You are on page 1of 62

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/241192587

Radar Remote Sensing of Soil Moisture

Article

CITATIONS READS

3 50

5 authors, including:

Paolo Pampaloni Emanuele Santi


Italian National Research Council Italian National Research Council
204 PUBLICATIONS 2,924 CITATIONS 147 PUBLICATIONS 1,189 CITATIONS

SEE PROFILE SEE PROFILE

S. Paloscia Simone Pettinato


Italian National Research Council Italian National Research Council
233 PUBLICATIONS 2,835 CITATIONS 111 PUBLICATIONS 752 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

LEIMON: Land Monitoring with Navigation Signals View project

HYDRO-COSMO: The retrieval and monitoring of Land Hydrological parameters for Risk and Water
Resources View project

All content following this page was uploaded by S. Paloscia on 21 February 2014.

The user has requested enhancement of the downloaded file.


Radar Remote Sensing of Soil Moisture
Deliverable No. 8, D4-WP2

Envisnow EVG1-CT-2001-00052

Paolo Pampaloni, Emanuele Santi,


Simonetta Paloscia, Simone Pettinato, Pietro Poggi

30 October 2004
EnviSnow

Document Title
EnviSnow – SMC algorithms

Author(s)
Paolo Pampaloni, Emanuele Santi, Simonetta Paloscia,Simone Pettinato, Pietro Poggi (IFAC-CNR)

Project Project no
EnviSnow EVG1-CT-2001-00052
ISSN Document No. Document Type Status/Availability
ISBN
8, D4-WP2 Report, Methodology Restricted
Date Version No. of Pages
October 30, 2004 1.0 61
Keywords
Soil, SAR, Envisat, ASAR, ERS, Soil moisture

Notices

Distribution
Restricted to a group specifies by the consortium (including the Commission Services)

Reviewed by Technical Responsible Quality manager


Paolo Pampaloni (IFAC)

Abstract/Summary:
The state of the art of radar backscattering in monitoring soil moisture has been reviewed. It has
been shown that the best observing parameters for this application are: L-band, HH pol, θ =25°
incidence angle. However, this configuration is not available in present satellite sensors, and the
study has been addressed to evaluating the potential of ENVISAT ASAR data. Several retrieval
algorithms based on empirical and statistical approaches have been implemented and their
performance compared in terms of accuracy, robustness and computing time. It has been shown
that, by using a properly trained Neural Network, the retrieval of soil moisture on bare and
vegetated soils is possible at C-band too. The algorithm has been tested on a flat agricultural site
(Scrivia watershed) and in a mountainous area (Cordevole watershed). On the first site, five
levels of soil moisture between 10-15% and 45% have been identified, with a mean error of the
order of 10%. On the mountainous watershed, soil was covered by a thick grass layer and, since
only VV polarized ASAR data were available, a correction for vegetation effects has been
necessary. In the latter case, at least four levels of soil moisture between 25% and 45% could be
identified, with a mean error of the order of 10%.

Deliverable No. 8, D4-WP2 2


EnviSnow

TABLE OF CONTENTS
1 INTRODUCTION .......................................................................................................................................... 4
2 BACKSCATTERING SENSITIVITY TO SOIL MOISTURE AND SURFACE ROUGHNESS........... 5
2.1 Sensitivity of SAR backscattering to soil parameters: experimental data ............................5
2.1.1 Sensitivity to soil moisture...........................................................................................5
2.1.2 Sensitivity to surface roughness...................................................................................9
2.2 Theoretical models..............................................................................................................10
2.2.1 Model validation with experimental data...................................................................11
2.3 Semi-empirical models .......................................................................................................16
2.3.1 OSU Model (1992).....................................................................................................16
2.3.2 Dubois, van Zyl, Engman model ..............................................................................16
2.3.3 OSU II model (2002) .................................................................................................17
2.4 OSU 2002 Model validation ...............................................................................................18
2.5 Sensitivity analysis..............................................................................................................19
2.5.1 Sensitivity to Ks .........................................................................................................19
2.5.2 Sensitivity to SMC .....................................................................................................20
2.6 Backscattering from frozen soil ..........................................................................................24
2.7 Effects of vegetation ...........................................................................................................25
3 RETRIEVAL ALGORITHMS.................................................................................................................... 32
3.1 Retrieval algorithms based on semi-empirical models ......................................................32
3.1.1 OSU Model ................................................................................................................32
3.1.2 Dubois Model (Dubois et al. 1995)............................................................................33
3.1.3 Shi Algorithm (Shi et al. 1997).................................................................................34
3.2 General inversion algorithms ..............................................................................................35
3.2.1 Bayesian approach .....................................................................................................35
3.2.2 Artificial Neural Network approach ..........................................................................38
3.2.3 Nelder-Mead minimization method ...........................................................................38
4 COMPARISON OF ALGORITHMS ......................................................................................................... 40
4.1 Retrieval tests using simulated data ....................................................................................40
4.2 Retrieval tests using measured data ....................................................................................42
4.3 General remarks on algorithms performance......................................................................43
5 THE ENVISNOW SOIL MOISTURE ALGORITHM ............................................................................. 44
6 VALIDATION OF THE ALGORITHM WITH EO DATA..................................................................... 46
6.1 SMC retrieval in flat areas ..................................................................................................47
6.2 SMC retrieval in mountain areas ........................................................................................50
6.2.1 The test area and the ground measurements ..............................................................50
6.2.2 Preliminary analysis...................................................................................................51
6.2.3 Retrieval of soil moisture: supervised approach ........................................................52
6.2.4 Correction for vegetation effects and unsupervised retrieval of soil moisture ..........53
7 CONCLUDING REMARKS ....................................................................................................................... 57
REFERENCES ............................................................................................................................................................... 58

Deliverable No. 8, D4-WP2 3


EnviSnow

1 INTRODUCTION

This report describes activities of and results from Work Package 2 – Task 3 Development of
algorithms for retrieval of bare soil parameters of the EnviSnow project.
The main objective of WP2 - Task 3 was the development of improved algorithms for retrieval of
moisture of bare soils from satellite-borne synthetic aperture radar (SAR) data. The activities
aimed at preparing for the use of ENVISAT ASAR for retrieving soil moisture of bare soils.
However, the developed methods are applicable also to other SAR sensors.
In addition to the original plan, the developed algorithm cover the case of vegetated surfaces too,
also in the case of non-optimal observation configuration.
Soil moisture (SMC) plays a critical role in the surface energy balance at the soil-atmosphere
interface and can be considered a key state variable that influences the redistribution of the radiant
energy and the runoff generation and percolation of water in soil. We know that local SMC
measurements are strongly affected by spatial variability, besides being time-consuming and
expensive. Moreover, the use of hydrological models for extending the forecast of soil moisture
over larger areas is not easy, and depends on the homogeneity of the selected areas and on the
information available on them (soil properties, i.e. hydraulic characteristics, and permeability,
together with meteorological and climatological data, etc.). The possibility of measuring soil
moisture on a large scale from satellites, with complete and frequent coverage of the Earth’s surface
is, therefore, extremely attractive.

The sensitivity to soil moisture of the backscattering coefficient measured at low microwave
frequencies is a well-known phenomenon already investigated by many scientists. Indeed, research
activities carried out worldwide in the past years have demonstrated that sensors operating in the
low frequency portion of the microwave spectrum (P- to L-band) are able to measure the moisture
of a soil layer, the depth of which depends on soil characteristics and moisture profile (e.g. Ulaby et
al. 1986, Dubois et al. 1995, Dobson et al. 1995). The most significant information was obtained by
combining different frequencies, polarizations, and incidence angles (Dobson et al. 1995, Baronti et
al. 1995, Ferrazzoli et al. 1997). Unfortunately, L-band is not available from current satellite
sensors, which moreover operate in a single-frequency band. Thus, research for the retrieval of soil
moisture has been focused on the potential of existing sensor characteristics, i.e. the C-band, which
is operational on ERS-2, RADARSAT, ENVISAT. C-band backscattering is still sensitive to SMC,
but it is significantly influenced by vegetation (especially at VV polarization) and surface roughness
too, so that the estimation of spatial variations of moisture with the accuracy requested in many
applications is still problematic, even using correcting procedures. However, much better results
can be obtained if one is interested in measuring temporal variations on a relatively wide area where
the average characteristics of roughness and vegetation cover remain almost unaltered.

The purpose of this task was to develop an algorithm to retrieve spatial and temporal variation of
soil moisture and surface roughness of bare soils from ENVISAT ASAR data. To do this, it is
necessary a solid knowledge of the interaction mechanism between microwaves and land surfaces.
Thus, we first investigated the sensitivity of backscattering coefficient to the two quantities of
interest by using experimental data collected at the same frequency polarizations and incidence
angles provided by ASAR, as well as simulations performed with the most advanced physical and
semi-empirical models. The second phase of this work included the analysis the most advanced
inversion approaches and the implementation of a robust retrieval algorithm. A review of active
microwave remote sensing of soil moisture has already been delivered (Paloscia et al 2002). This
report summarizes again the most significant aspects of this research, with a focus on the activities
performed in the framework of this contract.

