You are on page 1of 17

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO.

7, JULY 2011 2531

Retrieval of Soil Surface Parameters via a


Polarimetric Two-Scale Model
Antonio Iodice, Senior Member, IEEE, Antonio Natale, Student Member, IEEE, and
Daniele Riccio, Senior Member, IEEE

Abstract—We propose a polarimetric two-scale surface scatter- to as SPM) is a good candidate for slightly rough surfaces,
ing model employed to retrieve the surface parameters of bare soils since, according to it, the dependence on roughness cancels out
from polarimetric synthetic aperture radar or scatterometer data. in the copolarized (co-pol) ratio (i.e., the ratio of vv and hh
The scattering surface considered here is composed of slightly ◦ ◦
rough randomly tilted facets, for which the small perturbation normalized radar cross sections (NRCSs), σVV and σHH ), so
scattering method holds. The facet random tilt causes both a that this ratio only depends on the (complex) dielectric constant.
random variation of the local incidence angle and a random However, the SPM range of validity is limited to rough-
rotation of the local incidence plane around the line of sight. ness whose height deviations are small with respect to the
Unlike other similar already existing approaches, our method incident wavelength [1]. In addition, SPM fails to describe
considers both these stochastic effects in the analytical evaluation
of the facet scattering matrix, and their statistical modeling is cross-polarization and depolarization effects usually observed
derived from a proper statistical description of the scattering in actual data. Conversely, second-order SPM [1] and other
surface. In particular, we assume that the facet slope (i.e., the more refined methods (e.g., Integral Equation Method [2])
slope of the large-scale surface roughness) is a Gaussian random account for cross-polarization, but their formulations are not
variable, in agreement with both classical and fractal surface in closed form and are too involved to be usefully employed
models. The proposed scattering model is then used to retrieve
bare soil moisture and (large-scale) roughness from the co- and in surface parameter retrieval algorithms. A more manageable
cross-polarized ratios. The performance of the resulting retrieval expression of the scattered field, which allows extending the
algorithm and its validity limits are finally assessed by comparing SPM range of validity, is obtained by the so-called two-scale
obtained results to “in situ” measurements. To this aim, data models [1], [3]–[6]. However, classical two-scale models fail
from different measurement campaigns available in literature are to describe depolarization effects [1], [7], so that, in their
employed.
usual formulations, they cannot be usefully employed in surface
Index Terms—Polarimetry, retrieval of natural soil parameters, parameter retrieval algorithms.
rough surfaces, synthetic aperture radar (SAR). In order to widen the SPM range of validity and to account
for both cross-polarization and depolarization effects while
I. I NTRODUCTION maintaining a reasonably simple formulation, in [7], a different
approach, there termed X-Bragg, was recently presented. The
T HE FIELD scattered by a bare soil depends on both
surface roughness and soil dielectric constant, and hence,
on the terrain water content. Accordingly, it is possible, at least
model described in [7] is a more refined and more satisfactorily
formulated version of the tilted Bragg approximation intro-
in principle, to retrieve soil moisture and roughness from ap- duced in [8]. The X-Bragg model can be physically interpreted
propriate scattering measurements. These retrieval approaches as follows: The scattering surface is considered as composed
require multiple scattering measurements, supported by the use of rough randomly tilted facets, large with respect to the wave-
of general and accurate, but not too involved, scattering mod- length, but small with respect to the sensor resolution. Scatter-
els. As for measurements, multipolarimetric and/or multiangle ing from each rough facet is still evaluated by employing the
synthetic aperture radar (SAR) or scatterometer data can be SPM, but the facet random tilt causes both a random variation
used; as for scattering models, here, we focus on theoretical Δϑ of the local incidence angle ϑl with respect to the look angle
ones, among which the first-order small perturbation method ϑ and a random rotation β of the local incidence plane around
(SPM) [1] (often, as well as in the following, simply referred the line of sight. In [7], the random incidence angle variation
Δϑ is ignored, and the incidence plane angle of rotation β is
heuristically assumed to be uniformly distributed in an interval
Manuscript received December 1, 2009; revised September 29, 2010 (−β1 , β1 ), where the parameter β1 is used to characterize
and December 20, 2010; accepted January 9, 2011. Date of publication the surface roughness. The coherency matrix elements of the
February 17, 2011; date of current version June 24, 2011. This work was
supported in part by Agenzia Spaziale Italiana within COSMO/SkyMed AO, overall surface are then obtained from those of a single facet by
project 2202, and Carlo Gavazzi Space within the MORFEO Project under averaging the corresponding SPM expressions over β, and it is
Contract 2092A/08/15. shown that this introduces both cross-polarization (i.e., nonnull
The authors are with the Department of Biomedical, Electronic and
Telecommunication Engineering, University of Naples “Federico II,” 80125 cross-polarized (cross-pol) matrix elements) and depolarization
Naples, Italy (e-mail: iodice@unina.it; antonio.natale@unina.it; daniele. (i.e., nonunity copolarized (co-pol) correlation coefficient). It is
riccio@unina.it). also shown in [7] that the obtained formulations of coherency
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. matrix entropy H and mean scattering angle α [7] only depend
Digital Object Identifier 10.1109/TGRS.2011.2106792 on dielectric constant and roughness parameter β1 . These latter

0196-2892/$26.00 © 2011 IEEE

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
2532 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 7, JULY 2011

parameters can be therefore retrieved from H–α pairs com-


puted from fully polarimetric SAR data. Retrieval results over
bare or little vegetated soils are very promising, but not always
fully satisfactory.
In this paper, in order to improve the model, both the sim-
plifying assumptions of [7] (i.e., no Δϑ and uniform β) are
removed. First of all, here, the roughness within a facet is iden-
tified as the “small-scale” roughness and the roughness caused
by the random facets’ slopes as the “large-scale” roughness.
Then, a stochastic process describing the scattering surface
is used, which leads to the assumption that the facet slopes
along range and azimuth directions are independent zero-mean
Gaussian random variables with the same standard deviation σ.
This standard deviation is used to characterize the large-scale Fig. 1. From the (a) real surface to the (b) faceted one.
roughness. Owing to the available relations linking β and ϑl to
the surface slopes [9], NRCSs and the coherency matrix of the
overall surface are then obtained from those of a single facet
by averaging the corresponding SPM expressions directly over
the random facet slopes. It turns out that this also introduces
both cross-polarization and depolarization. In addition, H and
◦ ◦
α, and the co- and cross-pol NRCS ratios, i.e., σVV /σHH and
◦ ◦
σHV /σVV , turn out to be only very slightly dependent on
small-scale roughness and mainly depend only on dielectric
constant and large-scale roughness σ. Accordingly, these latter
parameters can be retrieved from H–α pairs or also from the
co- and cross-pol ratios. In fact, the H–α method requires fully Fig. 2. Three-dimensional sketch of a generic tilted facet.
polarimetric measurements, i.e., measurement of amplitude and
phase of all the scattering matrix entries, thus implying the the electromagnetic wavelength and than the correlation length
need of accurate interchannel calibration of both amplitude and of the small-scale roughness, but they are much smaller than
phase. Although fully polarimetric SAR sensors are available the sensor geometric resolution and than the correlation length
today (e.g., PALSAR, E-SAR, and TERRASAR-X) and accu- of the large-scale roughness.
rate calibration is state of the art today, there are still important Therefore, by using the reference system shown in Fig. 2, the
SAR systems (e.g., ENVISAT and COSMO/Skymed) that only randomly rough and randomly tilted facets are defined through
implement selectable dual-polarization (dual-pol). Therefore, the following:
it may be sometimes useful to resort to only-amplitude mea-
surements. This can be done by using NRCS-based co- and z(x, y) = a(x − xi ) + b(y − yi ) + zi + δ(x, y)
cross-pol ratios, which only need relative amplitude calibration. ∀(x, y) ∈ Di (1)
The co-pol/cross-pol method has the additional advantage that
selectable dual-pol systems with a very fast repeat–pass time where a = tan ω and b = tan γ are the local azimuth and range
(such as the COSMO/Skymed constellation) can be used: Two slopes, respectively, xi , yi , and zi are the coordinates of the ith
passes with very short time separation can be used to measure facet center, δ(x, y) describes the small-scale roughness, and
the two ratios. The co-pol/cross-pol method is preferred here Di is the ith facet domain.
mainly because it can be fully validated by using a wide variety Both large- and small-scale roughnesses are modeled as
of measurement campaign data available in literature, for which stochastic processes. In particular, we assume that the facet
only-amplitude data are reported, see Section IV. However, we slopes along range and azimuth directions are independent
stress again that the proposed model can be employed also identically distributed zero-mean σ 2 -variance Gaussian ran-
within the H–α method. dom variables, i.e., a, b ∼ N (0, σ 2 ). This assumption only
requires that the large-scale roughness is a Gaussian statisti-
II. P OLARIMETRIC T WO -S CALE M ODEL cally isotropic stationary-increment (i.e., locally homogeneous)
process, so that it is compatible with both classical surface mod-
A. Surface Model
els, employing stationary (i.e., homogeneous) processes [1],
A bare soil surface is considered here as composed of large- [10], and fractal surface descriptions, employing the stationary-
scale variations on which a small-scale roughness is superim- increment fractional Brownian motion (fBm) processes
posed, so that a two-scale model of the surface is employed. [11], [12].
Concerning the large-scale roughness, it is locally treated by With regard to the small-scale roughness δ(x, y), it is mod-
replacing the surface with a slightly rough tilted facet, whose eled as a zero-mean stochastic process whose height standard
slope is the same as that of the smoothed surface at the center deviation s is small compared with the electromagnetic wave-
of the pertinent facet (see Fig. 1). Facet sizes are greater than length λ. In the following, it is assumed that δ(x, y) is a

