In optical remote sensing, the reflected energy of the Earth surface is determined
by surface physical properties such as strong chlorophyll absorption and low reflectance
(and, transmittance) in visible bands, and high reflectance due to internal leaf structures in
near infrared (NIR). Optical remote sensing has been widely used to derive information
of vegetation properties such as fractional cover and green leaf area index (J asinski and
Eagleson 1990; Qi et al. 2000). A synergistic use of optical and microwave remote
sensing could quantitatively evaluate the effect of leaf canopy to SAR backscattering and,

therefore, could enhance the availability of biomass estimation.

In this chapter, several methods were used to estimate aboveground biomass in the study
area. Firstly, a simple regression model was built from the relationship between JERS-l
SAR backscattering coefficients and ground-measured biomass. Then, with the forest
structural parameters estimated in Chapter 5, the allometric equations were also applied

to calculate biomass distribution in the study area. Finally, the backscattering

148

contribution of leaves and their attenuation effects to other forest components were
calculated based on both JERS-l VNIR data and the microwave/optical synergistic model
built in Chapter 3. The modeled backscattering with leaf compensation was then

compared with ground measurements to estimation biomass distribution in the study area.

6.2 Biomass Estimation with a Simple Regression Method

In the past studies, a simple 0'0 ~biomass regression model was often built to estimate
biomass with SAR data (Luckman et al. 1997, Hashim and Kadir 1999, Santos, et al.
2003). In this section, the 0'0 in JERS-l SAR image at each study site was retrieved and
the 0'0 ~biomass at all study sites was plotted (Figure 6-1). It was expressed as:

035R“ 2 ——13.838 + 0.963 * ln(bi0mass — 5.487) (6-1)

Here the biomass is in the unit of ton/ha. The 0'0 was logarithmically related to biomass.

At degree of freedom = 29 (32 study sites — 3 model coefficients), the x2 of curve fitting

was 29.15, much smaller than the critical [3.0529 (46.19) at confidence level of 95%,

indicating that the curve fitting in Figure 6-1 was valid and the 0'0 and biomass were

logarithmically related.

With Eq.6-1, the biomass distribution in the study area can be estimated with J ERS-l
data (Figure 6-2). Unlike the forest fractional cover and leaf area index distribution, the
woody biomass distribution estimated in this regression model did not agree with the
forest type distribution. In the east of the watershed is Chiang Mai city and the forests are

inevitably disturbed by human activities at a higher intensity. The Mae Chaem Town is

149

the only major human settlements in the middle and east of the watershed and the forests
are less disturbed. This could be the possible reason that there had higher biomass
distribution in the west than that in the east. However, in lack of a second set of ground
measurements, it was impossible to validate this map. Moreover, the west part is in the
near-range of J ERS-l SAR sensor whereas the east part is in far—range. Although the

J ERS-l SAR data applied in this study has been calibrated and the pixels have been
changed from slant resolution to ground resolution, the mis-calibration in this high
mountainous area could also be a possible reason for the east-west variation of biomass

distribution.

6.3 Biomass Estimation with Compensation of leaf Attenuation

With the leaf area index (LAI) retrieved from JERS-l VNIR data, the attenuation effect of
leaf layer on other forest components could be quantified in the microwave/optical
synergistic canopy scattering model. Figure 6-3 is the LAI image created using the
method described in Chapter 5. The non-forest open areas are masked out in the analysis.
Since the J ERS-l data was acquired in the dry season, the dry dipterocarps and mixed
deciduous forests had low LAI values (around 1.0 or lower) while the evergreen forests

had high LAI values (4.0 or higher).

A microwave/optical canopy scattering model was developed in Chapter 5 to simulate the
backscattering from tropical forests. As shown in Figure 6-4a, there is a logarithm tic
relationship between biomass and total modeled backscattering coefficients at the study

sites. The modeled total backscattering scattering increases rapidly with biomass, then

150

quickly reaches the saturation point at biomass around 100 ton/ha, consistent with the

past studies (Le Toan et al. 1992, Dobson et al. 1992, Luckman et al. 1997).

The backscattering from each component of forests, however, changes differently with
the increase of biomass (Figure 6-4b). In L-band HH polarization, the backscattering
from branches plays the most important role and its variation with biomass is similar with
the total backscattering. Trunk-ground double bounce is the second important
backscattering component when biomass is lower. It decreases rapidly with higher
biomass and reaches infinity (-80dB in the model) when biomass > 400 ton/ha. Leaf
volume scattering also increases with biomass, but the residual is very high because leaf
amount is not directly related to woody biomass, especially in the dry season in the study
area. Branch-ground interaction, trunk volume scattering, and leaf volume scattering also
decreases with higher biomass but with a lower rate than that of the trunk-ground

interaction.