Deliverable No. 8, D4-WP2 4


EnviSnow

2 BACKSCATTERING SENSITIVITY TO SOIL MOISTURE AND


SURFACE ROUGHNESS

2.1 Sensitivity of SAR backscattering to soil parameters: experimental data

For the study of the interaction mechanisms between electromagnetic (e.m.) waves and soil, the
latter can be considered as a dielectric medium bounded by a rough surface. On a perfectly smooth
surface scattering of e.m waves is totally forward and depends on permittivity of the medium. On a
rough surface radiation is scattered in various directions and also generates backscattering. Thus,
two basic properties determine the backscatter response observed by the SAR system: the
permittivity of the medium and the roughness characteristics of the surface. Both parameters are in
turn related to different geophysical parameters of the soil. The soil permittivity is a function of the
sensor frequency (or wavelength) and depends on moisture, texture, bulk density, temperature, and
salinity of soil. The scattering pattern from rough surfaces depends primarily on surface roughness,
whereas the strength of the return is modulated by the dielectric properties determined, in this case,
by soil moisture. For natural surfaces, however, the influence of soil roughness is not yet fully
determined in a quantitative way.

An important aspect in the development and validation of retrieval algorithms is linking soil
moisture and roughness, to the requirements and needs of the various applications. In some cases
there are direct correlations between the parameters that can be retrieved using SAR data and those
needed in applications, in other cases the correlation is indirect and ancillary information is needed
to transform the SAR-sensed information into the information type needed by the application. It has
been found that C-band ENVISAT type data are sensitive to moisture in the upper 5 cm of the soil
medium. This result well matches the requirements of those process such as infiltration and
evapotranspiration that take place within these first few centimeters of the soil medium. In other
applications, where soil moisture profiles down to a depth of 1 meter are necessary SAR data must
be coupled to appropriate hydrological models.

The scattering theories used for interpreting radar backscatter from soil surfaces require to specify
the surface characteristics with one or more roughness parameter(s). The surface roughness is
generally characterized by some kind of height or height distribution in the vertical direction, and
by some kind of length or distribution of lengths in the horizontal direction. In general, for a
random surface, we assume that the surface height distribution is Gaussian or exponential, and the
most important vertical property for modeling the backscattering coefficient is the height standard
deviation (indicated with s, σ or HStD). The horizontal properties are usually described throughout
an autocorrelation function (ACF), power spectral density function or a structure function, which
are all related to each other.

2.1.1 Sensitivity to soil moisture

Several experiments carried out in the past years that pointed out a reasonable sensitivity of the
backscattering coefficient σo to soil moisture content (SMC) have been reviewed in Paloscia et al
(2002).

Deliverable No. 8, D4-WP2 5


EnviSnow

Waiting for the availabilty of ENVISAT-ASAR data, an in-depth analysis of the capabilities of C-
band data in measuring SMC was carried out, using all data at C-band collected with the ERS-2,
AIRSAR and SIR-C/XSAR over Montespertoli, Italy and Les Alpilles in Southern France. The
analysis of data collected on Montespertoli pointed out a correlation between σo at θ=23° and SMC
of bare soils characterized by moderate roughness (surface height standard deviation, Hstd < 3cm).
This correlation, although lower than the one obtained at L-band, was still remarkable at HH and
HV polarizations (σoHH = 0.57SMC-14.7 and σoHV 0.62SMC-25.2, both with a r2 = 0.46). Worse
results were obtained in VV polarization: σoVV 0.29SMC-11.7, with a r2 = 0.22. The same scarce
results were obtained on Les Alpilles with ERS-2 data for sunflower fields (r2=0.43) and even
worse for wheat fields, where C-band did not seem to be able to detect soil moisture variations
under the vegetation cover at all (Macelloni et al. 2002a, Macelloni et al 2002b).

At higher incidence angles (35-45°), the sensitivity to moisture can be completely cancelled, since
the effects of roughness and vegetation are stronger than at low incidence angles and destroy the
one due to moisture.

Figure 2.1 (a,b,c) shows σo as a function of SMC data classified in two classes of roughness: Hstd
≤1.5 cm, or >1.5 cm. From these diagrams we can observe that the backscattering is influenced by
both soil moisture and roughness and that the effects of these variables show a high synergy. In
fact, points characterized by high values of SMC also have high values of roughness (they
correspond to ploughed fields in the autumn season) and both quantities contribute to an increase of
the backscattering coefficient. On the contrary, points showing low values of roughness also have
low values of Hstd and therefore low values of backscattering too. The regression equations in this
case are:
2 2
Hstd ≤ 1.5 cm R Hstd > 1.5 cm R
σ°HH=0.46 SMC-13.0 0.35 σ°HH=0.51 SMC-12.8 0.33
σ°HV=0.48 SMC-22.6 0.27 σ°HV=0.61 SMC-25.0 0.39
σ°VV=0.35 SMC-12.9 0.15 σ°VV=0.16 SMC-9.15 0.04

We can note that, also in this case, the best correlations are obtained in HH and HV polarizations
and the poorest at VV pol.

Deliverable No. 8, D4-WP2 6


EnviSnow

SIR-C/MAC (Theta=20-25°)

5
(a)
0

1-1.5 cm

CHH
-5 >1.5 cm

-10

-15
0 10 20 30 40
SMC%

SIR-C/MAC (Theta=20-25°)

-5

(b)
-10
CHV

-15

-20

-25
0 10 20 30 40
SMC%

SIR-C/MAC (Theta=20-25°)

5
(c)
0
CVV

-5

-10

-15
0 10 20 30 40
SMC%

Figure 2.1: σ° at C-band (θ=25°), measured by means of AIRSAR sensor in 1991, SIR-C/X-SAR in 1994 and ERS in
1994 on bare or scarce vegetated soils on the agricultural area of Montespertoli at HH (a), HV (b) and VV (c)
polarizations, separated in classes of roughness: 1.5 cm < Hstd > 1.5 cm

Deliverable No. 8, D4-WP2 7


EnviSnow

0 0.5

σov 0.4
-5

soil moisture
σ °VV (dB)
0.3
-10
SM
0.2

-15
0.1

-20 0.0
23/1
23/2
23/3
23/4
23/5
23/6
23/7
23/8
23/9
23/10
23/11
Time

Figure 2.2: Temporal trends of ERS σ°VV and SMC for a single field on Les Alpilles area.

These results pointed out that the dependence of backscattering on the surface features of individual
fields limits the sensitivity of radar to surface moisture and the reliability of the measurement.
However, the temporal trends of σoVV, measured with ERS-2 on a single field of sunflowers on the
agricultural test site of Les Alpilles, was found to be in good agreement with the one of SMC (Fig.
2.2). Also the time variation of backscattering coefficient and SMC, both spatially averaged on the
whole area of Montespertoli, and represented in Fig. 2.3, show a good agreement between the two
variables. In these two cases the correlation coefficients were r2= 0.57 for Les Alpilles, where a
considerable spreading of data still existed, and r2=0.99 for Montespertoli, with a cumulative value
r2=0.62 for both areas (Macelloni et al. 2002b). It should be noted that, each point in the diagram
represents the average value of many pixels, so that the large number of pixels contained in the
averaged area compensated for the shortage of data used for the regression. Thus, we can state that
the correlation of C-band backscattering to spatial variations of soil moisture in non-homogeneous
areas is still questionable due to spurious effects introduced by surface roughness and vegetation
cover. However, temporal variation of SMC of an area containing various fields can be detected

Figure 2.3: Temporal trends of spatial-averaged σ°VV (ERS and SIR-C data) and SMC for Montespertoli area.

Deliverable No. 8, D4-WP2 8


EnviSnow

with a significant accuracy (Dobson et al., 1992, Beaudoin et al. 1990, Macelloni et al. 1999).
The regression equations and the correlation coefficients for the two groups of measurements are:

Test-site Regression equation R2


Les Alpilles σ°=17.31x SMC-11.3 0.57
Montespertoli σ°=38.76 x SMC-13.2 0.99
All measurements σ°=17.55 x SMC-10.9 0.62

2.1.2 Sensitivity to surface roughness

An example of radar sensitivity to surface roughness is given in Fig. 2.4 where σ°VV at C–band,
collected during both the MAC-Europe and SIR-C/X-SAR campaigns, is represented as a function
of Hstd (cm) in the case of bare or lightly vegetated fields (i.e. PWC, plant water
content, < 0.5 kg/m2).

C-BAND (VV)

0
y = 0,4012x - 8,6922
-2
R2 = 0,3269
Sigma-0 (dB)

-4

-6

-8

-10

-12
0 2 4 6 8
Hstd (cm)

Figure 2.4: The backscattering coefficient as a function of the surface Height Standard Deviation.

In Fig. 2.4, the spread of data was most likely due to different values of soil moisture which
changed with time. If we are interested in estimating surface roughness, we can smooth the effects
of the SMC variations in each field by a temporal averaging. In Fig. 2.5 σ° averaged in time for the
same fields was represented as a function of Hstd. We can note that the spread of points is sensibly
reduced and, if experimental points are separated in two classes of roughness: Hstd < 2 cm, and >
2cm, the correlation coefficient is rather high (0.54), at least for points with Hstd < 2cm.

