Professional Documents
Culture Documents
Master Thesis Xavier Banque
Master Thesis Xavier Banque
Mountainous
Regions with C-Band
Polarimetric SAR
Acknowledgments
The author of the present Master Thesis would like to acknowledge the
company Starlab Barcelona S.L. in which most of the work of this research has
been carried out. Likewise, he would like to acknowledge the master adviser
Carlos López for his guidance and tutorship along the development of the work
described hereafter. Finally, also Spanish CDTI (Centro para el Desarrollo
Tecnológico Industrial) has to be acknowledged since it provided the funding for
the research project that framed this research work carried out by a consortium
of Starlab Barcelona S.L. and UPC-TSC.
4
Abstract
This note summarizes the research work conducted in the frame of the Master
Thesis Snow Monitoring over Mountainous Regions with C-Band Polarimetric
SAR, targeting the polarimetric backscattering response of the snow-pack at
C-band. The scope of the research work carried out in this Master Thesis was
to characterize the polarimetric backscattering response of the snow-pack in
mountainous regions, in order to asses the feasibility of a quantitative monitoring
method based on C-Band polarimetric SAR measurements.
CONTENTS 6
Contents
1 Introduction 7
2 Polarimetry Introduction 9
2.1 SAR Polarimetry Introduction . . . . . . . . . . . . . . . . . . . 9
2.2 Introduction to Polarimetry . . . . . . . . . . . . . . . . . . . . . 10
2.2.1 Polarization Ellipse . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 Jones Vector . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.3 Stokes Vector . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.4 Covariance Matrix . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Scattering operators and polarimetric characterization . . . . . . 15
2.3.1 Scattering coefficient . . . . . . . . . . . . . . . . . . . . . 15
2.3.2 Scattering matrix . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.3 Polarimetric Coherency and Covariance Matrices . . . . . 18
2.3.4 Kennaugh Matrix. Huynen Decomposition . . . . . . . . 19
2.3.5 Canonical Scattering mechanisms . . . . . . . . . . . . . . 20
2.4 Polarimetric SAR Speckle Noise . . . . . . . . . . . . . . . . . . . 21
2.4.1 Speckle Noise Introduction . . . . . . . . . . . . . . . . . 21
2.4.2 Polarimetric Speckle Statistics . . . . . . . . . . . . . . . 22
2.5 Polarimetric decomposition theorems . . . . . . . . . . . . . . . . 22
2.5.1 Huynen decomposition . . . . . . . . . . . . . . . . . . . . 23
2.5.2 Freeman and Durden decomposition . . . . . . . . . . . . 24
2.5.3 Eigenvector-based decomposition . . . . . . . . . . . . . . 25
2.5.4 Pauli decomposition. Coherent Decomposition . . . . . . 25
2.6 SAR polarimetry Applications . . . . . . . . . . . . . . . . . . . . 26
1 Introduction
The current Master Thesis research work is motivated by the idea of Snow
Remote Sensing. During the last decades, Snow Monitoring has been an in-
teresting research topic for both the scientific and the remote sensing commu-
nity, due to the important role of the snow-pack in the hydrological cycle and
its many derivative applications for hydrology, meteorology and climate study.
Snow quantitative remote sensing is indeed still a research hot topic, since the
retrieve of quantitative snow physical parameters, such as Snow Water Equiva-
lent (SWE), remains not yet solved to a satisfactory operational level. Seasonal
snow in mountainous regions is for most of the countries in the north hemi-
sphere the main source of water reservoirs, therefore the estimation of the SWE
contained in the snow-pack would be a highly valuable information for water
management in those countries.
Taking the fact that snow is almost transparent to C-band SAR signal as
a starting point, the research work carried out in this Master Thesis aimed to
evaluate the impact of using fully polarimetric C-band SAR data in the im-
provement of snow monitoring with regards to single polarization performance.
Polarimetric backscattering response of the snow-pack might be more sensi-
tive to snow parameters, provided that polarimetry is sensitive to the different
scattering mechanisms of distributed targets, such as snow-pack volume scatter-
ing. With this purpose, a research project called EOSWAN, funded by CDTI,
has been carried out by a consortium formed by Starlab Barcelona S.L. and
TSC-UPC. Current space-borne C-band polarimetric SAR sensors and future
missions, such as Sentinel-1, would guarantee the feasibility of a potential snow
9
Conclusions and Further Steps The final section gathers the conclu-
sions and lessons learned during the development of the research work
and proposes some next steps to deepen in the topic of snow quantitative
remote sensing with C-band polarimetric SAR data.
10
2 Polarimetry Introduction
2.1 SAR Polarimetry Introduction
The SAR polarimetry technique aims to study the polarimetric response of a
scatterer to an incident polarized electromagnetic wave. Then, a scatterer is
characterized by the comparison of the polarimetric properties of the scattered
wave, with regards to those properties of the incident wave. Therefore, SAR
polarimetry is in pursuit of finding out determined parameters of the scatterers
from the knowledge of their polarimetric response.
The use of polarimetry in the frame of a SAR system adds some information
only available under polarimetric interpretation of the scattered signal, so that
we can talk about SAR Polarimetry. In order to take the maximum profit of the
polarimetric signature of a determined scatterer, this has had to be illuminated
by waves with both horizontal and vertical polarization and its scattered waves,
which will be a combination of both polarizations, have to be gathered also in
both horizontal and vertical polarization. In other words a fully polarimetric
system transmits and receives in the two linear polarizations, and there are few
instruments that can operate with such a polarimetric configuration. Therefore,
up to date, the amount of fully polarimetric SAR data available is reduced, and
polarimetric techniques are still a research field in a variety of target remote
sensing.
The polarimetric signature of a scatterer is to be analyzed in order to extract
the desired feature information from the observable. For instance, a scatterer
from whom no more information can be extracted under a non polarimetric
approach, can provide information about new parameters when observed in a
2.2 Introduction to Polarimetry 11
The ellipse orientation with respect to the x̂ axis that ranges from − pi
2 to
pi
2
E0x E0y
tan 2φ = 2 2 − E 2 cos δ with δ = δy − δx
E0x
(2.6)
0y
→
−
As time goes by the vector E (z0 , t) describes the elliptical shape. The ellipse
→
−
orientation is the angle of the vector E (z0 , t) with the x axis that evolves with
time ξ (t). The sign of the derivative of this magnitude determines the sense
of rotation, which, by convention, is evaluated when looking in the direction of
propagation.
E0x ejδx
→
− →
−
E = E (z) |z=0 = E (0) = (2.8)
E0y ejδy
where the α parameter is a constant phase term. In order to ease the calculations
with polarization vectors, the Jones vector can be rewritten by means of the
polarization algebra expressions, constructed from the unitary Pauli matrices
group.
10 10 01 0 −j
σ0 = σ = σ = σ = (2.10)
01 1 0 −1 2 10 3 j0
leading to the definition of three complex rotation matrices (SU(2)) of the special
unitary group whose expressions are:
cos (φ) − sin (φ)
U2 (φ) = = σ0 cos φ − jσ3 sin φ = e−jφσ3 (2.11a)
sin (φ) cos (φ)
2.2 Introduction to Polarimetry 14
cos (τ ) j sin (τ )
U2 (τ ) = = σ0 cos τ + jσ2 sin τ = ejτ σ2 (2.11b)
j sin (τ ) cos (τ )
jα
e 0
U2 (α) = = σ0 cos α + jσ1 sin α = ejασ1 (2.11c)
0 e−jα
obtained from the multiplicative properties of the quaternion formed by the
Pauli unitary matrices. The Jones vector in cartesian coordinates is thus ex-
pressed as a function of the special group matrices and a unit Jones vector
associated with the horizontal polarization x̂ = ûH :
cos(φ) − sin(φ) cos(τ )
E (x̂,ŷ) = Ae+jα = AU2 (φ) U2 (τ ) U2 (α) x̂ (2.12)
sin(φ) cos(φ) j sin(τ )
E1 |E2 = E T1 E ∗2 = 0
(2.13)
(2.20)
D E D E D E D E
1 g + g1 g −j g
= D 0 E D E D2 E D 3 E
2 g2 + j g3 g0 − g1
The diagonal elements of the wave covariance matrix present the intensities,
whereas the off-diagonal elements are the complex cross-correlation between Ex
and Ey , and the trace of the matrix is the total energy of the wave. In the
case that the off-diagonal elements are 0, no correlation exists between Ex and
Ey and therefore the wave is completely depolarized. On the other hand, for
2
|J| = 0 it follows that hJxx i hJyy i = |hJxy i| , which means that the correlation
between the two Jones vector components is maximum. The general case lies
between these two extreme cases, where |J| > 0 indicates a certain statistical
2.3 Scattering operators and polarimetric characterization 16
dependence between Ex and Ey that can be written in terms of the wave degree
of polarization (DoP) as follows:
q
2 2 2 12
hg1 i + hg2 i + hg3 i
hJi
DoP = = 1−4 (2.21)
hg0 i T r (J)
where
DoP = 0 is fully depolarized wave and
It is worth to point out that the wave covariance matrix elements depend on the
polarization basis of the Jones vector, and therefore to move the matrix from
one basis to another a special unitary similarity transformation is needed.
