03c

COMPUTATIONAL ELECTROMAGNETIC
SCATTERING MODELS FOR MICROWAVE

REMOTE SENSING
KUNG-HAU DING
Air Force Research Laboratory
Hanscom AFB, Massachusetts
LEUNG TSANG
University of Washington
Seattle, Washington
QIN LI
University of Washington
Seattle, Washington
1. INTRODUCTION
Modeling of electromagnetic wave scatterings by random
discrete scatterers and rough surfaces play an important
role in geoscience and remote sensing research. Since the
1970s, considerable theoretical efforts have been made to
elucidate and understand the scattering processes in-
volved in such problems, and various models have been
developed for microwave active and passive remote sen-
sing applications [17]. With the rapid advances in com-
puter technology and fast computational electromagnetics
algorithms, numerical simulations of scattering by ran-
dom media allow us to solve Maxwells equations exactly
without the limitations of analytical approximate models.
The numerical models can provide a valuable means for
evaluating the validity regimes of analytical scattering
theories; in addition, they can potentially aid in the future
development of extended analytical models. These theore-
tical and numerical models, which are commonly used for
tackling electromagnetic wave scattering problems and
for remote sensing applications, have been presented in
three volumes of books [57]. In this article, we will update
the development in the numerical scattering models for
discrete random scatterers, with emphasis on the applica-
tions of microwave remote sensing in snowcover, seafoam,
and vegetation canopy.
The frequency dependence of scattering by geophysical
media at microwave frequencies is an important topic
because multifrequency measurements are used in remote
sensing applications. In Section 2, we investigate rigor-
ously the frequency dependence of scattering by dense
media [79]. The approach used is based on the Monte
Carlo simulations where the three-dimensional solutions
of Maxwells equations are pursued [6]. The particle
positions are generated by deposition and bonding techni-
ques. The properties of absorption, scattering, and extinc-
tion are calculated for dense media consisting of sticky
and nonsticky particles. Numerical solutions of Maxwells
equations indicate that the frequency dependence of den-
sely packed sticky small particles is much weaker than
that of independent scattering. Numerical results are
illustrated using parameters of snow in microwave remote
sensing. Comparisons are made with extinction measure-
ments as a function of frequency.
In Section 3, polarimetric microwave emissions from
foam-covered ocean surfaces are studied. The foam is
treated as densely packed air bubbles coated with thin
seawater coating [1012]. The absorption, scattering, and
extinction coefcients are computed from the Monte Carlo
solutions of Maxwells equations for a collection of coated
particles. These quantities are then applied in the dense
media radiative transfer (DMRT) theory [2,7] to calculate
the polarimetric microwave emissivities of ocean surfaces
with foam cover. The theoretical results of Stokes bright-
ness temperatures with typical parameters of foam in
passive remote sensing at 10.8 and 36.5GHz are illustrated
and compared with experimental measurements [1012].
We present an efcient computational model for comput-
ing tree scattering at VHF/UHF frequencies in Section 4. A
structure model with dielectric cylinders is used to simulate
trees with bare branches. The method of moments (MoM) is
applied to solve the volume integral equation for the tree
scattering signatures. An efcient numerical algorithm
based on the sparse matrix iterative approach (SMIA) is
applied in solving the matrix equation iteratively [1315].
The SMIA decomposes the impedance matrix into a sparse
matrix for the near interactions, and a complementary
matrix for the far interactions among the cylindrical sub-
cells of the tree structure. The SMIA tree scattering model
is applied to calculate scattering from various simulated
trees with up to several hundreds of branches using a
laptop computer. Solutions obtained fromthe SMIAmethod
agree very well with the solutions obtained using exact
matrix inversion and the conjugate gradient method
(CGM). The key feature of the SMIA approach is that
very little iteration is required to obtain convergent solu-
tions, compared to the CGM, the SMIA approach may
reduce the number of iterations by a factor of 4100 [16].
In Section 5, a UV multilevel partitioning (UV-MLP)
method is presented for solving the volume scattering
problem [1720]. The method consists of setting up a
rank table of transmitting and receiving block sizes and
their separations. The table can be set up speedily using
coarsecoarse sampling. For a specic scattering problem
with given geometry, the scattering structure is parti-
tioned into multilevel blocks. By looking up the rank in the
pre-determined table, the impedance matrix for a given
transmitting and receiving block is expressed by a product
of U and V matrices. We demonstrate the method for two-
dimensional volume scattering by discrete scatterers.
Multiple scattering is cast into the FoldyLax equations
of partial waves [2,57]. We show that the UV decomposi-
tion can be applied directly to the impedance matrix of
partial waves of higher order than the usual lowest-order
Green function. Numerical results are illustrated for
randomly distributed cylinders with diameter of 1 wave-
length. For scattering by 1024 cylinders on a single PC
processor with 2.6-GHz CPU and 2 GB (gigabytes) of
memory, only 14 CPU minutes is needed to obtain the
numerical solution and, for 4096 cylinders, only 7.34 s is
needed for one matrixvector multiplication.
2. FREQUENCY DEPENDENCE OF SCATTERING BY DENSE
MEDIA WITH APPLICATIONS TO SNOW
For remote sensing applications, the measurements are
generally taken at multiple frequencies [2125] because
718 COMPUTATIONAL ELECTROMAGNETIC SCATTERING MODELS FOR MICROWAVE REMOTE SENSING
Previous Page
each frequency contains information. For instance, the
special sensor microwave/imager (SSM/I) radiometer mea-
sures at 19, 22, 37, and 85 GHz. Backscattering data are
usually measured at L, C, and X bands in the active
remote sensing case [14]. These multiple frequency mea-
surements are used in many applications, such as the
snow parameter retrievals, the surface type classication,
and metamorphism signature of snow. Thus, the fre-
quency dependence of scattering is an important issue
for making a better use of multifrequency data in remote
sensing. The extinction of dry snow is due primarily to
scattering rather than absorption. Since the snow grain
sizes are much smaller than wavelength, classical Ray-
leigh scattering theory indicates that the dependence on
scattering is frequency to the fourth power. The important
question of dense media is whether the scattering still
exhibits frequency to the fourth power when these small
particles are densely packed. Several approaches have
been used in studying the frequency dependence of scat-
tering. The rst is based on experimental measurements
and empirical tting. Hallikainen et al. [22] measured the
extinction coefcients of snow at 18, 35, 60, and 90 GHz.
The second approach is based on analytical modeling
using effective medium theories [2,7]. Analytical ap-
proaches that account for correlations between particles,
such as the quasicrystalline approximation (QCA) and
quasicrystalline approximation with coherent potential
(QCA-CP) [2,7], have been used to simulate the densely
packed media. The PercusYevick (PY) pair distribution
functions are used to model the relative location of
particles. The dense media radiative transfer equations
have been derived based on QCA for sticky particles
and found useful applications in microwave remote sen-
sing of snow [7,26]. At higher frequencies, particles be-
come comparable to wavelength and higher-order
multipoles have to be included [2729]. The third ap-
proach is a rigorous approach based on Monte Carlo
(MC) simulations of exact solutions of Maxwell equations
that govern microwave interactions exactly [6]. The
method consists of generating the positions of the particles
on a computer on the basis of prescribed rules. Once the
positions are generated, the Maxwell equations are solved
numerically. Numerical methods include the T-matrix
method [2,7], the nite-difference time-domain (FDTD)
method [30], and the volume integral equation method
[6]. Monte Carlo simulations have been applied to study
the extinction coefcients and the effective permittivities
for dense media [2735]. In 20012003, we conducted two-
and three-dimensional Monte Carlo simulations to inves-
tigate the frequency dependence of scattering by dense
media [79].
In this section, we describe the three-dimensional
dense medium scattering simulations for media with
small and moderate size particles. Particles with (sticky)
and without interparticle force (nonsticky) are both used
in simulations to study the frequency dependence of the
extinction coefcient. The T-matrix method is used, and
the effects of absorption are included. For typical snow
grains, the diameters are in millimeter ranges, and the
size parameter ka values are ranging from 0.2 to 1
depending on the frequencies. It is shown that densely
packed sticky particles exhibit a frequency dependence of
scattering coefcient weaker than that of independent
scattering. The methodology of dense media study consists
of studying the pair distribution functions of particle
positions, computer growth of dense media, and three-
dimensional solutions of Maxwell equations for the com-
puter-grown dense media. Section 2.1 brief describes the
pair distribution functions and the computer creation of
discrete random media with sticky particles. The three
types of formulations of FoldyLax multiple scattering
equations, based on exciting eld, scattered eld, and
internal eld, for spheres are presented in Section 2.2.
We also describe how to calculate scattering and absorp-
tion coefcients. Convergence tests are performed using
various numerical parameters to show that the results are
accurate. In Section 2.3, simulation convergence with
respect to the number of particles used, the number of
multipoles used, the box size used, and the number of
realizations are demonstrated. In Section 2.4, we present
results based on Monte Carlo simulations of densely
packed sticky and nonsticky dielectric spheres. The Monte
Carlo simulation results are compared with independent
scattering and QCA results. The results of Monte Carlo
simulations are in good agreement with QCA results.
Simulated results are shown for various size parameter
ka and stickiness parameter t for the four microwave
frequencies at 18, 35, 60, and 90 GHz using typical para-
meters in snow. Comparisons are also made with extinc-
tion measurements [22] as a function of frequency.
2.1. Pair Distribution Functions and Computer Generation of
Ensemble of Particle Positions for Finite-Size Sticky Particles
The pair distribution functions for various types of mate-
rials, which can be classied as gas, crystal, and liquid/
glass, are described in detail in three other texts [2,6,36].
The key observation is that liquid is a dynamical system
with moving molecules while glass is amorphous solid that
is static. Yet they have same probability density function
(pdf) and joint pdf of particle positions. They both have
short-range order with the same class of pair distribution
functions. Another comparison is using the Bernal ran-
dom closed-packed sphere model, which is a macroscopic
static situation, because in this model the spheres are
dropped into the container. On the other hand, the Per-
cusYevick (PY) hard-sphere model is based on molecular
dynamics of liquids using a hard-sphere potential. Yet
the Bernal random closed-packed sphere model and the
PY model give very similar pair distribution functions. In
practice, the Bernal random packing is of limited use,
due to its xed volume fraction at about 63%. From
these studies one can see that static situations can be
generated from dynamic situations because one can view
the dynamics as merely random walk or shufing of
particles to create realizations of random positions of
particles. In other words, shufing of particle positions
creates the static situation of various statistical ensem-
bles. The particles do not actually move in a natural snow
medium.
To understand the scattering of waves by dense dis-
crete random media, it is important to systematically
COMPUTATIONAL ELECTROMAGNETIC SCATTERING MODELS FOR MICROWAVE REMOTE SENSING 719
generate the ensemble of random positions of particles
using the procedure of computer growth. To generate
the static case of dense random media with fractional
volume between 20% and 50%, we have been using
shufing, which is viewed as a means of creating different
realizations rather than viewing particles as actually
moving in real life. The particles are rst placed periodi-
cally, which is possible up to 70%. Then the particles
are shufed, by random walk, to attain the random
positions. The results of particle positions actually look
quite realistic.
Aggregation is introduced by allowing particles to
bond. The diffusion-limited aggregation methods [37] are
not useful as they are limited to low fractional volume,
and the structure created is quite sparse not dense as in
snow grain. The fractional volume decreases as the
number of particles increases for fractals generated by
diffusion limited aggregation. The manner in which bond-
ing is done is actually static. The probabilities are
constructed based on the exposed surface area of the
particles. Then bonding is decided on the basis of these
probabilities [6,3840]. We call these particles sticky
because they bear similarities to the PercusYevick
liquid model of a sticky potential; the pair functions are
also similar.
Sticky particle models are used to consider cases
when the particles can have surface adhesion [6]. This
means that when the particles are randomly packed
together, they tend to form clusters and bonds with
each other. The adhesive force is parameterized using a
variable t that governs the degree of clustering. Analytic
solutions of the pair distribution functions can be calcu-
lated for particles of the same size [41] and for medium
with multiple sizes of particles [42]. They have distinct
pair distribution functions that are different from those
of nonsticky particles. To generate the positions of parti-
cles in the sticky particle model, a two-step procedure
is used. First, the Metropolis shufing technique is used
to provide random positions of the particles in a box.
Next we follow the work of Seaton and Glandt [38] and
Kranendonk and Frenkel [39], to allow the particles
to form bonds. There are four binding states for each
particle: unbonded, single bond, double bond, and triple
bond. Transition probabilities are calculated. The particles
are then selected to form bonds on the basis of
these transition probabilities [6,3840]. The sequence of
bonds can be formed by different particles resulting in
aggregates.
The sticky particle model is useful to describe
snow because the ice grains in snow are adhesive.
Furthermore, as metamorphism occurs, the ice grains
form bridges. The appropriateness of the sticky particle
model can be veried by reexamining the pair distribution
functions in snow and comparing them with the sticky
particle model. It has been observed that grain size
in snow follows a lognormal size distribution [43]. The
large size grain can be viewed as a result of aggregates.
The study of relation between particle size distribution
and the alternative viewpoint of aggregation of small
particles to form larger size grains will be studied in the
future.
In the following simulations, all the particles are
assumed to have the same size. However, since the
particles can form bonds and aggregates, the effective
sizes of the particles can be much larger. Furthermore,
the aggregates consist of a varying number of particles.
This gives an effective size distribution when the
medium is viewed as aggregates of particles. The particle
positions are generated for both nonsticky and sticky
particles. After the particle positions are generated,
the conjugate gradient method is used to solve the
FoldyLax multiple scattering matrix equations described
in Section 2.2.
2.2. Numerical Solution of Maxwell Equations Based on
FoldyLax Formulation
The nal step of simulation is to solve Maxwell equations
of these realizations to study the coherent wave and
incoherent waves by averaging the electromagnetic elds
and intensities over realizations. In the dispersion rela-
tion for the coherent wave, the coherent wave has an
effective propagation constant that has a real part and an
imaginary part. The real part corresponds to effective
permittivity. The imaginary part accounts for loss due to
absorption and scattering. The scattering loss is due to the
conversion of coherent waves to incoherent waves. It
can be calculated by angular integration of the incoherent
wave intensities. In the simulations, Maxwell equations
are solved exactly. Then we calculate incoherent waves
by subtracting the coherent wave from the scattered
wave. Absorption is calculated by nding the absorption
loss of each particle, based on the solution of internal
elds of each particle. Combining scattering and absorp-
tion give the extinction of the dense random media.
However, for dry snow, the extinction is crucial and the
results of extinction are shown in Section 2.4. The proce-
dures are used for the four frequencies: 18, 37, 60, and
90 GHz.
Consider a medium consisting of N spheres, with radii
a and permittivities e
p
, centered at " rr
1
; " rr
2
; . . . ; " rr
N
. An in-
cident wave with wavevector
"
kk
i
is impinging onto the
medium. The solution of Maxwells equations can be cast
into the FoldyLax multiple scattering equations. In
matrix form [2,6,9]
" ww
(q)
=
N
p=1; pOq
s(kr
q
r
p
)T
(p)
" ww
(p)
e
i
"
kk
i
.
" rrq
" aa
inc
(1)
where " ww
(q)
are the exciting eld coefcients, q=1; 2; . . . ; N,
s(kr
q
r
p
) is the vector spherical wave transformation
matrix, T
(p)
the T matrix representing scattering from
the scatterer p, and " aa
inc
the incident eld coefcients. The
physical interpretation of Eq. (1) is that the exciting eld
at particle q is equal to the incident eld plus the scattered
elds from all other particles except itself.
The relationships between exciting eld coefcients
w
(M)
mn
and w
(N)
mn
, scattered eld coefcients a
S(M)
mn
and a
S(N)
mn
,
and internal eld coefcients c
(M)
mn
and c
(N)
mn
are as follows [6,9]
a
S(M)
mn
=T
(M)
n
w
(M)
mn
a
S(N)
mn
=T
(N)
n
w
(N)
mn
(2)
a
S(M)
mn
=S
(M)
n
c
(M)
mn
a
S(N)
mn
=S
(N)
n
c
(N)
mn
(3)
c
(M)
mn
=B
(M)
n
w
(M)
mn
c
(N)
mn
=B
(N)
n
w
(N)
mn
(4)
with T
(M)
n
, T
(N)
n
, S
(M)
n
, S
(N)
n
, B
(M)
n
, and B
(N)
n
as given by
T
(M)
n
=
j
n
(k
p
a)[kaj
n
(ka)]
/
j
n
(ka)[k
p
aj
n
(k
p
a)]
/
j
n
(k
p
a)[ka h
n
(ka)]
/
h
n
(ka)[k
p
aj
n
(k
p
a)]
/
T
(N)
n
=
(k
p
a)
2
j
n
(k
p
a)[ka j
n
(ka)]
/
(ka)
2
j
n
(ka)[k
p
aj
n
(k
p
a)]
/
(k
p
a)
2
j
n
(k
p
a)[kah
n
(ka)]
/
(ka)
2
h
n
(ka)[k
p
aj
n
(k
p
a)]
/
(5)
S
(M)
n
= ika j
n
(ka)[k
p
aj
n
(k
p
a)]
/
j
n
(k
p
a)[ka j
n
(ka)]
/
_
S
(N)
n
=
i
k
p
a
(k
p
a)
2
j
n
(k
p
a)[ka j
n
(ka)]
/
_
(ka)
2
j
n
(ka)[k
p
aj
n
(k
p
a)]
/
(6)
B
(M)
n
=
i
ka
1
j
n
(k
p
a)[kah
n
(ka)]
/
h
n
(ka)[k
p
aj
n
(k
p
a)]
/
B
(N)
n
=
1
(k
p
a)
2
j
n
(k
p
a)[ka h
n
(ka)]
/
(ka)
2
h
n
(ka)[k
p
aj
n
(k
p
a)]
/
(7)
The unknowns in (1) are exciting eld coefcients. We
can rewrite the equation for w
(M)
mn
and w
(N)
mn
as
w
(M)(q)
mn
=e
i
"
kk
i
.
" rrq
a
(M)
mn

N
p=1; pOq
mn
[A
mnmn
(kr
q
r
p
)
T
(M)
n
w
(M)(p)
mn
B
mnmn
(kr
q
r
p
)T
(N)
n
w
(N)(p)
mn
]
w
(N)(q)
mn
=e
i
"
kk
i
.
" rr
q
a
(N)
mn

N
p=1; pOq
mn
[B
mnmn
(kr
q
r
p
)
T
(M)
n
w
(M)(p)
mn
A
mnmn
(kr
q
r
p
)T
(N)
n
w
(N)(p)
mn
]
(8)
where a
(M)
mn
and a
(N)
mn
are the incident eld coefcients and
A
mnmn
and B
mnmn
are the spherical wave transformation
between two particles [2,6,9]. Equation (8) is the exciting
eld formulation of multiple scattering equations.
We can also derive multiple scattering equations for the
scattered and internal elds. Multiplying T-matrix coef-
cients T
(M)
n
or T
(N)
n
on both sides of (8) and applying the
relationship between exciting eld and scattered eld, we
then have the FoldyLax equations for scattered eld
coefcients:
a
S(M)(q)
mn
=e
i
"
kk
i
.
" rr
q
T
(M)
n
a
(M)
mn
T
(M)
n
N
p=1; pOq
mn
[A
mnmn
(kr
q
r
p
)a
S(M)(p)
mn
B
mnmn
(kr
q
r
p
)a
S(N)(p)
mn
]
a
S(N)(q)
mn
=e
i
"
kk
i
.
" rrq
T
(N)
n
a
(N)
mn
T
(N)
n
N
p=1; pOq
mn
[B
mnmn
(kr
q
r
p
)a
S(M)(p)
mn
A
mnmn
(kr
q
r
p
)a
S(N)(p)
mn
]
(9)
The unknowns of (9) are the scattered eld coefcients.
Thus, it is called the scattered eld formulation. Replacing
T
(M)
n
w
(M)
mn
and T
(N)
n
w
(N)
mn
with S
(M)
n
c
(M)
mn
and S
(N)
n
c
(N)
mn
, respec-
tively, in (9) and multiplying respective B
(M)
n
and B
(N)
n
on
both sides, we then have the internal eld formulation of
FoldyLax equations:
c
(M)(q)
mn
=e
i
"
kk
i
.
" rrq
B
(M)
n
a
(M)
mn
B
(M)
n
N
p=1; pOq
mn
[A
mnmn
(kr
q
r
p
) S
(M)
m
c
(M)(p)
mn
B
mnmn
(kr
q
r
p
) S
(N)
n
c
(N)(p)
mn
]
c
(N)(q)
mn
=e
i
"
kk
i
.
" rr
q
B
(N)
n
a
(N)
mn
B
(N)
n
N
p=1; pOq
mn
[B
mnmn
(kr
q
r
p
) S
(M)
m
c
(M)(p)
mn
A
mnmn
(kr
q
r
p
) S
(N)
n
c
(N)(p)
mn
]
(10)
We have studied the condition numbers of the various
types of FoldyLax formulations [8,9]. The scattered and
exciting eld formulations have similar condition num-
bers; however, the matrices based on scattered eld for-
mulations have much larger condition number than do
those based on internal eld formulations for the same set
of parameters. Thus, the internal eld formulation is
applied in the three-dimensional scattering simulations.
After solving the multiple scattering equations, the
scattered eld coefcients a
S(M)
mn
and a
S(N)
mn
can be calcu-
lated, and the scattered far-eld coefcients in the scatter-
ing direction
^
kk
s
= ^ xx sin y
s
cos f
s
^ yy sin y
s
sin f
s
^ zz cos y
s
are given by
"
EE
s
(y
s
; f
s
) =
e
ikr
kr
N
q =1
e
i
^
kk
s
.
" rr
q
mn
g
mn
[a
S(M)(q)
mn
"
CC
mn
(y
s
; f
s
)i
n1
a
S(N)(q)
mn
"
BB
mn
(y
s
; f
s
)i
n
] (11)
where
"
BB
mn
(y; f) and
"
CC
mn
(y; f) are vector spherical wave-
functions and g
mn
is a coefcient [2,6].
The scattered eld can be further decomposed into
coherent and incoherent parts. The scattering coefcient
arises from the incoherent elds because the coherent
wave propagates in the forward transmitted direction and
does not contribute to scattering or extinction. For Monte
Carlo simulation with N
r
realizations, the coherent
scattered eld
"
EE
s
) is calculated by averaging over realiza-
tions
"
E E
s
(y
s
; f
s
)) =
1
N
r
Nr
s =1
"
EE
s
s
(y
s
; f
s
) (12)
where
"
EE
s
s
is the scattered eld for realization s and
angular brackets represent averaging over realizations.
The incoherent scattered eld is given by
"
EE
s
s
(y
s
; f
s
) =
"
EE
s
s
(y
s
; f
s
)
"
EE
s
(y
s
; f
s
)) (13)
The bistatic scattering cross section is proportional to the
averaged incoherent intensity:
s
b
(
^
kk
s
;
^
kk
i
) =r
2
[
"
EE
s
s
(y
s
; f
s
)[
2
)
[
"
EE
inc
[
2
(14)
The scattering cross section is the result of integrating the
bistatic cross section over solid angles. The scattering
coefcient is the scattering cross section per unit volume
k
s
=
s
s
V
=
1
V
_
p
0
dy
s
siny
s
_
2p
0
df
s
s
b
(
^
kk
s
;
^
kk
i
) (15)
where V is the sample volume in the simulation. The
power absorbed by particle q can be written as [5]
W
(q)
a
=
1
2Zk
2
mn
[[w
(M)(q)
mn
[
2
(Re T
(M)
n
[T
(M)
n
[
2
)
[w
(N)(q)
mn
[
2
(Re T
(N)
n
[T
(N)
n
[
2
)]
(16)
The absorption coefcient is expressed as
k
a
=
1
V
N
q=1
W
(q)
a
W
inc
=
1
Vk
2
[
"
EE
inc
[
2
N
q =1
mn
[[w
(M)(q)
mn
[
2
(Re T
(M)
n
[T
(M)
n
[
2
) [w
(N)(q)
mn
[
2
(Re T
(N)
n
[T
(N)
n
[
2
)] (17)
2.3. Results of Using Different Number of Particles
In Monte Carlo simulations, it is important to demon-
strate convergence of the results with respect to the
numbers of particles used. Physically, this means that
the simulated extinction coefcients are extinction per
unit volume and should not be dependent on the number
of particles used. The number of particles used in the
simulations is chosen on the basis of frequencies, volume
fractions, and particle diameters. The particle positions
are rst generated within a unit cube for selected frac-
tional volume and number of particles. The positions are
then scaled according to the actual particle size. It has the
following relationship
f
v
=
N
part
V
4
3
pa
3
(18)
where f
v
is the fractional volume, N
part
the number of
particles within the simulated volume V, a the radius of
particle, and V=L
3
. In general, to ensure that a sufcient
amount of incoherent elds is generated as a result of
scattering in the sample volume, L has to be larger than
(or at least close to) one wavelength l. Table 1 lists the
results of extinction coefcients for ka=0.2, f
v
=20%, and
e
p
=(3:2 i0:01)e
0
with different numbers of particles. The
simulated box size L is also shown in the table in the unit
of wavelength. As illustrated in the table, results show
convergence with respect to the number of particles and
with respect to box size.
Monte Carlo simulations are performed for different
frequencies and physical parameters, including particle
size, permittivity, and stickiness. Results based on inde-
pendent scattering are also shown. The numerical para-
meters, such as N
max
, number of particles, and number of
realizations, are chosen such that results converge. The
maximum number of particles used in simulations is 2000,
and up to 50 realizations are used. The simulation results
for permittivities e
p
with different loss tangents, e
p
=
e
/
p
ie
//
p
, are shown in Fig. 1. Nonsticky particles (i.e.,
lacking interparticle forces) with diameter 1.2 mm occupy-
ing 20% of volume fraction are used in Fig. 1. Results of
scattering coefcient k
s
for e
p
=(3:2 i0:01)e
0
and e
p
=
(3:2 i0:02)e
0
are quite similar, as shown in Fig. 1a. The
scattering coefcient based on independent scattering
assumption is larger than that calculated based on QCA
or Monte Carlo (MC) simulations. The absorption rate k
a
varies because of the difference in the imaginary part of e
p
,
as indicated in Fig. 1b, where k
a
is larger for larger e
//
p
. MC
simulations predict higher absorption rate than do the
other methods. Generally, independent scattering gives
the largest extinction coefcient while QCA predicts the
smallest.
2.4. Frequency Dependence of Sticky and
Nonsticky Particles
Figure 2 shows the results for different stickiness para-
meter t. Particles with diameter 1.2mm occupying 20% of
volume with permittivity e
p
=3:2e
0
are used. The extinc-
tion coefcient and scattering coefcient are the same
because there is no absorption in this case. The smaller
the t is, the larger the interparticle forces are. Three
stickiness parameters are used in the simulations:
t =0:1; t =0:5, and nonsticky, which corresponds to
t =o. For t =0:1, however, the scattering coefcient is
larger than independent scattering at 18, 35, and 60 GHz.
It is smaller than independent scattering at 90 GHz. This
is because for small t, the particles strongly tend to adhere
Table 1. Scattering Coefcients and Box Sizes for
Different Number of Particles
Number of Particles L in l Units k
e
/k
10 0.189 6.75463910
5
100 0.407 1.60508610
4
1000 0.877 1.96329410
4
2000 1.105 1.99860810
4
and form clusters. The result is that the particles have
larger effective sizes. The effect of the property is exhibited
at low frequencies. At higher frequency, the clustering
effect of scattering declines and the particles scatter more
like the usual dense media. Therefore, the frequency
dependence decreases as t gets smaller. Results for
t =0:5 and non-sticky particles are about the same. This
means that the stickiness represents little difference in
scattering except for small stickiness parameter t. Table 2
lists the scattering coefcients for various stickiness para-
meters at selected frequencies. The scattering coefcient
at 18 GHz is used as reference. To study the frequency
dependence, we dene n such that
k
s
(at f
2
)
k
s
(at f
1
)
=
f
2
f
1
_ _
n
(19)
where f
1
and f
2
are the two frequencies and k
s
is the
scattering coefcient at a specic frequency and n is a
power law to be determined. On the basis of the simulated
10
2
10
0
10
2
10
1
20 40 60 80 100
(a)
Frequency (GHz)
s

(
c
m
1
)
10
1
10
2
10
3
10
4
20 40 60 80 100
(b)
Frequency (GHz)
s

(
c
m
1
)
20 40 60 80 100
(d)
Frequency (GHz)
10
2
10
0
10
2
10
4
20 40 60 80 100
(c)
Frequency (GHz)
s

(
c
m
1
)
1
0.8
0.6
0.4
0.2
0
independent - e
r
-32 + r0.01
QCA- e
r
- 3.2 + r0.01
Mode cario - e
r
- 3.2+ r 0.01
Independent - e
r
-3.2+r0.002
QCA -er + 3.2+ r.002
Mode cario -e
r
- 3.2+r0.002
A
l
b
e
d
o
Figure 1. Results of different particle
permittivities: e
p
=(3:2 i0:01)e
0
and e
p
=
(3:2 i0:02)e
0
particles occupy 20% of
volume fraction with diameter 1.2mm:
(a) scattering coefcient; (b) absorption
rate; (c) extinction coefcient; (d) albedo
( ~ oo) as a function of frequency. (This gure
is available in full color at http://www.
mrw.interscience.wiley.com/erfme.)
10 20 30 40 50 60 70 80 90 100
10
1
10
0
10
1
10
2
10
3
10
4
e

(
c
m
1
)
Frequency (GHz)
Indepemdemt scattering
QCA - Nonsticky
QCA - sticky = 0.5
Monte Carlo - Nonsticky
Monte Carlo - sticky = 0.5
Monte Carlo - sticky = 0.1
QCA - Sticky =0.1
Figure 2. Extinction coefcient in reciprocal centimeters (cm
1
)
as a function of frequency in gigahertz (GHz) of different sticki-
ness parameter t =0:1; t =0:2, and N. Particles occupy 20% of
volume fraction with diameter 1.2mm and e
p
=3:2e
0
. (This gure
is available in full color at http://www.mrw.interscience.wiley.
com/erfme.)
Table 2. Scattering Coefcients in (dB/m) for Various Stickiness Parameters at Selected Frequencies
Monte Carlo QCA
Frequency (GHz) t =0.1 t =0.5 t =0.1 t =0.5 Independent Scattering
18 12.051 0.738 9.450 0.571 1.376
35 82.761 13.203 63.856 8.172 20.359
60 326.643 116.167 297.824 72.251 185.546
90 707.985 443.461 957.044 354.834 862.510
results, the frequency dependence n is calculated by
n=
log
k
s
(at f
2
)
k
s
(at f
1
)
_ _
log
f
2
f
1
_ _ (20)
The results for f
1
=18 GHz are shown in Table 3. The
frequency dependence n for stickiness t =0:1 is less than
that of nonsticky particles and t =0:5. For nonsticky
particles, dense media scattering is less than independent
scattering. This reduction of scattering for dense media
becomes a smaller effect at higher frequency. Thus results
in Table 3 can exhibit index larger than 4. Note that a
least-square t with snow extinction measurements gives
an index n between 2 and 3 [22]. The index is consistent
with the simulated index n of sticky particles.
Figure 3 shows the results for nonsticky particle with
different sizes. Particles occupying 20% of volume with
permittivity e
p
=3:2e
0
are used. The extinction coefcients
versus frequency are plotted. We note that scattering
increases with particle size, and the scattering by Monte
Carlo simulation is less than independent scattering.
However, the difference gets smaller at high frequencies.
Figure 4 shows results for a sticky particle with t =0:1,
and the diameters of particles are 1, 1.2, and 1.5mm.
These represent typical ice grain sizes in snow. Note that
scattering increases with particle size. For sticky particles,
the simulation results are larger than independent scat-
tering at low frequencies and are smaller than indepen-
dent scattering at high frequencies. At 90GHz, the
extinction is large. Thus at such high frequency, passive
remote sensing will not be useful to determine snow depth.
We next compare Monte Carlo simulation results with
that of QCA. For nonsticky particles, the extinction coef-
cient calculated from QCA is smaller than that from MC at
low frequencies while the extinction coefcients of MC and
QCA are comparable at higher frequency. For sticky
particles, larger particles generally have higher extinction
coefcients. We note that the QCA results are generally
much closer to simulations than independent scattering.
The QCA formulation has been used in dense media
formulation and applied to passive remote sensing of
spaceborne sensors [7,26].
In the results described above, we have not taken into
account the dependence of loss tangent on frequency
because the computed results of extinction are dominated
by scattering rather than absorption. However, in com-
parison with experimental data in Fig. 5, we include the
dependence of loss tangent on frequency. The permittivity
of ice as a function of frequency can be found in Ref. 5. In
Fig. 5 we compare Monte Carlo simulation, independent
scattering, and snow extinction measurement data at 18,
35, 60, and 90 GHz [22]. We use ground truth mean grain
diameter of 1mm and snow density of 0.315 g/cm
3
[22].
Table 3. Frequency Dependence n Based on Scattering Coefcients from Table 2
Monte Carlo QCA
Frequency (GHz) t =0.1 t =0.5 t =0.1 t =0.5 Independent Scattering
18
35 2.898 4.338 2.873 4.001 4.038
60 2.741 4.202 2.866 4.020 4.063
90 2.531 3.976 2.869 3.996 4.051
10 20 30 40 50 60 70 80 90 100
Frequency (GHz)
e

