You are on page 1of 10

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 38, NO.

2, MARCH 2000 641

A Fractal-Based Theoretical Framework for Retrieval


of Surface Parameters from Electromagnetic
Backscattering Data
Giorgio Franceschetti, Fellow, IEEE, Antonio Iodice, Member, IEEE, Stefania Maddaluno, and
Daniele Riccio, Senior Member, IEEE

Abstract—A new theoretical framework for retrieval of di- difficulty of devising a surface description suitable to natural
electric and geometric surface parameters from electromagnetic surfaces and at the same time sufficiently simple to lead to a
backscattering data is presented. Use is made of fractal geometry reasonable inversion technique.
for description of the surface. The suggested scheme leads to a
two-step procedure that allows us to retrieve the surface complex Recently, backscattering models based on the fractal geom-
permittivity and the fractal parameters. Limits of validity for etry for the surface description have been presented [16]–[27].
the procedure applicability are presented. Robustness of the These models allow us to evaluate the backscattering coefficient
method is checked against input data measurement errors and [or the normalized radar cross section (NRCS)] in terms of rel-
electromagnetic model inaccuracies. Key points for application of evant fractal parameters. These parameters have been proven to
the new framework to synthetic aperture radar (SAR) data are
discussed. shed light on natural surface characteristics [28]–[32]. More-
over, only two fractal parameters are sufficient to provide a reli-
Index Terms—Fractals, parameter retrieval, remote sensing, able description of a natural surface [28]–[36]. Since the fewer
scattering.
the number of parameters, the better the retrieval procedure, we
believe it is worthwhile to explore if a more efficient retrieval
I. INTRODUCTION procedure can be devised by using such new fractal models.
Hence, in this paper, we first present an overall view of
R ETRIEVAL of land surface parameters from backscat-
tered data can be performed by using theoretical [1]–[5]
as well as empirical (or semiempirical) models [6]–[9]. This
retrieval procedures. Then, we show that the SPM and the
backscattering model, based on the fractal geometry for the sur-
paper is concerned with the former approach. face description, leads to a new, efficient, and fully theoretical
Sound-direct electromagnetic theoretical backscattering retrieval framework. Limits of validity are presented, as well as
models are now available for evaluating the field backscattered relevant considerations on its applicability to synthetic aperture
from random rough surfaces [10]–[15]. Accordingly, it is radar (SAR) data. Guidelines for assessing the inversion scheme
worthwhile to explore if these direct models can be inverted robustness with respect to input data measurement errors and
in order to retrieve the surface parameters from backscattered electromagnetic model inaccuracies are presented. Reasons
data. The starting point is to make reference to available that do not allow introduction of a full theoretical scheme based
theoretical backscattering evaluation techniques [10]–[15]. on PO or the more general IEM are also discussed.
Three of these are: the integral equation method (IEM), which We restrict our analysis to retrieval of surface parameters and
is applicable over a wide range of surface roughnesses; the do not consider any volumetric interaction between the sensed
physical optics method (PO), which can be applied to large scene and the radar signal. This is due to the very cumbersome
roughness scale; and the small perturbation method (SPM), parameter dependence that volumetric backscattering models
which can be applied to small roughness scale. exhibit. This involved dependence practically does not allow the
Previous work in the area of surface parameter retrieval has design of any fully theoretical retrieval framework.
shown that inversion of theoretical models leads to promising, The paper is organized as follows. Section II is devoted
but never completely satisfactory, results. As a matter of fact, to assessing the relationship between surface parameters and
the effectiveness of such methods is primarily reduced by the backscattering coefficients. Section III examines the number
of parameters to be retrieved. The influence of radar system
parameters is also analyzed. It turns out that the crucial point
Manuscript received May 4, 1999; revised September 23, 1999. relies on the description of the surface correlation properties
G. Franceschetti is with the Dipartimento di Ingegneria Elettronica e delle
Telecomunicazioni, Universitá di Napoli “Federico II,” Naples, Italy, and is [10]–[15]. Moreover, estimation of the surface correlation
also with the University of California at Los Angeles, Los Angeles 90032 USA poses several unresolved problems. Hence, in Section IV, we
(e-mail: francesc@irece1.irece.na.cnr.it). move to a new description of the surface based on the fractal
A. Iodice is with the Dipartimento di Ingegneria Elettronica e delle Teleco-
municazioni, Universitá di Napoli “Federico II,” Naples, Italy, and is also with geometry. In this case, modeling of a random rough surface
Telespazio S.p.A., Rome, Italy (e-mail: iodicea@irece1.irece.na.cnr.it). can be performed in a very appropriate way. In addition, some
S. Maddaluno and D. Riccio are with the Dipartimento di Ingegneria Elet- direct backscattering models are also available to evaluate the
tronica e delle Telecomunicazioni, Universitá di Napoli “Federico II,” Naples,
Italy (e-mail: riccio@diesun.die.unina.it). backscattering coefficients of a fractal surface. This allows us
Publisher Item Identifier S 0196-2892(00)02082-9. to outline a new retrieval scheme for the description of the
0196–2892/00$10.00 © 2000 IEEE
642 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 38, NO. 2, MARCH 2000

