You are on page 1of 129

ESTIMATING SNOW ACCUMULATION FROM

INSAR CORRELATION OBSERVATIONS

A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF ELECTRICAL
ENGINEERING
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

Shadi Oveisgharan
September 2007
© Copyright by Shadi Oveisgharan 2007
All Rights Reserved

ii
I certify that I have read this dissertation and that, in my opinion, it is fully
adequate in scope and quality as a dissertation for the degree of Doctor of
Philosophy.

(Howard A. Zebker) Principal Advisor

I certify that I have read this dissertation and that, in my opinion, it is fully
adequate in scope and quality as a dissertation for the degree of Doctor of
Philosophy.

(Umran Inan)

I certify that I have read this dissertation and that, in my opinion, it is fully
adequate in scope and quality as a dissertation for the degree of Doctor of
Philosophy.

(Ivan Linscott)

Approved for the University Committee on Graduate Studies.

iii
iv
To My Family

v
vi
Abstract

Snow accumulation in remote regions such as Greenland and Antarctica is a key factor for
estimating Earth’s ice mass balance. In situ data are sparse, hence it is useful to derive snow
accumulation from remote sensing observations, such as from microwave thermal emission
and from radar brightness. These data are usually interpreted using electromagnetic mod-
els in which volume scattering is the dominant mechanism. The main limitation of this
approach is that microwave brightness is not well-related to backscatter if the ice sheet
is layered. Because larger grain size and thicker annual layers both increase radar image
brightness, the first corresponding to lower accumulation rate and the second to higher ac-
cumulation rate, models of radar brightness alone cannot accurately reflect accumulation.
Consideration of interferometric radar correlation measurements also can resolve this am-
biguity. Here we introduce an ice scattering model that relates InSAR correlation and radar
brightness to both ice grain size and hoar layer spacing in the dry snow zone of Greenland.
We use this model and ERS satellite radar observations to derive several parameters related
to snow accumulation rates in a small area in the dry snow zone. These parameters show
agreement with four in situ core accumulation rate measurements in this area, while models
using only radar brightness data do not match the observed variation in accumulation rates.
It is necessary to model both surface and volume scattering mechanism to predict the
accumulation rate in the dry snow zone of Greenland accurately. In this work we process
data covering several base stations in the dry snow zone of Greenland, acquired at two
different incidence angles, in order to estimate the surface and volume scattering terms.
The angle diversity is particularly helpful in characterizing the surface term. The detailed
analysis of InSAR radar backscatter in Greenland better describes the physical surface
conditions, so that we may better understand Earth’s changing climate.

vii
viii
Acknowledgements

Though only my name appears on the cover of this dissertation, in fact a great many people
have contributed to it. I owe my gratitude to all those people who have made this disserta-
tion possible and because of whom my graduate experience and life at Stanford has been
one that I will cherish forever.
My deepest gratitude goes to my advisor, Professor Howard Zebker. Five and a half
years ago when I first entered his office to ask about his research projects, I found an
extraordinary person who I right away wished would agree to be my advisor. I have been
amazingly fortunate to have had an advisor who gave me the freedom to explore on my own,
and at the same time the guidance to recover when my steps faltered. His patience and
support helped me go through many critical circumstances, the qualifying exams among
them, and finally finish this dissertation. Were it not for his great advising, I would not be
so interested in research. I hope that one day I would become as good a mentor for students
as Howard has been for me.
I would like to thank Professor Umran Inan for accepting to be my associate advisor
and for his useful comments. My gratitude also goes to Dr. Ivan Linscott for serving as my
third reader, and for his great group meetings which always gave me new insights to look
at the problems.
Professor Len Tyler is one of the best teachers that I have had in my life. I’ve always
learned enthusiastically from the way he looked at problems and from the long discussions
in his office. I would also like to thank the chair of my oral exam committee and my great
teacher at the Physics department, Professor Patricia Burchat. I cannot imagine anybody
whom I would have liked more to chair my orals committee.
I am also indebted to the members of the Radar Remote Sensing Group in STARLab. I

ix
worked with these smart and hardworking people during the course of my graduate studies
and I learned a lot from each one of them. My sincere thanks go to my groupmates, Leif
Harcke, Ana Bertran-Ortiz, Fayaz Onn, Lauren Wye, Sang-Ho Yun, Andy Hooper, Piyush
Shanker, Joern Hoffman, Sigurjon Johnson and Noa Bechor.
I am grateful to the student community of STARLab, especially Hrefna Gunnarsdot-
tir, Fraser Thomson, Kerri Kusza, Sarah Katharine Harriman, and Ahmed Kamal Sultan
Salem, for their continual assistance, in all matters, technical and non-technical, and for
their invaluable support. I would like to specifically thank Lauren Wye, Sarah Katharine
Harriman and Kerri Kusza for accepting to proofread my thesis and give useful feedbacks.
I had the opportunity here at Stanford to meet many great Persian friends. Their support
and care helped me overcome my not too few moments of stress and depression during my
research, and stay focused on my graduate study. I greatly value their friendship and I
deeply appreciate their belief in me. I would like to thank my cousin Hamid, and his
family, Leili and Kimia, for their love and guidance and being always there for us. Hamid,
a Stanford graduate himself, was the reason I applied to Stanford in the first place.
I appreciate the financial support from the Polar Program at NASA that funded the
research which is the subject of this dissertation.
Most importantly, none of this would have been possible without the unfailing love,
support, and patience of my family, to whom I owe everything. My immediate family,
to whom I dedicate this dissertation, has been a constant source of love, concern, support
and strength all these years. I would like to express my heart-felt gratitude to my family,
my father, my mother, my grandmother "maamaan joon", Sheida, Shahram, Shahab and
Shayan. My extended family has aided and encouraged me throughout this endeavor.
Last but not least, I’d like to thank Vahbod. My favorite poet, Rumi, better expresses
my feelings (Rumi et al., 2004):
Some commentary clarifies,
but with love, silence is clearer.
A pen went scribbling along,
but when it tried to write love, it broke...

x
Contents

Abstract vii

Acknowledgements ix

1 Ice Sheets and Snow 1


1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Zones in Earth’s ice sheets . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Ice and snow depth profiles . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.1 Density Profile of the snow . . . . . . . . . . . . . . . . . . . . . . 4
1.3.2 Grain and temperature profiles through depth . . . . . . . . . . . . 6
1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Synopsis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2 SAR and InSAR Remote Sensing 13


2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Radar Remote Sensing of Greenland . . . . . . . . . . . . . . . . . . . . . 13
2.3 SAR and InSAR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.1 Introduction to SAR . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.2 Interferometric SAR . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

xi
3 InSAR Observables 27
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Introduction to backscattering power . . . . . . . . . . . . . . . . . . . . . 28
3.2.1 Normalized backscattered power images . . . . . . . . . . . . . . . 30
3.3 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.1 Spatial correlation . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3.2 Surface correlation . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3.3 Volume correlation . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.4 The spectral approach to interferogram generation . . . . . . . . . 39
3.3.5 Thermal correlation . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3.6 Temporal correlation . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4 Modeling InSAR Echoes From Ice Sheets 45


4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2 Volume backscattering models . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2.1 The Tsang and Mätzler backscattering models . . . . . . . . . . . . 46
4.2.2 Numerical calculation of backscattering cross section and correlation 49
4.2.3 Results from the volume-only backscattering model . . . . . . . . . 50
4.3 Layered backscattering models . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.1 Modeling hoar layers . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.3.2 Numerical calculation of backscattering cross section and correla-
tion considering hoar layering . . . . . . . . . . . . . . . . . . . . 52
4.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.4.1 Interpretation of Greenland InSAR data . . . . . . . . . . . . . . . 56
4.4.2 Normalized backscattered power vs. incidence angle . . . . . . . . 59
4.4.3 Penetration depth in the dry snow zone . . . . . . . . . . . . . . . 61
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.5.1 Effect of density variations profile in our model . . . . . . . . . . . 65
4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

xii
5 Surface Scattering 69
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2 Surface scattering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.3 Effect of surface scattering on correlation . . . . . . . . . . . . . . . . . . 76
5.3.1 Effect of ratio of surface to volume scatter, b, on correlation . . . . 77
5.3.2 Effect of b on backscattered power . . . . . . . . . . . . . . . . . . 81
5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.4.1 Estimating b and σs,dB . . . . . . . . . . . . . . . . . . . . . . . . 82
∆dB
5.4.2 Observed ∆θ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
∆dB
5.4.3 Comparing σs,dB and ∆θ . . . . . . . . . . . . . . . . . . . . . . . 87
5.5 Accumulation rate and slope of backscattered power . . . . . . . . . . . . 92
5.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

6 Conclusions 101
6.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

Bibliography 105

xiii
xiv
List of Tables

2.1 Important ERS Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 19

xv
xvi
List of Figures

1.1 Mass balance for a typical glacier. . . . . . . . . . . . . . . . . . . . . . . 2


1.2 Different zones in the accumulation area. . . . . . . . . . . . . . . . . . . 3
1.3 Density profile through depth . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Effect of temperature and accumulation rate, on the density profile vs. depth 5
1.5 Geometry of grain cell migration . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 Modeled grain growth profile through depth . . . . . . . . . . . . . . . . . 8
1.7 Modeled temperature profile through depth . . . . . . . . . . . . . . . . . 9
1.8 Density profile for the first few meters of snow . . . . . . . . . . . . . . . 10

2.1 Accumulation map of Greenland derived from in situ measurements . . . . 14


2.2 Radar intensity image of Greenland . . . . . . . . . . . . . . . . . . . . . 15
2.3 Comparison of accumulation estimates derived from the backscattered power
and in situ base station accumulation data from four base stations . . . . . . 16
2.4 Angular variation in backscattering from the ERS scatterometer near the
summit in Greenland . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5 SAR viewing geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.6 InSAR images of Greenland . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.7 InSAR viewing geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.1 SAR viewing geometry (Hoen, 2001) . . . . . . . . . . . . . . . . . . . . 28


3.2 ERS sensor distance to the ground and incidence angle across the swath . . 31
3.3 2-way gain of the ERS-1 antenna vs. incidence angle (Laur et al., 2003). . . 32
3.4 Calibration factor of ERS power data vs. incidence angle. . . . . . . . . . . 33
3.5 InSAR viewing geometry (Hoen, 2001) . . . . . . . . . . . . . . . . . . . 34

xvii
3.6 (a) ρsur f ace vs. B⊥ and (b) ρsur f ace vs. θi . . . . . . . . . . . . . . . . . . . 37
3.7 ρvolume vs. B⊥ for grain radii equal to 0.8 and 1.8. . . . . . . . . . . . . . . 39
3.8 ρthermal vs. σ 0 (dB). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

4.1 ρ vs. σ 0 (dB) for different grain radii for models of Mätzler (1998) (dots)
and Tsang et al. (1985) (solid curve) . . . . . . . . . . . . . . . . . . . . 50
4.2 Geometry of hoar layers inside the firn . . . . . . . . . . . . . . . . . . . . 53
4.3 ρ vs. σ 0 (dB) using our model . . . . . . . . . . . . . . . . . . . . . . . . 54
4.4 Data acquisition area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.5 (a) σ 0 and (b) Correlation map in dry snow zone . . . . . . . . . . . . . . 58
4.6 Correlation vs. normalized power backscattered for the detail area of interest. 59
4.7 Hoar layer spacing map in dry snow zone. Four ◦s are in situ stations. . . . 60
4.8 R0 map in dry snow zone . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.9 Comparison of accumulation from in situ Data and volume backscatter
model and Surface + backscatter model . . . . . . . . . . . . . . . . . . . 62
4.10 Angular variation in backscattering from the ERS scatterometer near the
summit in Greenland . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.11 Our model prediction of penetration depth for grain radius variation be-
tween 0.8 and 2.2 mm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.12 Effect of accumulation rate on our model’s prediction . . . . . . . . . . . . 65
4.13 Effect of temperature on our model’s prediction . . . . . . . . . . . . . . . 66

5.1 Map of 14 different frames of ERS1-2 tandem pairs . . . . . . . . . . . . . 70


5.2 Azimuth averaged calibrated normalized backscattered power vs. inci-
dence angle for different frames of ERS1-2 tandem pairs . . . . . . . . . . 71
5.3 Averaged calibrated backscattered power vs. incidence angle for a frame
of RADARSAT-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.4 Volume correlation vs. angle for ERS1-2 data, using Mätzler’s volume
scattering model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.5 Averaged volume correlation as a function of incidence angle . . . . . . . . 75
5.6 Geometry of a layer at depth z0 in the firn. . . . . . . . . . . . . . . . . . . 76
5.7 Total correlation ρt vs. the coefficient b . . . . . . . . . . . . . . . . . . . 78

xviii
5.8 Estimated backscattered power as a function of incidence angle . . . . . . . 79
5.9 Total correlation vs. incidence angle . . . . . . . . . . . . . . . . . . . . . 80
5.10 ρ vs. σ 0 (dB) at different incidence angles. . . . . . . . . . . . . . . . . . . 83
5.11 ρ vs. σ 0 (dB) for h = 20cm at near and far incidence angles . . . . . . . . . 84
5.12 Slope of backscattered power and inferred surface scattering vs. angle for
four in situ stations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.13 Slope of backscattered power and inferred surface scattering vs. angle for
seven in situ stations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.14 Slope of backscattered power and inferred surface scattering vs. angle for
four in situ stations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.15 Slope of backscattered power and inferred surface scattering vs. angle for
eight in situ stations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.16 Geometry of in situ stations covered by two ERS data . . . . . . . . . . . . 89
5.17 Map of four ERS scenes . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.18 Normalized backscattered power and correlation of an area at two angles . . 91
5.19 Slope of backscattered power and inferred surface scattering vs. angle for
the area in Figure 5.18 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.20 Comparison of Scatterometer B values with in situ accumulation data . . . 93
5.21 Slope of backscattered power vs. accumulation rate for In Situ base station
in Figures 5.12, 5.13, 5.15 and 5.14 . . . . . . . . . . . . . . . . . . . . . 97
5.22 Inferred B for the in situ stations in Figure 5.12. . . . . . . . . . . . . . . . 98
5.23 Inferred B for the in situ stations in Figure 5.13. . . . . . . . . . . . . . . . 98
5.24 Inferred B for the in situ stations in Figure 5.15. . . . . . . . . . . . . . . . 99
5.25 Inferred B for the in situ stations in Figure 5.14. . . . . . . . . . . . . . . . 99

xix
xx
Chapter 1

Ice Sheets and Snow

1.1 Overview
Earth’s great ice sheets play an important role in controlling and diagnosing the global
climate. Most global warming models include prediction of the climatic conditions on the
ice sheets.
Beyond their effect on climate, the Greenland and Antarctica ice sheets are the largest
reservoirs of fresh water in the world, representing 5% of the Earth’s land surface and with
an average ice depth greater than 2000 m. Nearly 80% of the Earth’s total fresh water is
captured in the ice sheets and glaciers. Of this fraction, Antarctica and Greenland contain
89% and 10% respectively. The remaining 1% is captured in all other glaciers and small
ice caps. Greenland’s ice sheets, however, play an additional role in controlling the global
climate because of their low temperature. The ice sheet in Greenland is a 2000 km long,
2-3 km high heat sink for the relatively warm winds accompanying the Gulf Stream which
significantly influence the circulation of atmosphere (Bindschadler, 1998). Their height,
highly reflective surfaces, and low incident sunlight combine to make the Greenland ice
sheets among the coldest places on the Earth.
In addition, changes in the behavior of the ice sheets can lead to important natural
catastrophes. If all of the frozen water on Earth were to melt, the sea level around the
planet would rise by about 70 m, inundating many of the world’s cities. A less catastrophic
and more plausible loss of a smaller portion of the Earth’s ice cover would still threaten

1
2 CHAPTER 1. ICE SHEETS AND SNOW

coastal urban areas. The mass balance of the Earth’s great polar ice sheets is a key factor
in the determination of sea level change.

1.2 Zones in Earth’s ice sheets


  

Acc
Sno
ll
* *
* *
* * * Euilibrium
* * Li
  

M Abl
altitude


Flo  O tl
Glaci r

Bedrock

Figure 1.1: Mass balance for a typical glacier.

Current uncertainties in mass balance are comparable to recently observed sea level
rise. Mass balance, or whether the ice sheets are growing or receding, depends on the
relative rates of ice accumulation, primarily through precipitation, and wasting processes
such as glacier and ice stream flow or meltwater runoff. In calculating mass balance, the
ice sheet is commonly divided into areas of gain and loss, which are typically defined by
changes in altitude, as depicted in Figure 1.1. The accumulation area, containing the region
of our study, is where mass, in the form of fallen snow, is added to the glacier throughout
the year. The ablation area is the region where the ice sheet has a net loss of snow and ice
by run-off, calving, sublimation or other processes. Many ice balance estimates have been
derived by monitoring the position of the equilibrium line, the elevation that divides the two
areas. In this work, we address the accumulation side of the mass balance equation. Due to
1.2. ZONES IN EARTH’S ICE SHEETS 3

its low mean temperature, almost the entire Greenland ice sheet is an accumulation area.
In this Chapter, we present the important physical characteristics of the ice sheet that
we will be measuring using radar remote sensing data. We develop a radar scattering model
which relates interferometric radar remote sensing (InSAR) data to field observable proper-
ties of snow proportional to accumulation in the dry snow zone, where significant accumu-
lation occurs. We introduce different zones in the ice sheets later in this section including
the dry snow zone. The densification process of the freshly fallen snow is discussed in Sec-
tion 1.3.1. The snow profiles in Sections 1.3.1 and 1.3.2 are used in the following Chapters
to develop scattering models and analyze the backscattering properties of the firn.

Dry Snow Zone

Percolation Zone

Wet Snow Zone

Ice Zone

Figure 1.2: Different zones in the accumulation area.

There are four different zones within the accumulation area, as shown in Figure 1.2. The
dry snow zone, at the highest altitudes and latitudes, is the coldest interior region where no
melting occurs, even during the height of summer. Lower in elevation is the percolation
zone, which is not as cold as the dry snow zone. Some melting on the surface of the perco-
lation zone occurs in the summer. The melted water percolates some distance into the snow
at temperatures below 0◦ C, before it refreezes into ice layers and ice lenses. Further down
the glacier, we eventually reach a point where, by the end of summer, all the snow deposited
since the end of the previous summer has been melted. Finally, at the lowest elevation of
the accumulation zone is the ice zone, where the layers merge to form a continuous mass
4 CHAPTER 1. ICE SHEETS AND SNOW

because of the large amount of melt water. Annual layers are very difficult to detect in this
zone (Benson, 1962).

1.3 Ice and snow depth profiles

1.3.1 Density Profile of the snow

100

80
Depth, m

60

40

20

0
0.4 0.5 0.6 0.7 0.8 0.9 1
Density, g/cm3

Figure 1.3: Density profile through depth as modeled semi-empirically (Herron and Lang-
way, 1980)

The transformation of snow into glacial ice is one of the fundamental processes in
glaciology. The material in the intermediate stage of this transformation is called firn.
Three stages of densification have been identified in the dry snow zone. The most rapid
stage is the first stage of densification, from initial snow density to the critical density of
about 0.55 g cm−3 . The dominant mechanism in this densification stage is considered to
be grain settling and packing due to the increase in overlying weight. In the second stage,
density increases more slowly with depth until all intercommunicating passages become
closed to form individual bubbles at densities of 0.82-0.84 g cm−3 . In the third stage,
1.3. ICE AND SNOW DEPTH PROFILES 5

compression of the bubbles increases the density. The C-band radar signals used in this
thesis, only penetrate through the first two stages of densification, as will be discussed in
Chapter 4. A semi-empirical model is used to explain the increase of the firn density with
depth in these two stages (Herron and Langway, 1980). Given the temperature, T , in Kelvin
and the initial snow density ρ0 at a given site, it is possible to calculate the density at depth
z through Equation 1.2 as,

1 1
Acc. rate = 20 cm-w.e.
T = -15 Acc. rate = 40 cm-w.e.
Acc. rate = 60 cm-w.e.
0.8 0.8

3
 
T = -30
3

 
ρ,g
ρ,g

0.6 T = -50 0.6


f

0.4 0.4

0 20 40 60 80 100 0 20 40 60 80 100
Firn Depth (m) Firn Depth (m)

( (b )

Figure 1.4: Effect of (a) temperature and (b) accumulation rate, on the density profile
vs. depth as modeled semi-empirically (Herron and Langway, 1980). As we can see,
accumulation rate affects the density profile after a depth of about 17 m, while different
average temperatures change the profile from the surface.
ρi Z j
ρf = (1.1)
1 +Zj
where ρ f and ρi = 0.917 are the density of the firn and pure ice, respectively, and the initial
stage of densification Z j is defined as (Herron and Langway, 1980)
  
ρ0
Z j = exp ρi k1 z + ln (1.2)
ρi − ρ0

where
10160
k1 = 11e− 8.314T (1.3)

Note that the density in this stage is independent of accumulation rate. The depth at which
6 CHAPTER 1. ICE SHEETS AND SNOW

the "critical density", 0.55 g cm−3 , is reached is given by (Herron and Langway, 1980)
    
1 0.55 ρ0
z0.55 = ln − ln (1.4)
ρi k0 ρi − 0.55 ρi − ρ0

Given the snow accumulation rate, A, in meters-of-water-equivalent, for the second stage
of densification Z j is equal to (Herron and Langway, 1980)
  
ρi k2 (z − z0.55 ) 0.55
Zj = exp √ + ln (1.5)
A ρi − 0.55
21400
k2 = 575e− 8.314T (1.6)

Figure 1.3 shows the density versus depth profile for the temperature of -30◦ C and
accumulation rate of 20 cm water equivalent (cm w.e.). As we can see in this figure, in
the first stage, the firn density increases linearly with depth, while in the second stage it
increases exponentially.
Using Equations 1.1–1.6, we show the effects of temperature and accumulation rate
on the density profile in Figures 1.4.a and 1.4.b, respectively. In Figure 1.4.a, we depict
the density profile of snow through depth for different values of average temperature with
accumulation rate of 20 cm-w.e. Note that temperature changes the density profile in both
stages. In Figure 1.4.b, we show the effect of accumulation rate on the density versus depth
profile in the dry snow zone. The average temperature for this figure is -30◦ C. We can see
in these figures that in general, colder temperature and higher accumulation rate decrease
the speed of densification with depth. As discuss in Chapter 4, most of the backscattering in
the dry snow zone of Greenland occurs in the first 20–30 m of snow depth. Therefore, the
effect of temperature variation on density, and consequently on backscattering in C-band,
is a greater contributor than the effect of accumulation rate variation.