where U2 is the 2x2 unitary matrix of the SU(2) group containing the two unit
orthogonal eigenvectors u1 and u2 , and λ1 ≥ λ2 ≥ 0 are the two non negative
real eigenvalues. Wave entropy and wave anisotropy can be defined now as
measures of the correlated wave structure of the wave covariance matrix J as
follows:
2
λ1 − λ2 X λi
Aw = Hw = − pi log2 pi with pi = (2.23)
λ1 + λ2 i=1
λ1 + λ2
Both the entropy (Hw ) and the anisotropy (Aw ) range from 0 to 1with:
For a completely polarized wave λ2 = 0 : Hw = 0 and Aw = 1
if the target is smaller than the radar footprint, the power exchange be-
tween the wave and the scatterer is characterized by the radar cross-
section, which is defined using the radar equation as follows:
In case that the radar footprint is smaller than the illuminated target, this
is interpreted as set of statistically identical point targets. Therefore, in
such a case, the total power received is the contribution of all these point
targets in the illuminated area, and thus the power is obtained integrating
in this A0 area:
ZZ
PT GT (θ, φ) 0 AER (θ, φ)
PR = σ ds (2.26)
A0 4πrT2 2
4πrR
where σ 0 is the averaged radar cross-section per unit area, also called the
scattering coefficient. It represents the ratio of the statistically averaged
scattered power density to the average incident power density over the
illuminated surface with
→− 2
E S
hσi 4πr2
σ0 = = → (2.27)
A0 A0 − 2
E I
Although, the polarization is taken into account in the previous expression, this
is only an electromagnetic waves power based parameter, not exploiting the
vector nature of the electromagnetic waves. In order to use this vector nature,
the scattering process in the target of interest can be described, given the Jones
vectors of the scattered and incident waves, as follows:
e−jkr e−jkr S11 S12
ES = SE I = EI (2.29)
r r S21 S22
With regards to the incident and scattered orthogonal bases, the scattering
process can be described in a general form as:
R
Sφs φi Sφs θi EφI
Eφ
= (2.30)
EθR Sθs φi Sθs θi EθI
for the bistatic case, whereas in the monostatic case, the scattering matrix
is simetrical due to the reciprocity given by the geometry of the problem
(SXY = SY X ) and therefore, since the Pauli spin matrix basis has only
three matrices now, the resulting scattering k vector is:
1 T
k = √ SXX + SY Y SXX − SY Y 2SXY (2.36)
2
Lexicographic matrix basis set {ΨL }
1 0 0 1 0 0 0 0
{Ψp } = 2 2 2 2 (2.37)
0 0 0 0 1 0 0 1
that results in the Ω scattering vector:
1 T
Ω = √ SXX SXY SY X SY Y (2.38)
2
that corresponds to the bistatic case. Regarding the monostatic case, with
the same idea used in the k vector case, the Ω vector for the monostatic
backscattering case is:
1 √ T
Ω = √ SXX 2SXY SY Y (2.39)
2
These two different scattering vectors are related to each other by means of a
transformation matrix, which is constrained to be unitary due to the fact that
the Span of the scattering matrix has to be the same with independence of the
basis chosen. Hereafter the transformation expressions for both the bistatic an
the monostatic cases.
k = U4(L→P ) Ω and k = U3(L→P ) Ω (2.40)
where the expressions of the transformation matrices escape to the scope of this
polarimetry introduction note.
For the bistatic backscattering configuration both matrices are 4x4, and their
expressions are:
2 ∗ ∗ ∗
* |k1 | k1 k2 k1 k3 k1 k4 +
D E 2
k2 k1∗ |k2 | k2 k3∗ k2 k4∗
T4 = k · k∗T = k3 k ∗ k3 k ∗ |k3 |2 k3 k ∗
(2.42)
1 2 4
2
k4 k1∗ k4 k2∗ k4 k3∗ |k4 |
2.3 Scattering operators and polarimetric characterization 20
2
* |Ω1 | Ω1 Ω∗2 Ω1 Ω∗3 Ω1 Ω∗4 +
2
Ω2 Ω∗1 |Ω2 | Ω2 Ω∗3 Ω2 Ω∗4
D E
C4 = k · k∗T =
Ω3 Ω∗ 2
(2.43)
1 Ω3 Ω∗2 |Ω3 | Ω3 Ω∗
4
2
Ω4 Ω∗1 Ω4 Ω∗2 Ω4 Ω∗3 |Ω4 |
Regarding the monostatic case, the expressions are analogous, but using the
dimension 3 k and Ω scattering vectors, therefore the obtained Coherency and
Covariance matrices are 3x3.
Both Coherency and Covariance matrices are by construction Hermitian
semidefinite positive matrices with real nonnegative eigenvalues and orthogonal
eigenvectors. The transformation matrices relating the two scattering vectors
can be used to derive one matrix from the other one, as in the following example:
D E D E
∗T
T = k · k∗T = (UL→P Ω) · (UL→P Ω)
D E
= U(L→P ) Ω · Ω∗T U∗T
(L→P )
= U(L→P ) CU−1
(L→P ) (2.44)
These two matrices are in general terms the starting point for the polarimetric
analysis of a target. They contain information of the target themselves, and
they are used in the polarimetric decomposition methods of the targets
These two matrices can be diagonalized calculating their eigenvectors and
eigenvalues, which as stated before are nonegative values. The expressions of
the decomposed matrices are:
T = UP ΣP U−1
P and C = UC ΣC U−1
C (2.45)
where ΣP and ΣC are diagonal matrices containing the eigenvalues and UP and
UC are the unitary matrices with the three unit orthogonal eigenvectors. The
value of the eigenvalues provide information about the scattering mechanism,
precisely the level of polarization of the scatterers.
The canonical scattering mechanisms or scatterers are: the sphere or flat plate
or trihedral which are equivalent, the horizontal dipole, the oriented dipole, the
dihedral, the right helix and the left helix.
2.4 Polarimetric SAR Speckle Noise 22
The real and imaginary components of each return are random variables
statistically independent, and Gaussian distributed with zero mean and variance
σ 2 /2. Therefore the joint probability density function (PDF) is the result of
the product of the two independent
p gaussian random magnitudes, leading to a
received amplitude A = x2 + y 2 with a Rayleigh PDF as follows:
1 2 2
px,y (x, y) = px (x) py (y) = √ e−(x +y2)/σ (2.50)
πσ
2.5 Polarimetric decomposition theorems 23
the same PDF can be written as a function of the complex magnitude amplitude
A and argument θ and integrated over a θ interval of [−π, π] to obtain the PDF
for the amplitude:
2A 2 2
p (A) = 2 e−(A /σ ) A ≥ 0 (2.51)
σ
√
which is a Rayleigh distribution with mean σ π/2 and variance (4 − π) σ 2 /4.
Besides the amplitude of the received signal, the signal intensity I can be defined
as the square of the amplitude I = x2 + y 2 , having a negative exponential PDF
.
1 I
p (I) = 2 e− σ2 I ≥ 0 (2.52)
σ
The classical method to reduce the speckle noise is averaging several inde-
pendent reflectivity estimations, moving from single-look to multi-look images.
This is done dividing the synthetic aperture in N smaller sub-apertures which
are processed independently and averaged a posteriori as N independent sam-
ples. The result of this√ technique is the reduction of the standard deviation
of the speckle by a 1/ N factor, at expense of reducing the azimuth resolu-
tion by a factor of 1/N . The averaged multi-look amplitude, resulting from the
convolution of the N single-look Rayleigh PDFs has a Chi-squared distribution
which is narrower than the original single-look distribution. Another speckle
reduction technique is the averaging of neighboring pixels, that also worsens the
resolution.
For heterogeneous mediums, the Rayleigh model for the amplitude of the
received signal often fails. It is then used k-distribution model which derives
to a Rayleigh as the medium becomes homogeneous. The k-distribution model
is made up as the product of a gamma distribution, accounting for the texture
medium property, and a Rayleigh distribution. For homogeneous mediums the
variance of the gamma distribution decreases, with mean value equal to 1.
The information extracted from the parameters is about the total scattered
power from the regular and the depolarized components, the scattered power
from the symmetric and the nonsymetric components and the coupling between
symmetric and nonsymmetric.