(
c
m
1
)
QCA - diameter = 1.0 mm
Monte Carlo - diameter = 1.0 mm
10
1
10
0
10
1
10
2
10
3
10
4
Figure 3. Extinction coefcient in cm
1
of different particle sizes
as a function of frequency in GHz for nonsticky particles. Particles
occupy 20% of volume with permittivity e
p
=3:2e
0
. (This gure is
available in full color at http://www.mrw.interscience.wiley.com/
erfme.)
10 20 30 40 50 60 70 80 90 100
Frequency (GHz)
e

(
c
m
1
)
10
1
10
0
10
1
10
2
10
3
Figure 4. Extinction coefcient in cm
1
of different particle sizes
as a function of frequency in GHz for sticky particles (t =0.1).
Particles occupy 20% of volume with permittivity e
p
=3:2e
0
. (This
gure is available in full color at http://www.mrw.interscience.
wiley.com/erfme.)
The simulation is based on a volume fraction of 0.4134.
The permittivity of ice grains chosen is e
p
=3:2e
0
ie
//
s
. The
loss tangent depends on frequency and e
//
s
=e
0
=0:007,
0.009, 0.011, and 0.014, respectively, at 18, 35, 60, and
90 GHz, which correspond to a salinity of 0.12 parts per
thousand [44]. The frequency dependence as exhibited by
Monte Carlo simulation is in reasonable agreement with
experiment and is much better than that of independent
scattering.
In this section we present the Monte Carlo simulations
performed for various frequencies, particle sizes, particle
permittivities, and cases with different stickiness para-
meters t. Monte Carlo simulation results compare well
with QCA results. Scattering based on the sticky particle
assumption has a much weaker frequency dependence
than does that of independent scattering of frequency to
the fourth power. This frequency dependence has been
shown to better match the observed frequency dependence
of snow terrain [7,26]. For problems of very complicated
geometry, Monte Carlo simulations provide the correct
solutions for Maxwell equations. The results of simula-
tions can be tabulated and further combined with dense
media radiative transfer theory. The dense media radia-
tive transfer theory has been used extensively for map-
ping of snow depth using spaceborne radiometer.
3. POLARIMETRIC PASSIVE MICROWAVE REMOTE
SENSING OF WIND VECTORS WITH FOAM-COVERED
OCEAN SURFACE
Fully polarimetric microwave remote sensing means that
all four Stokes parameters are measured [45]. There has
been an increasing interest in the applications of polari-
metric microwave radiometers for ocean wind remote
sensing [4649]. It is known that the presence of foam on
the ocean surface has a profound effect on the brightness
temperatures measured by microwave radiometers.
Although foam typically covers only a few percent of sea
surfaces, increasing foam coverage on the sea surface can
substantially increase the sea surface emissivity [5052].
However, there is a great concern on the impact of foam on
the retrieval of the ocean surface wind vector from satel-
lite-mounted microwave instruments, which is due in
large part to the difculty in making measurements at
high-wind conditions when signicant foam coverage is
present. In the past, empirical microwave emissivity
models [48,50,51,5355] were used to estimate the effect
of the foam above the ocean surface on the passive
microwave remote sensing measurements. These are em-
pirical tting procedures using experimental data and the
empirical models do not take into account the physical
microstructure of foam and the foam-layer thickness. The
subject of foam dynamics has also attracted great atten-
tion. Huang and Jin discussed a composite model of foam
scatterers and two-scale wind-driven rough sea surface
[52]. Controlled eld experiments were performed to
measure foam dynamics and the microwave emissivity of
calm seawater [56,57].
More recently, physically based approaches, based on
the quasicrystalline approximation (QCA) dense media
model or Monte Carlo simulations, have been developed
to account for the microstructure of foam [1012]. The
model treats the foam as densely packed air bubbles
coated with thin seawater coating. In order to model
high-density packing, a face-centered-cubic (fcc) structure
is used to place the air bubbles. It was shown that the
polarization and frequency of the brightness temperatures
depend on the physical microstructure properties of foam
and the foam-layer thickness. In this section, we present
the polarimetric microwave emission model for foam-
covered ocean surfaces. The dense media model of foam
is applied to calculate the values of the complex effective
propagation constants, the extinction coefcients, and the
albedo. These are used to describe the characteristics of
the foam layer. In Section 3.1, we describe the physical
and geometric model of foam. The Monte Carlo simulation
by solving Maxwells equations and DMRT theory are
summarized in Sections 3.2 and 3.3, respectively. In the
Monte Carlo simulations, the volume integral equation
has been used. The absorption, scattering, and extinction
coefcients are calculated, and the simulation results for
emissivity with typical foam parameters at 10.8 and
36.5 GHz are presented in Section 3.4. Comparisons are
also made with experimental data [57] for vertical and
horizontal polarizations. The Monte Carlo simulation
results with those based on QCA are compared in Section
3.5. The effects of boundary roughness of ocean surface
can be included in the boundary conditions of dense media
radiative transfer (DMRT) theory by using the second-
order small perturbation method (SPM) [2,7,46]. A fully
polarimetric passive model for wind-generated and foam-
covered rough sea surfaces, using the empirical Durden
Vesecky spectrum [58], is presented in Ref. 12.
3.1. Description of Foam Model
Figure 6 shows a video micrograph of the bubble structure
of articially generated foam on the surface of Chesapeake
Bay [57]. Analysis of this and similar images shows that
10 20 30 40 50 60 70 80 90 100
Frequency (GHz)
E
x
t
i
n
c
t
i
o
n

(
d
B
/
m
)
10
3
10
2
10
1
10
0
Monte Carlo simulation
Independent
Snow extention experiment
Figure 5. Comparison of Monte Carlo simulation results with
experimental data and independent scattering. (This gure is avail-
able in full color at http://www.mrw.interscience.wiley.com/erfme.)
the void fraction of this foam is 8090% in most cases. To
simplify the foam model, the foam is assumed to be
composed of densely packed spherical air bubbles with
thin seawater coating [1012]. The physical and geometric
structure of a spherical dielectric-coated particle is illu-
strated in Fig. 7, where s
j
denotes that the coated particle
is of the jth species, s
j
=1; 2; . . . ; L. The core of the coated
particle is air, and the shell is seawater. It is assumed that
all the particles have the same outer radius a
s
j
, but need
not be identical in coating thicknesses (a
s
j
b
s
j
). e
a
sj
and
e
b
sj
are permittivities within the shell and in the core of the
coated particle, respectively.
To achieve high-density packing, the coated particles
are arranged in a FCC lattice structure that has a packing
density of p=
18
_
- 74% [11,12]. The air regions include
the core regions of the coated particles and the interstitial
space between the coated particles. For example, let V be
the total volume of the foam and N be the number of
coated particles. The jth coated particles has a inner
radius b
j
and outer radius a
j
. The fractional volume of
coated particles is about 74%; the fractional volume of
seawater f
w
is then
f
w
=
1
V
N
j =1
4p
3
(a
3
j
b
3
j
) (21)
By choosing values for the inner and outer radii, the
foam void fraction (1.0 f
w
) can be on the order of 90%,
which is in agreement with experimental measurements
of articially generated foam [57].
3.2. Absorption and Extinction Based on Monte Carlo
Simulation by Solving Maxwells Equations
Consider the thermal emission from a layered medium
with coated particles embedded in a background medium
of air, as indicated in Fig. 8. The layer consists of coated
particles (region 1), and covers a half-space of ocean
(region 2). The scattering coefcient and absorption coef-
cient of a collection of N coated particles are dened
respectively as scattering cross section per unit volume
and absorption cross section per unit volume. In the
Monte Carlo simulation approach, we place N particles
in volume V, and calculate the absorption and scattering of
these N particles collectively by solving Maxwells equa-
tions and then averaging over realizations and dividing
them by the volume V. These are done for large N and the
results of the scattering and absorption coefcients com-
puted in this manner converge for large N.
A volume integral equation derived from Maxwells
equations is used to solve scattering and absorption for
the N particles [6,11,12]. The volume integral equation
has the internal electric eld in the coated region as the
unknown. The internal eld
"
EE
i
( " rr) in the seawater coating
Figure 6. Videomicrograph of articially generated bubbles in
foamon the surface of Chesapeake Bay. The scale in the upper left
corner shows a distance of 1 mm.
b
sj
a
sj
c
bsj
c
asj
Figure 7. Spherical dielectric-coated particle; a
sj
and b
sj
are the
outer and inner radii of the spherical shells, respectively, and e
asj
and e
bsj
are permittivities within the shell and in the core of the
coated particle, respectively.
Radiometer
Region 0
z=0
Region 1
z=d
Region 2 Ocean
Foam
Figure 8. Geometric conguration for thermal emission from
foam-covered ocean. The foam layer is region 1 and is absorptive
and scattering. Region 2 is the ocean.
region of particle i, " rr c V
i
, is given by
"
EE
i
( " rr) =
"
E E
e
( " rr) k
2

N
j =1
_
V
j
d" rr
/
g( " rr; " rr
/
)(e
rj
1)
"
EE
j
( " rr
/
)
N
j =1
V
_
V
j
d" rr
/
V
/
g( " rr; " rr
/
)(e
rj
1)
"
EE
j
( " rr
/
)
(22)
where
"
EE
e
( " rr) is the incident eld, g( " rr; " rr
/
) the free space
scalar Greens function, e
rj
the relative permittivity of
particle within the coating, V
j
the volume occupied by
particle j, and i =1; 2; . . . ; N. We expand the internal eld
in the coating region of particle j,
"
EE
j
( " rr), into three basis
functions. The basis functions are from the electrostatic
solutions of the coating region of a coated sphere [11]. The
subscript j is suppressed in Eqs. (23) and (24)
"
EE
j
( " rr) =
3
a =1
c
a
( " rr)
"
ff
a
( " rr) (23)
"
ff
1
( " rr) =
1 2e
r
3e
r
^ xx
1 e
r
3e
r
b
3
^ rr
2
r
3
sin y cos f
_

1
r
3
(
^
yy cos y cos f
^
f f sin f)
_
"
ff
2
( " rr) =
1 2e
r
3e
r
^ yy
1 e
r
3e
r
b
3
^ rr
2
r
3
sin y sin f
_

1
r
3
(
^
yy cos y sin f
^
ff cos f)
_
"
ff
3
( " rr) =
1 2e
r
3e
r
^ zz
1 e
r
3e
r
b
3
^ rr
2
r
3
cos f
^
yy
1
r
3
sin y
_ _
(24)
where " rr c V
j
. Letting
" qq
ja
( " rr) =k
2
_
V
j
d" rr
/
g( " rr; " rr
/
)(e
rj
1)
"
ff
ja
( " rr
/
)
V
_
V
j
d" rr
/
V
/
g( " rr; " rr
/
)(e
rj
1)
"
ff
ja
( " rr
/
)
(25)
we can express the integral equation (22) as
3
a =1
c
ia
"
ff
ia
( " rr) =
"
EE
ex
( " rr)
N
j =1; jOi
3
a =1
c
ja
" qq
ja
( " rr)
3
a =1
c
ia
" qq
ia
( " rr)
(26)
We then apply the Galerkin method to rewrite (26) in a
linear system of equations for the coefcients
3
a =1
c
ia
_
V
i
d" rr
"
ff
ib
( " rr)
.
[
"
ff
ia
( " rr) " qq
ia
( " rr)]
=
_
V
i
d" rr
"
ff
ib
( " rr)
. "
EE
e
( " rr
i
)
N
j =1; jOi
3
a =1
c
ja
_
V
i
d" rr
"
ff
ib
( " rr)
.
" qq
ja
( " rr)
(27)
with i =1; 2; . . . ; N and b=1; 2; 3. Taking the small particle
assumption into account, we can make approximations
of (27):
3
a =1
c
ia
_
V
i
d" rr
"
ff
ib
( " rr)
. "
ff
ia
( " rr) " qq
ia
( " rr)
_ _
=
_
V
i
d" rr
"
ff
ib
( " rr)
. "
EE
e
( " rr
i
)
N
j =1; jOi
3
a =1
k
2
(e
rj
1) c
ja
_
V
i
d" rr
"
ff
ib
( " rr)
.
G( " rr
i
; " rr
j
)
_
V
j
d" rr
/
"
ff
ja
( " rr
/
)
(28)
After simplications, (28) can be written as
c
ib
K
i
= " ss
ib
. "
EE
e
( " rr
i
)
N
j =1; jOi
3
a =1
k
2
(e
rj
1)
c
ja
" ss
ib
.
G( " rr
i
; " rr
j
)
.
" ss
ja
(29)
where
K
i
=
2e
ri
1
3e
ri
.
1
9e
ri
.
4p
3
(a
3
i
b
3
i
)
(2e
ri
1)(2e
ri
) 2b
3
i
(e
ri
1)
2
a
3
i
(30)
and
" ss
i1
=
_
V
i
d" rr
"
ff
i1
( " rr) =
2e
ri
1
3e
ri
.
4p
3
(a
3
i
b
3
i
) ^ xx
" ss
i2
=
_
V
i
d" rr
"
ff
i2
( " rr) =
2e
ri
1
3e
ri
.
4p
3
(a
3
i
b
3
i
) ^ yy
" ss
i3
=
_
V
i
d" rr
"
ff
i3
( " rr) =
2e
ri
1
3e
ri
.
4p
3
(a
3
i
b
3
i
) ^ zz
(31)
After solving the matrix equations, the power absorbed by
N coated particles can be calculated by
P
abs
=
1
2
o
N
j =1
_
V
j
d" rre
//
rj
( " rr)[
"
EE
j
( " rr)[
2
(32)
and the absorption coefcient is the absorption per unit
volume
k
abs
=
P
abs
1
2Z
[
"
EE
e
[
2
V
(33)
where Z is the free-space wave impedance. The scattered
far eld can be calculated as following
"
E E
s
( " rr) =
k
2
e
ikr
4pr
( ^ vv
s
^ vv
s

^
hh
s
^
hh
s
)
N
j =1
_
V
j
d" rr
(e
rj
1)
"
EE
j
( " rr) e
i
"
kks
.
" rr
j
(34)
Then the scattered power is
P
s
=
1
2Z
_
2p
0
df
s
_
p
0
dy
s
sin y
s
([E
vs
[
2
[E
hs
[
2
) (35)
where E
vs
and E
hs
are vertical and horizontal polarized
components of the scattering elds, respectively. The
scattering coefcient is scattering cross section per unit
volume
k
s
=
P
inco
s
1
2Z
[
"
EE
ex
[
2
V
(36)
where P
inco
s
is the incoherent part of the scattered power.
The coherent wave is obtained by averaging the scat-
tered eld over the Monte Carlo realizations [6,11]. To
obtain the incoherent power, we have to subtract the
coherent intensity. Thus
P
inco
s
=
1
2Z
_
2p
0
df
s
_
p
0
dy
s
sin y
s
{[E
vs
E
vs
)[
2
) [E
hs
E
hs
)[
2
)]
(37)
where the angular brackets represent averaging over
realizations. The extinction coefcient is k
e
=k
s
k
abs
,
and the albedo is ~ oo=k
s
=k
e
. The effective permittivity
can be calculated as follows. In the forward direction
"
kk
s
=
"
kk
i
, the scattered eld in the incident polarization
can be written as E
s
=Fe
ikr
=r. Then the effective propaga-
tion constant K is given by
K =Re
k
2
4pF
V
_
_ _
i
k
e
2
_ _
(38)
The effective permittivity is expressed as
e
eff
=
K
2
k
2
(39)
The effective permittivity of foam will be used in the
Fresnel reection coefcients for the airfoam at inter-
face and the foamocean at interface in the dense media
radiative transfer theory. The calculated scattering and
absorption coefcients are then substituted into dense
media radiative transfer equations. A Rayleigh phase
matrix is used in the dense media radiative transfer
equations [2,7].
3.3. Dense Media Radiative Theory
Consider the thermal emission problem of a foam layer
covered wind-roughened ocean surfaces, as indicated in
Fig. 8. The foam layer consists of coated dielectric particles
in region 1. The DMRT equations for passive remote
sensing in region 1 can be written as follows
d
dz
"
II
u
(z; m; f) = k
es
"
II
u
(z; m; f)
k
as
C
"
TT
"
FF
u
(z; m; f)
d
dz
"
II
d
(z; m; f) = k
es
"
II
d
(z; m; f)
k
as
C
"
TT
"
FF
d
(z; m; f)
(40)
"
FF
u
(z; m; f) =
k
ss
4p
_
1
0
dm
/
_
2p
0
df
/
[P(m; f; m
/
; f
/
)
"
II
u
(z; m
/
; f
/
)
P(m; f; m
/
; f
/
)
"
II
d
(z; m
/
; f
/
)]
"
FF
d
(z; m; f) =
k
ss
4p
_
1
0
dm
/
_
2p
0
df
/
[P(m; f; m
/
; f
/
)
"
II
u
(z; m
/
; f
/
)
P(m; f; m
/
; f
/
)
"
II
d
(z; m
/
; f
/
)]
(41)
where
"
II
u
(z; m; f) and
"
II
d
(z; m; f) represent, respectively,
upward- and downward-going specic intensities and 41
matrices containing the four Stokes parameters; m = cos y,
m
/
= cos y
/
, k
es
=k
e
= cos y, k
ss
=k
s
= cos y; k
a
and k
s
are the
absorption and scattering coefcient matrices taken to be
diagonal respectively; C=K
b
K
/2
=(l
2
k
2
), where K
b
is the
Boltzmanns constant [2,5]; P(m; f; m
/
; f
/
) is the Rayleigh
scattering phase matrix; and
"
TT is the temperature prole
in the layer. The boundary condition at z =0 is
"
II
d
(z =0; y; f) =R(y)
"
II
u
(z =0; y; f) (42)
where R(y) is a reection matrix of the at surface at the
airfoam interface
R(y)
=
[R
v
(y)[
2
0 0 0
0 [R
h
(y)[
2
0 0
0 0 Re(R
v
(y)R
+
h
(y)) Im(R
v
(y)R
+
h
(y))
0 0 Im(R
v
(y)R
+
h
(y)) Re(R
v
(y)R
+
h
(y))
_
_
_
_
(43)
in which R
v
and R
h
are the Fresnel reectivities with
effective propagation constant K and effective relative
permittivity e
eff
for dense media.
The rough surface boundary condition at z = d is
bistatic and is determined by the second-order SPM
"
II
u
(z = d; m; f) =G
c
(y; f)
"
II
d
(z = d; m; f)
K
/2
m
_
1
0
dm
/
_
2p
0
df
/
m
/2
W(
"
kk
J

"
kk
/
J
)
G
ic
(m; f;m
/
; f
/
)
"
II
d
(z =d; m
/
; f
/
) " eeCT
g
(44)
where G
c
and G
ic
respectively are the coherent and
incoherent scattering phase matrixes of the rough surface
at z = d.
G
c;ic
=
[f
vv
[
2
) [f
vh
[
2
) Ref
vv
f
+
vh
) Imf
vv
f
+
vh
)
[f
hv
[
2
) [f
hh
[
2
) Ref
hv
f
+
hh
) Imf
hv
f
+
hh
)
2 Ref
vv
f
+
hv
) 2 Ref
vh
f
+
hh
) Ref
vv
f
+
hh
f
vh
f
+
hv
) Imf
vh
f
+
hv
f
vv
f
+
hh
)
2Imf
vv
f
+
hv
) 2 Imf
vh
f
+
hh
) Imf
vv
f
+
hh
f
vh
f
+
hv
) Ref
vv
f
+
hh
f
vh
f
+
hv
)
_
_
_
_
(45)
For G
c
, f
aa
=f
(0)
aa
f
(2)
aa
, f
ab
=f
(2)
ab
, and for G
ic
, f
ab
=f
(1)
ab
. Here,
symbols a and b represent vertical and horizontal. f
(0)
vv
and
f
(0)
hh
are Fresnel reection coefcients for vertical and
horizontal polarizations with the zeroth-order elds con-
sidered. f
(1)
ab
and f
(2)
ab
are scattering coefcients of the
a-polarized component of the rst- and second-order scat-
tered elds with b-polarized incident eld, respectively,
which are given in Ref. 7. The zeroth- and second-order
elds give the coherent reection coefcients of the sur-
faces. f
(1)
ab
gives the incoherent polarimetric bistatic scat-
tering coefcient due to the rst-order scattered eld.
They are derived using the second-order small-perturba-
tion method (SPM) [2,7]. T
g
is the temperature of the half-
space below the foam and " ee are the emissivities of the
lower boundary. The differential equations in (40) have
standard solutions of the form
"
II
u
(z) =
"
I I
u
(d)e
k
es
(z d)

_
z
d
dz
/
[
"
FF
u
(z
/
)
k
as
C
"
TT ] e
kes(zz
/
)
(46)
"
II
d
(z) =
"
II
d
(0)e
kesz
_
0
z
dz
/
[
"
FF
d
(z
/
)
k
as
C
"
TT] e
k
es
(zz
/
)
(47)
Next, we incorporate the rough surface boundary condi-
tions into a form suitable for iterative solutions. Substi-
tuting the boundary conditions into (46) and (47), we
have
"
II
u
(z) =e
kes(z d)
G
c
(m; f)
"
II
d
(z = d; m; f)
e
k
es
(z d)
K
/2
m
_
1
0
dm
/
_
2p
0
df
/
m
/
2W(
"
kk
J

"
kk
/
J
)
G
ic
(m; f; m
/
; f
/
)
"
II
d
(z = d; m
/
; f
/
)
" eeCT
g
e
kes(z d)

_
z
d
dz
/
[
"
FF
u
(z
/
) k
as
C
"
TT] e
kes(zz
/
)
(48)
"
II
d
(z) =
"
"
RR
"
RR(y)
"
II
u
(z =0; y; f) e
kesz
_
0
z
dz
/
[
"
FF
d
(z
/
) k
as
C
"
TT] e
k
es
(zz
/
)
(49)
The source terms in (48) are the upward temperature
originating from the layer temperature prole. We assume
a temperature prole T=T
g
_
z
d
dz
/
k
as
CTe
k
es
(zz
/
)
=C
k
a
k
e
T
g
[1 e
k
es
(z d)
] (50)
The contribution from the lower half-space is
" eeCT
g
exp[k
es
(z d)]. The source term in (49) is the down-
ward temperature:
_
0
z
dz
/
k
as
CT e
k
es
(zz
/
)
=C
k
a
k
e
T
g
[1 e
k
es
z
] (51)
All other terms in (48) that depend on the upward and
downward temperatures can be evaluated using these
three source terms. Using (41), (50), and (51), and after
taking integration over z, we have
"
VV(z =0) =
_
0
d
dz
/
"
FF
u
(z
/
) e
kesz
/
-
1
4p
k
s
k
a
k
e
k
e
_
1
0
dm
/
_
2p
0
df
/
P(m; f; m
/
; f
/
)
c
0
"
TT " eeCT
g
k
e
k
a
"
TT
_ _
c
1
e
k
e
d=m
/
_ _
1
4p
k
s
k
a
k
e
k
e
_
1
0
dm
/
_
2p
0
df
/
P(m; f;
m
/
; f
/
) c
0
"
TT
"
TT(c
0
c
2
)
(52)
where c
0
=1 exp(k
es
d), c
1
={1 exp[k
es
d(1 m=m
/
)]]=
(1 m=m
/
), and c
2
= 1 exp[k
es
d(1m=m
/
)]
_ _
=(1m=m
/
).
Another term in (48) accounting for upward scattering of
the downward temperature by the lower boundary at
z = d can be evaluated using the source term given by
(51) as
SG(z =0) =G
c
(m; f)
"
II
d
(z = d; y; f) e
k
es
d
K
/ 2
m
_
1
0
dm
/
_
2p
0
df
/
m
/2
W(
"
kk
J

"
k
/
k
/
J
) G
ic
(m; f;
m
/
; f
/
)
"
II
d
(z = d; m
/
; f
/
)e
k
es
d
- C
k
a
k
e
G
c
(m; f)
"
TT(1 e
kesd
)e
kesd
C
k
a
k
e
K
/ 2
m
_
1
0
dm
/
_
2p
0
df
/
m
/ 2
W(
"
kk
J

"
k
/
k
/
J
)
G
ic
(m; f; m
/
; f
/
)
"
TT(1 e
kesd=m
/
)e
kesd
(53)
Thus the complete expression for the upward temperature
at z =0 within the layer is
"
II
(1)
u
(z =0) =C
k
a
k
e
"
TT(1 e
k
es
d
) " eeCT
g
e
k
es
d
"
VV(z =0) SG(z =0)
(54)
This is the rst-order solution of
"
II
u
. The second order
solution for the upward temperature at z =0 is
"
II
(2)
u
(z =0) =C
k
a
k
e
"
TT(1 e
k
es
d
) " eeCT
g
e
k
es
d
"
VV(z =0) SG
(2)
(z =0)
(55)
where
SG
(2)
(z =0) =G
c
(m; f)
"
II
d
(z = d; y; f) e
kesd
K
/ 2
m
_
1
0
dm
/
_
2p
0
df
/
m
/ 2
W(
"
kk
J