With the microwave/optical canopy scattering model and the tree height map deve10ped
in Chapter 5, the attenuation factor, I, at each study site could be calculated from Eq.5-13
to 5-15 when LAI was known. The attenuation from leaf layer can also be compensated
and the scattering from the woody components (branches and trunks) be modeled. The
leaf-compensated backscattering also had a logarithmic relationship with biomass. Figure
6-5 showed the logarithmic regression between biomass and J ERS-l SAR measured,
modeled, and modeled after attenuation compensation in L-band and HH polarization

(same system parameters as J ERS-l SAR). The JERS-l backscattering increases quickly

151

with biomass, then it slows down at biomass of around 100 ton/ha and tends to saturate.
The modeled backscattering is about ldB lower than JERS-l, but the trend with biomass
is similar. The modeled backscattering with leaf attenuation compensation is more
sensitive to biomass. With limited study sites, the increase of backscattering does not
slow down until biomass reaches 200 ton/ha. There is no obvious saturation in the

modeled backscattering (no-leaf) in Figure 6—5.

At lower biomass, the green leaves have significant contribution to the total
backscattering and, therefore, the modeled backscattering is higher than the one after leaf
attenuation compensation. When biomass is higher than 300 ton/ha, the leaf attenuation
to the woody components (branches, trunks) becomes significant and therefore, the

modeled backscattering is lower than the one after leaf attenuation compensation.

The coefficients of these curve fittings and their statistical tests are listed in Table 6-1.

For a chi-square test, with degree of freedom=29 (32 - 3 coefficients in the model), the

critical value, 1505,29 , is 46.19 at 95% confidence level. The x2 values of 03A,, and
030d 6, in Table 6-1 are much smaller than 150529 and, therefore, the curve fitting lines are

valid in Figure 6-5. The x2 values of agwde,("o_,wf) is very close to the 130529. One

possible reason is that when the leaf attenuation is compensated, the topographic effect

on the SAR backscattering and its interaction with ground surface becomes more

significant, which introduces higher uncertainty to 0310de,(,,0_,mf) .

152

Figure 6-6 is the leaf scattering (Figure 6-4a) and leaf attenuation (Figure 6-4b) images in

the study area in the unit of dB. The leaf scattering component (0780]) is the combined

intensity of leaf volume scattering and leaf-ground interaction. The leaf attenuation factor
(1:) is the attenuation to the radar signals reaching or scattering from the branches, trunks,
and soil ground. The leaf scattering ranged from —32 dB in dry dipterocarps to —10 dB in
tropical evergreen forests. The leaf attenuation (r) ranged from 0.99 in dry dipterocarps
to 0.19 in tropical evergreen forests. The leaf scattering and its attenuation effects are
positively correlated to LA] values. When LAI increased in the study areas, both the leaf
scattering and its attenuation to woody structures increased. As a result, the value of
attenuation factor (I) was lower. Dry dipterocarp forests have lowest leaf scattering and
highest attenuation factor (r) values. In mixed deciduous forests, the leaf scattering is
higher and t is lower. Tropical evergreen forests have highest leaf scattering and lowest 1:

values.

With JERS-l SAR data, LAI data, and the microwave/optical synergistic canopy

scattering model, the leaf scattering and its attenuation to woody forests could be

calculated. The scattering coefficient (a ) from woody forests (branches, trunks) is

woody

then quantified by subtracting leaf scattering (am) from the total backscattering (CSAR)

then compensated for the two-way leaf attenuation (t). It is the combined contribution of
volume scattering from branches and trunks and their interaction with ground surface. In

power unit, 0' could be expressed as:

woody

_ USA}? —Uleaf (6 5)
woody _ z_2 -

0'

153

The woody scattering distribution in the study area was shown in Figure 6-7. It is obvious
that after leaf attenuation compensation, the radar scattering from woody forests
(branches and trunks) is more significantly influenced by topography in the study area.
Tropical evergreen forests have higher scattering from woody structures, but the values

vary greatly because of the topographic effect.

With the regression model described in Figure 6-5 and Table 6—1, the biomass
distribution was estimated with the relationship between the woody forest scattering and
biomass (Figure 6-8). The topographic effect in tropical evergreen forests was
overwhelming so that the estimated biomass in these forests was even lower than dry
dipterocarps and mixed deciduous forests. From both Figure 6-2 and Figure 6-8, the
regression models could not successfully estimate the biomass in the study area,
especially in tropical evergreen forests where the topographic effects were significant.
The results are either overestimated or underestimated in different forests. There was

always a trend of increasing biomass from west to east of the study area.