surface from the backscattered field. In Section V, it is shown Evaluation of the backscattering coefficients according to the
that, considering PO and SPM, the retrieval framework can SPM leads to
be split into a two-step procedure. The first step allows us to
evaluate the real and imaginary part of the dielectric constant. (2.3)
The second step, presented in Section VI, allows us to evaluate
the geometric parameters of the surface expressed by means where is given by (2.2) with , and
of the fractal geometry. This last step is limited to the SPM.
A conclusion and future recommendations are provided in
Section VII.
(2.4)
II. THEORETICAL BACKSCATTERING MODELS: THE CLASSICAL
APPROACH
In order to devise a fully theoretical framework for the re-
trieval of surface parameters from backscattered data, it is con- Finally, evaluation of the backscattering coefficients ac-
venient, as a preliminary, to state key results in the available cording to the PO solution leads to
backscattering models that allow us to compute the scattered
field whenever an appropriate set of surface parameters has been
provided.
Theoretical backscattering models, usually employed for land
surface applications [10]–[15], are IEM, PO, and SPM. Limits (2.5)
of validity exist for each model [10]–[15], [37], [38]; these limits
obviously restrict the class of surfaces to which each model ap- where is a complex coefficient that depends on the dielectric
plies. permittivity and on the incidence angle. Its complete expression
Each theoretical model can lead to an analytical evaluation can be found in [10], [11], [14].
of the backscattering coefficients only when the surface is de- It is not difficult to show [14] that PO and SPM can be ob-
scribed in terms of geometric and dielectric properties. The ge- tained from the IEM under some restrictive hypotheses.
ometrical description is normally obtained by considering the
surface as an isotropic, stationary, random process character- III. SURFACE PARAMETERS RETRIEVAL FRAMEWORKS
ized to the first order by means of a Gaussian probability den-
sity function (pdf) with zero-mean and -standard deviation. It In a theoretical retrieval scheme, the number of surface pa-
is characterized to the second order by means of an appropriate rameters to be retrieved is obviously limited to the number of
autocorrelation function or corresponding power spectrum . the same parameters appearing in the corresponding direct so-
The dielectric description is normally performed by considering lution. Note that also the radar parameters are present in any
the sensor (in the free space) that illuminates a semi-infinite ho- direct solution.
mogeneous space characterized by a complex permittivity (nor- Radar parameters in (2.1)–(2.5) are the electromagnetic
malized to the free space case) and limited by the wavenumber , the incidence angle over the mean reference
rough surface described above. plane and the transmitting and receiving polarization with
For these surfaces, according to the IEM model and ne- respect to the radar antenna reference plane.
glecting multiple scattering, the backscattering coefficients Note that models presented in Section II rely on the assump-
can be written as [14] tion of a zero mean surface profile. Accordingly, in (2.1)–(2.5),
no depolarization is present. Extension to include the effect of a
mean plane slope can be found in [39] and does not affect main
results presented in this paper.
(2.1) The electromagnetic wavenumber and the transmitting and
where receiving polarizations are known and do not need any retrieval
polarization of the transmitted field; procedure. Conversely, this is true for the incidence angle only
polarization of the received field; if the surface mean plane is known. This is what we assume in
incident wavenumber; the following, and we do not include in the retrieval proce-
incidence angle over the surface mean plane. dure as a parameter to be estimated. This assumption is reason-
In addition, in (2.1) able in flat areas and can always be accepted if a macroscopic
surface profile (DEM), sampled at the radar resolution scale, is
(2.2) provided. We do not assume any other a priori in situ knowl-
edge. Furthermore, in the SAR case, this (macroscopic) DEM
where is the zero-order Bessel function and is the surface can be obtained from backscattered data if an interferometric
normalized autocorrelation function. Finally, are complex SAR configuration is available [40].
coefficients that depend on the complex dielectric permittivity Under the above hypotheses, only the surface (microscopic)
and on the incidence angle. Their complete expression can be parameters are to be estimated. In any theoretical model re-
found in [14]. ported in Section II, the number of surface parameters to be
FRANCESCHETTI et al.: FRACTAL-BASED THEORETICAL FRAMEWORK FOR RETRIEVAL OF SURFACE PARAMETERS 643

retrieved changes according to the employed second-order geo- and . Incidentally, we note that different choices can be
metric characterization of the surface. In fact, it turns out from made for the set of parameters representing the fBm surface
(2.1)–(2.5) that the retrieval procedure must deal with two di- [28]–[32]. For instance, the fractal dimension and the
electric parameters, and , while the number of geometrical standard deviation of the height increments at unitary dis-
parameters is greater than one, the first being and the others placement can be used without any restriction on the following
arising from the selected autocorrelation function for the sur- analysis. In fact, and are related to and
face.
For a single-parameter autocorrelation function (for instance,
Gaussian or exponential form), the only additional parameter (4.3)
to be retrieved is the surface correlation length . However,
single-parameter autocorrelation forms have been proved to be
unsatisfactory in the modeling of natural surfaces [14], [25], Once the surface has been described by means of an fBm
[26]. In fact, estimation of the correlation length over natural random process, it is not straightforward to obtain the corre-
surface profiles has been proved to be unstable, in the sense that sponding evaluation of the scattered field, because the fulfilling
the estimated correlation length changes according to the se- of model hypotheses [20]–[23], mathematical developments
lected length profile [14], [41]. In addition, it is difficult to assess [20]–[23], limits of validity [21]–[23], and related issues
a sound relationship between the backscattering coefficient and needs to be addressed. In particular, we stress that a surface
the correlation length. In fact, many discrepancies arise in com- satisfying (4.1) for any small value of the distance would not
paring theoretical models with experimental results in the field be differentiable in any point, and its root mean square (RMS)
[41], [42]. In addition, some empirical backscattering models slope would be infinite. On the other hand, a surface satisfying
[6] that have been applied for the retrieval of surface parame- (4.1) for any large value of the distance would suffer from the
ters do not take into account any correlation length dependence. infinite variance problem (infrared catastrophe). Fortunately,
Two-parameter autocorrelation functions have also been pro- real natural surfaces satisfy (4.1) only in a limited (although
posed in the direct modeling of the backscattering from random wide) range of values (or, equivalently, of spatial frequen-
rough surfaces. In this case, surfaces whose characteristics may cies). In addition, whenever a scattering problem is in order,
vary according to the considered scale can also be modeled [14]. the range of distances to be considered is limited on one side
However, in this case, the number of surface parameters to be by the illuminated patch size (pixel size), and on the other by
retrieved becomes higher. the incident electromagnetic wavelength (values of much
Hence, correlation length discrepancies in its use and diffi- smaller than need not be considered). Therefore, bandlimited
culties in its estimation call for new and different models for fBm surfaces must be used [16], [19]–[23], and hence, the
surface description. These models are discussed in Section IV. lower and upper cutoff frequencies must be introduced. Limits
of validity of scattering models obviously depend on such
IV. THEORETICAL BACKSCATTERING MODELS: THE FRACTAL cutoff frequencies [22], [23]. However, here we want to stress
APPROACH again that these latter parameters are often not related to surface
properties, but rather to its illumination (illuminated patch size,
A. Surface Description wavelength), and therefore, they are a priori known.
It has been proven that the fractal geometry provides a sound
and reliable description of natural surfaces [28]–[32]. However,
B. Scattering Models
in order to apply the fractal geometry to remote sensing, it is
mandatory to devise a fractal model for the surface that al- Details on evaluating the backscattered field from an fBm sur-
lows a possibly approximate evaluation of the scattered electro- face can be found in the available literature. Here, we only re-
magnetic field. A viable fractal model for the surface descrip- port the main results for SPM and PO. Analytical solutions in
tion is the fractional Brownian motion (fBm). By definition, an closed form for the scattering from an fBm surface have been
isotropic surface belongs to the class of fBm if [32] recently obtained by using the IEM [25], [26]. However, they
are too involved to be used in a retrieval scheme as they are.
Furthermore, the IEM, even if applied to nonfractal surfaces, de-
serves several problems (see also Section V) in the development
of retrieval procedure. As of now, no fully theoretical inversion
(4.1) scheme exists for this model (in [9], linear approximations of
(2.1) are used).
where is the Hurst coefficient, is the topothesy and
Recent advances in the electromagnetic scattering theory
(4.2) from fBm surfaces have shown that the SPM backscattering
coefficient of an fBm isotropic random rough surface (except
Hence, an fBm surface is characterized by stationary, that at near-vertical incidence) can be expressed in terms of
Gaussian-distributed height increments, while the surface is not fractal parameters, as [21], [22]:
stationary.1 Its description relies on only the two parameters
1Note that any measured fBm surface height standard deviation depends on
(4.4)
the distance  for which it is determined [25], [32].
644 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 38, NO. 2, MARCH 2000