"
kk
/
J
)
G
ic
(m; f; m
/
; f
/
)
"
II
d
(z = d; m
/
; f
/
)e
k
es
d
(56)
The brightness temperatures
"
TT
B
in the direction (y
0
; f
0
),
where y
0
= sin
1
(K
/
sin y=k) as related to y by Snells law,
are given by
"
TT
B
(y
0
; f
0
) =T(y)
"
II(z =0; y; f) (57)
T(y) =
1 [R
v
(y)[
2
0 0 0
0 1 [R
h
(y)[
2
0 0
0 0
cos y
0
cos y
Re(T
v
(y)T
+
h
(y))
cos y
0
cos y
Im(T
v
(y)T
+
h
(y))
0 0
cos y
0
cos y
Im(T
v
(y)T
+
h
(y))
cos y
0
cos y
Re(T
v
(y)T
+
h
(y))
_
_
_
_
(58)
where T
v
(y) =1R
v
(y), and T
h
(y) =1 R
h
(y).
3.4. Numerical Simulations of Emissivities for Foam-Covered
Flat Ocean Surfaces
In this section, we illustrate the numerical results of the
emissivity, based on a model of coated particles in a FCC
structure, for a foam-covered at ocean surface and com-
pare with these results some experimental measurements.
The absorption rate, scattering rate, and effective permit-
tivity are rst calculated using the Monte Carlo simula-
tion. Subsequently, these quantities are used to compute
the emissivity. The foam parameters are summarized as
follows: a
j
and b
j
, the respective outer and inner radii of
air bubble j; N, number of air bubbles; V, total sample
volume in the Monte Carlo simulations; f
w
, the volume
fraction of seawater in foam; e
w
, the permittivity of
seawater; and y, the observation angle. Note that the
permittivity of seawater e
w
is a function of frequency and
other physical parameters, such as the temperature and
salinity. In the following simulations, the permittivities of
seawater at 10.8 and 36.5 GHz are 49.149 i40.105 and
13.448 i24.784, respectively [60]. The experimental mea-
surements are conducted at the Chesapeake Bay Detach-
ment at 10.8 and 36.5 GHz. The diameter of air bubble
ranges from 500 to 5000 mm with median between 900 and
1000 mm [57].
In the Monte Carlo simulations, different realizations
of the sample of spheres are obtained by rotations of the
sample volume. From the results of the different realiza-
tions, the coherent elds and the incoherent elds are
calculated. The parameters used for Monte Carlo simula-
tions are summarized in Table 4. The total number of
coated air bubbles, which are arranged in a face-centered-
cubic structure, is N=N
/
N
//
=500, where two species of
coated air bubbles are used. They have the same outer
radii but have different inner radii, b
/
and b
//
. We choose N
/
bubbles randomly of inner radii b
/
, and the rest have inner
radii of b
//
. Seven realizations are generated by rotations
of the sample. The respective absorption rate, scattering
rate, extinction rate, albedo, and effective permittivity
calculated from Monte Carlo simulations are given in
Tables 57. Figures 9 and 10 plot the microwave emissiv-
ity dependence on the observation angle at 10.8 and
36.5 GHz for vertical and horizontal polarizations, respec-
tively. The foam parameters are listed in Table 4. As the
size of the bubbles increases, the scattering coefcient
increases, and the albedo also increases. The increase in
albedo causes the corresponding brightness temperatures
to decrease.
The microwave emissivities, at 10.8 and 36.5GHz and
for vertical and horizontal polarizations, as a function of
the thickness of foam layer with different sizes of coated
air bubbles are plotted in Figs. 1113. In actual foam, the
air bubbles have a size distribution with mean diameter
about 1mm. However, scattering increases with particle
sizes and the effective scattering mean size can be sub-
stantially larger than the mean size. In our simulations,
we use single size particles. Three radii of air bubbles, 1.0,
0.5, and 0.25 mm, are used and represent the effective
scattering mean. In actual foam, the coating thicknesses
vary and also vary as a function of depth. For convenience,
we have used two coating thicknesses. The observation
angle is y =531. From these gures, we can see that as the
foam layer thickness increases, the emissivity increases
correspondingly and then saturates at a particular thick-
ness of the foam layer, for both horizontal and vertical
polarizations. The saturation point of horizontal polariza-
tion is slightly larger than that of vertical polarization. For
layer thickness larger than the saturation thickness, the
difference of emissivity between the two frequencies in-
creases as the size of coated air bubbles increases.
In Figs. 14 and 15, the simulation results of microwave
emissivities are compared with the experimental data as a
function of observation angle. An average foam layer
thickness 2.8cm is used. The parameters used for Monte
Carlo simulation are the same as those used in Figs. 911,
with a coated air bubble radius of 1.0 mm. In the experi-
ment, the emissivities of horizontal and vertical polariza-
tions are measured, at 10.8 and 36.5GHz, for a foam layer
with a mean thickness of 2.8cm. To facilitate the compar-
isons, we list in Table 8, the experimental data, the DMRT
model results, and the airocean half-space results for the
emissivities at 10.8 and 36.5 GHz. From Figs. 14 and 15,
we see that Monte Carlo simulations produce results in
reasonably good agreement with experimental measure-
ments. Both simulations and experiments indicate that
absorption at 10.8GHz is appreciable. In addition, the
simulations show that emissivities at 10.8 and 36.5 GHz
are comparable. The absorption coefcient at 36.5 GHz is
larger than that at 10.8GHz. However, scattering has a
signicant effect at 36.5 GHz. The results are in good
agreement at small angles of incidence. At large incidence
angles, the difference increases. We are presently studying
the renement of the model by investigating realistic foam
generation algorithms that can improve the model.
3.5. Comparison with Quasicrystalline Approximation
In this section, we compare the Monte Carlo results with
those based on QCA, in which wave scattering and emis-
sion in a medium consisting of densely packed coated
particles are solved by using QCA in combination with
the DMRT theory [10]. It is assumed that there are two
species of air bubbles that have the same outer radius a,
but have different coating thickness with inner radii of b
1
and b
2
. The parameters of the foam model are given in
Table 9. Emissivities of horizontal and vertical polariza-
tions are simulated at 10.8 and 36.5 GHz.
The quantities for foam layer calculated by QCA and by
Monte Carlo simulations are shown in Tables 10 and 11,
respectively. Figure 16 is based on Monte Carlo simula-
tions. We can see that the emissivities at 10.8 and
36.5 GHz are comparable. This feature is consistent with
experimental measurements [57]. On the other hand, as
shown in Fig. 17, which is based on QCA, the emissivities
at 36.5GHz are higher than at 10.8 GHz. For dense media
consisting of particles densely packed, there are two
different cases. They are differentiated by a large loss
tangent and small loss tangent where the loss tangent is
s=(oe) and s is the conductivity of the medium. A large loss
tangent represents the case where the conductive current
is much larger than the displacement current, while the
reverse is true for the small loss tangent case. It has been
shown by extensive simulations [27] that QCA is valid for
the small loss tangent case. The QCA theory has been
Table 4. Parameters for Monte Carlo Simulations Shown in Figs. 913
a (mm) b
/
(mm) N
/
b
/ /
(mm) N
/ /
V (mm
3
) f
w
(%)
Figures 911 1.0 0.4472 75 0.99795 425 2828 10.5
Figure 12 0.5 0.2271 75 0.49885 425 353.6 10.5
Figure 13 0.25 0.1285 75 0.24945 425 44.19 10.5
Table 5. Numerical Results from Monte Carlo Simulations
for Figs. 911
Parameter At 10.8GHz At 36.5GHz
Absorption rate k
abs
(cm
1
) 0.2849 0.8854
Scattering rate k
s
(cm
1
) 0.01201 0.5738
Extinction rate k
e
(cm
1
) 0.2969 1.4592
Albedo 0.04045 0.3933
Effective permittivity e
eff
1.448i0.158 1.158i0.206
for Fig. 12
Absorption rate k
abs
(cm
1
) 0.3009 0.9150
Scattering rate k
s
(cm
1
) 2.55210
3
0.1380
Extinction rate k
e
(cm
1
) 0.3035 1.0530
Albedo 8.40310
3
0.1310
Effective permittivity 1.508i0.165 1.284i0.156
successful in dry snow [7,26]. However, the QCA theory is
less successful when conductive current dominates [59].
The conductive current can go through several connected
particles. On the other hand, QCA is limited to pair
distribution functions. Thus the applicability of QCA is
also dependent on frequency through the dependence on
s=(oe). We can see that the results of QCA are in better
agreement with those of Monte Carlo simulations at high
frequency of 36.5 GHz, and less successful at low fre-
quency of 10.8 GHz. However, the experiment reported
in Refs. 11 and 57 was at 19.0GHz and QCA was reason-
ably successful. Since Monte Carlo simulations provide
the exact solution of Maxwell equations, this model will be
applied to calculate the four Stokes parameters in the next
section.
Four Stokes parameters of brightness temperature for
foam-covered rough ocean surface are theoretically ana-
lyzed. Important features are shown by the results at 10.8,
19.0, and 36.5 GHz for the four Stokes parameters. It is
important to determine, with respect to these Stokes
parameters, how the foam affects brightness temperatures
and retrieval of the ocean surface wind vector. The rst
two Stokes parameters are increased with the presence of
foam, and the third and fourth parameters are reduced.
The azimuthal variations of polarimetric brightness tem-
perature are also illustrated. The rst two Stokes para-
meters are even functions of f; the last two parameters
are odd functions. Emissions with various windspeeds and
foam layer thickness are also studied. The four Stokes
parameters of brightness temperature are dependent on
windspeeds and foam thickness.
4. TREE SCATTERING MODEL AT UHF/VHF USING SPARSE
MATRIX ITERATIVE APPROACH
Electromagnetic wave interaction with plants plays an
important role in microwave remote sensing applications.
Various scattering models have been developed to study
wave propagation and scattering in vegetation canopy
[17]. In the past, vector radiative transfer theory was
used extensively to compute the backscattering coef-
cients of forests for comparison with microwave back-
scatter measurements [14]. While the radiative transfer
approach can handle complicated geophysical structures,
it ignores coherent wave interactions within the tree
structure. A coherent addition approximation (CAA)
scattering model has been applied to take into account
the effects of coherent scattering and branching structure
1
0.9
0.8
0.7
0.6
20 30 40 50 60 70
Observation angle (degree)
E
m
i
s
s
i
v
i
t
y
36.5G V
10.8G V
Figure 9. Emissivity as a function of observation angle for
vertical polarization. The radius of the coated air bubble is
1.0 mm. (This gure is available in full color at http://www.
for Fig. 13
Absorption rate k
abs
(cm
1
) 0.2686 0.9042
Scattering rate k
s
(cm
1
) 6.52110
5
0.01553
Extinction rate k
e
(cm
1
) 0.2686 0.9197
Albedo 2.42710
4
0.01689
eff
1.478i0.144 1.356i0.140
1
0.7
0.8
0.9
0.6
20 30 40 50 60 70
E
m
i
s
s
i
v
i
t
y
36.5G H
10.8G H
Figure 10. Emissivity as a function of observation angle for
horizontal polarization. The radius of the coated air bubble is
1.0 mm. (This gure is available in full color at http://www.
1
0.8
0.6
0.4
0.2
0 2 4 6 8
Thickness of foam layer (cm)
36.5G V
36.5G H
10.8G V
10.8G H
E
m
i
s
s
i
v
i
t
y
Figure 11. Emissivity as a function of the thickness of the foam
layer at observation angle 531. The radius of the coated air bubble
is 1.0mm. (This gure is available in full color at http://www.
[6,61]. The volume and surface integral equation formula-
tions with the method of moments computational models
have also been employed to investigate tree scattering
[6,6264]. Monte Carlo simulations and tree geometric
models based on Lindenmayer systems [65] have also been
applied to take into account the random nature of trees
[6,16,63]. Foliage attenuation has driven the application of
radar frequency down to the VHF/UHF region. At these
low frequencies, it is useful to have a clear understanding
of the coherent interaction effects of tree branches and
trunk of a single tree or a small group of trees. In this
section we describe a computationally efcient model for
computing tree scattering at VHF/UHF frequencies. At
these low frequencies, the effects of scattering by tree
leaves can be neglected. A structure model with dielectric
cylinders is employed to simulate trees with bare
branches. The method of moments is used to calculate
tree scattering signatures by discretizing the volume
integral equation and transforming it into a matrix equa-
tion [6]. An efcient numerical algorithm based on the
sparse matrix iterative approach (SMIA) [1316] is used to
solve the matrix equation. In the SMIA method, the
impedance matrix of the resultant matrix equation is
decomposed into a sparse matrix for the near interactions
and a complementary matrix, for the far interactions
among the cylindrical subcells of the tree structures.
Using a direct sparse solver to estimate the strong inter-
action part, we iteratively included the weak interaction
contribution to update the solution. The key feature of this
approach is that very little iteration is required to obtain
convergent solutions. The SMIA technique has been ap-
plied in combination with the canonical grid method to
solve rough surface scattering [6,13] and analyze micro-
strip interconnects [14,15]. Section 4.1 gives the volume
integral equation formulation and the MoM treatment for
the tree scattering problem, and describes the SMIA
method. In Section 4.2, some numerical results are illu-
strated for scattering from simulated trees and demon-
strate that the SMIA method can signicantly reduce the
number of iterations when compared to the conventional
conjugate gradient method (CGM).
1
0.8
0.6
0.4
0.2
0 2 4 6 8
36.5G V
36.5G H
10.8G V
10.8G H
E
m
i
s
s
i
v
i
t
y
is 0.5 mm. (This gure is available in full color at http://
www.mrw.interscience.wiley.com/erfme.)
1
0.8
0.6
0.4
0.2
0 2 4 6 8
36.5G V
36.5G H
10.8G V
10.8G H
E
m
i
s
s
i
v
i
t
y
is 0.25mm. (This gure is available in full color at http://
www.mrw.interscience.wiley.com/erfme.)
1
0.9
0.8
0.7
0.6
20 30 40 50 60 70
E
m
i
s
s
i
v
i
t
y
DMRT Tv
DMRT Th
Experiment Tv
Experiment Th
Figure 14. Comparison between the simulation results and
measurements of microwave emissivity at 10.8GHz for horizontal
polarization and vertical polarization, respectively. (This gure is
erfme.)
1
0.9
0.8
0.7
0.6
20 30 40 50 60 70
E
m
i
s
s
i
v
i
t
y
DMRT Tv
DMRT Th
Experiment Tv
Experiment Th
Figure 15. Comparison between the simulation results and
measurements of microwave emissivity at 36.5GHz for horizontal
polarization and vertical polarization, respectively. (This gure is
erfme.)
4.1. Formulation and Sparse Matrix Iterative Approach
The discrete-dipole approximation (DDA) has been
applied to calculate electromagnetic scattering and ab-
sorption by trees [6,16,63]. At VHF/UHF frequencies, the
tree trunks and branches can be modeled as thin dielectric
circular cylinders. In the discretization process, each tree
branch is divided into cylindrically shaped subcells with
the same radius. The volume integral equation for the
electric eld is given by
"
EE( " rr) =
"
E E
inc
( " rr)
k
2
0
e
0
_
V
G
0
( " rr; " rr
/
)
.
[e
p
( " rr
/
) e
0
]
"
EE( " rr
/
) d" rr
/
(59)
where
"
EE
inc
( " rr) is the incident eld, e
p
( " rr) is the permittivity
distribution within region V and k
0
, e
0
, and G
0
( " rr; " rr
/
)
are the wavenumber, permittivity, and free-space
dyadic Green function, respectively. If we subdivide
the volume into small elemental volumes DV
i
, each cen-
tered at " rr
i
, i =1; 2; . . . ; N, the discretized version of (59)
becomes
" pp
i
=a
i
. "
EE
inc
i
a
i
.
k
2
0
e
0
N
j =1; jOi
_
DV
j
d" rr
/
G
0
( " rr
i
; " rr
/
)
DV
j
.
" pp
j
(60)
where
"
EE
inc
i
, " pp
i
=DV
i
(e
pi
e
0
)
"
EE
i
, and
"
EE
i
are the respective
vectors of the incident eld, dipole moment, and
electric eld at the center " rr
i
of the ith elemental
volume DV
i
. The summation over j runs over all the
subcells ( jOi) that form the tree structure. e
pi
is the
permittivity of the ith elemental volume, which is
assumed to be constant within DV
i
. In (60), the dyad
a
i
= ^ xx^ xxa
ix
^ yy^ yya
iy
^ zz^ zza
iz
is a diagonal matrix whose
elements, for a vertical circular cylindrical cell of radius
a
i
and length
i
, are given by [6]
a
ix
=Dv
i
(e
pi
e
0
)
1
1
e
pi
e
0
1
_ _
(L
ix
k
2
0
D
ix
)
a
iy
=Dv
i
(e
pi
e
0
)
1
1
e
pi
e
0
1
_ _
(L
iy
k
2
0
D
iy
)
a
iz
=Dv
i
(e
pi
e
0
)
1
1
e
pi
e
0
1
_ _
(L
iz
k
2
0
D
iz
)
(61)
where
L
ix
=L
iy
=

i
2
2
i
4a
2
i
_
L
iz
=1

i
2
2
i
4a
2
i
_
(62)
and
D
ix
=D
iy
=
a
2
i
8
ln
2
i
4a
2
i
_

i
2
i
4a
2
i
_

i

i
8
2
i
4a
2
i
_

i
_ _
i
k
0
a
2
i

i
6
D
iz
=
a
2
i
4
ln
2
i
4a
2
i
_

i
2
i
4a
2
i
_

i
i
k
0
a
2
i

i
6
(63)
Table 8. Emissivities of Experimental Data, DMRT Model, and Air-Ocean Half-Space Results
Experiment DMRT Model AirOcean Half Space Results
y(1) V H V H V H
a. Results at 10.8GHz
30 0.920 0.900 0.915 0.897 0.422 0.337
35 0.930 0.890 0.923 0.898 0.440 0.322
40 0.955 0.875 0.930 0.899 0.462 0.305
45 0.945 0.860 0.939 0.898 0.489 0.285
50 0.945 0.815 0.946 0.894 0.523 0.263
55 0.929 0.805 0.950 0.885 0.564 0.239
60 0.919 0.780 0.951 0.871 0.615 0.211
b. Results at 35.6GHz
30 0.910 0.880 0.916 0.908 0.526 0.429
40 0.940 0.885 0.916 0.900 0.570 0.391
45 0.945 0.875 0.915 0.895 0.599 0.367
55 0.935 0.850 0.911 0.877 0.676 0.310
60 0.912 0.785 0.905 0.863 0.725 0.276
Table 9. Parameters of FoamModel for Simulation Results
in Figs. 16 and 17
d 2a 2b
1
2b
2
f f
w
f
1
f
2
3 cm 2000mm 1220mm 1995.9mm 74% 10.3% 12.9% 61.1%
As can be seen from (59), we need to consider only those
volumes that have e
pi
Oe
0
. The matrix equation (60) is of
dimension 3N3N, where N is the total number of small
cylindrical cells. The factor 3 arises from the Cartesian
components of the polarization vector.
Equation (60) is the DDA matrix equation, which can
be expressed as
Z
. "
XX =
"
CC (64)
where the vector
"
C C corresponds to the known incident eld
and
"
XX is the unknown polarization vector to be solved
numerically. In general, the impedance matrix Z resulting
from the discretization of the volume integral equation is
large, dense, and complex. Its elements are proportional to
the integrals of the Green dyad as shown in (60). Usually
an iterative solver, such as conjugate gradient method
(CGM), is coupled with some kind of preconditioning
strategy to improve the spectral properties of the impe-
dance matrix Z such that the matrix equation (64) can be
solved much more rapidly [66]. Several preconditioners,
such as blockdiagonal [67], sparse approximate inverse
(SPAI) [68], and incomplete LU (ILU) factorization [69],
have been considered in the electromagnetic wave scatter-
ing problem.
The sparse matrix iterative approach (SMIA) technique
[6,1316] has been applied to solve the matrix equation
(64). Equation (64) is solved iteratively by decomposing
the impedance matrix Z into two parts
Z=Z
(s)
Z
(w)
(65)
where Z
(s)
, which is a sparse matrix, includes the near or
strong interactions among the close-by cylindrical sub-
cells, and Z
(w)
, which is a complementary matrix account-
ing for the far or weak interactions, contains the
remaining elements of the impedance matrix Z.
The numerical procedure uses a sparse solver to estimate
the rst-order solution
"
XX
(1)
Z
(s)
. "
XX
(1)
=
"
CC (66)
and iteratively includes the weak interaction contribution
to update the higher-order solution
"
XX
(n1)
:
Z
(s)
. "
XX
(n1)
=
"
CC Z
(w)
. "
XX
(n)
(67)
The iterations are carried out until the error norm criter-
ion
| Z
. "
XX
(n)

"
CC |
|
"
CC |
_d (68)
is satised. d has been set equal to 110
4
for the
numerical computations presented in Section 4.2.
Table 10. Quantities of Foam Based on QCA from Fig. 18
Absorption rate k
abs
(cm
1
) 0.1763 0.9042
Scattering rate k
s
(cm
1
) 0.0567 0.01553
Extinction rate k
e
(cm
1
) 0.2330 0.9197
Albedo 0.2435 0.01689
eff
1.4573i0.1245 1.356i0.140
Table 11. Quantities of Foam Based on Monte Carlo
Simulations from Fig. 16
Absorption rate k
abs
(cm
1
) 0.2686 0.9042
Scattering rate k
s
(cm
1
) 6.52110
5
0.01553
Extinction rate k
e
(cm
1
) 0.2686 0.9197
Albedo 2.42710
4
0.01689
eff
1.478i0.144 1.356i0.140
1
0.9
0.8
0.7
0.6
0 10 20 30 40 50 60 70
10.8 T
v
10.8 T
h
36.5 T
v
36.5 T
h
E
m
i
s
s
i
v
i
t
y
Figure 17. Emissivities as a function of observation angle for
vertical and horizontal polarizations at 10.8 and 36.5GHz, using
the model based on QCA.
1
0.9
0.8
0.7
0.6
0 10 20 30 40 50 60 70
10.8 Tv
10.8 Th
36.5 tv
36.5Th
E
m
i
s
s
i
v
i
t
y
Figure 16. Emissivities as a function of observation angle for
vertical and horizontal polarizations at 10.8 and 36.5GHz, using
the model based on Monte Carlo simulations.
4.2. Numerical Results
As shown in Fig. 18, six trees are generated on the basis of
the Lindenmayer systems [6,65]. The tree characteristics
are given in Table 12. We can see that the tree height
varies from 2.67 to 5.65m, and the total number of
branches, including the trunk, can be as much as several
hundreds. The total number of discretized cylindrical
subcells N and the total number of unknown polarizations
3N are also given. The total number of unknowns ranges
from several hundreds for a simpler tree to several
thousands for a more complicated tree. We can see that
the number of unknowns is much smaller when compared
to using a surface integral equation approach for calculat-
ing the tree scattering [64]. In addition, Table 12 includes
the expansions of the trees in the x and y directions, the
variations of the radii and lengths of branches for each
tree, and the woody volume of each tree.
A number of numerical examples for the polarimetric
bistatic scattering from the simulated trees are presented
in the following. The permittivities used are
e
p
=9:62 i4:332 and e
p
=33:357i21:408, which are cal-
culated fromthe formula developed by Ulaby et al. [3]. The
two permittivities are for gravimetric moisture contents of
M
g
=26% and M
g
=68% at 300MHz, respectively. The
radiowavelength is l =1m, and the incidence angle is
y
i
=45
. The bistatic scattering cross sections of trees 5

and 6 for e
p
=9:62 i4:332 are shown in Figs. 19 and 20,
respectively. We also use the solutions for an exact full
matrix inversion (INV) and the conjugate gradient method
(CGM) as baselines to illustrate the accuracy of the SMIA
approach. We can see that solutions of the SMIA method
agree very well with the solutions of exact matrix inver-
sion and CGM.
Figure 21 shows the number of iterations required to
achieve convergence, and compares the convergence rates
for solving the scattering from trees 5 and 6 using the
CGM and SMIA techniques. Note that CGM requires
thousands of iterations for the percentage error to reduce
to 10
2
% while the SMIA requires only about 10 itera-
tions. In Table 13 we present a comparison of the condition
number of the impedance matrix Z, the number of itera-
tions for horizontal and vertical-polarization incidence,
matrix lling, solving, and total computation time for the
scattering solutions of six trees. The computational time is
based on a Dell Latitude C800 with a 512M RAM memory
and 848MHz microprocessor. It is noted that, because of
their large condition numbers, trees 3 and 4 require a
large number of iterations to converge by using the
conventional CGM, while the SMIA method takes only a
few iterations to reach a convergent solution. It is clear
from Table 13 that, compared to the CGM, the SMIA
technique reduces the number of iterations by a factor of
4100 and provides a much faster numerical solution
scheme for computing tree scattering while keeping the
sufcient accuracy.
In Figs. 22 and 23 we plot the respective bistatic
scattering cross sections of trees 5 and 6 for a larger tree
permittivity e
p
=33:357i21:408. These plots show that
solutions based on the SMIA method agree very well with
the exact matrix inversion for this high-dielectric-con-
stant case. As shown in Table 14, because of the large
matrix condition numbers in this high-dielectric-contrast
case, scattering results for the CGM are not computed.
Compared to Table 13, the SMIA method requires more
iteration to converge for scattering from trees with larger
permittivities. However, SMIA is still more efcient than
the full matrix inversion in matrix solving.
6
4
2
0
2
2
2
2
0
0
6
4
2
0
2
2
2
2
0
0
Tree 2 Tree 1
6
4
2
0
2
2
2
2
0
0
Tree 4
6
4
2
0
2
2
2
2
0
0
Tree 5
6
4
2
0
2
2
2
2
0
0
Tree 3
6
4
2
0
2
2
2
2
0
0
Tree 6
Figure 18. Simulated tree samples using
the Lindenmayer systems. (This gure is
available in full color at http://www.mrw.
interscience.wiley.com/erfme.)
We have illustrated the use of SMIA numerical tree
scattering model to calculate scattering from simulated
trees with up to several hundred branches. We use the
solutions of an exact full matrix inversion and the con-
jugate gradient method (CGM) as baselines to illustrate
the accuracy and efciency of the SMIA algorithm. Solu-
tions based on the SMIA method agree very well with the
exact matrix inversion and the CGM. Compared to
the CGM, the SMIA approach can signicantly reduce
the number of iterations and provides a much faster
numerical scheme for computing tree scattering.
5. WAVE SCATTERING BY DISCRETE SCATTERERS USING
UV MULTILEVEL PARTITION METHOD
With the advent of modern computers and the develop-
ment of fast numerical methods, Monte Carlo simulations
of volume scattering problems have become an attractive
approach. A common method that has been used in
numerical simulations is the volume integral equation
method and its solution by the method of moments
(MoM). Conventional implementation of the MoM
requires O(N
3
) operations and O(N
2
) computer
memory storage. Several fast numerical methods have
been used, such as the sparse matrix canonical grid
(SMCG) method [6,7072] and the fast multipole
method (FMM) [73,74], and a characteristic basis
function has been used to solve volume scattering
problems [75]. In all these approaches, subsectional basis
functions and the lowest-order Green function or the
lowest-order partial wave are used to construct the
impedance matrix. In Ref. 71, small circular discrete
scatterers were used for the SMCG method and the matrix
multiply operation was extended to partial waves higher
than the lowest order.
Table 12. Simulated Tree Characteristics
Tree 1 Tree 2 Tree 3 Tree 4 Tree 5 Tree 6
Number of cylinders 7 15 31 63 127 255
Tree height (m) 2.66996 3.47377 3.83806 4.93697 5.48677 5.65101
Minimum x coordinate (m) 0 0.2577 0.4513 0.7715 0.8045 1.1348
Maximum x coordinate (m) 0.27391 0.7793 0.72995 1.51781 1.05277 1.74531
Minimum y coordinate (m) 0.7967 1.2325 1.4731 1.9889 2.0793 2.2819
Maximum y coordinate (m) 0.25491 0.36842 0.95329 1.07843 1.618 1.29136
Maximum cylinder length (m) 1 1 1 1 1 1
Minimum cylinder length (m) 0.34548 0.25109 0.15404 0.11553 0.04543 0.03565
Maximum cylinder radius (m) 0.05 0.05 0.05 0.05 0.05 0.05
Minimum cylinder radius (m) 0.03782 0.03244 0.03281 0.02636 0.03074 0.02448
Total cylinder volume (m
3
) 0.02636 0.03935 0.05961 0.0812 0.13313 0.14928
Number of cells 58 120 191 433 554 1077
Number of unknowns 174 360 573 1299 1662 3231
Tree 5 (HH)
0
5
10
15
20
25
30
35
40
50 0 50
Tree 5 (VV)
0
5
10
15
20
25
30
50 0 50
Tree 5 (VH)
0
5
10
15
20
25
30
50 0 50
S
c
a
t
t
e
r
i
n
g

c
r
o
s
s

s
e
c
t
i
o
n

(
d
B
s
m
)
S
c
a
t
t
e
r
i
n
g

C
r
o
s
s

S
e
c
t
i
o
n

(
d
B
s
m
)
S
c
a
t
t
e
r
i
n
g

C
r
o
s
s

S
e
c
t
i
o
n

(
d
B
s
m
)
Scattering angle (degree) Scattering angle (degree) Scattering angle (degree)
INV
SMIA
CGM
Figure 19. Comparison of the
bistatic scattering cross sections of
tree 5 calculated by the full matrix
inversion (INV), conjugate gradient
method (CGM), and sparse matrix
iterative approach (SMIA). The wa-
velength is l =1 m, the incidence
angle y
i
=451, and the permittivity
e
p
=9:62i4:332. (This gure is
available in full color at http://www.
In this section, we present a UV method using multi-
level partitioning (UV-MLP) to solve volume scattering
problems [20]. The UV method has also been used to solve
rough surface scattering problems [18,19]. The method
consists of setting up a table of transmitting and receiving
block sizes and their separations. For a specic scattering
problem with given geometry, the scattering structure is
partitioned into multilevel blocks. We will illustrate the
UV method for the case of scattering by a large number of
circular cylinders. The cylinders are moderate in size with
diameter comparable to wavelength. The problem has
many physical applications, such as interesting resonant
scattering and weak and strong localization occur when
particles of wavelength sizes are densely packed together
[7]. For this kind of problems, the FoldyLax multiple
scattering equations can be used with higher-order partial
waves as basis functions. The impedance matrices are
thus in terms of Green functions of higher-order partial
waves. It is shown that the UV decomposition can be
applied directly to the impedance matrix of partial waves
of higher order. We also apply the multilevel matrix
partitioning (MLP) to partition the partial waves impe-
dance matrix Z. By looking up the rank in the predeter-
mined table, the partial waves impedance matrix for
a given transmitting and receiving block of discrete
scatterers is expressed into a UV matrix product.
In Section 5.2, we describe the problem of independent
rank determination. We give the FoldyLax equations of
scattering by a conglomeration of circular cylinders in
Section 5.3. In Section 5.4, we describe the multilevel
partitioning process and the UV method, in Section 5.5.
In Section 5.6, the computational complexity of the
proposed algorithm is derived. In Section 5.7, we demon-
strate the technique for a problem with up to 4096
cylinders, each with diameter equal to 1 wavelength. If a
volumetric subsectional element is used with MoM for the
problem of 1024 cylinders, using 100 points per square
wavelength, the number of unknowns will be approxi-
mately 80,425. Using the present method of multilevel UV,
the solution is completed in 14min using a single PC
processor of 2.6GHz. The numerical results are illustrated
and discussed.
Tree 6(VV)
5
0
5
10
10
15
20
25
30
35
50 0 50
S
c
a
t
t
e
r
i
n
g

c
r
o
s
s

s
e
c
t
i
o
n

(
d
B
s
m
)
Tree 6(VH)
5
10
15
20
25
30
50 0 50
S
c
a
t
t
e
r
i
n
g

c
r
o
s
s

s
e
c
t
i
o
n

(
d
B
s
m
)
Tree 6(HH)
5
0
5
10
15
50 0 50
S
c
a
t
t
e
r
i
n
g

c
r
o
s
s

s
e
c
t
i
o
n

(
d
B
s
m
)
INV
SMIA
CGM
Scattering angle (degree) Scattering angle (degree) Scattering angle (degree)
Figure 20. Comparison of the
bistatic scattering cross sections of
inversion (INV), conjugate gradient
method (CGM) and sparse matrix
iterative approach (SMIA). The wave-
length is l =1 m, the incidence
angle y
i
e
p
=9:62i4:332. (This gure is avail-
able in full color at http://www.mrw.
H Polarization V Polarization
10
2
10
1
10
0

10
1
10
2
10
3
10
4
%

E
r
r
o
r
10
2
10
1
10
0

10
1
10
2
10
3
10
4
%

E
r
r
o
r
10
0
10
2
Number of iterations
10
0
10
2
Number of iterations
SMIA(Tree 5)
SMIA(Tree 6)
CGM (Tree 6)
CGM (Tree 5)
SMIA(Tree 5)
SMIA(Tree 6)
CGM (Tree 6)
CGM (Tree 5)
Figure 21. Comparison of the convergence rates for calculating
scattering from trees 5 and 6 using the conjugate gradient method
(CGM) and sparse matrix iterative approach (SMIA). (This gure
is available in full color at http://www.mrw.interscience.wiley.
com/erfme.)
5.1. Problem Independent Rank Table
We describe two methods of predetermination of rank
used in volume scattering problems. Consider two square
blocks each of side S. The center is separated by distance R
as shown in Fig. 24. Because of multilevel portioning, we
have 2S_R_C
D
S, where C
D
is a constant depending on
the dimension of the problem.
5.1.1. Method of Volume Discretization or Boundary
Discretization
5.1.1.1. Volume Discretization. We discretize the blocks
by 100 points per square wavelength. Let there be N
points in the transmitting block and N points in the
receiving block. Let receiving points be denoted by " rr
p
,
p=1; 2; . . . ; N, and the transmitting points denoted by " rr
q
,
q =1; 2; . . . ; N. The Green function in a two-dimensional
problem is given by
Z
pq
=
i
4
H
(1)
0
(k[ " rr
p
" rr
q
[) (69)
Then the impedance matrix is of dimension NN. We
can use the singular value decomposition (SVD) method to
determine the rank. Let s
1
be the largest singular value,
and let the singular values be arranged in decreasing
magnitude. Given a threshold e, the rank r is such that
[s
r 1
=s
1
[_e.
5.1.1.2. Enclosing Boundary Radiation. By Huygens
principle, the equivalent sources are the boundary sur-
faces that are the four sides of the block. By choosing the
boundary radiation, it will reduce the number of points
and yet get similar rank for the impedance matrix.
5.1.1.3. Coarse Sampling. As block size increases,
the method described becomes prohibitive. However, we
know from previous experience that the rank is much
smaller than the size of the matrix. Thus we can use much
coarser sampling. We use the terminology in which
dense sampling is the usual 10 points per wavelength
for boundary or volume sampling and coarse sampling
Table 13. SMIA and CGM Compared
Condition number 53127.99 4084.755 521019.4 165276.5 1975.732 1481.116
Method CGM SMIA CGM SMIA CGM SMIA CGM SMIA CGM SMIA CGM SMIA
Matrix lling (s) 12.77 10.30 31.60 25.85 57.29 49.43 169.64 139.31 238.80 196.23 715.84 591.04
Number of iterations (H) 750 6 1207 6 10586 7 11454 7 1634 11 3783 8
Number of iterations (V) 788 6 1209 6 216 7 11471 6 1648 11 3616 8
Matrix solving (s) 17.05 0.173 118.50 0.769 2620.7 2.374 14437. 13.628 3403.2 30.673 28906. 124.98
Total time (s) 29.89 10.53 150.19 27.06 2678.1 51.95 14607. 153.56 3642.4 227.26 29623. 717.08
Tree 5 (HH) Tree 5 (VV) Tree 5 (VH)
INV
SMIA
0
-5
-10
-15
-20
-25
-30
-50 0 50
Scattering angle (degree)
-50 0 50
-50 0 50
S
c
a
t
t
e
r
i
n
g

c
r
o
s
s

s
e
c
t
i
o
n

(
d
B
s
m
)
S
c
a
t
t
e
r
i
n
g

c
r
o
s
s

s
e
c
t
i
o
n

(
d
B
s
m
)
S
c
a
t
t
e
r
i
n
g

c
r
o
s
s

s
e
c
t
i
o
n

(
d
B
s
m
)
5 5
0
-5
-10
-15
-20
-25
0
-5
-10
-15
-20
-25
-30
-35
-40
Figure 22. Comparison of the bi-
static scattering cross sections of
inversion (INV) and sparse matrix
iterative approach (SMIA). The wave-
length is l =1m, the incidence angle
y
i
=451, and the permittivity e
p
=
33:357i21:408. (This gure is avail-
able in full color at http://www.mrw.
means that we pick the number of points to be slightly
greater than the rank. This means that we have a priori
knowledge of roughly what the rank is on the basis of
the numerical experiments carried out. Coarse sampling
means that we pick the number of points several
times larger than the rank. In forming a matrix, we
need to select the number of points in the receiving region
and a set of points in the transmitting region. Thus
we have (receiving sampling)(transmitting sampling).
This means that we can have densecoarse sampling,
meaning dense in the receiving region and coarse in
the transmitting region etc. To establish the rank of a
matrix, we nd that it is sufcient to have coarsecoarse
sampling.
5.1.2. Method of Using Two Circular Cylinders. We can
replace the transmitting and receiving blocks by two
cylinders of diameter D
b
=2S=

p
_
, as illustrated in
Fig. 25, where both the cylinder and the square will
have the same area. The centers of the receiving cylinder
and the transmitting cylinder are respectively " rr
r
and " rr
t
. If
partial waves are used to describe the waves from the
transmitting cylinder to the receiving cylinder, we can
truncate the partial waves at N
b
c
=0:75 kD
b
. We use the
subscript b to denote block, reecting the fact that the
purpose of these two large equivalent cylinders is only to
set up the rank table and not the physical size of the many
cylinders in the multiple scattering problems. From
FoldyLax equations [6], for two equivalent receiving
and transmitting cylinders, the interaction matrix has
the following form
w
(r)
m
=incidence wave
N
b
c
n=N
b
c
H
(1)
(mn)
(k[ " rr
r
" rr
t
[) e
i(mn)f
r
r
r
t
T
n
w
(t)
n
(70)
where m= N
b
c
; N
b
c
1; . . . ; N
b
c
. Then the FoldyLax
partial wave impedance matrix Z
b
is of dimension
Tree 6(HH) Tree 6(HH) Tree 6(VV)
5
5
5
0
0
-5
-5
10
-50 0 50
Scattering angles (degree)
-50 0 50
S
c
a
t
t
e
r
i
n
g

