Evaluation of a rough soil surface description with ASAR-ENVISAT radar data M. Zribi a, * , N. Baghdadi b , N. Holah b , O. Fafin a , C. Gue ´rin a a CETP/CNRS, 10-12 av. de l’Europe, 78140 Ve ´lizy, France b Bureau de Recherches Ge ´ologiques et Minie `res (BRGM), ARN/ATL 3 av C. Guillemin, B.P. 6009, 45060 Orle ´ans Cedex 2, France Received 21 July 2004; received in revised form 24 November 2004; accepted 28 November 2004 Abstract The input roughness parameters for electromagnetic backscattering modelling need to be accurate to estimate radar measurements correctly over bare soils, particularly in agricultural environments. This paper proposes to evaluate the roughness description in terms of several characterisations through a correlation function using a numerical backscattering model. The experimental database used in this study is based on ASAR-ENVISAT experimental campaigns in the Beauce region (France). Two presentations of the surface height correlation function are proposed in this study. The first one, referred to as the ba functionQ fits the experimental correlation functions up to the correlation length, while the second one, the b(a,b ) functionQ, fits the correlation function for scales corresponding to positive values. A relationship is proposed between the rms height of soil surface and the shape of the correlation function. Using the a function, comparisons between radar measurements for high incidence angles and simulations based on the numerical backscattering model (moment method) show a good agreement for soil surfaces with an rms height smaller than 2 cm with medium and high soil moisture. D 2004 Elsevier Inc. All rights reserved. Keywords: Radar; ENVISAT; ASAR; Roughness; Backscattering 1. Introduction Soil moisture and roughness play a key role in hydro- logical and climate studies. Considerable effort has been devoted in active microwave remote sensing to study radar backscattering response from natural surfaces (e.g. Bagh- dadi et al., 2002, 2004; Jackson et al., 1996; Ulaby et al., 1986; Zribi & Dechambre, 2002). Electromagnetic back- scattering models (Kirchoff models, the small perturbation model (Ulaby et al., 1986)) and more recently, the Integral Equation Model (IEM, Fung et al., 1992) have been developed to study this question. However, different experimental measurements have shown that they were restricted to smooth or very rough soil surfaces. These difficulties are attributed to two factors: the first one is the soil roughness description based only on two surface parameters (the rms height (s ) and the correlation length (l )) and generally an exponential correlation function for all surface types; the second one is the physical approximations introduced in these models. For example, the small perturbation model (SPM) is valid only for very smooth soils, Kirchoff approximations correspond to very rough surfaces, and for IEM, it is still difficult to identify a large validity domain for real agricultural soils in spite of the improvements developed over the past few years (Chen et al., 2000; Wu et al., 2001). The developed backscattering models generally neglect the volumetric contribution of soil. This hypothesis limits their validation to low surface moisture, where wave penetration is large (Fung, 1994). In that context, over the last few years, different studies have tried to introduce a more complete description of soil surface roughness for forward studies (Davidson et al., 2000; Li et al., 2002; Mattia & Le Toan, 1999; Mattia et al., 2003; 0034-4257/$ - see front matter D 2004 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2004.11.014 * Corresponding author. Tel.: +33 1 39 25 49 34; fax: +33 1 39 25 49 22. E-mail address: [email protected] (M. Zribi). Remote Sensing of Environment 95 (2005) 67 – 76 www.elsevier.com/locate/rse
10
Embed
Evaluation of a rough soil surface description with ASAR-ENVISAT radar data
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
www.elsevier.com/locate/rse
Remote Sensing of Environ
Evaluation of a rough soil surface description with
ASAR-ENVISAT radar data
M. Zribia,*, N. Baghdadib, N. Holahb, O. Fafina, C. Guerina
aCETP/CNRS, 10-12 av. de l’Europe, 78140 Velizy, FrancebBureau de Recherches Geologiques et Minieres (BRGM), ARN/ATL 3 av C. Guillemin, B.P. 6009, 45060 Orleans Cedex 2, France
Received 21 July 2004; received in revised form 24 November 2004; accepted 28 November 2004
Abstract
The input roughness parameters for electromagnetic backscattering modelling need to be accurate to estimate radar measurements
correctly over bare soils, particularly in agricultural environments. This paper proposes to evaluate the roughness description in terms of
several characterisations through a correlation function using a numerical backscattering model. The experimental database used in this
study is based on ASAR-ENVISAT experimental campaigns in the Beauce region (France).