For the Bragg scatterer modeling the surface, the cross-polarized compo-
nents are negligible, and therefore the scattering matrix of these compo-
nent is:
RH 0
S= (2.53)
0 RV
This Sinclair matrix leads to a scattering covariance matrix for this com-
ponent called C3S
where the subindex T stands for the Trunk of the dihedral like structure
and G stands for the ground. Notice that the Trunk and Ground different
reflection coefficients are accounted in the scattering matrix expression, as
well as the propagation factors included by the exponentials accounting
for possible attenuation and phase change effects. As well as in the surface
case, this Sinclair matrix leads to a covariance matrix for this component
C3D .
the volume scattering component, generally facing forestry targets, comes
from a cloud of dipoles and its scattering matrix depends on the rotation
angle θ of the horizontal dipole around the radar line of sight. It leads to
the covariance matrix for this volume scattering component hC3V iθ
The advantage of using this decomposition method is that it is based on a
physical model, and not only mathematical. However two strong limitations
reduce its usage: the three components used in the model apply only for a
determined kind of scenes, and the reflection symmetry assumption is not always
∗
fulfilled. Reflection symmetry hSHH SHV i = hSHV SV∗ V i = 0 assumes symmetry
in a plane normal to the line of sight. This is: if there is contribution of a
scatterer P in one side of the plane, there is an exact contribution of a scatterer Q
at the same position at the opposite side of the plane. Some reflection scenarios
do not have this property.
2.5 Polarimetric decomposition theorems 26
three single scattering mechanisms: the first is the single scattering from a
plane surface, the second and the third are double-bounce scatterings from
dihedral structures with relative orientation 0 and 45 respectively. The Pauli
decomposition associates a color to each of the components of the Span of the
scattering matrix that is written as follows:
2 2 2 2 2 2
Span = |SHH | + 2 |SHV | + |SV V | = |a| + |b| + |c| (2.60)
2 2 2
Being |a| red, |b| green and |c| blue, leads to a decomposition result like the
one shown in the following figure:
where d is snow depth in meters and ρ is snow density in Kg per cubic meter.
Therefore, in order to estimate the SWE of a snow-pack, two parameters have
to be estimated: the snow depth and the snow density in order to obtain the
SWE estimation as the product of them.
The snow water equivalent of a snow-pack is one of the most important
measurements of a snow-pack from a hydrological standpoint because it is a
direct measure of the liquid water content of a snow-pack. When some or all
of this water is released during the seasonal melt, it becomes an important
component of local surface water and groundwater budgets. In some areas of
the world, snow melt provides the majority of annual groundwater recharge,
so monitoring of snow water equivalent of snow-packs is an important part
of water resources management. However, SWE measurement is tedious and
costly, since it consists of melting a known volume of snow in order to weight the
water contained in it, in field measurement campaigns, that cannot be carried
out in vast areas. Therefore SWE cannot be extensively studied for large scale
applications which are the ones of scientific interest. From this need of mapping
in higher spatial and frequency scales the SWE parameter, arises the interest of
the remote sensing community in the estimation of this snow-pack parameter.
SIR-C/X SAR
The Shuttle imaging radar-C and X-band synthetic aperture radar, in
Figure 3.1, was a cooperative space shuttle experiment between NASA,
German Space Agency (DARA) and Italian Space Agency (ASI). This was
a SAR instrument mounted aboard NASAs space Shuttle that flew two
campaigns in 1994. It was the first fully polarimetric space-borne SAR,
that led to the obtaining of the first quad-pol image data sets from space
at L and C bands.
ENVISAT ASAR
ESAs ENVISAT launched in 2002, in Figure 3.2, carries the Advanced
Synthetic Aperture Radar (ASAR), in addition to other remote-sensing
instruments. ASAR instrument operates at C-band offering very good ca-
pabilities in terms of coverage, range incidence angles, modes of operation
and polarization. Though it does not consist of a fully-polarimetric in-
strument, it provides partially polarimetric products comprising two non-
coherent images of the same scene in the selectable polarization combina-
tion: HH/VV, HH/HV or VV/VH.
ALOS-PALSAR
The Japanese Aerospace Exploration Agency (JAXA) Advanced Land
Observing Satellite (ALOS), in Figure 3.3, is specially designed for land
remote sensing, with applications such as land coverage observation, dis-
aster monitoring and resource surveying. Launched in 2006, it carries
the phased array type L-band SAR (PALSAR), whose polarimetric ex-
perimental mode provides L-band full polarized images of swath 20-65Km
wide with a resolution of 24-89m. In partially polarization mode it can
achieve improved images with a resolution up to 14 m.
RADARSAT-2
This Canadian Space Agency (CSA) satellite mission matches the main
CSA challenges of monitoring the environment and managing the natural
resources. The RADARSAT satellites are a major data source for remote
sensing scientific as well as commercial applications. RADARSAT-2, in
Figure 3.4, was launched in December 2007 as a public-private partnership
between CSA and McDonald, Dettwiler and Associates Ltd. (MDA). It
3.1 Introduction 34
operates at C-band fully polarimetric and can achieve a very high resolu-
tion imaging of 8m and full flexibility in the selection of the polarimetric
configuration. It consists of the first instrument with quad-pol capabili-
ties providing full polarimetric products in normal configuration modes,
being a valuable tool in the remote sensing of geophysical properties of
the earths surface.
TerraSAR-X - TanDEM-X
The german radar satellite TerraSAR-X, in Figure 3.5, was launched in
June 2007 as a public-private partnership between German Ministry of Ed-
ucation and Science, the German Aerospace Center (DLR) and EADS As-
trium GmbH with an estimated life-cycle of 5 years. TerraSAR-X satellite
mission’s objectives are the provision of high quality multi-mode X-band
SAR data for scientific applications as well as establishing a commercial
EO market to develop a sustainable EO services business. In normal oper-
ation modes, TerraSAR-X can provide dual polarized data, and addition-
ally quad polarized data is available in experimental mode. In ScanSAR
operation mode, it provides 100 km wide swath images with a resolu-
tion of 16m. TanDEM-X opens a new era in space borne radar remote
sensing. The first bi-static SAR mission, is formed by adding a second,
almost identical spacecraft, to TerraSAR-X and flying the two satellites in
a closely controlled formation, with the primary mission of generating a
consistent global digital elevation model with an unprecedented accuracy.
The flight formation of two satellites with polarimetric capabilities allows
a set o polarimetric configurations in the acquired images.
Future Instruments
Since SAR polarimetry has been revealed as a promising remote sensing
technique, there will be further polarimetric SAR space-borne instruments
in the near future. An example is the CoreH2O mission of ESA’s Earth Ex-
plorer program, for cold regions hydrology observations, which is currently
in ESA’s phase A. This mission envisions a twin-frequency dual-polarized
SAR instrument specially developed for the forecasting of water supply
from snow cover and glaciers and, in general, for the modeling of water
and energy cycles at high latitudes. ESA’s Sentinel-1 will be the succes-
sor of ENVISAT’s ASAR as a C-band dual-polarization SAR, and is to
3.2 Snow interaction with EM waves 35
The parameters that determine the nature of each kind of snow are intro-
duced as follows:
Dielectric constant: The snow dielectric constant real and imaginary parts
are obtained from the air and the ice dielectric constant, since snow is an
inhomogeneous air and ice mixture. For dry snow, the dielectric con-
stants is almost independent of the frequency and the temperature at the
microwaves band of the spectrum, and the imaginary component is neg-
ligible whereas the real one is a function of the density. In the wet snow
case, the imaginary component of the dielectric constant is strongly re-
lated with the liquid water content of snow, increasing rapidly with the
presence of liquid water in the snow-pack. The complex dielectric constant
is a weighted average of the dielectric constants of the snow components:
air, ice and liquid water
ε = ε0 + iε00 (3.4)
Snow is a mixture of air and ice and its permitivity is a function of the
snow density ρ and the relative permitivities of air and ice. Snow is classified
in two types depending on the amount of liquid water content per unit volume,
since its characteristics and the scattering signature of the snow-pack strongly
depends on it.
Dry snow, which has no liquid water content, has a small absorption coef-
ficient allowing microwaves to propagate over long distances, having pro-
nounced scattering effects in the interaction with the snow-pack volume.
The real part of the permitivity ε0 of dry snow only depends on snow
density which in natural conditions ranges [2]:
0.2 < ρ < 0.5g/cm3 (3.5)
On the other hand, wet snow, with presence of liquid water content, has
a strong absorption coefficient attenuating the microwaves in a very short
distance. In such a situation, scattering is limited to a thin layer close to
the surface called surface scattering. This is due to the strong dependence
of the imaginary part of the permitivity ε00 with the amount of liquid water
in the snow volume. Indeed, it can raise an order of magnitude due to
an increase of a 0.5% of the liquid water volume, drastically reducing the
penetration depth [2].
Another interesting analysis involves the incidence angle in a mono-static
radar configuration:
For small incidence angles (close to nadir) surface scattering is the domi-
nating phenomena.
For bigger incidence angles volume scattering becomes more important,
since the surface contribution is moving to forward scattering.
3.2 Snow interaction with EM waves 38
This document will mainly focus on dry snow, since this is the scope of this
research project, and therefore, from now on, when talking about snow, we will
be referring to dry snow.