where (4.8) and (4.10) can be considered as useful guidelines to decide


the applicability of the SPM.
(4.5) We note that, for fixed values of the fractal parameters and
and of the incident wavelength , (4.8) defines a maximum
Similarly, the PO backscattering coefficient of an fBm area on the fractal surface, for which SPM is valid. Furthermore,
isotropic random rough surface can be expressed, in terms of for fixed values of , , and of the illuminated area, (4.8) and
fractal parameters, as2 (4.10) define a minimum incident wavelength for which SPM is
valid.
We finally observe that for high values of the incident wave-
length, the penetration depth is not negligible, and terrain ho-
mogeneity must also be checked if a surface scattering model is
to be applied.
(4.6)
V. SURFACE DIELECTRIC PARAMETERS RETRIEVAL
see [27] for perfectly conducting surfaces and [20] and [23] for In the general case, for the surface parameter retrieval, a min-
dielectric surfaces. imization scheme with at least four unknown parameters must
be devised (see equations in Section II and discussion in Section
C. Limits of Validity III). However, the overall procedure is dramatically simplified
if the backscattering coefficient can be expressed as the product
It is well known that SPM can be used if the surface height
of two factors, the first dependent on polarization but indepen-
standard deviation is much smaller than wavelength and the
dent of surface roughness and vice versa (the second dependent
RMS slope is not high, while PO can be used if the surface
on roughness but independent of polarization). In fact, in this
mean radius of curvature is much larger than the wavelength.
case, the ratio between the backscattering coefficient at and
However, these validity limits must be expressed in terms of the
polarization (copolarized ratio) is independent of roughness,
fractal parameters appearing in (4.4)–(4.6). This point is cur-
and a minimization over only two real parameters must be per-
rently under study, and with regard to the PO approach, the
formed to retrieve the complex dielectric constant. This condi-
reader can refer to the discussion reported in [23]. Here, we
tion is verified by SPM and PO solutions, but is not verified by
focus our attention on the SPM, which is actually used in the
the IEM approach (see (2.1), (2.3), and (2.5)). Furthermore, in
proposed retrieval procedure (see Sections V and VI).
the PO case, the dependence on polarization is too weak, so that
The surface height standard deviation of an fBm surface de-
the SPM is the most suitable solution for this retrieval scheme.
pends on its linear size. In particular, it can be shown (e.g., [22])
Obviously, applicability of the procedure will be limited to sur-
that
faces satisfying SPM validity limits. Furthermore, we explicitly
note that the complex dielectric constant retrieval algorithm is
(4.7)
independent of the employed geometric surface description.
From (2.3) and (2.4) we get
where is the lower cutoff frequency related to the illumi-
nated patch size . As a consequence, in order
to assess the validity of the SPM approach applied to fBm sur-
faces, we must require that
(5.1)
(4.8)
By analyzing the three-dimensional (3-D) plots of this ratio
With regard to the RMS slope, it is related to the small scale for fixed values of (see Fig. 1), we note that the copolarized
limit . In particular, it can be shown (see, e.g., [23]) that the ratio decreases as increases, and that for a fixed , it decreases
slope variance is as and increase. Furthermore, dependence on is much
weaker than dependence on , especially if is greater than 10.
(4.9) If multipolarization measured data at two or more incidence
angles are available, the complex dielectric constant can be es-
where is related to the incident wavelength ( ). timated by minimizing the following function:
As a consequence, in order to assess the validity of the SPM
approach applied to fBm surfaces, we must require that (5.2)

(4.10)
where is the ratio of measured backscattering coefficient at
The determination of more precise limits calls for the use of and polarization.
numerical methods [44] and is currently under study. However, We verified that, due to the monotonic behavior of versus
2For 0 < H < 1=2, expression (4.6) can be used only for small values of and for fixed , a simple steepest-descent algorithm is
kT sin #=(kT cos #) [23], [27]. able to recover the correct complex dielectric constant if the
FRANCESCHETTI et al.: FRACTAL-BASED THEORETICAL FRAMEWORK FOR RETRIEVAL OF SURFACE PARAMETERS 645

(a) (b)

(c) (d)

Fig. 1. Plots of f for (a) # = 20 , (b) # = 30 , (c) # = 40 , and (d) # = 50 .