Two presentations of the surface height correlation function are proposed in this study. The first one, referred to as the ba functionQ fitsthe experimental correlation functions up to the correlation length, while the second one, the b(a,b) functionQ, fits the correlation function
for scales corresponding to positive values. A relationship is proposed between the rms height of soil surface and the shape of the
correlation function. Using the a function, comparisons between radar measurements for high incidence angles and simulations based on
the numerical backscattering model (moment method) show a good agreement for soil surfaces with an rms height smaller than 2 cm with
Fig. 3. Comparison between experimental correlation function, a correlation function, (a, b) correlation function, exponential and gaussian correlation
functions for two agricultural test fields.
0
0,2
0,4
0,6
0,8
1
1,2
1,4
1,6
1,8
2
0 1 2 3 40,5 1,5 2,5 3,5 4,5
Rms height, s (cm)
α pa
ram
eter
09/02/2003
23/09/2003
26/09/2003
09/10/2003
09/11/2003
11/12/2003
Orgeval 94
06/05/2004
α=0.84log(s)+0.96 R=0.78, RMSE=0.19
Fig. 4. Variation of the a parameter as a function of the rms height s for different experimental campaigns.
M. Zribi et al. / Remote Sensing of Environment 95 (2005) 67–7670
M. Zribi et al. / Remote Sensing of Environment 95 (2005) 67–76 71
profiles, the shape of the correlation function is not usually
stable with a constant power, particularly beyond the
correlation length scale. Coefficients a and b are computed
by a mean square approach for positive correlation function
values. When b is close to zero, this means that the
correlation function could be simply fitted by a function
with a constant power. On the other hand, when b is large,
the shape of the function beyond the correlation length scale
corresponds more to a gaussian correlation function. This
case is largely present in real agricultural surfaces, as will be
illustrated below.
As an example, Fig. 3 shows a comparison of two
experimental correlation functions and the two proposed
correlation functions, where the gaussian and exponential
0
0,2
0,4
0,6
0,8
1
1,2
1,4
1,6
1,8
2
0 0,2 0,4 0,6 0,8
a par
α pa
ram
eter
(a)
(b)
-0,2
0
0,2
0,4
0,6
0,8
1
0 0,5 1 1,5
Rms he
b pa
ram
eter
Fig. 5. (a) Comparison between a and a parameters for test fields. (b) Beha
correlation functions are also plotted. We observe a good
agreement with the two functions (a and (a,b) functions) in
the first scales up to the correlation length. For values
between the correlation length scale and the scale corre-
sponding to a zero correlation, the second function fits better
the experimental function, particularly for Fig. 3b.
For all test field measurements, the different roughness
parameters are computed. Fig. 4 shows the a parameter as
a function of the rms height s for different campaigns.
Here, roughness measurements are illustrated for cam-
paigns in the Villamblain site. We have plotted on the
same graph other measurements made at the Orgeval site
(in 1994 with SIRC/XSAR campaign, Zribi et al., 1997).
The results clearly show an increase in the a parameter as
1 1,2 1,4 1,6 1,8 2
ameter
2 2,5 3 3,5 4
ight (cm)
viour of b parameter as a function of the rms height s for test fields.
M. Zribi et al. / Remote Sensing of Environment 95 (2005) 67–7672
a function of the rms height before an approximate settling
at high roughness values. This could be explained by the
fact that for small roughness values we have generally
more small clods and therefore a larger power at high
frequencies, and therefore a smaller a. High rms height
corresponds generally to ploughed soils with principally
large clods and then a limited power at high frequencies.
The second important observation deals with the
behaviour of a for the different campaigns. The a values
have a general tendency to increase with the oldness of the
agricultural tillage and soil moisture, except for rough
surfaces (sN1.5 cm). This behaviour is due particularly to
the effect of rain on the degradation of small clods (high-
frequency structures) and then the increase in a. For
surfaces with large rms height, the limited high-frequency
power explains the absence of tendency for a.In fact, the smallest values are generally retrieved
approximately for the dates 23/09/2003 and 26/09/2003.
These dates correspond to new agricultural tillage and
small or medium values of soil moisture (mean Mv equals
to 7 and 18%). For 09/10/2003 and 03/11/2003 measure-
ments, with rainfall after the end of September measure-
ments, soil has been degraded which induced an increase
in the a parameter. The date 11/12/2003 corresponds to
new tillage made at the beginning of December but
followed by large rainfall (mean Mv of about 30%).