c
r
o
s
s

s
e
c
t
i
o
n

(
d
b
m
s
)
S
c
a
t
t
e
r
i
n
g

c
r
o
s
s

s
e
c
t
i
o
n

(
d
b
m
s
)
S
c
a
t
t
e
r
i
n
g

c
r
o
s
s

s
e
c
t
i
o
n

(
d
b
m
s
)
-10
-10
-10
0
-5
-50 0 50
-15
-15
-20
-30
-25
-35
INV
SMIA
Figure 23. Comparison of the bi-
static scattering cross sections of
inversion (INV) and sparse matrix
iterative approach (SMIA). The wa-
velength is l =1 m, the incidence
angle y
i
e
p
=33:357i21:408. (This gure is
available in full color at http://www.
Table 14. SMIA and INV Compared
Condition number 130733.0 10351.14 1161364.0 706516.4 5229.153 26051.72
Method INV SMIA INV SMIA INV SMIA INV SMIA INV SMIA INV SMIA
Matrix lling (s) 10.32 10.28 26.03 25.84 47.30 47.16 139.89 139.25 196.96 196.24 590.67 596.81
Number of iterations (H) 0 10 0 11 0 15 0 19 0 22 0 68
Number of iterations (V) 0 10 0 13 0 15 0 19 0 20 0 60
Matrix solving (s) 0.200 0.250 2.245 1.541 9.085 4.116 104.93 25.106 220.46 45.667 1612.0 442.72
Total time (s) 10.56 10.57 28.35 27.47 56.84 51.44 245.10 164.64 417.77 242.27 2203.3 1040.2
(2N
b
c
1) (2N
b
c
1) and the elements are given by
(Z
b
)
mn
=H
(1)
(mn)
(k[ " rr
r
" rr
t
[) e
i(mn)f
r
r
r
t
(71)
where m; n= N
b
c
; N
b
c
1; . . . ; N
b
c
. We can apply the
SVD to determine the rank of Z
b
. This is a fast method
because N
b
c
{N. This has a slightly larger rank because
their smallest separation is slightly less.
On the basis of the two techniques described above, a
rank table is given in Table 15, where the superscripts
1, 2, 3, and 4 are used to indicate the different cases of
(1) area sampling (100 points per square wavelength),
(2) boundary (line) dense sampling (10 points per wave-
length), (3) boundary (line) coarse sampling (4 points per
wavelength), and (4) cylindrical wave expansion, respec-
tively. The threshold chosen for rank determination is
1 10
5
. It is seen that all of them gives essentially same
rank for same physical problem.
5.2. FoldyLax Partial-Wave Equation
Consider N
p
cylinders of permittivity e
p
and radius a
distributed arbitrarily in a square, as illustrated in
Fig. 26. Let a plane TM wave be incident on them. The
FoldyLax multiple scattering equation in terms of partial
wave expansions is expressed as [6,20]
c
(N)(q)
n
=
i
n
e
inf
i
k
e
i
"
kk
i
.
" rr
q
E
vi
B
(N)
n
B
(N)
n
Nc
n
/
=N
c
N
p
p=1; pOq
H
(1)
(nn
/
)
(k[ " rr
p
" rr
q
[)
e
i(nn
/
)f
r
p
r
q
S
(N)
n
/ c
(N)(p)
n
/
(72)
where c
(N)(q)
n
is internal eld coefcient, " rr
q
is the position of
the qth cylinder and B
(N)
n
and S
(N)
n
/ are given by [6,20]
B
(N)
n
=B
(N)
n
=
2ik
pak
p
1
k
p
H
(1)
n
(ka)J
/
n
(k
p
a) kH
(1)
n
/
(ka)J
n
(k
p
a)
(73)
S
(N)
n
=S
(N)
n
=
ipak
p
2k
[k
p
J
n
(ka)J
/
n
(k
p
a)
kJ
/
n
(ka)J
n
(k
p
a)]
(74)
Parameter k
p
is wavenumber inside the particle. Each
particle has N
c
partial waves. Note that the number N
c
is
different from N
b
c
, which denotes rank determination for
block structure. For each particle, the partial wave is
truncated at N
c
=1.5ka.
The FoldyLax equation is a matrix equation with
the dimension of N=N
p
(2N
c
1). Each particle has
(2N
c
1) coefcients. The impedance matrix for the
particle with itself is a (2N
c
1) (2N
c
1) unit matrix.
The impedance matrix for two different particles p and
S
R
Figure 24. Illustration of rank determination.
1.13S
R
Figure 25. Illustration of rank determination with FoldyLax
equation.
Table 15. Rank Table of Volume Scattering
L
x
=L
y
(l) Points in Block Distance (l) Rank
1.0 100 2.00 9
1
, 9
2
, 9
3
, 8
4
1.0 100 4.24 6, 6, 6, NA
2.0 400 4.00 10, 11, 11, 11
2.0 400 8.48 7, 7, 7, NA
4.0 1,600 8.00 12, 13, 13, 15
4.0 1,600 17.0 8, 9, 9, NA
8.0 6,400 16.0 NA, 18, 18, 23
16.0 25,600 32.0 NA, 26, 26, 37
32.0 102,400 64.0 NA, 41, 41, 63
Figure 26. Illustration of many random distributed cylinders.
q,
"
"
ZZ
"
ZZ
pq
, is of dimension (2N
c
1) (2N
c
1), and the
matrix elements are given by B
(N)
n
H
(1)
(nn
/
)
(k[ " rr
p
" rr
q
[)
exp[i(n n
/
)f
r
p
r
q
]S
(N)
n
/ .
5.3. Matrix Compression for Partial Waves
Suppose that the cylinders are of diameter 1 wavelength
and are densely packed as in Fig. 26. Then N
c
=0.75(2p) =4,
so that there are nine partial waves for each particle.
Consider two blocks of side S wavelengths, with the
distance separated by distance R, as shown in Fig. 27,
the number of cylinders in each block is approximately S
2
.
The dimension of the impedance matrix using partial
waves between the two blocks is S
2
(2N
c
1) =9S
2
.
For example, if S=32 wavelengths, then the impe-
dance matrix of the blocks is of dimension (9)(32 32) =
9216. However, the rank is only 41 from Table 15. We
observe the following:
1. If volume discretization is used, and assuming 100
volumetric elements per square wavelength over the
areas of the cylinders, the dimension of the impe-
dance matrix between the two blocks is 80,425.
2. Since the particles of the physical problem are of 1
wavelength diameter, using partial wave expansion,
the dimension of the impedance matrix between
these two blocks is only 9216, which is substantially
less than that of volume discretization. This reduc-
tion is even more for three-dimensional problems.
3. Whether volume discretization or partial wave ex-
pansion for the 1l cylinders, the rank is the same
and is 41.
4. Since the use of partial waves reduces the dimension
of the matrix, it is advantageous to use partial
waves.
5. It is important to recognize that UV method can
directly be applied to the compression of the partial
wave impedance matrix of dimension 9216.
6. The impedance matrix is of the form of
B
(N)
n
H
(1)
(nn
/
)
(k[ " rr
p
" rr
q
[) exp[i(n n
/
)f
r
p
r
q
]S
(N)
n
/ . Only
the factor H
(1)
(nn
/
)
(k[ " rr
p
" rr
q
[) exp[i(n n
/
)f
r
p
r
q
]
depends on both particles. Thus the compression
needs be done only for this term. After the compres-
sion is done, we can postmultiply by B
(N)
n
and
premultiply by S
(N)
n
/ .
After the internal eld coefcients are determined, the
scattered eld in direction f
s
is
E
s
=
2k
pr
e
i(krp=4)
N
p
q =1
N
c
n=N
c
S
(N)
n
c
q
n
e
in(f
s
p=2)
e
i
"
kks
.
" rr
q
(75)
The bistatic scattering coefcient as dened in Ref. 2 can
be calculated.
5.4. Multilevel Partition Process
For the physical problem of the cylinders distributed in
Fig. 28, we use the same techniques to do multilevel
partitioning as in the case of three-dimensional rough
surface scattering [19]. However, the matrix equation is in
terms of partial waves form and there are (2N
c
1) partial
waves for each cylinder. If there are M
r
cylinders in the
receiving group and M
t
cylinders in the transmitting
groups, the size of the interaction matrix is
M
r
(2N
c
1) M
t
(2N
c
1).
Assuming that we have a square area as shown in
Fig. 28a, we rst split it into four blocks as shown in
Fig. 28b. Each of the four blocks is a subgroup at the Pth
level, which has the largest group size. Then we split each
subblock into another four small groups as shown in
Fig. 28c. This splitting process is continued until we reach
the smallest group size, which is at the rst level. We
decompose the full impedance matrix in Eq. (76) as the
sum of P sparse matrices as follows:
Z=Z
(0)
Z
(1)
Z
(2)
Z
(P1)
(76)
Matrix Z
(0)
includes all the interactions among neighbor-
ing groups (including the self group) at the rst level.
Matrix Z
(1)
includes all the interactions among neighbor-
ing groups at the second level and consists of blocks in the
rst level. Similarly, matrix Z
(i)
includes all the interac-
tions among neighboring groups at the (i 1)th level, but
these consist of blocks in the ith level.
To facilitate understanding of the multilevel partition-
ing process, we give an example that has 64 subgroups at
the rst level. Assume each group has M elements. For
this example, the highest level is P=3. Thus
Z=Z
(0)
Z
(1)
Z
(2)
and the impedance matrix of Z is
64 M64 M and has 4096 MM blocks.

R
S
Figure 27. Illustration of interactions of many cylinders within
two blocks.
(a) (b) (c)
Figure 28. Illustration of multilevel partitioning process.
In Fig. 29a, the 64 subgroups of the rst level are
expressed as 1
1
; 2
1
; . . . ; 64
1
. E. In Fig. 29b, the level 2
groups are shown and are represented as 1
2
; 2
2
; . . . ; 16
2
.
Each of the level 2 groups has four level 1 groups. For
example, the group 3
2
has 4 groups of 5
1
, 6
1
, 13
1
, and 14
1
.
In Fig. 29c, the level 3 groups are shown, denoted as 1
3
, 2
3
,
3
3
, and 4
3
. Each of the level 3 groups has four level 2
groups. For example, the group 2
3
has four groups of 3
2
, 4
2
,
7
2
, and 8
2
.
We use m
i
to represent the group m of the level i. Then
the matrix Z
m
i
n
i
represents the interactions between the
receiving group m and the transmitting group n of the ith
level. Since the four level 1 groups form a level 2 group
and four level 2 groups form a level 3 group, the dimension
of Z
m
1
n
1
is MM, and the dimension of Z
m
2
n
2
is 4M4M.
Examples are as follows:
Z
3
2
5
2
=
Z
5
1
17
1
Z
5
1
18
1
Z
5
1
25
1
Z
5
1
26
1
Z
6
1
17
1
Z
6
1
18
1
Z
6
1
25
1
Z
6
1
26
1
Z
13
1
17
1
Z
13
1
18
1
Z
13
1
25
1
Z
13
1
26
1
Z
14
1
17
1
Z
14
1
18
1
Z
14
1
25
1
Z
14
1
26
1
_
_
_
_
(77)
Z
6
2
5
2
=
Z
19
1
17
1
Z
19
1
18
1
Z
19
1
25
1
Z
19
1
26
1
Z
20
1
17
1
Z
20
1
18
1
Z
20
1
25
1
Z
20
1
26
1
Z
271171
Z
271181
Z
271251
Z
271261
Z
28
1
17
1
Z
28
1
18
1
Z
28
1
25
1
Z
28
1
26
1
_
_
_
_
(78)
Z
1
3
2
3
=
Z
1
2
3
2
Z
1
2
4
2
Z
1
2
7
2
Z
1
2
8
2
Z
2
2
3
2
Z
2
2
4
2
Z
2
2
7
2
Z
2
2
8
2
Z
5
2
3
2
Z
5
2
4
2
Z
5
2
7
2
Z
5
2
8
2
Z
6
2
3
2
Z
6
2
4
2
Z
6
2
7
2
Z
6
2
8
2
_
_
_
_
(79)
In Z
(0)
, we select interactions of the level 1 groups with
their nearest neighbors. For example, 20
1
has eight neigh-
bors of 11
1
, 12
1
, 13
1
, 19
1
, 21
1
, 27
1
, 28
1
, and 29
1
. Thus Z
(0)
includes Z
11
1
20
1
, Z
12
1
20
1
, Z
13
1
20
1
Z
19
1
20
1
, Z
20
1
20
1
, Z
21
1
20
1
Z
27
1
20
1
, Z
28
1
20
1
, and Z
29
1
20
1
, a total of nine matrices. Note
that self-interaction is also included here. We note that
(1) block size of Z
m
1
n
1
is MM, (2) there are 36 interior
rst-level 1 groups of blocks =36 9 =324 (MM)
blocks, (3) 24 edge rst-level groups of six blocks =24
6 =144 (MM) blocks, and (4) four corner rst-level
groups of four blocks =4 4 =16 (MM) blocks. Thus
Z
(0)
has total of 324 144 16 =484 (MM) blocks.
In Z
(1)
, we select the interactions between level 2
groups and their nearest neighbors. For example, we
need to include Z
6252
. However, we need to exclude those
that have been included Z
(0)
. We dene the impedance
matrix primes, for example
Z
/
6
2
5
2
=
Z
19
1
17
1
0
1
Z
19
1
25
1
0
1
Z
20
1
17
1
Z
20
1
18
1
Z
20
1
25
1
Z
20
1
26
1
Z
27
1
17
1
0
1
Z
27
1
25
1
0
1
Z
28
1
17
1
Z
28
1
18
1
Z
28
1
25
1
Z
28
1
26
1
_
_
_
_
(80)
where 0
1
is the zero matrix of dimension (MM). Thus
Z
(1)
includes the entire Z
/
m
2
n
2
where m
2
and n
2
are
neighbors. As shown in Eq. (80), the matrix Z
/
m
2
n
2
consists
of blocks with size of MM, and each block consists of a
transmitting region and a receiving region that are not
neighbors of each other. For example, in Z
19
1
17
1
, the
receiving region is 19
1
and the transmitting region is
17
1
, and the two are not neighbors of each other; also,
the separation R between the transmitting and the receiv-
ing regions is at a minimum of 2S for Z
19
1
17
1
, where S is
the block size. It is at a maximum of R=3
2
_
S for Z
44
11
17
1
that is in Z
/
10
2
5
2
. Thus Z
(1)
includes (1) four interior level 2
groups that have eight neighbors, four of which have 12
blocks and four of which have 15 blocks [the total number
of blocks is 4(4 12 4 15) =432]; (2) eight edge level
2 groups that have ve neighbors, three of which have 12
blocks, while two of them have 15 blocks [the total number
of blocks is 8(312 2 15) =528]; and (3) four corner
level 2 groups that have three neighbors, two of which
have 12 blocks and one that has 15 blocks [the total
number of blocks is 4 (212 115) =156]. Thus ma-
trix Z
(1)
includes 432 528 156=1116 (MM) blocks.
In Z
(2)
, we select the interactions between level 3
groups and their nearest neighbors. For example, we
need to include Z
1
3
2
3
. However, some of the interactions
have already been included in Z
(0)
and Z
(1)
and therefore
3
3 4
3
1
3

2
3
1
2
2
2
3
2
4
2
5
2
6
2
7
2
8
2
12
2
9
2
10
2
11
2
16
2
13
2
14
2
15
2
59
1
51
1

43
1
35
1

27
1
19
1
11
1
3
1
58
1

50
1

42
1

34
1
26
1

18
1

10
1
2
1
60
1

52
1
44
1

36
1
28
1

20
1

12
1

4
1
61
1

53
1

45
1

37
1

29
1

21
1
13
1
5
1
62
1

54
1

46
1

38
1
30
1

22
1

14
1

6
1
57
1

49
1

41
1

33
1

25
1
17
1
9
1
1
1
63
1

55
1

47
1

39
1

31
1
23
1
15
1

7
1
64
1

56
1

48
1

40
1
32
1

24
1

16
1

8
1
(a) (c) (b)
Figure 29. Illustration of multilevel parti-
tioning process with 64 subgroups at the rst
level.
need to be excluded. We dene the prime impedance
matrices, for example
Z
/
1
3
2
3
=
Z
1
2
3
2
Z
1
2
4
2
Z
1
2
7
2
Z
1
2
8
2
0
2
Z
2
2
4
2
0
2
Z
2
2
8
2
Z
5
2
3
2
Z
5
2
4
2
Z
5
2
7
2
Z
5
2
8
2
0
2
Z
6
2
4
2
0
2
Z
6
2
8
2
_
_
_
_
(81)
Note that the building blocks of Z
/
1
3
2
3
are level 2 blocks
and are of size (4M4 M) =16 MM. This is important,
as it indicates that as the level increases, the block size
will grow by 4 times as we go up each level. We have four
level 3 groups, and each of them has two neighbors with 12
level 2 blocks and one neighbor with 15 level 2 blocks.
Thus, we have 4(2 121 15) =156 level 2 blocks.
Since each level 2 block has 16 level 1 blocks, we then have
a total of 16 156 =2496 level 1 blocks in Z
(2)
.
Thus the total count is 4841116 2496 =4096, and
all the level 1 blocks are counted exactly once. To sum-
marize, the block size in Z
(0)
is of MM, in Z
(1)
is of MM,
and in Z
(2)
is of 4 M4M. In applying UV decomposition,
the UV is applied to each block for that level. Each block
consists of a transmitting region and a receiving region
that are not neighbors of each other. However, their
separations are within a restricted range as indicated
before.
5.5. UV Method
The matrix Z
m
i
n
i
, which represents the interactions of two
nonneighbor groups m
i
and n
i
, can be represented by UV
decomposition. The matrix Z
m
i
n
i
is of dimensions
4
(i1)
M(2N
c
1) 4
(i1)
M(2N
c
1). The rank of Z
m
i
n
i
, r, is
much smaller than 4
(i1)
M(2N
c
1). For simpler notation,
we denote Z
m
i
n
i
by A, which has dimension of NN and
rank of r with r{N. To decompose A by the SVD and
GramSchmidt processes will be CPU and memory-inten-
sive because N is large. Note that in Section 5.6, the SVD
is applied to nd r. But because of coarsecoarse sampling,
the selected matrices are roughly of dimension r r only.
To perform UV decomposition, we need only r independent
columns of A and r independent rows of A.
Let the column of A be denoted by " aa
l
, where
l =1; 2; . . . ; N, where N=4
(i1)
M(2N
c
1). The matrix ele-
ment A
mn
is the mth element of the column vector " aa
n
,
A
mn
=( " aa
n
)
m
. Note that the N columns are the number of
cylinders 4
(i1)
M times the number of partial waves
(2N
c
1) in the transmitting region. However there are
only r independent columns. In the transmitting region,
we select r columns from the N columns. The r elements
must be uniformly distributed in the transmitting region.
If the rank r is less than the number of cylinders inside the
transmitting area, we select the zeroth-order partial wave
of r cylinders uniformly distributed. If the rank r is larger
than the number of cylinders 4
(i1)
M, we select the zeroth-
order harmonics of all the cylinders rst, and then select
the rst-order partial wave of 4
(i1)
M uniformly distrib-
uted cylinders.
We compute the r columns " uu
l
, l =1; 2; . . . ; r. Each col-
umn is of dimension of N and coincides with a column of A
U
ml
=( " uu
l
)
m
=A
mp(l)
(82)
where m=1; 2; . . . ; N and p(l) is a column index of A that
depends on l. Note that to get U
ml
, one needs to go through
all the Nelements in the receiving region. Thus the matrix
U has rN elements.
Because of linear independence, any general column " aa
m
of A is a linear combination of " uu
l
, that is
" aa
m
=
r
l =1
v
lm
" uu
l
(83)
where m=1; 2; . . . ; N, and v
lm
are the coefcients to be
determined.
We pick r rows of A, which has total of rNelements. The
r rows correspond to r elements in the receiving group.
The criterion for picking up elements is the same as in the
transmitting area. We rst put these rows in a matrix of R,
its matrix elements are given as R
map
=A
m(ma)p
, with
p=1; 2; . . . ; N and m
a
=1; 2; . . . ; r. We next pick the m
a
rows in " uu
l
, l =1; 2; . . . ; r. That will give us an r r matrix
named
~
UU with
~
UU
m
a
n
a
=( " uu
n
a
)
m(m
a
)
. Then we set
R
m
a
l
=
r
na =1
~
UU
m
a
n
a
v
n
a
l
(84)
with l =1; 2; . . . ; N. In matrix notation, V is of dimension
r N, R is of dimension r N,
~
UU is of dimension r r,
R=
~
UUV, and V =
~
UU
1
R. This completes the UV decompo-
sition of A, and
A=UV (85)
To summarize, we take r columns of A and then r rows of
N, representing a total of 2rNelements. We need to take the
inverse of an r r matrix
~
UU, and a matrix multiplication of
an r r matrix with an r N matrix. The computational
and memory efciency is achieved when r5N.
5.6. Computational Complexity Analysis
5.6.1. Multilevel Group Sizes and Number of Groups.
The centers of the cylinders are generated in a square
area with
N
p
_
points in both x and y directions, such that
the total number of cylinders is N
p
. Note that each
cylinder has a nite radius of a. We use various subgroup
sizes at P level to partition the whole area as follows:
1. At the Pth level, split the whole area as four groups.
Each group has N
p
/4 cylinders. Let L
P
be the num-
ber of groups and M
P
be the number of cylinders of
each group at the Pth level. Thus, L
P
=4 =2
2(PP1)
,
and M
P
=N
p
=4 =N
p
=L
P
.
2. Split each group into four subgroups and continuing
this partitioning, at the ith level, we have
L
i
=2
2(Pi 1)
subgroups and M
i
=N
p
=L
i
cylinders of
each subgroup at the ith level, respectively.
3. At the rst level, we have respectively L
1
=2
2P
level
1 subgroups and M
1
=N
p
/L
1
cylinders of each sub-
group.
5.6.2. Cost Function at ith Level. In Z
(i)
, we select the
interactions between level (i 1) groups and their nearest
neighbors. However, some of the interactions have already
been included in the lower-level groups and must be
excluded from them. There are ve kinds of neighboring
groups:
1. Groups Sharing Only One Common Point. As shown
in Fig. 30a, for the neighboring groups of m
(i 1)
and
n
(i 1)
, only one common point is shared between
them. The impedance matrix that will be calculated is
Z
/
m
(i 1)
n
(i 1)
=
Z
m
(1)
i
n
(1)
i
Z
m
(1)
i
n
(2)
i
Z
m
(1)
i
n
(3)
i
Z
m
(1)
i
n
(4)
i
Z
m
(2)
i
n
(1)
i
0
i
Z
m
(2)
i
n
(3)
i
Z
m
(2)
i
n
(4)
i
Z
m
(3)
i
n
(1)
i
Z
m
(3)
i
n
(2)
i
Z
m
(3)
i
n
(3)
i
Z
m
(3)
i
n
(4)
i
Z
m
(4)
i
n
(1)
i
Z
m
(4)
i
n
(2)
i
Z
m
(4)
i
n
(3)
i
Z
m
(4)
i
n
(4)
i
_
_
_
_
(86)
The dimension of matrix Z
m
i
n
i
is of M
i
(2N
c
1) M
i
(2N
c
1). The computational steps for Z
m
i
n
i
. "
bb
M
i
through the UV is of 2(2N
c
1) M
i
r
i
, where r
i
is
the rank of Z
m
i
n
i
. In here we assume that the rank r
i
of Z
m
i
n
i
is the same for all block interactions at the
same level. There are 15 nonzero matrices of Z
m
i
n
i
in
the Z
/
m
(i 1)
n
(i 1)
. The total computational steps for
Z
/
m
(i 1)
n
(i 1)
. "
bb
/
M
(i 1)
are 30(2N
c
1)M
i
r
i
.
2. Groups Sharing One Common Edge. As shown in
Fig. 30b, for the neighboring groups of m
(i 1)
and
n
(i 1)
, one common edge is shared between them.
The impedance matrix that will be calculated is
Z
/
m
(i 1)
n
(i 1)
=
Z
m
(1)
i
n
(1)
i
0
i
Z
m
(1)
i
n
(3)
i
0
i
Z
m
(2)
i
n
(1)
i
Z
m
(2)
i
n
(2)
i
Z
m
(2)
i
n
(3)
i
Z
m
(2)
i
n
(4)
i
Z
m
(3)
i
n
(1)
i
0
i
Z
m
(3)
i
n
(3)
i
0
i
Z
m
(4)
i
n
(1)
i
Z
m
(4)
i
n
(2)
i
Z
m
(4)
i
n
(3)
i
Z
m
(4)
i
n
(4)
i
_
_
_
_
(87)
There are 12 nonzero matrices of Z
m
i
n
i
in Z
/
m
(i 1)
n
(i 1)
.
The total computational steps for Z
/
m
(i 1)
n
(i 1)
. "
bb
/
M
(i 1)
are 24(2N
c
1)M
i
r
i
.
3. Interior Groups. Interior groups have eight neigh-
bors, four of which share only one common point and
four others that share one common edge. There are
(
L
(i 1)
_
2)
2
interior groups at the (i 1)th level.
Thus computational steps for the interior groups at
the ith level are (
L
(i 1)
_
2)
2
[430(2N
c
1)
M
i
r
i
424(2N
c
1)M
i
r
i
]=216(
L
(i1)
_
2)
2
(2N
c
1)
M
i
r
i
.
4. Edge Groups. Edge groups have ve neighbors; two
of them share only one common point and three of
them share one common edge. There are
4(
L
(i 1)
_
2) edge groups at the (i 1)th level.
Thus computational steps for the edge groups at
the ith level are 4(
L
(i 1)
_
2)(2 30M
i
r
i

3 24M
i
r
i
)(2N
c
1) =528(
L
(i 1)
_
2)(2N
c
1)M
i
r
i
.
5. Corner Groups. Corner groups have three neighbors.
One of them shares only one common point and two
of them share one common edge. There are four
corner groups at the (i 1)th level. Thus computa-
tional steps for the corner groups at the ith level are
4(130M
i
r
i
2 24M
i
r
i
)(2N
c
1)=312(2N
c
1)M
i
r
i
.
Thus the total computational step at the ith level is the
sum of 216(
L
(i 1)
_
2)
2
M
i
r
i
(2N
c
1) for the interior
groups, 528(
L
(i 1)
_
2)M
i
r
i
(2N
c
1) for the edge groups,
n
(i+1)
n
i
(3)
n
i
(1)
n
i
(2)
n
i
(4)
n
i
(3)
n
i
(1)
n
i
(2)
n
i
(4) m
i
(3)
m
i
(1)
m
i
(2)
m
i
(4)
m
(i+1)
n
(i +1)
m
(i +1)
m
i
(3)
m
i
(1)
m
i
(2)
m
i
(4)
(b) (a)
Figure 30. Illustration of interactions of mul-
tiple cylinders within two blocks.
and 312M
i
r
i
(2N
c
1) for the corner groups, which is
equal to
2M
i
r
i
108L
(i 1)
168
L
(i 1)
_
60
_ _
(2N
c
1) (88)
5.6.3. Cost Function for the Multilevel UV Method (Non-
near). The total computational steps for nonneighbor
block interactions through the multilevel UV decomposi-
tion are
P1
i =1
2r
i
M
i
108L
(i 1)
168
L
(i 1)
_
60
_ _
2N
c
1 ( ) (89)
Generally speaking, the rank of each level block will
increase with increase in block size. For simplicity, we
assume r
i
=r to be a constant here, and substituting this r
in Eq. (89), we have
27rN(2N
b
c
1) log
2
N
M
1
_ _
60rN(2N
b
c
1)
120rM
1
(2N
c
1) (90)
5.6.4. Cost Function for Near-Field Interactions in Z
(0)
.
All the near-eld interactions are included in the matrix
Z
(0)
, which is computed directly as the original impedance
matrix. The block size in Z
(0)
is of dimension of (2N
c
1)M
1
(2N
c
1)M
1
. The computational steps for the block and
block interaction are (2N
c
1)
2
M
1
2
.
1. Interior Group. The interior group has nine blocks
(including self-interaction). There are (
L
1
_
2)
2
interior groups at the rst level. Thus the computa-
tional steps for the interior groups at the rst level
are 9(
L
1
_
2)
2
M
2
1
(2N
c
1)
2
.
2. Edge Groups. Each edge group has six blocks (in-
cluding self-interaction). There are 4(
L
1
_
2) edge
groups at the rst level. Thus the computational
steps for the edge groups at the rst level are
6(
L
1
_
2)M
2
1
(2N
c
1)
2
.
3. Corner Groups. Each corner group has four blocks
(including self). There are four corner groups at the
rst level. Thus the computational steps for the
corner groups at the rst level are 4M
2
1
(2N
c
1)
2
.
The total computational step for near-eld interaction
is then the sum of 9(
L
1
_
2)
2
M
2
1
(2N
c
1)
2
for the interior
groups, 6(
L
1
_
2)M
2
1
(2N
c
1)
2
for the edge groups, and
4M
2
1
(2N
c
1)
2
for the corner groups, which is equal to
(9NM
1
12N
1=2
M
3=2
1
4M
2
1
)(2N
c
1)
2
(91)
5.6.5. Cost Function. The total computational step for
multilevel partitioning UVis sum of the near and nonnear
interactions and is given by
27rNlog
2
N
M
1
60rN 120rM
1
_ _
(2N
c
1)
(9NM
1
12N
1=2
M
3=2
1
4M
2
1
)(2N
c
1)
2
(92)
5.7. Numerical Results and Discussion
The numerical simulation results are presented and com-
pared in terms of the matrixvector product and solutions
of the FoldyLax equation with an iterative solver. Simu-
lations are based on the periodic distributions of many
cylinders and the random distribution of many cylinders.
We assume that all the cylinders have the same diameter
of 1 wavelength. The horizontal and vertical separations
between the centers of two neighboring cylinders are
taken as 1.3 wavelengths for periodic distribution. For
the case of random distribution, the center of each cylin-
der is moved randomly within the range from 0.15 to
0.15 wavelength in both horizontal and vertical directions.
The relative dielectric constant of all the cylinders is 3.2 in
Fig. 29 and is equal to 1.5 in Figs. 30 and 31. The computer
used in the simulations is a Pentium single processor of
2.6 GHz with 2 GB (gigabyte) memory.
5.7.1. Rank Determination for FoldyLax Equation. First,
we illustrate the rank determination for the problem. The
cylinders are rst grouped with the multilevel partition-
ing process, and then we know the size of each group. On
the basis of the coarsecoarse boundary sampling, we can
calculate the ranks of interaction groups with the SVD
method. Since the UV is not as strict as the SVD method,
we actually use a rank a little bit higher than that
calculated when constructing the UV matrix.
5.7.2. CPU for the UV. In Table 16, we list the CPU
time based on the UV method for different numbers of
cylinders. The diameter of each cylinder is 1 wavelength,
and the number of partial waves is 9. This means that
there are nine partial waves unknowns to be determined
for each cylinder. The numbers of cylinders are 64, 256,
1024, and 4096. The preprocessing time of the UV method
is the time used to construct UV matrix and also the
time for rank determination. The CPU time for matrix
vector multiplication means the CPU time spent for one
time matrixvector multiplication. For the case of 4096
cylinders, which corresponds to 36,864 partial wave un-
knowns, it takes less than B7.34 s for the matrixvector
multiplication. This case corresponds to 321,700 equiva-
lent numbers of volumetric unknowns.
5.7.3. Comparisons of the MatrixVector Product. The
results for 4096 cylinders are shown in Fig. 31. Because of
the limited computer resource, the direct multiplication is
not possible for this case. Thus we make comparison
between the SVD-based QR method and the UV method.
We use the product fromthe SVD method as the horizontal
values and the product from the UV method as the vertical
values. Figure 31a shows results for the real part and Fig.
31b, is for the imaginary part. We can see that all the
points are along a 11 line, which means that the products
from both methods are matched well for this case.
In Table 16, we also list the relative norm errors for the
matrixvector product compared with direct matrixvec-
tor multiplication for all the cases cited above. The errors
inside parentheses for 1024 cylinders indicate random
distribution. It can be seen that the errors are pretty
small. This means that an accurate matrixvector product
can be reached through the UV method.
5.7.4. Comparisons of the Solutions of FoldyLax Equation
for Different Numbers of Cylinders. Using the proposed UV
method, we have solved the FoldyLax equation with an
iterative solver for randomly distributed cylinders. The
relative dielectric constant of all the cylinders is 1.5 and
the convergent condition for the iterative approach is set
up at 1 10
3
. We have done this for 1024 cylinders. For
this case, we have made comparisons with the solutions
from the SVD-based QR method. Good agreement was
obtained.
In Table 17, we list the CPU time and relative norm
errors for each solution. We also show the number of CG
iterations and CPU time for the rank determination. The
CPU time for rank determination is included in the second
column in brackets. The numbers of iterations are listed in
column 3 and in parentheses. The total CPU for 64
cylinders is 3s with a total of 36s for 256 cylinders. For
1024 cylinders, preprocessing time is 121s, taking only
32 s for rank determination based on the coarsecoarse
sampling. The CPU for iteration method is 711 s, which
needs 330 iterations to reach the convergence. The rela-
tive errors are 0.0014 for 64 cylinders, 0.0078 for 256
cylinders, and 0.0189 for 1024 cylinders. Thus, using the
UV method, we can solve the partial wave equation very
efciently and accurately. However, the number of CG
iterations grows with the number of cylinders. This prop-
erty of the Foldy-Lax partial wave equation needs to be
investigated further.
5.7.5. Bistatic Scattering Coefcients for Different Num-
bers of Cylinders. After the solutions of the FoldyLax
equation, we can calculate the scattering coefcients.
Figure 32 shows the bistatic scattering coefcients for
cylinders totaling 1024. The incidence angle is 301.
The diameter of all cylinders is 1.0 wavelength and the
Periodic distribution
real part
4096 cylinders
15
10
5
0
5
10
15
15 10 5 0 5 10 15
Product of matrix-vector based on the SVD
15 10 5 0 5 10 15
Product of matrix-vector based on the SVD
P
r
o
d
u
c
t