6.4 Biomass Estimation with Synergistic Model and Allometric

Equations
The modeled tree height and stand density distributions in the study area were described
in Chapter 5. A least-square optimization technique was applied in model inversion while
the leaf scattering and its attenuation to woody forests (branches, trunks) were quantified
with J ERS-l VNIR data. Also, as described in both Chapter 2 and Chapter 5, the

diameter at breast height (DBH) was linearly correlated with tree height. With the tree

154

height distribution map in Chapter 5, the DBH distribution was modeled (Figure 6-9).
Similar as the tree height distribution, most of dry dipterocarps have lower DBH values
(around 15-20 centimeters). Mixed deciduous forests have a wider range of DBH values
between 15 -40 centimeters. The tropical evergreen forests have the highest heterogeneity.
Some of the areas in tropical evergreen forests could have DBH values larger than 45
centimeters while in some isolated areas they are smaller than 10 centimeters. There is
also a trend of DBH values increasing from west to east of the watershed. The trend is not

obvious in other forest types.

From the modeled tree height, stand density, and DBH distributions, the aboveground
forest woody biomass in the watershed could be calculated with the allometric equations
described in Chapter 2. In Figure 6-10, the dry dipterocarp forests have the lowest
biomass (SO-150 ton/ha). There are also some isolated areas with biomass less than 50
ton/ha. The biomass in the areas close to human settlements is lower than the ones far
away. Mixed deciduous forests have higher biomass ranging from 50 to 200 ton/ha. In

some isolated areas it could reach 400 ton/ha.

The biomass distribution in tropical evergreen forests has the highest heterogeneity than
any other forest types. The biomass values range from 50 ton/ha to more than 400 ton/ha.
In some areas the biomass estimation is saturated. In some isolated areas, the biomass
distribution is even less than 50 ton/ha. The biomass estimation in tropical evergreen
forests is questionable because of two reasons: first, tropical evergreen forests locate atop

of mountains in the study area. In some areas, the topographic relief is so high and the

155

slopes are so steep that the topographic effect to the J ERS-l SAR image cannot be
corrected. It introduces high error of model inversion in the areas. Second, the modeled
stand density in tropical evergreen forests saturates at’around 500 trees/ha (Chapter 5),
and therefore, the accuracy of the biomass estimation in these areas are lower than dry

dipterocarps and mixed deciduous forests.

The modeled biomass values at the study sites were compared with ground-measured
values (Figure 6-11). The modeled and measured biomass values scattered along the 1:1
line. The biomass values at most of the study sites were less than 200 ton/ha. Both the
modeled and measured biomass values at moist evergreen forests were higher than 350
ton/ha. There were one pine transition site and one dry evergreen site at which the
modeled biomass was underestimated. The site (SZ-14) was an outlier at which the
modeled biomass was highly overestimated. This overestimation came from the model
saturation. The site was also an outlier in the scatterplot of modeled and measured tree
height at study sites in Chapter 5 (Figure 5-1 3a). There was no study site with biomass
between 200 to 350 ton/ha in both modeled and measured methods. The total root-mean-
square error (RMSE) was 121 ton/ha with this outlier (82-14), and. 88 ton/ha when the

outlier was removed.

When compared in different forest types (Figure 6—12), the modeled biomass was
obviously overestimated at dry dipterocarps study sites because of the overestimation in
the modeled tree height at study sites. In mixed deciduous, pine transition, and dry

evergreen forests (when the outlier S2-l4 was removed), the average values of modeled

156

biomass fitted well with measured values. However, the standard error of modeled
biomass in pine transition was very high, indicating high heterogeneity in the zone. Even
when the outlier (S2-14) was removed, the standard error of dry evergreen was still
slightly higher than that of dry dipteorcarps and mixed deciduous. Both the modeled and
measured biomass values at moist evergreen forest sites were higher than 400 ton/ha. The

model tended to saturate and therefore the standard error was very low.

6.5 Conclusions and Discussions

In this chapter, several methods for biomass estimation with microwave and optical
remotely sensed data were examined. Firstly, a simple regression model was developed
with the observed J ERS-l SAR backscattering coefficients and ground-measured
biomass in the study sites. The model indicated low biomass in open areas and dry
dipterocarp forests close to the human settlements, and high biomass values in mixed
deciduous forests. However, due to the severe topographic effect to the JERS-l SAR
data, the estimated biomass in tropical evergreen forests was low and varied greatly.
There was an obvious trend of increasing biomass from east to west of the study area.
The possible reason comes from the mis-calibration of JERS-l SAR sensor. With the
limited study sites and biomass measurements, it was impossible to validate the biomass

map derived with this method.