As already introduced, the back-scattering response of the snow-pack is the
result of the contribution of different scattering phenomena: underlying ground
surface back-scattering, volume back-scattering from the snow layer, air-snow
interface back-scattering, and multiple contributions from double and triple
bounce between the snow-pack and the soil surface. The following Figure 3.7
from [7] gives a clear graphical idea of the multiple scattering mechanisms that
compose snow-pack back-scattering.
The low dielectric contrast between air and dry snow-pack and the smooth-
ness of the snow surface lead to a snow surface back-scattering coefficient neg-
ligible ([2] and [1]), and therefore the scattering signature of the dry snow pack
is the addition of the volume scattering contribution and the soil surface back-
scattering contribution. This is the idea that drives the most recent research
works in snow remote sensing with PolSAR: understanding the received back-
scattered power as the addition of different scattering mechanisms out of which
we can infer different informations about the snow-pack. The volume scattering
contribution is driven by the size of the ice crystals relative to the wavelength
and snow particles orientation [2]. The back-scattering coefficient of the snow-
pack is insensitive to the snow surface roughness for dry snow [9].
In most of the cases, the soil surface scattering is considered as Brag scat-
tering or Small Perturbation Method (SPM) [1, 17], since it is considered that
surface roughness is slight with regards to the incident wavelength, and snow-
pack volume backscattering is modeled as a Rayleigh scattering resulting of a
random oriented cloud of particles whose scattered intensity in the far field is
the addition of the individual intensities of each particle [1]. As already said,
volume scattering in the snow-pack is mainly governed by the size of the ice
crystals relative to the incident wavelength [2]: the bigger the snow particles
with regards to the incident wavelength, the higher the interaction of the elec-
tromagnetic waves with the snow grains, and therefore the higher the volume
scattering. Therefore, as the incident wave frequency is reduced the volume
scattering effect becomes more important. Another usual assumption in the
snow-pack scattering response modeling is the Born approximation that estab-
lishes that the scattering generated out of the interactions of the particles with
the diffused radiation is negligible. Born approximation can be assumed since
snow-pack medium fluctuation is very little. Snow presents continuously dis-
tributed volume absorption and volume scattering and is therefore described by
3.2 Snow interaction with EM waves 39
the Radiative Transfer Method (RTM), that describes the waves interaction as
they travel across the volume, taking into account all the different scattering
mechanisms.
At this point of the document we can already point out an interesting con-
cept. For quantitative snow remote sensing, snow-pack volume has to be sensed
in order to get information out of it. This means that microwaves have to pen-
etrate the volume and at same time interact with the snow grains so that some
information about the snow-pack can be extracted from the scattered power.
From what has been said in Section 3.2.1 and in previous paragraph of the cur-
rent Section 3.2.2, one can realize that there must be an interesting trade-off
between penetration capacity and sensitivity to the volume of snow particles. In
[7] is stated that at C-band the volume scattering is the dominant phenomena,
whereas at L-band the ground scattering is the dominant mechanism. However,
this statement is contradicted in other literature sources [?].
due to the volume scattering that results in a power dispersion and therefore
less power is back to the antennas. On the other hand, C-Band is not sensitive
to re-frost layers of the snow-pack. However, more recent publications [7] do
find a reasonable sensitivity to the snow-pack volume at C-Band. This difference
might be driven by the amount of liquid water content in the snow-pack, since
the extinction coefficient at C-Band increases dramatically with the liquid water
content, decreasing the backscattered power. Effort needs to be put on this issue
in order to clarify the snow behavior at C-Band.
In PolSAR there are a set of decomposition methods called model-based in-
coherent decompositions that try to identify the main physical components of a
back-scattering scenario in order to find out the contribution of each component
in the total retrieved back-scattered power [4, 3].
The research work carried out in [9, 10] in 2000 fixed the basis for a method
that has been used in many other research works later. This is based on
the fact that although it does not significantly modify the back-scattering
response, the presence of snow affects the underlying ground backscatter
as well as the relation between the backscatter at different polarimetric
configurations. This initial research work has been the inspiration for
further works such as [6] or [14], and even [15], though this last one is a
more complex evolution of the original idea [9].
Their rationale is stated as follows: the effects of dry snow cover on surface
back-scattering compared to the surface back-scattering are:
Due to refraction with the snow, the incidence angle at snow-ground in-
terface is smaller than the incidence angle at air-snow interface (Snell’s
law).
The first two effects in the list above result in a change of the sensor observ-
ables. Both effects depend on the dielectric constant of the snow-pack, which
can be related to the density by means of Looyenga’s semi-empirical formula
in 3.3. we can assume that at L-band the volume scattering is negligible, and
therefore the backscattered power can be expressed by
t 2 s
σpp (k0 , θi )) = Tpp (θi )σpp (k1 , θr )) (3.6)
s
where σpp (k1 , θr )) is the ground surface back-scattering given by the Integral
Equation Method (IEM) model, k1 is the incident wave number at the snow-
√
ground interface with k1 = k0 εs , θi is the incidence angle at air-snow interface
which is related to the incidence angle at the snow-ground interface by Snell s
law,
√ √
log10 ( σhh + σvv ) = a(θr , k1 ) + b(θr , k1 ) log 10 (σhh + σvv ) +
+ c(θr , k1 ) log 10 (σhh ) + d(θr , k1 ) log 10 σσhh
vv
+
2
σhh
+ e(θr , k1 ) log 10 σvv (3.8)
This equation represents the relationship between HH and VV back-scattering
coefficients at a given incidence angle and wavenumber k1 . Note that the coef-
ficients a,b,c,d and e depend only on the incidence angle and the wavenumber.
Further investigation in the scope of this research is needed since the rationale
for the obtaining of the previous equation is not clear in the documentation
found so far.
The algorithm was validated by means of the comparison of the estimations
obtained from SIR-C L-band images and field snow density measurements, show-
ing an absolute RMS of 42 Kg/m3 and a relative error of 13%.
Further research work was carried out with the same idea than the here-
above explained [13] but also moving to C-Band data [6, 14]. The idea of
the techniques based on C-Band polarimetric data is the assumption that the
total back-scattering coefficient for a dry snow-pack is the sum of two main
v sg
contributions: volume scattering σpp and snow-ground backscattering σpp
t v sg
σpp (k, θ)) = σpp (k1 , θr )) + σpp (k1 , θr )) (3.9)
The development of the algorithm for snow density estimation is based on
the first order volume scattering and the Integral Equation Method (IEM). The
effect on the incidence angle change depending on the snow density found at
L-band [9], applies also for C-Band, that has lower penetration capability and
thus is even more sensitive than L-Band to this effect. Therefore, an algorithm
can be performed to estimate snow density from C-Band ASAR alternative
polarized HH and VV data. Using HH and VV backscattering coefficients for the
snow-ground backscattering, and volume backscattering ratios for the volume
backscattering, a model can be obtained with only two unknowns: the dielectric
constant and the incidence angle that can be related to the dielectric constant
with Fresnel Transmission coefficients and Snell’s law. The model validation
was carried out by comparing the snow density estimations obtained from the
ASAR polarimetric images with field measurements, with a mean absolute error
of 21.2 Kg/m3 . One of the main advantages of the model is that it does not
require knowledge about the surface properties.
Finally, in the same research group, [15] also based on C-Band dual co-
polarization data, comes out with an enhanced technique based in the same
idea, that gives as output an estimation of the snow-pack dielectric constant. It
is interesting the fact that the local incidence angle is corrected on a pixel basis
since in a mountainous region, the local incidence angle has a strong variation,
and it has to be estimated for each pixel prior to its processing.
wet or dry is the monitored snow. Three very interesting ideas can be extracted
out of this work:
The main idea is the development of a method based in the combination of
Earth Observation data (ENVISAT ASAR) and a meteorological model
in order to enhance the snow-pack characterization. The good spatial
resolution of the SAR data can significantly improve the capacity of snow
parameters estimation.
The use of a multilayer-snow electromagnetic backscatter model allows
a much more accurate modeling of the scattering phenomena since it
takes into account the physical vertical morphology of the snow-pack, that
presents some layers whose presence depends on the atmospheric and me-
teorological conditions at each given time.
The application of the already mentioned idea about the extinction coef-
ficient increasing with the presence of liquid water within the snow-pack
at C-Band. This principle is used in order to identify the wet snow zones
by means of a simple ratio value on the comparison of winter co-pol SAR
images with reference summer co-pol SAR images.
σwinter
< −3dB (3.10)
σwinter
If for each pixel the condition in equation 3.10 is fulfilled, the pixel contains
wet snow.
SW E = ρK(ρ)R (3.11)
where ρ is the snow density and ρK(ρ) is the slope of the regression function
relating SWE with R that has to be estimated.