1.3.2 Grain and temperature profiles through depth

The average ice grain size increases throughout the densification process. Depending on
the firn depth, snow grain growth is controlled by different processes. In the upper firn
where most scattering occurs for C-band radar (Section 4.4.3), grain growth is dominated
1.3. ICE AND SNOW DEPTH PROFILES 7

Freshly-fallen snow Older, buried snow

Diffusion from
small grains to
larger grains

Figure 1.5: Geometry of grain cell migration. The net direction of the movement of
molecules is governed by the thermodynamic principle that the free energy of the system
tends to a minimum by reduction in surface area. The larger grains are “radar-brighter”
than the smaller grains (Paterson, 1994).

by vapor transport processes, driven by the firn temperature gradient (Paterson, 1994). In
this process, the smaller grains diffuse to larger grains to minimize the free energy of the
system by reducing the total surface area (Figure 1.5). The growth rate is much higher in
the first few meters of snow and then decreases. Figure 1.6 shows the grain size radius
variation as a function of depth (Jordan, 1991). The rate at which the grains are buried
below the surface is proportional to the accumulation rate. For high accumulation rates,
grains are buried rapidly and spend less time in the upper firn and as a result, grain sizes
are smaller than for low accumulation rate. This property can affect how radar images of
the dry snow zone are interpreted, as we discuss in Section 2.2.
The depth at which the seasonal and long-term changes in surface temperature pen-
etrates through the snow can be analyzed by heat transfer theory (Paterson, 1994). At a
depth of about 10 m, the measured temperature is roughly equal to the mean annual air
temperature (Figure 1.7). Above 10 m, firn temperatures are strongly modulated by the
seasonal temperature cycle.
Under some circumstances, the change in physical state of recently fallen snow pro-
ceeds in a very different way than the cycle just described. Normally in autumn, when the
surface is cooling rapidly and the underlying layers are still relatively warm, evaporation
takes place in the lower layers. The vapor rises and condenses to form depth hoar crystals in
8 CHAPTER 1. ICE SHEETS AND SNOW

1.6

1.4

1.2
Grain Radius (mm)

0.8

0.6
10 cm/yr w.e.
0.4 20 cm/yr w.e.
30 cm/yr w.e.
0.2
40 cm/yr w.e.
0
0 5 10 15 20
Firn Depth (m)

Figure 1.6: Modeled grain growth profile through depth as a function of accumulation rate
(Jordan, 1991)

the cold upper layers (Paterson, 1994). In central Greenland, however, depth hoar appears
to form in the summer by radiative heating of near-surface snow (Alley et al., 1990). Depth
hoar is the most coarse-grained type of firn that can be formed in the absence of melt water.
The average grain size is about 2 to 5 mm. Foundation of depth hoar leads to the annual
layers observed in ice cores and snow trenches. Below about 6 m depth, these low density
layers (0.1 to 0.3 g/cm3 ) collapse. In Figure 1.8, we show the density profile for the first 6
m of snow at 72.21◦ N, 38.48◦ W in the dry snow zone of Greenland. The thin layers with
low density correspond to depth hoar layers, for instance at depth 2.8 m. Since each depth
hoar layer marks the onset of the summer season in a given year, the spacing between these
layers can be related to snow accumulated each year (Alley et al., 1990; Wismann et al.,
1997).
1.4. CONTRIBUTIONS 9

15

10
Variation ∆ T from Tavg(C)

−5 Summer
Spring

−10 Winter
Autumn

−15
0 2 4 6 8 10 12 14 16 18 20
Firn Depth (m)

Figure 1.7: Modeled temperature profile through depth (Paterson, 1994)

In this Chapter, we have introduced the important physical characteristics of the Green-
land ice sheet that we will be measuring using radar remote sensing data. We discussed
different zones in the ice sheets and the densification process of the freshly fallen snow. In
the following Chapters, we will use the snow profiles in Sections 1.3.1 and 1.3.2 to develop
scattering models and analyze the backscattering properties of the firn. The two important
structural features in the dry snow zone that are indicative of the accumulation rate are the
grain size and the spacing between the hoar layers. Importance of each of these will be
discussed in Chapter 4.

1.4 Contributions

The contributions of the research presented in this dissertation are:


10 CHAPTER 1. ICE SHEETS AND SNOW

0.8

0.7

0.6
ρ, g/cm3

0.5

0.4

0.3

0.2
0 1 2 3 4 5 6 7
Depth, m

Figure 1.8: Density profile for the first few meters of snow at 72.21 N and 38.48 W in the
dry snow zone of Greenland. The thin layers with low density correspond to hoar layers.

(1) We show that greater grain radius and greater layer spacing both increase backscat-
tered power in observations of ice sheets. The former corresponds to lower accu-
mulation rate, whereas the latter corresponds to higher accumulation rate. Hence,
we cannot reliably estimate accumulation from backscattered power alone. We show
that consideration of radar image correlation together with backscattered power mea-
surements can resolve this ambiguity.

(2) We present a radar scattering model that can retrieve snow accumulation parameters
more accurately than existing remote sensing methods. We introduce an interfero-
metric ice scattering model relating InSAR correlation and radar brightness to both
ice grain size and hoar layer spacing in the dry snow zone of Greenland. We use
this model and ERS satellite radar observations to derive several parameters related
to snow accumulation rates in a small area in the dry snow zone. These parameters
show agreement with four in situ core accumulation rate measurements in this area,
while models using only radar brightness data do not match the observed variation in
accumulation rates.
1.5. SYNOPSIS 11

(3) Our model also demonstrates that surface scattering plays a significant role in model-
ing backscattered power from dry snow. We generalize the surface scattering term in
the interferometric ice scattering model in the dry snow zone to apply the theory to a
wider range of imaging geometries. We process new data sets over each base station
with two different incidence angles in order to better estimate the surface scattering
contribution. We introduce the parameter b to quantify the surface scattering and
its effect on correlation. The value of surface scattering inferred under this model
closely parallels empirical observations.

(4) We show that the slope of surface scattering, B, from our model, helps derive the
accumulation rate. We show that higher values of accumulation rate correspond to
higher values of B. This gives another method to estimate the accumulation rate in
addition to estimating the layer spacing.

1.5 Synopsis
Chapter 2 presents background on different techniques to monitor the accumulation rate
of the dry snow zone of Greenland, including a remote sensing technique. In Chapter 2
we describe microwave radar measurement methods and how they are related to studies of
the Earth’s ice sheets. We introduce Synthetic Aperture Radar (SAR) and Interferometric
Synthetic Aperture Radar. We show that there is useful information in both the amplitude
SAR and the interferometric phase InSAR images.
In Chapter 3, we show that two radar observables permit us to probe the sub-surface
of the ice sheet from space. In this Chapter we study these two important InSAR observ-
ables: normalized backscattered power, σ 0 , and interferometric correlation, ρ , which is the
coherence between corresponding pixels in two different SAR images. We also develop a
model that relates correlation to both the backscattering characteristics of the medium and
geometry of the antenna with respect to the ground.
We analyze in Chapter 4 two types of radar scattering models for deriving the accumu-
lation rates in the dry snow zone of Greenland. First, we present a volume scattering model
and compare its predictions to ERS backscattered power and correlation data. We find that
12 CHAPTER 1. ICE SHEETS AND SNOW

this model does not provide an accurate match to the data. Then, we propose a volume
and surface scattering model which we find explains the radar data more accurately, and
resolves some of the ambiguities encountered in estimating accumulation rate using the
first model.
Chapter 5 presents surface scattering in the dry snow zone of Greenland in more detail.
We review the inconsistencies in measured backscattered power and correlation that occur
if we ignore surface scattering. A new parameter b, the ratio of surface backscattered
power to volume backscattered power, is defined to quantify the effect of surface scattering
on correlation. We then show how our estimate of surface scattering helps us to better
estimate the accumulation rate in the dry snow zone of Greenland. We summarize and
conclude in Chapter 6.
Chapter 2

SAR and InSAR Remote Sensing

2.1 Introduction
In this Chapter we show how remote sensing data can be exploited to estimate the accu-
mulation rate over the remote and cold regions of Greenland, an important factor in overall
mass balance. We describe microwave radar measurement methods and how they are re-
lated to studies of the Earth’s ice sheets. We introduce Synthetic Aperture Radar (SAR)
and Interferometric Synthetic Aperture Radar (InSAR) in Section 2.3, and explain various
parameters used in SAR/InSAR imaging systems. We show that there is useful information
in both the amplitude SAR and the interferometric phase InSAR images. Later Chapters
will expand the InSAR discussion and introduce models of InSAR backscatter that we will
use to estimate snow accumulation.

2.2 Radar Remote Sensing of Greenland and Snow Accu-


mulation
Our specific goal in this work is to develop a model relating interferometric radar remote
sensing (InSAR) data to field observable properties of snow proportional to accumulation
in the dry snow zone, where significant accumulation occurs. The dry snow zone, at the
highest altitudes and latitudes, is the coldest interior region where no melting occurs, even

13
14 CHAPTER 2. SAR AND INSAR REMOTE SENSING

Accumulation Rate
below 10
10 - 15
15 - 20
20 - 25
25 - 30
30 - 40
40 - 50
50 - 60
60 - 80
80 - 100
100 - 120
over 120
Water equiv., cm

Figure 2.1: Example accumulation map of Greenland derived from in situ measurements
at core and pit stations shown by dots (Ohmura and Reeh, 1991). The inset square in this
map is the area of our study. ⋆s are meteorological stations.

during the height of summer.

The literature reports many estimates of accumulation derived from sparse in situ mea-
surements (e.g. Ohmura and Reeh, 1991, Bales et al., 2001 during PARCA, Program for
Regional Climate Assessment ). Figure 2.1 depicts an accumulation map of Greenland, de-
rived from such in situ measurements at stations shown by dots (Ohmura and Reeh, 1991).
2.2. RADAR REMOTE SENSING OF GREENLAND 15

These sparse ground station data are contoured to form an estimate of accumulation vari-
ations. The number of data points is limited, hence the resolution of such a map is very
poor.

Figure 2.2: Radar intensity image of Greenland from power backscattered InSAR data from
Fahnestock et al. (1993).

A higher resolution approach to estimating accumulation results from the use of radar
backscatter data. Figure 2.2 shows a radar intensity image of Greenland presented by
Fahnestock et al. (1993). In this figure the dry snow zone appears dark, and the darker
area tends to correspond to higher accumulation rates. Hence an accumulation estimate
follows from contouring the intensity map. Detailed comparison of accumulation rates de-
rived from in situ stations and reduction of the backscatter data shows that these quantities
can disagree significantly, as can be seen in Figure 2.3. This figure depicts in situ data from
16 CHAPTER 2. SAR AND INSAR REMOTE SENSING

50
In situ
40 accumulation
rate (cm-we/
30 year)
20 Normalized
power
10 backscattered
0 (dB)
Inferred acc.
-10 rate (cm-we/
-20 year), Vol.
site a site b site c site d scatter only

Figure 2.3: Comparison of accumulation estimates derived from the backscattered


power(violet) using a model of volume backscattering only, and in situ base station ac-
cumulation data from four base stations located in the inset box of Figure 2.1. The in
situ accumulation data disagree with the inferred accumulation rate from volume scattering
only.

four base stations (Ohmura and Reeh, 1991) located in the inset box of Figure 2.1. This
box is centered at about 75.9 ◦ N and 36.5 ◦ W. The data are compared to accumulation es-
timates derived from the backscattered power using a model of volume backscattering as
explained in Section 4.2. Similar results have frequently appeared in the literature. It is
clear that for this region the inferred accumulation rates and the in situ measured values
strongly disagree. For instance, Site a is the radar-darkest site and has the lowest accumu-
lation rate among the four base stations; this result disagrees with the general trend of σ 0
in the dry snow zone. As we show below, it is difficult to produce an accurate accumulation
map using only backscattered power and volume-scattering models.

Other evidence that volume scatter models are inadequate to fully explain echoes from
the ice sheets can be seen in Figure 2.4. Here, the angular variation in backscattering from
ERS scatterometer is shown near the summit in Greenland (73.25 N, 37.28 W) (Ashcraft
2.2. RADAR REMOTE SENSING OF GREENLAND 17

−10

−11
σ0(dB)

−12

−13
20 22 24 26
θ
i

Figure 2.4: Angular variation in backscattering from the ERS scatterometer near the sum-
mit in Greenland, shown by plus sign (Ashcraft and Long, 2006). This angular variation
of backscattered power in the dry snow zone cannot be explained by a volume-scatter-only
model (dashed line). Therefore, surface scattering is also an important factor in the dry
snow zone in the C-band for low incidence angles.

and Long, 2006). It is apparent in Figure 2.4 that a volume scattering model does not
well match the angular dependence of echoes from ice in the dry snow zone. Although
volume scattering is the dominant factor of backscattering in the dry snow zone, surface
backscattering terms (in addition to volume backscattering) are needed to better explain
these observations. As we see in Section 4.4, reflections at the internal surfaces add up to
match the observed angular variation. Therefore volume scattering only cannot explain the
observed total backscattered power.
18 CHAPTER 2. SAR AND INSAR REMOTE SENSING

2.3 SAR and InSAR

2.3.1 Introduction to SAR


Spaceborne SAR systems transmit and receive microwave signals at cm-dm wavelengths.
Because of their long wavelengths compared to the particles in atmospheric clouds and
dust, they pass through the atmosphere nearly unaffected, while transmission of optical
signals are highly dependent on the weather conditions. This allows radar systems to oper-
ate year-round and through nearly any weather conditions.

v S

*
ac
'()
tr

% &t /a
l
H #$
R az

 !" an+,--1
.
S

f //tp 0 nt

Figure 2.5: SAR viewing geometry. The antenna footprint is the antenna beam pattern
projected on the ground. The swath width is the length of the antenna footprint in the
ground range direction. Other parameters are described in the text.

Typically, a synthetic aperture radar illuminates the ground surface in a side looking
fashion as depicted in Figure 2.5. The antenna footprint is the area on the ground illu-
minated by the radar antenna beam. The projection of the antenna footprint in the range
direction defines the width of the image swath, typically 100 km. As the SAR moves at a
2.3. SAR AND INSAR 19

Parameter Value
Wavelength (λ ), m 0.0566 (C-Band)
Altitude (H), km 796
Spacecraft velocity (vSAR ), m/s 7541
Pulse repetition frequency (PRF), Hz 1679
Pulse length (τ ), µ sec 37.1
Range sampling rate ( fs ), MHz 18.96
Range chirp slope (s), Hz/sec 0.42 × 1012
Range to first pixel (R0 ), km 838
Incidence angle range (θi ), deg 19.4–26.4
Slant Range resolution, m 10
Azimuth resolution, m 5.84
Scene size on the ground, km2 ∼ 100 × 100
Orbit Repeat time, days 35

Table 2.1: Important ERS Parameters

velocity vSAR along its flight track at altitude H above the ground, it transmits a series of
microwave pulses at a rate denoted the pulse repetition frequency (PRF). All of the SAR
observations presented in this thesis are European Research Satellites (ERS) data with a
wavelength of λ = 5.67 cm (C-Band). We have acquired these data sets from the European
Space Agency (ESA). The two satellites, ERS-1 and ERS-2, were put in orbit in 1991 and
1995, respectively. Table 2.1 lists the important parameters of ERS data.
Transmitted signals scatter off natural and man-made objects on the ground. The am-
plitude and phase of the scattered signal are determined by the position of the scatterers
relative to the SAR sensor and the electrical properties of all of the scatterers in the illumi-
nated area on the ground. The SAR system receives and samples the echo at a sampling
frequency rate fs .
The resolution of a SAR image is defined as the ability to distinguish different features
close to each other. The minimum distance between two distinguishable scatterers on the
ground is called the resolution size. The resolution size is described by the 2-D impulse
response function, W (x, R), where R and x are the distances from the SAR sensor to the
center of resolution element in range and azimuth directions, respectively. Each resolution
cell, or resel, has an associated impulse response that extends to infinity in the range and
20 CHAPTER 2. SAR AND INSAR REMOTE SENSING

azimuth directions. The resolution size is often approximated by a measure of the width of
the impulse response, usually the width where the impulse response falls down to half of
its maximum.

The resel size however should not be confused with the pixel size. The smallest element
of an image is called a pixel. The pixel size in the range/azimuth directions, whether in the
raw or in the processed image, is defined by the spacing between adjacent pixels in that
direction. A SAR image is a two-dimensional matrix of complex values in both the range
and azimuth coordinates. The amplitude and phase of a pixel is a superposition of echoes
from many scatterers within a resolution cell centered on that pixel.
vSAR re
The azimuth pixel size is equal to PRF ( H+re ) (Curlander and McDonough, 1991),
where re is the Earth’s radius. The azimuth pixel size is about 4 m for ERS data. The
c c
slant range pixel size is given by 2 fs and the ground range pixel size is 2 fs sin θi (Curlander
and McDonough, 1991); it is about 20 m for ERS data. Here θi and c are incidence angle
and the speed of light, respectively. Typically, 5 azimuth pixels are averaged together in
order to create pixels with approximately equal range and azimuth dimensions. This results
in a final 20m × 20m pixel. Averaging of adjacent pixels is referred to as “taking looks”.

The resolutions in the range and azimuth directions are poor for unprocessed SAR data.
The unprocessed resolution, with no matched filter processing in the range direction, is

δ runprocessed = 2 which is about 5 km for ERS data, where τ is the pulse length. The
resolution in the along-track (azimuth) direction before SAR processing is determined by

the antenna footprint. The unprocessed azimuth resolution is equal to La which is about 5
km for ERS data. Here R is the distance between the satellite and the illuminated area, La
is the length of the antenna parallel to the azimuth direction, and λ is the wavelength of the
signal.

With signal processing techniques one can achieve an image with several-meter reso-
lution as given above. High range resolution is often obtained by using a linear-frequency
modulated waveform, a chirp signal for ERS, to increase the bandwidth of the transmitted
signal. Cross-correlating the received echoes with a matched filter tuned to the transmit-
ted pulse, we can improve our range resolution to δ rfocused = c
2BW which is less than 10
m in the ERS case. A similar process is used in the azimuth direction. The reference
2.3. SAR AND INSAR 21

function used in the matched filter is generated based on the Doppler frequency shift phe-
nomenon. Because of the relative motion of the satellite with respect to the ground, echoes
from the target in front of the moving sensor are shifted to higher frequencies, whereas
those reflected from behind are shifted to lower frequencies, compared to the transmitted
frequency. This frequency spread behaves like a chirp signal, and the same matched filter-
ing method can be employed. This matched filtering makes the resolution in the azimuth
La
direction equal to 2, which is around 5 m for ERS. Therefore, we can synthesize a larger
aperture by combining the echoes coherently.