Deliverable No. 8, D4-WP2 9


EnviSnow

-2

σ ° HH (dB)
-4

-6

-8

-10
0 1 2 3 4 5

HStD (cm)

Figure 2.5: Time-averaged C- band σ°HH of a single field as a function of surface Height Standard Deviation.
The correlation coefficient for all measurements is r2 = 0.37. By disregarding measurements with HStD >
2cm is r2 = 0.54.

The effect of roughness is stronger at high incidence angles than at low incidence angles, and at VV
polarization. This observation lead to the hypothesis that the ratio between the backscattering at HH
and VV polarization could be a good indicator of the surface roughness, at least for the range of
Hstd < 2cm.

2.2 Theoretical models

In spite of their complexity, only theoretical models can yield a significant understanding of the
interaction between the electromagnetic waves and the earth's surface. However, the exact solution
of equations governing rough surface scattering is not yet available and several approximate
methods have been developed with different ranges of validity. The asymptotic models usually
taken into consideration are: Small Perturbation (SP), Kirchhoff Physical Optics (PO), Kirchhoff
Geometrical Optics (GO) (Tsang et al. 1985, Ulaby et al. 1982), Integral Equation (IEM) (Fung
1994). The limits of validity of these approaches have been frequently discussed in literature, and
although an exact criterion has not yet been established we can assume the limits represented in
Macelloni et al. (2002).

The Kirchhoff approximation (KA), which is valid when the radius of curvature at every point on
the surface is much larger than the wavelength, assumes that the fields at any point on the surface
are equal to the fields that would be present on an infinite tangent plane at that point. When the
wavelength is much smaller than the surface height standard deviation (GO), we can assume that
scattering occurs only along directions for which there are specular points on the surface excluding
local diffraction effects. In this approach the coherent component vanishes and the cross-polarized
backscattering coefficient is zero since the Fresnel coefficients are evaluated at normal incidence. A
second order Kirchhoff approximation, which extends the validity range of KA to very rough
surface was developed by Ishimaru and Chen (1990). The Small Perturbation approximation
requires small surface rms height and slope with respects to the wavelength. The scattered fields can
be solved by using the Extended Boundary Condition method with the perturbation method (Tsang
et al. 1985). A similar approach is followed in the phase perturbation methods (Winebrenner and
Ishimaru 1985), which extends the SPM validity region to higher values of the HStD, provided that
the slope remains relatively small. The IEM (Fung 1994) is based on an approximate solution of a
Deliverable No. 8, D4-WP2 10
EnviSnow

pair of integral equations governing the surface scattering. It has a wider validity domain than KA
and SP and approximates SP at low frequencies and GO at high frequencies. In a successive work ,
Fung et al (1996) improved the model for estimating multiple scattering terms, which are essential
to simulate the cross-polarised backscattering coefficient.

2.2.1 Model validation with experimental data

Validation on artificial surfaces

The comparison of theoretical simulations with data collected on natural terrains suffers from
certain significant limitations. Firstly, the theoretical autocorrelation functions used in the models
only partially fit the real statistics of natural terrains. Secondly, experimental data usually available
from remote sensors mostly concern the backscattering coefficient measured within a limited range
of incidence angles, whereas a basic parameter for evaluating the model performances is the bistatic
scattering coefficient, measured at various incidence and scattering angles on the whole upper half-
space. A research aiming at validating electromagnetic models was carried out at the European
Microwave Signature Laboratory (EMSL) of the Joint Research Centre of Ispra, Italy, by measuring
scattering and emission from experimental models realized with the same surface and dielectric
characteristics used in the theoretical simulations (Macelloni et al. 2000). It should be noted that,
here too, the comparison of scattering coefficients with experimental data could only be carried out
in actual fact on a limited numbers of points.

Figure 2.6: “Smooth” surface (σ = 0.4 cm, L = 6 cm): simulated and measured bistatic
scattering coefficient as a function of frequency, at scattering angle θs = 0: (a) incidence angle
θi = - 20° and (b) incidence angle θi = - 40°. (After Macelloni et al. 2000)

Deliverable No. 8, D4-WP2 11


EnviSnow

Three different experimental models, composed by an artificial homogeneous dielectric shaped with
a suitable mould, were realized. All the models had azimuthal isotropic random surfaces with
Gaussian height distribution, two of them (smooth and rough) were characterized by a Gaussian
autocorrelation function with σ = 0.4 cm and 2.5 cm respectively and correlation lengths = 6 cm
for both, while a third model, with intermediate roughness degree, had been realized with a
composite (Gaussian-exponential) correlation function with σ = 0.9 cm and correlation lengths =
1.8 and 3.0. The polarimetric scattering measurements were carried out on the sample placed in the
center of the EMSL anechoic chamber (radius 10 m) at various incidence and scattering angles
between 10 and 50 degrees. The absolute calibration accuracy was better than +/- 0.5 dB and the
system noise level for typical measurement settings (averaging factor =128, time window span = 10
ns) ranges from -50 dBsm at 2 GHz to -40 dBsm at 18 GHz.

The comparison between the measurements and the model calculations showed that even when the
statistics of the observed surface is well determined and known, some discrepancy can exist
between models and measurements. However if the limits of validity are respected, the
disagreement is, in general, quite small and may be insignificant for practical applications. It was
confirmed that the Integral Equation Model (IEM) has a wider range of applicability with respect to
other classical approximations (Small Perturbation, Geometrical and Physical Optics). For smooth
surfaces (0.1 < ks <1.5, 1.5 <kl < 22.6) (k is the electromagnetic wavenumber = 2π/λ, s= σ is the
height standard deviation, and l the correlation length) (Fig. 2.6) the Small Perturbation reproduced
experimental data quite well up to 10 GHz (ks = 0.83, kL = 12.6 at incidence angle θi = 20 degrees,
and up to 6 GHz (ks = 0.5. kl = 7.5) at θi = 40 degrees. Physical Optics (PO) and IEM gave
similar results at least up to 11 GHz for θi = 20 degrees, and up to 9 GHz for θi = 20 degrees. But,
unlike what we expected, the results at 5 GHz gave better agreement than those at 10 GHz. For each
incidence angle, there was a frequency limit (10 GHz at θi = 20 degrees, 8 GHz at θi = 30 degrees)
above which, as the frequency increased, the measured backscattering tended to a constant value,
whereas model prediction indicated a decrease. For “medium rough” surfaces (0.23 < ks< 3.4) (Fig.
2.7) a good agreement was obtained between experimental and theoretical data at θi = 40 degrees
by changing, in the model, the local angle θl according to an empirical function of θi given in
Macelloni et al. (2000). However, at a smaller incidence angle (θi = 20 degrees) the accordance was
less convincing, since the model underestimated the measurements, especially at the higher
frequencies. Finally for “rough surfaces” (0.64 < ks < 9.4) both GO and IEM (in the form of “large
scale roughness”) well reproduced experimental data on the whole tested frequency range (Fig. 2.8).

Figure 2.7: Mixed Gaussian-exponential “medium rough” surface (σ = 0.9, l = 1.8 cm, L=3.0 cm). The
simulated and measured backscattering coefficient as a function of frequency at θi = 40°. The model has been
run with three different approximations for the local incidence angleθ1 (dashed: θ1 = θi, continuous: θ1 = 0 ,
dot-dashed = empirical function of θi given in Macelloni et al. (2000).
(a) HH polarization and (b) VV polarization. (After Macelloni et al. 2000).
Deliverable No. 8, D4-WP2 12
EnviSnow

Figure 2.8: “Rough” surface (σ = 2.5, L = 6 cm). The simulated and measured backscattering
coefficient as a function of frequency at (a) θi = 20° and (b) 40°.
(After Macelloni et al 2000)

Also the modified multiple scattering IEM was successfully validated by EMSL data.

Validation of IEM on natural terrains

Due to the wider range of validity of IEM with respect to other models, validation with
experimental data collected on natural terrains was performed for this model only. In the model the
dielectric constant of soil was simulated by using a polynomial fit (Hallikainen et al 1985). We
assumed a soil with a mean value of texture (70% sand and 30% silt+clay), moisture SMC=15%
and an exponential autocorrelation function with correlation length lc ranging between 4 and 10 cm.
Fig. 2.9 shows that the model (continuous lines) reproduces quite well the data up to ks = 2 where
multiple scattering effects can be neglected while at C-band it shows a maximum not present in the
experimental data, probably due to the multiple roughness scales which are usually present in
ploughed soils. It should be noted that, whereas at L-band σo gradually increases with ks, at C-band
the saturation confirms that the same surface can appear rougher at C- than at L-band. A value of ks
= 0.28 in the L-band data range can be considered a threshold value for separating smooth from
rough soils, in accordance with the Fraunhofer criterium applied for an incidence angle of 45°. The
model was implemented for simulating the backscattering coefficient both at L and C bands as a
function of SMC for different values of Hstd, by using an exponential autocorrelation function and
a correlation length, Lc = 6 cm, a soil density, d = 1.2 g/cm3, and a soil composition of sand (20%)
and clay (30%). In Fig. 2.10 the backscattering coefficient at L-band at HH pol. is shown as a
function of SMC for different values of Hstd (from 1 cm to 3 cm) for two incidence angles: a) 25°
and b) 35°. Experimental points were subdivided in two classes: Hstd < 2cm and > 2cm. Although
there is a saturation at Hstd value of 2cm, the model is able to follow fairly well the experimental

Deliverable No. 8, D4-WP2 13


EnviSnow

data , at least up to a Hstd value of 2 cm. Experimental points with Hstd <2cm are between the

Figure 2.9: σ° (HH pol. θ = 45o ) as a function of the Normalized Height Standard
Deviation ks = HStD x 2π /λ. Continuous lines represent IEM model run at L- and
C-bands with an exponential autocorrelation function and correlation length
ranging between 4 and 10 cm. (SMC=15%)

model lines of 1 and 1.5 cm and those with Hstd > 2cm are all above the line of 1.5 cm.