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
IODICE et al.: RETRIEVAL OF SOIL SURFACE PARAMETERS VIA A POLARIMETRIC TWO-SCALE MODEL 2533

Since we have assumed s  λ, we can use the first-order


SPM and evaluate the field scattered by the facet as the sum
of a zeroth-order (or coherent or mean) field, which is the field
scattered by a smooth facet, and a first-order (or incoherent)
field, which is a zero-mean random variable (due to the facet
random roughness). The coherent field is negligible, except that
in a very narrow cone around ϑl = 0 (i.e., a = 0 and b = tan ϑ)
whose angular width is of the order of λ/L, its expression
depends on (smooth) facet shape, orientation, and size and it
s
is of no concern here. The incoherent first-order field Epq is
given by [1], [7], [10]

s
Epi k 2 cos2 ϑl
Epq = χpq (ϑl , β)δ̃(0, 2k sin ϑl ) exp(−jkr)
Fig. 3. Rotation of local incidence plane due to facet tilt.
πr
(5)
(band-limited) fBm process, so that its power spectral density where Epi is the incident field, k = 2π/λ is the wavenumber,
(PSD) is [12], [13]  x , κy ) is the Fourier trans-
r is the radar-to-target distance, δ(κ
S0 S0n form of δ(x, y), p and q are the polarizations of the incident
W (κ) = (2+2H
= s2 = s2 Wn (κ) (2)
κ t) κ (2+2H t) and scattered fields, respectively, and can each stand for H
 (horizontal) or V (vertical), χpq (ϑl , β) are the elements of the
where κ = κ2x + κ2y , κx and κy are the Fourier mates of x and matrix
y, respectively, 0 < Ht < 1 is the Hurst coefficient, related to  
FH (ϑl ) 0
the roughness fractal dimension D by the relation D = 3 − Ht , χ(ϑl , β) = R2 (β) · · R2 −1 (β) (6)
0 FV (ϑl )
S0 is directly proportional to the roughness variance s2 via the
 
dimensional facet-size-dependent constant S0n , the expression cos β sin β
R2 (β) = (7)
of which is reported in [12], and Wn (κ) is the normalized PSD. − sin β cos β
This small-scale-roughness description is chosen here because
there is a wide consensus on the fact that fBm is a proper model is the 2 × 2 unitary rotation matrix, and FH and FV are
for natural surfaces (see, e.g., [11] and the references therein): the SPM (or Bragg) coefficients for H and V polarizations,
In fact, the small-scale roughness of most natural bare soils respectively
turns out to satisfy (2) with Ht ranging from about 0.4 to about ⎧ √
0.8. However, other surface models, and hence spectra, can be ⎪
⎨ FH =
cos ϑl − ε−sin2 ϑl

cos ϑl + ε−sin2 ϑl
used in our approach as well. Note finally that, due to the finite (8)

⎩ FV = (ε − 1)

sin2 ϑl −ε(1+sin2 ϑl )
√ 2
facet size, (2) holds for κ values greater than about 2π/L, where ε cos ϑl + ε−sin2 ϑl
L is the facet linear size.
with ε being the soil relative dielectric constant (or relative
permittivity).
B. Single Tilted-Facet Return
The rotation matrix R2 accounts for the rotation of the local
Now, let us assume that the sensor illuminates the scene with polarization reference system with respect to the global one,
a (global) incidence angle ϑ (see Fig. 3), and let us consider the which is caused by the incidence plane rotation and gives rise
field scattered by a single tilted rough facet. Due to the facet to the cross-polarization effect. This effect is often not included
tilt, the local incidence angle ϑl (see Fig. 3) is different from in usual two-scale models [1], but it is considered in [7]. On
the global one and is related to it and to the facet slopes by [9] the other hand, in [7], ϑl is approximated by ϑ in (5), (6), and
cos ϑ + b sin ϑ (8). As already underlined, we here consider both the incidence
cos ϑl = √ . (3) plane rotation and the variations of ϑl .
1 + a 2 + b2
The facet’s (incoherent) NRCS is defined as
In addition, the facet tilt causes a rotation of the local inci- 
dence plane around the look direction k̂ by an angle β related 4πr2 Epq s 2
0 |δ
to the facet slopes and to the global incidence angle by1 [9] σpq = 2 (9)

A Ep i
a
tan β = . (4)
−b cos ϑ + sin ϑ where A is the facet’s area and the symbol f |ξ stands for “the
mean of f with respect to the random variable ξ.” Accordingly,
1 From (3), it turns out that, for a small facet’s slope, | cos ϑ − cos ϑ| ∼
l = in (9), the square modulus of the scattered field is averaged over
sin ϑ · |Δϑ| ∼ = b sin ϑ so that the standard deviation of Δϑ is approximately
σ, whereas from (4), tan β ∼ =β∼ = a/ sin ϑ so that the standard deviation of β the facet (i.e., small-scale) roughness. By using (5) in (9), we
is approximately σ/ sin ϑ, which is larger than σ. This is in agreement with the get [1], [10]
intuition that the rotation angle β shows a wider distribution than the incidence
angle variation Δϑ, but this also shows that the latter is not negligible with 4 4
respect to the former.
0
σpq = k cos4 ϑl |χpq |2 s2 Wn (2k sin ϑl ). (10)
π

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
2534 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 7, JULY 2011

Full expressions of NRCS, obtained by using (6) and (7) in C. Total Scattered Power
(10), are
If the large-scale-roughness height variations are larger than
⎧ 0 4 4 the wavelength and the facet size is larger than small-scale-
⎪ 4 2
⎪ σHH = π k cos ϑl s Wn (2k sin ϑl )

⎪ 2 roughness correlation length, then the returns from different


⎪ × FH (ϑl ) cos2 β + FV (ϑl ) sin2 β facets are uncorrelated, and the NRCS and coherency matrix
⎨ 0
σVV = π4 k 4 cos4 ϑl s2 Wn (2k sin ϑl ) of the whole surface can be obtained by averaging those of a
2 (11)


⎪ × FV (ϑl ) cos2 β + FH (ϑl ) sin2 β single facet over β and ϑl or, equivalently, over a and b.

⎪ σ 0 = σVH
0
= π4 k 4 cos4 ϑl s2 Wn (2k sin ϑl )