(a) (b)

Figure 2.6: (a) An ERS-1 radar power image from the north coast of Greenland (centered
at 80.40N and 23.95W). Depending on the different scattering properties of the different
forms of water present, the image looks brighter or darker. (b) SAR interferogram over the
same area. The perpendicular baseline between the two radar observation is 170 m.

Figure 2.6.a shows an example SAR image over the north coast of Greenland acquired
from ERS-1 in January 30th, 1996. The variation in the observed brightness is due to the
different backscattering properties of the different forms of water at C-band. The bright
region at the bottom of this image is the start of the percolation zone, where ice pipes
and lenses dominate the scattering as explained in Section 1.2. The very dark region at
22 CHAPTER 2. SAR AND INSAR REMOTE SENSING

the top right of the image is the Arctic ocean, which, due to its smooth surface and to
the electrically lossy nature of sea water, scatters very little to the radar. The gray region
corresponds to the wet-snow zone, where there is an absence of large structures in the firn.

2.3.2 Interferometric SAR

As stated in Section 2.3.1, the SAR image may be represented as a 2D complex matrix. The
phase of the returned echo contains information on the distance between the sensor and the
ground as well as the reflection properties of the scatterers in the resolution cell at that
wavelength and viewing geometry. SAR interferometry consists of cross-correlating two
such complex SAR images of the same terrain to get more information about the terrain. In
order to provide additional information, at least one of the parameters of the second image
should be different, such as flight path, or time of observation. The most well-known
interferometry is the one in which the positions of two observations are slightly different.
Therefore, in interferometric SAR (InSAR), two SAR (InSAR) images are acquired by
antennas A1 and A2 over the same area but from slightly different viewing geometries as
shown in Figure 2.7; or at different times. Here we use spatial diversity and do not extract
information from temporal changes. The flight direction here is out of the page. In the
single-pass mode, the antennas are physically operated simultaneously on a single platform,
as in the TOPSAR and SRTM instruments. In repeat-pass interferometry on the other hand,
the terrain can be imaged by a single SAR antenna at two different times corresponding to
the satellite’s orbit period. Because the satellite orbits do not repeat exactly, the viewing
geometry is a little bit different. Since ERS1 and ERS2 operated simultaneously for several
years, they were used in conjunction in what is called tandem mode. The interval between
their passes was one day. The effect of time in interferometry is discussed in Section 3.3.6.
In Figure 2.7, the altitude of the first antenna above the ground is H and θ1 and θ2
are the incidence angles for radars one and two. R1 and R2 are the distances from the
antennas to the same point of the ground. The separation distance between their flight path
is called the baseline, B, and its projection perpendicular to the radar line-of-sight is called
the perpendicular baseline B⊥ .
A complex signal returned from a pixel located at range R and azimuth x can be written
2.3. SAR AND INSAR 23

A2

∆R
B
A1
B
R 2

R 1

θ1

H θ2

Figure 2.7: InSAR viewing geometry. The two circles denote the locations of the antennas
on their passes. The imaging pixel is at height z and its distances to antenna one and two
are R1 and R2 , respectively. B⊥ is the component of the baseline perpendicular to the look
direction.

as
s(R, x) = |s(R, x)|e jφ (R,x) (2.1)

where φ is the phase of the pixel. The phase of the signal can be rewritten as

−4π
φ (R, x) = R + φscat (R, x) (2.2)
λ

where the first term on the right side of Equation 2.2 is the phase due to the round-trip travel
between the antenna and the center of the pixel and φscat is the phase due to the scattering
from different scatterers in the resolution cell.

An interferogram is formed by multiplying one SAR image by the complex conjugate


of the other. Figure 2.6.b shows an example interferogram over the same area as the SAR
image in Figure 2.6.a. This interferogram is formed from ERS1-2 tandem pair images
24 CHAPTER 2. SAR AND INSAR REMOTE SENSING

acquired on January 30th and 31st, 1996. The colored fringes show the interferometric
phase values between 0 and 2π .
The interferometric phase is

−4π
φint (R, x) = ∆R + φscat,1 (R1 , x1 ) − φscat,2 (R2 , x2 ) (2.3)
λ

where the difference between the ranges, ∆R, is shown in Figure 2.7. Because the view-
ing geometry changes very little in terms of scattering properties of the medium, φscat,1 ≈
φscat,2 . However, as we see in the following Chapters, the small difference between φscat,1
and φscat,2 is the main point of this thesis. This difference gives some useful information
about the scattering properties of the medium. We consider this phase difference as part
of a residual phase, φnoise . The total interferometric phase of a pixel is a superposition of
phases due to several processes

φint (R, x) = φtopo + φdefo + φprop + φnoise (2.4)

where φtopo is the phase due to the height of the resolution cell above a reference surface for
instance z in Figure 2.7. The topographic phase is given by (Zebker and Goldstein, 1986)

−4π B⊥
φtopo = z (2.5)
λ R1 sin θ

Here θ is the look angle of the satellite, nominally 23◦ for ERS-1, and z is the height of the
resolution cell above the reference surface.
φdefo in Figure 2.4 is the phase due to the amount of coherent line-of-sight motion that
occurs between observations. In the case of coherent motion all the scatterers in a resolution
cell have the same velocity vector. Therefore InSAR is most sensitive to large-scale (> 100
m) crustal deformation or glacial flow.
φprop in Equation 2.4 is the signal phase shift due to signal propagation in the ionosphere
and atmosphere (Onn and Zebker, 2006). Any other sources of interferometric phase are
lumped into φnoise . The specific application of InSAR determines which terms in Equation
2.4 are important. To make an accurate geophysical measurement from one of the three
terms on the right hand side of Equation 2.4, that term must be determined by combining
2.4. SUMMARY 25

other interferograms or by a priori information. While φprop is small over Greenland, φdefo
can be large and should be considered in glacier motion. The topography of the interior
of Greenland, which is the area of our study in this thesis, varies very little inside a scene
(100km × 100km) and so φtopo can often be neglected. We show in the next Chapters,
however, that there is much useful information in φnoise .

2.4 Summary
In this Chapter, we first discussed how radar remote sensing helps us study the cold and
remote ice sheets of Greenland. We showed that the intensity image of the dry snow zone
of Greenland is insufficient by itself to estimate the accumulation rate. In the following
Chapters, we quantify the information contained in the φnoise by the interferometric cor-
relation. We show that the InSAR observables, intensity and correlation, together resolve
ambiguities in estimating the accumulation rate.
26 CHAPTER 2. SAR AND INSAR REMOTE SENSING
Chapter 3

InSAR Observables: Normalized


Backscattered Power and Correlation

3.1 Introduction
In the previous Chapters we showed how radar remote sensing methods yield data that
lead to a better understanding of the facies and changes in the Greenland ice sheets. We
also showed that volume scattering is significant, which allows us to infer properties of
the ice beneath the topmost surface. We now show that two radar observables permit us
to probe the sub-surface of the ice sheet from space. This is important, because as we
see in Chapter 4, knowledge of vertical structure of the snow allows us to estimate the
accumulation rate, which is the most sought-after mass balance term. In this Chapter we
study these two important InSAR observables: normalized backscattered power, σ 0 , and
correlation, ρ . Much of the material and derivations related to these observables in this
Chapter is a duplication of Hoen (2001)’s dissertation as referenced in the text.
We first describe in Section 3.2 the backscattered power in more detail. In order to
model the backscattered power in the following Chapters, we show how to normalize it to
form σ 0 in Section 3.2.1.
In Section 3.3, we introduce a new InSAR observable called interferometric correlation,
which is the coherence between corresponding pixels in two different SAR images. We
study the correlation and how it depends on the physical separation of the two observations

27
28 CHAPTER 3. INSAR OBSERVABLES

forming the interferograms, i.e., the spatial correlation function, in Section 3.3.1. We also
develop a model that relates spatial correlation to both the backscattering characteristics
of the medium and geometry of the antenna with respect to the ground. In Sections 3.3.2-
3.3.6, different sources of the correlation are discussed.

3.2 Introduction to backscattering power

Ro
z
θ
y
x {xo , yo , zo }

f(x,y,z)
ε

Figure 3.1: SAR viewing geometry (Hoen, 2001)

First, we model the expected return from a single representative pixel of a focused SAR
image. Radiowave scattering from terrain is often restricted to surface scatter, however in
the case of ice and snow terrain, the wave readily penetrates the surface. Therefore, we
consider scattering from a volume as defined by a reflectivity function f (x, y, z). In the
antenna far field, the echo from a scatterer at position (x, y, z) (see Figure 3.1) is
 4π √ 
sscatterer (x, y, z) = f (x, y, z) exp − j (R0 + y sin θ − ε z cos θr ) Wa (x)Wr (y, z) (3.1)
λ

where the reflectivity f is zero for all z > 0, Wr (y, z) is the radar range impulse response,
3.2. INTRODUCTION TO BACKSCATTERING POWER 29

and Wa (x) is the impulse response in the azimuth direction. The phase in the exponential
accounts for two way wave travel, with θ (23◦ in ERS data) and θr as the incident and
refracted angles, respectively. By assuming Gaussian distributed scatterers, the reflectivity
function is a white random process which has an autocorrelation function given by

< f (x, y, z) f ∗ (x′ , y′ , z′ ) >= σv (x, y, z)δ (x − x′ , y − y′ , z − z′ ) (3.2)

where σv (x, y, z) is the volumetric radar cross section. From Equations 3.1 and 3.2, the
expected power received from the representative pixel reduces to (Zebker and Villasenor,
1992; Hoen, 2001)
Z +∞ Z +∞ Z +∞

< ss >= σv (x, y, z)|Wa(x)|2 |Wr (y, z)|2dxdydz (3.3)
−∞ −∞ −∞

If the medium exhibits volume scatter, then the depth dependence of the volume cross-
section must be taken into account. For a volume-scattering only model of ice sheets, we
can assume that the medium is an isotropic, infinite half space with extinction coefficient
κe , in units of m−1 . The extinction coefficient embodies both scattering and absorption
losses. With this assumption the volumetric backscatter coefficient for a negative z within
the medium becomes

2zκe
σv (x, y, z) = σv0 e cos θr (3.4)
κe = κs + κa (3.5)

where κs and κa are the scattering and absorption coefficients, respectively. In Equation 3.4,
σv0 is the average cross section per unit volume, and the propagation losses are contained in
2
the exponential, where the factor cos θr in the exponent accounts for the two way off-vertical
propagation of the wave within the medium. As we will see in Section 4.2.1, σv , κs and κa
also depend on depth and thus we express backscattered power as (Hoen, 2001)
Z 0
δ xδ r 2
Rz κe (z′ )
0 [ cos θr (z′ ) ]dz


< ss > = σv0 (z)e dz (3.6)
| sin θ | −∞

where δ r and δ x are the range and azimuth resolution, equal to 10 m and 6m for ERS
30 CHAPTER 3. INSAR OBSERVABLES

data, respectively. Often a SAR image is normalized to produce unit-less σ 0 values and
thus remove explicit dependence on resolution. Backscatter power images of the ice sheets
usually possess normalized σ 0 values ranging from almost unity down to -15 dB.

3.2.1 Normalized backscattered power images

The backscattered power images presented in this dissertation are from the same data used
for InSAR observations. Pixel i, j of the backscattered power image is given by averaging
the backscattered power of the two SAR images 1 and 2, according to
q
P(i, j) = < s1 (i, j)s∗1(i, j) > · < s2 (i, j)s∗2(i, j) > (3.7)

The backscattered power received at the antenna is given by the radar equation as (Curlan-
der and McDonough, 1991)

Pt G 1
Pr = · σ 0 Ai · Ar (3.8)
4 π R0
2 4π R20

where Pr and Pt are the received and transmitted powers, respectively, G is the antenna gain,
Ar is the effective area of the antenna, σ 0 is the normalized cross section of the target, and
Ai is illuminated surface area of the target, respectively. For SAR power images we can
rewrite Equation 3.8 as

Pt G 0 R0 λ cτ 1 λG
Pr = · σ ( ) · (3.9)
4π R20 La 2 sin θ 4π R20 4π

where the term in the parentheses is the illuminated area of a pixel in the SAR image and
last term in the right side of the equation is Ar in terms of G (Curlander and McDonough,
1991). In this equation, λ is the wavelength of the microwave signal, τ is the pulse length,
and La is the antenna length in the azimuth direction. As mentioned above, in order to ac-
count for geometric and radiometric distortions, radar power images are often calibrated to
form σ 0 . In order to calibrate the ERS power images, we need to find the proper calibration
3.3. CORRELATION 31

factor for ERS data. From Equation 3.9, we know that

R30 sin θ
σ 0 ∝ Pr (3.10)
G2

28

880
26
R (km)

24

θi
860
0

22

20
840

18
0 100 200 300 400 0 100 200 300 400
Pixel number pixel number
(a) (b)
Figure 3.2: (a) The distance between the antenna and the ground and (b) the incidence
angle, along the range for ERS data. We will use these profiles to calibrate the power
images to σ 0 .

Figures 3.2.a and 3.2.b, show the incidence angle, θ , and the distance between the
antenna and the resolution cell on the ground, R0 , of the ERS data for different pixels across
range. The two way antenna gain as a function of ERS incidence angle, G2 , is shown in
Figure 3.3 (Laur et al., 2003). Using the profiles in Figures 3.2 and 3.3 in Equation 3.10,
we determine the calibration factor along the range for ERS data which is shown in Figure
3.4. The ratio between the value found in Equation 3.10 and true σ 0 from a known target is
the calibration factor. We calibrate our power images with σ 0 using the calibration factor
in Figure 3.4.

3.3 Correlation
SAR interferometry images are formed by cross correlating two complex SAR images of
the same terrain, and provide more information about the terrain than is available in a single
image. For a second SAR image to provide additional information, at least one imaging
32 CHAPTER 3. INSAR OBSERVABLES

0
Gain (dB)

−1

−2

−3
20 22 24 26 28
θi

Figure 3.3: 2-way gain of the ERS-1 antenna vs. incidence angle (Laur et al., 2003).

parameter must be different, such as flight path, time of observation or wavelength. In


the ERS system the cross track positions of the two observations are often slightly dif-
ferent. This leads to another observable beside power and phase, called correlation. The
correlation of an interferogram is a statistical comparison between pixels in two different
SAR images. Mathematically, the correlation ρ between two signals s1 and s2 acquired on
subsequent radar passes over the same area, is defined as

| < s1 s∗2 > |


ρ=p (3.11)
< s1 s∗1 >< s2 s∗2 >

In practice, the correlation is evaluated by approximating the expectation by a spatial aver-


age of neighboring pixels,

|ΣN s1 (k)s∗2(k)|
ρ=q
e k=1
(3.12)
ΣN
k=1 s1 (k)s1 (k) Σk=1 s2 (k)s2 (k)
∗ N ∗
3.3. CORRELATION 33

−9

Calibration Factor (dB)


−10

−11

−12

−13

−14
20 22 24 26 28
θi

Figure 3.4: Calibration factor of ERS power data vs. incidence angle.

where ρe is the measured correlation and N is the total number of pixels averaged in azimuth
and range.
The correlation coefficient depends on imaging geometry (such as incidence angle, per-
pendicular baseline, etc.) and surface properties (such as correlation length of the surface,
auto correlation function of the surface, etc.).
The total correlation, i.e. the observed correlation image, consists of the product of the
individual correlations (Zebker and Villasenor, 1992)

ρ = ρspatial · ρthermal · ρtemporal · · · · (3.13)

As we will see in Section 3.3.1, ρspatial itself is divided into ρvolume and ρsur f ace . We will
show how to estimate different kinds of decorrelation in Sections 3.3.1-3.3.6. We will also
see that the only source of the correlation that contains the information about the medium
is ρvolume . In this thesis, we will use ρvolume along with σ 0 to estimate the accumulation
rate in the dry snow zone of Greenland. We will remove the other sources of decorrelation
34 CHAPTER 3. INSAR OBSERVABLES

A2
A1
B
B

R2 θ1
R1
θ2

z
y
x
ε
Resolution Element

Figure 3.5: InSAR viewing geometry (Hoen, 2001)

from the total correlation to be left with ρvolume .

3.3.1 Spatial correlation

Repeat-pass ERS observations are separated in both time and space. We will study the
effect of time on correlation in Section 3.3.6 and assume for the moment the observations
are simultaneous. As shown, in Figure 3.5, the two antennas forming the interferograms are
physically separated with two different incidence angles. We call the correlation dependent
on this angular separation spatial correlation.
Consider the viewing geometry shown in Figure 3.5 where A1 and A2 are the two anten-
nas forming the interferometer, with the angle of incidence for each θ1 and θ2 respectively.
We again assume a scattering medium with the reflectivity function f (x, y, z), and include
an effective permittivity for the firn ε . The distances from the center of the resolution
element to the radar at the two imaging positions are R1 and R2 .
We assume that the scatterers are randomly spread within a refracting half space. Ex-
tending the approach derived by Zebker and Villasenor (1992), the signal received at the
3.3. CORRELATION 35

antenna A1 from a resolution cell can be expressed by the sum of complex echoes from
different scatterers in antenna far-field
Z 0 Z ∞Z ∞ √

(R1 +y sin θ1 −z ε cos θr1 )
s1 = f (x, y, z)e− j λ ·Wx (x)Wr (y, z)dxdydz (3.14)
−∞ −∞ −∞

where θr1 denotes the refracted angle in the medium. Likewise, the corresponding signal
received at antenna A2 is
Z 0 Z ∞Z ∞ √
4π (R
s2 = f (x, y, z)e− j λ 2 +y sin θ2 −z ε cos θr2 )
·Wx (x)Wr (y, z)dxdydz (3.15)
−∞ −∞ −∞

The cross-correlation of these two signals is


Z Z Z Z Z Z
− 4λπ (R1 −R2 )
s1 s∗2 = e f (x, y, z) f ∗ (x′ , y′ , z′ )Wx (x)Wx∗ (x′ )Wr (y, z)Wr∗ (y′ , z′ )
4π √
(y sin θ1 −y′ sin θ2 − ε (z cos θr1 −z′ cos θr2 ))
e− j λ dxdydzdx′ dy′ dz′ (3.16)

By assuming uniformly distributed and uncorrelated scattering centers as stated in Equation


3.2, we can simplify Equation 3.16 to (Hoen, 2001)
Z Z Z
4π 4π (y cos θ δ θ +z√ε sin θ
rδ θ )
< s1 s∗2 >= e− λ (R1 −R2 ) σv (x, y, z)e− j λ ·|Wx (x)|2 |Wr (y, z)|2dxdydz

As an approximation to the physical situation, we also assume that the volumetric cross
section of the scatterers within each resolution cell varies with depth only. We can then
express < s1 s∗2 > as (Hoen, 2001)
Z Z
− 4λπ (R1 −R2 ) δr 2 − j2πκy y
< s1 s∗2 >= e ·δx· |Wr (y)| e dy σv (z)e− j2πκzz z dz (3.17)
sin θ

where
2 δr
κy = cos θ δ θ (3.18)
λ sin θ
and √
2 εδ θr
κzz = (3.19)
λ sin θr
36 CHAPTER 3. INSAR OBSERVABLES

Here δ θ = θ2 − θ1 = B⊥
R where B⊥ is the perpendicular baseline in Figure 3.5. Conse-
quently, substituting Equation 3.17 into Equation 3.11, we can express the spatial correla-
tion as R +∞ R0
|Wr (y)|2 e− j2πκy y dy −∞ σv (z)e
− j2πκzz z dz
ρspatial = | −∞
R +∞ |·| R0 | (3.20)
−∞ |Wr (y)|2 dy −∞ σv (z)dz
The first term in Equation 3.20 depends only on surface scattering characteristics and is
called ρsur f ace . The second term on the other hand, describes the scattering medium,
specifically the vertical dispersion of scatterers producing the return echo. We call this
term ρvolume . The usefulness of ρsur f ace and ρvolume is that we gain information about the
inside of the scattering medium using ρvolume . In the next section we will show how these
correlations behave in the dry snow zone.