The method applied is as follows:
3.4 Coherent PolSAR Data Methods 44
SW Emeasured
a= with a = ρK(ρ) (3.13)
Rmeasured
The retrieved data from the radar instrument composes the scattering
matrix 3.2, whose information can be put in a vectorial form under the
assumption of mono-static radar configuration (reciprocity applies Sxy =
Syx ), using the Lexicographic matrix basis set (see [3] and [4]) leading to
the lexicographic target vector:
h √ iT
K = Sxx 2Sxy Syy (3.14)
D
2
E √
∗
∗
|Sxx | 2 Sxx Sxy Sxx Syy
√
D
2
E √
C = K · K ∗T = 2 hSxy Sxx
∗ ∗
with ∗T being the conjugate transpose, ∗ being the transpose and h·i being
the ensemble average
3.4 Coherent PolSAR Data Methods 45
In addition, ice and snow have the same physical components. Obviously,
proportions and morphologies are different and thus, so it is the dielectric con-
stant, which is the most important parameter driving the interaction between
EM waves and ice or snow.
In [17] the approach chosen to infer properties of the ice pack consists of
decomposing the polarimetric covariance matrix in three submatrices, repre-
senting three different scattering sources: ground surface scattering, ice volume
and sastrugi (wind-induced features is the snow-firn upper layer). Then the
covariance matrix results as:
Section intends to gather the most interesting ideas out of that group of refer-
ences, that could apply for the snow case.
Some interesting research work has been done with PolSAR in the field of
vegetation remote sensing [19, 21]. From the physical point of view, vegetation
is seen as a volume, called canopy. This is the aboveground portion of plant
community or crop. Since the snow-pack is also seen as a volume of air particles
in an air background, concepts from the PolSAR vegetation volume remote
sensing might be also interesting for the snow case.
These two works [19, 21] have the same starting point idea, which is to model
separately the ground and the volume contributions to the total backscatter-
ing of the vegetation covered soil. In [21] a Polarimetric Interferometry SAR
(PolInSAR) approach is used, and the vegetation is seen as a random volume
over ground (RVoG). The RVoG is a model used in PolInSAR to asses the two-
layer scattering coherence problem (vegetation and soil). It considers a random
oriented volume, and therefore the propagation is independent of the polariza-
tion and depends only on the extinction coefficient. It is remarkable from this
job the approach of separating ground and volume contributions out of the total
backscattered power in order to infer information about the volume of vegeta-
tion, since this separation idea seems to be present in all the situations where a
volume scattering is present.
The work carried out in [19] is based on PolSAR. The most relevant idea
regarding the snow-pack monitoring is the inclusion of a model for an oriented
volume. They rely in a probability density function to characterize the particles
distribution, which is completely defined by a mean orientation angle and the
power of a cosine-square function. They come out with a decomposition of the
volume covariance matrix which is the addition of the oriented volume contri-
bution plus the random oriented contribution on a weighted manner depending
on the scatterers volume characteristics.
Finally, another interesting decomposition of PolSAR data is proposed in
[22]. This consists of a four-component decomposition to take into account
the cases when reflection symmetry cannot be applied, which is normally in
the case of man-made targets. This decomposition adds to the well known
surface, double-bounce and volume scattering components, a fourth helicoidal
mechanism. It is a suitable decomposition method for those images with man-
made targets as well as natural distributed targets.
3.5 Conclusions
After the state-of-the-art review concerning SAR and polarimetric SAR for snow
remote sensing, it is clear that snow remote sensing is a very interesting research
topic, on-going for the last decades, due to its scientific applications on climate
study as well as hydrology, and water resources management applications, such
as flooding or water reservoirs forecasting. This Section aims to summarize
all the relevant conclusions out of the literature review, as well as the ideas to
initiate the research work on PolSAR for snow quantitative remote sensing.
Concerning conclusions:
It seems, out of the review, that the general idea to face the volume
scattering case with a model-based decomposition combined with an EM
polarimetric model of the snow scattering problem; involving volume un-
derlying surface scattering plus snow-pack volume scattering, plus other
mechanisms scattering if any, is the appropriate one for the scope of snow
3.5 Conclusions 48
The following subsections will describe the research work carried out, from
the experiment planning, to the data processing and analysis, and finally the
snow monitoring features obtained. Summarizing, the work accomplished fits
the classical structure of a remote sensing research project, with a ground mea-
surements campaign providing with ground truth, some satellite data acquisition
in coordination with the ground campaigns, that produce the polarimetric SAR
data to be processed, and finally some ancillary data supporting the assump-
tions during the data analysis and interpretation of the results. The processing
of the data acquired aimed to find some feature sensitive to the snowed scenarios
that could be used to retrieve snow physical parameters. Finally, an effort to
synthesize the lessons learned out of the gathered data and its processing and
analysis leads to the conclusions and further steps section.
area of around 25 Km2 . SAR images were acquired in Single Look Complex
(SLC) configuration, so that no loss of information affects SAR data before
own processing. The Pauli decomposition RGB representation for the Pla de
Beret Zone in the three different acquisitions can be appreciated in the following
Figure 4.2. Henceforth in the document, those figures containing SAR images
or derived products have the azimuth dimension in the vertical axis and range
dimension in the horizontal axis:
Note that blue color corresponds to the polarimetric channel for plane sur-
face scattering, red corresponds to double-bounce scattering and finally green is
associated with the volumetric scattering (the non-diagonal components of the
scattering S matrix [3]). By means of comparing these snapshot images with
a zenithal optical image (see Figure 4.3) from the same zone of Pla de Beret
we can identify different elements in the scenario, such as the parking zone of
the ski resort and the road that gets to it, the small canyons on the right side
of the parking area or the tele-ski path as a straight line crossing the images
horizontally.
Concerning the ground truth measurements, the two campaigns carried out
consisted of snow measurements of snow depth and snow density, as well as a
general assessment of the snow status carried out by a snow expert who was
present in both campaigns. The approximately 30 measurements per campaign
covered the Pla de Beret zone, with more or less equi-spaced measured spots.
Fortunately, snow conditions changed significantly from one date to the other,
with the snow depth increasing a 100% and the snow density increasing 34%
from one scene to the other. In Table 1, a summary of the average values
obtained during the campaigns is presented, the Difference column refers to the
differences between the two snowy acquisitions.
The ground measurements campaign was carried out in the frame of a project
called AGORA, which was a research project led by Starlab Barcelona S.L. and
funded by ACC10 targeting the use of remote sensing data for the improvement
of the hydrological modelling. The author of the current Master Thesis would
like to thank the AGORA team for the collaboration during the planning of the
4.2 Description of the Experiment 52
campaigns and the SAR data acquisitions and the valuable data provided.
Figure 4.3: Ortho image from Pla de Beret (Institut Cartografic de Catalunya).
homogeneous distributed target. Therefore, mean intensity values, and its cor-
responding variance, were computed for the following polarimetric channels:
HH, HV, VV, HH+VV, HH-VV and Span for the three SAR acquisitions. The
a priori expected result according to the literature, was finding bigger Span
in the snowy scenes, due to the increase of the intensity in HV and HH+VV
channels. Obtained mean values in dB for the different channel intensities are
summarized in the following Table 2:
Table 2: First analysis results summary for Pla de Beret zone (Test Zone 1)
Surface is not strictly flat, and the roughness might be high enough so
that no Bragg [2] surface scattering can be assumed in our data analysis.
In the next Figure 4.4 , a picture of the soil can be appreciated, with two
kinds of roughness: small scale roughness due to grass and small variations
and a bigger scale roughness with random small hill like perturbations.
After the Pla de Beret site visit no reasonable explanation could be found
to the fact that February and October scenes have very similar backscattering
in most of the channels and that March scene, with a thicker snow-pack, has
around 5dB lower backscattering than the other scenes.
A potential hypothesis was assuming that March scene snow was not com-
pletely dry. Past research results at C-band with wet (Liquid Water Content
4.3 Snow Backscattering Results Analysis 56
> 1% per unit volume [2]) snow at an incidence angle of 40 degrees [5] show
that from dry to wet snow backscattering power falls around 5dB. In order to
prove this hypothesis, meteorological information of the site from the interest
period was needed for the estimation of the snow status during the satellite
data acquisition. The Confedereación Hidrográfica del Ebro has a permanent
automatic meteorological station at Pla de Beret, and daily basis information
was obtained about temperature and precipitation for the period of interest.
Taking into account the fact that satellite data acquisition for the three scenes
was around 18h, and with the knowledge of the meteorological data about tem-
perature from the acquisition days (see Table 3), the following assumptions
were done. In the February acquisition the temperature was below 0 during all
the day, as well as the two days before, guaranteeing dry snow status of the
snowpack, whereas in the March case, temperature was above 0 during most
of the day, and in addition max temperature time was only 90 minutes before
the acquisition time, meaning no time enough for the snow to refreeze before
acquisition time. Again, March 2 days prior to the acquisition showed similar
meteorological features in temperature. This means that we can assume that
March scene’s snow status was wet.