Secondly, with the J ERS-l VNIR optical data and the microwave/optical synergistic
canopy scattering model developed in Chapter 5, the volume scattering from leaf layer in

forests and its interaction with ground surface was quantified, and its attenuation factor

157

(1:) was calculated. Then the scattering from woody forests (branches, trunks) was
modeled by subtracting the modeled leaf scattering from the observed JERS-l SAR
backscattering coefficients and compensating the attenuation from leaf layers. According
to the relationship between the woody scattering and biomass over the study sites, the
woody scattering was more sensitive to biomass when the attenuation from leaf layer was
compensated. A similar regression model was built based on the relationship between the
modeled woody scattering and ground-measured biomass over the study sites. It was
applied to estimate the biomass distribution in the study area. However, when the leaf
scattering was removed and its attenuation to the woody forests compensated, the
backscattering coefficients were more severely affected by topographic variation. The
modeled biomass also revealed a trend of increasing biomass from east to west of the

study area. Similarly, the biomass map in this method cannot be validated.

Thirdly, with the tree height and stand density distributions with the microwave/optical
synergistic canopy scattering model in Chapter 5, the DBH distribution was estimated
and the biomass was calculated with the allometric equations. The biomass map in this
method clearly showed the lower biomass distribution in dry dipterocarp forests (SO-150
ton/ha, less than 100 ton/ha in most areas), medium distribution in mixed deciduous
forests (SO-200 ton/ha, greater than 100 ton/ha in most areas), and high biomass
distribution in tropical evergreen forests (50-500 ton/ha and higher). Tropical evergreen
forests showed a high heterogeneity in biomass distribution. It possessed the lowest
biomass in the areas with high relief and steep slopes where the topographic effects

cannot be corrected. As mentioned in Chapter 5, the stand density estimation with model

158

inversion did not work in tropical evergreen forests and saturated at around 500 trees/ha.

Therefore, the biomass estimation in tropical evergreen forests tends to saturate.

The estimated biomass in the third method was compared with the ground—measured
biomass in all study sites. The root-mean-square error of the biomass estimation in the
study sites was 88 ton/ha. The biomass was overestimated in dry dipterocarp forests. The
model worked well in mixed deciduous forest. But the pine transition and dry evergreen
forests, which belong to tropical evergreen forests in the forest type map, had high

standard error, indicating high heterogeneity in these forests.

In conclusion, the synergistic use of microwave and optical remotely sensed data in a
canopy scattering model provided a new approach of biomass estimation in tropical
forests. However, this method is highly limited by the poor quality of SAR data and
geometric mismatch between SAR and DEM data. In mountainous areas where the
topographic effect to the SAR data cannot be corrected, the biomass estimation is very
questionable. The model could be applied to estimate forest biomass up to 200 ton/ha. In
forests with higher densities (biomass higher than 400 ton/ha), the SAR signal begins to

be saturated and is no longer sensitive to the biomass increments.

The accuracy of ground measurements is also critical to the validation of this method.
With limited study sites, the outlier of one site could result in 33 ton/ha of RMSE of
biomass estimation in the area. Moreover, no study sites with biomass between 200 to

350 ton/ha were visited. As a result, it is impossible to validate the biomass estimation in

159

 

this range. Since the ground-measured biomass at study sites with point-quadrant method
tend to be underestimated in heterogeneous forests, a better method for ground

measurements is needed in the future research.

160

6.6 References

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

Beaudoin, A., Le Toan, T., Goze, S., Nezry, E., Lopes, A. and Mougin, E. (1994).
Retrieval of forest biomass from SAR data. International Journal of Remote
Sensing: 15(14), 2777-2796.

Brown, S. and Lugo, A. E. (1992). Aboveground biomass estimates for tropical moist
forests of the Brazilian Amazon. INTERCIENCLA: 17(1), 8-18.

Brown, S., Gillespie, A. J. R. and Lugo, A. E. (1989). Biomass estimation methods for
tropical forests with applications to forest inventory data. Forest Science: 35(4),
881-902.

Dobson, M. C., Ulaby, F. T., Le Toan, T., Beaudoin, A. and Kasischke, E. S. (1992).
Dependence of radar backscatter on coniferous forest biomass. IEEE Transactions
on Geoscience and Remote Sensing: 30, 412-415.

Dobson, M. C., Ulaby, F. T., Pierce, L. E., Sharick, T. L., Bergen, K. M., kellndorfer, J .,
Kendra, J. R., Li, B, Lin, Y. C., Nashashibi, A., Sarabandi, K. and Siqueira, P.
(1995). Estimation of forest biophysical characteristics in Northern Michigan with
SIR-C/X—SAR. IEEE Transactions on Geoscience and Remote Sensing: 33, 877-
895.

Frazer, G. W., Canham, C. D. and Lertzman, K. P. (1999). Gap Light Analyzer (GLA),
Version 2.0: Imaging software to extract canopy structure and gap light
transmission indices from true-color fisheye photographs, users manual and
program documentation. Copyright © 1999: Simon Fraser University, Burnaby,
British Columbia, Canada, and the Institute of Ecosystem Studies, Millbrook,
New York, USA.