⎪ With regard to the zeroth-order field, this leads us to ob-
⎩ HV
× |[FV (ϑl ) − FH (ϑl )] sin β cos β|2 . tain (independent of facets’ shape and size) the well-known
Kirchhoff Approach (KA) Geometrical Optics (GO) solution.
Note that these expressions do not hold at near-grazing an- This GO term is dominant at small global incidence angles, but
gles (i.e., ϑl ∼
= π/2), because first-order SPM does not account it is negligible at the usual SAR look angles for moderate values
for multiple scattering, and at near-vertical incidence, because of σ, so that we will ignore it.
(2) does not hold for κ values smaller than about 2π/L, i.e., Let us now move to the first-order SPM term, which is of
sin ϑl smaller than about λ/(2L), and, in addition, at these major interest here. If, as in [7], ϑl is approximated with ϑ and β
small local incidence angles, the coherent field is dominant. is assumed uniformly distributed, it is not difficult to verify, by
However, these two extreme situations are very unlikely to averaging over β, that the overall surface NRCS and coherency
happen for usual SAR system look angles and for moderate matrix have the expressions (16) and (17), shown at the bottom
facet root mean square (rms) slope σ. of the next page.
The facet’s coherency matrix is defined as [14] Note that, as expected, (17) coincides with the result of [7].
 Conversely, here, we want to start from the large-scale-
T = k P k †P (12) roughness statistical description of Section II-A. In order to
|δ average over β and ϑl , we should evaluate the joint probability
density function (pdf) of these two random variables, starting
where † denotes the transposed conjugate and k P is the 3-D from the one of a and b. However, although it is possible,
Pauli scattering vector invoking the fundamental random variable transformation the-
orem, to analytically compute the exact marginal pdf of β (see
1 [16]), the joint pdf of β and ϑl cannot be analytically evaluated.
k P = √ ( SHH + SVV , SHH − SVV , 2SHV )T (13)
2 Accordingly, we have to resort to the average over a and b,
which, first of all, implies that we have to express (11) and
with Spq being the elements of the facet’s scattering matrix, (15) in terms of a and b by using (3) and (4). This is quite
which, in the SPM case, can be expressed as straightforward, since it is easy to express all terms appearing
in (11) and (15) in terms of cos ϑl and tan β.
k 2 cos2 ϑl At this point, if f (a, b) is the facet generic NRCS or co-
Spq = χpq (ϑl , β)δ̃(0, 2k sin ϑl ). (14)
π herency matrix entry, then the expected value of f can be
computed as
Full expressions of the coherency matrix, obtained by using
(6), (7), (13), and (14) in (12), are given in (15), shown at the +∞ +∞  2 
bottom of the page. f (a, b) a + b2
f (a, b) |a,b = exp − da db.
The expressions in (15) have the same incidence angle range 2πσ 2 2σ 2
− cot(ϑ) −∞
of validity as those in (11). Finally, notice that, as the coherency
(18)
matrix for a single facet is a rank-one matrix, then just a surface
scattering mechanism is involved in its formation [7], [14], [15].
Of course, the ensemble averaging over β and ϑl will introduce Note that integration over b is restricted to the range
other degrees of freedom in the coherency matrix, which are − cot ϑ < b < +∞ to exclude values of ϑl greater than π/2
used to define quantities like entropy and mean scattering angle [see (3)], which correspond to shadowed facets (more exactly,
[14], [15]. these are the “self-shadowed” facets, whereas we are neglecting

(k cos ϑl )4 s2 Wn (2k sin ϑl )


T=
2π 2
⎛ ⎞
|FH (ϑl )+FV (ϑl )|2 (FH (ϑl )+FV (ϑl ))(FH (ϑl )−FV (ϑl ))∗ cos 2β −(FH (ϑl )+FV (ϑl ))(FH (ϑl )−FV (ϑl ))∗ sin 2β
×⎝(FH (ϑl )+FV (ϑl ))∗ (FH (ϑl )−FV (ϑl )) cos 2β |FH (ϑl )−FV (ϑl )|2 cos2 2β −|FH (ϑl )−FV (ϑl )|2 sin 2β cos 2β ⎠
−(FH (ϑl )+FV (ϑl ))∗ (FH (ϑl )−FV (ϑl )) sin 2β −|FH (ϑl )−FV (ϑl )|2 sin 2β cos 2β |FH (ϑl )−FV (ϑl )|2 sin2 2β

(15)

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
IODICE et al.: RETRIEVAL OF SOIL SURFACE PARAMETERS VIA A POLARIMETRIC TWO-SCALE MODEL 2535

the case of facets shadowed by other facets). However, the Hence, by using the formula to obtain moments of any order
probability of self-shadowed facets is for a zero-mean Gaussian variable Z ∼ N (0, σ 2 )
  
1 cot ϑ
Pr{b < − cot ϑ} = Erf c √ (19) Z n =
0, if n is odd
(22)
2 2σ 1 · 3 · 5 · . . . · (n − 1)σ n , if n is even
which is very small for non-near-grazing angles and moderate
values of σ. Accordingly, in this case, the integrals over a and it is possible to compute the averages of the NRCS and co-
b can both span the entire real axis. In addition, it must be herency matrix elements. As shown in Appendix A, considering
noted that, in a small region of the a, b plane around a = 0 and Taylor expansion terms up to the second order, this leads to the
b = tan ϑ, which corresponds to ϑl ∼ = 0, the first-order SPM following:
field expressions (11) or (15) do not hold. However, for non- ⎧    
⎪   Re{C0,0
HV
}−C0,0
HH
near-vertical incidence and, again, moderate values of σ, we ⎪ σ0
⎪ = 4
C HH
+ C HH
+2 +C HH 2
⎪ 0,2 σ
can ignore this effect. ⎪

HH |a,b π

0,0

2,0 sin2 ϑ
 
 0  Re{C0,0
HV
}−C0,0
VV
Unfortunately, even with the simplifications aforementioned, ⎪ σ = 4
C VV
+ C VV
+2 +C VV 2
⎪ VV |a,b π 0,0 2,0 sin2 ϑ 0,2 σ
the integral in (18) cannot be computed in closed form. How- ⎪
⎪  
 

⎩ σ0 σ2
HV |a,b = π C0,0 +C0,0 −2Re C0,0
4 HH VV HV
ever, assuming small values for facet slopes, the Taylor expan- sin2 ϑ
sion of f around a = 0 and b = 0 can be used.
(23)
Accordingly, let us consider the MacLaurin series of f (a, b) ⎛ ⎞
 n      t11 t12 0

+∞
1  n ∂ n f 1
T |a,b = 2 ⎝t∗12 t22 0⎠ (24)
k n−k
f (a, b) = a b
n=0
n! k ∂ak ∂bn−k a=b=0 2π
0 0 t33
k=0
(20)
where t11 , t12 , t22 , and t33 are denoted as in (25), shown at the
then, the expected value of f , recalling that a and b are pq
bottom of the page, in which Ck,n−k are the series expansion
independent and identically distributed, is
coefficients of the function (k cos ϑl )4 W Fp Fq∗ , which appears
f (a, b) |a,b in all NRCS and coherency matrix entries
 n     

 1  n ∂ n f  k   n−k  1 n ∂ n W k 4 cos4 ϑl Fp Fq∗
+∞
pq
= a b . (21) Ck,n−k = . (26)
n=0
n! k ∂ak ∂bn−k a=b=0 n! k ∂ak ∂bn−k
k=0 a=b=0

⎧   ! ! !"
⎪ |FH (ϑ)|2 |FV (ϑ)|2 |FH (ϑ)−FV (ϑ)|2

⎪ σ 0
= 4 4
k cos 4
ϑs 2
W n (2k sin ϑ) 1+ sin 2β1
+ 1− sin 2β1
− 1− sin 4β1
⎨ HH |β π  2 2 2β1
! 2 2β1
! 8 4β1
!"
|FV (ϑ)| |FH (ϑ)|2 |FH (ϑ)−FV (ϑ)|2
0 4 4 4 2
σVV |β = π k cos ϑs Wn (2k sin ϑ) 1+ sin 2β1
+ 1− sin 2β1
− 1− sin 4β1


2 2β1 2
! 2β1 8 4β1
⎩σ 0  = σ 0  = 4 k 4 cos4 ϑs2 Wn (2k sin ϑ) |FH (ϑ)−FV (ϑ)|2 1− sin 4β1

HV |β VH |β π 8 4β1

(16)
4 2
(k cos ϑ) s Wn(2k sin ϑ)
T |β =
2π 2
⎛ ⎞
|FH (ϑ)+FV (ϑ)|2 (FH (ϑ)+FV (ϑ))(FH (ϑ)−FV (ϑ))∗ sin2β2β 1
0
⎜ ! 1

×⎜(F (ϑ)−FV (ϑ))(FH (ϑ)+FV (ϑ))∗ sin2β2β 1
|FH (ϑ)−FV (ϑ)|2 12 1+ sin4β4β 1
0 ⎟
⎝ H 1 1
!⎠
0 0 |FH (ϑ)−FV (ϑ)|2 12 1− sin4β4β
1
1

(17)

⎧  HV 
HH  HV  2
⎪ HH VV VV HH VV HV
⎪ t11 = C0,0 +C0,0 +2Re C0,0 + C2,0 +C2,0 +C0,2 +C0,2 +2Re VVC2,0HH+C0,2 %σ HV

⎪  HV  &!

⎨ t12 = C0,0
HH
−C0,0
VV
−2jIm C0,0 + C2,0 HH
−C2,0
VV HH
+C0,2 −C0,2
VV C −C0,0
+2 0,0 +2jIm 2
C0,0
−C HV
−C HV
σ2
2
sin ϑ 2
sin ϑ 2,0 0,2
 HV  % &! (25)
⎪ C VV+C0,0HH
C HV


HH
t22 = C0,0 VV
+C0,0 −2Re C0,0 + C2,0 HH VV
+C2,0 HH
+C0,2 VV
+C0,2 −4 0,0 −2Re C2,0
HV HV
+C0,2 −4 sin0,0 σ2


 
sin2 ϑ 2ϑ

⎩ 2
HH
t33 = C0,0 VV
+C0,0 −2Re C0,0
HV
4 sinσ2 ϑ

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
2536 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 7, JULY 2011

equal tan β variances in the uniform β and Gaussian tan β


cases, we obtain
'
σ tan β1 √ σ
= − 1 ⇒ β1 ∼
= 3 . (28)
sin ϑ β1 sin ϑ

III. S OIL M OISTURE R ETRIEVAL


Once the description of the scattering phenomena and total
power evaluation are carried out, we can illustrate the methods
employed to retrieve the soil volumetric moisture content mv
from polarimetric SAR images. In fact, as mv affects the
dielectric constant values, it is possible to estimate the relative
permittivity ε from SAR data and then obtain the moisture
content using one of the several empirical or semiempirical
mixing models available in literature [17]–[20].