3.3.2 Surface correlation

The notation surface correlation is somewhat misleading because even for the case where
no surface scattering occurs, this term still exists in the total correlation. It is more accurate
to consider ρsur f ace as the decorrelation factor related to the y variable. As explained above,
surface correlation is defined as
R +∞
|Wr (y)|2e− j2πκy y dy
ρ =| −∞
R +∞ | (3.21)
−∞ |Wr (y)|2 dy

It is clear from this equation that ρsur f ace depends only on the surface scattering properties
and there is no x or z dependence. Note that surface correlation is the ratio of the Fourier
component of |Wr2 | at the frequency κy to its DC component. Assuming a spectral window
with a sinc impulse response, and setting δ θ = B⊥
R0 , the Fourier transform in Equation 3.21
results in a scaled triangle function
( 2|B |δ r
1 − R0 λ |⊥tan θ | |B⊥ | < B⊥,crit
ρsur f ace = (3.22)
0 |B⊥ | > B⊥,crit
3.3. CORRELATION 37

1 0.86

0.8
0.84

ρsurface
0.6
ρsurface

0.82
0.4

0.8
0.2

0 0.78
0 200 400 600 800 1000 1200 19 21 23 25 27
B⊥ θ
i
(a) (b)
Figure 3.6: (a) ρsur f ace vs. B⊥ and (b) ρsur f ace vs. θi .

where B⊥,crit is the baseline where the correlation goes to zero and is called the critical
baseline. The critical baseline is equal to

R0 λ | tan θ |
B⊥,crit = (3.23)
2δ r

For ERS data, B⊥,crit is about 1100 m.

In Figure 3.6.a we have shown how the surface correlation changes with respect to the
perpendicular baseline for ERS data. As expected, the coherence is unity when there is
no change in the geometry, i.e. B⊥ = 0. As the baseline increases, the surface correlation
decreases linearly. As we will see in Section 3.3.3, surface correlation decays with baseline
more slowly than the volume correlation.

As seen in Figure 3.2, both R0 and θi vary across the range. Therefore, the surface
correlation in Equation 3.22 changes with range also. Figure 3.6.b shows how the surface
correlation changes as a function of the incidence angle for the ERS data. The surface
correlation profile along the range in Figure 3.6.b is needed in next Chapters to remove the
surface correlation dependence from the observed total correlation.
38 CHAPTER 3. INSAR OBSERVABLES

3.3.3 Volume correlation

We now explore the second term in Equation 3.20, ρvolume . Volume correlation is defined
as R0 − j2πκzz z dz
−∞ σv (z)e
ρvolume = | R0 | (3.24)
−∞ σv (z)dz

Note that ρvolume is a measure of the extent of the scatterers contributing to the return echo
in the z direction. The greater the vertical extent of scattering, the faster the correlation
drops with increasing B⊥ . By using Equation 3.4 and considering the dependence of κs and
κa on depth (as will be shown in Section 4.2.1), ρvolume can be derived as (Hoen, 2001)

R z κe (z′ ) 4π R z ε (z′′ )δ θr (z′′ ) ′′
R 0 σ 0 (z)e2 0 cos θr (z )

′ dz − j λ 0
e sin θr (z′′ )
dz
dz

ρvolume = −∞ v R z κe (z′ ) (3.25)
R0 2 0 cos θr (z′ ) dz

−∞ σv (z)e
0 dz

As we can see in this equation, ρvolume gives information about the scattering medium,
specifically, the vertical dispersion of scatterers comprising the return echo. In the follow-
ing Chapters we will show how ρvolume , along with σ 0 , helps us gain some information
about the structure of the ice and snow through the depth.
Note that using the term volume correlation does not mean that the surface scattering
does not contribute at all, but that it depends on the depth z. For instance, as we will see in
Chapter 5, surface scattering from buried layers in the snow has important contributions to
ρvolume . In general, σv0 (z) in Equation 3.25 can be written as

σv0 (z) = gvolume (z) + ΣM


m=1 gsur f ace,m δ (z − zm ) (3.26)

where gvolume (z) is the contribution of volume scattering only in the backscattering, gsur f ace,m
is the surface backscattering from the mth layer, zm is the depth of the mth layer in the snow
and M is the number of layers in the snow. We will specify gvolume (z) and gsur f ace,m using
scattering models in Chapter 4.
Using Mätzler’s model (see Section 4.2.1), we calculate gvolume (z), κe and ε . We ignore
any surface scattering from the layers in the snow for the sake of simplicity. Substituting
these values in Equation 3.25, we depict ρvolume variation with respect to the perpendicular
3.3. CORRELATION 39

R = 1.8 mm
0.8 0

0.6
ρvolume

0.4

0.2
R0 = 0.8 mm

0
0 200 400 600 800 1000
B⊥ (m)

Figure 3.7: ρvolume vs. B⊥ for grain radii equal to 0.8 and 1.8.

baseline in Figure 3.7. In this figure, we plot two curves of Equation 3.25 for different grain
radii. As will see in Section 4.2.1, different snow average grain radius, R0 , leads to different
backscattering properties. The volume correlation in Figure 3.7 drops exponentially with
perpendicular baseline, while the surface correlation in Figure 3.6.a drops linearly. There-
fore, the usable range of B⊥ is constrained wherever the radio waves penetrates through the
ground. Because the volume correlation drops so quickly with the B⊥ , we will confine our
data to B⊥ < 250. Nonetheless, B⊥ must be high enough for good contrast in correlation
between the different areas in the dry snow zone. Therefore, we will only use correlation
data with the perpendicular baseline of 100 < B⊥ < 250 in the following Chapters.

3.3.4 The spectral approach to interferogram generation


Now we present another derivation of critical baseline based on a spectral approach. Here
we are going to show that the spectra of the signals s1 and s2 correspond to different bands
of the ground reflectivity’s spectrum. As shown before (Equation 3.14, 3.15), radar return
40 CHAPTER 3. INSAR OBSERVABLES

is a function of the complex reflectivity f (x, y, z)


Z Z √
2π f
(y sin θ − ε z cos θ )
s(2π f , θ ) = f (y, z)e− j2 c dydz (3.27)

Wavenumbers of the ground range, βy , and in elevation, βz , have the following form

2π f
βy = 2 sin θ (3.28)
c
2π f √
βz = 2 ε cos θ (3.29)
c

Thus the received signal can be regarded as a 2-D Fourier transform of the ground
reflectivity at the wavenumber given by Equations 3.28,3.29. Two SAR images with two
different angles (θ and θ + δ θ ) represent different bands of reflectivity spectrum.

In order to show how the spectrum of the backscattered signal is affected, we con-
sider the simple case where no penetration through the ground exists, i.e. s(2π f , θ ) =
R 2π f
βy y dy.
f (y)e− j2 c The variation of βy , δ βy , caused by a slight change of the looking
angle δ θ would be
2π f δ θ
δ βy = 2 cos θ (3.30)
c
Therefore, a look angle difference δ θ generates a shift and stretch in the terrain spectra. In
order to compare the shift in the reflectivity spectrum to the SAR bandwidth BW , we find
the equivalent frequency shift. From 3.28 we know that

2π f
δ βy = 2 sin θ (3.31)
c

Comparing this equation with Equation 3.30, we obtain the shift in the frequency (Gatelli
et al., 1994):
fδθ
δf =− (3.32)
tan θ
Using f = c
λ and δ θ = B⊥
R0 we have

cB⊥
δf =− (3.33)
R0 λ tan θ
3.3. CORRELATION 41

Note that Equation 3.33 does not mean that by changing the look angle of the SAR system
the bandwidth is shifted by δ f . Instead, it implies that by changing the look angle, the
backscattered signal contains different spectral component of the ground reflectivity spec-
trum. In other words, the specific spectral components of of the first signal are shifted by
δ f in the second spectrum. The images will be highly correlated if the shift in the spectrum
is not large. However, if the amount of the shift in the frequency spectrum is larger than the
bandwidth of the signal, then we will have no correlation between the images. Substituting
δ f = BW and δ r = c
2BW in Equation 3.33, we express the critical baseline as

R0 λ tan θ
B⊥,crit = (3.34)
2δ r

which is identical to Equation 3.23.

3.3.5 Thermal correlation

The fact that the noise in the two images does not correlate (white Gaussian noise) leads
to some decorrelation called ρthermal . By assuming white Gaussian noise in the receiver,
ρthermal is (Zebker and Villasenor, 1992)

1
ρthermal = 1
(3.35)
1 + SNR

where SNR is the signal to noise ratio of each pixel in the image.

For ERS data, the thermal noise equivalent, σnoise


0 , which is the σ 0 where we receive no
signal but the noise, is -23 dB. In Figure 3.8 we have shown ρthermal as a function of σ 0 .
As we can see in this figure, unless σ 0 is very low there is negligible thermal decorrelation.
For most of the ERS correlation images we will analyze in the dry snow zone of Greenland,
σ 0 is high enough to get a very high thermal correlation term (ρthermal ≈ 1). Yet, we will
remove thermal correlation from the total correlation in next Chapters using the Equation
3.35 and σ 0 of the image.
42 CHAPTER 3. INSAR OBSERVABLES

0.9

0.8
ρthermal

0.7

0.6

0.5
−25 −20 −15 −10 −5 0
0
σ (dB)

Figure 3.8: ρthermal vs. σ 0 (dB).

3.3.6 Temporal correlation

The correlation coefficient related to the changes in the scattering medium between the
observations is called ρtemporal . These changes could arise from deviations in the position,
size or shape of the scatterers or deviations in the dielectric properties of the scattering
medium (Zebker and Villasenor, 1992).
The firn in the dry snow zone shows a very pronounced decrease of coherence with
time at C-band. The main reason for the temporal variation is the changes in the surface
conditions, which can lead to significant surface decorrelation. The major factor in the
dry snow temporal decorrelation is the movement of snow/ice particles by the wind (Rott
and Siegel, 1996), and snow precipitation in the winter time. Although the backscattered
power is stable in time, and most of the signal originates from the volume of snow, the
surface term is important for phase changes and therefore for the correlation. For the wet
snow zone and percolation zone, liquid water plays an important role. Different amounts of
liquid water lead to different scattering properties and thus different temporal correlation
3.4. CONCLUSIONS 43

from one observation to another.


Areas of dense vegetation often suffer a high temporal decorrelation due to the move-
ments of leaves and branches in the wind, as well as their growth over time (Zebker and
Villasenor, 1992).
Oveisgharan and Zebker (2004) have shown that the correlation obtained from RADAR-
SAT-1 fine beam-1 mode was not high enough to be used in this study. The repeat cycle
of RADARSAT-1 is 24 days, which causes high temporal decorrelation. For the reminder
of this dissertation, we use data from the dry snow zone from the winter months, to avoid
melt possibilities. Also, the data that we have used in this work is from ERS-1,2 tandem
pair with separation time of one day. With such a short repeating and with observations in
the winter time over the dry snow zone, we can assume that ρtemporal ≈ 1.

3.4 Conclusions
In this Chapter we have examined both backscattered power and correlation in more detail.
We showed how volume correlation theoretically, provides information about the snow
structure below the surface. We have described other sources of correlation and demon-
strated methods to remove them from the total correlation, leaving the volume correlation
from which we will derive accumulation rates. The correlation data presented in the fol-
lowing Chapters is the volume correlation only with other sources of correlation removed.
Unless otherwise stated in the following Chapters, we mean volume correlation when the
word correlation is used.
44 CHAPTER 3. INSAR OBSERVABLES
Chapter 4

Modeling InSAR Echoes From Ice


Sheets

4.1 Introduction
In this Chapter, we analyze two types of radar scattering models for deriving the accumula-
tion rates in the dry snow zone of Greenland. First, we explain a volume scattering model in
Section 4.2 and compare its predictions to ERS backscattered power and correlation data.
We find that this model does not provide an accurate match to the data. Therefore, we
propose a volume and surface scattering model in Section 4.3 which we find explains the
radar data more accurately, and resolves some of the ambiguities encountered in estimating
accumulation rate using the first model.

4.2 Volume backscattering models


As explained in Section 2.2, volume scattering from snow grains has been proposed as the
dominant scattering mechanism for C-band observations in the dry snow zone of Green-
land. In this approach, fluctuations in backscattered power are related to the variation in
mean grain size, which is inversely related to the accumulation rate (Munk et al., 2003;
Tsang et al., 1985; Hoen and Zebker, 2000). The Measured microwave emission and
backscattering signature of snow-packs motivated the advanced theoretical modeling of

45
46 CHAPTER 4. MODELING INSAR ECHOES FROM ICE SHEETS

volume scattering in low-loss medium (Mätzler, 1998; Tsang and Kong, 1980, 1981; Tsang,
1987; Stogryn, 1986). For dense media, such as snow, the quantitative treatment of volume
scatter has been a difficult task. Because there is a relatively large difference between the di-
electric permittivities of air and ice, the conventional Born approximation is not sufficiently
accurate. Models have been proposed to cope with this situation, using strong fluctuation
theory (Tsang and Kong, 1981), or SFT, and by dense-medium radiative transfer theory
(Tsang et al., 1985), or DMRT. Applying these models to the Greenland ice sheet requires
a complex theory. However, we can use a simplified DMRT based on Tsang et al. (1985)
in this Chapter to test its results. In a simpler approach, Mätzler formulated an improved
Born approximation model using a quasi-static model for the internal field.

4.2.1 The Tsang and Mätzler backscattering models

As mentioned above, previous radar estimates of accumulation have used volume-only


backscatter models in the dry snow zone (Mätzler, 1998; Tsang et al., 1985). These models
are based on the grain migration process, whereby smaller snow grains grow larger with
time, so that older firn is composed of larger particles and hence is radar-bright (Figure
1.5). In areas of high accumulation, there is not time for significant migration as grains are
buried, so these areas remain radar-dark.
At common firn densities, the snow grains are so closely packed that the usual Rayleigh
model is not applicable. The DMRT model (Tsang et al., 1985) accounts for the interaction
~ s from neighboring grains, so that the fields from neighboring grains
of scattering fields E
are no longer considered independent, or, more formally,

< E~s,i E~s, j >i6= j 6= 0 (4.1)

where E~s,i is the scattered field from the i th grain. In the case of the DMRT model devel-
oped by Tsang et al. (1985), for scattering from a dense medium containing same size and
permittivity spheres, κs is

(1 − ν )4
κs = 2ν k04 R3 y2a 3 1 (4.2)
(1 − ν ya ) 2 (1 + 2ν )2 (1 + 2ν ya ) 2
4.2. VOLUME BACKSCATTERING MODELS 47

where
ε2 − ε1
ya = (4.3)
ε 2 + 2ε 1
ρf
Here ν = ρi is the volume fraction of the firn, ρ f and ρi = 0.917 cmg 3 are the densities of
firn and ice, respectively, and ε2 and ε1 = 1 are the real permittivities of dry snow and air

respectively. In Equation 4.2, R is the radius of snow particles, k0 = λ0 is the wavenumber
of the radiated field in air. ε2 is modeled by Mätzler (1987) as

ε2 = 1 + 1.7ρ f + 0.7ρ 2f (4.4)

For an idealized solid block of uniform pure ice (ν → 1), we might expect the volume
backscattering to approach 0 as Equation 4.2 predicts. In fact, as ν approaches to zero, κs
is reduced to 2ν k04 R3 ( εε22+2
−ε1 2
ε1 ) , which is the classical Rayleigh expression for the scattering
coefficient (Ishimaru, 1978).

Mätzler (1998) used the improved Born approximation model to compute volume back-
scatter from grain-size depth profiles. It is a simplified SFT where the mean field is assumed
to be given by quasi-static dielectric mixing theory. The advantage of this model is its
simplicity and its potential applicability to natural medium consisting of complex particles,
such as snow firn.

For freely arranged snow grains, a good assumption in the dry snow zone, the scattering
coefficient κs , backscattering cross section σ 0 and absorption coefficient κa are computed
using Equations 4.5-4.7 below, where by free arrangement we mean that the probability
of finding a particle anywhere outside of a test particle is equal to the volume fraction of
particles.

2R3 4
κs = ν (1 − ν )(ε2 − ε1 )2 K 2 k (4.5)
9 0
3κ s 2
σ0 = sin χ (4.6)
2
′′
κa = ν k0 ε2 K 2 (4.7)

ρf
Here ν = ρi , ρ f , ρi , ε2 , ε1 , R and k0 are defined in Tsang’s model above. The imaginary
48 CHAPTER 4. MODELING INSAR ECHOES FROM ICE SHEETS

′′
permittivity of dry snow, ε2 , can be expressed as (Mätzler, 1987)

′′ 1 p
ε2 = 1.59 × 106(0.52ρ f + 0.62ρ 2f )( + 1.23 × 10−14 f )e0.036T (4.8)
f

where f is the frequency of the radiated field and T is the temperature of the snow. Finally,
K 2 is computed from (Mätzler, 1998)
2
1 3 ε
∑ εa + (ε2 − ε1 )A j
a
K2 = (4.9)
3 j=1

In Equation 4.9, εa , called the apparent permittivity, is the permittivity of the ambient
background which appears locally to a given particle of permittivity ε2 and is defined by
Mätzler (1998) as
εa = εe f f (1 − A j ) + ε1 A j (4.10)

A j in Equations 4.9 and 4.10 is the depolarization factor along axis j. In dry snow A1 =
A2 = A, A3 = 1 − 2A and A is (Mätzler, 1998)

 0.5ν + 0.1 ν ≤ 0.33


A= 0.18 + 3.24(ν − 0.49)(ν − 0.49) 0.33 ≤ ν ≤ 0.71


 1
ν > 0.71
3

In Equation 4.10 εe f f , the effective permittivity, can be approximated in dry snow by


(Mätzler, 1998)
ν (ε2 − ε1 ) ∑3j=1 εa +A jε(aε2 −ε1 )
εe f f = ε1 + A
(4.11)
3 − ν (ε2 − ε1 ) ∑3j=1 εa +A j (εj
2 −ε1 )

Using Equations 4.5-4.11 we are ready to compute backscattered power (σ 0 ) and correla-
tion (ρ ) in the dry snow zone for volume-only backscatter.
It is important to note that in dry snow over the depth of C-band penetration the density,
grain radius and temperature of snow are not constant. These variations make κs , σ 0 and
κa functions of depth also, according to Equations (4.5-4.7). Typical profiles are depicted
in Figures 1.3, 1.6 and 1.7 respectively. These graphs are not unique and vary by location.
As will be discussed in Section 4.5.1, of these factors the only parameter with significant
4.2. VOLUME BACKSCATTERING MODELS 49

effect on accumulation estimates is the variation of temperature, which affects the density
profile shown in Figure 1.3.

4.2.2 Numerical calculation of backscattering cross section and corre-


lation

Since the scattering parameters change with depth, to evaluate the integrals for σ 0 and ρ
we sum layers using different scattering parameters for each layer. From Equations (3.4)
and (3.25) we have


σ0 = ∑ σv (n) (4.12)
n=1

1 ∞ ε (m)δ θr (m)
j 4λπ d ∑nm=1 sin θ (m)
ρ = 0 ∑ σv (n)e r (4.13)
σ n=1

where

3κs(n)d −2d ∑nm=1 κscos


(m)+κa (m)
σv (n) = e θr (m) (4.14)
2

Here d is the layer spacing, ε (m) is the effective permittivity of each layer computed from
(4.11), and κs and κa are derived using (4.5) and (4.7). θr (m) is the angle of the refracted
waves in each layer. From Snell’s law,

sin θi
θr (m) = arcsin( p ) (4.15)
ε (m)
B⊥ cos θi
δ θr (m) = p (4.16)
R0 ε (m) cos θr (m)

Hence knowledge of the density, temperature and grain growth profiles enables the
calculation of σ 0 and ρ for volume-only backscattering.
50 CHAPTER 4. MODELING INSAR ECHOES FROM ICE SHEETS

4.2.3 Results from the volume-only backscattering model


Using the volume backscattering model above we can compute the backscattered power
and correlation in the dry snow zone. As seen in Figure 1.6, the grain radius depends
on the accumulation rate. So the grain radius can be considered the dominant parameter
affecting σ 0 or ρ , if we assume that the temperature and density profiles do not vary much
in the dry snow zone. We use the density and temperature profiles shown in Figure 1.3 and
Figure 1.7 throughout our study area.

R0 = 2.3mm
0.8

0.6
ρ

0.4
R = 0.8mm
0

0.2

0
− 25 − 20 − 15 − 10 −5 0
σ0(dB)

Figure 4.1: ρ vs. σ 0 (dB) for different grain radii for models of Mätzler (1998) (dots) and
Tsang et al. (1985) (solid curve)

In Figure 4.1 we present the backscattered power and correlation for firn with an aver-
age grain radius in the interval 0.8 mm ≤ R0 ≤ 2.3 mm, where the dots represent 0.1mm
steps in R0 . As seen in the plot, greater grain radius, i.e. lower accumulation rate, cor-
responds to brighter returns and also higher correlation. This is a typical conclusion of
4.3. LAYERED BACKSCATTERING MODELS 51

volume scatter models, that brighter return corresponds to lower accumulation rates. As
we will see in next section this is not always true in measured data. A different volume
scatter model, Tsang et al. (1985) model, is also shown (solid line in Figure 4.1). As
we will discuss in Section 4.4.1, Tsang’s model predicts correlation values that are much
higher than are observed.