o
f

m
a
t
r
i
x
-
v
e
c
t
o
r

b
a
s
e
d

o
n

t
h
e

U
V
15
10
5
0
5
10
15
P
r
o
d
u
c
t

o
f

m
a
t
r
i
x
-
v
e
c
t
o
r

b
a
s
e
d

o
n

t
h
e

U
V
Periodic distribution
real part
4096 cylinders
(a)
(b)
Figure 31. Comparisons of the matrixvector products calcu-
lated by the SVD-based QR method and the UV method. The
number of cylinders is 4096 with a diameter of 1.0 wavelength.
(This gure is available in full color at http://www.mrw.
Table 16. CPU for the UV Method and Norm Error
Number of
Cylinders
Preprocessing
Time (s)
MatrixVector
Product (s) Norm Error
64 1 0.02 0.0033
256 14 0.13 0.0020
1024 115 1.03 0.0015 (0.0017)
4096 999 7.34 0.0127
Table 17. CPU for the UV-CG Method and Norm Error
Number of Cylinders Preprocessing Time (s) CG Time (s) Total CPU Time (s) Norm Error
64 2 (o1
a
) 1 (21
b
) 3 0.0014
256 15 (4) 21 (75) 36 0.0078
1024 121 (32) 711 (330) 832 0.0189
a
CPU time for rank determination.
b
Number of iterations for the conjugate gradient method.
relative permittivity is 1.5. The cylinders are randomly
distributed.
The UV-MLP method is presented for the rapid solution
of integral equation. Like SMCG and FMM methods, the
UV-MLP method emphasizes on the reduction of both
CPU and memory requirements for matrixvector multi-
plication. On the other hand, the method of characteristic
basis functions [75] emphasizes the drastic reduction in
basis functions.
BIBLIOGRAPHY
1. F. T. Ulaby, R. K. Moore, and A. K. Fung, Microwave Remote
Sensing: Active and Passive, Vols. 1 and 2, Addison-Wesley,
Reading, MA, 1981.
2. L. Tsang, J. A. Kong, and R. T. Shin, Theory of Microwave
Remote Sensing, Wiley-Interscience, New York, 1985.
3. F. T. Ulaby, R. K. Moore, and A. K. Fung, Microwave Remote
Sensing: Active and Passive, Vol. 3, Artech House, Norwood,
MA, 1986.
4. A. K. Fung, Microwave Scattering and Emission Models and
Their Applications, Artech House, Norwood, MA, 1994.
5. L. Tsang, J. A. Kong, and K. H. Ding, Scattering of Electro-
magnetic Waves, Vol. 1: Theory and Applications, Wiley-
Interscience, New York, 2000.
6. L. Tsang, J. A. Kong, K. H. Ding, and C. O. Ao, Scattering of
Electromagnetic Waves, Vol. 2: Numerical Simulations, Wiley-
Interscience, New York, 2001.
7. L. Tsang and J. A. Kong, Scattering of Electromagnetic Waves,
Vol. 3: Advanced Topics, Wiley-Interscience, New York, 2001.
8. J. Guo, L. Tsang, K. H. Ding, A. T. C. Chang, and C. T. Chen,
Frequency dependence of scattering by dense media of small
particles based on Monte Carlo simulation of Maxwells
equations, IEEE Trans. Geosci. Remote Sens. 40:153161
(Jan. 2002).
9. C. T. Chen, L. Tsang, J. Guo, A. T. C. Chang, and K. H. Ding,
Frequency dependence of scattering and extinction of dense
media based on three-dimensional simulations of Maxwells
equations with applications to snow, IEEE Trans. Geosci.
Remote Sens. 41:18441852 (Aug. 2003).
10. J. Guo, L. Tsang, W. Asher, K. H. Ding, and C. T. Chen,
Applications of dense media radiative transfer theory for
passive microwave remote sensing of foam covered ocean,
IEEE Trans. Geosci. Remote Sens. 39:10191027 (May 2001).
11. D. Chen, L. Tsang, L. Zhou, S. C. Reising, W. Asher, L. A. Rose,
K. St Germain, K. H. Ding, and C. T. Chen, Microwave
emission and scattering of foam based on Monte Carlo simu-
lations of dense media, IEEE Trans. Geosci. Remote Sens.
41:782790 (April 2003).
12. L. Zhou, L. Tsang, and D. Chen, Polarimetric microwave
remote sensing of wind vectors with foam covered rough
ocean surfaces, Radio Sci. 38(4):12-112-14 (DOI: 10.1029/
2002RS002764) (Aug. 2003).
13. K. Pak, L. Tsang, and J. T. Johnson, Numerical simulations
and backscattering enhancement of electromagnetic waves
from two-dimensional dielectric random rough surfaces with
the sparse matrix canonical grid method, J. Opt. Soc. Am. A
14:15151529 (1997).
14. C. H. Chan, C. M. Lin, L. Tsang, and Y. F. Leung, A sparse-
matrix/canonical grid method for analyzing microstrip struc-
tures, IEICE Trans. Electron. E80-C:13541359 (1997).
15. S. Q. Li, Y. Yu, C. H. Chan, K. F. Chan, and L. Tsang, A
sparse-matrix/canonical grid method for analyzing densely
packed interconnects, IEEE Trans. Microwave Theory Tech.
MTT-49:12211228 (2002).
16. K. H. Ding and L. Tsang, A sparse matrix iterative approach
for modeling tree scattering, Microwave Opt. Technol. Lett.
38(1):198202 (Aug. 2003).
17. H. G. Wang, C. H. Chan, L. Tsang, and V. Jandhyala, An
improved multi-level matrix QR factorization for large-scale
simulations on magnetoquasistatic analysis of integrated
circuits over multi-layered lossy substrates (in press).
18. L. Tsang, D. Chen, Q. Li, and V. Jandhyala, A fast numerical
solutions of wave scattering problem: Part I: 2D problem of
PEC surface scattering, Radio Sci. (2004).
19. L. Tsang, Q. Li, D. Chen, and V. Jandhyala, Wave scattering
with UV multi-level partitioning method Part II: 3D problem
of non-penetrable surface scattering, Radio Sci. (2004).
20. L. Tsang and Q. Li, Wave scattering with UV multi-level
partitioning method, wave scattering by discrete scatterers,
Microwave Opt. Tech. Lett. 41(5):354361 (June 2004).
21. A. T. C. Chang, J. L. Foster, and D. K. Hall, Nimbus-7 SMMR
derived global snow cover parameters, Ann. Glaciol. 9:3944
(1987).
22. M. T. Hallikainen, F. T. Ulaby, and T. E. V. Deventer, Extinc-
tion behavior of dry snow in the 18- to 90- GHz range, IEEE
Trans. Geosci. Remote Sens. 25:737745 (June 1987).
23. L. Wilson, L. Tsang, J. N. Hwang, and C. T. Chen, Mapping
snow water equivalent in mountainous areas by combining a
spatially distributed snow hydrology model with passive
microwave remote sensing data, IEEE Trans. Geosci. Remote
Sens. 37:690704 (March 1999).
24. S. Rosenfeld and N. C. Grody, Metamorphic signature of snow
revealed in SSM/I measurements, IEEE Trans. Geosci. Re-
mote Sens. 38:5363 (Jan. 2000).
25. C. T. Chen, B. Nijssen, J. Guo, L. Tsang, A. Wood, J. N. Hwang,
and D. P. Lettenmaier, Passive microwave remote sensing of
snow constrained by hydrological simulations, IEEE Trans.
Geosci. Remote Sens. 39:17441756 (Aug. 2001).
26. L. Tsang, C. T. Chen, A. T. C. Chang, J. Guo, and
K. H. Ding, Dense media radiative transfer theory based on
0 300
0
10
200 100
Scattering angle (degrees)
S
c
a
t
t
e
r
i
n
g

c
o
e
f
f
i
c
i
e
n
t

(
d
B
)
10
20
30
40
50
Figure 32. Bistatic scattering coefcients based on the UV
method. The number of cylinders is 1024, with radius of 1.0
wavelength. The incidence angle is 301. (This gure is available in
full color at http://www.mrw.interscience.wiley.com/erfme.)
quasicrystalline approximation with applications to passive
microwave remote sensing of snow, Radio Sci. 35(3):731749
(2000).
27. L. Tsang, C. E. Mandt, and K. H. Ding, Monte Carlo simula-
tions of the extinction rate of dense media with randomly
distributed dielectric spheres based on solution of Maxwells
equations, Opt. Lett. 17(5):314316 (1992).
28. L. M. Zurk, K. H. Ding, L. Tsang, and D. P. Winebrenner,
Monte Carlo simulations of the extinction rate of densely
packed spheres with clustered and nonclustered geometries,
J. Opt. Soc. Am. A 12(8):17721781 (1995).
29. L. M. Zurk, L. Tsang, and D. P. Winebrenner, Scattering
properties of dense media from Monte Carlo simulations with
applications to active remote sensing of snow, Radio Sci.
31:803819 (1996).
30. K. Karkkainen, A. Sihvola, and K. Nikoskinen, Analysis of a
three-dimensional dielectric mixture with nite difference
method, IEEE Trans. Geosci. Remote Sens. 39:10131018
(May 2001).
31. C. Mandt, Y. Kuga, L. Tsang, and A. Ishimaru, Microwave
propagation and scattering in a dense distribution of sphe-
rical particles: experiment and theory, Waves Random Media
2:225234 (1992).
32. C. C. Lu, W. C. Chew, and L. Tsang, The application of
recursive aggregate T-matrix algorithm in the Monte Carlo
simulations of the extinction rate of random distribution of
particles, Radio Sci. 30(1):2528 (1995).
33. Y. Nanbu and M. Tateiba, A comparative study of the
effective dielectric constant of a medium containing randomly
distributed dielectric spheres embedded in a homogeneous
background medium, Waves Random Media 6(4):347360
(1996).
34. B. E. Barrowes, C. O. Ao, F. L. Teixeira, J. A. Kong, and
L. Tsang, Monte Carlo simulation of electromagnetic wave
propagation in dense random media with dielectric spheroids,
IEICE Trans. Electron. E83-C:17971802 (2000).
35. P. R. Siqueira and K. Sarabandi, T-matrix determination
of effective permittivity for three-dimensional dense random
media, IEEE Trans. Anten. Propag. 48:317327 (2000).
36. S. M. Allen and E. L. Thomas, The Structure of Materials,
Wiley, New York, 1999.
37. T. Vicsek, Fractal Growth Phenomena, World Scientic Pub-
lishers, Singapore, 1992.
38. N. A. Seaton and E. D. Glandt, Monte Carlo simulation of
adhesive spheres, J. Chem. Phys. 87(3):17851790 (1987).
39. W. G. T. Kranendonk, and D. Frenkel, Simulation of the
adhesive-hard-sphere model, Mol. Phys. 64(3):403424
(1988).
40. K. H. Ding, L. Tsang, and S. E. Shih, Monte Carlo simulation
of particle positions for densely packed multispecies sticky
particles, Microwave Opt. Technol. Lett. 30:187192 (2001).
41. R. J. Baxter, Ornstein-Zernike relation and Percus-Yevick
approximation for uid mixtures, J Chem. Phys. 52:4559
4562 (1970).
42. J. W. Perram and E. R. Smith, A model for the examination of
phase behavior in multicomponent systems, Chem. Phys. Lett.
35:138140 (1975).
43. J. Shi, R. E. Davis, and J. Dozier, Stereological determination
of dry snow parameters for discrete microwave modeling,
Ann. Glaciol. 17:295299 (1993).
44. C. Matzler and U. Wegmuller, Dielectric properties of fresh-
water ice at microwave frequencies, J. Phys. D 20:16231630
(1987).
45. L. Tsang, Polarimetric passive microwave remote sensing of
random discrete scatterers and rough surfaces, J. Electro-
magn. Waves Appl. 5(1):4157 (1991).
46. S. H. Yueh, R. Kwok, F. K. Li, S. V. Nghiem, W. J. Wilson, and
J. A. Kong, Polarimetric passive remote sensing of ocean wind
vector, Radio Sci. 29:799814 (1994).
47. J. T. Johnson, J. A. Kong, R. T. Shin, S. H. Yueh, S. V. Nghiem,
and R. Kwok, Polarimetric thermal emission from rough
ocean surfaces, J. Electromagn. Waves Appl. 8:4359
(1994).
48. S. H. Yueh, Modeling of wind direction signals in polarimetric
sea surface brightness temperatures, IEEE Trans. Geosci.
Remote Sens. 35:14001418 (1997).
49. D. B. Kunkee and A. J. Gasiewski, Simulation of passive
microwave wind direction signatures over the ocean using an
asymmetric-wave geometrical optics model, Radio Sci. 32:
5978 (1997).
50. A. Stogryn, The emissivity of sea foam at microwave frequen-
cies, J. Geophys. Res. 77:16581666 (1972).
51. P. M. Smith, The emissivity of sea foam at 19 and 37GHz,
IEEE Trans. Geosci. Remote Sens. 26:541547 (1988).
52. X. Z. Huang and Y. Q. Jin, Scattering and emission from two-
scale randomly rough sea surface with foam scatterers, IEE
Proceedings - Microwaves, Antennas and Propagation 142:
109114 (April 1995).
53. G. F. Williams, Microwave emissivity measurements of bub-
bles and foam, IEEE Trans. Geosci. Electron. GE-9:221224
(1971).
54. T. T. Wilheit Jr, A model for the microwave emissivity of the
oceans surface as a function of wind speed, IEEE Trans.
Geosci. Electron. GE-17:244249 (1979).
55. P. C. Pandey and R. K. Kakar, An empirical microwave
emissivity model for a foam-covered sea, IEEE J. Ocean.
Eng. OE-7(3):135140 (1982).
56. W. Asher, Q. Wang, E. C. Monahan, and P. M. Smith, Estima-
tion of air-sea gas transfer velocities from apparent micro-
wave brightness temperature, Mar. Technol. Soc. J. 32(2):
3240 (1998).
57. L. A. Rose, W. Asher, S. C. Reising, P. W. Gaiser, K. M. St.
Germain, D. J. Dowgiallo, K. A. Horgan, G. Farquharson, and
E. J. Knapp, Radiometric measurements of the microwave
emissivity of foam, IEEE Trans. Geosci. Remote Sens.
40:26192625 (Dec. 2002).
58. S. P. Durden and J. F. Vesecky, A physical radar cross section
model for a wind driven sea with swell, IEEE J. Ocean. Eng.
OE-10:445451 (1985).
59. W. C. Chew, J. A. Friedrich, and R. Geiger, A multiple
scattering solution for the effective permittivity of a sphere
mixture, IEEE Trans. Geosci. Remote Sens. 28:207214
(1990).
60. L. A. Klein and C. T. Swift, An improved model for the
dielectric constant of sea water at microwave frequencies,
IEEE Trans. Anten. Propag. 25(1):104111 (1977).
61. S. H. Yueh, J. A. Kong, J. K. Jao, R. T. Shin, and T. Le Toan,
Branching model for vegetation, IEEE Trans. Geosci. Remote
Sens. 30:390402 (1992).
62. W. C. Au, L. Tsang, and J. A. Kong, Absorption enhancement
of scattering electromagnetic waves by dielectric cylinder
clusters, Microwave Opt. Technol. Lett. 7:454457 (1994).
63. Z. Chen, L. Tsang, and G. Zhang, Scattering of electromag-
netic waves by vegetation based on the wave approach and
the stochastic Lindenmayer system, Microwave Opt. Technol.
Lett. 8:3033 (1995).
64. J. He, N. Geng, L. Nguyen, and L. Carin, Rigorous modeling of
ultrawideband VHF scattering from tree trunks over at and
sloped terrain, IEEE Trans. Geosci. Remote Sens. 39:
21822193 (2001).
65. P. Prusinkiewicz and A. Lindenmayer, The Algorithmic
Beauty of Plants, Springer, New York, 1990.
66. G. H. Golub and C. F. Van Loan, Matrix Computations, 3rd
ed., Johns Hopkins Univ. Press, Baltimore, 1996.
67. G. Alleon, M. Benzi, and L. Giraud, Sparse approximate
inverse preconditioning for dense linear systems arising in
computational electromagnetics, Num. Algorithms 16:115
(1997).
68. J. M. Song, C. C. Lu, and W. C. Chew, Multilevel fast multi-
pole algorithm for electromagnetic scattering by large com-
plex objects, IEEE Trans. Anten. Propag. AP-45:14881493
(1997).
69. K. Sertel and J. L. Volakis, Incomplete LU preconditioner
for FMM implementation, Microwave Opt. Technol. Lett.
26:265267 (2000).
70. C. H. Chan and L. Tsang, A sparse-matrix canonical grid
method for scattering by many scatterers, Microwave Opt.
Technol. Lett. 8(2):114118 (Feb. 1995).
71. J. Guo, L. Tsang, A. T. C. Chang, Q. Li, C. C. Huang, and K. H.
Ding, Bistatic phase function and fast solution of scattering
by 2-dimensional random distributed scatterers, Microwave
Opt. Technol. Lett. 38(4):313317 (Aug. 2003).
72. B. E. Barrowes, C. Ao, F. L. Teixeira, and J. A. Kong, Sparse
matrix/canonical grid Method applied to 3-D dense medium
simulations, IEEE Trans. Anten. Propag. 51:4858 (2003).
73. C. C. Lu, J. M. Song, and W. C. Chew, A multi-level fast
multipole algorithm for solving 3D volume integral equations
of electromagnetic scattering, Proc. IEEE Antennas Propaga-
tion Int. Symp., 2000, pp. 18641867.
74. C. C. Lu, Site specic propagation modeling by multilevel fast
multipole algorithm, in B. Beker and Y. Chen, eds., Recent
Research Development in Microwave Theory and Techniques,
Research Signpost, Kerala, India, 2003.
75. Y. F. Sun, C. H. Chan, R. Mittra, and L. Tsang, Characteristic
basis function methods for solving large problems arising
from dense medium scattering, Proc. IEEE Antennas Propa-
gation Int. Symp., Columbus, OH, 2003.
CONFORMAL ANTENNAS
DIPAK L. SENGUPTA
University of Detroit Mercy
Detroit, Michigan
A conformal antenna may be dened as an antenna whose
radiating aperture conforms to the surface of the body on
which it is mounted. Ideally, such antennas are ush
mounted or low prole (i.e., they do not protrude appre-
ciably out of the mounting surface). Basic slot and micro-
strip (patch) antennas are typical examples of conformal
antenna elements. The term conformal array has no
unique definition. Kummer [1] denes it as an array
that is nonplanar. We shall assume here that a conformal
array consists of conformal (or low-prole) antenna ele-
ments placed on a nonplanar surface. The array surface is
not generally at the disposal of the antenna designer and
is often dictated by the specific application. For ground-
based application, a conformal phased array requiring
coverage over 3601 in azimuth (omnidirectional coverage)
or coverage over a hemisphere the array surface may be
cylindrical or spherical, respectively. For conformal arrays
on aircraft, missiles, satellites, and surface ships, the ar-
ray shape may assume another form dictated by the con-
tour of the vehicle. Basic slot and microstrip antennas are
extensively discussed in the literaturefor example, the
textbook by Balanis [2] is a typical reference. These an-
tennas provide ideal performance only when they are
mounted on planar surfaces. During conformal applica-
tion the curvature of the mounting surface can affect their
impedance and radiation properties; such effects must be
taken into account during the design of such antennas.
The need for conformal phased arrays for aircraft and
missile applications, and for ground-based arrays with
omnidirectional coverage in azimuth or complete hemi-
spherical coverage in space, has grown continually with
requirements that emphasize maximum utilization of
available space and minimum cost. Many of the develop-
ments in conformal arrays have been extensions of the
concepts for planar phased arrays, which are extensively
discussed in the literature (for example, in Ref. 2). How-
ever, there are significant differences between planar and
conformal arrays that must be taken into account during
the design of the latter. The individual elements on curved
bodies point in different directions that make it necessary
to turn off those elements that radiate primarily away
from the desired beam direction. For this reason also, one
cannot factor out the element pattern out of the total ra-
diation patternthis makes the conformal array analysis
and synthesis more difcult. The element orientation may
also cause severe crosspolarization. In addition, the mu-
tual coupling effects between the elements can be severe in
some cases.
Within the limitations of space allowed, it is not possi-
ble to describe here every aspect of conformal antennas
and antenna arrays. Instead, we shall at rst describe
briefly certain aspects of a few basic antennas that are
commonly used either singly or as array elements for con-
formal applications. Then we give brief descriptions of a
selected number of conformal antennas and antenna ar-
rays. Specifically, this article describes the following:
1. The specific considerations that must be given to the
performance of basic slot and microstrip or patch
antenna elements when mounted on nonplanar con-
ducting surfaces
2. A selected number of conformal antennas: for exam-
ple, microstrip conformal antennas and dielectric-
lled edge slot (DFES) antennas
3. A class of wraparound antennas and antenna arrays
4. Cylindrical and spherical phased arrays used
for omnidirectional and hemispherical coverage,
respectively.
750 CONFORMAL ANTENNAS
All of these antennas have found practical applications.
Detailed descriptions of their development, design proce-
dures, and analysis of their performance are described in
the references cited at appropriate places.
Literature on conformal antennas is vast and ranges
from technical journal articles to numerous textbooks and
specialized books, of which Refs. 28 are typical.
1. BASIC ANTENNA ELEMENTS
Elementary slot and microstrip patch antennas are com-
monly used singly or as array elements for conformal ap-
plication. However, these radiators provide ideal
performance only when they use plane conducting surfac-
es. Ideal theory can be used when the radii of curvature of
the surfaces are large compared to the operating wave-
length. In other cases both the impedance and radiation
characteristics may be affected significantly.
1.1. Slots on Curved Surfaces
The radiation patterns of slot antennas can be signicant-
ly altered by the curved mounting surface. Pathak and
Kouyoumjian [9] give a convenient extension of the geo-
metric theory of diffraction (GTD) for apertures in curved
surfaces. Figure 1, taken from Ref. 9, shows the patterns
of an axial slot element on perfectly conducting circular
cylinders of various radii; the results indicate the accuracy
of the approximate theory. The effects of the cylinder ra-
dius on the patterns shown in Fig. 1 should be noticed. A
similar slot on a at ground plane would have a constant
pattern from f=0 to 1801. Radiation patterns of slots on a
variety of other generalized surfaces are discussed in Refs.
1012. Mailloux [13] summarizes some of the results of
Pathak and Kouyoumjian [9] shown in Fig. 2, which gives
the radiated power pattern in the upper half-plane
(y_90
) for an innitesimal slot in a cylinder of radius a.

The angular extent of the transition zone is on the order of
(k
0
a)
1/3
on each side of the shadow boundary, where
k
0
=2p/l
0
is the propagation constant in free space. The
results indicate that above the transition zone (i.e., the
illuminated zone) the circumferentially polarized radia-
tion is nearly constant but the axially polarized radiation
has a cos y pattern. Compared with the eld strength in
the y =01 direction, the eld strengths in the y =901 area
are found to be about 0.7 and 0.4(2/k
0
a)
1/3
for circumfe-
rential and axial polarizations, respectively. It should be
noted that in the case of at surface the eld reduces to
zero in the y =901 area.
1.2. Slots on Metallic Cones
Slots and slot arrays on metallic cones are found advan-
tageous to use for missile or missilelike bodies. For ef-
cient design of such arrays, the self- and mutual
admittances must be taken into account. Theoretical and
experimental investigation of slot antennas on metallic
cones are discussed in Ref. 14, where the effects of scat-
tering from a sharp tip on the mutual admittances have
been investigated for pairs of circumferential and radial
slots on a semi-innite metallic cone. The base of the con-
ical model used in the experimental study was terminated
in a spherical cap to minimize scattering from the nite
length of the apparatus. The two slot antennas congura-
tions considered are shown in Fig. 3. Self- and mutual ad-
mittance expressions for pairs of slots shown in Fig. 3 have
been derived by Golden et al. [14], and the results have
been conrmed by measurements. These admittance re-
sults can be immediately applied to determine the aper-
ture voltages required for the analysis of N-element slots
on cones.
In Ref. 14 the circumferential slot results illustrate in-
terference effects between the direct coupling from slot to
slot via the geodesic path over the conical surface and the
tip backscattering. For the radial slot conguration, the
results indicate negligible tip scattering effects. Golden
and Stewart [15] have found that the current distribution
near a slot for a sharp cone can be approximated by the
distribution on an equivalent cylinder if scattering from
the apex (on tip) is small. Thus, the mutual admittance
between two slots can be approximately calculated by us-
ing a cylindrical model with the same local radii of curva-
ture as the cone, provided the wave scattering from either
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0
0 20 40 60 80 100 120 140 160 180
Azimuth angle ( ) degrees
Exact solution
Creeping wave
Surface ray modes
R
e
l
a
t
i
v
e

a
m
p
l
i
t
u
d
e
Slot
k =1.5
k = 2.0
k = 3.0
k = 6.0
Figure 1. Patterns of a thin axial slot on a perfectly conducting

cylinder. (After Ref. 9.)
Axial
polarization
Circumferential
polarization
Transition
zone
f (0) = cos
f (0) = 1
g(0)/2~0.7
0.38
2
k
0
a
( )
1/3
0
Figure 2. Approximate pattern of a thin slot on a conducting
cylinder of radius a; k
0
is the free-space propagation constant.
(After Ref. 6.)
CONFORMAL ANTENNAS 751
the tip or the base region of the vehicle is negligible. The
slotted cone and equivalent cylinder are shown in Fig. 4,
which reveals that the cylinder has a radius equal to the
radius of the circular cross section of the cone midway be-
tween the two slots antennas. For small-angle cones
(y
0
B1801), the radial separation of the slots on the cone
can be equated to the axial separation of the slots on the
equivalent cylinder.
Mutual coupling (|S
12
| parameter) results versus az-
imuthal separation for two circumferential and axial slots
on a cylinder are shown in Figs. 5 and 6, respectively. The
mutual coupling between two radial slots on a 12.21 half-
angle cone is shown in Fig. 7 as a function of frequencies.
Figure 8 shows the mutual coupling versus frequencies for
circumferential slots on a 12.21 (half-angle) cone. The
results illustrate the interference effects between the di-
rect and tip-scattered components. The mutual coupling
between circumferential slots on an 111 (half-angle) cone
is shown in Fig. 9. Using the results given in Ref. 14, it
may be concluded that for the case of circumferential slots
(radial electric elds) the tip-scattered portion of the az-
imuthal magnetic eld at the slot aperture can be ex-
pressed in terms of an appropriate diffraction coefcient;
in the case of radial slots (azimuthal electric elds) there
is no radial component of the magnetic eld in the far eld
of tip and therefore no contribution to the mutual admit-
tance. More detailed results and discussions are given in
Refs. 14 and 15.
1.3. Microstrips on Curved Surfaces
Microstrip or patch is a popular low-prole, ush-mounted
antenna developed in the 1970s. Detailed descriptions of
Circumferential slots
or
Radial slots
b
a
b
a
r
3
r
1
r
2
0
Figure 3. Slotted-cone geometry.