The same model was applied to C-band data too, and the results are shown in the diagrams of
Fig. 2.11, where the backscattering in HH pol. is shown vs. SMC for two incidence angles a) 25°
and b) 35°. In this case the effect of saturation is even more marked than at L-band and, as soon as
the Hstd exceeds 1 cm, the model shows a saturation. The criteria used for classifying a surface can
be taken as follows

Smooth Hstd < λ /(25 cos θ)


Rough Hstd > λ/(4 cos θ)
Intermediate λ /(25 cos θ) < Hstd < λ/(4 cos θ)

This means that, at C-band, a surface with a Hstd > 1.5 cm (at θ=40°) should be considered
“rough”.

Deliverable No. 8, D4-WP2 14


EnviSnow

L HH (Θ =25°) ->SMC L HH (Θ=35°) -> SM C

0 0
Hstd=1 cm

-5 -5 Hstd=1.5 cm
Hstd=2 cm
-10
σ° (dB)

-10 Hstd=3 cm
Hstd<2cm
-15 -15
Hstd>2cm

-20 -20

-25 -25
0 10 20 30 0 10 20 30

SMCg% SMCg%

Figure 2.10: σ° at L-band (a: θ=25°, b: θ=35°) as a function of gravimetric SMC (%). Lines represent the IEM
model computed for different values of Hstd (1 cm, 1.5 cm, 2 cm and 3 cm). Points represent the experimental data
subdivided in two roughness classes

4 0
Hstd 0.2
2 (a) -2 (b)
Hstd 0.4
0 -4
Hstd 0.6
-2 Hstd 0.8 -6
C HH

-4 Hstd 1.0 -8
-6 Hstd 1.2 -10
-8 1-1.5 -12
>1.5
-10 -14
<1
-12 -16
0 10 20 30 40 0 10 20 30 40
SMC%

Figure 2.11: σ°HH at C-band (a: θ=25°, b: θ=35°) as a function of gravimetric SMC (%).
Lines represent the IEM model computed for different values of Hstd (0.2 cm, 0.4 cm, 0.6 cm, 0.8 cm,
1.0 cm, and 1.2). Points represent the experimental data subdivided in three roughness classes

Deliverable No. 8, D4-WP2 15


EnviSnow

2.3 Semi-empirical models


Semi-empirical models rely on physical considerations but contain fitting parameters which can be
derived from a large amount of experimental data. They are, in general, simple to handle and can be
easily inverted. However, their relevance is questionable, since they account for the surface
roughness through only one parameter, i.e. the Height Standard Deviation, while it is known that a
rough surface cannot be satisfactorily characterised by a single parameter. Several semi-empirical
models are available to predict scattering from rough soil. Some of them have the potential of
estimating soil parameters by directly inverting backscattering measurements (Shi et al. 1992, Oh et
al. 1992, Dubois et al. 1995, Deroin et al. 1997, Oh et al. 2002).

2.3.1 OSU Model (1992)

The semi-empirical model (OSU) suggested by Oh et al. (1992) relates the two backscattering
parameters p = σ ohh σ ovv and q = σ ohv σ vv
o
to surface roughness, described by ks, and to soil
reflectivity, which in turn depends on the volumetric soil moisture content mv . The relevant
relations given in Oh et al. (1992) are
1 / 3 Γo
σ ° hh  2θ 
p= = 1−   e − ks (2.3.1.1)
σ ° vv π 

where θ is the incidence angle (in radians), and

σ ° hv
q= = 0.23 Γo (1 − e − ks )
σ ° vv (2.3.1.2)

where Γ0 is the Fresnel reflectivity at nadir.

2.3.2 Dubois, van Zyl, Engman model

An empirical model, which can be inverted to retrieve soil moisture from the co-polar
backscattering only was developed by Dubois et al (1995). The model is based on the following
equations:

cos1.5 ϑ 0.028ε tan ϑ


σ hh0 = 10 − 2.75 10 (ks sin ϑ )1.4 λ0.7 (2.3.2.1)
sin 5 ϑ
cos 3 ϑ 0.046ε tan ϑ
σ vv0 = 10 − 2.35 10 (ks sin ϑ )1.1 λ0.7 (2.3.2.2)
sin 3 ϑ

Deliverable No. 8, D4-WP2 16


EnviSnow

2.3.3 OSU II model (2002)

In a recent paper, Oh et al. (2002) presented an improved version of their semi-empirical model.
Based on existing scattering models and data sets measured by polarimetric scatterometers and the
JPL AirSAR, the parameters of the co-polarized phase-difference probability density function,
namely the degree of correlation α and the co-polarized phase-difference ζ, in addition to the
backscattering coefficients σhh, σvv, and σhv, were modeled empirically in terms of the volumetric
soil moisture content and the surface roughness parameters

VH-polarized backscattering coefficient

The simulated cross-polar backscattering coefficient is represented by the following equation:

σ vho = am vb (cos θ ) c [1 − exp[ − d ( ks ) e ] (2.3.3.1)

agrees very well with the measurements, especially with regard to its dependence on θ .

The magnitudes of constants a, b, c, d and e were determined through data fitting, using the
database, by applying the minimum mean square error (MMSE) technique. The process led to the
following values a = 0.11, b = 0.7, c = 2.2, d = 0.32, and e = 1.8.

Co-polarized backscattering coefficient

Based on the analysis of data set the selected functional forms of the cross-pol. and co-pol. ratios,
defined as in the OSU model, were the following:

c
σ o
s 
q = vh
= a  + sin b θ  {1 − exp[ − d ( ks ) e ]} (2.3.3.2)
σ o
vv  l 

amv − b
σo  θ 
p ≡ hho = 1 − 
σ vv

 90° 
[
exp − c(ks )
d
] (2.3.3.3)

where a, b, c, d and e are constants to be determined by applying the MMSE fitting to the data
base.

Consequently the co-polar backscattering coefficients have been expressed as follows

σ 0 . 11 m v0 . 7 cos 2 . 2 θ {1 − exp[ − 0 . 32 ( ks ) 1 . 8 ]}
σ o
vv = vh
= (2.3.3.4)
q 0 . 10 [ ls + sin( 1 . 3θ ]1 . 2 {1 − exp[ − 0 . 9 ( ks ) 0 . 8 ]}

p o
σ hho = pσ vvo = σ vh .
q (2.3.3.5)

Deliverable No. 8, D4-WP2 17


EnviSnow

Degree of correlation and co-polarized phase difference

Following similar approaches the degree of correlation and the co-polarized phase difference have
been modelled as follows:

−e
α = 1 − (a + bkl + cmv )(sin θ ) d ( ks ) (2.3.3.6)

s
ς = ( a + bm v − c )θ (2.3.3.7)
l

2.4 OSU 2002 Model validation

Figs. 2.12 (a, HV pol., and b, VV pol.) show a comparison between OSU model and the IEM for an
incidence angle of 45o and a soil moisture content mv = 0.13.The figures also include data measured
with scatterometers and SAR. For the IEM computation, an approximate backscattering model for
surfaces with small and moderate roughness with exponential ACF was used.

Figure 2.12: Measurements of σ°HV (a) and σ°VV as a function of normalized Height Standard
Deviation compared with model simulations (IEM and OSU 2002). (After Oh et al. 2002)

We see that OSU model agrees well with experimental data across the entire range of ks and agrees
with IEM in restricted ranges.

Deliverable No. 8, D4-WP2 18


EnviSnow

2.5 Sensitivity analysis


Once a model has been validated with experimental data, it is useful to investigate the potential of
SAR backscattering in measuring the desired quantities through a sensitivity analysis. Some
examples of results obtained with the Integral Equation Models for the co-polar terms and with the
semi-empirical OSU model for the cross-polar are shown in the following sections.

2.5.1 Sensitivity to Ks

In Figs. 2.13 the backscattering coefficient at C-band (HH,VV and HV polarizations) is represented
as a function of ks for different values of SMC, and two kl values (5 cm and 15 cm).

C- band - 23 degs - KL = 5

0
Backscattering Coefficient (dB)

-5 (a)
-10

-15

-20

-25

-30
0 0.5 1 1.5 2 2.5
KS

VV - smc = 5 HH - smc =5 VV - smc=10 HH - smc =10 VV - smc = 15 HH - smc=15


VV smc=25 HH smc =25 HV - smc=5 HV - smc=10 HV - smc=15 HV - smc=25

C- band - 23 degs - KL = 15

0
Backscattering Coefficient (dB)

-5 (b)
-10

-15

-20

-25

-30
0 0.5 1 1.5 2 2.5
KS

VV - smc = 5 HH - smc =5 VV - smc=10 HH - smc =10 VV - smc = 15 HH - smc=15


VV smc=25 HH smc =25 HV - smc =5 HV - smc =10 HV - smc =15 HV - smc =25

Figure 2.13: σ° at C-band (HH,VV and HV polarizations, θ=23°) as a function of ks for different values
of SMC (5, 10, 15, and 25%) and for two values of kl (a: 5cm, b:15 cm).