4.3 Layered backscattering models

We have previously shown in Figure 2.3 that radar brightness and accumulation are not
simply related, hence the above model cannot accurately retrieve accumulation from radar
measurements. Previous attempts to add radar correlation data to the model (Hoen, 2001)
are likewise insufficient, as measurements of backscatter (σ 0 ) and correlation (ρ ) are con-
strained to lie on a single curve in Figure 4.1. Real data are not nearly so well-behaved
(Hoen and Zebker, 2000). Our approach here is to introduce a model with both volume
scatter and also surface scatter from hoar layers at depth. We will find that such a model
using both backscatter and correlation data can indeed reproduce observed accumulation
rates. Among all the various layers at depth in the dry snow zone, wind crusts, granular
layers, sastrugi layers and so forth, hoar layer interfaces present the strongest reflections
because of their high dielectric contrasts. These reflections can be modeled using rough sur-
face scattering methods. As discussed above, hoar layers are formed during the summer,
with thicknesses varying from a few mm to 7 cm. The hoar layer is the most coarsely-
grained type of firn that forms in the absence of melt water, with an average grain size in
the range of 2 to 5mm. A layer of depth hoar is highly porous and its density is low, typi-
cally 0.1 to 0.3 cmg 3 . Below about 6 m depth, hoar layers typically collapse. Since each depth
hoar layer marks the onset of the summer season in a given year, the spacing between these
layers can be related to snow accumulated each year (Alley et al., 1990; Wismann et al.,
1997).
52 CHAPTER 4. MODELING INSAR ECHOES FROM ICE SHEETS

4.3.1 Modeling hoar layers


For the purposes of this paper, we assume annual hoar layers to a depth of 6 m, with a
volume-scattering medium below. Layering is most visible in the first few meters of depth
and rapidly decreases after 4 to 6 m (West, 1994). We note that below this 6 m depth no
layering is typically observed. Thus we use a 6 m depth of layering here although we have
investigated values from 4 to 6 m. A value of 6 m results in a reasonable fit to the data we
will present later and is an acceptable value for the dry snow zone. We assume a variable
depth hoar spacing of h cm and a fixed thickness of 2 cm (Abdalati and Steffen, 1998; Alley
et al., 1990).We show ice sheet physical parameters in Figure 4.2. Under this model, two
kinds of scattering processes occur. First, volume backscatter in the firn and hoar layers
occurs, and second, surface backscatter from layers with different permittivities occurs,
such as at the hoar and firn interface. The hoar layer spacing, h, depends on accumulation
rate – higher accumulation rates imply greater h (Wismann et al., 1997) and with a fixed
6 m depth of layering in our model, larger values of h also imply a smaller number of
layers in the firn (also observed in Rott et al., 1993). The permittivity of the hoar also
depends on accumulation. Rott et al. (1993) found that in regions of lower accumulation
the dielectric contrast of the interface is larger than in higher accumulation areas. The hoar
layering thickness h is inversely related to dielectric contrast between the hoar and firn
layers, because greater h increases the density of the hoar layer, increasing the permittivity
of the hoar. Hence changing h as a parameter in our model changes the dielectric contrast
inversely.

4.3.2 Numerical calculation of backscattering cross section and corre-


lation considering hoar layering
In our model we include volume backscattering as described in Section 4.2.2, and also sur-
face backscatter due to dielectric contrast between firn and hoar layers. The average density
profile in Figure 1.3 is used to compute volume backscattering, and density changes due
to hoar layers are used to derive surface backscattering. The influence of interferometric
phase on the observed correlation follows from the propagation factor in the exponential
in Equation 4.18. For simplicity, we restrict our model to the single scattering case and
4.3. LAYERED BACKSCATTERING MODELS 53

Figure 4.2: Hoar layer region model is 6 m thick with hoar spacing and thickness of h cm
and 2cm respectively. Below this depth we assume no layers are present.

neglect multiple scattering or coherent effects such as resonances between layers. The total
backscattering coefficient and correlation from multiple layers forming firn and hoar are
described by (Oveisgharan and Zebker, 2007):

σ0 = ∑ σv(n) (4.17)
n=1

1 ∞ π ε (m)δ θr (m)

ρ = 0 ∑ σv (n)e λ ∑m=1 sin θr (m)
4
j d n
(4.18)
σ n=1
h 3κ (n)d i  n
1
+ σh0 (n) exp − 2d ∑
s
σv (n) =
2 m=1 cos θr (m)

(κs (m) + κa(m)) Πn−1 m=1 γm (θr (m))
2
(4.19)

Here all parameters are as defined in Section 4.2.2. Additionally, σh0 is the surface backscat-
tering coefficient of firn-hoar or hoar-firn interfaces, and γm is the transmission coefficient
of the m-th layer. σh0 and γm are equal to zero and one respectively wherever there is no
hoar-firn or firn-hoar interface.
54 CHAPTER 4. MODELING INSAR ECHOES FROM ICE SHEETS

h=30cm
h=20cm
h=15cm
0.8
h=10cm

0.6
ρ

0.4

0.2

0
- 20 -15 - 10 -5 0
σ0(dB)

Figure 4.3: ρ vs. σ 0 (dB) for different final grain radii and hoar layer spacing. Each dotted
curve corresponds to a different value of h: 10,15, 20 and 30 cm. Grain radii also change
from 0.8 to 2.3 mm in 0.1 mm steps along each curve as depicted by dots.

When both the surface standard deviation, σl , and correlation length, l, are smaller than
the wavelength, small perturbation methods can be used for calculating the surface scatter-
ing component of backscattering as will be explained below (Equations 4.20–4.22). The
surface standard deviation, σl , includes the frequency components of the surface responsi-
ble for scattering at a given electromagnetic wavelength. Typical values observed for snow
surfaces in the Arctic are l = 5 cm and σl = 3 mm (Bingham, 1997). Equations 4.20-4.22
below are not the only representation possible for calculating the contribution of surface
scatter, σh0 , in Equation 4.19. For example, another possible model for surface scattering
is Kirchoff surface scattering under a scalar approximation (Ulaby et al., 1981; Drinkwater
et al., 2001) with an exponential height correlation function. The dry snow surface scatter-
ing physical characteristics lie between the scattering conditions constrained by these two
4.3. LAYERED BACKSCATTERING MODELS 55

models. We, however, use small perturbation method for calculating the surface scattering
component of backscattering in our model following Ashcraft and Long (2006) and Hoen
(2001). This is a reasonable approach, because of the low surface roughness of the layers
at C-band (Long and Drinkwater, 1994).

Thus, in our model we use the first-order small perturbation backscattering coefficient
for a slightly rough surface at VV polarization (Ulaby et al., 1981)

2π 4 2 4
σh0 = 8( ) σl cos θ |αvv |2W (2ki sin θ ) (4.20)
λi


where λi and θ are the wave number and angle in the incident medium, σl2 is the variance
of surface height, αvv is the VV polarization amplitude given by

sin2 θ − εr (1 + sin2 θ )
αvv = (εr − 1)  2 (4.21)
p
εr cos θ + εr − sin θ 2

εr is the ratio of permittivities between the media at the hoar-firn or firn-hoar interface
ε (below)
forming the rough surface (εr = ε (above) ), and W is the Fourier transform of the interface
autocorrelation function, which we assume to be Gaussian, specifically

1
W (2ki sin θ ) = l 2 e−(ki l sin θ )
2
(4.22)
2

Here l is the correlation length of surface. For hoar layers in the dry snow zone, we use
l = 5 cm and σl = 3 mm as observed by Long and Drinkwater (1994).

We present calculations from our model using the parameters of Section 4.3.2 for sur-
face backscattering and those of Section 4.2.2 for volume scattering.

Looking back to Equations 4.17 to 4.19, the parameters we vary are the grain radius,
R0 , and hoar layer spacing, h. For 0.8 mm ≤ R0 ≤ 2.3 mm and 10 ≤ h ≤ 30 cm, typical
grain size and layer spacing in the dry snow zone, we plot both backscatter and correlation
in Figure 4.3. In contrast to Figure 4.1, we now have a family of curves covering most
pairwise values of these observables. Thus using our model any observed combination of
ρ and σ 0 implies both a layer thickness and a mean grain size.
56 CHAPTER 4. MODELING INSAR ECHOES FROM ICE SHEETS

4.4 Results

4.4.1 Interpretation of Greenland InSAR data


We obtained data from an area in Greenland’s dry snow zone, as shown on the map in Fig-
ure 4.4. The center location of this area is 75.9 N, 36.5 W. ERS-1/2 tandem pair data were
acquired over the strip on 15-16 Feb. 1996. Temporal decorrelation is removed from this
correlation map by comparison with a second ERS-1/2 tandem pair with 10 m baseline from
11-12 Jan. (not shown here); its decorrelation is dominated by one day temporal decorre-
lation because the spatial baseline is near 0. Thermal decorrelation and other sources of
decorrelation are removed by signal to noise ratio analysis, leaving only the volume corre-
lation, as described by Hoen (2001). Also shown are the radar intensity data. On the left is
backscattered power over three different snow zones: the coastal region, the bright perco-
lation zone and the darker dry snow zone in the interior where we focus our work. At right
is a correlation image from the 15-16 Feb. tandem pair, where the baseline ranges from
134m at the coast down to 124m in the interior.
Detailed dry snow zone volume correlation and power backscatter maps are shown on
Figures 4.5(a) and 4.5(b) from the indicated inset region. To examine the range of σ 0 and ρ̃
for this region, we plot backscatter intensity and correlation data in Figure 4.6 and compare
to predictions of the Mätzler and Tsang models shown in Figure 4.1. Though neither model
matches all of the data well, Mätzler’s model follows the data more closely. Tsang’s model
predicts correlations that are significantly too high. Hence for the purpose of our study, we
will adopt Mätzler’s approach for volume backscatter.
Each resolution cell in the inset region yields one pair of correlation and normalized
power backscattered (ρ̃ , σ 0 ), as plotted in Figure 4.6. The observed range of σ 0 and ρ̃ in
Figure 4.6 exceed that covered by the Mätzler volume scatter model alone, again demon-
strating the volume model’s inaccuracy for estimating accumulation. A sensitivity analysis
(Section 4.5.1) shows that the only parameters in the dry snow zone whose variation can
affect the volume-only scattering model appreciably are grain radius or temperature. But,
Figure 4.1 shows that varying the grain radius can cover only a small portion of our data in
Figure 4.6. Topography changes very little in the interior of Greenland limiting the range of
temperatures as well. Maximum temperature variations in the dry snow zone also change
4.4. RESULTS 57

Figure 4.4: Data acquisition area is a strip in Greenland map with size 800 × 100 km
ranging from coast to interior. Power and correlation of this strip are shown at right. The
region comprises 8 ERS scenes from the rocky coastal area through the percolation zone
into the dry snow zone. Our area of interest in this paper is the inset square in the dry snow
zone (Hoen, 2001).

the model results very little as will be discussed in 4.5. The DEM data available from
58 CHAPTER 4. MODELING INSAR ECHOES FROM ICE SHEETS

NSIDC has about a 3 km resolution in this area and does not show minor surface undula-
tions due to wind. Hence, wind effects on accumulation are likely negligible.
Using the data shown in Figures 4.5(a) and 4.5(b), we derive maps of hoar layer thick-
ness and grain size, Figures 4.7 and 4.8 respectively, using our improved model (Equations
4.17-4.19). For each (ρ̃ , σ 0 ) data pair our model predicts a unique (h, R0 ) according to Fig-
ure 4.3, which we display in the new maps. We expect the estimated hoar spacing to be
proportional to accumulation rate and that the estimated grain radius values vary inversely
proportionally to accumulation rate. The correlation between these two maps is −0.66, so
that each behaves as we would expect, on a gross scale, at least.
For the four reference in situ stations in the area of our study shown in Figure 2.1, we
can now compare the layer thickness as a proxy for accumulation rate with the in situ ob-
servations in Figure 4.9. For completeness, we include the accumulation estimates derived
from the volume backscatter-radar brightness method also. The best indication of the use-
fulness of our new approach is shown in this figure: the new model estimates much more
closely follow the observations. The in situ data were collected a long time ago, in 1912-
1913, but unfortunately no more recent data for the area of our study are available. They are
still the best in situ data for this region (Munk et al., 2003; Ohmura and Reeh, 1991). There
are some limitations in these estimations but these are the only field data available for the
area. Future work should concentrate in areas where more recent data with traceable error
characteristics exists.

-6 079
-7 0 965

78 -8 423 0 96

0 955
100 km
σ (
100

-9
0 95

0 945
- 10

0 94
- 11

- 12
100 56 100 km
(a) (b)

Figure 4.5: (a) σ 0 and (b) Correlation map in dry snow zone
4.4. RESULTS 59

0.8

0.6
ρ

0.4

0.2

0
−20 −15 −10 −5 0
σ0(dB)

Figure 4.6: Correlation vs. normalized power backscattered for the detail area of interest.

4.4.2 Normalized backscattered power vs. incidence angle

As we discussed in Section 2.2, the volume scatter models are inadequate to fully explain
echoes from the ice sheets. In Figure 4.10, we show the angular variation in backscattering
from ERS scatterometer near the summit in Greenland (73.25 N, 37.28 W), plotted as +
signs (Ashcraft and Long, 2006). It is apparent that a volume-scattering-only model, shown
here by dashed line, does not match the angular dependence of echoes from ice in the dry
snow zone. In this figure, we have also shown our model prediction that backscattered
power falls off with incidence angle. As we can see our model rate of decrease of is a very
good fit to the scatterometer data. Clearly the dry snow does not scatter solely like a diffuse
volume in the ERS range of incidence angle at C-band. The variation of our ERS SAR data
with respect to incidence angle in Chapter 5 also show the same form of variation as ERS
scatterometer data.
60 CHAPTER 4. MODELING INSAR ECHOES FROM ICE SHEETS

30

25

h(cm)
100 km

20

15

10

100 km

Figure 4.7: Hoar layer spacing map in dry snow zone. Four ◦s are in situ stations.

Although volume scattering is the dominant factor for backscattering in the dry snow
zone, surface backscattering from the hoar layer interfaces in the snow, in addition to vol-
ume backscattering from the snow packs and grains, is needed as in our model to better
explain these observations. Reflections at the internal surfaces plus the volume scattering
term add up to the observed angular variation. It is clear that volume scattering only cannot
explain the total backscattered power.

The model presented here uses a very simple model for surface scatter, yet it can explain
fairly well the ERS1-2 InSAR observations. In Chapter 5, we show in greater detail how
surface scattering changes both backscattered power and correlation in more details. It will
enable us to apply our model to a wider look angles.
4.4. RESULTS 61

2.2

1.8
100 km

1.6

R (mm)
0
1.4

1.2

0.8
100 km

Figure 4.8: R0 map in dry snow zone

4.4.3 Penetration depth in the dry snow zone

In Equation 3.4, the effects of propagation loss are contained in the exponential term. The
extinction coefficient κe is equivalent to 1/d penetration (Hoen, 2001), where d penetration is the
penetration depth at which the one-way power falls to 1/e. For estimation of penetration
depth, we must also take into account the off-vertical factor cos θr . For ERS observations
over the dry snow zone of Greenland, this factor is about 0.96. Therefore the measured
penetration depth is almost the same as for vertical incidence. At longer wavelengths (2m),
airborne radar systems are capable of sounding the bedrock topography of Greenland ice
sheets (Gogineni et al., 1998) with a round trip distance in the ice of 4-6 km. At C-band
however, Mätzler (1987) reports a penetration depth of 25 m in the dry alpine snow and
Rott et al. (1993) found a C-band penetration depth of 22 meters in Antarctic firn. Hoen
(2001) shows that the penetration depth of the Greenland ice sheet ranges from about 10 to
30 m.
We showed in Section 1.3.2 that the density and grain size and consequently scatter-
ing properties change through the depth. Hence, there is no constant κe that can be used
62 CHAPTER 4. MODELING INSAR ECHOES FROM ICE SHEETS

40 In situ
Inferred,
Accumulation 30
volume
rate
scatter only
(cm-we/year) 20
Hoar layer
10 spacing from
InSAR (cm)
0
site site site site
a b c d

Figure 4.9: Comparison of accumulation from in situ Data and volume backscatter model
and Surface + backscatter model. Volume backscattered model only opposes completely
the in situ accumulation data while hoar layer spacing follow it closely. Since hoar layer
spacing is proportional to accumulation rate, our improved model at scattering in dry snow
using hoar layers explains the accumulation rate more accurately.

accurately to calculate penetration depth. In Figure 4.11, therefore, we use the κe profile
as a function of depth to calculate penetration depth. The penetration depth predicted by
our model and shown in Figure 4.11 varies between 12 to 32 m for grain radius variation
from 0.8 to 2.2 mm. Because by definition most of the backscattering from the snow occurs
above the penetration depth, the firn scattering properties below the penetration depth do
not affect our results, as we will see below.

4.5 Discussion

We have presented a model that retrieves snow accumulation parameters more accurately
than existing remote sensing methods over a limited region in Greenland. To apply our
method to the entire ice sheet, we in addition have to consider the affects of variations in
many of the parameters we assumed in this reduction.
4.5. DISCUSSION 63

−10

−11
σ0(dB)

−12

−13
20 22 24 26
θ
i

Figure 4.10: Angular variation in backscattering from the ERS scatterometer near the sum-
mit in Greenland, shown by plus sign (Ashcraft and Long, 2006). This angular variation
of backscattered power in the dry snow zone cannot be explained by a volume-scatter-only
model (dashed line). However, it can be modeled by a combination of rough surface scat-
tering and volume scattering effects (dashed-dot line). Therefore, surface scattering is also
an important factor in the dry snow zone in the C-band for low incidence angles.

Volume scattering rather than surface scattering is the dominant factor for σ 0 within the
dry snow zone. Therefore σsur
0
f ace from these layers will be small compared to the σvolume ,
0

except at very low incidence angles, where the effect of surface scattering is noticeably
higher (Drinkwater et al., 2001; Ashcraft and Long, 2006). So as long as surface scattering
is a second-order effect, the transmission coefficients of the layers are more important than
surface backscattering from those layers. Most of the observed backscattered power is
from volume scattering, especially from below the hoar layer depth. Therefore, as the
number of interfaces in the layered part of the snow increases, more power is scattered
away by the layers, and less power is transmitted to the snow below resulting in less power
backscattered. To a first approximation, lower accumulation rates correspond to larger grain
64 CHAPTER 4. MODELING INSAR ECHOES FROM ICE SHEETS

32

28
Penetration Depth (m)

24

20

16

12
0.8 1 1.2 1.4 1.6 1.8 2 2.2
Grain Radius (mm)

Figure 4.11: Our model prediction of penetration depth for grain radius variation between
0.8 and 2.2 mm.

sizes and so to brighter areas. However, lower accumulation rates imply to more layers in
the first 6 meters of snow and hence darker areas. In areas with similar accumulation rates
and consequently similar grain size, layering is required to model the observed differences
in accumulation rate, whereas volume scattering only models cannot.
The permittivity of the firn above and below each layer is nearly unchanged with slight
variation in the position of the layer. Hence, the transmissivity of the layer changes very
little, and has almost no effect on the total backscattering. Therefore the number of layers
is a much more important parameter than the position of layers in the top 6m in terms of
its effect on radar backscatter.
In Equation 4.18, σv (n) appears in both the numerator and denominator, and so the
3κs d
effects of 2 + σh0 and γm cancel out to some extent and are therefore less important than
the exponential loss. This point can be seen more clearly if κs , κa , εr , . . . are kept constant
through depth. Surface scattering has no effect on the exponential loss and hence, the
correlation depends much more strongly on the grain size rather than on h in Figure 4.3.
Rough surface characteristics such as l and σl also have little effect on the final result
4.5. DISCUSSION 65

because volume backscattering dominates. Finally, the proportionality factor relating hoar
layer spacing in our model to accumulation rate should be investigated in more detail. For
now, we assume we can multiply spacing by firn density to obtain the accumulation rate.
Next, we describe the effect of density variations on our model.

4.5.1 Effect of density variations profile in our model


In our model we use the data depicted in Figure 1.3 for the variation of density through the
depth. However, variation of mean annual temperature and accumulation rates can change
this behavior (Herron and Langway, 1980).

0.8

0.6
ρ

0.4

0.2 Acc. rate = 20 cm−w.e.


Acc. rate = 40 cm−w.e.
Acc. rate = 60 cm−w.e.
0
−20 −15 −10 −5 0
σ0

Figure 4.12: Effect of accumulation rate on our model’s prediction. As shown in Figure
1.4.b, density vs. depth profile changes with accumulation rate. Because the change in
density profile is below the penetration depth for C-band, it has almost no effect in our
modeling.