A. Co- and Cross-pol Method


The co- and cross-pol ratios are defined as the ratio of the
co-pol NRCS and the ratio of cross-pol ones

⎪ σVV
0
|a,b

⎨ Co-pol = σ0
HH |a,b
(29)

⎪ σHV
0

⎩ Cross-pol = 0 |a,b .
σVV |a,b
Fig. 4. Co-pol correlation coefficient versus σ for the (solid line) X-Bragg and
PTSM with (dotted–dashed line) ε = 5 and (dashed line) ε = 20 at (a) 35◦ and If the X-Bragg NRCS formulation (16) is used, it is clear that
(b) 55◦ incidence angles. the dependence on small-scale-roughness parameters s and Ht
cancels out in the ratios in (29), so that the latter only depend
pq pq pq on the large-scale-roughness parameter σ and on the relative
The full expressions of C0,0 , C2,0 , and C0,2 are reported in
Appendix B. dielectric constant ε. Conversely, if the PTSM NRCS formula-
In the following, we will refer to our expressions (23)–(25) tion (23) is used, the dependence on the small-roughness height
as the “polarimetric two-scale model” (PTSM) solution, to be standard deviation s still cancels out in the co- and cross-pol
compared with the X-Bragg solution [7], expressed by (16) ratios (29), but this is not the case for the dependence on Ht .
and (17). With regard to its validity limits, by summarizing However, as we will show later, the dependence on Ht of the
the previous discussion, we can state that the model holds if ratios in (29) is weak, so that these ratios mainly depend only
the small-scale-roughness height standard deviation s is much on the large-scale-roughness parameter σ and on the relative
smaller than the wavelength and if the large-scale-roughness dielectric constant ε. Accordingly, if we neglect the imaginary
slope standard deviation σ is small enough. The exact value part of ε, as it is reasonable for usual soils (or also if we assume
of the superior limit for σ is discussed in the next section. In that there is a relationship relating the real and imaginary parts
addition, the model does not hold at small incidence angles, for of ε at the considered frequency, as in [19]), it is possible to
which the GO contribution is dominant. retrieve σ and ε from measured co- and cross-pol ratios.
Finally, we want to note explicitly that both X-Bragg and Although the ratios’ dependence on σ and ε is too involved
PTSM account for depolarization. To show this, in Fig. 4, plots to be analytically inverted, both co- and cross-pol ratios exhibit
of the co-pol correlation coefficient monotonic trends versus σ and ε, so that it is possible to perform
an effective inversion by building up numerical charts where co-
|t12 | and cross-pol loci for different values of σ and ε are plotted. An
γ=√ (27)
t11 t22 example of such a chart is reported in Fig. 5, which shows co-
and cross-pol loci parameterized with respect to the electric per-
versus σ are shown for both models. These plots show that, mittivity, ranging from 2 to 22 with a step equal to 2, and to the
for nonzero σ values, the correlation coefficient is smaller than large-scale-roughness parameter σ, ranging from 0 up to 0.4.
unity and it decreases as σ increases. To obtain the plots in By entering this chart with values of co- and cross-pol ratios
Fig. 4, σ and β1 have been related to each other by noting that, computed from measured SAR (or, more in general, scattering)
in the first-order limit, the PTSM hypotheses on the scattering data, one can directly read the corresponding retrieved values
surface imply that tan β is a zero-mean Gaussian random of ε and σ. Of course, this can be automatically performed by a
variable with standard deviation σ/ sin ϑ [16]. Then, assuming proper software code.

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
IODICE et al.: RETRIEVAL OF SOIL SURFACE PARAMETERS VIA A POLARIMETRIC TWO-SCALE MODEL 2537

Fig. 5. PTSM co-pol–cross-pol for Ht = 0.5 and ϑ = 45◦ .

Note that, in the chart of Fig. 5, rms slope values σ as large change of Ht , spanning the entire range of values actually
as 0.4 are considered. This may seem inappropriate, if one encountered for natural soils, just turns into a slight variation
considers that employed second-order NRCS expansions hold of (co- and cross-pol) loci. The only exception is at small
only for small values of σ. We have numerically verified that incidence angles (Fig. 7), for large values of ε and/or σ.
differences between higher and second-order terms of NRCS However, for such large values of σ, the employed approxi-
remain below 10% only up to values of σ of about 0.2; however, mations are not well verified in any case, and for large values
this range of validity extends up to more than 0.4 if we consider of ε, rather large variations of ε correspond to rather small
co-pol ratios and up to about 0.3 for cross-pol ratios. This means variations of mv [17]–[20]. Therefore, we can conclude that
that, fortunately, higher order effects tend to cancel out in the these charts benefit from an overall robustness with respect to
ratios, at least in the range of σ values from about 0.2 to 0.4. variations of small-roughness parameters, above all for large
It is worth noting that, if we assume that only surface scatter- radar look angles. Of course, the fact that co- and cross-pol
ing occurs, co-pol–cross-pol charts bring the same information ratios are not affected by the small-scale roughness does not
deducible from the H–α charts of [7]. In particular, the cross- imply that the single polarimetric channel return is not affected
pol ratio plays a role similar to that of the entropy H, while the by small-scale roughness also: Dependence of the NRCS on
co-pol ratio plays a role similar to that of the mean scattering small-scale roughness is clearly expressed in (10). In particular,
angle α. by using (23), it can be verified that, for co-pol channels at
In order to emphasize the effect of considering the local intermediate incidence angles, scattering is mainly dependent
incidence angle random variations, a comparison between co- on small-scale roughness (although this dependence cancels
pol–cross-pol charts obtained using both PTSM, i.e., (23), and out in the co-pol ratio, as shown previously), whereas for
X-Bragg, i.e., (16), methods is provided in Fig. 6. This figure the cross-pol channel, the large-scale roughness controls the
shows that differences between the two charts are significant: overall level of the scattered field and the small-scale roughness
In particular, co-pol ratios in the dashed-line chart, pertinent to controls the behavior of scattering as a function of the incidence
the X-Bragg method, decrease quite softly in function of the angle.
large-scale roughness, whereas in the solid-line chart, relevant
to our PTSM formulation, the corresponding co-pol ratios
decrease more rapidly, most likely because they account for B. H–α Method
an additional randomness inherent to the stochastic drift of the Although, in the next section, we will use the co- and cross-
local incidence angle. In the next section (Section IV), we will pol ratio method, it is important to note that our PTSM can
show that our PTSM formulation provides a better match with be used in the H–α retrieval approach as well. We recall that
ground measurements than the X-Bragg one. the entropy H is related to the eigenvalues of the coherency
Finally, let us verify that, as already mentioned, the value matrix and mainly measures the “degree of randomness” of
of the small-scale-roughness Hurst coefficient Ht does not the scattering process, whereas the mean scattering angle α
significantly affect the PTSM co- and cross-pol ratios. To this is related to eigenvalues and eigenvectors of the coherency
aim, in Figs. 7 and 8, the co-pol–cross-pol charts for Ht = 0.4 matrix, depends on the kinds of mechanisms (single and double
(dashed line) and Ht = 0.8 (solid line) are depicted for ϑ = 30◦ bounce and volumetric) involved in the scattering process, and,
and ϑ = 60◦ , respectively. These figures show that a significant in the surface scattering case, mainly depends on the surface

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
2538 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 7, JULY 2011

Fig. 6. Co-pol–cross-pol charts for (a) ϑ = 35◦ and (b) ϑ = 55◦ derived from (dashed line) X-Bragg and (solid line, Ht = 0.5) PTSM methods.

dielectric constant and large-scale roughness. Full definitions Note that, in the charts in Figs. 9 and 10, only rms slope
and interpretations of these parameters can be found in [7], [14], values σ not larger than about 0.2 are considered. We have
[15], and [21]. numerically verified that, at variance with the co-pol/cross-pol
An example of a PTSM-based H–α chart is shown in Fig. 9, case, differences between higher and second-order PTSM terms
whereas a comparison between PTSM and X-Bragg charts is of H and α remain below 10% only up to values of σ of
shown in Fig. 10. about 0.2; therefore, PTSM-based H–α charts cannot be used
For the H–α charts, considerations similar to those made for for higher values of σ. However, this is usually sufficient for
the co-pol and cross-pol ones can be made: The dependence on L-band data and moderate roughness, according to what was
small-roughness parameters exactly cancels out by using the obtained in Section IV-A (see last column of Tables I and II) and
X-Bragg formulation (17), whereas it approximately cancels according to [22], where it is reported that a value of β1 (there
out by using the PTSM formulation (24) and (25). Moreover, called “δ”) of π/6 is usually obtained. According to (28), this
significant differences are present in the H–α charts obtained value corresponds to values of σ of about 0.2 for intermediate
by the two methods. incidence angles.