Table 3: Temperature in Pla de Beret zone for February and March acquisition
days
Figure 4.5: Temporal evelution of the Temperature in Pla de Beret during the
period of interest)
Pla de Beret
– Number: 1
– Geo-coordinates: lon 0.966◦ and lat 42.732◦
– Altitude: 1860 m
– Area: 60 x 300 px (trapezoid)
4.4 Snow Monitoring 58
Notice that the two forested zones (2 and 3) are located in the Pla de Beret,
so that all the ground truth snow measurements information as well as the
meteorologic data can be also used in the analysis of these two zones. Concerning
test zones 4 and 5, both quite flat alpine meadows, they were chosen in order
to have, on one side, a meadow (5) similar to Pla de Beret in altitude, so
that assumptions about similar meteorological and snow conditions could apply
provided that the two meadows are separated approximately 4.5 Km, and on
the other side, a 400 m. higher meadow, in which by assuming the rule of the
temperature decrease with the height increase, the temperature can be assumed
to be between 2 and 4 degrees lower than in Pla de Beret, and the quantity
of snow is expected to be higher. Finally, test zone 6 was chosen as ground
truth, being a urban zone its polarimetric backscattering response was expected
to be know a priori. The following Pauli RGB representation of the February
acquisition shows the situation of the different 6 test zones within the image
(see Figure 4.6).
Analyzing data on a test zone basis, the idea of the forested zones was to
be able to notice the effect of the trees in the volumetric scattering component,
masking the potential snow contribution to the volume scattering component
having in addition two very different local incidence angles in each forested test
zone, allowing the observation in the polarimetric backscattering response of the
change of angle of incidence effect. The estimation of the local incidence angle
in each zone has been done with an easy trigonometric transformation having
topographic basic information from the site: the slope of each forested zone can
be estimated from the ratio of increase in height with the distance. In Figure
4.7 a simple diagram depicting the geography of the north part of Pla de Beret
shows graphically the method employed for the local incidence angle estimation.
Obtaining the topography information from a detailed map (Vissir application
from Institut Cartogràfic de Catalunya http://www.icc.cat/vissir3/), the calcu-
lation of the local incidence angle simply involves an arctangent operation and
a difference as shown in equations 4.1 and 4.2.
70 ◦ 250
α = atan = 21.8 β = atan = 24.0◦ (4.1)
200 560
0 00
θi = 46.6◦ + α = 68.4◦ θi = 46.6◦ + β = 22.6◦ (4.2)
Where α and β are respectively the angle of the terrain slopes in Bosc de
Parros and Er Anherar, obtained from the topographic information, as shown
in Figure 4.7, where it can be seen as well the different local incidence angles
due to the different terrain slopes.
Figure 4.7: Diagram for the different incidence angles according to the terrain
slopes
Figure 4.8: PauliRGB representation of the co-registerd data images. Axis units
are pixels and colormap is as in 4.2
Note that test zone 3 is the one where the forest effect on volume and double
bounce is bigger, probably due to the highest incidence angle that leads to a
bigger volumetric contribution since the signal path interacts longer with the
vegetation. It is worth to check the dominance of the double bounce component
in all the scenes and test zones with regards to the volumetric component.
Finally, it is interesting to realize that, regardless of the test zone, both the
volumetric and the double bounce components are bigger in the no snow scene
(October) than in the wet (March) or dry (February) snow scenes. This fact
4.4 Snow Monitoring 63
scattering component in the no snow scene for the three meadows analyzed (1,4
and 5). It was not expected to be as high as the surface scattering component,
since the initial idea was that, being a mountain grass covered surface, the
surface scattering would be much higher than the other scattering mechanisms.
In what the polarimetric H/A/ᾱ decomposition is concerned the behavior
of Entropy and ᾱ is the same that was already observed in the test zones 1,2
and 3, suggesting a pattern that varies with the presence and/or type of snow
independently of the kind of target that the radar is illuminating: Entropy and
ᾱ decrease with the presence of snow, and are lower for the dry than for the
wet snow case. This phenomena might be indicating that the presence of snow
reduces the randomness level of the backscattering mechanism.
In March acquisition there was probably wet snow in the area illuminated
by the radar. All the polarimetric channels in all the test zones from
1 to 5 are around 5dB lower, which could be perfectly explained with
the wet snow hypothesis [5] which is reinforced with the interpretation of
the meteorologic data available from Pla de Beret during the acquisitions
period.
From [8], the idea that the presence of snow increases the volume scattering
mechanism intensity was tried to assess with our data. Only in test zone
4, the HV intensity was bigger for the dry snow scene than for the no
snow scene, probably due to a significantly thick snowpack due to the
higher altitude of this test zone. This is aligned with the fact that in [8]
a 2 m depth snowpack was used as input parameter of the simulations.
4.4 Snow Monitoring 65
Therefore, since in the acquisitions used in this research work there was a
shallow snowpack, the volumetric mechanism increase is not an observed
effect.
Concerning Pauli surface component HH+VV, a slight increase of it can
be observed in most of the test zones. It happens for the test zones 1,2 and
2 2
4, 5 that: |HH + V V |drysnow > |HH + V V |nosnow (between 0.4dB and
3.5dB) which could be due to the reduction of the incidence angle in the
surface scattering with the presence of dry snow caused by the refraction
of the incident signal [9]. Test zone 3 shows almost no variation on this
component with the presence or not of dry snow.
Regarding Pauli double bounce component HH-VV, despite it could not
be observed a general trend, it happens for the test zones 1, 2 and 3 that:
2 2
|HH − V V |drysnow < |HH − V V |nosnow (approximately 1dB). It seems
however that this effect is not happening in test zones 4 and 5, but the
opposite.
Therefore, and despite the effect on HH-VV couldn’t be found in all the
test zones, putting together the information of the last two items, it came
up a new idea that could be used for dry snow detection, since with the
presence of dry snow, on one side the HH+VV component increases and
on the other side the HH-VV component decreases. This leads to:
2 2
∆ = |HH + V V | − |HH − V V | (4.3)
and it happens that ∆drysnow > ∆nosnow around 1dB for all the test zones
from 1 to 5 (see diff columns in Table 9 for February and October scenes
respectively). Though these results don’t show big differences between
dry snow case and no snow and therefore they are not robust, they might
be indicating a research field to be explored. Care has to be taken since
HH-VV is sensitive to polarization orientation angle changes, therefore
data should be corrected to compensate this effect.
About the polarimetric H/A/ᾱ decomposition, the most interesting fact
is realizing how the presence of snow reduces the H and ᾱ parameters in
all the test zones from 1 to 5. This suggests that the presence of snow
reduces the randomness of the scattering phenomena regardless of the
kind of underlying target. The reduction of the Entropy in the snowy
scenes can be easily related with the observed slight decrease in the HH-
VV component and the slight increase of the HH+VV component with the
presence of snow, since this involves HH+VV being a relatively stronger
scattering mechanism in the total Span contribution, thus a reduction of
the randomness of the backscattering.
Figure 4.9: ∆ (dB) representation of the co-registerd data for the three different
scenes. Values range from -13dB to 23dB and axis units are pixels
snow (diff-17) image and no snow (diff-15) can be identified visually, despite not
being as obvious as in the wet snow case (diff-13).