During the Orgeval campaign, the highest a values are
observed. The measurements were made in April 1994
after a long period without agricultural tillage and strong
rainfall (mean soil moisture of about 35%). The date of 06/
05/2004 corresponds to the case of very smooth soil
surfaces with small a parameters.
To conclude on these results, it may be observed that
the a parameter might be one of the tools for quantifying
the evolution of soil surface structure and degradation, an
important parameter for agronomic and erosion studies
(Boiffin & Monnier, 1985; Le Bissonnais et al., 1989).
An empirical logarithmic relationship is proposed to link
the rms height to the a parameter with a high correlation
-20
0
20
40
60
80
100
120
0 200 400Distanc
Hei
ght (
mm
)
Fig. 6. Generation of three surfaces with s=0,7 cm
coefficient (R=0.78) and a root mean square error equal to
0.19. It is written as:
a ¼ 0:84log sð Þ þ 0:96 ð1Þ
This relationship could be considered as a tool for
eliminating one parameter (the shape of the correlation
function) in the inversion of radar measurements.
The a and b parameters of the second proposed function
are also computed over the whole set of measured soil
profiles. As observed in Fig. 5a, a and a are generally close,
particularly for surfaces with very small b values. On the
other hand, the second parameter b is not very stable. This is
illustrated in Fig. 5b, where b is plotted as function of the
rms height. We do not observe any significant correlation
between the rms height and the b parameter.
3.2. Simulation of soil surfaces
In order to study the backscattering behaviour of soil
surfaces and to estimate the difference between the two
proposed functions, we chose to use a numerical model
based on the moment method. This numerical approach
needs a generation of soil surface as input to roughness
description.
For the simulation of surface, we used the approach
described by Fung and Chen (1985), as follows:
The surface heights are written as:
h kð Þ ¼Xi¼M
i¼�M
W ið ÞX iþ kð Þ ð2Þ
where X(i) is a gaussian random variable N(0,1) and W( j) is
the weighting function given by W ið Þ ¼ F�1tffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiF C ið Þ½ �
pb,
where C(i) is the correlation function and F[] denotes the
Fourier transform operator.
For the proposed functions, there is no obvious analytical
Fourier transformation of the correlation functions. There-
fore, Fast Fourier Transformation (FFT) is used to compute
Fig. 7. Comparisons between backscattering simulations with the two types of correlation functions: a function and (a,b) function, for two incidence angles 268and 428.
M. Zribi et al. / Remote Sensing of Environment 95 (2005) 67–76 73
Fig. 6 shows different examples of surface generations
with different values of a (a=1 (exponential case) a=1.3,a=1.6). We observe, from the three illustrated examples,
the effect of this coefficient on the high-frequency
structures. The highest a coefficient corresponds to the
lowest high frequency power, that is, a surface with fewer
small clods.
3.3. Numerical simulation modelling
As noted previously, analytical models are often
restricted to a small validity domain. A numerical back-
scattering model based on an exact solution seems to be
the best way to study the contribution of the present
roughness description to backscattering simulations, espe-
cially because of the improvement in computation speed.
-12
-10
-8
-6
-4
-2
0
2
-12 -10 -8 -6simulations with m
sim
ulta
tions
with
par
amet
eris
ed a
lpha
(dB
)
RMSE (26˚)=0.55 dBRMSE (42˚)=0.7 dB
Fig. 8. Comparisons between backscattering simulations with measured roughness
a parameter from Eq. (1), for two incidence angles 268 and 428.
A moment method based on the integral equations
resolution is applied over the simulated surfaces for HH
polarisation and a dielectric case (Chen et al., 1989;
Harrington, 1968).
The soil surface is characterised by a dielectric
constant directly linked to surface soil moisture, computed
using the empirical approach developed by Hallikainen
et al. (1985).
Different test simulations have shown that a numerical
resolution with a sampling cell size of k/10 is sufficient for
the convergence of results. This is in agreement with other
numerical studies (Chen et al., 1989; Fung & Chen, 1985).
In order to simulate the radar signal, the backscattering
fields are computed over 100 profiles. Profiles are taken to
be sufficiently long (25 wavelengths) to consider rough
surfaces as infinite. Finally, a statistical average is
-4 -2 0 2easured alpha (dB)
incidence= 26˚incidence= 42˚
parameters (s, l and a) and simulations with measured s and l and retrieved