SPM holds and measured backscattering data are not affected by Calibration errors over the two copolarized channels can be
errors. This test was performed by replacing in (5.2) measured modeled as follows [8]:
data with theoretically evaluated data (5.1) at a prescribed
value of the complex dielectric constant and applying
the minimization algorithm to the resulting function. Obviously, (5.3)
this test was performed for several values of .
Let us now analyze the behavior of the proposed method with where and are the measured backscattering coefficients
respect to measurement errors. For polarimetric SAR data, the is the absolute calibration error, and is the relative cali-
main error sources are speckle noise and calibration errors. bration error. From (5.3), we get
With regard to speckle noise, we note that the SPM approx-
imation leads to a correlation coefficient between and
channels equal to unity [43], so that multiplicative speckle (5.4)
noises on the two channels are equal and cancel out in the
copolarized ratio. Therefore, in principle, speckle noise does which shows that only the relative calibration errors affect the
not affect the complex dielectric constant retrieval algorithm. (complex) dielectric constant retrieval algorithm.
In practice [43], the correlation coefficient between and In order to test the effect of relative calibration errors on the
channels is slightly less than unity (0.8 0.9), and the speckle estimation accuracy, we substituted (5.4) in (5.2) with
noise slightly decreases estimation accuracy. In the following dB, and again used a steepest-descent minimization algorithm.
analysis, we neglect this marginal effect, which can be made Results are reported in Table I (second row) for three different
very small by using multilook data. values of the complex dielectric constant, which is all the values
646 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 38, NO. 2, MARCH 2000

TABLE I TABLE II
ESTIMATION OF COMPLEX PERMITTIVITY MEASURED BACKSCATTERING COEFFICIENTS USED TO TEST THE PROPOSED
FROM BACKSCATTERED DATA AFFECTED BY RELATIVE CALIBRATION ERRORS RETRIEVAL ALGORITHM

actually measured over bare soil surfaces described in [6]. It can


has been obtained in the estimation of , whereas the errors on
be noted that the estimated is always zero, while is un-
are appreciably reduced.
derestimated, and the estimation errors increase as increases.
We now prove that the effect of relative calibration errors can be Finally, an example of application to real data is in order.
reduced if we use a modified copolarized ratio, defined by the This allows us to test the performance of the proposed retrieval
following relation: method with respect to inaccuracies of the SPM. Therefore, we
chose a surface that is within the validity region of SPM, but
close to its limit. Furthermore, in order to minimize other error
(5.5) sources, high accuracy scatterometer data are used. In partic-
ular, POLARSCAT measured data at 1.5 GHz are taken from
The corresponding function to minimize is [6] and are reported in Table II. In [6], the surface roughness is
described by using a classical (nonfractal) model, and accord-
ingly, the roughness is characterized by its standard deviation.
(5.6)
The measured value of the height standard deviation reported
in [6] is such that , so that SPM should hold [14]
By using (5.3) in (5.5) we get even if we are close to its validity limit. However, the value of
reported in [6] depends (see Section IV) on the length of the
profile used to measure it, that is, 1 m [6]. The linear size of the
(5.7) area illuminated by POLARSCAT at 1.5 GHz can be derived
from the radar data reported in [6] and is about 3 m. In view
of (4.7), this means that the value of we should consider is
Since is usually close to unity, we can assume
slightly higher than 0.13, the exact value depending on and
.
with (5.8)
The value of the complex dielectric constant obtained by in
situ measurements is [6]. Unfortunately, a signifi-
Accordingly, we get
cant underestimation is obtained by our procedure. In fact, the
retrieved dielectric constant is about . Analogous results
(5.9) are obtained by using data relevant to the same surface with a
different moisture content. The measured value of the complex
dielectric constant is [6], whereas the retrieved one is
In dB, about . This underestimation is due to the fact that the mea-
sured separation between copolarized channels is smaller with
respect to the one predicted by the SPM. This is in agreement
with the fact that numerical simulations [44] have shown that as
(5.10) the height standard deviation increases, the separation between
and backscattering coefficient decreases from the value
predicted by SPM to the one predicted by PO. In addition, this
If we compare this expression with the analogous expression for separation can be further reduced by the presence of unmodeled
, i.e., multiple scattering. This latter effect could be avoided by re-
moving the multiple scattering component from measured data
(5.11) by using the Cloude’s target decomposition theorem [9], [45].
This point is currently under investigation.
we conclude that the effect of calibration errors is smaller on In conclusion, use of the SPM and of modified backscat-
than on . tering copolarized ratio leads to a dielectric constant estimation
In order to test the effect of relative calibration errors on the scheme that correctly retrieves the actual value of the complex
estimation accuracy obtained by using the function, we sub- dielectric constant if the SPM holds and backscattering mea-
stituted (5.7) in (5.6) with dB, and used again a sured data are not affected by measurement errors. The quoted
steepest-descent minimization algorithm. Results are reported estimation scheme is insensitive to any absolute calibration
in Table I (third row) for the same three values of the complex error, are virtually unaffected by speckle noise, and is quite
dielectric constant used above. We note that no improvement robust with respect to relative calibration errors. However, it
FRANCESCHETTI et al.: FRACTAL-BASED THEORETICAL FRAMEWORK FOR RETRIEVAL OF SURFACE PARAMETERS 647