Notice that the images displayed in Figure 4.9 have been obtained from
Speckle filtered Coherency matrices. A 7x7 Boxcar filter was used in order to
be able of constructing statistically robust ∆ images from the first and second
diagonal elements of the filtered Coherency matrix (T11 and T22 )
4.4
Date 17/02/2011
Zone Magnitude HH HV VV HH+VV HH-VV Span
Mean Linear 2,53E+06 4,04E+05 2,70E+06 7,70E+05 2,02E+05 6,04E+06
Snow Monitoring
Date 13/03/2011
Zone Magnitude HH HV VV HH+VV HH-VV Span
Mean Linear 3,3860E+13 1,3716E+11 2,5326E+13 5,6654E+11 7,3529E+09 1,1427E+14
Snow Monitoring
Date 15/10/2011
Zone Magnitude HH HV VV HH+VV HH-VV Span
Mean Linear 1600100 304370 1638800 701260 240220 3847700
Snow Monitoring
Var Linear 0,007 0,014 26,570 0,009 0,014 36,875 0,006 0,013 31,059
Mean Linear 0,734 0,264 32,802 0,809 0,260 39,123 0,818 0,242 39,480
2 Var Linear 0,009 0,009 39,201 0,009 0,011 51,726 0,007 0,009 40,410
Mean Linear 0,815 0,266 39,772 0,861 0,258 44,387 0,868 0,264 44,482
3 Var Linear 0,004 0,011 26,810 0,004 0,011 25,697 0,003 0,011 24,190
Mean Linear 0,648 0,290 29,493 0,625 0,322 28,525 0,761 0,252 35,169
4 Var Linear 0,011 0,015 47,734 0,012 0,013 46,057 0,006 0,011 28,584
Mean Linear 0,785 0,298 37,700 0,790 0,338 38,866 0,832 0,260 41,072
5 Var Linear 0,007 0,014 39,170 0,008 0,014 50,032 0,006 0,010 39,224
Mean Linear 0,729 0,465 44,436 0,731 0,475 43,466 0,733 0,450 41,575
6 Var Linear 0,024 0,048 135,760 0,024 0,046 138,260 0,022 0,047 108,560
70
4.4
Table 9: First and Second Pauli Channels difference for the three scens
Var Log (dB) 136,98 123,72 13,26 136,92 126,4 10,52 136,24 126,1 10,14
3 Mean Log (dB) 60,216 55,838 4,378 60,173 56,702 3,471 60,336 56,926 3,41
Var Log (dB) 122,25 113,33 8,92 123,07 115,3 7,77 122,88 115,44 7,44
4 Mean Log (dB) 64,067 57,453 6,614 59,904 52,821 7,083 60,599 55,237 5,362
Var Log (dB) 131,24 118,85 12,39 131,09 110,71 20,38 124,23 114,43 9,8
5 Mean Log (dB) 59,559 54,724 4,835 53,58 49,447 4,133 58,495 54,415 4,08
Var Log (dB) 123,48 111,49 11,99 112,32 102,75 9,57 120,59 110,7 9,89
6 Mean Log (dB) 69,983 69,51 0,473 69,85 68,549 1,301 69,509 67,307 2,202
Var Log (dB) 153,47 154,22 -0,75 152,39 150,32 2,07 152,16 147,25 4,91
71
4.4 Snow Monitoring 72
∗
Θ = [Arg (hSRR SLL i) + π] /4 (4.4)
where the Arg() based on the circular polarization copol components can
4.4 Snow Monitoring 73
∗
−4Re (h(SHH − S ) S i)
∗
Arg (hSRR SLL i) = tan−1 D E V V D HV E (4.5)
2 2
− |SHH − SV V | + 4 |SHV |
0 cos (Θ) sin (Θ) Shh Shv cos (Θ) −sin (Θ)
S = (4.6)
−sin (Θ) cos (Θ) Svh Svv sin (Θ) cos (Θ)
The above method was applied to the PolSAR dataset using a 3x3 box
pixel averaging in the estimation of the angle shift, in order to reduce
some noise in the estimated values (with this average, slopes of around
24 m2 were obtained). Resulting data analyzed in the presented 6 test
zones for the sake of comparison, was quite similar to the original one.
It was realized that the angle shift about the LOS for the different test
zones is similar. The real change between the test zones is in the incidence
angle and this does not almost affect the polarization state. HV channel
is the only one that suffered a drop of around 1dB in all the test zones
and temporal scenes, indicating that it was probably overestimated in
previous analysis. However, since the change applied the same way for
all the zones and scenes and since the data analysis is based on change
detection techniques, the ratios computed would remain very similar and
therefore the terrain slope correction for PolSAR data was assumed a
second order phenomena, to be taken into account in further research
work with a wider scope.
Finally, in the results interpretation, the main topics of the work carried
out were on one hand looking for a reasonable physical description of the ∆
parameter, showing an increase in the dry snowy scenes, and on the other
hand the deep analysis of the H/A/ᾱ decomposition, and if possible, find-
ing synergies between these two topics. Indeed, the fact that the double
bounce polarimetric component was surprisingly important, with respect
to surface component, in alpine meadow test zones was the initial point
of the survey. Having such significant levels of double bounce scattering
intensity, together with the high Entropy (suggesting some depolariza-
tion) and ᾱ values, that classified the alpine meadows in the H/ᾱ plane
as random surface and/or anisotropic particles volume diffusion, it was
questioned the initial assumption that the snow back-scattering problem
had to be focused mainly on T11 (surface) and T33 (volume). Therefore,
the initial hypothesis that alpine meadows at C-band could be modeled
with Bragg scattering was concluded not to be correct. In fact, this can be
clearly stated having a look at the pairs H/ᾱ for each test zone, figuring
out that none falls into the Bragg surface zone of Figure 4.10, regardless
of the time of the year scene we work with.
Deploying in the H/ᾱ plane the pairs of Entropy and ᾱ values of the 6
different test zones present in Table 8 for the three different scenes, it
came up a remarkable result. As it can be observed in Figure 4.11 all
4.4 Snow Monitoring 74
the test zones show a similar increasing slope function except the 6th one
(urban zone). Notice that the dots in the bottom left of each function
correspond to the March scene and the ones in the top right part of the
functions are from the October scene. This means that the presence of
dry snow with respect to the no snow scene decreases the Entropy and the
ᾱ parameters at C-band, thus decreases the randomness of the scattering
phenomena and surprisingly moves the ᾱ parameter to a more random
surface scattering phenomena. In yellow, as expected due to its more
deterministic nature, the urban test zone shows no Entropy variation and
almost to ᾱ variation among the different scenes. Recall that scenes 1
to 5 have different characteristics of height, surface coverage and slope,
and therefore the scattering phenomena in each zone might be different.
However, concerning the presence of snow, the resulting H/ᾱ plot is similar
for all the test zones. This remarkable behavior found in the curves for
test zones 1 to 5 might be further investigated as a means of determining
dry snow covered pixel with a change detection based technique.
With the Bragg surface hypothesis discarded, since probably the rough-
ness of the alpine meadows at C-band does not fulfill the Bragg surface
condition of KS < 0.3, where K is the wave number and s is the rms sur-
face roughness, the X-Bragg model hypothesis was then considered. The
model to be used has to be suitable to deal with rougher surfaces and the
X-Bragg [25, 4] consists of a surface scattering model that accounts for
high rms roughness surfaces. It introduces roughness disturbance model-
ing the surface as a reflection symmetric depolarizer by rotating the bragg
coherency matrix [T ] an angle β about the LOS. The width of the proba-
bility distribution function of β of the model is proportional to the surface
roughness. Opposite to the Bragg Coherency matrix case, there is cross-
polarization, and therefore the polarimetric coherence is ≤ 1. Assuming
an X-Bragg surface, which involves rough surfaces enabling more than just
surface scattering mechanisms, the presence of double bounce scattering
in the test zones could be explained.
4.5 Preliminary Results 75
As final for this section, it could be stated from the empirical observations
in this work, that the presence of dry snow has a smoothing effect on the
underlying surface, involving the slight decrease of double bounce scatter-
ing mechanism and the slight increase of the surface scattering mechanism,
with a reduction of the Entropy parameter.
snow ground truth scene, and ∆b represents the difference between the wet snow
scene and the non snow ground truth scene. See in Figure 4.12 what these two
differences look like.
Figure 4.12: :7x7 Boxcar obtained ∆ differences in dB in Pla de Beret Zone for
dry and wet snow scenes respect to no snow ground truth scene. Axis units are
pixels.
First of all note the difference between the dry and the wet scenes differences,
suggesting sensitivity to the kind of snow. Notice as well, concerning differences
between ∆a and ∆b , how the wet snow scene difference shows values lower than
those of the dry snow scene difference. Individually analyzed, looking in the
∆a image at Pla de Beret zone, around the pixel (x=1700, y=2700), with an
upper right diagonal shape, it can be noticed that most values are above 1dB,
indicating dry snow covered area. Indeed, having a look at the ∆0217 image
in Figure 4.9 it can be sensed the shape of the mountains and the valleys in
between them, in the zone illuminated by the radar, as it can be confirmed if
attention is put in the comparison of ∆ in Figure 4.9 and Figure 4.13, which
is the topographic map of the PolSAR data covered zone. In order to better
appreciate whether this topography like shape exists, indicating snow in the
higher parts of the map, i.e. the mountains, a magnitude filter was applied to
∆a with a binary mask showing those pixels with a value higher than 1dB, that
might be dry snow covered. The result of this masking can be seen in Figure
4.14.
It is worth to point that the threshold at 1dB has been decided as such with
a very few amount of ground truth data. Recall that this value was adopted
out of the results in Table 9, and further work should be done with a complete
ground truth covering the whole image, in order to optimize this threshold for
the classifier. In addition some empirical essays have been conducted looking
for the value of the threshold showing better visual results with regards to the
topography map in Figure 4.13. Again, it is worth to mention that a layover-
shadowing mask should be applied to this obtained above-1dB filtered image
(Figure 4.14), in order to assess which are the pixels with reliable values and
which are not. Nonetheless, it can be clearly identified in Figure 4.9 the valley
4.5 Preliminary Results 77
Figure 4.14: :Diff-a image filtered to those pixels with a value higher than 1dB,
indicating the presence of dry snow. White and black pixels indicate snow and
no snow respectively. Axis units are pixels.
Vall d’Aran in the bottom-left area of the figure images, as well as the shape of
the mountains that we suppose covered by dry snow.
As the reader might have already seen, the results displayed in the previous
images 4.12 and 4.14 are quite noisy. In order to try to reduce the noise level,
the ∆ parameter was recomputed using a 11x11 pixel Boxcar filter. Results
from the more filtered ∆ are shown as follows in Figure 4.15.