Hashim M. and Kadir, W. H. W. (1999). Comparison of JERS-l and Radarsat synthetic
radar data for mapping mangrove and its biomass. In the 20th Asian Conference
on Remote Sensing, Nov. 22-25, 1999, Hong Kong, China. Poster session 1, 1-5.

Henderson, F. M. and Lewis, A. J. (1998). Principles and applications of imaging
RADAR, Manual of remote sensing, vol.2, third edition. John Wiley &Sons, Inc.

Le Toan, T., Beaudoin, A. and Guyon, D. (1992). Relating forest biomass to SAR data.
IEEE Transactions on Geoscience and Remote Sensing: 30, 403-411.

Luckman A. J. (1997). The effects of topography on mechanisms of radar backscatter
from a coniferous forest and upland pasture. IEEE Trgnjsactions on Geoscience
and Remote Sensing: 36(5), 1830-1834.

161

Luckman, A., Baker, J ., Honzak, M. and Lucas, R. (1998). Tropical forest biomass
density estimation using JERS-l SAR: seasonal variation, confidence limits, and
application to image mosaics. Remote Sensing of Environment: 63, 126-139.

Luckman, A., Baker, J ., Kuplich, T. M., Yanasse, C. C. F. and Frery, A. C. (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.

 

Melon, P., Martinez, J. M., Le Toan, T., Ulander, L. M. H. and Beaudoin, A. (2001). On
the retrieving of forest stem volume from VHF SAR data: observation and
modeling. IEEE Transactions on Geoscience and Remote Sensing: 39(11), 2364-
2372.

Peterson, D. L., Spanner, M. A., Running, S. W. and Teuber, K. B. (1987). Relationship
of thematic mapper simulator data to leaf area index of temperate coniferous
forest. Remote Sensing of Environment: 22, 323-341.

 

Pierce, L. E., Sarabandi, K. S. and Ulaby, F. T. (1994). Application of an artificial neural
network in canopy scattering inversion. International Journal of Remote Sensing:
15(16), 3263-3270.

Qi, J ., Cabot, F., Moran, M. S. and Dedieu, G. (1995). Biophysical parameter estimations
using multidirectional spectral measurements. Remote Sensing of Environment:
54, 71-83.

Qi, J ., Chehbouni, A., Huete, A. R. and Kerr, Y. (1994). A modified soil adjusted
vegetation index (MSAVI). Remote Sensing of Environment: 48, 119-126.

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

Ranson, K. J. and Sun G. (1997). An evaluation of AIRSAR and SIR-C/X-SAR images
for mapping northern forest attributes in Maine, USA. Remote Sensing of
Environment: 59, 203—222.

 

Saatchi, S. S. and McDonald, K. C. (1997). Coherent effects on microwave
backscattering models for forest canopies. IEEE Transactions on Geoscience and
Remote Sensing: 35, 1032-1044.

Santos, J. R., Freitas, C. C., Araujo, J. S., Dutra, L. V., Mura, J. C., Gama, F. F., Soler, L.
S., and Sant’Anna S. J. S. (2003). Airborne P-band SAR applied to the
aboveground biomass studies in Brazilian Tropical Rainforest. Remote Sensing of
Environment: 87, 482-493.

162

Sun, G., Simonett, D. and Strahler, A. (1991). A radar backscatter model for
discontinuous coniferous forests. IEEE Transactions on Geoscience and Remote

Sensing: 29, 639-650.

Thailand Land Use and Land Cover Change Case Study (1997). Southeast Asia, IHDP,
IGBP, WCRP program.

Turner, D. R, Cohen, W. B., Kennedy, R. E., Fassnacht, K. S. and Briggs, J. M. (1999).
Relationships between leaf area index and Landsat TM spectral vegetation indices
across three temperate zone sites. Remote Sensing of Environment: 70, 52-68.

Ulaby, F .T., Sarabandi, K., Mcdonald, K., Whitt, M. and Dobson, M. C. (1990).
Michigan microwave canopy scattering model. International Journal of Remote
Sensing: 11(2), 1223-1253.

163

Table 6-1 Coefficients of 0'0 ~biomass logarithmic curve fitting and their statistic
tests.

 

 

 

 

 

 

 

 

 

0' 321R Grand e! Grind el(no-leaf)
m0 -14.22 -14.43 -19.09
ml 1.04 0.93 1.73
m2 -3.29 -11.10 -7.14
X2 24.22 27.58 46.08
Correlation coefficient (R) 0.75 0.79 0.79
Equation 0'0 = m0 + ml * ln(bi0mass + m2)

 

 

 

 

164

JERS sigmHH

 

0 100 200 300 400 500 600
biomass

Figure 6-1 Scatterplot and curve fitting of 035R“ ~biomass at all study sites.

 

 

 

 

300
>300

 

Figure 6—2 Biomass distribution estimated with a simple 0'3““ ~biomass regression

model.