is very sensitive to the validity of the solution we accept for TABLE III
ESTIMATED FRACTAL PARAMETERS FOR DIFFERENT NUMBERS OF LOOKS.
the scattered field, in this case, SPM. This is to be expected, TRUE VALUES: H = 0:6 AND S = 40 dB0
because the retrieval procedure is based on an analytical
solution: should the latter be invalid, the retrieval algorithm
does not converge to the right values of the dielectric constant.
Furthermore, the imaginary part of the dielectric constant
is not correctly retrieved whenever the relative calibration
error is at least 0.3 dB. A further study on this point is in
progress. However, the latter is not a great limitation, because
it is possible to use a semi-empirical model [45], [47] that
relates the soil moisture content to the real part of the dielectric
constant. If needed, the same semi-empirical model can be used
to obtain the imaginary part of the dielectric constant from the
soil moisture content. frequencies and incidence angles is limited to the validity
region of the SPM model.
Speckle noise over the backscattering coefficient can be mod-
VI. SURFACE ROUGHNESS PARAMETERS RETRIEVAL
eled as follows:
Once the dielectric constant is known, the retrieval of geo-
metric surface parameters is in order. In this case, use of the (6.3)
fractal fBm model is particularly convenient, not only because
it is very suitable for natural surface description, but also be- where is a Chi-squared distributed random variable with
cause it leads to a very simple estimation scheme based on a parameter ( being the number of looks). Accordingly, the
linear regression. As a matter of fact, (4.4) can be rewritten as mean of is 1, and its variance is . By using this speckle
model and (4.4) with , dB, and , we
simulated measured data at 1 GHz, polarization for different
numbers of looks and applied the retrieval method based on the
(6.1) linear regression of (6.1). Results are summarized in Table III.
Note that for single look data, a completely unreliable estimate
of fractal parameters is obtained. For , estimates
Therefore, points whose coordinates are , with are biased and somehow inaccurate, although reasonable. Good
results are obtained for equal or greater than 64.
Finally, in order to test the performance of the proposed re-
trieval method with respect to the accuracy of the SPM model,
(6.2)
we applied the linear regression of (6.1) to the multi-incidence
angle data reported in Table II and tried to retrieve the fractal
lay on a straight line whose slope is , and whose inter- parameters of the actual surface considered in the previous sec-
section with the axis is . tion. By employing the polarization data and using (4.3) and
Therefore, if the dielectric constant is known, and if multifre- (4.5), we get and mm. In spite of the SPM
quency and/or multi-incidence angle data are available, a linear limitations, both values are within the range of values obtained
regression can be performed to evaluate and and hence for natural surfaces by in situ measurements and reported in lit-
(see (4.3) and (4.5)), and . Note that whenever multifre- erature [30], [35]. By employing the polarization data, we
quency data are used, the dielectric constant must be known obtain slightly different, but still reasonable, values: ,
at every different employed frequency. If this is not the case, a mm.
model to account for the dependence on frequency [46] must Unfortunately, in situ-measured fractal parameters of the con-
be used. sidered surface are not available, and therefore, a direct com-
We verified that the above simple linear regression algorithm parison between retrieved and actual fractal parameters is not
is able to recover the correct fBm fractal parameters if the SPM possible. However, by using (4.5) and (4.7), we can evaluate the
model holds and measured backscattering data are not affected fractal profile standard deviation corresponding to the retrieved
by errors. fractal parameters, and then compare it to the one reported in
Let us now analyze the behavior of the proposed method [6]. Note that in (4.7), a value of the profile length must also
with respect to measurement errors (i.e., calibration errors and be inserted. In the case at hand, it is appropriate to use a pro-
speckle). It follows from (6.1) that both relative and absolute file length equal to the one of the laser profilometer used in [6]
calibration errors affect the geometric parameters retrieval. (i.e., 1 m). It turns out that retrieved fractal parameters lead to
However, if the calibration error is the same on all employed a value of equal to about 0.2, in rather good agreement with
data, it does not affect the estimate of (and hence, of ), and the value reported in [6].
the error on is equal to the calibration error. With regard to At this point, we have to verify that the retrieved values of sur-
the speckle noise, it can be reduced by increasing the number face fractal parameters satisfy the SPM limits of validity ((4.8),
of looks and increasing the number of employed frequencies (4.10)). Note that in (4.8), a value of the illuminated patch size
or incidence angles. Note however, that the range of usable must also be inserted. In the case at hand, it is appropriate to use
648 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 38, NO. 2, MARCH 2000