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
IODICE et al.: RETRIEVAL OF SOIL SURFACE PARAMETERS VIA A POLARIMETRIC TWO-SCALE MODEL 2539

Fig. 7. Charts for ϑ = 30◦ , (dashed line) Ht = 0.4 and (solid line) Ht = 0.8.

Fig. 8. Charts for ϑ = 60◦ , (dashed line) Ht = 0.4 and (solid line) Ht = 0.8.

IV. M ETHOD VALIDATION observed in both dry and wet conditions. Such scattering and
in situ data were reported and used, for instance, in [2] and
In order to validate the retrieval method, a wide variety
[23]. We here use those data to validate our PTSM-based
of scattering data at different frequencies, incidence angles,
co-pol/cross-pol retrieval approach, to explore its validity lim-
surface roughnesses, and soil moisture contents, in conjunc-
its, and to compare it with its X-Bragg counterpart. A similar
tion with the corresponding ground measurements, should be
comparison is also performed on AIRSAR data corresponding
available over bare or little vegetated soils. This is not the case
to the Washita measurement campaign data, which are freely
for fully polarimetric SAR data, so that we do not examine
available on the Web [24].
here the H–α retrieval approach, but it is the case for polari-
metric NRCSs. In fact, HH, VV, and HV NRCS data at L-,
C-, and X-bands acquired by the University of Michigan’s
A. POLARSCAT Data, Moderate Roughness
LCX POLARSCAT [23] are available, together with “in situ”
measurements, for incidence angles varying from 10◦ to 70◦ Here, the favorable case of L-band data and moderate rough-
and for four bare soil surfaces with different roughness, each ness is considered. To this aim, L-band POLARSCAT data

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
2540 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 7, JULY 2011

Fig. 9. PTSM H–α chart for Ht = 0.5 and ϑ = 45◦ .

relevant to the slightly rough bare soil surfaces 1 and 2 of [23] precision is ±0.4 dB [23]). The rms value of differences be-
are employed. tween retrieved and in situ measured ε values is 6.01 for the
Let us first analyze surface 1 in the “wet” and “dry” cases. X-Bragg method and 3.24 for the PTSM method. In addition,
Corresponding in situ measured parameters are reported in the the correlation coefficient r between retrieved and in situ
first rows of Tables I and II, where s is the height standard measured ε values is 0.655 for the X-Bragg method and 0.740
deviation measured over 1-m-long profiles [23] and εref is the for the PTSM method. These quantitative results confirm that
measured relative dielectric constant in the top 4-cm soil layer PTSM ε estimates are in better agreement with ground mea-
[23]. Unfortunately, large-scale roughness was not measured surements than X-Bragg ones.
in [23], so that only the ε retrieval can be verified and no With regard to the σ estimates, by comparing the sixth and
direct validation of the σ retrieval is possible. Tables I and II seventh columns of Tables I and II (however, similar results are
collect the retrieval results obtained via both the X-Bragg and obtained also for surface 2), it is clear that they are smaller for
the PTSM (with Ht = 0.5) by considering co- and cross-pol the PTSM, and this seems more reasonable for the considered
ratios at incidence angles from 20◦ to 60◦ (the smallest and surface (which is rather smooth according to [23]). We note that
highest incidence angles are discarded for the reasons explained there is a slight difference between the σ values obtained for the
in Section II-C). In these tables, as in Tables III–V, a dash wet and dry cases, and this indicates that the model is still not
indicates that the considered co-pol and cross-pol values do not fully appropriate (in fact, roughness should be the same in both
correspond to any physically acceptable value of ε, so that the cases). However, this difference is smaller for the PTSM than
inversion algorithm automatically assumes that the model does for the X-Bragg.
not hold in these cases.
By comparing the retrieved permittivity values reported in
B. POLARSCAT Data, High Roughness
the fourth (for X-Bragg) and fifth (for PTSM) columns of
Tables I and II with the measured values reported in the first In order to explore the potentiality of PTSM beyond the SPM
row, it is clear that the PTSM ε retrieval results are in better validity limits, the unfavorable case of very rough (with respect
agreement with ground measurements than the X-Bragg ones to the wavelength) surfaces is also considered here. To this end,
in all of the considered cases, except that at 30◦ for the “dry” we consider X-band data relevant to surface 1 and both L- and
surface. X-band data relevant to surface 4 of [23], which are very rough
In order to obtain an overall view of the retrieval methods’ at both frequencies. Of course, this is a very challenging case
performances, previously described results, and analogous re- for an SPM-based approach, and in fact, no valid results are
sults obtained for the surface 2 of [23] [together with results obtained for incidence angles from 20◦ to 40◦ (see Tables III–V,
from AIRSAR data (see Section IV-C)], are shown in Fig. 11(a) which have the same format as Tables I and II). However, a
for X-Bragg and Fig. 11(b) for PTSM. In general, a significant reasonable agreement with the measured value of ε is obtained
underestimation of ε is obtained by the X-Bragg method, by the PTSM at 50◦ and 60◦ incidence angles. These results
whereas a reasonable or good agreement with measured ε is can be explained by considering that, on the one hand, for
obtained by the PTSM. Fluctuations of ε estimates are probably such high values of σ, the co-pol ratio is very near to 1 up
due to measurement uncertainty (POLARSCAT measurement to about 40◦ (as foreseen by the GO, which, for such a high

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
IODICE et al.: RETRIEVAL OF SOIL SURFACE PARAMETERS VIA A POLARIMETRIC TWO-SCALE MODEL 2541

Fig. 10. H–α charts for (a) ϑ = 35◦ and (b) ϑ = 55◦ derived from (dashed line) X-Bragg and (solid line, Ht = 0.5) PTSM methods.

TABLE I TABLE II
R ETRIEVAL R ESULTS FOR S URFACE 1-W ET AT L-BAND . A DASH “-” R ETRIEVAL R ESULTS FOR S URFACE 1-D RY AT L-BAND . A DASH “-”
I S S HOWN FOR R ETRIEVED ε VALUES > 20 OR < 2 I S S HOWN FOR R ETRIEVED ε VALUES > 20 OR < 2

roughness, is dominant up to intermediate incidence angles; to these incidence angle values, and that, on the other hand,
see also Section IV-D), so that no information is provided the separation between small- and large-scale roughnesses is
by the co-pol ratio and no useful results can be obtained up frequency dependent, so that, at X-band, the facet size can be

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
2542 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 7, JULY 2011

TABLE III
R ETRIEVAL R ESULTS FOR S URFACE 1-W ET AT X-BAND . A DASH “-”
I S S HOWN FOR R ETRIEVED ε VALUES > 20 OR < 2

TABLE IV
R ETRIEVAL R ESULTS FOR S URFACE 4-W ET AT L-BAND . A DASH “-”
I S S HOWN FOR R ETRIEVED ε VALUES > 20 OR < 2

TABLE V
R ETRIEVAL R ESULTS FOR S URFACE 4-W ET AT X-BAND . A DASH “-”
I S S HOWN FOR R ETRIEVED ε VALUES > 20 OR < 2

Fig. 11. (a) X-Bragg and (b) PTSM L-band retrieval results for (diamonds)
surface 1 and (triangles) surface 2 of [23] and for (squares) the Little Washita
basin [24]. For X-Bragg, rms error is 6.01 and correlation coefficient r is 0.655.
For PTSM, rms error is 3.24 and correlation coefficient r is 0.740.