165

 

 

 

 

 

Figure 6-3 The LA] distribution derived from J ERS-l VNIR data in the study area.

166

 

 

 

 

 

 

 

 

 

 

 

 

-7
-8 T . .. . .
I 90 o
g '9 ‘ a? . .
.9 - - ‘5
m 10 ’ . .
'5 o
2 -11 - .9
a.) O
s 12 _
E - O
-13 -
—14 I I
0 200 400 600
biomass
(a)
0
_ x
A -10 8° 35 my XQ 0°
:0 0g >$<
3 _20 _
o
E
U)
'71 -30 ~
3 x oleaf
% _40 _ + Aleaf-soil
g x xbranch
A
_50 - + xbranch—soil
.5 otrunk
-60 ' 0 I I +trunk-soil
0 200 400 600
biomass

(b)
Figure 6-4 Relationships between ground-measured biomass and modeled

backscattering coefficients of forests (a) and their components (b).

167

 

 

-1 3 . . -

 

backscattering coefficient (dB)

 

 

 

 

 

I .

444/ A-7JERS=18AR .
'5' —-—- Model

'15'I ------- ModeunoIean

-16’

 

160 260 360 460 560 600
biomass (ton/ha)

Figure 6-5 Logarithmic curve fitting of SAR observed and modeled backscattering

coefficients with biomass.

 

 

 

 

 

 

 

Figure 6-6 Leaf scattering (a) and its attenuation to the woody forests (b).

168

 

 

 

 

 

Figure 6—7 Woody scattering in the study area.

 

 

 

 

 

Figure 6-8 Biomass distribution with woody forest scattering in a regression model.

169

 

  
 

DBH (cm)
<15
15

 

 

 

Figure 6-9 Modeled DBH distribution in the watershed.

 

 

 

   

300
>300

 

Figure 6-10 Forest biomass distribution from model inversion in the watershed.

170

600

 

RMSE=121 ton/ha (with outlier)

75‘ RMSE=88 ton/ha (without outlier)
E 500 — 9
g I ’ o
a
70’ 400 — 82-14 ’ °
to
E
2 300 ~
.0
g 200 ~ .3 ‘ , .
g o .9 o
O 100 ~ ‘ o.’
e 14‘

O I I | I l

 

 

 

0 100 200 300 400 500 600
measured biomass (ton/ha)

Figure 6-11 Scatterplot of modeled and measured biomass at study sites.

 

 

 

 

500 I I I I
a modeled

A 400 T I measured
3 . standard error
5 300 —
8 200 -
E f "
.9 If.
_D :14 . . I,

100— I I

dl')’ mixed pine dry moist
dipt trans ever ever

Figure 6-12 Average values of measured and modeled biomass (without outlier) and

standard error of the modeled biomass in each forest type.

171

Chapter 7

Conclusions and Future Envisions

7.1 Concluding Remarks

The objective of this study is to quantitatively estimate biophysical attributes using both
optical and microwave (SAR) remote sensing techniques in tropical forests. Several
physical, semi-empirical, and empirical models were built to address this objective. The
study area is tropical environment in Mae Chaem Watershed, ChiangMai, Thailand. The
field measurements were described in Chapter 2. The biophysical attributes estimated in
this study include forest fractional cover (Chapter 3), leaf area index (Chapter 4), tree

height and stand density (Chapter 5), and aboveground biomass (Chapter 6).

A linear unmixing model was built to estimate forest fractional cover (Chapter 3). Instead
of using the spectral reflectance in the linear unmixing models that were commonly
applied in the past studies, a vegetation index — MSA VI, which is most linearly related to
the green leaf abundance in tropical forests when LA] is less than 4.0 — was applied in this
study. Only two components were assumed in each pixel of the remote sensing imagery:
tree canopy and open area. The forest fractional cover thus represents the coverage of the
forest canopy in each pixel. A forest fractional cover map was built with a Landsat ETM+
image acquired in the dry season in the study area. It was validated with both ground-
measured fractional cover at 32 study sites (R2=0.76) and high-resolution (1 meter)
IKONOS calculated fractional cover data at 400 randomly selected locations (R2=O.7O).

The fractional cover map was also adjusted to the wet season in which all forest types

172

 

were flourishing and the seasonal variation of deciduous species in the forests was

reduced.

Leaf area index is highly correlated to forest fractional cover. A modified Gaussian
regression model was built to estimate leaf area index with forest fractional cover
(Chapter 4). Since it is impossible to measure leaf area index with LAI—2000 at the study
sites in the watershed, additional ground data were acquired in northern Michigan where
several forest areas were measured at different seasons to represent the forest conditions
in the study area in Thailand. Both LAI-2000 and fisheye photos were taken to measure
leaf area index and forest fractional cover in these forests. The regression model was
examined with a x2 goodness-of-fit test and then applied to the study area in Thailand.
The leaf area index in both dry and wet seasons were mapped in the study area. A narrow
zone, possibly the pine transition zone between mixed deciduous and evergreen forests
was shown in the leaf area map in the dry season. The leaf area index in the wet season
was much higher in the study area. But some forests not far from the large open areas
(villages, agricultural areas) at lower elevation have very low values even in the wet

season, indicating intense human disturbances in these forests.