Deliverable No. 8, D4-WP2 19


EnviSnow

In general we can note that :


• σhv increases with ks, when kl and/or SMC increase.
• σhh and σvv increase with ks for all SMC values and for higher value of kl (kl = 15), whilst
for kl = 5 (or lower), σhh and σvv increase with ks, for ks ≤ 1 and decrease for ks ≥ 1.

2.5.2 Sensitivity to SMC

The backscattering coefficient at HH, VV and HV polarizations is represented as a function of SMC


for different ks values in Fig. 2.14, at an incidence angle of 35°, and in Fig. 2.15 for an incidence
angle of 23°.

C- band - 35 degs - KL = 5

0
Backscattering Coefficient (dB)

-5
(a)
-10

-15

-20

-25

-30
0 5 10 15 20 25 30 35 40
SMC (%)

VV - ks = 0.2 HH - ks =0.2 VV - ks=0.5 HH - ks =0.5


VV - ks = 1 HH - ks=1 VV ks=1.5 HH ks =1.5

C- band - 35 degs - KL = 5

-5
Backscattering Coefficient (dB)

-10 (b)
-15

-20

-25

-30
0 5 10 15 20 25 30 35 40
SMC (%)

HV - ks =0.2 HV - ks =0.5 HV - ks=1 HV ks =1.5

Figure 2.14: The backscattering coefficient at C-band is represented as function of SMC (%) for
different ks values, at an incidence angle of 35°. a) HH and VV polarizations; b) HV polarization.

Deliverable No. 8, D4-WP2 20


EnviSnow

C- band - 23 degs - KL = 5
Backscattering Coefficient (dB)

0
(a)
-5

-10

-15

-20
0 5 10 15 20 25 30 35 40
SMC (%)

VV - ks = 0.2 HH - ks =0.2 VV - ks=0.5 HH - ks =0.5


VV - ks = 1 HH - ks=1 VV ks=1.5 HH ks =1.5

C- band - 23 degs - KL = 5

0
Backscattering Coefficient (dB)

-5 (b)
-10

-15

-20

-25

-30
0 5 10 15 20 25 30 35 40
SMC (%)

HV - ks =0.2 HV - ks =0.5 HV - ks=1 HV ks =1.5

Figure 2.15: The same as in the previous diagrams, but at an incidence angle of 23°.

Deliverable No. 8, D4-WP2 21


EnviSnow

We can observe that, for each polarization, angle, and ks value, σ° increases when the soil moisture
increases. Similar results were obtained for different values of kl. From previous figures, we can
observe that, for soil moisture higher than 10%, it is reasonable to interpolate the backscattering
coefficient with a linear trend.

Figs. 2.16 (a and b) represent the sensitivity of backscattering coefficient to SMC (slope of the
regression line between σ° and the SMC) versus ks for different kl values at θ = 35°, and VV (a)
and HH (b) polarizations

Theta = 35 degs - VV

0.108

0.107

0.106 kl = 5
SMC -slope

0.105 kl = 8
kl = 10
0.104 kl =12
0.103 kl = 14

0.102

0.101
0 0.5 1 1.5 2 2.5
ks

Figure 2.16a: The backscattering coefficient sensitivity to SMC (represented by the slope of the regression line
between σ° and the SMC) versus ks for different kl values at VV polarization and θ = 35°.

Theta = 35 degs - HH

0.0730

0.0725 kl = 5
SMC -slope

kl = 8
kl = 10
kl =12
0.0720
kl = 14

0.0715
0 0.5 1 1.5 2 2.5
ks

Figure 2.16b: The backscattering coefficient sensitivity to SMC (represented by the slope of the
regression line between σ° and the SMC) versus ks for different kl values at HH polarization and θ =
35°.

Deliverable No. 8, D4-WP2 22


EnviSnow

From these diagrams we can see that on bare soils the effect of soil roughness is more important at
VV than at HH polarization.

The comparison between the simulated backscattering coefficients at L and C bands for both linear
and cross polarizations shows that, whilst the σ° level is higher at C band than at L-band, the
sensitivity to soil moisture remains almost the same at both frequencies (Fig. 2.17). However, it
should be noted that this behaviour is strongly altered by the presence of vegetation, which affects
much more C-band backscattering.

C- L band - 23 degs - l = 5 cm - s = 1cm

0
Backscattering Coefficient (dB)

-5
(a)
-10
Chv
-15
Lhv
-20

-25

-30
0 5 10 15 20 25 30 35 40
SMC (%)

C- L band - 23 degs - l = 5 cm - s = 1cm

0
(b)
Backscattering Coefficient (dB)

-2
-4
Cvv
-6
Chh
-8
Lvv
-10
Lhh
-12
-14
-16
0 5 10 15 20 25 30 35 40
SMC (%)

Figure 2.17: A comparison between the simulated trends of σ° at L and C bands (at θ=23°) as a
function of SMC (%). a) HV polarization, b) HH and VV polarizations.

Deliverable No. 8, D4-WP2 23


EnviSnow

2.6 Backscattering from frozen soil


We investigated the effects on the soil temperature on the backscattering variations by using data
obtained from the Joint Research Center in Ispra, Italy. These data consist of a set of backscattering
coefficients measured over a soil sample placed in an anechoic chamber (Experiment Series
NVT30-32)”. The experiment aimed at investigating the polarimetric radar response of frozen soil
as function of soil temperature during the transition to thawed condition. The sample under test was
contained in a cylinder of 2 m in diameter and 0.4 m in depth, placed in the center of the
measurement chamber (radius 10 m). The target was rotated in steps of 5° about the z-axis to obtain
independent scatterometric data samples. The soil used in the test was sandy loam with about 10%
of clay, 30% of loam and 60% of sand. Two moisture levels were considered, corresponding to the
experiments labeled NVT31 and NVT32. The vertical temperature profile was measured at different
depths (z = -10, -18 , -33 , -58 and -108 mm) by using calibrated thermocouples (accuracy: about
0.1°C , sensitivity: 0.1°C ). The soil sample, thermally isolated by the container's walls, was
refrigerated from top to down by using the metallic mould as heat sink (kept at -25°C by a suitable
refrigeration circuit embedded in the metal) until a minimum temperature of -10°C was reached in
all the sample's volume. The mould was subsequently removed from the surface in order to perform
the radar measurements. The backscattering measurements were performed using a frequency-
stepped CW mode (step = 11.25 MHz), in the range 1.0 - 10 GHz, acquiring full polarimetric data.
Data were calibrated using the standard monostatic full polarimetric procedure. The calibration
accuracy was better than 0.5 dB at 0 dBsm. The system noise level ranged fom about -50 dBsm at 1
GHz to -40 dBsm at 10 GHz. As an example, Fig. 2.18 shows backscattering coefficients (VV
polarization) as a function of the temperature of the upper part of the soil sample at 5.3 GHz and the
temporal trends of the temperatures measured by the probes at different depths (right).

4
-2
2
VV 10 mm
18 mm
5.3 GHz 0
33 mm
-3
23 °
Temperature [°C]

-2 58 mm

-4
-4
σ0dB

-6

-8
-5
-10

-12
-6
-14

-16

-7 -18
-16 -14 -12 -10 -8 -6 -4 -2 0 2 0 20 40 60 80 100 120 140 160

Temperature [C] Time [hours]

Figure 2.18: Backscattering coefficients (VV polarization) as a function of the temperature of the upper part
of the soil sample at 5.3 GHz and the temporal trends of the temperatures measured by the probes at
different depths (right). (Data Courtesy: JRC – ISPRA).

Deliverable No. 8, D4-WP2 24


EnviSnow

The relationship between backscattering coefficients at HH and VV polarizations and soil


temperature was approximated by:

σ o (T ,θ i , λ , d ) = A(θ , λ , d ) + B (θ , λ , d )T ( d ) + C (θ , λ , d )T 2 ( d ) + D (θ , λ , d )T 3 ( d ) (3.6.1)

Where T(d) is the temperature at depth d in the soil, and the coefficients A, B C and D, computed
from JRC experimental data are given in Table 2.1.

θ Depth Freq. A B C D R2

θΙ =23 °

D=6 cm
5.3 GHz 0.548 6.14 10-2 3.64·10-3 6·10-5 0.95
10 GHz 0.14 1.61·10-2 9.75·10-4 1.76·10-5 0.96
θΙ =45 °
5.3 GHz 0.082 1.08·10-2 7.15·10-4 1.37·10-5 0.97
10 GHz 0.096 1.36·10-2 7.59 10-4 1.08·10-7 0.96
θΙ =23 °
D=0 cm
5.3 GHz 0.487 5.98 10-2 3.96·10-3 5.82·10-5 0.98
10 GHz 0.1225 1.34·10-2 4.84·10-4 -1.45·10-5 0.96
θΙ =45 °
5.3 GHz 0.123 1.44·10-2 1.1·10-3 2.5·10-5 0.97
10 GHz 0.072 1.21·10-2 1.2·10-3 3.8·10-5 0.96

Table 2.1: Coefficients computed from experimental data for the formula in Eq. (3.6.1).