Using the different density profile curves shown in Figure 1.4.b, we plot our model
estimates of (σ 0 ,ρ ) in Figure 4.12. In this figure, hoar layer spacing is 20 cm and grain
radius varies between 0.8 mm to 2.3 mm. As we can see in this figure, the variation of
66 CHAPTER 4. MODELING INSAR ECHOES FROM ICE SHEETS

density due to the accumulation rate variation changes our model results very little. This
is because over the depths that radar signals penetrate, the density profile does not change
appreciably as a function of the accumulation rate. Accumulation rate primarily alters the
density of firn below the observed penetration depth, as seen in Figure 1.4.b.

0.8

0.6
ρ

0.4 T = −15
T = −30

0.2 T = −50

0
−20 −15 −10 −5 0
0
σ

Figure 4.13: Effect of temperature on our model’s prediction. As shown in Figure 1.4.a,
density vs. depth profile changes with temperature. Because the change in density profile
is in penetration depth for C-band, it changes our modeling results.

We can use the different density profile curves in Figure 1.4.a to study the effect of
temperature variations in our model. Green, black and red dotted curves in Figure 4.13
correspond to our model prediction of (σ 0 ,ρ ) for average air temperature, T , equal to -15,
-30, -50. Hoar layer spacing is again 20 cm and grain radii change between 0.8 mm to 2.3
mm. Variations of one degree in temperature lead to changes in σ 0 and ρ by about 1.16%
and 1% respectively. This sensitivity to temperature will not affect the data presented here,
as the study area is small and temperature variations will also be small. While an almost
constant temperature is appropriate for this area, to apply our model to the entirety of
Greenland we would include temperature variation with location.
4.6. CONCLUSIONS 67

4.6 Conclusions

Our model adds scattering from hoar layers to volume scatter to accommodate a wider
range of physical conditions in the ice sheet. As shown in Figure 4.3, the new model cov-
ers the range of observed dry snow data in Figure 4.6 through variation of hoar spacing
and grain radius over the range of typical values in the dry snow zone. We have shown
that the volume-only model cannot adequately explain variation in measurements of ρ̃ and
σ 0 . Adding radar InSAR correlation to the backscatter measurements supports a more com-
plete model, with much better agreement with field observations. There are of course many
reasons that data may disagree with prediction. Uncertainties in the data may be underes-
timated, systematic effects may have been neglected, or the model may be incomplete, for
example. We found that adding a scattering contribution from layer interfaces resolves the
discrepancy in this instance. Our InSAR scattering model predicts that in areas of simi-
lar accumulation rate, and thus similar grain size, layering can better explain the observed
differences in accumulation than those derived from the volume only model. While apply-
ing this algorithm to the entirety of Greenland remains future work, in a global sense our
model should give similar continent-wide trends to those published in the past. The model
is consistent with the general trend in the dry snow zone that the darker areas correspond
to smaller grain sizes and higher accumulation rates.
Correlation data are derived independently from backscatter observations, and having
access to both data types spreads the backscattered power/correlation pairs to form a two
dimensional map as shown in Figure 4.6. As seen in Figure 4.3, greater grain radius and
greater h both increase backscattered power. The former corresponds to less accumulation
while the latter corresponds to higher accumulation, so that backscattered power alone is
not reliable in estimating accumulation. While, to a first approximation, it is correct that
brighter areas correspond to less accumulation, this is not true in detail.
This is not the first time that hoar layer effects have been invoked to explain microwave
observations at the ice sheets. Abdalati and Steffen (1998) also show that accumulation
estimates based on passive microwave observations require successful parametrization of
the hoar formation characteristics. West (1994) and Winebrenner et al. (2001) also use
layering in the dry snow zone to explain passive microwave data. Here we have shown
68 CHAPTER 4. MODELING INSAR ECHOES FROM ICE SHEETS

that accumulation estimates based on active radar data also require hoar layers to explain
variations in the dry snow zone.
In this Chapter, we have used InSAR data to obtain a high resolution, accurate map
of hoar layer spacing, which is proportional to accumulation in a small part of Green-
land. These data significantly improve contours derived from sparse in situ stations, and
benefit from spatially-contiguous InSAR remote sensing data. We hope that by applying
this model to the entire dry snow zone, we will be able to test our model predictions of hoar
layer spacing, and lead to estimates of accumulation rate with high resolution and high
accuracy.
Chapter 5

Surface Scattering

5.1 Introduction
In this Chapter, we present surface scattering in the dry snow zone of Greenland in more
detail. In Section 5.2, we review the inconsistencies in measured backscattered power and
correlation that occur if we ignore surface scattering. In Section 5.3 we define a parameter
b, the ratio of surface backscattered power to volume backscattered power, to quantify
the effect of surface scattering on correlation. We explain how to infer the parameter b
in Section 5.4 and we compare our estimate of surface scattering using b with the actual
surface scattering data. We then show in Section 5.5 how our estimate of surface scattering
helps us to better predict the accumulation rate in the dry snow zone of Greenland. We
summarize and conclude in Section 5.6.

5.2 Surface scattering


As seen in previous Chapters, both surface and volume scattering mechanisms are needed
to explain the observed characteristics of the total scattering response in the dry snow zone
of Greenland. To apply our model to a wider range of imaging geometries we find that we
need a better estimate of surface scattering as a function of look angle. Since volume scat-
tering does not change significantly over the limited range of incidence angles (Figure 4.10)
available in ERS data, we assumed that most observed variation in the total backscattered

69
70 CHAPTER 5. SURFACE SCATTERING

power is due to the surface scattering component. To apply our model to data acquired over
a wider range of imaging geometry, we will extend the model of Chapter 4 as described
below. Of course, the more accurately we can measure surface and volume scattering, the
more accurately can model accumulation rate.

Figure 5.1: Map of 14 different frames of ERS1-2 tandem pairs in the dry snow zone of
Greenland used in our study. Acquisition time of the data set is January to February of
1996. Small boxes denote one ERS scene, approximately 100 × 100 km.

In this Chapter we apply our model to a larger area in the dry snow zone of Greenland,
again using ERS1-2 tandem pair data from ESA, the European Space Agency (Figure 5.1).
We chose fourteen different frames to cover as many of the available in situ base stations
as possible. The acquisition time of these data set range from January to February of 1996.
We also use RADARSAT-1 data, ordered from the Alaska Satellite Facility, to study the
variation of normalized backscattered power with incidence angle. The correlation obtained
5.2. SURFACE SCATTERING 71

from the RADARSAT-1 data, however, was not high enough for accurate accumulation rate
retrieval in our modeling as we were not able to separate temporal decorrelation effects
from the volume scattering effects. The repeat cycle of RADARSAT-1 is 24 days, and leads
to relatively high temporal decorrelation. Rott and Siegel (1996) have shown that, although
the magnitude of the backscattering coefficient is stable in time in the dry snow zone, a
very pronounced decrease of coherence with time is observed in this area. As previously
described, the main reason for the temporal variability is change in surface conditions.
Snow precipitation and wind in the dry snow zone in the winter can lead to significant
decorrelation.

−6
σ0(dB)

−8

−10

−12
20.77 23.18 25.32 27.25
θi

Figure 5.2: Azimuth averaged calibrated normalized backscattered power vs. incidence
angle for different frames of ERS1-2 tandem pairs. Normalized intensity roughly changes
1-3 dB with incidence angle across the swath.

First, we examine the effect of the surface scattering component on σ 0 and ρ as seen
in the new data sets. In Figure 5.2, we plot the averaged calibrated normalized backscatter
power as a function of incidence angle for several different ERS1-2 tandem pairs. Different
colors correspond to different frames in this figure. The variation of the backscattered
72 CHAPTER 5. SURFACE SCATTERING

power with incidence angle, is ∼ 2dB across an ERS swath. As we will show in Figure
5.8.a and previously showed in Figure 4.10, the volume component of the backscattered
power changes very little over this incidence angle range. Therefore, observed variation
in the backscattered power with respect to incidence angle in Figure 5.2 is dominated by
surface scattering variations (also shown in Figure 5.8.a).
For a set of overall larger incidence angles such as for RADARSAT-1, the observed
variation is less, as can be seen in Figure 5.3.a. This is expected from most models of
surface and volume scatter. In this figure we plot the average normalized backscattered
power vs. incidence angle for a frame of RADARSAT-1 data. Normalized backscattered
power changes less than 0.5 dB with incidence angle across the RADARSAT swath. Our
model prediction of the volume and surface scattering component of the backscattered
power (Section 4.3), are depicted by a red dashed line in Figure 5.3.a. and black solid
line in Figure 5.3.b, respectively. The surface scattering component in Figure 5.3.b is very
small compared to the volume scattering component. Hence, its variation with respect to
incidence angle does not change the total backscattered power variation with increasing the
incidence angle. Therefore, the surface scattering is negligible for RADARSAT data but

−12.5 −46

−13
σsurface(dB)

−50
σ (dB)

−13.5
0

−14
−54

−14.5
37 38 39 40 37 38 39 40
θ θ
i i
(a) (b)
Figure 5.3: (a) Averaged calibrated backscattered power vs. incidence angle for a frame
of RADARSAT-1 data in the dry snow zone of Greenland (black line). Red dashed-line
is the estimated volume scattering component of backscattered power vs. angle, using our
model in Section 4.3. (b) Estimated surface scattering component of backscattered power
vs. incidence angle, using our model in Section 4.3.
5.2. SURFACE SCATTERING 73

cannot be neglected for ERS data.

As we discussed in previous Chapters and observed above, surface scattering has an im-
portant impact on backscattering in the dry snow zone. To better understand and quantify
the effect of surface scattering on correlation, we first model the case with no surface scat-
tering and compare the predicted correlation with what we observe in our data. We assume
a uniform medium with constant permittivity, grain size, and other non-varying parameters
and include no surface scattering.

From Equation 3.20, we have


R0 − j2πκzz z dz
−∞ σv (z)e
ρvolume = | R0 | (5.1)
−∞ σv (z)dz

We can see that the volume correlation is the ratio of the Fourier component of σv (z)
at the frequency κzz to its DC component. Assuming as before an exponential loss for
backscattered power,
2zκe
σv (z) = σv0 e cos θr (5.2)

and the simple case of constant density and grain size through depth, volume correlation is
2κe
proportional to the Fourier transform of eaz u(−z), where a = cos θr , thus

σv0
a− j2π f 1
ρvolume = | σ 0 | f =κ = 2√εδ θr =| | √
j2π f f =κzz = 2 εδ θr
(5.3)
a
v zz λ sin θr 1− a
λ sin θr

In the equations above, the volume-backscatter coefficient σv0 does not explicitly appear
in the correlation, but it is proportional to scattering coefficient, κs (see Equation 4.6).
From Equation 4.5 we know that κs ≈ 0 for small grain sizes and therefore the extinction
coefficient, κe = κs + κa , is dominated by absorption coefficient, κa , rather than κs . Hence
the effect of grain size on a and consequently on correlation is insignificant for small grain
sizes. Therefore for small grain radii, the correlation in Figure 4.1 changes for changes in
grain radius.
74 CHAPTER 5. SURFACE SCATTERING

After applying Snell’s law and letting δ θ = B⊥


R0 , Equation 5.3 simplifies to

1
ρvolume = q √ (5.4)
2π ε B⊥ 2
1 + ( R λ tan θ κe )
0

0.6

0.55

0.5
ρvol

0.45

0.4

0.35
18 20 22 24 26 28
θi

Figure 5.4: Volume correlation vs. angle for ERS1-2 data, using Mätzler’s volume scatter-
ing model. Note that correlation increases with angle.

Using ERS parameters and Mätzler’s volume scattering model (Section 4.2.1) in Equa-
tion 5.4, we can plot volume correlation variation as a function of angle to yield Figure
5.4. In this figure, note that correlation increases with incidence angle. In Figure 5.5 we
show the averaged volume correlation vs. incidence angle for the fourteen different scenes
in Figure 5.1 and see that the variation of correlation with respect to incidence angle in
general, is not similar to Figure 5.4. Recall that the effects of thermal and surface corre-
lation are removed in this correlation calculation, as explained in Sections 3.3.5 and 3.3.2,
respectively. Thus volume correlation dominates the figure. Different colors in this figure
correspond to different scenes and are the same as the colors in Figure 5.2. Although the
5.2. SURFACE SCATTERING 75

0.75

0.7

0.65

0.6
ρ

0.55

0.5

0.45
20.77 23.18 25.32 27.25
θi

Figure 5.5: Averaged volume correlation as a function of incidence angle after correcting
for the effect of thermal and surface correlation. The observed correlation is much higher
than what is expected from only volume-scattering at low incidence angles.

volume correlation increases with range for some scenes, it decreases for most. For ex-
ample, the variation of the volume correlation vs. incidence angle is similar to variation
shown in Figure 5.4 for the cyan and blue dashed lines, while it is completely opposite for
the green dashed and yellow dashed lines. For the other scenes in Figure 5.5, correlation at
low incidence angles is much higher than expected if we consider volume scattering only.
This implies that the volume-only model applies for some surface conditions, but not for
most. Therefore, the volume-only scattering model is insufficient to explain the high value
of correlation at low incidence angles as well as the variation of correlation with incidence
angle.

In the next section, we show that by including the effects of surface scattering to the
volume scattering equations shown here we can better reproduce the observed σ 0 and ρ
variation with incidence angle. This gives insight into the structure of the ice.
76 CHAPTER 5. SURFACE SCATTERING

5.3 Effect of surface scattering on correlation


As seen in the previous section, the variation of correlation with incidence angle of ERS
data cannot be explained using volume scattering only. As in Chapter 4, here we examine
the effect of adding a surface scattering term to the correlation calculation. We first add a
single layer at depth z0 to the uniform volume of snow, as shown in Figure 5.6. Under these
conditions, Equation 5.2 is rewritten as

z
0

Figure 5.6: Geometry of a layer at depth z0 .

2zκe
σv (z) = [σv0 + σh0 δ (z − z0 )]e cos θr (5.5)

and therefore,
σv0 0 −az0 e j2π f z0
a− j2π f + σh e
ρt = |
σv0
| f =κ = 2√εδ θr (5.6)
a + σh e
0 −az0 zz λ sin θr

1 1
ρt = | j2π f
+ be j2π f z0 | × | | (5.7)
1− 1+b
a
where we define a parameter b as the ratio of surface backscattered power to volume
backscattered power:
R0
σh0 e−az0 −∞ σh e δ (z − z0 )dz
0 az σsur
0
f ace
b= = R0 = (5.8)
σv0
−∞ σv e dz
0 az σvolume
0
a
5.3. EFFECT OF SURFACE SCATTERING ON CORRELATION 77

Looking back to Equation 5.3, we note that the first ratio in the first magnitude term in
1
Equation 5.7, , is the same as the correlation term in Equation 5.3. This is the
1− j2aπ f
correlation in the case where no layering exists. We call this term ρvolume−only and rewrite
Equation 5.7 as
1
ρt = | | × |ρvolume−only + be j2π f z0 | (5.9)
1+b
As can be seen in this equation, both surface and volume backscattered power appear in
the equation for b and consequently in the correlation equation as well. This implies that
even for small grain sizes where σv0 ≈ 0 (Equations 4.5, 4.6), σv0 affects the total correla-
tion by changing b. As explained in the previous section, in contrast for the volume-only
scattering case, there is no correlation dependence on the backscattering coefficient and
therefore on grain size. This might be seen to imply that ρvolume−only does not change
much with grain radius. In the more complete development model shown here, however,
volume backscattered power appears in the denominator of Equation 5.8, in the form of pa-
rameter b. Therefore, even for small grain sizes, correlation depends on grain radius when
the more complete model is used.
We examine the effect of the parameter b on correlation and backscattered power in
Sections 5.3.1 and 5.3.2, respectively.

5.3.1 Effect of ratio of surface to volume scatter, b, on correlation


Here we show how the parameter b in Equation 5.9 changes the correlation. Examining
the different values of ε from 1.3 (ρ f = 0.2 cmg 3 ) to 3.15 (pure ice), f varies between 0.026
and 0.038. Also, from Section 4.3 we know that z0 can be any value up to 6 meters depth.
Considering the correct magnitude and phase for the terms in Equation 5.9, we plot the
total correlation in Figure 5.7 as a function of b for different values of ρvolume−only . For b =
0, the total correlation equals ρvolume−only . The total correlation increases with b after the
minimum in the curve.
As discussed in Section 5.2, the surface scattering mechanism in the dry snow zone of
Greenland is not negligible at the ERS incidence angles. Correlation can be higher than
expected from volume scattering only, with the precise amount dependent on the ratio of
surface to volume scattering, b. This effect can thus explain the high values of correlation
78 CHAPTER 5. SURFACE SCATTERING

ρ = 0.9
0.8 Vol−Only

0.6
ρt

0.4

ρVol−Only = 0.1
0.2

0
0 2 4 6 8 10
b

Figure 5.7: Total correlation ρt vs. the coefficient b. Different curves in this figure corre-
spond to different values of ρvolume−only in steps of 0.1. The total correlation increases with
b after the minimum in the curve.

shown in Figure 5.5. On the other hand, as we discussed in Section 5.2, the shapes of green
and yellow dashed lines in Figure 5.5 differ markedly from the correlation predicted by
volume-only scattering (Figure 5.4). Examining the backscattered power of the two scenes
in Figure 5.2, we note an approximate 4 dB variation in σ 0 from near to far range. This
variation with incidence angle is probably caused by surface scattering (Rott and Rack,
1995; Long and Drinkwater, 1994; Ashcraft and Long, 2006). Below, we show that an
increase of surface scattering compared to volume scattering, i.e. high b, can better explain
the shape of green and yellow dashed lines correlation with respect to incidence angle.
In Figure 5.8.a, the black solid line shows the backscattered power as a function of
incidence angle using our model in Section 4.3. The corresponding grain radius and hoar
layer spacing in this figure are 0.8 mm and 20 cm, respectively. As seen in this figure,
the normalized backscattered power changes from about -7.5 dB to about -12 dB. The
surface and volume scattering component of our model prediction of backscattered power,
5.3. EFFECT OF SURFACE SCATTERING ON CORRELATION 79

is also shown in Figure 5.8.a by blue and red dashed lines. Note that the volume scattering
component changes very little along the incidence angle while surface scattering drops
about 10 dB from near to far range. We use these surface and volume scattering variations
σsur
0
f ace
in part (a) of Figure 5.8 to calculate b = σvolume
0 . Figure 5.8.b shows the inferred values
of b using the model predictions in part (a). Because the volume scattering component
of backscattered power is roughly invariant with incidence angle, we can expect that b is
proportional to the surface scattering component, thus it is larger at lower incidence angles
than at larger incidence angles.
Figure 5.9 shows the predicted effect of b on the volume-only correlation shown in
Figure 5.4. We substitude the ρvolume−only (Figure 5.4) and b (Figure 5.8.b) in Equation
5.9, to calculate the total correlation as a function of incidence angle. The total expected
correlation as a function of incidence angle is shown in Figure 5.9. Note that unlike the
volume-only correlation shown in Figure 5.4, the total correlation now decreases as inci-
dence angle increases. This is because of very high values of b at low incidence angles as
compared to its value at high incidence angles. Therefore, parameter b not only increases
the volume-only correlation but also can change the shape of the correlation variation with

−6

−10 σ
0
σ (dB)

0 2
σ
b

volume
0

−14

1
0
σsurface
−18

0
18 20 22 24 26 28 18 20 22 24 26 28
θ θ
i i
(a) (b)
Figure 5.8: (a) Estimated backscattered power (black solid-line) as a function of incidence
angle using our model in Section 4.3 for grain radius R0 equal to 0.8 mm and h = 20 cm.
Blue and red dashed line in this figure correspond to the surface and the volume component
σsur
0
f ace
of backscattered power vs. incidence angle, respectively. (b) Calculated b = σvolume
0 using
the surface and volume backscattered power in part (a).
80 CHAPTER 5. SURFACE SCATTERING

0.9

0.8
ρt

0.7

0.6
18 20 22 24 26 28
θi

Figure 5.9: Total correlation vs. incidence angle using Equation 5.9. The volume only
correlation, ρvolume , and b used in this equation are shown in Figures 5.4 and 5.8.b, respec-
tively.

respect to incidence angle.