To further retrieve forest structural information and aboveground biomass, a
microwave/optical synergistic radiative transfer model was built in Chapter 5. The
volume scattering from woody components (branches, trunks) and their interaction with
ground surface were simulated while the leaf scattering was quantified with leaf area

index fi'om JERS-l VNIR optical data. The root-mean-square error (RMSE) of the

173

model was 0.94 dB compared with the J ERS—l SAR backscattering coefficients. Forest
structural parameters, such as tree height and stand density, were estimated by model
inversion with least-square-error optimization techniques. The error of model inversion in
most of the study areas was less than 1 dB. In areas with high relief and steep slopes, the
topographic effects on J ERS-l SAR backscattering coefficients cannot be corrected and
the error of model inversion could be higher than 4 dB, introducing high error to the
forest structural estimation of tree height and stand density. The RMSE of tree height
estimation was 3.8 meter and that of stand density estimation was 299 trees/ha. The tree
height in young secondary forests, mostly dry dipterocarp forests, was overestimated
while the estimation in other forests fitted well with ground measurements. The stand
density estimation in tropical evergreen forests did not work. It saturated at around 500
trees/ha. In accordance with the ground measurements, the modeled tree height was

negatively correlated with the modeled stand density.

Several methods of biomass estimation with J ERS-l SAR and VNIR data were examined
in Chapter 6. A simple regression model was firstly built to estimate biomass from JERS-
1 SAR backscattering coefficients. The model was examined with a )8 test. It revealed a
trend of increasing biomass distribution from east to west of the study area that was
possibly the mis-calibration in the SAR data when converting slant—range to ground
pixels. The estimated biomass of tropical evergreen forests was very low and varied
greatly. With J ERS-l VNIR data, the leaf scattering and its attenuation to the woody
forests (branches and trunks) were quantified. Then the woody forest scattering after the

compensation of leaf attenuation was modeled. A new regression model was built to

174

estimate biomass with woody scattering in the study area. The woody forest scattering
was more sensitive to the biomass, but the topographic effect to the scattering also
became higher when the leaf attenuation was removed. The modeled biomass in tropical
evergreen forests was even lower with higher heterogeneity. None of the two methods
above could be validated in lack of additional ground measurements. The third method of
biomass estimation was allometric equations. With the microwave/optical synergistic
model and J ERS-l SAR and VNIR data, the tree height and stand density were estimated
in Chapter 5. The DBH was estimated from its linear relation to tree height. Therefore,
the biomass was able to be calculated with allometric equations. The biomass in this
method was overestimated in dry dipterocarp forests, but fitted well with ground
measurements in mixed deciduous forests. The biomass estimation in tropical evergreen
forests is highly heterogeneous because of the topographic effect and the mis-estimation
of stand density in Chapter 5. The root-mean-square error of the biomass estimation in

the study sites was 88 ton/ha.

In a summery, with both optical and microwave remote sensing imagery and radiative
transfer model, the forest fractional cover, leaf area index, tree height and stand density,
and aboveground biomass were mapped in the study area. Aside from forest fractional
cover, each other biophysical attribute is mainly (such as leaf area index) or partially
(such as tree height and stand density, and aboveground biomass) from the results in
earlier chapters. The error propagation between these attributes will be studied in the near

future.

175

7.2 Challenges

The estimation of biophysical attributes in tropical forests mountainous areas was very
challenging. The ground measurements are time and labor consuming and are often
limited because of the physical difficulty of access to the study sites. Moreover, the
ground measurements are not actual ground “truth” in many situations. For example, the
ground-measured forest fractional cover and leaf area index with fisheye photo and LAI-
2000 are actually projected cover and foliage area index. They are the combined
contribution from both green leaves and branches, trunks and any other non-open
components. On the other hand, the fractional cover and leaf area index estimation with
remotely sensed data are based on the spectral responses that are mostly sensitive to
green leaves. The validation of remote sensing estimated fractional cover and leaf area

index with ground measurements is thus biased.

Another challenge of ground measurements is to measure the tree structural parameters
such as tree height and stand density. The fixed-radius plot method often provides the
accurate measure of these parameters on the study area, but is too time- and labor-
consuming and not realistic in forests. The point-quadrant plot method is a more efficient
way of ground measurements, but often results in high bias to the ground data. In the late
succession of forests, there are usually many young trees that surround the old ones and,
therefore, there is high possibility of picking up young trees in each quadrant. The
measured tree height and stand density are thus underestimated. Assuming these data as

ground “truth”, the estimated biophysical parameters from remotely sensed data are often

176

overestimated. In the future studies, the ground measurements with both methods at same

study sites could be compared and a possible compensation model could be developed.