4.5 Preliminary Results 78
Figure 4.15: :∆ (11x11 Boxcar used) representation of the co-registerd data for
the three different scenes. Values range from -13dB to 23dB and axis units are
pixels
Computing again the differences of ∆ as in Eq 4.7, for the dry and wet snow
scenes with respect to the no snow ground truth scene, the results, shown in
Figure 4.16, are quite similar to those in Figure 4.12 obtained with a smaller
Boxcar filter. In this case, the whole image is shown so that the geographical
valleys can be appreciated in the bottom-left part of the ∆a image on the left
of the Figure 4.16. Again, manually adjusting a threshold so that the masked
image looks like similar to the topography of the illuminated area, it has been
optimized the dry snow mask shown in Figure 4.17. Notice that its aspect is
less noisy than the one in Figure 4.14, as expected taking into account that
more averaging has been applied to compute the ∆ matrices that led to this
difference image, however the dry snow mask still looks too noisy. Probably a
better Speckle filtering method would lead to better results, since Boxcar kind
of filter has very poor results in the non-homogeneities of the image. Further
word should include better Speckle filtering techniques in order to compute more
reliably the ∆ parameter.
4.5 Preliminary Results 79
Figure 4.16: :11x11 Boxcar obtained ∆ differences in dB for dry and wet snow
scenes respect to no snow ground truth scene. Axis units are pixels.
Figure 4.17: :Dry snow covered pixels mask. White and black pixels indicate
snow and no snow respectively. Axis units are pixels.
with dry snow or not is described in Eq 4.8. The values of the no snow to
dry snow thresholds for Entropy (0.925) and ᾱ (0.9) have been decided as such
observing the values represented in Figure 4.11.
Figure 4.18: :H/ᾱ based dry snow covered pixels mask. White and black pixels
indicate snow and no snow respectively. Axis units are pixels.
Better Speckle filtering techniques are needed to further assess the poten-
tial trajectory of the presented approaches. Using less noisy Coherency
Matrix diagonal components T11 and T22 , the resulting parameters would
be much less noisier than they are right now. Advanced polarimetric
speckle filtering techniques are needed in order to be able to filter a very
in-homogeneous data such as the one used for this research work, and
being sure that the source data used to compute ∆ and H/A/ᾱ decom-
position is optimized from the noise reduction point of view. In such a
case, results in detecting dry snow covered area and differentiating with
wet snow would give a more realistic idea of the potential extent of these
polarimetric presented parameters or methods.
It might be appropriate as well to think about some kind of post-filtering
technique for the obtained results using simple decision thresholds. Some
computational intelligence techniques for classification might be studied
in order to cluster in different well defined zones, the obtained boolean dry
snow maps. Neural Networks might be useful in this case since probably
some patterns of ancillary data variables can be found in the contour
conditions of each pixel bringing valuable information for deciding about
the presence and kind of snow.
A wider dataset and ground truth would allow a much more realistic as-
sessment of the potential of the classification results. This means having
more fully polarimetric C-band SAR images, covering many different con-
ditions snow-packs, as well as an extended ground truth information about
the snow-pack and the snow covered area, allowing a robust validation of
the results obtained. An interesting first approach could be using MODIS
hyper-spectral data of the Pyrenees zone to compute the Normalized Dif-
ferential Snow Index (NDSI) with MODIS bands 4 and 6. This would
provide with an initial ground truth map of Snow Covered Area (SCA),
although with very poor spatial resolution of 500m per pixel.
Getting back to the initial goal of this Master thesis, recall that this targeted
Snow Water Equivalent (SWE). The idea was quantitatively sensing some snow
parameter related with SWE, such as density or permittivity, using Polarimet-
ric SAR data at C-band. Unfortunately no robust conclusion can be extracted
about this main goal, since despite C-band polarimetric SAR data presented
sensitivity to the presence of snow, the research carried out so far can not assess
the potential for quantitative snow remote sensing. Further research needs to
be conducted on this issue, with a more ambitious approach, in order to have a
much wider dataset, so that snow remote sensing potential with C-band polari-
metric SAR data can be correctly determined. The global conclusion achieved
84
after this Master thesis research work is that, despite the widely extended idea
that C-band microwaves are not specially sensitive to the presence of snow,
Polarimetry adds some additional information whose contribution is yet to be
carefully investigated because it seems that it increases the snow remote sensing
capabilities of C-band SAR.
REFERENCES 85
References
[1] Christian Matzler, Applications of the interaction of Microwaves with the
Natural Snow Cover, Remote Sensig Reviews, Vol 2, 1987. 28, 30, 34, 35,
37, 48, 53
[2] Fawwaz T. Ulaby, Richard K. Moore, Adrian K. Fung, Microwave Remote
Sensing: Active and Passive, from Theory to Applications,Artech House,
1986. 34, 35, 36, 37, 48, 54, 55
[3] Jong-Sen Lee, Eric Pottier, Polarimetric Radar Imaging. From basics to
Applications, CRC Press, 2009. 11, 12, 21, 30, 38, 39, 43, 50, 73
[4] S. R. Cloude, Polarisation. Applications in Remote Sensing , Oxford Uni-
versity Press, 2010. 30, 39, 43, 73
[5] Tazio Strozzi, Shi, Christion Matzler, Backscattering Measurements of
Alpine Snowcovers at 5.3 and 35 GHz, IEEE Transactions on Geoscience
and Remote Sensing, May 1998. 38, 53, 55, 63
[6] Singh, G. Venkataraman, G., Snow density estimation using polarimetric
ASAR data, Recent Advances in Microwave Theory and Applications, 2008.
MICROWAVE 2008. International Conference on . 35, 39, 41, 44
[7] Audrey Martini, Teledetection d’un couvert neigeux en milieux alpins a par-
tir de donnes SAR polarimtriques multi-frquentielles et multi-temporelles,
PhD Thesis, Universite Rennes, 2005. 37, 38, 39, 48
[8] A. Martini et al., Polarimetric Study of Scattering from Dry Snow Cover
in Alpine Areas, Proceedings IGARSS 2003. 53, 63, 71
[9] Jiancheng Shi, Jeff Dozier, Estimation of Snow Water Equivalent using
SIR-C/X SAR Part I: Inferring snow density and subsurface properties,
IEEE Transactions on Geoscience and Remote Sensing, Nov 2000. 37, 39,
40, 41, 53, 64
[10] Jiancheng Shi, Jeff Dozier, Estimation of Snow Water Equivalent using
SIR-C/X SAR Part II: Inferring snow depth and particle size, IEEE Trans-
actions on Geoscience and Remote Sensing, Nov 2000. 39
[11] Monique Bernier et al., The Potential of Time Series of C-Band SAR Data
to Monitor dry Shallow Snow Cover, IEEE Transactions on Geoscience and
Remote Sensing, Jan 1998. 39, 42, 47
[12] Nicolas Longepe et al., snow-pack Characterization in Mountainous Regions
Using C-Band SAR Data and a Meteorological Model, IEEE Transactions
on Geoscience and Remote Sensing, Feb 2009. 39, 41
[13] Zhen Li et al., Estimation of Snow Density with L-Band Polarimetric SAR
data, IEEE IGARS Proceedings, 2000. 41
[14] Singh, G. Venkataraman, G., Snow density estimation using Polarimetric
ASAR data, IEEE IGARS Proceedings, 2009. 39, 41
[15] Snehmani et al., Development of an inversion algorithm for dry snow
density estimation and its application with ENVISAT-ASAR dual co-
polarization datal, Geocarto International Taylor & Francis, Dec 2010. 39,
41, 45
REFERENCES 86
[16] Jacob J. van Zyl et al, Requirements for model-based polarimetric decom-
positions, IEEE IGARS Proceedings, 2008. 44
[17] Jayanti J. Sharma et al., Polarimetric Decomposition Over Glacier Ice Us-
ing Long-Wavelength Airborne PolSAR, IEEE Transactions on Geoscience
and Remote Sensing, Jan 2011. 37, 44, 45
[18] Jayanti J. Sharma et al., Characterization of Oriented Volumes in Glacier
Ice and Extinction Inversion with PolInSAR, Proceedings PolInSAR, 2009.
[21] Maxim Neuman et al., Forest Parameter Retreival using a General Repeat-
pass Polarimetric Interferometric Vegetation Mode, IEEE IGARS Proceed-
ings, 2009. 46
[22] Yoshio Yamaguchi et al., Four-Component Scattering Model for Polari-
metric SAR Image Decomposition, IEEE Transactions on Geoscience and
Remote Sensing, Aug 2005. Geophys. Res. Lett., Vol. 31, 2004. 46
[23] Nagler, T. and Rott, H., Retrieval of wet snow by means of multitemporal
SAR data, IEEE Transactions on Geoscience and Remote Sensing, Aug
2002. 30
[24] Jong-Sen Lee et al., Polrimetric SAR Data Compensation for Terrain Az-
imuth Slope Variation, IEEE Transactions on Geoscience and Remote Sens-
ing, Sep 2000. 71
[25] Irena Hajnsek et al., Inversion of Surface Parameters from Polarimetric
SAR, IEEE Transactions on Geoscience and Remote Sensing, Apr 2003. 73