Circumferential slots
or
radial slots
Metal cone
Equivalent
cylinder
Direct ray
Direct ray
Tap-scattered ray
z
0
z
0
2

0
Figure 4. Slotted cone and equivalent cylinder.

0 20 40 60 80 100 120
0
z
0
, deg
Theory
Measurements, z
0
=10.16 cm
Measurements, z
0
=7.62 cm
Measurements, z
0
=5.08 cm
Frequency = 9.0 GHz
70
60
50
40
30
20
10
0
2
0

l
o
g
1
0

S
1
2

,

d
B
Figure 5. Mutual coupling for circumferential slots on cylinder,
r
0
=5.057cm. (After Ref. 14.)
0
z
0
0
, deg
Theory
Measurements, z
0
= 3.81 cm
Frequency = 9.0 GHz
80
70
60
50
40
30
20
10
2
0

l
o
g
1
0

S
1
2

,

d
B
0 30 60 90 120 150 180 210
Figure 6. Mutual coupling for axial slots on cylinder, r
0
=
5.057cm. (After Ref. 14.)
the research and development of microstrip antennas can
be found in Refs. 16 and 17. Such antennas generally use a
metallic patch on a dielectric substrate backed by a planar
ground plane, and they are excited either by a stripline or
a coaxial line. The shape of the patch can be rectangular,
circular, or some other shape, in general, of which the rst
two are the most popular. We shall mostly describe the
basic rectangular patch antenna whose one dimension is
l/2 at the operating wavelength in the substance and the
other dimension is slightly less than the former. Ideally,
such antennas produce similar E- and H-plane patterns
that have maxima in the broadside direction; generally,
the polarization is linear and parallel to the patch plane
but they can be designed to produce circular polarization
also. For conformal applications, it is necessary to take
into account the effects of nonplanar surfaces on the per-
formance of such antennas.
1.4. CylindricalRectangular Patch Antenna
The geometry of a rectangular microstrip patch antenna
mounted on a conducting cylinder is shown in Fig. 10.
Resonant frequencies and radiation characteristics of this
antenna are discussed in Refs. 18 and 19. For thin sub-
strate satisfying h5a, Luk et al. [19] give the following
expression for the (transverse magnetic mode with respect
to r) TM
r
resonant frequencies for the antenna
f
mn
=
c
2

e
r
_
m
2(ah)y
1
_ _
2
n
2b
_ _
2
_ _
1=2
(1)
where c is the velocity of light in free space, e
r
is the
dielectric constant of the substrate, and m, n=0, 1, 2,y,
but m=na0. Equation (1) indicates that if the dimen-
sions of the patchthat is, 2(ah)y and 2bare xed, the
resonant frequencies of the TM
r
modes are not affected by
the curvature of the thin substrate. However, to account
for fringing elds, effective values of the dimensions are to
be used in Eq. (11), as mentioned by Carver and Mink [20].
Luk et al. [19] discuss the E- and H-plane radiation pat-
terns produced by the antenna using e
r
=1.06, e
r
=2.32,
and different values a. It is found that the patterns are not
sensitive to the thickness. For a curved patch, there is
significant radiation in the lower hemisphere for the TM
01
mode; the deviation from the at patch results increases
for larger value of e
r
. Compared to the TM
10
mode, there is
less radiation in the lower hemisphere for the TM
01
mode.
Wong and Ke [21] describe the design of this antenna for
0
45
40
35
30
25
20
2
0

l
o
g
1
0

S
1
2

,

d
B
8.5 9.0 9.5
Cylinder calculations
Cone measurements
= 60.8 deg
c
= 12.2 deg
r
1
= r
2
= 45.53 cm
Figure 7. Mutual coupling for radial slots versus frequency,
r
0
=9.622cm. (After Ref. 14.)
0
2
0

l
o
g
1
0

S
1
2

,

d
B
Cone measurements
r
1
r
2 = = 45.53 cm
r
3 = 92.39 cm
= 60.8 deg
c
= 12.2 deg
8.5 9.0 9.5
90
80
70
60
Figure 8. Mutual coupling for circumferential slots versus fre-
quency, r
0
=9.622cm. (After Ref. 14.)
2
0

l
o
g
1
0

S
1
2

,

d
B
Cone calculations
Cone measurements
0
r
2 =
r
1 = 27.03 cm
25.88 cm
r
3 = 77.47 cm
= 80 deg
c
= 11 deg
80
70
60
50
8.2 8.4 8.6 8.8 9.0 9.2 9.4
Signal frequency, GHz
Figure 9. Mutual coupling for circumferential slots versus fre-
quency; r
0
=5.041cm. (After Ref. 14.)
z
x
a
h
2b
y
w
1
z
a
Figure 10. Geometry of a cylindricalrectangular microstrip
patch antenna.
circular polarization by using the TM
01
and TM
10
modes
excited by a single coaxial feed located on a diagonal line
and the operating frequency chosen between the two low-
est frequencies f
01
and f
10
given by Eq. (1).
Kashiwa et al. [22] describe the application of a strip-
line-fed cylindrically curved rectangular patch antenna as
a small, portable antenna for mobile communication. It
has been found that near the resonant frequency the real
part of the input impedance approaches 50O. The radia-
tion patterns near the broadside direction are found to be
similar to those of the equivalent planar antenna; however,
significant differences have been found in large off-broad-
side directions.
Radiation patterns of a cavity-backed microstrip patch
antenna on a cylindrical body of arbitrary cross section
have been investigated theoretically and experimentally
by Jin et al. [23]. The nite-element method is used to
characterize the microstrip patch antennas, and then the
reciprocity theorem is applied in conjunction with a two-
dimensional method of moments to calculate the radiated
eld. The method can be extended to characterize the ra-
diation patterns of conformal microstrip patch antennas
on general three-dimensional bodies.
1.5. Microstrip Patch Antennas on Conical Surfaces
The use of microstrip antennas on conical surfaces is of
interest for aerospace vehicles with portions of their bod-
ies conically shaped. Performance of a basic rectangular
patch antenna on a metallic cone has been investigated
theoretically by Descardeci and Giarola [24]. In the anal-
ysis the substrate thickness is assumed to be very small
compared with the distance of the patch to the cone apex,
and the curvature radius of the cone surface large com-
pared with the operating wavelength. The capacitive ef-
fects and losses associated with surface waves have been
neglected. Except for these assumptions, the cavity model
analysis used is general and applies to any conical surface.
Within the approximations made, the resonant frequency
is not significantly affected by the conical surface. How-
ever, the radiation pattern is affected, with a consequent
inuence on the input impedance and the total quality
factor. Details can be found in Ref. 24.
2. CONFORMAL ANTENNAS
2.1. The Omnidirectional Microstrip Antenna
The omnidirectional (also known as nondirectional or
omni) microstrip or spiral slot antenna discussed in
Refs. 2527 is essentially a short-circuited quarter-wave-
length microstrip patch wrapped around a cylindrical sur-
face to form a spiral, as shown in Fig. 11. The cylinder is
an epoxy berglass dielectric, and the copper conduction is
added using an electroless platingmaskingelectroplat-
ing technique. The lower end and the insider of the patch
are similarly plated to form a short circuit and ground
plane. The spiral slot antenna has a height and diameter
of 0.06l
0
but, unlike conventional small antennas, has
well-matched input voltage standing-wave ratio (VSWR)
of less than 2: 1 over a 2% bandwidth at 238 MHz. The
radiation patterns are similar to those of a dipole oriented
parallel to the cylinder axis, and the 1 dB gain indicates
an efciency of better than 50%. The spiral slot has also
been developed for 42-MHz application in which the an-
tenna has to be contained in a 0.04l
0
0.15l
0
cylindrical
volume.
2.2. Dielectric-Filled Edge Slot Antenna
A class of circumferential slot antennas, called the DFES
antennas, that are ideally suited for conformal mounting
on conducting bodies of revolution has been described by
Schaubert et al. [28]. As shown in Fig. 12, the simplest
form of the antenna consists of a disk of dielectric sub-
strate that is copper-coated on both sides and mounted
between the two halves of a conducting body so that the
radiating aperture coincides with the surface. The anten-
na is excited at the center by a coaxial line whose outer
conductor is connected to the lower conducting surface,
and the inner conductor is extended through the dielectric
and nally connected to the conducting surface at the up-
per end of the substrate. The input reection coefcients of
the antenna are found to assume minimum values at some
discrete frequencies, called the operating frequencies,
w
h
w
g/4
(a) (b)
Figure 11. (a) Linear short-circuited l
g
/4 microstrip resonator;
(b) omnidirectional microstrip antenna: a cylindrical l/4 micro-
strip resonator. (After Ref. 27.)
Input connector
Copperclad
Inductive posts
plated through
Soldered
Dielectric
(Teflon fiberglass)
Figure 12. Two-element edge slot antenna.
where the antenna also radiates most efciently. The
DFES antenna can be tuned for a desired operating fre-
quency by using a number of axially oriented passive me-
tallic posts. The antenna without the tuning posts is
referred to as the basic DFES antenna, which general-
ly provides the highest operating frequency. By varying
the number and location of the inductive posts, the oper-
ating frequency of the antenna can be tuned over a 6: 1
range; instantaneous bandwidths of 3% are typical. The-
ory and design of basic and tuned DFES antennas mount-
ed on a conducting cylinder have been developed and
discussed by Sengupta and Martins-Camelo [29]. The ra-
diation patterns of edge slot excited conducting bodies of
revolution display a high degree of azimuthal symmetry.
The radiation pattern of DFES antennas is strongly in-
uenced by the body on which it is mounted. The patterns
in Fig. 13 are typical of the performance of the antenna
when mounted on a conducting cylinder.
Sometimes it is not possible to place a at disk across
the body, and at times the antenna must be mounted near
the top of a conical body where the diameter is not suf-
cient to build an antenna operating at the desired fre-
quency. In such cases the planar disk can be deformed
(symmetrically) to t in the available space and to operate
at the required frequency. A conical edge slot antenna is
described in Ref. 28, and its radiation patterns are shown
in Fig. 14. The DFES antenna is a versatile and useful
radiator. Because the azimuthal symmetric radiation pat-
tern can be obtained at any desired frequency within a
very wide range, system designers are not restricted in
their choice of operating frequency. Also, DFES antennas
can be integrated into a variety of structures because their
shape can be varied to conform to the body and the avail-
able space.
2.3. Microstrip Wraparound Antennas
Microstrip wraparound antennas consisting of continuous
metal strips that wrap around missiles, rockets, and sat-
ellites can provide omnidirectional coverage. Various
forms of such antennas are described in Refs. 3034. Mun-
son [30] proposes a continuous radiator for linear polar-
ization, as shown in Fig. 15, which shows that the
microstrip feed network is a parallel (or corporate) feed
network where two-way power splits are equal phase to all
of the feedpoints. The number of power divisions can be 2,
4, 8, 16, and so on. The specific number of feeds and power
divisions required is dictated by the microstrip radiator.
The number of feedpoints N
F
must exceed the number of
wavelengths in the dielectric in the L direction (i.e., N
F
).
The following design relations can be used for the antenna
shown in Fig. 15:
l =
l
0
e
r
_ (2)
w=
l
0
2

e
r
_ =
l
2
(3)
L=pD (4)
L
D
=
L(e
r
)
1=2
l
0
(5)
N
F
> L
D
(6)
Withl
0
=wavelengthinfree space (7)
The pattern coverage of the omnidirectional antenna
shown in Fig. 15 depends on the diameter of the missile. A
270
180
0
90
270
180
0
90
270
180
0
90
270
180
0
90
270
180
0
90
270
180
0
90
0
+
2 dBi
+
2 dBi
+
2 dBi
Relative power Relative power
10 10 20 20 30 30
10 10 20 20 30 30
10 10 20 20 30 30 10 10 20 20 30 30
10 10 20 20 30 30
10 10 20 20 30 30
Relative power
Relative power
Relative power Relative power
15.2 cm
7.6 cm
+
2 dBi
+
2 dBi
+
2 dBi
2 Elements
F = 660 MHz
2 Elements
F = 1210 MHz
4 Elements
F = 2000 MHz
Figure 13. Radiation patterns of a 7.6-cm DFES antenna mount-
ed on a cylinder. (After Ref. 28.)
0 0
90 90
180 180
270 270
3
0
2
0
1
0
3
0
2
0
1
0
7 dBi
Elevation
pattern
E ( )

+0
Azimuthal
pattern
E ( ) 90
F = 6330 MHz
Figure 14. Radiation patterns of four-element edge slot antenna
mounted on conical base. (After Ref. 28.)
typical measured E-plane pattern of a wraparound anten-
na mounted on an 8-in. (203-mm) cylinder given in Ref. 31
is reproduced in Fig. 16. The limiting factor in the omni-
directional coverage is a hole at the tip and tail of the
missile that gets narrower as the diameter of the missile
increases.
Reference 33 studies radiation patterns of wraparound
microstrip antenna on a spherical body for different radii
of the conducting sphere, frequencies, dielectric constant,
and thickness of the dielectric. Specifically, the antenna
studied consists of a metal strip of width d wrapped
around a conducting sphere of radius a covered with a di-
electric substrate of chosen thickness d. A j-symmetric
transverse electromagnetic mode of excitation is used. The
parameter d is kept equal to half a wavelength (l) inside
the dielectric for constructive interference to occur in the
broadside direction. The following comments summarize
the ndings of the investigation reported in Ref. 33:
1. Radiation patterns are almost independent of the
pressure of the patch when the radius of the sphere
is much larger than the strip width.
2. The larger the radius of the sphere (abl
0
), the bet-
ter the omnidirectional pattern.
3. The dielectric constant e
r
does not have significant
inuence on the pattern shape. The radiation inten-
sity tends to increase with increase of the dielectric
constant.
4. The shape of the radiation patterns remains almost
unchanged for different substrate thickness (h) for
h5l
0
. However, sidelobe levels increase with in-
crease of e
r
.
Radiation patterns of rectangular microstrip patches
arrayed circumferentially on a circular cylinder (wrap-
around array) have been computed in Ref. 34. Both axial
and circumferential patches, using axial and circumfer-
ential modes of excitation, respectively, have been used. In
general, it has been found that the number of circumfer-
ential patches required for a given ripple in the gain pat-
tern is considerably less than that required in the axial
case, thus simplifying the feed network for the former
case. Results given in Ref. 34 compared favorably with
reported measurements.
2.4. A Patch Array for Aircraft
A patch array designed for an aircraft to satellite commu-
nication link is described by Sanford [35] and is shown in
Fig. 17. Eight patches are mounted together with the
phase shifting and feeding circuitry to scan the beam in
the elevation direction. Designed for operation at 1.5GHz,
the array, including the radome, is 3.6mm thick. Element
phasing was optimized for maximum multipath rejection
at low scan angles and to account for the curvature of the
mounting surface. Each element was pointed in a different
direction and has an inherent phase error relative to the
center elements. A digital computer was used to determine
how the design parameters actually affect the perfor-
mance of the array. The spacing of the array elements
must be greater than 0.32l (in free space) because the
physical size of the radiating element on Teflon berglass
requires thin space. The spacing required to prevent the
D
Wrapped
to shape
Wrapped
on missile
L = D
W
Input Feed
points
Microstrip
radiator
Feed
network
Figure 15. Microstrip wraparound antenna. (After Ref. 30.)
0 10
20
30
40
50
60
70
80
90
100
110
120
130
140
150
160
170
180
170
160
150
140
130
120
110
100
270
90
350
10
340
20 330
30
320
40
310
50
300
60
290
70
280
80
Figure 16. Measured E-plane pattern of the 8-in. (203-mm)
wraparound microstrip antenna. The antenna pattern is a gure
of revolution about the missile axis. (After Ref. 31.)
Phase shifting
circuitry
Patch
radiators
Aircraft
body
Figure 17. Conformal array for aircraft application. (After
Ref. 35.)
formation of grating lobes is given by
D_
l
1 siny
(8)
where D is the separation distance between the patch el-
ement, l is the wavelength in free space, and y is the
maximum beamsteering angle. For a 501 maximum steer-
ing angle, D must be less than 0.57l.
2.5. Concentric Microstrip Ring Arrays
Bhattacharyya and Garg [36] describe the design of a con-
centric annular ring microstrip antenna array that can be
excited by means of a single feed by interconnecting two
consecutive rings with an impedance transformer. The
feasibility of such an antenna is based on the observation
that annular rings with different mean radii can be de-
signed to resonate at the same frequency for the TM
12
mode. An impedance bandwidth of about 5% for VSWR_2
has been reported in Ref. 36. It has been found possible to
control the principal plane patterns for concentric arrays
independently of each other by appropriately designing
the feed system.
Saha-Misra and Chowdhury [37] describe electromag-
netically fed concentric microstrip ring arrays using the
log-periodic principle that have been reported to have in-
creased impedance and radiation pattern bandwidths.
Specifically, circular, square, and triangular concentric
rings have been investigated. Generally, these antennas
work at multiple bands of frequencies with some bands
having larger bandwidths than standard microstrip an-
tennas. With a nonuniformly spaced concentric annular
ring array, almost 20% bandwidth for VSWR_2 has been
reported. A planar, wideband feed for a slot spiral antenna
has been described by Nurnberger and Volakis [38]. The
antenna has been developed for operation at very high
frequency (VHF) frequencies. In contrast to most tradi-
tional printed spiral antenna designs, the one reported in
Ref. 38 incorporates a completely planar spiral microstrip
balun feed, thereby making it attractive for a variety of
conformal applications.
3. CONFORMAL ARRAYS
Antenna arrays conforming to a nonplanar surface are
suitable and may even be a requirement for a number of
applications. For example, phased arrays of ush-mounted
elements conformally mounted on the surface of an air-
craft or missile reduce the aerodynamic drag and hence
are preferable. Also, in some cases a nonplanar array sur-
face may provide some natural advantage for broad-beam
coverage in space. Spherical, cylindrical, and conical
arrays have been developed for ground, airborne, and mis-
sile applications. We shall consider here the class of con-
formal arrays where the radiating surface is nonplanar
with a radius of curvature large compared to the operating
wavelength. Conformal arrays that are highly curved are
generally difcult to design because of the following
reason [1,3,4,8]:
1. Array elements point in different directions, and so
it is often necessary to switch off those elements that
radiate primarily away from the desired direction of
radiation. This, in turn, requires more sophisticated
switching mechanisms for activation of elements.
2. The fact that element patterns cannot be factored
out of the total radiation pattern makes the analysis
and synthesis of such antennas more complicated.
3. Mutual coupling effects can be very severe and dif-
cult to ascertain.
4. Nonplanar arrangement of elements may give rise
to severe cross-polarization effects.
3.1. Spherical Arrays
Certain applications require phased arrays capable of
steering the beam over a complete hemisphere. For this
requirement a spherical array surface seems to provide
some natural advantage for beam steering. Schrank [39]
discusses the manner in which an array of radiating ele-
ments placed on a sphere provides a natural conguration
for obtaining hemispherical coverage with nearly identical
highly directive beams. A spherical phased array consist-
ing of circularly polarized at spiral antenna elements has
been developed by Sengupta et al. [40,41]. Theoretical de-
sign and other considerations are given in Ref. 42, and
experimental fabrication and results are given in Ref. 41.
As described in Ref. 40, a special element distribution was
obtained from the consideration of icosahedron geometry
resulting in a best possible uniformity of element spacing.
It was found that the array could operate with widely
spaced elements. The special element distribution devel-
oped for this purpose considerably suppressed the grating
lobes in the pattern and thereby made the array signi-
cantly broadband. Figure 18 shows the icosahedron geom-
etry of element locations on a spherical surface. The choice
of circularly polarized elements made the antenna beam
retain its circular polarization fairly well over the entire
range of beam steering directions [39,40]. Experimental
results given in Ref. 41 demonstrate the capability of a
spherical array of 16 at spiral antennas over a frequency
range of 0.63 GHz. The work reported in Ref. 41 used
manual control of phase and illuminated aperture area;
consequently, the results obtained were limited in scope.
However, with the availability of modern sophisticated
Figure 18. Icosahedron geometry of element locations. (After
Ref. 39.)
computer control mechanisms, it seems that such spher-
ical arrays could provide almost complete hemispherical
phased coverage over a broad band.
3.2. Cylindrical Arrays
Conformal elements such as microstrips and slots arrayed
around the circumference of large metal cylinders have
been used to obtain omnidirectional pattern coverage.
Such coverage may also be obtained with the help of wrap-
around antenna, as discussed earlier. References 4246
show that an array of slots equally spaced around the cir-
cumference of a cylinder can produce a pattern with very
low ripple. Croswell and Knop [43] have obtained exten-
sive numerical data using realistic patterns for slots
on perfectly conducting planes. In such arrays the design
parameters are the numbers of elements, radiating ele-
ments, and feed network. The number of elements is cho-
sen to provide a nearly omnidirectional pattern; the
minimum number of elements is determined by the
allowed amplitude ripple. The evaluation of the amplitude
ripple can be given in terms of the uctuation, which is
dened as the ratio of maximum |F| to minimum |F|,
where F is the total far-eld pattern of an S-element cir-
cular array and is given by [42]
F =S
N
n=0
A
n
(j)
n
d
n
d
z
n
[J
0
(z) 2(j)SJ
S
(z) cos Sj] (9)
where J
S
(z) is the Bessel function of the rst kind of order
S, z =k
0
a sin y; k
0
is the free-space wavenumber; a is the
radius of the circular array; y,j are the usual coordinates;
the z axis is the axis of the cylinder; and N_S.
The preceding expression assumes that the single ele-
ment pattern f(j
1
) can be expressed by a Fourier cosine
series
f (j
/
) =
o
n=0
A
n
cos
n
j
/
(10)
A practical single-element pattern can be approximated
by
f (j
/
) =
1 cos j
/
=2
2
2 3cos j
/
cos 2j
/
6
_
_
(11)
Pattern uctuations as a function of size of cylinder and
number of elements for the preceding two single-element
patterns are given in Ref. 43. Cylindrical phased arrays,
where selected sections are illuminated to provide a beam
in a certain direction, are sometimes found advantageous
to use for some requirements. Sophisticated types of elec-
tronic switches for such circular arrays are based on a
concept originally proposed by Shelton [45] and developed
by Sheleg [46]. The antenna uses a Butler matrix-fed cir-
cular array with xed phase shifters to execute current
modes around the array and variable phase shifters to
provide continuous scanning of the radiated beam over
3601. The operation was experimentally demonstrated
with a 32-dipole circular array.
The principles involved in scanning a multimode array
are readily seen by considering a continuous distribution
of current, as described by Sheleg [46]. Figure 19 shows
the conguration of a continuous cylindrical sheet of ver-
tical current elements around a vertical conducting cylin-
der of radius a. Referring to Fig. 19, consider a current
distribution I(a) to be the sum of a nite number of con-
tinuous current modes I
n
e
jna
with N_n_N. The radia-
tion pattern for I(a) =I
n
e
jna
is then given by
E(j) =
N
n=N
C
n
e
inj
(12)
where C
n
are complex constants given by
C
n
=2pKj
n
I
n
J
n
2pa
l
_ _
(13)
where K is a constant, l is the wavelength of operation,
and J
n
is the Bessel function dened earlier. If, in the an-
tenna being considered, it is desired that the pattern mode
be equal in magnitude and be in phase at j=0, the exci-
tation of the current modes must be
I
n
=
1
2pKj
n
J
n
2pa
l
_ _ (14)
Under this condition, the radiation pattern is given by
E(j) =
N
n=N
e
jnj
= sin
2N1
2
j
_ _
sin
j
2
(15)
If the phase difference between the adjacent modes is
j
0
(i.e., multiply I
n
by e
jnj
], the resultant radiation
z
x
y
= 0
J = N1
J = N
J = 1
J = 2
x
x
x
x
Figure 19. Coordinates for continuous cylindrical sheet of verti-
cal current elements.
pattern is
E(j) =
sin[(2N1)(j j
0
)=2]
sin[(j j
0
)=2]
(16)
which indicates a beam in the j
0
direction. As described
by Sheleg [46], it was possible to excite simultaneously
and independently all the modes both positive and nega-
tive n, from 0 to N/2, by connecting a single ring on N
elements to the outputs of a Butler matrix. A schematic
diagram of a scanning multimode array is shown in
Fig. 20. The desired phase and amplitude distribution is
established over the inputs of the Butler matrix by xed
phase shifts and a corporate structure. Once the pencil-
beam pattern is formed at some azimuth angle, it is
scanned just as in a linear array; the mode amplitudes
are held xed and a linear phase progression is set up on
the mode inputs by operating the variable phase shifters.
3.3. Conical Array
Kummer [1] discusses a number of difculties associated
with antenna pattern synthesis utilizing conical surfaces.
An array on a conical surface generally looks different at
different aspect angles; also, the geometry is such that all
elements do not contribute equally to the main beam di-
rection, thereby causing cross-polarization problems. In
spite of this, for their obvious applications to missile and
other similar vehicles, conical arrays have been consid-
ered for conformal array development. Theoretical and
experimental investigations of various aspects of conical
arrays are discussed in Refs. 15, 47, and 48. The experi-
mental studies of Munger et al. [48] provide some data on
the characteristics of several conical arrays. Balzano and
Dowling [47] developed an effective method to evaluate
the pattern of elements in a conical array. The method
takes into account the mutual coupling between array
elements and aperture matching conditions. By properly
matching the array aperture, the radiation in a certain
direction can be substantially increased, thus allowing the
designer to meet specific design goals in the application of
conical arrays to airborne or missileborne systems. More-
over, it has been shown that in some cases, the element
pattern can be approximated by much simpler planar and
cylindrical models.
BIBLIOGRAPHY
1. W. H. Kummer, Preface, IEEE Trans. Antennas Propag.
AP-22:13 (1974).
2. C. A. Balanis, Antenna Theory, 2nd ed., Wiley, New York,
1997.
3. R. C. Johnson and H. Jasik, eds., Antenna Engineering Hand-
book, 2nd ed., McGraw-Hill, New York, 1984.
4. R. C. Hansen, Significant Phased Array Papers, Artech
House, Dedham, MA, 1973.
5. Y. T. Lo and S. W. Lee, eds., Antenna Handbook: Theory,
Applications, and Design, Van Nostrand Reinhold, New York,
1988.
6. R. J. Mailloux, Phased Array Antenna Handbook, Artech
House, Boston, 1994.
7. W. H. Kummer, guest editor, special issue on conformal
arrays, IEEE Trans. Anten. Propag. AP-22:1150 (1974).
8. R. C. Hansen, Microwave Scanning Antennas, Vols. IIII,
Academic Press, New York, 1964.
9. P. H. Pathak and R. G. Kouyoumjian, An analysis of the
radiation from aperture in curved surfaces by the geometrical
theory of diffraction, Proc. IEEE 62:14381447 (1974).
10. P. H. Pathak et al., A uniform GTD solution for the radiation
from sources on a curvex surface, IEEE Trans. Anten. Propag.
AP-29:602622 (1981).
11. W. D. Burnside, R. J. Marhefka, and C. L. Yu, Roll plane
analysis of on-aircraft antennas, IEEE Trans. Anten. Propag.
AP-21:780786 (1973).
12. W. D. Burnside et al., A study of KC-135 aircraft antenna
patterns, IEEE Trans. Anten. Propag. AP-23:309316 (1975).
13. R. J. Mailloux, Conformal and low-prole arrays, in Ref. 3.
14. K. E. Golden, G. E. Stewart, and D. C. Pridmore-Brown,
Approximation techniques for the mutual admittance of slot
antennas on metallic cones, IEEE Trans. Anten. Propag.
AP-22:4348 (1974).
15. K. E. Golden and G. F. Stewart, Self and mutual admittance
of rectangular slot antennas in the presence of inhomoge-
neous plasma layer, IEEE Trans. Anten. Propag. AP-17:
763771 (1969).
16. D. C. Chang, IEEE Trans. Anten. Propag. AP-29:1182
(1981).
17. I. J. Bahl and P. Bhartia, Microstrip Antennas, Artech House,
Norwood, MA, 1980.
18. C. M. Krowne, Cylindrical-rectangular microstrip antenna,
IEEE Trans. Anten. Propag. AP-31:194199 (1983).
19. K.-M. Luk, K.-F. Lee, and J. S. Dahele, Analysis of the
cylindrical-rectangular patch antenna, IEEE Trans. Anten.
Propag. AP-37:143147 (1989).
20. K. R. Carver and J. W. Mink, Microstrip antenna technology,
Power divider
1
2

N
N1
Butler matrix
Equal line
lengths
Variable phase
shifters
Fixed phase
shifters
Figure 20. Schematic diagram of scanning multimode array.
(After Ref. 46.)
21. K.-L. Wong and S.-Y. Ke, Cylindrical-rectangular microstrip
for circular polarization, IEEE Trans. Anten. Propag. AP-41:
246249 (1993).
22. T. Kashiwa, T. Onishi, and I. Fukai, Analysis of microstrip
antennas on a curved surface using the conformal grids FD-
TD method, IEEE Trans. Anten. Propag. AP-42:423427
(1994).
23. J. M. Jin et al., Calculation of radiation patterns or microstrip
antennas on cylindrical bodies of arbitrary cross section,
24. J. R. Descardeci and A. J. Giarola, Microstrip antenna on a
conical surface, IEEE Trans. Anten. Propag. AP-40:460463
(1992).
25. H. S. Jones, Design of Dielectric-Loaded Circumferential Slot
Antennas of Arbitrary Size for Conical and Cylindrical Bod-
ies, Report HDL-TR-1684, Harry Diamond Laboratories,
Adelphi, MD, 1974.
26. D. H. Schaubert, A. R. Sindons, and F. G. Farrar, The spiral
slot: A unique microstrip antenna, Proc. 1978 Antenna Applic.
Symp., Univ. of Illinois Allerton Conf., Harry Diamond Lab-
oratories, Oct. 1978.
27. A. D. Krall et al., The omni microstrip antenna: A new small
antenna, IEEE Trans. Anten. Propag. AP-27:850853 (1979).
28. D. H. Schaubert, H. S. Jones, and F. Reggia, Conformal di-
electric-lled edge-slot antennas with inductive-post turning,
29. D. L. Sengupta and L. F. Martins-Camelo, Theory of dielec-
tric-lled edge-slot antennas, IEEE Trans. Anten. Propag.
AP-28:481490 (1980).
30. K. F. Munson, Conformal microstrip antennas and microstrip
phased arrays, IEEE Trans. Anten. Propag. AP-22:7478
(1974).
31. R. F. Munson, Omnidirectional microstrip arrays, pp. 719,
721 of Ref. 3.
32. S. B. D. A. Fonseca and A. J. Giarola, Analysis of microstrip
wraparound antennas using dyadic Greens functions, IEEE
Trans. Anten. Propag. AP-31:248253 (1983).
33. A. Das, S. K. Das, and M.-S. Narasimhan, Radiation
characteristics of wraparound microstrip antenna on spheri-
cal body, IEEE Trans. Anten. Propag. AP-39:10311034
(1991).
34. I. Jayakumar et al., A conformal cylindrical microstrip array
for producing omnidirectional radiation pattern, IEEE Trans.
Anten. Propag. AP-34:12581261 (1986).
35. G. G. Sanford, Conformal microstrip phased array for aircraft
tests with ATS-6, IEEE Trans. Anten. Propag. AP-26:642646
(1978).
36. A. K. Bhattacharyya and R. Garg, Input impedance of annu-
lar ring microstrip antenna using circuit theory approach,
37. I. Saha-Misra and S. K. Chowdhury, Concentric microstrip
ring antenna: Theory and experiment, J. Electromagn. Wave
Appl. 10:439450 (1996).
38. M. W. Nurnberger and J. L. Volakis, A new planar feed for slot
spiral antennas, IEEE Trans. Anten. Propag. AP-44:130131
(1996).
39. H. E. Schrank, Basic theoretical aspects of spherical
phased arrays, in A. A. Oliner and G. H. Knitted, eds., Phased
Array Antennas, Artech House, Dedham, MA, 1972, pp.
323327.
40. D. L. Sengupta, T. M. Smith, and R. W. Larson, Radiation
characteristics of spherical array of circularly polarized
elements, IEEE Trans. Anten. Propag. AP-16:17 (1968).
41. D. L. Sengupta, J. E. Ferris, and T. M. Smith, Experimental
study of a spherical array of circularly polarized elements,
Proc. IEEE 56:20482051 (1968).
42. T. S. Chu, On the use of uniform circular arrays to obtain
omnidirectional patterns, IEEE Trans. Anten. Propag. AP-
7:436438 (1959).
43. W. F. Croswell and C. M. Knop, On the use of an array of
circumferential slots on a large cylinder as an omnidirectional
antenna, IEEE Trans Anten. Propag. AP-14:394396 (1966).
44. W. F. Croswell and C. R. Cockrell, An omnidirectional micro-
wave antenna for use on spacecraft, IEEE Trans. Anten.
Propag. AP-17:459466 (1969).
45. P. Shelton, Application of hybrid matrices to various multi-
mode and multibeam antenna systems, Proc. IEEE Washing-
ton Chapter P GAP Meeting, March 1965.
46. B. Sheleg, A matrix-fed circular array for continuous scan-
ning, Proc. IEEE 56:20162027 (1968).
47. Q. Balzano and T. B. Dowling, Mutual coupling analysis
of arrays of apertures on cones, IEEE Trans. Anten. Propag.
AP-22: 9297 (1974).
48. A. D. Munger et al., Conical array studies, IEEE Trans.
Anten. Propag. AP-22:3542 (1974).
CONFORMAL MAPPING TECHNIQUES
ERDEM YAZGAN
VOLKAN AKAN
Hacettepe University
Beytepe, Ankara, Turkey
1. INTRODUCTION
Conformal mapping (MP) is the transformation of a region
dened in a complex plane to a distinct region in the ref-
erence plane of another complex variable by using a prop-
er transformation function. Since 1867, conformal
mapping techniques (CMTs) have been used by many
mathematicians and engineers. Many engineering appli-
cations such as boundary value problems and electromag-
netic eld problems in which the components of the elds
are conned in a dened plane can be tackled by using
CMTs. Here the crucial point is that a complex and cum-
bersome structure can be transformed into a simpler and
well-known one. When the transformation is done, an im-
portant characteristic is retained: the angle between lines
of ux and equipotential lines or transversity of elds (e.g.,
electric and magnetic elds). Also, there is a wide area for
CMT applications such as heat, uid ow problems, and
electromagnetic problems. According to the known gener-
al literature as reviewed in detail by Schinzinger [1].
Routh [2] has transformed vibrating membrane structure
into simpler congurations, and in 1900. Carter [3] pub-
lished a note on airgap and interpolar induction as the
rst application of conformal mapping to a real engineer-
ing problem. After then, usage of CMTs began to spread in
760 CONFORMAL MAPPING TECHNIQUES
very large areas such as wave propagation between irreg-
ular boundaries [47], acoustic channels [8], radio propa-
gation around Earth [9,10], reection from waveguide
corners [11], calculation of quasistatic parameters of co-
planar striplines, coplanar waveguides, and microstrip
lines, electrooptical applications, and also resonance fre-
quency calculations of microstrip patch antennas and res-
onators [1234].
Potential and ux functions are dened below with re-
lated mathematical equations, followed by a detailed def-
inition and discussion of the basic properties of conformal
mapping.
In Section 4, the SchwarzChristoffel transformation,
which is used widely for quasistatic models, is dened.
Finally, examples of CM applications in electromagnetic
problems are illustrated and explained.
2. POTENTIAL FUNCTION AND FLUX FUNCTION AS
CONJUGATE FUNCTIONS
The basic mathematical definitions are presented below in
order to introduce the potential function and the ux
function as conjugate functions.
If two complex numbers are considered
z =x jy
r =x jy
(1)
where
j =
1
_
it is very simple to show that z and r satisfy Laplace equa-
tion. The real parts and the imaginary parts of these num-
bers satisfy the Laplace equation.
If F=ujv is a function of z =x jy as
F =G(x jy)
F =G(z)
(2)
This can be rewritten as
@F
@x
=
@G(z)
@z
@z
@x
@F
@x
=G
/
(z)
(3)
Then
@
2
F
@x
2
=G
//
(z) (4)
and for the imaginary part
@F
@y
=
@G(z)
@z
@z
@y
@F
@y
=jG
/
(z)
(5)
and
@
2
F
@y
2
=j
@G
/
(z)
@z
@z
@y
@
2
F
@y
2
=j
2
@G
/
(z)
@z
@
2
F
@y
2
= G
//
(z)
(6)
From Eqs. (4) and (6) one can obtain
@
2
F
@x
2