Topographic effect in SAR images was the largest obstacle in microwave remote sensing.
The quality of topographic correction to the SAR images depends on both the system
characteristics of SAR sensor, the DEM data, and the target characteristics. In
mountainous areas with high relief and steep slopes, the topographic effects cannot be
corrected at all. The mis-correction results in high errors to the estimation of forest

biophysical attributes.

There are thousands of computing iterations in the model inversion and the canopy
scattering model is called at each iteration for each pixel of the SAR and optical images.
Therefore, the computation time is also very challenging in the estimation of forest
structural parameters. A possible solution is to do the model inversion and to estimate
biophysical attributes only at core locations in the study area. The attributes at other

pixels can be estimated with linear or polynomial interpolation.

When SAR and optical remote sensing data are applied synergistically, the acquisition of
SAR and optical data at the same time becomes critical. Although quite a few optical
sensors are available by now, there are only ERS-2, Radarsat, and Envisat together with
some Airborne sensors (for example, AIRSAR) still operative. SAR sensors in both ERS-
2 and Radarsat are in C-band and their application in tropical forests is thus very limited.

Fortunately, with high temporal resolution of optical data in the past years, it is possible

177

to develop leaf area index maps on an annual base, which could provide important

information for the microwave/optical synergistic model.

In the near future, the only opportunity to acquire SAR/optical imagery at the same time
is the Advanced Land Observing Satellite (ALOS) designed by NASDA in Japan. It will
be launched in 2004 boarding one radar sensor and two optical sensors: Phrase Array
type L-band Synthetic Aperture Radar (PALSAR, a follow-up of J ERS-l SAR),
Advanced Visible and Near Infrared Radiometer type-2 (AVNIR-2, a follow-up of J ERS—
1 VNIR), and Panchromatic Remote-sensing Instrument for Stereo Mapping (PRISM).
While PALSAR and AVNIR-2 acquire SAR and optical data almost simultaneously, the
PRISM sensor collects high-resolution elevation data that could improve the topographic
correction on both SAR and Optical data. If launched successfully, ALOS data could

provide valuable information in studies of forests of mountainous areas.

7.3 Potential Applications for Other Studies

The results in this study could provide some quantitative information for the studies of
land use / land cover change and human-environment interactions. The variation of the
biophysical attributes distribution in the study area is related to the extent of human
disturbance and forest recovery, together with the difference in forest types, soil texture,
and topography. For example, the biophysical parameters had a pattern of rapid increase
from downhill to uphill in Mae Chaem Watershed. This phenomenon revealed the effects
of human activities on land use and land cover dynamics. The forests at downhill suffer

from intense human disturbance of burning for agriculture and cutting for firewood.

178

Therefore, the values of the biophysical attributes are very low. At higher elevation, due
to the geographical difficulty, clear-cuts for small area agriculture fields and selective
logging for valuable trees become the major human activities. Consequently, the forest
biophysical attributes have higher values. On the top of the mountain, the forests are only
accessible to highland tribes whose activities are very limited and therefore, the

biophysical distribution of these forests varies little.

When applied to a larger scale, this research could also serve as a guide for ecosystem

assessment and further decision-making for environment protection.

For a long time, the binary forest/non-forest mapping has been applied in the assessment
of carbon cycle and global climate change studies. However, different forests, such as
mature forests, selectively logged forests, highly degraded and fragmented forests, and
secondly regrowed forests which have different fractional cover and biomass, play
different roles in carbon sequestration. The methods developed in this study can provide
quantitative estimation of fractional cover and biomass in the forests that are important
input parameters in carbon and climate models. With this information, we can assess the

net carbon flux in forest ecosystems at higher accuracy.

For future studies, the methods developed in this study could be applied to estimate
biophysical attributes in other vegetated ecosystems like temperate forests, agricultural
cropping systems, and rangelands. In each ecosystem, only the probability distribution

functions of the vegetation components and their geophysical and biophysical parameters

179

in the synergistic radiative transfer model need to be re-evaluated. The model could also
be modified and applied to SAR data at different polarizations and radar frequencies in

the estimate of biophysical attributes in different ecosystems.

180

 

.1. Ear...)

 

33.3.... v

.343 .1: :

Eli H.
W}.-.
l. i

HHS

 

 

 

 

15,/53:2

e

 

 

 

 

 

 

 

 

 

 

 

D. n

 

 

 

.u .N .52....
(‘12..

.t ..
. 2w”

 

J.
: z: .4
: a}?