an illuminated patch size equal to the POLARSCAT one (at 1.5 With regard to complex permittivity retrieval, it turned out
GHz, i.e., about 3 m, see Section V). It turns out that both (4.8) that the proposed estimation scheme is insensitive to any abso-
and (4.10) are hardly satisfied, thus confirming that the consid- lute calibration error, is virtually unaffected by speckle noise,
ered surface is very close to SPM validity limit. and is quite robust with respect to relative calibration errors.
In conclusion, use of the fBm fractal surface model within the However, it is very sensitive to inaccuracies of the scattering
SPM leads to a very simple surface parameter retrieval scheme model. Furthermore, the imaginary part of the dielectric con-
based on a linear regression that obviously correctly retrieves stant is not correctly retrieved if a 0.3 dB relative calibration
the actual values of the fBm fractal parameters if the SPM holds error is present. However, the latter is not a great limitation, be-
and backscattering measured data are not affected by measure- cause it is possible to use a semi-empirical model that relates
ment errors. In addition, the fBm fractal parameters estimation imaginary and real parts of the dielectric constant through the
scheme is virtually unaffected by the speckle noise if the number soil moisture content.
of look is at least 64 and seem to be robust regarding the inac- With regard to fractal geometric parameters retrieval, results
curacies of the scattering model. Furthermore, the fBm fractal can be summarized by noting that the fBm fractal parameters es-
dimension estimation scheme is insensitive to calibration errors timation scheme is practically unaffected by the speckle noise if
if the latter do not vary with the incidence angle. It is worth the number of look is at least 64 and seem to be robust regarding
recalling that the surface fractal dimension is a meaningful pa- the inaccuracies of the scattering model. Furthermore, the fBm
rameter, since it is a measure of surface roughness [28]–[32], fractal dimension estimation scheme is insensitive to calibration
and that fractal parameters have been successfully used for ter- errors if the latter do not vary with the incidence angle. This
rain classification [33]–[36]. latter result plays an interesting role in assessing the reliability
Finally, note that an approach for the estimation of the sur- of fractal dimension-based procedures used for terrain classifi-
face spectrum similar to the one we propose was presented in cation from radar data.
[2]. However, in that paper, no reference was made to fractal Finally, as a general comment, we stress again that our proce-
geometry, and the importance of the power spectrum exponent dure can be safely applied only to surfaces that satisfy SPM va-
coefficient was not recognized. lidity limits. Reasons that do not allow the introduction of a full
theoretical scheme based on PO model or the more general IEM
model have also been discussed. Extension of our procedure to a
VII. CONCLUSIONS wider class of surfaces would imply that it is acceptable to leave
a fully theoretical approach and to use, for instance, semi-empir-
A new, fully theoretical framework for the retrieval of surface ical models like the one reported in [9] combined with the use of
electromagnetic and geometric parameters has been presented. the fBm surface description. Another possibility in the future is
The model relies on use of SPM and the fractal geometry. In to develop new and reliable direct scattering models tailored on
particular, the random rough surface is described by means of an fractal surfaces not fulfilling the SPM criteria that should also
fBm fractal process. The retrieval scheme requires availability be amenable to devising a theoretical inversion procedure.
of multipolarization data acquired at different incidence angles As a guideline for future work, an extensive comparison of the
and/or frequencies. retrieval algorithm results with in situ measurements is advis-
Use of SPM is motivated by the fact that it leads to an in- able. Interestingly, at present, no in situ measurements of fractal
version technique that can be split into two simpler steps. In surface parameters, combined with the corresponding scattering
the first one, the complex permittivity is estimated, and in the measurements, are available. Announced availability in the near
second one, the geometric surface parameters are retrieved. On future [48] of a database involving in situ roughness profile data
the other hand, use of fractal geometry is preferred because it and corresponding scattering data will allow the direct measure-
is a very suitable tool for natural surface description, and be- ment of surface fractal parameters, thus making it possible to
cause its use in conjunction with SPM allows a straightforward perform a complete assessment of the proposed algorithm per-
retrieval scheme based on a linear regression. As a first result, formance.
it turns out that use of SPM, the modified backscattering copo-
larized ratio, and the fBm fractal model for the surface profile
description, lead to complex dielectric constant and fractal pa-
rameters estimation schemes that correctly retrieve the actual REFERENCES
values, provided that SPM holds and backscattering measured [1] T. Mo, J. R. Wang, and T. J. Schmugge, “Estimation of surface rough-
data are affected by negligible measurement errors. Conditions ness parameters from dual-frequency measurements of radar backscat-
for assessing the applicability of the SPM model to fractal sur- tering coefficients,” IEEE Trans. Geosci. Remote Sensing, vol. 26, pp.
574–579, Sept. 1988.
faces has also been provided. [2] J. J. van Zyl, C. F. Brunette, and T. G. Farr, “Inference of surface power
Performances of the proposed method in the presence of spectra from inversion of multifrequency polarimetric radar data,” Geo-
calibration errors and speckle noise have been tested by means phys. Res. Lett., vol. 18, pp. 1787–1790, 1991.
[3] D. L. Evans, T. G. Farr, and J. J. van Zyl, “Estimates of the surface rough-
of theoretical analysis and simulations. In addition, the effect ness derived from synthetic aperture radar (SAR) data,” IEEE Trans.
of model inaccuracies has been tested by applying the retrieval Geosci. Remote Sensing, vol. 30, pp. 382–389, Mar. 1992.
scheme to a set of measured data relevant to a surface that [4] E. Altese, O. Bolognani, M. Mancini, and P. A. Troch, “Retrieving soil
moisture over bare soil from ERS-1 synthetic aperture radar data: Sensi-
only marginally falls within the validity limits required by tivity analysis based on a theoretical surface scattering model and field
the SPM. data,” Water Resources Res., vol. 32, no. 3, pp. 653–661, 1996.
FRANCESCHETTI et al.: FRACTAL-BASED THEORETICAL FRAMEWORK FOR RETRIEVAL OF SURFACE PARAMETERS 649