C. AIRSAR Data
assumed to be a few centimeters and the microscopic roughness
of the facet is still compatible with SPM. Finally, an example of application to SAR data is presented.
We also note that the retrieved facets’ (i.e., large-scale) L-band AIRSAR data are considered, which were acquired on
rms slope σ for the surface 1-wet moves from about 0.15 at several different days during a measurement campaign in the
L-band to about 0.33 at X-band (see Tables I and III). This is Little Washita basin in June 1992 [24]. We select the bare soil
consistent with the following reasonable assumptions: 1) As field labeled as AG002 in [24] (see Fig. 12), for which the
already noted, the separation between small- and large-scale volumetric soil moisture content was monitored in situ, and
roughnesses is frequency dependent, so that in our model the we evaluate the average NRCS for the different polarimetric
facet size L is directly proportional to the wavelength λ, and channels over three 30 × 30 pixel windows within the consid-
2) the surface roughness can be modeled by an fBm process ered field. Co-pol–cross-pol charts for the X-Bragg and PTSM
with Ht ∼ = 0.55, according to which chords joining points on are reported in Fig. 13(a) and (b), respectively. The retrieved
the surface at fixed distance L have an rms slope σ proportional values of the relative dielectric constant are converted into
to (1/L)1−Ht ∼ = (1/L)0.45 [12], [13]. In fact, from assumptions volumetric moisture values mv by using the mixing model of
1) and 2), it follows that the ratio between the retrieved values of [19], with percentages of sand and clay equal to 45.5 and 13.4,
σ at frequencies f1 and f2 is (f1 /f2 )0.45 , i.e., for f1 = 9.5 GHz respectively [24]. The retrieved results for X-Bragg and PTSM
and f2 = 1.5 GHz, about 2.2, which is in agreement with the are reported in Table VI, together with in situ measured values.
measured ratio 0.33/0.15. These results are also presented in graphic form in Fig. 11,

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
IODICE et al.: RETRIEVAL OF SOIL SURFACE PARAMETERS VIA A POLARIMETRIC TWO-SCALE MODEL 2543

experimental data. Note, however, that a reasonable agreement


with measured data is obtained for the co- and cross-pol ratios
(i.e., the vertical distances between different polarizations in
the plots of NRCS expressed in decibels) for incidence angles
greater than 40◦ .

V. C ONCLUSION
A polarimetric two-scale surface scattering model (PTSM)
has been introduced and employed to retrieve the surface
parameters of bare soils from polarimetric SAR data. Unlike
other similar existing approaches, here, large-scale-roughness-
induced random variations of both local incidence angle and
plane have been considered, and their statistical modeling has
not been arbitrarily chosen, but it has been derived from a
proper statistical description of the scattering surface. The
resulting scattering model can be used to retrieve soil moisture
Fig. 12. L-band HH-polarization AIRSAR image acquired on June 10, 1992 and large-scale roughness by using co- and cross-pol ratios
over the Little Washita basin. The yellow rectangle indicates the considered or the H-alpha decomposition. The former method has been
field. used on scatterometer and SAR data relevant to different bare
soils for which the ground measurements are available. Good
retrieval results have been obtained at L-band for moderately
together with POLARSCAT data. It is clear that, in this case,
rough surfaces at incidence angles ranging from 30◦ to 60◦ ,
better results are again obtained via the PTSM rather than via
and reasonable results have been achieved even at X-band and
the X-Bragg.
for rough surfaces, but only at incidence angles of 50◦ –60◦ .
Conversely, as expected, no valid results have been obtained
D. Direct Model Validation in the other considered cases, i.e., 20◦ at L-band for mod-
erately rough surfaces, 20◦ –40◦ at X-band for all surfaces,
Retrieval results presented in the previous sections can be and 20◦ –40◦ at L-band for very rough surfaces. However, in
also considered as an indirect validation of the PTSM scattering practically all of the considered cases, results obtained by using
model. However, they only show that the PTSM model provides the PTSM have turned out to be in better agreement with mea-
a good evaluation of co- and cross-pol ratios: In order to provide surements than those obtained by using already existing similar
a more direct validation of the PTSM scattering model that also models.
allows us to verify if this model is able to properly describe In conclusion, results are very promising, and future work
the behavior of the NRCS as a function of the incidence is needed, first of all, to validate the use of the PTSM also in
angle, PTSM backscattering results, obtained from (23), are conjunction with the H–α method. To this aim, measurement
directly compared here with experimental backscattering data. campaigns are planned at times of acquisition of available SAR
In particular, L-band measured backscattering data relevant to sensors [25]. In addition, in order to extend soil moisture re-
surfaces 1 and 2 of [23] are considered in Fig. 14(a) and (b). trieval also to vegetated areas, an effort can be made to include
The employed surface parameters are the in situ measured the PTSM in one of the available physical-based polarimetric
relative permittivity and the σ values obtained in Section IV-A; scattering decomposition methods [22], [26], [27].
with regard to the small-roughness parameters, they are the Ht
value obtained at the end of Section IV-B and a value of S0
obtained by fitting measured data (in fact, a variation of S0 A PPENDIX A
only corresponds to a rigid vertical shift on the plot of NRCS
In this Appendix, we explicitly derive averages over a and b
expressed in decibels). All of these employed surface param-
for NRCS and coherency matrix entries by using (21) and (22).
eters are summarized in Table VII. Plots of Fig. 14(a) and
Let us note that every β rotation term involved in (11) and
(b) show a very good agreement between PTSM results and
(15) can be written by means of these two quantities
measured backscattering data for all of the three polarization
channels. ( 2 tan2 β a2
sin β = 1+tan 2 β = a2 +(sin ϑ−b cos ϑ)2
Finally, the very challenging case of X-band measured tan β 2a(sin ϑ−b cos ϑ) (30)
backscattering data relevant to surface 1 is considered in sin 2β = 2 1+tan 2 β = a2 +(sin ϑ−b cos ϑ)2 .

Fig. 14(c). Corresponding surface parameters are also reported


in Table VII. In this case, a surprisingly good agreement of the Hence, by expanding both of them up to the second or-
angular behavior of PTSM results with the one of measured der, i.e.,
data is obtained at HV polarization (apart from a small vertical  2
sin β ∼
2

shift), whereas this does not apply to HH and VV polarizations, = sina2 ϑ


(31)
for which there is not a good agreement between PTSM and sin 2β ∼ 2a
= sin b
ϑ 1 + tan ϑ

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
2544 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 7, JULY 2011

Fig. 13. Co-pol–cross-pol charts for the (a) X-Bragg and (b) PTSM with Ht = 0.5 for the AIRSAR–Washita data. Acquisitions on (diamond) June 10, 1992,
(hexagon) June 13, 1992, (triangle) June 14, 1992, (square) June 16, 1992, and (circle) June 18, 1992.

TABLE VI Then, using series (32) and (34) in (11) and neglecting
R ETRIEVAL R ESULTS FOR AIRSAR DATA
terms of order higher than the second, we get the following

MacLaurin expansions for σHH :

π 0
σ = (k cos ϑl )4 W |FH cos2 β + FV sin2 β|2
4 HH
)
= (k cos ϑl )4 W |FH |2 cos4 β + 2Re {FH FV∗ }
*
× cos2 β sin2 β + |FV |2 sin4 β
  
∼ a2
= (k cos ϑl )4 W |FH |2 1 − 2 2
sin ϑ

a2
we can easily expand any other function of β appearing in (11) + 2Re {FH FV∗ }
sin2 ϑ
and (15), thus obtaining
HH

= C0,0 + C0,1 HH HH 2
b + C2,0 a + C0,2HH 2
b
⎧ 4

⎪ sin β = o(β 2 ) ∼ =0  

⎪ a2
⎪ cos4 β = (1 − sin2 β)2 ∼
2

⎪ = 1 − 2 sina2 ϑ × 1−2 2


1 2 sin ϑ
⎪ 2 ∼ a2
⎨ (sin β cos β) = 2 sin 2β = sin2 ϑ
 HV  a2
cos 2β = 1 − 2 sin2 β ∼
2
= 1 − 2 sina2 ϑ + 2Re C0,0 HV
+ C1,0 HV 2
b + C2,0 HV 2
a + C0,2 b

⎪ 2 ∼ sin2 ϑ
⎪ a2
⎪ cos 2β = (1 − 2 sin β) = 1 − 4 sin2 ϑ
2 2



⎪ ∼ 2a ∼ HH HH

⎪ sin 2β cos 2β = sin 2β(1 − 2 sin β) = sin ϑ 1 + tan ϑ
2 b = C0,0 + C0,1 b

⎩ 2 +  HV  ,
sin 2β ∼
2
= 4 sina2 ϑ . HH
C0,0 Re C0,0
(32) + HH
C2,0 −2 +2 a2 +C0,2
HH 2
b .
sin2 ϑ sin2 ϑ
Concerning the Taylor expansions of functions depending on
ϑl , NRCSs contain terms proportional to (k cos ϑl )4 W Fp Fq∗ ,
so by setting Following similar steps for the other two channels and using
  n
4 4

∗ (22) to compute the averages over a and b, the NRCS expres-
pq 1 n ∂ W k cos ϑ l F p F q sions reported in (23) are obtained.
Ck,n−k = (33)
n! k k
∂a ∂b n−k A similar procedure can be used for the coherency
a=b=0
matrix: Also its elements contain terms proportional to
(k cos ϑl )4 W Fp Fq∗ , so we can evaluate each entry of (15) using
pq pq
and noting that C1,0 = C1,1 = 0∀p, q ∈ {H, V }, we obtain the
general second-order expansion coefficients in (33) and expansions in (32) and (34). Then, after
ensemble averaging, computed via (22), we obtain the matrix
W k 4 cos4 ϑl Fp Fq∗ ∼ pq pq pq pq
= C0,0 + C0,1 b + C2,0 a2 + C0,2 b2 . (34) reported in (24) and (25).