2.7 Effects of vegetation

For soils covered with vegetation the total measured backscattering is a combination of
contributions from soil and the vegetation layer. The most important terms are (Fig. 2.19):

Deliverable No. 8, D4-WP2 25


EnviSnow

1. direct backscattering from soil attenuated by vegetation,


2. direct backscattering from the vegetation layer (leaves and stems),
3. soil- vegetation and vegetation-soil interaction (double scattering),
4. soil-vegetation- soil interaction

Figure 2.19: The four main scattering mechanisms in a


vegetation covered soil.

In this case the backscattering can be modelled by using the radiative transfer theory, which can
take into consideration the effects of total vegetation biomass as well as plant shape and size. For a
canopy layer represented as an ensemble of sparsely distributed elements (randomly oriented disks,
simulating leaves, and nearly vertical cylinders simulating stalks and stems) with permittivity εv
embedded in a medium with permittivity εo (air) upon a homogeneous dielectric half space of
permittivity εs and with a rough surface, the backscattering coefficients σppo and σpqo for co- and
cross- polarized terms (p, q = vertical V, horizontal H polarization) are expressed as follows (Tsang
et al. 1981):

− 2 Kep d Secθ P pps − 2 Kep d Secθ


σ pp
o
= σspp e + 2 Kep (1 − e )

+ 2 rsp Pppm d Secθ e − 2 Kep d Secθ

rsp2 Ppps e− 2 Kep d Secθ − 2 Kep d Secθ


+ 2Kep (1 − e )

Deliverable No. 8, D4-WP2 26


EnviSnow

P pqs
σ pq
o
= σspq e − (Kep +Keq) d Secθ
+ Kep+Keq (1 − e − ( Kep +Keq) d Secθ )

rsp Ppqm e − Kep d Secθ − Kep d Secθ − Keq d Secθ


+ (e −e )
-Kep + Keq

rsq Ppqm e − Keq d Secθ − Kep d Secθ − Keq d Secθ


+ (e −e )
-Kep + Keq

rsp rsq Ppqs e−( Kep + Keq) d Secθ


+ Kep + Keq (1 − e −( Kep + Keq) d Secθ )

where :
σspp, σspq = bare soil backscattering at co- and cross-polarization
rsp = soil reflectivity
kep = extinction coefficient of vegetation (m-1)
d = height of vegetation layer (m)
Ppqs, Ppqm = scattering matrix elements per unit volume of disks and cylinders for pq polarization,
computed for direct and double scattering, respectively, within the reference frame.

It should be noted that, since discs and cylinders are mixed within a unique layer, kep, Ppqs and Ppqm
are weighted averages, with weights given by the densities of the two kinds of scatterers.

We can assume that, for the considered composite medium, the permittivity tensors at all points are
symmetric and then reciprocity holds. Therefore, the cross polarized terms σVH, and σHV are equal
(Tsang et al. 1985).The scattering amplitudes and the extinction cross sections of circular and
elliptic dielectric disks were computed by using the model proposed by Karam in (Karam 1998).
The extinction cross sections were obtained from the scattering amplitude tensor elements by
applying the forward scattering theorem (eq. 37 in Karam 1998). Since the crop stems considered
were much longer and much thinner than the electromagnetic wavelength, scattering by a finite
length cylinder was computed by using the infinite cylinder approximation, i.e. by assuming that
the cylinder responds to an incoming wave as though it were infinite in length (Tsang et al. 1992 ).
The extinction cross sections were computed from the forward scattering theorem also in this case
(Tsang et al 1981).

This model approach was validated by comparing model simulations with experimental data
collected on crops with a similar leaf area index, but very different geometrical characteristics:
Sunflowers, characterized by tall stems and by almost circular leaves, Wheat, which has thin stems
and elliptical leaves (Macelloni et al 2001).

Deliverable No. 8, D4-WP2 27


EnviSnow

The model made it possible separating soil contribution from the total backscattering and
computing the extinction coefficient ke (m-1 ) of a vegetation layer. As an example Fig. 2.20 shows,
as a reference, the backscattering coefficient as a function of SMC of a medium rough bare soil,
while Figs. 2.21 and 2.22 shows the cases of the same soil covered with well developed crops of
wheat and sunflowers.

Bare Soil

-5
σ ° (dB)

-10
HH
VV

-15

-20
0 10 20 30 40 50

SMC%

Figure 2.20: Modeled backscattering coefficient (C-band, θ = 25o) as a


function of soil moisture (SMC) for a soil with HStD = 0.8 cm , and
correlation length = 8 cm.

Deliverable No. 8, D4-WP2 28


EnviSnow

Wheat: d=200, H=80cm

-10
HHSoil

σ °(dB) -20 VVSoil


HH tot
-30 VV Tot

-40
0 10 20 30 40 50

SMC%

Figure 2.21: Modeled backscattering coefficient (C-band, θ = 25o, HH and VV pol) as a


function of soil moisture (SMC) for the soil of Fig. 2.20 covered with wheat with density of
200 plants/m2 and plant height = 80 cm ; HH soil and VV soil represent the contribution of
soil attenuated by vegetation and separated by the total backscattering HH tot and VV tot .

Sunflower: d=6, H=1m

-5
HHSoil
σ ° (dB)

-10 VVSoil
HH tot

-15 VV Tot

-20
0 10 20 30 40 50

SMC%

Figure 2.22: Modeled backscattering coefficient (C-band, θ = 25o, HH and VV pol) as a


function of soil moisture (SMC) for the soil of Fig. 2.20 covered with sunflowers with density
of 6 plants/m2 and plant height = 100 cm ; HH soil and VV soil represent the contribution of
soil attenuated by vegetation and separated by the total backscattering HH tot and VV tot.

Deliverable No. 8, D4-WP2 29


EnviSnow

The case of a natural terrain covered by grass is represented in Figs. 2.23 and 2.24 for two level of
total biomass.

Grass: d=200, H=30cm

-10
σ ° (dB)

HHSoil
VVSoil
-20 HH Tot
VV Tot

-30
0 10 20 30 40 50
SMC%

Figure 2.23: Modeled backscattering coefficient (C-band, θ = 25o, HH and VV pol) as a function of soil
moisture (SMC) for the soil of Fig. 2.20 covered with grass with density of 200 blades/m2 and height = 30 cm;
HH soil and VV soil represent the contribution of soil attenuated by vegetation and separated by the total
backscattering HH tot and VV tot.

Grass: d=200, H=60cm

-10
σ ° (dB)

HHSoil
-20
VVSoil
HH tot
-30
VV Tot

-40
0 10 20 30 40 50

SMC%

Figure 2.24: Modeled backscattering coefficient (C-band, θ = 25o, HH and VV pol) as a function of soil
moisture (SMC) for the soil of Fig. 2.20 covered with grass with density of 200 blades/m2 and height = 60 cm;
HH soil and VV soil represent the contribution of soil attenuated by vegetation and separated by the total
backscattering HH tot and VV tot.

Deliverable No. 8, D4-WP2 30


EnviSnow

All these diagrams show that, in general, the presence of vegetation changes significantly the level
of backscattering especially at VV polarization. However, the sensitivity to soil moisture is not
cancelled. This analysis suggested the possibility of retrieving soil moisture under vegetated
canopies. Indeed, provided some information on crop biomass and type is available, the effect of
vegetation can be estimated and its extinction effect compensated.

The extinction of vegetation components (leaves and stalks of sunflower and wheat), is represented
in Fig. 2.25, which represent the extinction coefficient as a function of biomass represented by the
leaf area index (LAI in m2/m2). We note that, in general, the extinction :
- is higher at the highest frequency,
- changes with LAI for cylinders at vertical polarization, and is almost independent of LAI at
horizontal polarization.
- is almost independent of LAI for disks, except in the early stage.
- is higher for cylinders at VV polarization and for disks at HH polarization except in the early
growth stage of sunflower.

In the case of wheat, where there is a high number of stalks per square meter, the extinction of
cylinders at vertical polarization shows a steep increase in the very early stage of growth followed
by a decrease due to a reduction of the number of plant elements per unit volume during the
growing phase of crop. For sunflowers, the extinction of stalks at VV polarization continuously
increases with LAI at both frequencies and polarizations, whilst the extinction of leaves shows a
maximum. The extinction of grass can be estimated by assuming the blades similar to thin
cylinders as for wheat.

Figure 2.25: Extinction coefficient of leaves (dotted line) and stalks


(continuous line) for sunflowers (left) and wheat (right).

Deliverable No. 8, D4-WP2 31


EnviSnow

3 RETRIEVAL ALGORITHMS

The complexity of the algorithms to be developed for the retrieving of soil moisture and/or soil
roughness depends on the auxiliary information available, and on the selected direct models (i.e. the
number of equations describing the radar backscatter at different frequencies and/or or
polarisations). For soil moisture retrieval, different schemes have been studied, according to the
possibility to reduce the unknown parameters to just the soil moisture, and to provide information
on the disturbing factors by using ancillary data.
The retrieval of soil moisture from single frequency/polarization data such as those from ERS-1/2
SAR requires some additional information on surface characteristics (roughness and vegetation
cover). The latter can be obtained from a-priori knowledge such as agricultural practices or from
other remote sensing data. For example, optical data can be used first to classify the scene into bare
soil, soil with sparse vegetation, and soil with dense vegetation. The inversion algorithm can be
subsequently applied on those pixels, which correspond to bare soil and soil with sparse vegetation.