@
2
F
@y
2
=0 (7)
Examination of these in terms of u and v yields
@
2
F
@x
2
=
@
2
u
@x
2
j
@
2
v
@x
2
(8)
and
@
2
F
@y
2
=
@
2
u
@y
2
j
@
2
v
@y
2
(9)
If Eqs. (8) and (9) are summed, and using Eq. (7), the
following equation is obtained:
@
2
u
@x
2

@
2
u
@y
2
j
@
2
v
@x
2

@
2
v
@y
2
_ _
=0 (10)
Equation (10) shows that
@
2
u
@x
2

@
2
u
@y
2
=0 (11)
and
@
2
v
@x
2

@
2
v
@y
2
=0 (12)
From these equations it can be seen that real and imag-
inary parts of any function of a complex variable x jy
satisfy Laplaces equation.
Briefly, if F is a function of z where F=ujv and z =x
jy, this leads to
V
2
u=0; V
2
v =0 (13)
where u is the potential function and the conjugate func-
tion v is the ux function.
CONFORMAL MAPPING TECHNIQUES 761
The relations between potential and ux functions,
known as the CauchyRiemann equations, are very
important.
Again
F =ujv (F =G(z))
z =x jy
and
@F
@x
=
@u
@x
j
@v
@x
(14)
Also, one can write
@F
@x
=
@G(z)
@z
@z
@x
@F
@x
=
@G(z)
@z
(15)
from which
@u
@x
j
@v
@x
=
@G(z)
@z
(16)
and
@F
@y
=
@u
@y
j
@v
@y
@F
@y
=
@G(z)
@z
@z
@y
@F
@y
=j
@G(z)
@z
so
@u
@y
j
@v
@y
=j
@G(z)
@z
(17)
If Eq. (17) is multiplied by j and then summed with
Eq. (16), the following equation is obtained:
@u
@x

@v
@y
_ _
j
@u
@y

@v
@x
_ _
=0 (18)
Then
@u
@x
=
@v
@y
(19)
@v
@x
=
@u
@y
(20)
Equations (19) and (20) are CauchyRiemann equations.
It must be emphasized that these equations are crucial
points of conjugate functions.
3. CONFORMAL MAPPING
After explaining the conjugate functions for CM, we can
now dene conformal mapping. Kreyszig [35] dened CM
as The mapping dened by an analytic function f
(z)
is
conformal, except at critical points, that is, points at which
the derivative f
(z)
is zero. In other words, if the function is
analytic at point z and f(z) =df/dz a0, then the transfor-
mation is said to be conformal. The most important prop-
erty of CM is illustrated in Fig. 1.
As seen in Fig. 1, following CM, curves 1 and 2 form the
same angle c, the angle between their oriented tangents
at the intersection point. Note that this is the same angle
in both magnitude and direction. One also determine
whether a mapping function is analytic using the
CauchyRiemann equations. If these equations [(19) and
(20)] are satised by a complex function, this complex
function is analytic.
As mentioned in Ref. 1 for CM, scaled factors must be
taken into account. For this case, examination of inni-
tesimal displacements between p and w planes, where
p=x jy and w=f jy, will be benecial:
df - dl
f
or dl
f
=h
f
(f; y)df (21)
dy - dl
y
or dl
y
=h
y
(f; y)dy (22)
For the static case we can use constant lines F and V (see
Fig. 2) as ux lines and equipotentials, and for wave prop-
agation E and H elds can be parallel to these constant
lines. As shown in Ref. 1 for relating curl and divergence
between planes p and w, regardless of time, the curvilin-
ear squares are helpful. So in p plane dl
f
dl
y
is written as
df dy, where dl
f
=dl
y
=dp=
2
_
and d
f
=d
y
=dw=
2
_
:
h
f
=
dl
f
df
=h
y
=
dl
y
dy
=h=
dp
dw
(23)
Here h is a function of p or w and h can be written as
h=
dp
dw
= g
/
wp
(w)
(24)
1
h
=
dw
dp
= g
/
pw
(p)
(25)
It must be noted that CM requires that scaling factors be
equal in all directions because of the CauchyRiemann

Curve 1
Curve 2
Curve 1*
Curve 2*
Figure 1. Conformal mapping for curves 1 and 2 and their
images after transformation.
equations. In order to obtain equal scaling factors in all
directions, m is introduced instead of h as in Ref. 1. For
instance, transformation from the Cartesian coordinates
to the cylindrical coordinates r, f, z (not the same as f
with w=fjy plane) is x =r cos f y =r sin f z =z
h
r
, h
f
, and h
z
are expressed as h
r
=1, h
f
=r and h
z
=1.
Now, lets write m
r
, m
f
, and m
z
such that m directly as-
signs for df a factor 1/r:
m
r
=h
r
; m
f
=
h
f
r
m
z
=h
z
(26)
If axially uniform conguration is taken into account, then
m
z
=1 (27)
and
m
r
=m
f
=m (28)
This is also the general case for any coordinate transfor-
mations in CM. This generalization is shown in Table 1.
If Table 1 is examined carefully, an important point is
seen, namely, that the eld components can be rearranged
as follows after transformation:
E
f
=mE
x
; H
f
=mH
x
E
y
=mE
Z
; H
y
=mH
Z
(29)
where f=x, y =Z, and q =B. It is understood (see Fig. 3)
that the effect of transformation has been reected to
elds and a new set of coordinates have been formed as
f, y, q. As demonstrated in Ref. 1, the new medium ob-
tained has characteristics s, m, and e that remain un-
changed in the w plane, but along the z direction they are
seen as (see also Table 1)
s
z
=m
2
s; m
z
=m
2
m; e
z
=m
2
e
z
(30)
On the basis of this equation, the new medium introduced
is anisotropic. Moreover, it can be inhomogeneous if m is a
function of coordinates [m
(x,y)
or m
(f,y)
]. Some complexities
occur because for this reason. These complexities can be
resolved by simplication of the boundary. In addition,
y
F
1
V
1
V
2
p = x + jy
x
z = g
wp
(w)
w = g
pw
(p)
F
2
F
1
V
1
V
2
F
2
dw
dp
dl
0
0
dl
[
d[
w = [ + i0
d0
[
Figure 2. Transformation between p and w
planes.
Table 1. Field Equations
Equation p Plane, with x, y, z as Coordinates w Plane, with x, Z, z as coordinates
Metric coefcient m
x
=1 m
x
=m
m
y
=1 m
Z
=m
m
z
=1 m
z
=1
V
"
EE @E
z
@y

@E
y
@z
= m
@H
x
@t
@E
z
@Z

@(mE
Z
)
@B
= mm
@H
x
@t
@E
x
@z

@E
z
@x
= m
@H
y
@t
@(mE
x
)
@B

@E
B
@x
= mm
@H
Z
@t
@E
y
@x

@E
x
@y
= m
@H
z
@t
@(mE
Z
)
@x

@(mE
x
)
@Z
= mm
2
@H
B
@t
V
"
HH @H
z
@y

@H
y
@z
=s E
x
e
@E
x
@t
@H
z
@Z

@(mH
Z
)
@B
=s mE
x
em
@E
x
@t
@H
x
@z

@H
z
@x
=sE
y
e
@E
y
@t
@(mH
x
)
@B

@H
B
@x
=smE
Z
em
@E
Z
@t
@H
y
@x

@H
x
@y
=sE
z
e
@E
z
@t
@(mH
Z
)
@x

@(mH
x
)
@Z
=sm
2
E
B
em
2
@E
B
@t
V +
"
DD
e
@E
x
@x
e
@E
y
@y
e
@E
z
@z
=0 e
@(mE
x
)
@x
e
@(mE
Z
)
@Z
em
2
@E
B
@B
=0
V +
"
BB
m
@H
x
@x
m
@H
y
@y
m
@H
z
@z
=0 m
@(mH
x
)
@x
m
@(mH
Z
)
@Z
mm
2
@H
B
@B
=0
Source: Ref. 1.
depending on the features of m, some approximations can
be used in many applications.
If wave propagation is considered, one can write again
the equations in phasor form as shown in Table 1. They
can also written for TE, TM, and TEM cases. Table 2
shows these equations and also includes wave equations.
If Table 2 is examined carefully, the vanishing anisot-
ropy (note that there is no change along the z direction)
can be seen, and this leads to a simplication.
In many problems the main purpose is to nd the char-
acteristic impedance of a transmission line or propagation
constant. For these cases, application of local boundary
conditions will be adequate. There is no need to obtain the
entire solution for the wave propagation. As is known, mis
generally variable but an effective m can be used in such
problems.
4. SCHWARZCHRISTOFFEL TRANSFORMATION
The SchwarzChristoffel transformation opens out the in-
terior of a polygon in a dened plane into half of another
plane, obtaining a simpler structure with a known solu-
tion. Its a powerful method and is often used in transmis-
sion-line problems (coplanar stripline and waveguide).
This method involves a procedure between two structures
(the original structure and a transformed simpler one) in
order to obtain the transformation equation. As shown in
Fig. 4, the polygon has been opened between points A and
B. One must note that the opening point is not one of the
corners. If opening is required at a corner, then this corner
is not taken into account for transformation. Here, the
conformal transformation equation is
w=f (z) (31)
The transformation equation is obtained by using the
well-known relation [36]
dw
dz
=M(z x
1
)
(a
1
=p)1
(z x
2
)
(a
2
=p)1
(z x
3
)
(a
3
=p)1

(32a)
where M is a complex constant, x
n
are real number rep-
resenting points on the x axis of the z plane in Fig. 4b and
satisfying x
1
ox
2
o?ox
N1
ox
N
, and a
N
are interior an-
gles of the polygon in Fig. 4a. Note that the number of
factors must be equal to the number of sides of the polygon
plus constant M (as a factor). Integrating Eq. (32a) with
y
p w
p
0
[ x
Figure 3. Demonstration of changing elds
and coordinates after transformation.
Table 2. Basic Equations for Propagating Wave in Transformed Plane
a
TM Wave TE Wave TEM Wave
E
z
=H
f
=H
y
=0 H
z
=E
f
=E
y
=0 H
f
=E
f
=0
@H
z
@y
=(s jwe)E
f
@E
z
@y
= jwmH
f
@E
z
@y
=0
@H
z
@y
=0
@H
z
@f
= (s jwe)E
y
@E
z
@f
=jwmH
f
@E
z
@f
=jwmH
f
@H
z
@f
= (sjwe)E
y
@E
f
@f

@E
y
@y
=0 @H
f
@f

@H
y
@y
=0
@E
y
@y
=0
@H
y
@y
=0
@E
y
@f

@E
f
@y
= jwmm
2
H
z
@H
y
@f

@H
f
@y
=(s jwe)m
2
E
z
@E
y
@f
= jwmm
2
H
z
@H
y
@f
=(s jwe)m
2
E
z
Wave equation: Wave equation: Wave equation:
@
2
H
z
@f
2

@
2
H
z
@y
2
jwm(s jwe)m
2
H
z
=0 @
2
E
z
@f
2

@
2
E
z
@y
2
jwm(s jwe)m
2
E
z
=0
@
2
H
z
@f
2
jwm(s jwe)m
2
H
z
=0
m=m(f; y) m=m(f; y) m=(f; y)
Source: Ref. 1.
a
With no variation in z plane.
respect to z gives
w=f z ( )
=M
_

N
n=1
(z x
1
)
(an=p
1
)
dz Q
(32b)
where Q is an arbitrary constant. It determines the posi-
tion of the polygon in the w plane. This mapping is known
as the SchwarzChristoffel transformation. Briefly, as
seen from Fig. 4, the interior region of the polygon in the
z plane is transformed into the upper half-plane in the p
plane. Afterward, using another transformation, the sim-
pler structure can be obtained.
5. APPLICATIONS OF CM
Some CM applications in electromagnetic problems are
given below, such as effective resistance calculation, cal-
culation of quasistatic parameters of asymmetric coplanar
waveguide, elliptical and cylindrical striplines, calculation
of resonant frequency of rectangular patch antennas,
Earth-attening procedures for radiowave propagation,
and, nally, bent rectangular waveguides.
5.1. Effective Resistance Calculation of a Current Path
between Two Planar Electrodes
In the example given by Gibbs [36], two electrodes are
semiinnite (longitudinally). Also, between these elec-
trodes there exists a known voltage value. Here the prob-
lem is to nd the effective resistance of the current path
per unit length. This problem is illustrated in Fig. 5.
After transformation the region shown in Fig. 6 is ob-
tained. The points a
/
, b
/
, c
/
, and d
/
are obtained via the
SchwarzChristoffel transformation relation
dw
dz
=M(z a)
a
a
/
p
( )1
(z 1)
a
b
/
p
( )1
(z 1)
a
c
/
p
( )1
(z d)
a
d
/
p
( )1
(33)
where
d= a and a=
1
k
a
a
/ =a
b
/ =a
c
/ =a
d
/ =
p
2
(34)
so
dw
dz
=M(z a)
(1=2)
(z 1)
(1=2)
(z 1)
(1=2)
(z a)
(1=2)
(35)
and
w=M
_
dz
(z
2
1)
1=2
(z
2
a
2
)
1=2
(36)
jv
jy
x =
x = x = +
:
1
:
2
x
1
a d c
O
b
x
4
x
2
x
3
:
3
:
4
A
B
D
C
w-plane
z-plane
(a)
(b)
u
x
Figure 4. The SchwarzChristoffel transformation, (a) polygon
in w plane; (b) transformed structure in z plane.
jy
z-plane
x
a b c d
1/k 1 1 1/k
O
Figure 5. Cross-sectional view of semiinnite conductors in
z plane.
d
jv
w-plane
w = K
u
w = K
w = K+jK w = K+jK
a
b c O
Figure 6. Mapping of electrodes using the SchwarzChristoffel
transformation.
Then, letting k =1/a, the integral can be formed as follows
w=C
_
dz
(1 z
2
)
1=2
(1 k
2
z
2
)
1=2
(37)
where C= kM. Equation (37) is the elliptic integral.
From definition of the complete elliptic integral and its
complementary value, the following equations can formu-
lated:
At
b
/
w=
_
1
0
dz
(1 z
2
)
1=2
(1 k
2
z
2
)
1=2
=K
At
a
/
w=
_
1=k
0
dz
(1 z
2
)
1=2
(1 k
2
z
2
)
1=2
w=
_
1
0
dz
(1 z
2
)
1=2
(1 k
2
z
2
)
1=2
j
_
1=k
1
dz
j(1 z
2
)
1=2
(1 k
2
z
2
)
1=2
(38)
w=K jK
/
Also, points c
/
and d
/
can be expressed as
w= K for c
/
w= K jK
/
for d
/
(39)
If points a, b, c, and d are chosen as a=2, b =1, c = 1,
and d= 2, then, from 1/k =2, we see that k is equal
to 0.5.
In the w plane the resistance between |b
/
a
/
| and |c
/
d
/
|
per unit length is
R=r
c
/
b
/
[ [
b
/
a
/
[ [
(O) (40)
The effective resistance between electrodes is invariant
under CM, so
R=r
c
/
b
/
[ [
b
/
a
/
[ [
R=r
2K
K
/
(O)
(41)
where K is the complete elliptic integral of rst kind and
K
/
is complementary to K. For k =0.5, we obtain
R=1:56 r(O) (42)
The application of CM has been shown above for the
calculation of effective resistance.
5.2. Calculations of quasistatic parameters of asymmetric
coplanar waveguide
In this section we consider another example, which has
been studied via the quasi-TEM approximation [24]. In
this study, the characteristic impedance and the effective
permittivity relations of ACPW (asymmetric coplanar
waveguide) have been derived by the authors. The cross
section of ACPW is shown in Figs. 7 and 8.
In Fig. 7 dielectricair interfaces are assumed to be
magnetic walls. Figure 8 shows a detailed version of Fig. 7
(coordinates have been dened).
Using the method of superposition of partial capaci-
tances, analytical expressions for the total line capaci-
tance C, the characteristic impedance Z
0
, and the effective
dielectric constant e
eff
of quasi-TEM lines can be obtained.
We can begin with calculation of total capacitance per unit
length:
C=C
a
C
d
=C
a
e
r
C
0e
r
=(C
a
C
0e
r
) e
r
C
0e
r
C
0e
r
=C
0
C
0er
(e
r
1)
(43)
Here, C
a
is the air capacitance, C
d
refers to capacitance of
the dielectric part, C
0er
refers to capacitance of the dielec-
tric part without substrate, and C
0
refers to summation of
capacitances C
a
and C
0er
. Now we can calculate C
0
by rst
transforming the upper half-plane of Fig. 8 into the rect-
angular region as shown in Fig. 9. This mapping was done
by the SchwarzChristoffel transformation, the integral
equation of which is
W=
_
t
t
0
dt
(t t
1
)(t t
2
)(t t
3
)(t t
4
)
_ (44)
h
g
2
g
1
Magnetic walls
w
c
r
Figure 7. The original ACPW.
Im (t)
Re (t)
E
H G
F
t
1
t
2
t
3
t
4

0
I
A B C D
c
r
Figure 8. Cross section of the original ACPW in t plane.
Then C
0
is given by
C
0
=2e
0
K(k
0
)
K(k
/
0
)
(45)
and
k
0
=
1 (g
1
g
2
)=w
_
1 (g
1
=w)
_
1 (g
2
=w)
_
(46)
As a second step, the calculation of the dielectric capaci-
tance is given.
Since the dielectric part of the structure is nite, we
rst transform the dielectric region (shaded area) in Fig. 8
is transformed into the lower-half region of Fig. 10 using
the mapping function
z = sinh
pt
2h
_ _
(47)
We can then transform the lower half-plane shown
in Fig. 10 into the parallel-plate capacitor using the
SchwarzChristoffel transformation in the same way
as we did earlier. The integral equation for this transfor-
mation is
w=
_
z
z
0
dz
(z z
1
)(z z
2
)(z z
3
)(z z
4
)
_ (48)
Then
C
0e
r
=e
0
K(k
d
)
K(k
/
d
)
(49)
where
k
d
=
2(G
1
G
2
)=W
(1 G
1
=W)(1 G
2
=W)
(50)
and
W= sinh
pw
4h
_ _
G
1
= sinh
p(w2g
1
)
4h
_ _
G
2
= sinh
p(w2g
2
)
4h
_ _
(51)
Now, the total line capacitance relation per unit length
can be written
C=C
0
C
0er
(e
r
1)
C=2e
0
K(k
0
)
K(k
/
0
)
e
0
(e
r
1)
K(k
d
)
K(k
/
d
)
(52)
and the effective permittivity is given by
e
eff
=
C
C
0
=1
e
r
1
2
K(k
d
)
K(k
/
d
)
K(k
/
0
)
K(k
0
)
(53)
and as seen from Eq. (53), the lling factor is
q =
K(k
d
)
K(k
/
d
)
K(k
/
0
)
K(k
0
)
(54)
Also, the characteristic impedance can be calculated easily
by using the quasistatic approximation as follows:
Z
0
=
60p
e
eff
_
K(k
/
0
)
K(k
0
)
(55)
One must note that the quasi-TEM approximation is valid
only at low frequencies. But, because of their dispersion
characteristics, CPWs and CPSs with cross-sectional di-
mensions that do not exceed the substrate thickness are
slightly sensitive to changes in frequency for almost the
entire microwave region.
Thus, the accuracy of this method shows good agree-
ment with other methods and experimental results that
Im (W)
B
W (plane)
Re (W)
A
C
D
Figure 9. Transformation into a parallel-plate capacitor.
F E
z-plane
Re (z)
Im (z)
z
1
z
2
0
z
3
z
4
G, H
A
c
r
B C D
I
Figure 10. Intermediate mapping for the dielectric region (into z
plane).
are available in the literature. The results can be seen in
Table 3.
Having completed these basic planar problems, we now
study more complicated geometries such as elliptical and
cylindrical transmission lines.
5.3. Calculations of Quasistatic Parameters of Elliptical,
Cylindrical, and Asymmetrical Coplanar Striplines
In Fig. 11, a cross section of an elliptical coplanar stripline
(ECPS) with innitely small arc strips on the outer sur-
face of a dielectric substrate (having relative dielectric e
r
)
is shown. As seen, the two ellipses are given as confocal.
Semimajor and semiminor axes of the elliptical cylin-
ders are a
1
, a
2
, b
1
, and b
2
, respectively.
The focal distance of the two confocal elliptic cylinders
is given as follows:
c =[a
2
1
b
2
1
]
1=2
=[a
2
2
b
2
2
]
1=2
(56)
In order to transform the ECPS in the x plane in Fig. 11
into a cylindrical coplanar stripline (CCPS) in the z plane
as shown in Fig. 12, the following formula is used:
z =
1
c
(x (
x
2
c
2
_
) (57)
The radii of the cylindrical structure in Fig. 12 are
r
1
=
a
1
b
1
a
1
b
1
_ _
1
2
; r
2
=
a
2
b
2
a
2
b
2
_ _
1
2
(58)
By using the mapping function
u=j ln
z
r
2
p
2
(59)
the CCPS is transformed into the asymmetric coplanar
striplines (ACPS) with a nite-boundary dielectric sub-
strate on the u plane as seen in Fig. 13.
After the nite boundary substrate ACPS is obtained as
shown in Fig. 13, the total capacitance per unit length of
this structure is equal to the sum of the free-space capac-
itance C
0
after replacing the dielectric substrate by air
and the capacitance C
d
of the substrate having dielectric
constant (e
r
1) as done in the previous example. Thus,
the structure without the substrate in Fig. 13 is trans-
formed into the rectangular region in Fig. 14 using the
Table 3. Comparison of Characteristic Impedance (in X) Results
a
Cross Section Dimensions Calculated Measured
ACPW w (mm) g1 (mm) g2 (mm) Ref. 24 Ref. 15 Ref. 15
1 747 123 1060 51.77 51.78 51.5
2 737 257 991 59.88 59.88 57.5
3 735 356 843 62.96 62.96 61.1
4 1250 196 1756 53.51 53.53 52.0
5 1248 406 1548 62.37 62.38 62.4
6 1244 575 1386 67.16 67.17 66.3
Source: Ref. 24.
a
Obtained in Refs. 15 and 24, where e
r
=9.9 and h=635mm.
Im()
Re()
a2
b2
0
0
1
0
2
a1
b1
c
r
[ [
Figure 11. Cross-sectional illustration of ECPS.
[ [
0
1
Im (z)
Re (z)
0
r
1
r
2
c
r
0
2
Figure 12. Cross-sectional illustration of CCPS.
SchwarzChristoffel transformation as given in Ref. 26.
Here
W=
_
u
u
0
du
(u u
3
)(u u
4
)(u u
5
)(u u
6
)
_ (60)
C
0
=2e
0
K(k
/
0
)
K(
0
)
(61)
and modular k
0
can be written as
k
0
=
2s(2s w
1
w
2
)
(2s w
1
)(2s w
2
)
_ _
1
2
(62)
K(k
0
) is the complete elliptic integral of the rst kind with
modulus k
0
and k
0
/
=(1 k
0
2
)
1/2
.
In order to compute the dielectric capacitance C
d
, the
dielectric region in Fig. 13 is rst transformed into the
lower-half region in Fig. 15 by using the mapping function
as seen in Ref. 26:
t =sn (K(k)=p)m; k
_ _
(63)
Then, the structure obtained in Fig. 15 is again mapped
into the rectangular region in Fig. 14 using the Schwarz
Christoffel transformation
W=
_
t
t
0
dt
(t t
3
)(t t
4
)(t t
5
)(t t
6
)
_ (64a)
where
t
3
=s
t
w
t2
; t
4
=s
t
t
5
= s
t
; t
6
= s
t
w
t1
(64b)
So
C
d
=e
0
(e
r
1)
K(k
/
d
)
K(k
d
)
(65)
where
k
d
=
2t
4
(t
3
t
6
)
(t
3
t
4
)(t
4
t
6
)
_ _
1
2
(66a)
and
t
3
=s
t
w
t2
; t
4
=s
t
t
6
= s
t
w
t1
(66b)
After this, the total line capacitance C and the
characteristic impedance Z
0
per unit length can be
1
2 3 4 5 6 7
8
h
w
2
S S
Im (u)
Re (u)
jln(r
1
/r
2
)
w
1

Figure 13. Cross-sectional illustration of ACPS.
Im(w)
3
5
4
6
Re(w)
Figure 14. Transformation into a parallel-plate capacitor.
Re(t)
Im(t)
W
t2
1 2 3 6 7 4 5 8
W
t1
S
t
S
t

Figure 15. The intermediate transformation for the dielectric
region.
obtained as
C=2e
0
e
eff
K(k
/
0
)
K(k
0
)
(67)
Z
0
=
60p
e
eff
_
K(k
0
)
K(k
/
0
)
(68)
where the effective permittivity e
eff
is given by
e
eff
=1
e
r
1
2
K(k
0
)
K(k
/
0
)
K(k
/
d
)
K(k
d
)
(69)
The formulation derived is used to see the effect of w, s, h
in the variation of e
eff
and Z
0
of ECPS, CCPS, and ACPS.
Some results for the characteristic impedance and
the effective dielectric constants can be seen in Figs. 16a
and 16b, where nally, the variation of the characteristic
impedance and the effective dielectric constant with
respect to y
1
/(y
1
y
2
) is seen for ECPS and CCPS, respec-
tively.
5.4. Resonant Frequency Calculation of Rectangular Patch
Antennas with Multidielectric Layers
Zhong et al. [31] have reported that the longlasting prob-
lem of analyzing rectangular patch antennas with multi-
dielectric layers has been solved (for calculation of
resonant frequency) by CMT and the transmission-line
model. According to this approach, the microstrip patch
antenna is taken up as a microstrip line. In addition, when
the quasi-TEM wave propagates in this line, a quasistatic
value of e
e
of e
eff
(effective dielectric constant) can be ob-
tained by using CMT [31]. The rectangular patch is shown
in Fig. 17. Using the Wheeler transformation [32] the z
plane in Fig. 17 is mapped into Fig. 18.
After one transformation and approximation (shown in
Figs. 18 and 19, respectively), the quasistatic effective
permittivity is obtained as follows
e
e
=e
r1
e
r2
(q
1
q
2
)
2
e
r1
q
2
e
r2
q
1
e
r3
(1 q
1
q
2
)
2
e
r3
(1 q
1
q
2
q
3
) q
3
(70)
where q
1
, q
2
, and q
3
are the lling factors, which are de-
ned by the ratio of partial areas S
1
, S
2
, and S
3
to the total
area S
c
, respectively, as shown in Figs. 18 and 19.
These can be written for
w
h
12
_1 (wide microstrip line) as
q
1
=
S
1
S
c
=
h
1
2h
12
1
p
4

h
12
w
e
_
ln
2w
e
h
1
sin
ph
1
2h
12
_ _ _
cos
ph
1
2h
12
_ ___
(71)
q
2
=
S
2
S
c
=1 q
1

h
12
2w
e
ln
pw
e
h
12
1
_ _
(72)
q
3
=
S
3
S
c
=1 q
1
q
2

h
12
v
e
2w
e
ln
2w
e
2h
13
h
12
v
e
.
cos
pv
e
2h
12
_ _
sin
pv
e
2h
12
_ _ _ _
(73)
Z
0
(ohm)
320
300
280
260
240
220
200
180
160
0 0.2 0.4
(a)
(b)
0.6 0.8
0
1
/(0
1
+0
2
)
[ = 40
[ = 40
[ = 30
[ = 30
[ = 20
[ = 20
c
eff
2.3
2.2
2.1
2
1.9
1.8
0 0.2 0.4 0.6 0.8
0
1
/(0
1
+0
2
)
Figure 16. Z
0
and e
eff
for ECPS and CCPS versus y
1
/(y
1
y
2
)
[e
r
=3.78, y
2
=301, (a
2
b
2
)/(a
1
b
1
) =3]: (a) characteristic impe-
dance; (b) effective dielectric constant (from Ref. 26).
where the w
e
(effective linewidth) and the quantity V
e
are
given [32] by
w
e
=w
2h
12
p
ln 17:08
w
2h
12
0:92
_ _ _ _
(74)
v
e
=
2h
12
p
tan
1
2p
pw
e
4h
12
(h
13
h
12
)
_ _
(75)
and e
eff
can be determined as mentioned by Zhong et al.
[31] by
e
eff
=e
/
r

e
/
r
e
e
1 P(f )
(76)
This last equation (76) has been developed for a micro-
strip line on one substrate, but here the multilayer line is
taken as a normal line having substrate with e
e
and equiv-
alent relative dielectric constant e
/
r
. The relation between
e
e
and e
/
r
is as below
e
e
=
1
2
[e
/
r
1 (e
/
r
1) A] (77)
A= 1
12h
12
w
_ _
(1=2)
(78)
and
e
/
r
=
2e
e
1 A
1 A
(79)
and P (f), which depends on frequency, is given [33] as
P(f ) =P
1
P
2
(0:1844 P
3
P
4
)10fh
12
[ ]
1:5763
(80)
in which
P
1
=0:27488
0:6315
0:525
(1 0:157fh
12
)
20
_ _
u
/
0:065683 exp(8:7513u
/
)
P
2
=0:33622 1 exp(0:03442e
/
r
)
_
P
3
=0:0363 exp(4:6u
/
)
1 exp
fh
12
3:87
_ _
4:97
_ _ _ _
P
4
=1 2:751
1 exp
e
r
15:916
_ _
8
_ _ _ _
(81)
L
w
(a)
(b)
Im(z) z-plane
c
r0
= 1
c
r3
c
r2
h
13
h
3
h
2
h
1
Re(z)
h
12
c
r1
Figure 17. The original structure: (a) top view; (b) cross-section-
al view (from Ref. 31).
w-plane
u
jv
h
12
S
0
S
1
0 u
e
w
e
/2
S
3
S
2
c
r0 c
r1
c
r2
c
r3
v
e
Figure 18. The boundary curves after transformation (from
Ref. 31).
jv
w-plane
u
0 w
e
/2
S
0
S
1
S
3
S
2
c
r2
c
r3
c
r0
c
r1
h
12
Figure 19. Approximation of the individual dielectric areas in
Fig. 18 (from Ref. 31).
Eventually, the resonant frequency f
r
can be written as
f
r
=
c
2 L2DL ( )

e
eff
_ (82)
Experimental and reported methods [31] are compared in
Table 4.
The comparison of results between calculation and
experiments indicates the evaluation accuracy of CMT
application in calculation of the resonance frequency.
5.5. The Earth-Flattening Procedures in Radiowave
Propagation via CMT
Another application of CM that has been used for Earth-
attening procedures in radio propagation problems by
Richter [9] is discussed here. In this example wave propa-
gation along the spherical Earth is the main subject, but
Earth is assumed to be a cylinder; that is, Earths curvature
transverse to the direction of propagation can be
neglected. Also, it is mentioned in this article that this is
a frequently valid assumption in practice. The mapping
function for the cylinder can be applied to the spherical
geometry, thus significantly simplifying the differential
equations. Figures 20 and 21 illustrate this transformation.
Specifically, this mapping is the transformation from
the polar into the Cartesian coordinate system by
q=jb
2
ln
p
b
1
_ _
or p=b
1
exp
jq
b
2
_ _
(83)
where
j = 1 [ ]
1=2
; p=Be
jj
; x =Bj (84)
and
q=x jz (85)
Then
z =b
2
ln
z
b
1
and x =b
2
j (86)
For z =a, z =0, so b
1
=a, and if z =ha for z =h, then
b
2
=
h
ln 1
h
a
_ _ (87)
In Fig. 20 the p plane is considered as a Riemann surface
and its range is 0_j_2p; in the q plane the range is
0 _xo2pb
2
. The CM relation between p and q is
F(x; z) =
dz
dz
=
dx
dx
(88)
and
F(z) =
a
h
ln 1
h
a
_ _ _ _
1
h
a
_ _
z=h
(89)
Here, it must be noted that F is only a function of z and is
independent of x. Therefore this case simplies the solu-
tion for wave propagation.
Table 4. Comparison between Experimental Results and
Results of Present Method Where L=33.25, h
1
=0, h
2
=3.18,
w=32.25, e
r2
=e
r3
=2.32
Cover Thickness h
3
(mm)
Experiment f
r
(GHz)
Calculated f
r0
(GHz)
3.18 2.749 2.718
6.36 2.704 2.701
9.54 2.682 2.688
12.72 2.670 2.678
Source: Ref. 31.
h
a

p-plane
Figure 20. Polar coordinate system (original structure) (from

Ref. 9).
z
h
x
q-plane
Figure 21. Cartesian coordinate system after transformation
(from Ref. 9).
In Eq. (89) F(z) has been expanded for (
h
a
){1 in Ref. 9 as
follows:
F(z) =1
z
a
h
2a
(90)
Now, TE and TM waves, which are guided by Earth, can
be examined by using CM. Here, Earth is assumed to be a
perfectly conducting cylinder surrounded by an isotropic,
nonconducting medium, and its permittivity e(z) is a func-
tion of only z.
5.5.1. TM Case. In p plane the (Fig. 20) one can express
the components of the TM wave as E
z
; E
x
and H
Z
and in
the q plane (Fig. 21) after transformation these compo-
nents can be written as E
z
, E
x
, and H
y
, respectively. It
must be noted that the coordinates Z and y are perpendic-
ular to these planes (p plane and q planes, respectively).
The following partial-differential equation can be reached
using Maxwell equations for E
z
; E
x
; H
Z
:
F
2
(z) =m(z); e(z) =e
/
(z)
H
Z
=H
y
; F(z)E
x
=E
x
; F(z)E
z
=E
z
(91)
Then
d
2
H
y
(z)
dz
2