[5] G. R. Taylor, A. H. Mah, F. A. Kruse, K. S. Kierein-Young, and R. D. [35] T. R. Austin, A. W. England, and G. H. Wakefield, “Special problems
Bennet, “The extraction of soil properties from AIRSAR data,” Int. J. in the estimation of power-law spectra as applied to topographical mod-
Remote Sensing, vol. 17, no. 3, 1996. eling,” IEEE Trans. Geosci. Remote Sensing, vol. 32, pp. 928–939, July
[6] Y. Oh, K. Sarabandi, and F. T. Ulaby, “An empirical model and an inver- 1994.
sion technique for radar scattering from bare soil surfaces,” IEEE Trans. [36] M. Coltelli, G. Fornaro, G. Franceschetti, R. Lanari, M. Migliaccio,
Geosci. Remote Sensing, vol. 30, pp. 370–381, Mar. 1992. J. R. Moreira, K. P. Papathanassiou, G. Puglisi, D. Riccio, and M.
[7] , “An inversion algorithm for retrieving soil moisture and surface Schwäbisch, “SIR-C/X-SAR multifrequency multipass interferometry:
roughness from polarimetric radar observation,” in Proc. IGARSS’94, A new tool for geological interpretation,” J. Geophys. Res., vol. 101,
Pasadena, CA, pp. 1582–1584. pp. 23 127–23 148, 1996.
[8] P. C. Dubois, J. J. van Zyl, and T. Engman, “Measuring soil moisture [37] M. F. Chew and A. K. Fung, “A numerical study of the regions of va-
with imaging radars,” IEEE Trans. Geosci. Remote Sensing, vol. 33, pp. lidity of the Kirchhoff and small-perturbation rough surface scattering
915–926, July 1995. models,” Radio Sci., vol. 23, pp. 163–170, Mar.–Apr. 1988.
[9] J. Shi, J. Wang, A. Y. Hsu, P. E. O’Neill, and E. Engman, “Estimation of [38] A. Collaro, G. Franceschetti, M. Migliaccio, and D. Riccio, “Gaussian
bare surface soil moisture and surface roughness parameter using L-band rough surfaces and Kirchhoff approximation,” IEEE Trans. Antennas
SAR image data,” IEEE Trans. Geosci. Remote Sensing, vol. 35, pp. Propagat., vol. AP-47, pp. 392–398, Feb. 1999.
1254–1266, Sept. 1997. [39] G. Franceschetti, M. Migliaccio, D. Riccio, and G. Schirinzi, “SARAS:
[10] F. T. Ulaby, R. K. Moore, and A. K. Fung, Microwave Remote A raw signal simulator,” IEEE Trans. Geosci. Remote Sensing, vol. 30,
Sensing. Reading, MA: Addison-Wesley, 1982, vol. 2. pp. 110–123, 1992.
[11] L. Tsang, J. A. Kong, and R. T. Shin, Theory of Microwave Remote [40] H. A. Zebker, C. H. Werner, P. A. Rosen, and S. Hensley, “Accuracy
Sensing. New York: Wiley, 1985. of topographic maps derived from ERS-1 interferometric radar,” IEEE
[12] P. Beckmann and A. Spizzichino, The Scattering of Electromagnetic Trans. Geosci. Remote Sensing, vol. 32, pp. 823–836, July 1994.
Waves from Rough Surfaces. Norwood, MA: Artech House, 1987. [41] Y. Oh and Y. C. Kay, “Condition for precise measurements of soil surface
[13] A. K. Fung, Z. Li, and K. S. Chen, “Backscattering from a randomly roughness,” IEEE Trans. Geosci. Remote Sensing, vol. 36, pp. 691–695,
rough dielectric surface,” IEEE Trans. Geosci. Remote Sensing, vol. 30, May 1998.
pp. 356–369, Mar. 1992. [42] E. T. Engman and J. R. Wang, “Evaluating roughness models of radar
[14] A. K. Fung, Microwave Scattering and Emission. Models and Their Ap- backscatter,” IEEE Trans. Geosci. Remote Sensing, vol. 25, pp. 709–713,
plications. Norwood, MA: Artech House, 1994. Nov. 1987.
[15] A. Ishimaru, Wave Propagation and Scattering in Random [43] M. Borgeaud and J. Noll, “Analysis of theoretical surface scattering
Media. New York: Academic, 1993. models for polarimetric microwave remote sensing of bare soils,” Int.
[16] D. L. Jaggard, “On fractal electrodynamics,” in Recent Advances in J. Remote Sensing, vol. 15, no. 14, pp. 2931–2942, 1994.
Electromagnetic Theory, H. N. Kritikos and D. L. Jaggard, Eds. Berlin, [44] M. F. Chen and S. Y. Bai, “Computer simulation of wave scattering from
Germany: Springer Verlag, 1990, pp. 183–223. a dielectric random surface in two dimensions-cylindrical case,” J. Elec-
[17] Y. Agnon and M. Stiassnie, “Remote sensing of the roughness of a fractal tromagn. Waves Applicat., vol. 4, no. 10, pp. 963–982, 1990.
sea surface,” J. Geophys. Res., vol. 96, pp. 12 773–12 779, July 1991. [45] S. R. Cloude, “Uniqueness of target decomposition theorems in radar
[18] J. Chen, T. K. L. Lo, H. Leung, and J. Litva, “The use of fractals for polarimetry,” in Direct and Inverse Methods in Radar Polarimetry: Part
modeling EM waves scattering from rough sea surface,” IEEE Trans. 1. Dordrecht, The Netherlands: Kluwer, 1992, pp. 266–296.
Geosci. Remote Sensing, vol. 34, pp. 966–972, July 1996. [46] M. T. Hallikainen, F. T. Ulaby, M. C. Dobson, M. A. El-Rayes, and
[19] G. Franceschetti, M. Migliaccio, and D. Riccio, “An electromagnetic L. Wo, “Microwave dielectric behavior of wet soil—Part I: Empirical
fractal-based model for the study of fading,” Radio Sci., vol. 31, pp. models and experimental observations,” IEEE Trans. Geosci. Remote
1749–1759, Nov.–Dec. 1996. Sensing, vol. GE-23, pp. 25–34, Jan. 1985.
[20] G. Franceschetti, A. Iodice, M. Migliaccio, and D. Riccio, “Backscat- [47] N. R. Peplinski, F. T. Ulaby, and M. C. Dobson, “Dielectric properties of
tering from an fBm surface,” in Proc. URSI Electromagnetic Theory soils in the 0.3–1.3-GHz range,” IEEE Trans. Geosci. Remote Sensing,
Symp., Thessaloniki, Greece, 1998, pp. 692–694. vol. 33, pp. 803–807, May 1995.
[21] , “Scattering from a fractal sea surface,” Atti della Fondazione [48] M. Borgeaud and A. Bellini, “A database for electromagnetic scattering
Ronchi., vol. LIV, no. 3–4, pp. 455–462, May–Aug. 1999. studies of bare soil surfaces,” in Proc. IGARSS’98, Seattle, WA, pp.
[22] , “Fractals and the small perturbation scattering model,” Radio Sci- 1197–1199.
ence, vol. 34, pp. 1043–1056, Sept./Oct. 1999.
[23] , “Scattering from natural rough surfaces modeled by fractional
Brownian motion two-dimensional processes,” IEEE Trans. Antennas
Propagat., vol. 47, pp. 1405–1415, Sept. 1999.
[24] M. V. Berry and T. M. Blackwell, “Diffractal echoes,” J. Phys. A Math.
Gen., vol. 14, pp. 3101–3110, 1981.
[25] F. Mattia and T. Le Toan, “Backscattering properties of multi-scale
rough surfaces,” J. Electromagn. Waves Applicat., vol. 13, pp. 491–526,
1999. Giorgio Franceschetti (S’60–M’62–SM’85–F’88)
[26] A. T. Manninen, “Multiscale surface roughness and backscattering,” was born in Italy.
Prog. Electromag. Res., vol. 16, pp. 175–203, 1997. He was appointed Full Professor of Electromag-
[27] O. Y. Yordanov and K. Ivanova, “Kirchhoff diffractals,” J. Phys. A Math. netic Wave Theory at the University of Naples,
Gen., vol. 27, pp. 5979–5993, 1994. Italy, in 1968. He was a Visiting Professor with
[28] B. B. Mandelbrot, The Fractal Geometry of Nature. New York: W. H. the University of Illinois, Champaign, in 1976 and
Freeman, 1983. 1977, the University of California, Los Angeles
[29] R. F. Voss, “Random fractal forgeries,” in Fundamental Algorithms for (UCLA), in 1980 and 1982, National Somali
Computer Graphics, R. A. Earnshaw, Ed. Berlin, Germany: Springer University, Somalia, in 1984, and the University of
Verlag, 1985, pp. 805–835. Santiago de Compostela, Spain, in 1995. He was
[30] S. R. Brown and C. H. Scholz, “Broad band study of the topography a Research Associate with the California Institute
of natural rock surfaces,” J. Geophys. Res., vol. 90, pp. 12 575–12 582, of Technology, Pasadena, in 1981 and 1983. He was a Director of IRECE, a
Dec. 1985. Research Institute of CNR (the Italian National Council of Research) and a
[31] P. Flandrin, “On the spectrum of fractional Brownian motions,” IEEE Member of the Board of the Italian Space Agency (ASI). He is currently an
Trans. Inform. Theory, vol. 35, pp. 197–199, Jan. 1989. Adjunct Professor at UCLA. He has published several books and more than
[32] K. Falconer, Fractal Geometry. Chichester, U.K.: Wiley, 1990. 130 refereed papers in the field of applied electromagnetics (reflector antennas,
[33] K. Yocoya, K. Yamamoto, and N. Funakubo, “Fractal-based analysis transient phenomena, shielding, nonlinear propagation, and scattering) and,
and interpolation of 3-D natural surface shapes and their application to more recently, in the field of SAR data processing and simulation. He has
terrain modeling,” Comput. Vis., Graph., Image Processing, vol. 46, pp. lectured in several summer schools in China, the U.K., The Netherlands, Italy,
284–302, 1989. Spain, Sweden, and the United States.
[34] C. V. Stewart, B. Moghaddam, K. J. Hintz, and L. M. Novak, “Fractional Prof. Franceschetti was a Fulbright Scholar with the California Institute of
Brownian motion models for synthetic aperture radar imagery scene seg- Technology in 1973. He was the recipient of several national and international
mentation,” Proc. IEEE, vol. 81, pp. 1511–1522, 1993. awards.
650 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 38, NO. 2, MARCH 2000