Due to different sensitivity of observation parameters (frequency, polarization, incidence angle) to


surface roughness and vegetation cover, the use of multi-frequency/polarization data is potentially a
good tool for retrieving soil moisture without a-priori knowledge on surface characteristics. Also
the interferometric coherence, which is more sensitive to changes in soil roughness and vegetation
than in moisture, can be of significant interest.

3.1 Retrieval algorithms based on semi-empirical models


The inversion of analytical EM models is in general a complicated procedure. For this reason
several approaches based on semi-empirical models have been developed.

3.1.1 OSU Model

According to Oh et al (1992) eqs. (2.3.1.1) and (2.3.1.2) provide a good fit to data for 0.1 < ks < 6
and 2.6 < klc < 19.7 (lc is the correlation length of the surface roughness) and allow the retrieval of
both ks and soil reflectivity through an iterative procedure Γo. A comparison of q measured at L–
and C–bands on bare and lightly vegetated fields with its values computed from eq.(2.3.1.1)
indicates that, although experimental data are rather scattered, the trend is broadly consistent with
the model. This observation supported the attempt to retrieve ks from the ratios
o o o o
σ hv σ vv and σ hh σ vv measured at L–band. Fig. 3.1a compares ks retrieved from θ = 35° data
with ground truth roughness measurements of single fields on Montespertoli agricultural area in
Tuscany, Italy. The OSU model was then used to estimate Γ0 and, in turn, the SMC of single fields
by using a soil permittivity model (Dobson et al. 1985, Hallikainen et al. 1985),. The retrieved
values of SMC are compared with ground truth in Fig. 3.1b. The obtained result appears rather
satisfactory (r2 = 0.55), in spite of some dispersion, probably mainly due to parameter p, whose
small values tend to amplify the fluctuations of Γ0.

Deliverable No. 8, D4-WP2 32


EnviSnow

Figure 3.1: a) ks of bare or lightly vegetated (LAI<1, PWC<0.5 kg m-2) fields retrieved by OSU model from C-
band (stars) and L-band (squares) data versus ks obtained from ground measurements.
b) SMCg (%) of bare or lightly vegetated (LAI<1, PWC<0.5 kg m-2) fields retrieved by successive use of OSU and
Dobson et al. models from L-band data versus SMCg (%) measured at ground.

The updated version of the OSU model agrees with experimental observations over a wide range of
soil surface conditions, and also agrees with the IEM and geometrical optics model over their
individual regions of validity, thereby encompassing the full range of surface roughness
encountered under natural conditions. However, the surface roughness spectrum (and the
correlation length) is not accounted for in the direct forward model, so that the validity of the
model for a wider range of experimental data remains questionable. Indeed, in a former study,
simulations to evaluate the sensitivity of σ°HH and σ°vv to the surface roughness spectrum have
shown that variations as large as 10 dB are created by various assumptions on the surface
correlation function shape (Souyris et al. 1994).

3.1.2 Dubois Model (Dubois et al. 1995)

Unlike the OSU model, Dubois model does not require the knowledge of cross-polarized term. As
for the OSU model, roughness is characterized only by the rms height s, consequently, the same
questions hold. The algorithm has been optimized for bare surfaces and requires two co-polarized
channels at a frequency between 1.5 and 11 GHz. It has been successfully tested on several data sets
collected by AIRSAR and SIR-C over the Little Wishita Watershed in Southwest Oklahoma. Best
results are obtained for kh = 2.5, a soil moisture content = 35%, and an incidence angle θ = 30
degrees. An example of retrieval results is given in Fig. 3.2.

Deliverable No. 8, D4-WP2 33


EnviSnow

Figure 3.2: Comparison of retrieved and ground measured data (After Dubois et al. 1997)

3.1.3 Shi Algorithm (Shi et al. 1997)

Shi et al. (1997) developed an algorithm based on a fit of the single-scattering Integral Equation
Method (IEM). The algorithm provides estimation of soil moisture and surface roughness parameter
(a combination of rms roughness height and surface power spectrum) from quad-polarized synthetic
aperture radar (SAR) measurements. The algorithm was applied to a series of measurements
acquired at L-band (1.25 GHz) from both AIRSAR (Airborne Synthetic Aperture Radar operated by
the Jet Propulsion Laboratory) and SIR-C (Spaceborne Imaging Radar-C) over a well managed
watershed in southwest Oklahoma. It was found that the two co-polarized backscattering
coefficients and their combinations would provide the best input to the algorithm for estimating soil
moisture and roughness parameters. Application of the inversion algorithm to the co-polarized
measurements of both AIRSAR and SIR-C resulted in estimated values of soil moisture and
roughness parameter for bare and short-vegetated fields that compared favorably with those
sampled on the ground. The root-mean-square (rms) errors of the comparison were found to be
3.4% and 1.9 dB for soil moisture and surface roughness parameter, respectively. An example of
the retrieval results is given in Fig. 3.3.

Figure 3.3: Comparison between the inferred and ground measured soil moisture and surface
roughness parameter from all available SIR-C (+) and AIRSAR(*) L-band images.

Deliverable No. 8, D4-WP2 34


EnviSnow

3.2 General inversion algorithms

Once the forward models have been validated, more general inversion algorithms can be developed
based on several techniques. In particular these include Neural Networks and Bayesian approaches.

The possibility to invert soil moisture using polarimetric features has been investigated at MIT, in
order to find the necessary information (multi-frequency and multi polarization) needed for a
correct soil moisture reconstruction. Theoretical simulations (Wang et al. 1993) showed that a
Neural Network can be used to retrieve soil moisture provided the unknown parameters are within
the training range. This study was limited to the validity domain of SPM. In a further study (Souyris
et al. 1994), using IEM and SPM as forward models, the soil moisture retrieval obtained with
neural nets directly driven by single frequency co-polarized backscattering coefficients σ°hh and
σ°vv, was found not optimal. The neural net was therefore trained with several combinations of
multi-frequency (P band and L band) or polarimetric data. In both cases, the introduction of the
ratio between the two copol backscattering coefficients greatly improved the inversion accuracy.
The estimated uncertainty on the retrieved soil moisture was found to be 10% when one frequency,
co-polarized data are used in the training process, and dropped to 3% when both co- and
cross-polarized data were considered. Finally, when the neural net was trained with multifrequency
(P band and L band) co-polarized data, the maximum uncertainty on the retrieved soil moisture was
1.5%. The question of how neural nets work with real measured data and how many frequencies
and polarimetric data are needed to retrieve the soil moisture has still to be investigated, in a more
general approach.

To another extent, an inversion algorithm based on Kalman approach was tested with satisfactory
results on data acquired during the AVIOSAR 580 Campaign on the Matera. test site (De Carolis et
al. 1994). In this study Geometrical Optics was used as direct forward model, but the Kalman
approach is flexible enough to accept other models.

Alternative classification techniques based on a stochastic formulation of the inversion problem


have been reported in the literature, such as the Bayesian approach (Hunt, 1977).

3.2.1 Bayesian approach

In general, the retrieval of soil and vegetation parameters from radar measurements is a typical ill
posed problem, due to the non linearity of the relationships between remote sensing measurements
and ground parameters and because, in general, more than one combination of soil parameters (soil
moisture, roughness etc.) has the same electromagnetic response.

Multi-sensor techniques are able to discriminate the different contributions of the soil features to the
global system response and in fact, by using different frequencies, polarization and incidence angle,
it is possible to improve the extraction of information and the classification accuracy.

However, the availability of multi-frequency and multi-angle sensors in the satellite and airborne
platforms actually operating is limited, reducing the application of the inversion algorithms.
Therefore, in order to minimize the uncertainties and enhance the performance of the retrieval of
soil parameters from remote sensing data, statistical approaches have been introduced instead of the
deterministic inversion of the theoretical relationships.

Deliverable No. 8, D4-WP2 35


EnviSnow

In this context, the Bayesian inference technique plays a fundamental role: this technique, based on
the Bayes theorem, updates the likelihood of an event, given a previous likelihood estimate and
additional evidence. In general, supposing H1, H2,....,Hn exhaustive and mutually exclusive
hypotheses (as an example: values of soil moisture, roughness, vegetation biomass etc.) that give to
a certain evidence E which has occurred (as an example: a radar measure), the Bayesian theorem
can be expressed as:

P( E | H j ) P( H j )
P( H j | E ) = (3.2.1.1)
∑ P( E | H
j
j ) P( H j )

where:
I. ∑ P( H
j
j ) =1

II. P( H j | E ) is the posterior probability, given the evidence E, that Hj is true;


III. P( H j ) is the prior and unconditional probability that is true;
IV. P( E | H j ) is the probability of observing E, given Hj.