1
e
/
(z)
de
/
(z)
dz
dH
y
(z)
dz
[b
2
0
e
/
(z)F
2
(z) b
2
] H
y
(z) =0
(92)
where
b=
2p
l
; b
2
0
=w
2
m
0
e
0
; and l =wavelength (93)
5.5.2. TE Case. The TE wave components are
H
z
; H
x
; andE
Z
in p plane and H
z
; H
x
; andE
y
in the q
plane. Using the following relations for the TE wave
e(z) =e
/
(z); F
2
(z) =e
//
(z)
E
y
=E
Z
; H
x
=F(z)H
x
; H
z
=F(z)H
z
(94)
then similarly
d
2
E
y
(z)
dz
2
[b
2
0
e
/
(z)F
2
(z) b
2
]E
y
(z) =0 (95)
Equation (92) indicates that following transformation in
Fig. 21, the permeability m(z) =K
2
(z) and permittivity e
/
(z)
are space-dependent.
Richter [9] has explained the meaning of K
2
(z), in
which the elements of volume of both systems are differ-
ent; however, the magnetic eld strengths are equal.
Thus, there exists a ctitious permeability so that the to-
tal magnetic energy would be equal in both coordinate
systems. In addition, in Eq. (95) there is space-dependent
permittivity [i.e., e
/
(z)
.
e
//
(z)]. Here again, the reason is the
same as the previous one for introducing a ctitious per-
mittivity.
If Eqs. (92) and (95) are written in another form where
b=b
0
F(h) cos y
h
(96)
then y
h
is the angle of the wave normal to the x axis at
height h. Therefore, Eqs. (92) and (95) can be written as
follows:
d
2
f
dz
2
B
2
0
t
2
f =0 (97)
It must be noted that e
[
(z) is assumed to be a constant.
Here t is
t
2
=F
2
(z)
b
b
0
_ _
2
=F
2
(z) F
2
(h) cos
2
y
h
(98)
Finally, the differential equations can be solved by using
different methods as mentioned in Ref. 9.
5.6. Bent Rectangular Waveguide
The procedure of analyzing bent rectangular waveguide
has been summarized [9] by using CM. The analyzed
waveguide geometry in z =x jy plane is seen in Fig. 22.
As mentioned in this article, in one case the wave travels
in the plane of irregularity. The transformed structure in
the w=ujv plane is shown in Fig. 23.
The wave equation in the transformed plane is
@
2
Q
@u
2

@
2
Q
@v
2
m
2
k
2
Q=0 (99)
where Q represents the electrical or magnetic components
of the wave and the metric coefcient m (see Section 3) is
m=1g(u; v) = f
/
(ujv)
2
p
2
=b
2
1g(u; v) = cosh u cos v [ ]
2a
cosh(u t) [ ]
a
cosh(ut) cos y [ ]
a
(100)
N
5
= e
j (2:)
N
3
=
2:
:
:
N
4
N
0
N
1
N
2
N
6
b
b
y
x
Figure 22. The original bent waveguide in z plane (from Ref. 9).
Here f
/
(ujv) denotes the rst derivative of the transfor-
mation function f(w) (obtained using the SchwarzChrist-
offel transformation in Ref. 9), 2pa is the total angle of the
bend, and t is a parameter that depends on a and the ratio
[N
4
N
0
[=[N
4
N
6
[. Also, k can be written as
k=
2b
l
1
l
0
2a
_ _
2
_ _
1=2
= j G
b
p
G=j2pl
1
0
(1 l
2
0
a
2
=4)
1=2
g
2
i
=i
2
k
2
; i =0; 1; 2; 3; . . . ; g
0
=jk
(101)
where l
0
is the free-space wavelength and a is the largest
of the rectangular cross section of the waveguide. This
application method can be used in other irregularities
such as inserts in the channel as argued in Ref. 34.
6. SUMMARY
A limited number of examples of conformal mapping tech-
niques (CMTs) have been presented in this article. How-
ever, as seen from each example, analytical closed-form
expressions have been obtained easily via these CMTs.
These equations can be implemented in CAD tools accu-
rately. Other methods such as the nite-difference meth-
od, integral equation method, and full-wave analysis are
unwieldy, since they require excessive amounts of com-
puter memory and calculation time.
As shown from a comparison of the results of conformal
mapping with other methods and measured values from
experiments reported in the general literature, accuracy
of the CM techniques is very good. Also, it has been men-
tioned many articles that since this method is analytical,
the relation between parameters to be calculated and
physical parameters can be interpreted easily.
BIBLIOGRAPHY
1. R. Schinzinger, Conformal transformations in the presence of
eld components along a third axispart I, Int. J. Electric.
Eng. Educ. 13:7689 (1976).
2. E. J. Routh, The advanced part of a treatise on the dynamics
of a system of rigid bodies, Part II of a Treatise on the Whole
Subject, 6th ed., Dover, 1905, pp. 461467.
3. F. W. Carter, A note on airgap and interpolar induction,
J. Inst. Electric. Eng. 29:925 (1900).
4. H. H. Meinke, Die Anwendung der konformen abbildung auf
Wellenfelder, Z. Angew. Physik. 1:245252 (1949).
5. H. H. Meinke, Ein allgemeines Loesungsverfahren fuer inho-
mogene Zylindersymmetrizche Wellenfelder, Z. Angew. Phys.
1:509516 (1949).
6. R. Piloty Jun, Die Anwendung der konformen Abbildung auf
die Feldgleichungen in inhomogenen Rechteckrohren, Z.
Angew. Phys. 1:441448 (1949).
7. H. H. Meinke, Das Verhalten elektromagnetischer Wellen in-
stark inhomogenen Leitungsbauelementen, Z. Angew. Phys.
2:473478 (1950).
8. P. A. Laura, Calculations of eigenvalues for uniform uid
waveguides with complicated cross sections, J. Acoust. Soc.
Am. 42:2126 (1967).
9. J. H. Richter, Application conformal mapping to Earth-
atenning procedures in radio propagation problems, Radio
Sci. 1:14351438 (1966).
10. R. A. Pappert, A numerical study of VLF mode structure and
polarization below an anisotropic ionosphere, Radio Sci.
3:219233 (1968).
11. S. O. Rice, Reection from corners in rectangular wave
guidesconformal transformation, Bell Syst. Tech. J.
28:104135 (1949).
12. W. Hilberg, From approximations to exact relations for char-
acteristic impedances, IEEE Trans. Microwave Theory Tech.
MTT-17 (5):259265 (May 1969).
13. C. Veyres and V. F. Hanna, Extension of the application of
conformal mapping technique to coplanar lines with nite
dimensions, Int. J. Electron. 48(1):4756 (1980).
14. V. F. Hanna and D. Thebault, Analysis of asymmetrical
coplanar waveguides, Int. J. Electron. 50 (3):221224 (1981).
15. V. F. Hanna and D. Thebault, Theoretical and experimental
investigation of asymmetric coplanar waveguides, IEEE
Trans. Microwave Theory Tech. MTT-32(12):16491651
(Dec. 1984).
16. G. Ghione and C. Naldi, Analytical formulas for coplanar
lines in hybrid and monolithic MICs, Electron. Lett. 20(4):
179181 (Feb. 1984).
17. G. Ghione and C. Naldi, Coplanar waveguides for MMIC
applications: Effect of upper shielding, conductor backing,
nite extent ground planes, and line-to-line coupling, IEEE
Trans. Microwave Theory Tech. MTT-35(3):260267 (March
1987).
18. G. H. Owyang and T. T. Wu, The approximate parameters of
slot lines and their complement, IRE Trans. Anten. Propag.
4955 (Jan. 1958).
19. G. Ghione, A CAD-oriented analytical model for the losses of
general asymmetric coplanar lines in hybrid and monolithic
MICs, IRE Trans. Microwave Theory Tech. 41(9):14991510
(Sept. 1993).
20. L. Zeng, Y. Wang, Accurate solutions of elliptical and cylin-
drical striplines and microstrip lines, IEEE Trans. Microwave
Theory Tech. MTT-34(2):259264 (Feb. 1986).
21. J. Svacina, Analaysis of multilayer microstrip lines by a con-
formal mapping method, IEEE Trans. Microwave Theory
Tech. 40(4):769772 (April 1992).
22. S. Rawal, D. R. Jackson, An exact TEM calculation of loss in a
stripline of arbitrary dimensions, IEEE Trans. Microwave
Theory Tech. 39(4):694699 (April 1991).
23. R. N. Simons, Coplanar Waveguide Circuits, Components,
and Systems, Wiley, New York, 2001.
u u
v = 0
v =
u = 0
u = t u = t
N
5
N
1
N
2
N
3
N
4

Figure 23. The waveguide after transformation [z =x jy =
f(ujv) =f(w)) in w plane (from Ref. 9).
24. C. Karpuz, A. Goru r, H. Goru r, and M. Alkan, Fast and simple
analytical expressions for quasistatic parameters of asym-
metric coplanar lines, Microwave Opt. Technol. Lett. 9(6):
334336 (Aug. 20, 1995).
25. A. Goru r, C. Karpuz, and M. Alkan, Quasistatic TEM char-
acterisitcs of overlayed supported asymmetric coplanar wave-
guides, Int. J. Microwave Millimeter Wave Comput. Aided
Eng. 6 (5):297304 (1996).
26. V. Akan and E. Yazgan, A simple formulation for quasistatic
parameters of elliptical, cylindrical and asymmetrical copla-
nar strip lines, Microwave Opt. Technol. Lett. 41(1):1821
(April 5, 2004).
27. V. Akan and E. Yazgan, Quasistatic TEM characteristics
of multilayer elliptical and cylindrical coplanar waveguides,
Microwave Opt. Technol. Lett. 42(4):317322 (Aug. 20, 2004).
28. G. Ghione, M. Goana, G. Madonna, G. Omegna, M. Pirola, S.
Boss, D. Frassati, and A. Perasso, Microwave modelling and
characterization of thick coplanar waveguides on oxide-coat-
ed lithium niobate substrates for electrooptical applications,
IEEE Trans. Microwave Theory Tech. 47(12):22872293 (Dec.
1999).
29. M. Goana, F. Bertazzi, P. Caravelli, G. Ghione, and T. A. Dris-
coll, A general conformal-mapping approach to the optimum
electrode design of coplanar waveguides with arbitrary cross
section, IEEE Trans. Microwave Theory Tech. 49(9):1573
1580 (Sept. 2001).
30. J. T. Bernhard, and C. J. Tousignant, Resonant frequencies
of rectangular microstrip antennas with ush and spaced
dielectric superstrates, IEEE Trans. Anten. Propag. 47(2)
(Feb. 1999).
31. S. S. Zhong, G. Liu and G. Qasim, Closed form expressions for
resonant frequency of rectangular patch antennas with mul-
tidielectric layers, IEEE Trans. Anten. Propag. 42(9):1360
1363 (Sept. 1994).
32. H. A. Wheeler, Transmission-line properties of parallel wide
strips by a conformal-mapping approximation, IEEE Trans.
Microwave Theory Tech. MTT-12:280287 (March 1964).
33. M. Kirschining, R. H. Jansen, Accurate model for effective
dielectric constant of microstrip with validity up to millime-
ter-wave frequencies, Electron. Lett. 18:272273 (March
1982).
34. R. Schinzinger, Conformal transformations in the presence of
eld components along a third axispart II, Int. J. Electric.
Eng. Educ. 13:127131 (1976).
35. E. Kreyszig, Advanced Engineering Mathematics, Wiley,
Singapore, 1993.
36. W. J. Gibbs, Conformal Mappings in Electrical Engineering,
Chapman & Hall, London, 1958.
CONICAL ANTENNAS
SABURO ADACHI
Tohoku Institute of Technology
Sendai, Japan
This article describes characteristics such as input imped-
ance, radiation pattern, and directivity gain of conical an-
tennas consisting of solid conducting cones, conducting
conical plates, or their modications. Conical conducting
structures on which simple antenna elements such as di-
pole and a slot are mounted are also described. An impor-
tant general feature of conical antennas is their lack of
sensitivity to frequency variation, that is, their broadband
characteristics. Note that a conical horn antenna is de-
scribed in HORN ANTENNAS.
The history of the practical use of conical antennas is
long. Sir Oliver Lodge constructed a biconical antenna in
1897 and made a wireless communication experiment,
while a single cone antenna on the ground and a fan
(at triangular) antenna were used by Marconi and
others. The history of the theory of conical antennas is
also long. The spherical coordinate is one of the few coor-
dinates for which the method of variable separation in
electromagnetic eld problems can be applied. The conical
surface is dened by a constant polar angle y =c in the
spherical coordinate system. For this reason conical an-
tennas and biconical antennas, in particular, have been
extensively investigated by Schelkunoff [1,2], Smith [3],
Tai [4,5], Papas and King [6], and many others.
In the limit as c-0 and p, the biconical antenna is re-
duced to a vanishingly thin linear antenna that is very
sensitive to frequency (that is, of narrow bandwidth). As
the cone angle increases, the antenna shows broadband
characteristics that make it useful for practical applica-
tions.
Figure 1 shows an innitely long, symmetric biconical
conductor with a half-cone angle c and an innitesimally
small feeding gap (d gap). The antenna is assumed to be
located in free space. This structure can support the trans-
verse electromagnetic (TEM) transmission line mode, that
is, the outward-propagating principal spherical-wave
mode expressed by
H
f
=
1
r sin y
H
0
e
jbr
(1)
E
f
=ZH
f
(2)
where b =2p/l (l: wavelength) is the free-space wavenum-
ber and Z =
m
0
=e
0
_
is the free-space wave impedance. The
P(r, , )
y
z
x
Figure 1. Innite biconical conductor fed by a d-gap generator.
CONICAL ANTENNAS 775
characteristic impedance K of the biconical transmission
line is given by the ratio of the transmission voltage (i.e.,
the integral of E
y
along the cone meridian) to the conduc-
tion current along the cone as follows:
K =
Z
p
ln cot
c
2
_ _
(3)
Figure 2 shows the characteristic impedance K versus the
half-cone angle c.
1. BICONICAL ANTENNAS
Figure 3 shows the geometry of the biconical antenna. The
conducting antenna surfaces are dened by the conical
surfaces y =c and y =pc, and the two spherical end
surfaces at r =a. The analytical procedure of the biconical
antenna will be outlined below. In region I, the electric and
magnetic elds are represented as a sum of the outward-
and inward-traveling TEM principal modes and an in-
nite number of complementary (higher) transverse mag-
netic (TM) modes. In region II, the elds are represented
by an innite series of complementary radiating modes.
Boundary conditions on the aperture indicated by the
dashed lines in Fig. 3 and the end surfaces of the cone at
r =a are used to obtain an innite set of linear algebraic
equations from which the amplitudes of the complemen-
tary modes and the principal mode reected back at the
aperture are determined.
The input admittance of the biconical antenna is rep-
resented by the equivalent transmission line circuit as
shown in Fig. 4, where K is the characteristic impedance
given by Eq. (3). The terminal admittance Y
t
represents
the effect of the truncation of the biconical transmission
line at r =a, that is, the transformation of the outward-
traveling TEM mode into the complementary modes in
both regions and the reected TEM mode, which eventu-
ally determines the input admittance of the biconical an-
tenna Y
i
.
Schelkunoff [2] has formulated the abovementioned
boundary value problem rigorously and has discussed in
detail special cases of a vanishingly thin antenna and a
very wide-angle cone, or a spherical antenna with a very
narrow equatorial gap. Tai [4] has obtained the exact an-
alytical solution of the terminal admittance of the vanish-
ingly thin antenna, which has been found to be identical to
the expression obtained ingeniously by Schelkunoff.
Tai [5] has also made an important contribution to the
development of the theory for biconical antennas by ap-
plying Schwingers variational method. He has given the
rst-order variational numerical solution for the specific
wide cone angles c=39.231, 57.431, and 66.061.
The more recent development of computers has made
feasible the numerical solution of Schelkunoff s formula-
tion. However, it is still not easy to solve the innite set of
linear determining equations with reasonable accuracy
because of slow convergence of the innite series when the
cone-angle decreases. For example, 15 or more modes for
c=51 [7], and 13 modes for c=51 [8] are necessary for
computation of the input impedance. A conical monopole
above an image plane driven by a coaxial line has been
numerically analyzed by using the nite difference time
domain method [9].
When the upper half-cone of the biconical antenna is
mounted on an innite conducting plane (ground plane),
the antenna forms a conical unipole having one half of the
input impedance of the biconical antenna. Figures 5 and 6
(from Brown and Woodward, Jr. [10]) show respectively
the measured input resistance and reactance of a conical
unipole having at caps instead of spherical caps. It is
clear that the antenna tends to have a constant input
C
h
a
r
a
c
t
e
r
i
s
t
i
c

i
m
p
e
d
a
n
c
e

K

(

)
800
700
600
500
400
300
200
100
0
10 0 20 30 40 50 60 70 80
Half-cone angle (deg)
K = 120 In cot
2
Figure 2. Characteristic impedance of a biconical transmission

line.
a
Figure 3. Symmetric biconical antenna.
K
a
Y
t
Y
i
Figure 4. Equivalent circuit of a symmetric biconical antenna.
776 CONICAL ANTENNAS
resistance and a small reactance around zero versus fre-
quency, showing broadband characteristics as the cone
angle is increased.
The radiation pattern of the biconical antenna has been
computed by Papas and King [6] and by Bevensee [11].
Figure 7 shows the far-zone electric eld pattern [6] for
the cone angle of c=301. It is found that the patterns are
not much different from those of a straight wire antenna.
Theoretical analysis of biconical antennas loaded with
and/or immersed in dielectric, lossy, and ferromagnetic
materials has been provided by Schelkunoff [2], Tai [4],
400
300
200
100
0
R
e
s
i
s
t
a
n
c
e

(

)
50 0 100 150 200 250 300

2 = 5
2
h
Antenna length h (deg)
30
40 50 60
90
10
5
10
20
30
40
50
60
80
90
20
30
40
50
60
70
80
90
20
5
70

Figure 5. Measured input resistance of a conical unipole versus
length in electrical degrees showing broadband characteristics
with increasing cone angle (from Ref. 10).
200
100
0
100
200
300
R
e
s
i
s
t
a
n
c
e

(

)
50 0 100 150 200 250 300

2
h
Antenna length h (deg)
2 = 5

5
10
20
30
90
40
50
60
5
70
10
20
30
50
60
80
40
90
10
20
50
60
90
30
40
70
80
90
5
10
60
20
30
40
50
Figure 6. Measured input reactance of a conical unipole versus

length in electrical degrees showing broadband characteristics
with increasing cone angle (from Ref. 10).
60
70
80
90
60
70
80
90
60
70
80
90
60
70
80
90
60
70
80
90
60
70
80
90
ka 1
ka = 2
ka = 1.5
ka = 4
ka = 3.5
ka = 6
ka = 5.5
ka = 3
ka = 2.5
ka = 5
ka = 4.5
0.25 0.5 0.75 1.0 1.25 0.25 0.5 0.75 1.0 1.25
0.25 0.5 0.75 1.0 1.25 0.25 0.5 0.75 1.0 1.25
0.25 0.5 0.75 1.0 1.25 0.25 0.5 0.75 1.0 1.25
(a)
(a) (b)
(c) (d)
(e) (f)
Figure 7. Far-zone electric eld patterns of
a biconical antenna, c=301. Patterns do not
change very much from ka=0 to kaEp.
Polk [12], and others. These topics are reviewed by Wait
[13]. The theory of an asymmetric biconical antenna was
also discussed by Schelkunoff [2]. The variational ap-
proach by Tai was extended to a semiinnite asymmetric
conical antenna consisting of an innite cone and a nite
cone [14].
To reduce wind resistance and/or weight, a solid bicon-
ical antenna can be replaced by a skeletal conducting wire
structure using several radial rods [15,16]. It has been
found, however, by analysis using the moment method
that a considerable number of wires (e.g., 16) is required to
approximate the solid biconical antenna.
2. ANTENNAS ON CONICAL STRUCTURE
In practice, a conical structure on which simple antenna
elements such as dipole, disk, cone, slot, or patch are
mounted to excite a cone is often used. The innite cone
excited with an axial dipole at the tip has shown [14] that
the strong radiation occurs along the small-angle cone
unless the dipole length is about a half-wavelength.
Figure 8 shows a nite wide-angle cone excited with a
quarter-wavelength long and a half-wavelength long di-
pole [17]. Note that the maximum radiation can be direct-
ed toward the horizontal plane by proper choice of a and y
0
at a desired frequency. A cone excited with an axially
symmetric circumferential slot close to the tip [18] shows
radiation characteristics similar to those of a dipole-excit-
ed cone, since a small circumferential slot (magnetic cur-
rent loop) is equivalent to a small axial electric dipole.
When the circumferential slot is not too close to the tip,
that is, apart by 2.5 wavelength (2c=301), the radiation
pattern shows a rather complicated lobe structure [18].
The cone excited with a circular disk at the tip is called a
discone antenna [19]. This antenna is fed with a coaxial
cable whose inner conductor terminates on the center of
the disk and whose outer conductor terminates on the tip
of the cone. The radiation pattern of the discone is similar
Figure 8. Far-zone electric eld pattern of a tip-excited conical
antenna, ba=50: (a) y
0
=901, l =l/4; (b) y
0
=103.81, l =l/4;
(c) y
0
=103.81, l =l/2; radiation beam can be directed to horizon
by slanting a cone downward (from Ref. 17).
7
6
5
4
3
2
1
5 4 6 7 8 9 10 11 12 13
V
S
W
R
Frequency (GHz)
0.635 cm
h
0.238 cm
L = 7.45 cm
L = 0.835 cm
Thin
polystyrene
disk
0.90 cm
0.238 cm
0.835 cm
Cone excited with
a cylinder element
Cone excited with
a conical element
Disk excited with a
cylindrical element
(dia of disk = dia
of cone base)
Figure 9. VSWR versus frequency
of a disk and a cone excited with a
cylindrical element or a conical ele-
ment showing that a cone excited
with a conical element is most broad-
band (from Ref. 17).
778 CONICAL ANTENNAS
to that of a dipole antenna, but its input impedance band-
width is exceedingly broad compared with an ordinary di-
pole antenna.
Figure 9 shows the measured input VSWR (for 57.6-O-
cable) versus frequency of the nite conical antenna with
different exciting elements at its tip [17]. The cone excited
with a conical element (an asymmetric biconical antenna)
indicates very broadband characteristics, particularly in
y
z
x
(a) (b)
2h
h
2
2
Figure 10. (a) Triangular plate antenna and (b) bowtie antenna.
700
600
500
400
300
200
100
0
R
e
s
i
s
t
a
n
c
e

(

)
R
e
s
i
s
t
a
n
c
e

(

)
0 0.5 1 2 1.5
0 0.5 1 2 1.5
= 15
400
200
0
200
400
600
Length of bowtie antenna 2h ( )
Length of bowtie antenna 2h ( )
(a)
(b)
= 45
= 30
= 60
= 15
= 45
= 30
= 60
Figure 11. Input impedance of a bowtie antenna showing the

broadband characteristics with increasing are angle (from
Ref. 20).
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
0
30
30
60
90
120
150
180
90
120
60
30
0
330 210
300
270
2h = 0.5
2h = 1.5
2h =
(a) x-y plane
(b) x-z plane
2h = 0.5
2h = 1.5
2h =
Figure 12. Far-zone electric eld pattern of a bowtie antenna,

c=301; (a) x y plane (b) xz plane (from Ref. 20).
lower-frequency regions, without affecting the radiation
pattern.
3. TRIANGULAR (BOWTIE) ANTENNAS
A triangular plate antenna above a conducting ground
plane and a bowtie antenna are shown in Figs. 10a and
10b. These antennas also possess broadband characteris-
tics, though not as broad as a solid conical antenna. The
theoretical characteristics of the bowtie antenna have
been obtained numerically [20] by using the method of -
nite-difference time domain (FDTD). Figures 11a and 11b
show the calculated input impedance. The input imped-
ance of the triangular plate antenna above the ground
plane is half of that of the bowtie antenna.
The far-zone electric eld patterns in the xy plane and
in the xz plane are shown in Figs. 12a and 12b, respec-
tively. Note that the radiation is enhanced in the direction
perpendicular to the antenna plate for the antenna length
2htl, because the radiation from the antenna surface
current is added in phase in that direction. The theoretical
directivity gain of the bowtie antenna in the direction of
the x axis is shown in dBi in Fig. 13 versus the antenna
length 2h/l for various cone angles [20]. It is noted here
that the bowtie antenna can also be simulated by several
radial wire rods as the solid biconical antenna.
BIBLIOGRAPHY
1. S. A. Schelkunoff, Principal and complementary waves in
antennas, Proc. IRE 34(1):2332 (1946).
2. S. A. Schelkunoff, Advanced Antenna Theory, Wiley, New
York, 1952.
3. P. D. P. Smith, The conical dipole of wide angle, J. Appl. Phys.
19(1):1123 (1948).
4. C. T. Tai, On the theory of biconical antennas, J. Appl. Phys.
19(12):11551160 (1948).
5. C. T. Tai, Application of variational principle to biconical
antennas, J. Appl. Phys. 20(11):10761084 (1949).
6. C. H. Papas and R. W. P. King, Radiation from wide-angle
conical antennas fed by a coaxial line, Proc. IRE 39(1):4951
(1951).
7. S. A. Saoudy and M. Hamid, Input admittance of a biconical
antenna with wide feed gap, IEEE Trans. Anten. Propag.
38(11):17841790 (1990).
8. V. Badii, K. Tomiyama, and D. M. Grimes, Biconical trans-
mitting antennas, a numerical analysis, Appl. Comput. Elect-
romagn. Soc. J. 5(1):6293 (1990).
9. J. G. Maloney, G. S. Smith, and W. R. Scott, Jr., Accurate
computation of the radiation from simple antennas using the
nite-difference time-domain method, IEEE Trans. Anten.
Propag. 38(7):10591068 (1990).
10. G. H. Brown and O. M. Woodward, Jr., Experimentally de-
termined radiation characteristics of conical and triangular
antennas, RCA Rev. 13(4):425452 (1952).
11. R. M. Bevensee, Handbook of Conical Antennas and Scatter-
ers, Gordon & Breach, New York, 1973.
12. C. Polk, Resonance and supergain effects in small ferromag-
netically or dielectrically loaded biconical antennas, IRE
Trans. Anten. Propag. 7(special suppl.):414423 (1959).
13. J. R. Wait, Electromagnetic radiation from conical structures,
in R. E. Collin and F. J. Zucker, eds., Antenna Theory,
McGraw-Hill, New York, 1969.
14. S. Adachi, A theoretical analysis of semi-innite conical an-
tennas, IEEE Trans. Anten. Propag. 8:534547 (l960).
15. C. E. Smith, C. M. Butler, and K. R. Umashanker, Charac-
teristics of a wire biconical antenna, Microwave J. 22(9):3740
(1979).
16. O. Givati and A. P. C. Fourie, Analysis of skeletal wire conical
antennas, IEEE Trans. Anten. Propag. 44:844858 (1996).
17. S. Adachi, R. G. Kouyoumjian, and R. G. Van Sickle, The nite
conical antenna, IEEE Trans. Anten. Propag. 7(special
suppl.):S406S411 (1959).
18. L. L. Bailin and S. Silver, Exterior electromagnetic boundary
problem for sphere and cones, IRE Trans. Anten. Propag.
4(1):516 (1956); corrections 4(3):313 (1957).
19. A. G. Kandoian, Three new antenna types and their applica-
tions, Proc. IRE 70W75W (1946).
20. Private communication from Y. He, T. Uno, and S. Adachi,
1997.
COPLANAR STRIPLINE (CPS) COMPONENTS
YOUNG-HO SUH
Mimix Broadband Inc.
Houston, Texas
1. INTRODUCTION
Coplanar stripline (CPS) is an attractive uniplanar trans-
mission line offering exibility in the design of planar mi-
crowave and millimeter-wave circuits, especially in
mounting the solid-state device in series or shunt with
no via holes. Its balanced structure is useful in applica-
tions such as printed dipole antenna feeding, rectennas,
uniplanar mixers [1], integrated optic traveling-wave
8
6
4
2
0
2
4
D
i
r
e
c
t
i
v
i
t
y

g
a
i
n

(
d
B
i
)
= 15
= 45
= 30
= 60
Length of bowtie antenna 2h ()
0.2 0.2 0.2 0.2 0.2 0.2 0.2 0.2
Figure 13. Directivity gain of a bowtie antenna in x direction

(from Ref. 20).
780 COPLANAR STRIPLINE (CPS) COMPONENTS
Next Page

03c

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

03c

Uploaded by

Copyright:

Available Formats

COMPUTATIONAL ELECTROMAGNETIC

SCATTERING MODELS FOR MICROWAVE

. The bistatic scattering cross sections of trees 5

) for an innitesimal slot in a cylinder of radius a.

Figure 1. Patterns of a thin axial slot on a perfectly conducting

Figure 3. Slotted-cone geometry.

Figure 4. Slotted cone and equivalent cylinder.

Figure 20. Polar coordinate system (original structure) (from

Figure 2. Characteristic impedance of a biconical transmission

50 0 100 150 200 250 300

50 0 100 150 200 250 300

Figure 6. Measured input reactance of a conical unipole versus

Figure 11. Input impedance of a bowtie antenna showing the

Figure 12. Far-zone electric eld pattern of a bowtie antenna,

Figure 13. Directivity gain of a bowtie antenna in x direction

You might also like