Antonio lodice (S’97–M’00) was born in Naples, Daniele Riccio (M’91–SM’99) was born in Naples,
Italy, on July 4, 1968. He received the Laurea degree Italy, on April 13, 1962. He received the Laurea De-
in electronic engineering and the Ph.D. degree in gree in electronic engineering from the University of
electronic engineering and computer science from Naples “Federico II,” Naples, in 1989.
the University of Naples “Federico II,” Naples, Italy, He was a Research Scientist at the “Istituto di
in 1993 and 1999, respectively. Ricerca sull’Elettromagnetismo e i Componenti
In 1995, he was with the Istituto di Ricerca per Elettronici” (IRECE), Naples, of the Italian National
l’Elettromagnetismo e i Componenti Elettronici Council of Research (CNR). Since 1994, he has been
(IRECE ), Naples, Italy, working under a Grant from a Research Scientist at the Department of Electronic
the Italian National Council of Research (CNR) and Telecommunication Engineering, University
for research in the field of remote sensing. Since of Naples “Federico II,” where he is a Lecturer of
1996, he has been with the Department of Electronic and Telecommunication Remote Sensing and Electromagnetic Diagnostic. From 1994 to 1995, he was
Engineering of the University of Naples “Federico II.” He is also currently a Guest Scientist with the DLR High-Frequency Institute, Munich, Germany.
with Telespazio S.p.A., Rome, Italy. Since 1998, he has also been a Research Scientist with IRECE. His main
His main research interests are in the field of SAR remote sensing: modeling research interests are in the fields of simulation and modeling of SAR signals
of electromagnetic scattering from natural surfaces, and SAR interferometry. relevant to terrestrial and oceanic scenes, as well as in the application of fractal
geometry to electromagnetic scattering and remote sensing.
Dr. Riccio has won several fellowships from private and public companies
(SIP, Selenia, CNR, CORISTA, CRATI) for research in the remote sensing field.
Stefania Maddaluno was born in Naples, Italy, on
September 24, 1972. She received the Laurea degree
in telecommunication engineering from the Univer-
sity of Naples “Federico II,” Naples, in 1998.
She joined the Department of Electronic and
Telecommunication Engineering, University of
Naples “Federico II,” where she was involved in
research in the field of remote sensing. Presently, she
is with Alenia Aerospace as an Expert in Avionics.

You might also like