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
IODICE et al.: RETRIEVAL OF SOIL SURFACE PARAMETERS VIA A POLARIMETRIC TWO-SCALE MODEL 2545

TABLE VII
S URFACE PARAMETERS E MPLOYED IN FIG. 14

thus obtaining


pq 1 ∂ 2 W (2k sin ϑl )(k cos ϑl )4 Fp (ϑl )Fq∗ (ϑl )
C0,2 =
2 ∂b2
a=b=0


2 4
1 ∂ W (2k sin ϑl )(k cos ϑl )
= Fp (ϑl )Fq∗ (ϑl )
2 ∂b2
a=b=0

1 2
∂ Fp (ϑl ) ∗
+ W (2k sin ϑl )(k cos ϑl )4 Fq (ϑl )
2 ∂b2 a=b=0

1 2 ∗
∂ Fq (ϑl )
+ W (2k sin ϑl )(k cos ϑl )4 Fp (ϑl )
2 ∂b2
a=b=0


4
∂ W (2k sin ϑl )(k cos ϑl ) ∂Fp (ϑl ) ∗
+ Fq (ϑl )
∂b ∂b
a=b=0


∂ W (2k sin ϑl )(k cos ϑl ) 4 ∗
∂Fq (ϑl )
+ Fp (ϑl )
∂b ∂b
a=b=0

∂F
∂Fp (ϑl ) q l (ϑ )
+ W (2k sin ϑl )(k cos ϑl )4
∂b ∂b a=b=0
(36)


pq 1 2
∂ W (2k sin ϑl )(k cos ϑl ) 4
Fp (ϑl )Fq∗ (ϑl )

C2,0 =
2 ∂a2
a=b=0


2 4
1 ∂ W (2k sin ϑl )(k cos ϑl )
= Fp (ϑl )Fq∗ (ϑl )
2 ∂a2
a=b=0

1 ∂ 2
F (ϑ )
Fq∗ (ϑl )
p l
+ W (2k sin ϑl )(k cos ϑl )4
Fig. 14. (Dots) Measured and (lines) PTSM backscattering coefficients at 2 ∂a2 a=b=0
(solid line) VV, (dashed line) HH, and (dotted–dashed line) HV polarizations
for (a) surface 1 at L-band, (b) surface 2 at L-band, and (c) surface 1 at X-band. 1 2 ∗
∂ Fq (ϑl )
Employed surface parameters are listed in Table VII. + W (2k sin ϑl )(k cos ϑl )4 Fp (ϑl )
2 ∂a2
a=b=0
(37)
A PPENDIX B
In this Appendix, we derive the complete expressions for where we have exploited the fact that, for p = {H, V }
pq
Ck,n−k used in previous expansions. First of all, notice that,
from (3), we have cos ϑl = cos ϑ in a = b = 0, so zeroth-order ⎧
⎪ ∂ 2 (W (2k sin ϑl )(k cos ϑl )4 ) ∂ 2 Fp (ϑl )
coefficients are ⎪
⎨ ∂a∂b = ∂a∂b a=b=0 =0
a=b=0

⎪ ∂ (W (2k sin ϑl )(k cos ϑl )4 ) ∂Fp (ϑl )
pq
C0,0 = W (2k sin ϑl )(k cos ϑl )4 Fp (ϑl )Fq∗ (ϑl )|a=b=0 ⎪
⎩ = ∂a =0 .
∂a a=b=0
a=b=0
= W (2k sin ϑ)(k cos ϑ)4 Fp (ϑ)Fq∗ (ϑ). (35) (38)
The expressions for the derivatives involved in previous coeffi-
pq pq
In order to obtain C2,0 and C0,2 , recalling the definition cients are explicitly reported in the equations, shown at the top
(33), we have to differentiate twice with respect to a and b, of the next page.

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
2546 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 49, NO. 7, JULY 2011



∂ W (2k sin ϑl )(k cos ϑl )4 S0 (2k)2−2Ht 3 − cos 2ϑ + Ht (1 + cos 2ϑ)
=
∂b 16 (sin ϑ)2Ht tan3 ϑ
a=b=0


∂ 2 W (2k sin ϑl )(k cos ϑl )4 S0 (2k)2−2Ht cos2 ϑ − 2 − Ht cos2 ϑ
=
∂a2 8 (sin ϑ)2Ht tan4 ϑ
a=b=0

  
∂ 2 W (2k sin ϑl )(k cos ϑl )4 S0 (2k)2−2Ht 2 3 6 5
= cos 2ϑ − + −H t 1+2 cos 2ϑ+ −
∂b2 8 (sin ϑ)2Ht sin2 ϑ sin4 ϑ sin2 ϑ sin4 ϑ
a=b=0
 
6 2
+ Ht 5 + cos 2ϑ −
2
+
sin2 ϑ sin4 ϑ
-
∂FH (ϑl ) 2 sin ϑ cos ϑ − ε − sin2 ϑ
=− -
∂b a=b=0 ε − 1 + cos ϑ cos ϑ + ε − sin2 ϑ
-
∂ 2 FH (ϑl ) 2 cos ϑ cos ϑ − ε − sin2 ϑ
= -
∂a2 a=b=0 ε − 1 + cos ϑ cos ϑ + ε − sin2 ϑ
√ √
∂ 2 FH (ϑl ) 3 + cos 4ϑ − 4 cos 2ϑ + 2ε (3 cos 2ϑ − 1) − 2 2 cos ϑ (ε − 1 + cos 2ϑ) 2ε − 1 + cos 2ϑ
= 
- !
∂b2 a=b=0 3
2 ε − sin2 ϑ cos ϑ + ε − sin2 ϑ

√ √
∂FV (ϑl ) ε (ε − 1) sin ϑ 2ε cos ϑ + 2 (2ε − 1) 2ε − 1 + cos 2ϑ
= - !3 · -
∂b a=b=0 ε − sin2 ϑ
ε cos ϑ + ε − sin2 ϑ

√ √
∂ 2 FV (ϑl ) ε (ε − 1) cos ϑ 2ε cos ϑ + 2 (2ε − 1) 2ε − 1 + cos 2ϑ
=− - !3 · -
∂a2 ε − sin2 ϑ
a=b=0 ε cos ϑ + ε − sin2 ϑ

∂ 2 FV (ϑl ) ε(ε − 1)
2 = -
∂b a=b=0
3
4(ε − sin ϑ) 2 (ε cos ϑ + ε − sin2 ϑ)4
2
% √ ) *
× 2 4ε − 2 + 2 cos 2ϑ −4ε(1 − cos 2ϑ) + ε2 (9 − 10 cos 2ϑ + cos 4ϑ) − 4ε3 (2 − cos 2ϑ)

+ [−2 cos ϑ + 3 cos 3ϑ − cos 5ϑ + 2ε(3 cos ϑ − 4 cos 3ϑ + cos 5ϑ)


*
+ ε2 (6 cos ϑ + cos 3ϑ + cos 5ϑ) − 2ε3 (13 cos ϑ − cos 3ϑ)

R EFERENCES [11] G. Franceschetti, A. Iodice, M. Migliaccio, and D. Riccio, “Scattering