The backscattered power shown by green and yellow dashed lines in Figure 5.5, varies
significantly with respect to incidence angle. Knowing that the volume component of the
backscattered power changes very little (Figure 5.8.a), we conclued that surface scattering
is very large compared to volume scattering (large b) at low incidence angles. Hence, b
values for near-swath incidence angles in these scenes are high because of the relatively
large values of surface scattering. This causes the correlation at low incidence angles to
be much higher than what is expected from volume-only scatter. Therefore the correlation
variation with incidence angle for these two scenes does not match the volume-only corre-
lation in Figure 5.4. Instead, it is similar to correlation variation in Figure 5.9. Hence, we
can better reproduce the data by adding surface scattering to volume scattering to describe
the shape of the correlation curves in Figure 5.5 by considering b, which has high values at
low incidence angles.
5.3. EFFECT OF SURFACE SCATTERING ON CORRELATION 81

5.3.2 Effect of b on backscattered power

Now we are ready to show how the b values help us estimate the surface scattering com-
ponent in the dry snow zone of Greenland. We divide the total normalized backscattered
power into surface and volume backscattering term, and obtain

σ 0 (θi ) = σvolume
0
+ σsur
0
f ace (θi ) = σvolume (1 + b(θi ))
0
(5.10)

For incidence angles larger than some θc , volume scattering dominates the total backscatter,
so that b(θc) ≈ 0 and σ 0 (θc ) ≈ σvolume . Volume backscattered power changes slowly with
incidence angle and

σ 0 (θi ) σvolume
0 (1 + b(θi))
≈ = 1 + b(θi ) (5.11)
σ ( θc )
0 σvolume
0

and
σ 0 ( θi )
10 log( ) = σdB
0
(θi ) − σdB
0
(θc ) = 10 log(1 + b(θi )) (5.12)
σ 0 ( θc )
We define σs,dB = σdB
0 (θ ) − σ 0 (θ ) as the difference between total backscattered power
i dB c
and volume backscattered power in dB. The value of σs,dB represents the effect of surface
scattering on the total backscattered power in dB. If the changes in total backscattered
power with incidence angle are modeled as linear in dB then,

∆dB
σs,dB = σdB
0
(θi ) − σdB
0
( θc ) = ( θi − θc ) (5.13)
∆θ
∆dB
In this case, the slope of the backscattered power ∆θ , is sufficient to describe σs,dB . We
will use this proportionality to compare our model estimate of σs,dB with the slope of the
∆dB
measured backscattered power, ∆θ in Section 5.4.3.

We next will show in Section 5.4.1, how we estimate b from InSAR observations. We
will calculate σs,dB using our estimated b values and Equation 5.12. Then we compare our
predicted σs,dB with the actual surface scattering slope in Section 5.4.3.
82 CHAPTER 5. SURFACE SCATTERING

5.4 Results
As discussed in the previous section, the value of b gives an estimate of the importance
of surface scattering for radar measurements of ice sheets. Knowing the surface scatter at
different look angles improves our previous model estimates to wider geometries. In this
section, we verify our model predictions of surface scattering using InSAR observations.

5.4.1 Estimating b and σs,dB


As discussed in Section 4.5, in areas with similar accumulation rates and consequently
similar grain size, models with layering be used to estimate the differences in accumulation
rate. Therefore, for each scene in the dry snow zone of Greenland, we first assume that the
grain radius changes very little in that specific scene because of the small area of the scenes
(100km × 100km).
As we will see later in this section, the average grain radius R0 of a scene determines
the value of ρvolume−only at different incidence angles in that scene. Consequently, we can
specify the b value of each pixel in the scene, by using the pixel’s correlation data, ρt ,
together with Figure 5.7. Knowing ρvolume−only , we specify the ρt versus b curve in Figure
5.7 that our data obey. The correlation of the pixel, ρt , determines the corresponding b
along that curve. For instance, if ρvolume−only = 0.9 for all the pixels in a specific incidence
angle inside a scene, then the top curve in Figure 5.7 is the proper ρt vs. b curve describing
those pixels. Then, for a pixel correlation of ρt = 0.73, for instance, b equals 2.
We next specify the grain radius of a scene in the dry snow zone. Then we will show
how to find b for each pixel in that scene.
In order to determine the grain radius of a scene, we use our framework in Section 4.3
to model the variation of the average backscattered power as a function of the incidence
angle within a scene using. Figure 5.10 shows ρvolume−only as a function of σ 0 , where
ρvolume−only is calculated using Mätzler’s model as described in Section 4.2.1, that is as-
suming no layering in the snow and thus σ 0 is calculated using our model in Section 4.3.
The reason we consider no layering for calculating the correlation in this plot is to study the
effect of surface scattering on the volume-only correlation as we explained in Section 5.3.
However, surface scattering from the layers is needed to increase the backscattered power
5.4. RESULTS 83

0.8
h = 10
h = 20
0.6 h = 30
ρ

0.4

0.2 θi = 25.4 θ = 19.4


i

0
−14 −12 −10 −8 −6 −4 −2
0
σ (dB)

Figure 5.10: ρ vs. σ 0 (dB) using our model in Section 4.3 and Mätzler’s model in Section
4.2.1, for calculating σ 0 and ρ respectively. The first three curves at the left correspond
to θi = 25.4/far-range and the three curves at the right correspond to θi = 19.4/near-range.
The blue, black and red dotted curves correspond to h equal to 10, 20 and 30, respectively.
Grain radii also change from 0.8 to 2.2 mm in 0.1 mm steps along each curve as depicted
by dots.

with decreasing the incidence angle. Therefore, we have shown Mätzler’s volume-only
correlation and our model estimate of backscattered power in this figure.
The first three curves at the left in Figure 5.10 correspond to θi = 25.4, and the three
curves at the right correspond to θi = 19.4. The blue, black and red dotted curves corre-
spond to h equal to 10, 20 and 30, respectively. As we can see in this figure, for different
(h, R0) pairs, shown by dots, we predict different values of σ 0 in both far and near range.
Note that ρvolume−only in Figure 5.10 increases from near to far incidence angles, as we have
also seen in Figure 5.4. Looking back to Figure 5.2, we can see that the average backscat-
tered power in different scenes changes from -5 to -8.5 in near-range and from -6 to -11 in
far-range. For the average of σ 0 at near and far incidence angles in each scene, our model
84 CHAPTER 5. SURFACE SCATTERING

(Figure 5.10) predicts a single pair (h, R0 ). As we discussed earlier in this section, we as-
sume that the grain radius of a scene, R0 , changes very little for the pixels in that specific
scene. We use this R0 to find the volume-only component of the correlation at different
incidence angles inside a scene by using Mätzler’s model (Section 4.2.1).
Note that the constant grain size assumption in a scene fails, if there is a very abrupt
change in the accumulation rates. An abrupt variation in the accumulation rate inside a
scene changes the grain size, and so the constant average grain size for the pixels in that
scene is invalid. Any shortcoming in this assumption can lead to inaccuracy in our estima-
tion of surface scattering.

0.8
θi = 25.4

0.6 θi = 19.4
ρ

0.4

0.2

0
−14 −12 −10 −8 −6 −4 −2
σ0(dB)

Figure 5.11: ρ vs. σ 0 (dB) for different final grain radii, using our model in Section 4.3 and
Mätzler’s model in Section 4.2.1, for calculating σ 0 and ρ respectively. The two dotted
curves correspond to different value of θi : 19.4, 25.4. Grain radii also change from 0.8 to
2.2 mm in 0.1 mm steps along each curve as depicted by dots. Hoar layer spacing, h, is
equal to 20 cm for both curves.

For instance, for the purple line in Figure 5.2, the average backscattered power changes
5.4. RESULTS 85

from about -5.5 dB at low incidence angle to about -7.5 dB at high incidence angle. As we
explained earlier in this section, our model in Section 4.3 predicts a unique (h, R0 ) for the
average of backscattered power observed in this scene at high and low incidence angle. In
this case the model prediction pair is (h = 20cm, R0 = 1.6mm). We depicted ρ vs. σ 0 (dB)
for different grain radii in Figure 5.11. The two dotted curves correspond to different value
of θi : 19.4, 25.4. Grain radii also change from 0.8 to 2.2 mm in 0.1 mm steps along each
curve as depicted by dots. Hoar layer spacing, h, is equal to 20 cm for both curves. As we
can see in this figure, for R0 = 1.6mm, σ 0 equals to -5.5 dB and -7.5 dB at near and far
incidence angles (19.4 and 25.4), respectively. Using R0 = 1.6mm, we can determine the
ρvolume−only for the ERS incidence angle interval (Section 4.2.1). Thus, for each pixel on
this scene, we have the total correlation, ρt , of that pixel from the data and ρvolume−only at
the pixel’s incidence angle. We use these values to find b from Figure 5.7, as we explained
earlier in this section.
Part (b) in Figures 5.19, 5.12, 5.14, 5.13 and 5.15 are σs,dB = 10 log(1 + b(θi )) inferred
from b. We will discuss this in more detail later in Section 5.4.3.

23774-2151 23774-2151
0.7 3.5

0.6
3
;:
∆θ

<
σs dB

0.5
∆d

2.5
0.4

0.3 2
20 22 24 20 22 24
θi θi
( => (b)

Figure 5.12: (a) Slope of backscattered power vs. angle using two backscattered power
scenes covering each in situ base station located at (71.3N,34.55W), (71.17N,35.85W),
(71.2N,35.95W) and (71.5N,35.88W). (b) Surface scattering inferred from the data using
the model.
86 CHAPTER 5. SURFACE SCATTERING

23832-2043 23832-2043
4.5

0.4
4
@?
∆θ

σs dB
∆d

0.3
3.5

0.2 3
20 22 24 26 20 22 24 26
θ θ
i i
(BC (b)

Figure 5.13: (a) Slope of backscattered power vs. angle using two backscattered power
scenes covering each in situ base station located at (76.32N,45.1W), (76.2N,45.77W),
(76.57N,45.32W), (76.63N,45.7W), (76.45N,46.52W), (76.73N,47.33W) and
(76.97N,46.98W). (b) Surface scattering inferred from the data using the model.

∆dB
5.4.2 Observed ∆θ

Now we compare our model predictions with the actual surface scattering data. In order
to find the actual surface scattering values, we used additional new datasets to cover each
base station with two different incidence angles. ERS data that cover the in situ stations
with two incidence angles were not available for all 14 scenes in Figure 5.1. ERS data with
different look angles were available only for four of these frames. Figure 5.17 shows the
map of these four scenes. The five digit number in this Figure is the ERS1 orbit number
and the four digit number is the frame number of the ERS data. We use the measured
∆dB
backscattered at the two incidence angles to obtain ∆θ .
Figure 5.16 shows the geometry of a base station covered with two different ERS
scenes. The main scene is in blue with two in situ stations shown by the circles. The
location of the main scene is depicted in Figure 5.17. As we can see, the right most station
in the main scene is located in near swath range and its incidence angle is lower (for the
descending right-looking ERS antenna), while it is in the far range and its incidence angle
is high for the scene on the right. The left most station in the main scene is in far range and
5.4. RESULTS 87

23774-2169 23774-2169

3
0 D6
GF
∆θ

σs dB
∆d

2
E
0D

1
22 24 26 22 24 26
θ θ
i i
(IJ (b )

Figure 5.14: (a) Slope of backscattered power vs. angle using two backscattered power
scenes covering each in situ base station located at (70.65N,35.83W), (70.87N,35.85W),
(70.63N,36.18W) and (70.67N,37.48W). (b) Surface scattering inferred from the data using
the model.

its incidence angle is high, and its incidence angle is low in the left hand scene. The ERS
backscattered powers of these two scenes give us the slope of the backscattered power,

∆dB σdB
0 (θ ) − σ 0 (θ )
i1 dB i2
= (5.14)
∆θ θi2 − θi1

where θi1 , θi2 are the two different incidence angles. Hence each in situ station on the four
ERS scenes in Figure 5.17, is covered with two different incidence angles. Using σ 0 at
∆dB ∆dB
these two incidence angles and Equation 5.14, we calculate ∆θ . The ∆θ of these base
stations are plotted in Figures 5.12.a, 5.14.a, 5.13.a and 5.15.a.

∆dB
5.4.3 Comparing σs,dB and ∆θ

As explained earlier in Section 5.4.1, for each base station we find b using our model and
the ERS observations. From the b value, we infer σs,dB using Equation 5.12. We show σs,dB
at the in situ station locations in Figures 5.12.b, 5.14.b, 5.13.b and 5.15.b. From Equation
∆dB ∆dB
5.13, σs,dB is proportional to ∆θ . On the other hand, we show ∆θ of those base stations,
88 CHAPTER 5. SURFACE SCATTERING

23846-2097 23846-2097
0.5 4.5

0.4
LK
∆θ

σs dB
4
∆d

0.3

0.2 3.5
18 20 22 24 26 18 20 22 24 26
θi θi
(NO (b)

Figure 5.15: (a) Slope of backscattered power vs. angle using two backscattered
power scenes covering each in situ base station located at (73.87,41.8), (74.38,41.42),
(74.32,41.98), (74.22,42.17), (74.2,42.53), (74.13,43.05), (74.03,43.52) and (73.98,44.27).
(b) Surface scattering inferred from the data using the model.

using σ 0 at two incidence angles.


∆dB
Hence for each in situ station in a scene, we show ∆θ in part (a) of the figure by a
specific color, and σs,dB in part (b) with the corresponding color. Note that while part (a) is
the actual surface scattering data, as expressed by the slope of σ 0 , and part (b) is the model
prediction.
∆dB
In Figure 5.12 we show ∆θ and σs,dB of four base stations as a function of incidence
∆dB
angle in scene 23774-2151. Assuming the proportionality of σs,dB and ∆θ from Equation
5.13, we can compare these values. We can see in this figure that the trend in variation in
part (a) and (b) are the same for these four base stations. In other words, the red station has
the highest surface scattering in both (a) and (b). The blue, green and purple stations have
successively lower surface scattering in both (a) and (b). Hence our prediction of the trend
of surface scattering seems correct for these four in situ stations.
In Figure 5.13, our model estimate of surface scattering is compared with the actual
data of surface scattering for seven different in situ stations in scene 23832-2043. Here
for the first five in situ stations, the general trend of the variation of surface scattering is
5.4. RESULTS 89

Figure 5.16: Geometry of in situ stations covered by two ERS data scenes at two different
incidence angles.

the same for both (a) and (b) except for the cyan station. The variation of the surface
scattering of the red and green stations are the same in both parts, but compared with other
stations our estimate is higher than the actual data. This problem, though, is reconciled
∆dB
when we apply ∆ θ ( θi − θc ) to the data in part (a). Because the incidence angles for the
∆dB
red and green stations are higher, θi − θc is smaller and, therefore, ∆ θ ( θi − θc ) for these
two stations is smaller than with lower incidence angles. So for these stations, the only
significant difference between our modeled and measured data occurs at the cyan station.
In Figure 5.14, the trend of the variation of surface scattering is the same for both parts.
In Figure 5.15, the last six stations behave similarly in part (a) and (b). The green and blue
stations, however, do not show the same behavior in both parts.
One of the scenes in Figure 5.1 has a neighboring scene available that covers a similar
ground area. In other words, we were able to find a scene with a large area covered by
another scene with a different look angle. This will allow us to compare the behavior of
every pixel in the area of overlap. The times of acquisition between these two images
are just 3 days apart. So, we can assume the correlation and backscattered power of the
intersection of these two scenes essentially remains the same. Figure 5.18 part (a), (b) and
5.18 part (c), (d) show the backscattered power and ERS1-2 tandem pair correlation of the
intersection area with the incidence angles equal to 21.55◦ and 24.28◦, respectively. The
perpendicular baselines are 224m and 204m, respectively. As we can see in parts (a) and
(c), the backscattered power is about 0.5 dB higher for the low incidence angle. This is
due to surface scattering as discussed earlier. The interesting point in the correlation of
90 CHAPTER 5. SURFACE SCATTERING

Figure 5.17: Map of four ERS scenes. The in situ base stations within theses scenes were
covered by another ERS scene at different incidence angles.

parts (b) and (d) is that, although the perpendicular baseline is higher in Figure 5.18.b, the
correlation is higher. According to Figure 5.4 volume only correlation should be lower in
Figure 5.18.b. The best explanation for this high correlation is the contribution of surface
scattering as we have discussed in Section 5.3. For the intersecting area of the two scenes,
∆dB
we find the slope of the backscattered power ∆θ , using Equation 5.14. We then estimate
surface scattering σs,dB as explained earlier in this section. In other words, we apply the
same method to each pixel in these data sets as we did on InSAR observations of the in
∆dB
situ stations. The ∆θ and inferred σs,dB of each point in the intersected area are depicted
in Figure 5.19 part (a) and (b), respectively. The correlation between these two images
5.4. RESULTS 91

- 7.5
0.7
-8

V S V
- 8.5
QR 0.6

79k
79k

ρ
σ(
0
-9
0.5
-9 5 P

53 TU 53 TU
(a) (b )
-7.5
0.7
-8
V S V
-8.5 QR 0.6
79k

79k

ρ
σ(
0

-9
0.5
-9 5 P

53 TU 53TU
(c) (WX

Figure 5.18: (a) Normalized backscattered power and (b) correlation, of an area
(74.18N,42.07W) at 21.55 degrees incidence. B⊥ = 224m in these two images. (c) Normal-
ized backscattered power and (d) correlation, of the same area in at 24.28 degrees incidence.
B⊥ = 204m in these two images. Comparing the backscattered power in (a) and (d), we can
see that surface scattering can increase the backscattered power for low incidence angles.
Although the image at part (b) has a longer perpendicular baseline than the images at part
(d) and has a lower incidence angle, the correlation is lower in part (d). This shows how
surface scattering can increase b and therefore the total correlation.

is 0.58. Hence, our model estimate of surface scattering and b closely follow the actual
surface scattering. Note that the image in part (a) is the calculated surface scattering data
from backscattered power while the image in part (b) is the inferred surface scattering using
i) our model and ii) the effect of surface scattering on correlation, b (as in Section 5.4.1).
92 CHAPTER 5. SURFACE SCATTERING

2.4
0.3
_ _ 2
0.2 ZY

∆d ∆θ
[

79k
79k

s dB
1.6

σ
0.1 1.2

0 0.8

53k ^ 53k ^
( \] (b)

Figure 5.19: (a) Slope of backscattered power vs. angle using the two backscattered power
scenes in Figures 5.18.a and 5.18.c. (b) Surface scattering inferred from the data.

Figures 5.12, 5.13, 5.14, 5.15 and 5.19 show that our model estimates of surface scatter-
ing agree qualitatively with surface scattering data. As we have seen in previous Chapters,
we need to know both surface and volume scattering of the snow to better estimate the
accumulation rate. Here we have shown a method to find the surface scattering for a wider
range of look angles. This will help us to improve our estimate of accumulation rate over a
wider geometry.
In next section, we show how the slope of the backscattered power yields estimates
the accumulation rate in the dry snow zone. We also show that estimating b enables us to
estimate accumulation rate.

5.5 Accumulation rate and slope of backscattered power


as a function of angle

In order to parameterize the σ 0 signature of the Greenland surface, Drinkwater et al. (2001)
introduced an empirical observation model,

σ 0 = A + B(θ − 40) (5.15)


5.5. ACCUMULATION RATE AND SLOPE OF BACKSCATTERED POWER 93

A (in dB) represents the normalized backscatter power at 40◦ . A key application of A is
in delineating Greenland ice facies (Fahnestock et al., 1993; Long and Drinkwater, 1994).
The dry snow zone is characterized by low A values throughout the center of Greenland.
Because the grains of the dry snow are relatively small, the microwave penetrates deep in
the snow and are absorbed, producing little backscatter. Further down-slope in the perco-
lation zone, the summer melt contributes to the formation of subsurface ice pipes and ice
lenses, which form when the wet snow freezes. This process increases the value of A.

800
Accumulation Rate (mm)

600

400

200

−0.22 −0.2 −0.18 −0.16 −0.14 −0.12


B (dB/deg)

Figure 5.20: Comparison of Scatterometer B values with in situ accumulation data (in mm
w.e.) in the dry snow zone of Greenland. The curve in this figure indicates the least squares
fit to the EScat B data (Drinkwater et al., 2001).