from natural rough surfaces modelled by fractional Brownian motion two-
[1] F. T. Ulaby, R. K. Moore, and A. K. Fung, Microwave Remote Sensing: dimensional processes,” IEEE Trans. Antennas Propag., vol. 47, no. 9,
Active and Passive. Reading, MA: Addison-Wesley, 1982. pp. 1405–1415, Sep. 1999.
[2] A. K. Fung, Z. Li, and K. S. Chen, “Backscattering from a randomly rough [12] G. Franceschetti, A. Iodice, and D. Riccio, “Fractal models for scatter-
dielectric surface,” IEEE Trans. Geosci. Remote Sens., vol. 30, no. 2, ing from natural surfaces,” in Scattering, R. Pike and P. Sabatier, Eds.
pp. 356–369, Mar. 1992. London, U.K.: Academic, 2001, pp. 467–485.
[3] D. E. Barrick and W. Peake, “A review of scattering from rough surfaces [13] K. Falconer, Fractal Geometry: Mathematical Foundations and
with different roughness scales,” Radio Sci., vol. 3, pp. 865–868, 1968. Applications. New York: Wiley, 1990.
[4] G. R. Valenzuela, “Theories for the interaction of electromagnetic waves [14] S. R. Claude, “Uniqueness of target decomposition theorems in radar
and oceanic waves. A review,” Bound. Layer Meteorol., vol. 13, no. 1–4, polarimetry,” in Direct and Inverse Methods in Radar Polarimetty, Part 1,
pp. 61–85, Jan. 1978. NATO-ARW. W. M. Boemer, et al., Eds. Norwell, MA: Kluwer, 1992,
[5] J. W. Wright, “A new model for sea clutter,” IEEE Trans. Antennas pp. 267–296.
Propag., vol. AP-16, no. 2, pp. 217–223, Mar. 1968. [15] I. Hajnsek, “Inversion of surface parameters from polarimetric SAR,”
[6] A. G. Voronovich, “A two-scale model from the point of view of the Ph.D. dissertation, DLR, Oberpfaffenhofen, Germany, 2001.
Small-Slope Approximation,” Waves Random Media, vol. 6, pp. 73–83, [16] F. del Monaco, N. De Quattro, A. Iodice, and A. Natale, “Extended small
1996. perturbation method and retrieval of natural surface parameters,” in Proc.
[7] I. Hajnsek, E. Pottier, and S. R. Cloude, “Inversion of surface parameters 6th Eur. Radar Conf., Rome, Italy, 2009, pp. 537–540.
from polarimetric SAR,” IEEE Trans. Geosci. Remote Sens., vol. 41, [17] J. R. Wang and T. J. Schmugge, “An empirical model for the com-
no. 4, pp. 727–744, Apr. 2003. plex dielectric permittivity of soils as a function of water content,”
[8] Y. Kim, J. van Zyl, and J. Shi, “A soil moisture algorithm using tilted- IEEE Trans. Geosci. Remote Sens., vol. GRS-18, no. 4, pp. 288–295,
Bragg approximation,” in Proc. IGARSS, 2003, pp. 1402–1404. Oct. 1980.
[9] J. S. Lee, D. L. Schuler, and T. L. Ainsworth, “Polarimetric SAR data [18] G. C. Topp, “Electromagnetic determination of soil water content: Mea-
compensation for terrain azimuth slope variation,” IEEE Trans. Geosci. surements in coaxial transmission lines,” Water Resour. Res., vol. 16,
Remote Sens., vol. 38, no. 5, pp. 2153–2163, Sep. 2000. no. 3, pp. 574–582, Jun. 1980.
[10] L. Tsang, J. Kong, and K. Ding, Scattering of Electromagnetic [19] H. T. Hallikainen, F. T. Ulaby, M. C. Dobson, M. A. El-Rayes, and
Waves—Theory and Applications. New York: Wiley, 2000. L. K. Wu, “Microwave dielectric behavior of wet soils—Part I:

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.
IODICE et al.: RETRIEVAL OF SOIL SURFACE PARAMETERS VIA A POLARIMETRIC TWO-SCALE MODEL 2547

Empirical models and experimental observations,” IEEE Trans. Geosci. Antonio Natale (S’09) was born in Naples, Italy,
Remote Sens., vol. GRS-23, no. 1, pp. 25–34, Jan. 1985. on July 3, 1982. He received the B.S. (cum laude)
[20] J. D. Miller and G. J. Gaskin, “Theta Probe ML2x principles of operation and M.S. degrees (cum laude) in telecommuni-
and application,” in MLURI Technical Note, 2nd ed. Cambridge, U.K.: cation engineering from the University of Naples
Delta-T Devices, 1999. “Federico II,” Naples, in 2005 and 2008, respec-
[21] J. Praks, E. C. Koeniguer, and M. T. Hallikainen, “Alternatives to tar- tively, where he is currently working toward the
get entropy and alpha angle in SAR polarimetry,” IEEE Trans. Geosci. Ph.D. degree in applied electromagnetism in the
Remote Sens., vol. 47, no. 7, pp. 2262–2274, Oct. 2009. Department of Biomedical, Electronic and Telecom-
[22] I. Hajnsek, T. Jagdhuber, H. Schön, and K. P. Papathanassiou, “Potential munication Engineering.
of estimating soil moisture under vegetation cover by means of PolSAR,” His research interests span the retrieval of natu-
IEEE Trans. Geosci. Remote Sens., vol. 47, no. 2, pp. 442–454, Feb. 2009. ral scene parameters from synthetic aperture radar
[23] Y. Oh, K. Sarabandi, and F. T. Ulaby, “An empirical model and an inver- (SAR) data, the theory of scattering from natural surfaces, as well as the mod-
sion technique for radar scattering from bare soil surfaces,” IEEE Trans. eling of bistatic SAR configurations and the diffraction theory, with application
Geosci. Remote Sens., vol. 30, no. 2, pp. 370–381, Mar. 1992. to electromagnetic propagation in urban areas.
[24] [Online]. Available: http://hydrolab.arsusda.gov/washita92/wash92.htm Dr. Natale was the recipient of the IEEE GRS South Italy Chapter Award
[25] G. Calamita, R. Carlà, A. Iodice, V. Lapenna, A. Natale, A. Mondini, for the best 2008 Italian thesis in remote sensing and the 2009 Sergei A.
A. Perrone, D. Riccio, L. Santurri, L. Candela, G. Ober, and F. Guzzetti, Schelkunoff Prize Paper Award from the IEEE Antennas and Propagation
“Stima dell’umidità del suolo da dati EO a supporto del monitoraggio del Society in 2009.
rischio da frana,” in Proc. Atti 13a Conferenza Nazionale ASITA, Bari,
Italy, 2009.
[26] A. Freeman and S. L. Durden, “A three-component scattering model for
polarimetric SAR data,” IEEE Trans. Geosci. Remote Sens., vol. 36, no. 3,
pp. 963–973, May 1998.
[27] Y. Yamaguchi, T. Moriyama, M. Ishido, and H. Yamada, “Four compo-
nent scattering model for polarimetric SAR image decomposition,” IEEE
Trans. Geosci. Remote Sens., vol. 43, no. 8, pp. 1699–1706, Aug. 2005. Daniele Riccio (M’91–SM’99) was born in Naples,
Italy. He received the Laurea degree (cum laude) in
electronic engineering from the University of Naples
Antonio Iodice (S’97–M’00–SM’04) was born in “Federico II,” Naples, in 1989.
Naples, Italy, on July 4, 1968. He received the He is a Professor of electromagnetics and remote
Laurea degree (cum laude) in electronic engineer- sensing with the Department of Biomedical, Elec-
ing and the Ph.D. degree in electronic engineer- tronic and Telecommunication Engineering, Univer-
ing and computer science from the University of sity of Naples “Federico II.” He is a member of the
Naples “Federico II,” Naples, in 1993 and 1999, Cassini Radar Science Team. He was a Research Sci-
respectively. entist with the Institute for Research on Electromag-
In 1995, he received a grant from the Italian netics and Electronic Components, Italian National
National Council of Research (CNR) to be spent at Council of Research (CNR). He also was a Guest Scientist with the German
the Istituto di Ricerca per l’Elettromagnetismo e i Aerospace Center (DLR), Munich, Germany, (in 1994 and 1995) and a Visiting
Componenti Elettronici, Naples, for research in the Professor with the Universitat Politecnica de Catalunya, Barcelona, Spain, in
field of remote sensing. He was with Telespazio S.p.A., Rome, Italy, from 1999 2006. His research interests are mainly focused on microwave remote sensing,
to 2000. Since 2000, he has been with the Department of Biomedical, Elec- synthetic aperture radar with emphasis on sensor design, data simulation,
tronic and Telecommunication Engineering, University of Naples “Federico and information retrieval for land, oceanic, and urban scenes, as well as on
II,” where he is currently a Professor of electromagnetics. His main research the application of fractal geometry to remote sensing and electromagnetic
interests are in the field of microwave remote sensing and electromagnetics: scattering from natural surfaces. His research works are published in three
modeling of electromagnetic scattering from natural surfaces and urban areas, books (including the text Scattering, Natural Surfaces, and Fractals) and more
simulation and processing of synthetic aperture radar (SAR) signals, SAR than 250 papers. He is an Associate Editor for the journals Remote Sensing,
interferometry, and electromagnetic propagation in urban areas. He is the Sensors, The Open Remote Sensing Journal, and Applied Remote Sensing
author or coauthor of about 170 papers published on refereed journals or on Journal.
proceedings of international and national conferences. Prof. Riccio was the recipient of the 2009 Sergei A. Schelkunoff Transaction
Prof. Iodice received the “2009 Sergei A. Schelkunoff Prize Paper Award” Prize Paper Award for the best paper published in year 2008 on the IEEE
from the IEEE Antennas and Propagation Society. T RANSACTIONS ON A NTENNAS AND P ROPAGATION.

Authorized licensed use limited to: ULAKBIM UASL - KOCAELI UNIVERSITESI. Downloaded on April 12,2023 at 04:31:19 UTC from IEEE Xplore. Restrictions apply.

You might also like