B (in dB deg−1 ) is an indication of the magnitude of the linear dependence of σ 0 on


incidence angle. B, as we can see from our development above, is an empirical value relat-
ing the relative contributions of volume and surface scattering. Greater contributions from
surface scattering as compared to volume scattering increase the magnitude of B. For ERS
|B| ≈ 0.3 dB deg−1 in the upper percolation zone, which is nearly double that observed
across the rest of the ice sheet (Ashcraft and Long, 2005). This suggests that the relative
94 CHAPTER 5. SURFACE SCATTERING

contribution from surface scattering is much more than for other areas. Variations of B
within the dry snow zone are attributed in the literature to variations in the accumulation
rate. In Figure 5.20, Drinkwater et al. (2001) compared the values of B with in situ accu-
mulation rate Q in the dry snow zone of Greenland. The EScat mode of the ERS Active
Microwave Instrument (AMI) collects C band VV backscattered power as a function of
angle. The curve in Figure 5.20 indicates the least squares fit to the EScat B data. This
figure shows that more negative B values correspond to regions of higher accumulation.
Note that this is a least squares fit to the scatterometer data, and the individual data are less
well behaved (Drinkwater et al., 2001, Figure 1.b).
If we assume that B remains constant over different incidence angles then we can relate
the Drinkwater et al. (2001) results to our own model development through

∆dB
B= (5.16)
∆θ

∆dB
In Figure 5.21, we plot our measured values of ∆θ for all of the in situ stations in
Figures 5.12-5.15 as a function of their accumulation rates. For clearer comparison, the
∆dB
∆θ values are separated by their incidence angles in this figure. Figure 5.21 confirms the
∆dB
finding of Drinkwater et al. (2001) in Figure 5.20. For instance in Figure 5.21.d, ∆θ as
a function of accumulation rate is depicted for all the in situ stations with 23◦ incidence
∆dB
angle. As we can see, the in situ stations with higher ∆θ have higher accumulation rates
in agreement with Figure 5.20.
σs,dB ∆dB
From Equation 5.13, we know that estimated θi −θc is equal to ∆θ , which is the defi-
nition of B in Equation 5.15. Hence, an estimate of σs,dB will not only improve our model
and make it applicable to wider geometries, but also gives us another way to estimate the
accumulation rate. In Figures 5.22 to 5.25, we have plotted our own estimates of B as a
function of accumulation rate. We estimate B by dividing σs,dB in Figures 5.12.b - 5.15.b,
by θi − θc . We used an angle between 40◦ to 50◦ for θc for the different scenes following
Ashcraft and Long (2006). This angle could differ for different in situ stations inside a
scene, but the overall result is highly dependent on the reference angle as the relationship
of σ 0 in dB with angle is nearly linear in this range.
We showed in Figure 5.16 that each in situ station is covered by two radar scenes.
5.6. CONCLUSIONS 95

The incidence angle of the base station is always higher in one than the other. Therefore,
both InSAR correlation and backscattered power observables are available at two incidence
angles. Hence, b and consequently σs,dB can be inferred at two incidence angles for each
base station. The B estimated using the σs,dB from the scenes with higher incidence angle
agrees better with the accumulation rate. This could be due to changes in the slope of the
surface scattering term for lower incidence angles. Hence in Figures 5.22 to 5.25, we have
used the data from the scenes with higher incidence angles. In some cases, this higher-
angle scene does not cover all the in situ stations. Therefore, there are three in situ stations
in Figures 5.13 and 5.14 that are not shown in Figures 5.23 and 5.25.
In Figure 5.22 to 5.25, we depict the estimated B as a function of accumulation rate for
the in situ stations of different frames. As we can see in these figures, higher accumulation
rate on average corresponds to higher B. The exceptions are three in situ stations in Figure
5.24 and the purple station in Figure 5.25. The three exceptions in Figure 5.24 are the green,
blue, and the purple circle station. We have also shown that the green and blue stations were
not well behaved in Figure 5.15.b. Hence, the exception of the green and blue stations in
Figure 5.24 is likely because of an incorrect estimate of surface scattering resulting from an
unknown physical condition that we have neglected to model. Thus in general it is safe to
say that higher accumulation rates increase estimated B. This is also what the fitted curve in
Figure 5.20 shows using the real surface scattering data. Note that the estimated values of B
in our figures are close to those in Figure 5.20 for the corresponding range of accumulation
rates of the in situ stations.

5.6 Conclusions
As is easily discerned in the average normalized backscattered power as a function of an-
gle plots (Figure 5.2), surface scattering plays a significant role in modeling backscattered
power from dry snow. In this Chapter, we demonstrated the effect of changes in surface
scattering on both correlation and the backscattered power. We infer that any model ig-
noring surface scattering cannot be accurate at C-band frequencies. We have defined a
parameter b to quantify the relative effect of surface scattering on InSAR observables. We
saw that the value of surface scattering inferred under our model σs,dB , closely parallels the
96 CHAPTER 5. SURFACE SCATTERING

∆dB
observed slope of the surface backscattering, ∆θ . This helps to generalize our model to
wider look angles and wider geometries.
We also compared the estimate of the slope of surface scattering, B, from our model,
with accumulation rate. We saw that higher values of accumulation rate qualitatively cor-
respond to higher values of B. This gives another method to estimate the accumulation rate
in addition to our previous method of estimating h. Combining the estimate of B with the
prediction of the layer spacing h, likely gives a better estimate of accumulation rate, but
this is work left for the future.
5.6. CONCLUSIONS 97

θi=20 θ =21
i
0.5 0.7

0.45 0.6
0.4
0.5

∆dB/∆θ
∆dB/∆θ

0.35
0.4
0.3

0.25 0.3

0.2 0.2
14 16 18 20 22 24 5 10 15 20 25 30
acc-rate(cm) acc-rate(cm)

(a) (b)
θi=22 θi=23
0.8 0.8

0.6 0.6
∆dB/∆θ

∆dB/∆θ

0.4 0.4

0.2 0.2

0 0
15 20 25 30 5 10 15 20 25 30
acc-rate(cm) acc-rate(cm)

(c) (d)
θi=24 θi=25
0.7 0.5

0.6
0.4
0.5
∆dB/∆θ

∆dB/∆θ

0.3
0.4
0.2
0.3

0.2 0.1
10 15 20 25 5 10 15 20 25 30
acc-rate(cm) acc-rate(cm)
(e) (f)

Figure 5.21: Slope of backscattered power vs. accumulation rate (in cm w.e.) for In Situ
base station in Figures 5.12, 5.13, 5.15 and 5.14. They have been separated based on their
incidence angle; (a) θi = 20, (b) θi = 21, (c) θi = 22, (d) θi = 23, (e) θi = 24 and (f) θi = 25
98 CHAPTER 5. SURFACE SCATTERING

23774−2151
0.34

0.32

0.3
B

0.28

0.26

0.24
21.5 22 22.5 23
acc rate (cm)

Figure 5.22: Inferred B for the in situ stations in Figure 5.12.

23832−2043
0.155

0.15
B

0.145

0.14

0.135
16 16.5 17 17.5 18
acc rate (cm)

Figure 5.23: Inferred B for the in situ stations in Figure 5.13.


5.6. CONCLUSIONS 99

23846−2097
0.3

0.28

0.26
B

0.24

0.22

0.2

0.18
15 20 25 30
acc rate (cm)

Figure 5.24: Inferred B for the in situ stations in Figure 5.15.

23774−2169
0.3

0.28
B

0.26

0.24

0.22
22 24 26 28 30
acc rate (cm)

Figure 5.25: Inferred B for the in situ stations in Figure 5.14.


100 CHAPTER 5. SURFACE SCATTERING
Chapter 6

Conclusions

Radar remote sensing is a powerful method for monitoring the icy polar regions on a conti-
nental scale and with regularity without requiring scientist to travel to these regions. In this
dissertation, we introduced an interferometric ice scattering model that relates InSAR ob-
servables to field observable properties of snow proportional to the accumulation rate in the
dry snow zone of Greenland. Snow accumulation in remote regions such as Greenland and
Antarctica is a key factor for estimating Earth’s ice mass balance. We verified the results
by comparing our solutions to in situ accumulation data, and also with empirical solutions
from spaceborne data as reported in the literature.
In the absence of enough in situ data, estimating snow accumulation from remote sens-
ing observations, such as microwave thermal emission and radar brightness can potentially
be very useful. In order to do so, these data need to be interpreted using electromagnetic
models. In most of the models proposed in the literature only volume scattering has been
considered, as it is the dominant mechanism. However, this approximation becomes less
accurate if the ice sheet is layered. This is because larger grain size and thicker annual
layers both increase radar image brightness, whereas the first corresponds to lower accu-
mulation rate and the second to higher accumulation rate, so models of radar brightness
alone have no chance of accurately reflecting accumulation. To resolve this ambiguity,
interferometric radar correlation measurements have been considered in this thesis. Here
we introduce an ice scattering model that relates InSAR correlation and radar brightness to
both ice grain size and hoar layer spacing in the dry snow zone of Greenland. We use this

101
102 CHAPTER 6. CONCLUSIONS

model and ERS satellite radar observations to derive several parameters related to snow
accumulation rates in a small area in the dry snow zone. These parameters show agreement
with four in situ core accumulation rate measurements in this area, while models using only
radar brightness data do not match the observed variation in accumulation rates.
We show that modeling both surface and volume scattering mechanisms is necessary
to accurately predict the accumulation rate in the dry snow zone of Greenland. In order
to estimate the surface and volume scattering terms, we process data covering several base
stations in the dry snow zone of Greenland acquired at two different incidence angles. The
angle diversity is particularly helpful in characterizing the surface term.

6.1 Contributions
The interferometric ice scattering model and its application to scattering in Greenland’s
dry snow zone that we used in this dissertation to estimate the accumulation rate are our
primary contributions. Specifically,

(1) We have shown that greater grain radius and greater layer spacing h both increase
backscattered power in observations of ice sheets. The former corresponds to lower
accumulation rate, whereas the latter corresponds to higher accumulation rate; Hence,
we cannot reliably estimate accumulation from backscattered power alone. To a first-
order approximation, brighter areas correspond to lower accumulation rates, but this
correspondence breaks down in more detailed analysis. We show that considera-
tion of correlation together with backscattered power measurements can resolve this
ambiguity.

(2) We have presented a model that can retrieve snow accumulation parameters more
accurately than existing remote sensing methods. We introduced an interferomet-
ric ice scattering model relating InSAR correlation and radar brightness to both ice
grain size and hoar layer spacing in the dry snow zone of Greenland. We used this
model and ERS satellite radar observations to derive several parameters related to
snow accumulation rates in a small area in the dry snow zone. These parameters
show agreement with four in situ core accumulation rate measurements in this area,
6.2. FUTURE WORK 103

while models using only radar brightness data do not match the observed variation in
accumulation rates. Our method is more accurate than conventional remote sensing
in areas of nearly uniform accumulation rate, and we can more precisely estimate
deviations in the accumulation rate.

(3) Our model also demonstrates that surface scattering plays a significant role in mod-
eling backscattered power from dry snow. Easily discerned in the average backscat-
tered power as a function of angle plots, any model ignoring surface scattering can-
not be accurate. We generalized the surface scattering term in the interferometric ice
scattering model in the dry snow zone to apply the theory to a wider range of imag-
ing geometries. We processed new data sets over each base station with two different
incidence angles in order to better estimate the surface scattering contribution. We
introduced the parameter b to quantify the relative intensity of surface scattering and
its effect on correlation. The value of surface scattering inferred under this model
closely parallels empirical determinations.

(4) We also showed that the slope of surface scattering, B, from our model, helps derive
the accumulation rate. We saw that higher values of accumulation rate correspond
to higher values of B. This gives another method to estimate the accumulation rate
in addition to estimating h. Combining the estimate of B with the prediction of the
layer spacing h, gives a better estimate of accumulation rate.

6.2 Future Work


This dissertation represents a step forward in estimating the accumulation rate in the dry
snow zone of polar ice and exploring some of the properties of these remote regions. The
approach to modeling scatter from ice sheets we developed is a useful tool for understand-
ing the Earth’s ice sheets. As with any new approach, of course, there are unsolved prob-
lems needing attention. There are several aspects of particular problems that are grounds
for interesting and relevant future work:

(1) As we show, backscattering from snow is a complex process. We still need even
better estimates of volume and surface scattering in the dry snow zones. Collecting
104 CHAPTER 6. CONCLUSIONS

remote sensing data at other incidence angles or other frequencies would help us im-
prove our understanding of volume and surface scattering. Instead of using a constant
grain radius, R0 , for the entire scene, we would be able to separate the volume and
surface component of backscattering for individual pixels in the image. We can also
use other volume and surface backscattering models with more rigorous solution and
fewer assumptions. Also, more recent field data can help to test our model estimates
of accumulation rate. At present, the number of in situ stations are very few within
our imaged areas, and the data from these fields are very old and possibly even out
of data.

(2) In section 3.3.6 we showed that the 24 days repeat cycle of RADARSAT-1 causes
high temporal decorrelation. Therefore, we could not apply our model to the RADAR-
SAT-1 data, which could give us the the information about the backscattering proper-
ties of the dry snow zone at higher incidence angles. Having InSAR data at different
incidence angles, frequencies, or polarizations, would improve our model and there-
fore our estimate of accumulation rate. Including scatterometry, radiometry and other
InSAR data at different frequencies also would give us a more accurate description
of the polar ice. A more comprehensive model would be able to explain more data
features.

(3) UAVSAR, a reconfigurable, polarimetric L-band synthetic aperture radar is specif-


ically designed to acquire airborne repeat track SAR data for differential interfero-
metric measurements. Using UAVSAR, we can investigate the temporal correlation
more accurately in the dry snow zone of Greenland. In addition, its look angle is
steerable between 25◦ − 60◦ and so it can be used to expand our model to a wider set
of geometries. Volume backscattering in the dry snow zone is smaller at L-band than
C-band frequency (section 4.2.1). Therefore, the relative surface to volume backscat-
tering becomes much larger in L-band than C-band. This can improve our estimates
of b and therefore surface and volume scattering.
Bibliography

Abdalati, W. and Steffen, K. (1998). Accumulation and Hoar effects on microwave emis-
sion in the Greenland ice-sheet dry-snow zones. Journal of Glaciology, 44(148):523–
531.

Alley, R. B., Saltzman, E. S., Cuffey, K. M., and Fitzpatrick, J. J. (1990). Summertime for-
mation of depth hoar in central Greenland. Geophysical Research Letters, 17(12):2393–
2396.

Ashcraft, I. S. and Long, D. G. (2005). Observation and Characterization of Radar


Backscatter Over Greenland. IEEE Transactions on Geoscience and Remote Sensing,
43(2):225–237.

Ashcraft, I. S. and Long, D. G. (2006). Relating Microwave Backscatter Azimuth Mod-


ulation to Surface Properties of the Greenland Ice Sheet. Journal of Glaciology,
52(177):257–266.

Benson, C. S. (1962). Stratigraphic studies in the snow and firn of the Greenland ice sheet.
Technical Report, SIPRE, U.S. Army Corps of Engineers, (70).

Bindschadler, R. (1998). Monitoring ice sheet behavior from space. Reviews of Geophysics,
36(1).

Bingham, A. W. (1997). Monitoring Arctic glaciers and ice caps using satellite remote
sensing. PhD thesis, University of Cambridge, Cambridge, England.

Curlander, J. C. and McDonough, R. N. (1991). Synthetic Aperture Radar Systems and


Signal Processing. Wiley, New York.

105
106 BIBLIOGRAPHY

Drinkwater, M. R., Long, D. G., and Bingham, A. W. (2001). Greenland snow accumula-
tion estimates from satellite radar scatterometer data. Journal of Geophysical Research,
106(D24):33935–33950.

Fahnestock, M., Kwok, R., and Jezek, K. (1993). Greenland ice sheet surface properties
and ice dynamics from ERS-1 SAR imagery. Science, 262:1530–1534.

Gatelli, F., Guarnieri, A. M., Parizzi, F., and Pasquali, P. (1994). The wavenumber shift in
SAR interferometry. IEEE Transactions on Geoscience and Remote Sensing, 32(4):855–
865.

Gogineni, P., Legarsky, J., and Thomas, R. (1998). Coherent radar depth sounder measure-
ments over the Greenland ice sheet. IEEE Proceedings of IGARSS, 4:2258–2260.

Herron, M. and Langway, C. (1980). Firn densification: An empirical model. Journal of


Glaciology, 25(93):373–385.

Hoen, E. and Zebker, H. (2000). Penetration Depth Inferred from Interferometric Volume
Decorrelation Observed over the Greenland Ice Sheet. IEEE Transactions on Geoscience
and Remote Sensing, 38(6):2571–2583.

Hoen, E. W. (2001). A Correlation-Based approach to modeling interferometric Radar


Observations of the Greenland Ice Sheet. PhD thesis, Stanford University.

Ishimaru, A. (1978). Wave propagation and scattering in a random media. Academic


Press, New York.

Jordan, R. (1991). A one dimensional temperature model for snow cover. CRREL Spec.
Rep. 91-16, U.S. Army Cold Region Res. and Eng. Lab., Hanover, N.H.

Laur, H., Bally, P., Meadows, P., and Sánchez, J. (2003). ERS SAR Calibration: Derivation
of σ 0 in ESA ERS SAR PRI Products. Technical Report 2.

Long, D. and Drinkwater, M. (1994). Greenland Ice Sheet surface properties observed by
the Seasat-A scatterometer at enhanced resolution. Journal of Glaciology, 40(1):213–
230.
BIBLIOGRAPHY 107

Mätzler, C. (1987). Application of the interaction of Microwave with the Natural Snow
Cover. Remote Sensing Reviews, 2:259–387.

Mätzler, C. (1998). Improved Born Approximation for Scattering of radiation in a granular


medium. Journal of Applied Physics, 83(11):6111–6117.

Munk, J., Jezek, K. C., Forster, R., and Gogineni, S. P. (2003). An accumulation map for
the Greenland dry-snow facies derived from spaceborne radar. Journal of Geophysical
Research, 108(D9):8–1–8–12.

Ohmura, A. and Reeh, N. (1991). New Precipitation and accumulation maps for Greenland.
Journal of Glaciology, 37:140–148.

Onn, F. and Zebker, H. A. (2006). Correction for interferometric synthetic aperture radar at-
mospheric phase artifacts using time series of zenith wet delay observations from a GPS
network . Journal of Geophysical Research, 111(B09102). doi:10.1029/2005JB004012.

Oveisgharan, S. and Zebker, H. A. (2004). A Snow Accumulation Map For the Dry Snow
Region of Greenland Derived from InSAR Correlation Observations . San Francisco.

Oveisgharan, S. and Zebker, H. A. (2007). Estimating Snow Accumulation From In-


SAR Correlation Observation. IEEE Transactions on Geoscience and Remote Sensing,
45(1):10–20.

Paterson, W. S. B. (1994). The Physics of Glaciers. Pergamon.

Rott, H. and Rack, W. (1995). Characterization of Antarctic firn by means of ERS-1 scat-
terometer measurements. Geoscience and Remote Sensing Symposium, 3:2041–2043.

Rott, H. and Siegel, A. (1996). Glaciological Studies in the Alps and in the Antarctica
Using ERS Interferometric SAR. FRINGE.

Rott, H., Sturm, K., and Miller, H. (1993). Active and passive microwave signature of
Antarctic firn by means of field measurements and satellite data. Annals of Glaciology,
17:337–343.
108 BIBLIOGRAPHY

Rumi, J., Barks, C., Moyne, J., and Arberry, A. (2004). The Essential Rumi. Penguin
Books Ltd, UK.

Stogryn, A. (1986). A study of the microwave brightness temperature of snow from the
point-of-view of strong fluctuation theory. IEEE Transactions on Geoscience and Re-
mote Sensing, 24(2):220–231.

Tsang, L. (1987). Passive remote sensing of dense nontenuous media . Journal of Electro-
magnetic Waves and Applications, 1(2):159–173.

Tsang, L., Kong, J., and Shin, R. (1985). Theory of Microwave Remote Sensing, Active and
Passive. Addison-Wesley, London.

Tsang, L. and Kong, J. A. (1980). Multiple scattering of electromagnetic waves by ran-


dom distributions of discrete scatterers with coherent potential and quantum mechanical
formulism. Journal of Applied Physics, 51(7):3465–3485.

Tsang, L. and Kong, J. A. (1981). Scattering of electromagnetic waves from random media
with strong permittivity fluctuations. Radio Science, 16(3):303–320.

Ulaby, F., Moore, R., and Fung, A. (1981). Microwave Remote Senising, Active and Pas-
sive. Addison-Wesley, London.

West, R. D. (1994). Microwave Emission From Polar Firn. PhD thesis, University of
Washington.

Winebrenner, D., Arthern, R., and Shuman, C. (2001). Mapping Greenland accumulation
rates using observations of thermal emission at 4.5-cm wavelength. Journal of Geophys-
ical Research, 106(D24):33919–33934.

Wismann, V., Winebrenner, D. P., Boehnke, K., and Arthern, R. J. (1997). Snow accumu-
lation on Greenland estimated from ERS scatterometer data. Geoscience and Remote
Sensing, 4:1823–1825.

Zebker, H. A. and Goldstein, R. M. (1986). Topographic mapping from interferometric


SAR observations . Journal of Geophysical Research, 91:4993–4999.
BIBLIOGRAPHY 109

Zebker, H. A. and Villasenor, J. (1992). Decorrelation in Interferometric Radar Echoes.


IEEE Transactions on Geoscience and Remote Sensing, 30(5):950–959.

You might also like