In fact, assuming to have different sensors measuring the same unknown object, by using the
observation of the i-sensor it is possible to derive the i-hypotesis about the identity of the observed
object, as an example by combining measurements and dedicated algorithms for the retrieval. The
reliability of these estimates is given by the ‘previous established’ uncertainty characteristics,
expressed as probability that the sensor would declare the object Oj, given that Oj is true: P(D1 | Oj).
Combining all the sensors declaration by using the Bayes formula, it is possible to recover the joint
probability for each possible entity Oj of the m total entities, based on the declarations of the n
observing sensors:

P(Oj | D1, D2, ...,Dn) j=1,...,m

Consequently, the association between observed object and sensor measurements can be obtained
by choosing the object whose joint probability P(Oj | D1, D2, ...,Dn) is the greatest. This criterion is
called maximum a posteriori probability (MAP).

Due to the complexity of relationships that link the soil parameters to remote sensing
measurements, several problems can be found in computing the prior likelihood P(Hj ) of the
observed parameters and in respecting the requirements of exaustivity and mutual exclusivity of the
Hj. Usually, when no prior information exists on the relative likelihood of Hj, the principle of
indifference is used, in which P(Hj) are set equal for all j. A typical way to compute the requested ‘a
priori’ information is to take in account theoretical values of the remote sensing parameters
corresponding to a certain combination of soil parameters and compare them with the
corresponding measures: the spread between theoretical results and measurements gives the density
functions which are incorporated in the inversion algorithm in terms of probability information.
Unfortunately such approach requires the computation of density error functions and in fact can be
applied only to large datasets.

Application to the SMC retrieval


The bayesian approach can be applied to the retrieval of soil parameters from single- or multi-
frequency and single- or multi-polarization radar measurements. For example, considering single
frequency, dual polarization data, the conditional density function P(ε,s,lc|σoVV ,σoHH ), which

Deliverable No. 8, D4-WP2 36


EnviSnow

represents the probability to retrieve soil permittivity (ε), standard deviation of heights (s) and
correlation length (lc) once given the measured backscattering, can be expressed as:

Pprior (ε , s, lc) Ppost (σ VV


0
, σ HH
0
| ε , s, lc)
P(ε,s,lc |σ o
VV ,σ HH )=
o
P(σ VV
0
, σ HH
0
)

where
Pprior (ε , s, lc) includes the available ‘a priori’ information about these parameters derived
from direct measurements or estimated from other instruments. When ‘a priori’ information
is unavailable, the ‘principle of indifference’ can be applied, by associating an uniform
density function to each of these parameters.
Ppost (σ VV
0
, σ HH
0
| ε , s, lc) can be expressed more conveniently by considering the radar
measurements as theoretical values affected by a multiplicative noise, which is modelled by
means of two random variables independent of the soil parameters ε, s and lc (eq. 3.2.1.1):

σ oVVmeas = R1σ oVVtheo


(3.2.1.1)
σ o HHmeas = R2σ o HHtheo

With this assumption, Ppost (σ VV


0
, σ HH
0
| ε , s, lc) becomes:

1
Ppost (σ VV
0
, σ HH
0
| ε , s, lc) = P ( R1 , R2 )
σ 0
VV ⋅ σ HH
0

The joint probability density function of the errors P(R1,R2) is modelled by means of two gaussian
distribution functions, as follows:
2 2 2 2
e − ( R1 − µ1 ) / 2σ1 e − ( R2 − µ2 ) / 2σ 2
P ( R1 , R2 ) = ⋅
2π σ 1 2π σ 2

where mean (µ1,2) and standard deviation (σ1,2) can be derived from the statistics of experimental
data.

The “optimal estimator” of the dielectric constant of soil, can be obtained from the measured values
of backscattering in HH and VV polarization as follows (Haddad et al. 1996, Notarnicola and Posa
2001):

1
ε
∫∫∫εP prior (ε , s, lc)
σ 0
⋅ σ HH
0
P( R1 , R2 )dε ds dlc
ε= , s ,lc VV
(3.2.1.2)
P(σ VV
0
, σ HH
0
)

Lastly, the soil moisture (SMC) values can be obtained from ε by inverting the Dobson et al. model
(1985).

Deliverable No. 8, D4-WP2 37


EnviSnow

It should be noted that, due to the complexity of the implemented equation, which includes a triple
integration operation, the computing time required by the method is of about 1 second for each
point (pixel), making this release of the algorithm problematic for real time prediction, when a ‘per
pixel’ retrieval from wide SAR images is required.

3.2.2 Artificial Neural Network approach

The algorithm based on the Artificial Neural Network (ANN) uses a feed-forward multi-layer
perceptron (MLP), with some hidden layers of neurons between the input and output. In MLPs
successive layers of neurons are fully interconnected, with trainable connection weights controlling
the strength of the connections. The input to a neuron in a given layer is the sum of all its incoming
connection weights, each one multiplied by the output of the connected neuron from the preceding
layer. The trainable offset value associated with the neuron is added to the sum, and the result is fed
into the activation function of the neuron. The latter is most commonly chosen as a non-linear
sigmoid function, which is also the form we used in the present work.

MLP ANNs can be trained to represent arbitrary input-output relations (Hornik et al. 1989). During
the training phase, training patterns are sequentially presented to the network and the
interconnecting weights of each neuron are adjusted according to a learning algorithm. The trained
ANN can be considered as a type of non-linear least mean square interpolation formula for the
discrete set of data points in the training set.
The algorithm chosen for the training phase was the back-propagation (BP) learning rule, an
iterative gradient descent algorithm designed to minimize the mean square error between the
desired target vectors and the actual output vectors.
After several tests, a configuration with two hidden layers of 10 perceptrons each was chosen as the
optimal one. The used ANN was trained by using backscattering coefficients as input and SMC
expected values as output.

3.2.3 Nelder-Mead minimization method

The Nelder-Mead simplex algorithm (Nelder and Mead 1965) is a popular search method for
multidimensional unconstrained minimization. It is a direct search method of optimization that
works moderately well for stochastic problems and is commonly used in non-linear regression
problems. It is computationally quite simple and works with a number of rules. The required
starting vector of n elements (where n is the number of variables of the function under test) is used
to construct a “simplex”, which is a shape with n +1 vertices. A simplex, sometimes called a
hypertetrahedron (Walters et al., 1991), is the generalization of a tetrahedral region of space to n
dimensions.

The simplex is so-called because it represents the simplest possible polytope in any given space.The
following diagram show the graphs schematizing the vertices connections for the n-simplexes with
n = 2 to 7.

Deliverable No. 8, D4-WP2 38


EnviSnow

Function values are evaluated at the vertices of the simplex, that is iteratively shrunk, as better
points are found until some desired bound is obtained.

The Nelder-Mead method was applied to the retrieval of the soil moisture content from ENVISAT
backscattering images as follows: a dataset of backscattering values σomeas was extracted; for each
backscattering value, a corresponding theoretical “true” σotheo was calculated by using the IEM,
starting from estimated initial values of soil permittivity (ε ), correlation length (lc), and height
standard deviation (s), required as inputs.

In the algorithm, these inputs were varied until the minimum of an appropriate error function was
reached. For example, in the case of single-frequency single-polarization data the function to be
minimized was:

F(ε,lc,s) = [abs(σomeas - σotheo)]

Input vector corresponding to the minimum was assumed as “optimum” and from the ε value the
corresponding SMC was calculated as usual by means of the inversion of the Dobson model.

Deliverable No. 8, D4-WP2 39


EnviSnow

4 COMPARISON OF ALGORITHMS

4.1 Retrieval tests using simulated data


In order to provide a first comparison of the developed algorithms (NN, Bayes, Nelder Iteration,
linear regression, 2nd order regression) we created a synthetic data base by simulating 3000 values
of single-polarization backscattering σo generated by multiplying the outputs of the Integral
Equation Model IEM (run at Envisat observation parameters) for a noise random variable with
mean and variance derived from the experimental dataset. The selected soil surface parameters
were: Height Standard Deviation Hstd = 0.9 cm, exponential ACF with correlation length L= 8 cm;
the moisture range was between 5 and 35% (Fig. 4.1). The results of the comparison are shown in
Fig 4.2a where six classes of moisture are separated and the outputs of the various algorithms are
compared with the true SMC. We can see that in general all the approaches give acceptable results.
The error in each class is in general lower than 15% (Fig 4.2b) except for the linear regression for
the lower values of moisture. The mean error is of the order of 10% or lower, except for the linear
regression that is affected by the poor results at low moisture value. In evaluating these results it
should be noted that both linear and 2nd order regressions are affected by local conditions and
cannot be generalized.

Figure 4.1: synthetic data base of backscattering values vs. SMC obtained by IEM model.

Deliverable No. 8, D4-WP2 40


EnviSnow

40 (a)
35

retrieved SMC %
30 NN
25 Nelder
20 linear
15 2nd order
10 bayes
5
0
5 - 10 10 - 15 15 - 20 20 - 25 25 - 30 30 - 35
SMC classes (%)

45 (b)
40
35 NN
30 Nelder
error %

25
linear
20
15 2nd order
10 Bayes
5
0
5 - 10 10 - 15 15 - 20 20 - 25 25-30 30 -35
SMC classes (%)

Bayes
(c)
2nd order

linear

Nelder

NN

0 5 10 15
mean error%

Figure 4.2: (a) comparison between algorithm results; (b) error of the tested retrieval algorithms in
each SMC class; (c) mean SMC retrieval error for the tested algorithms.

Deliverable No. 8, D4-WP2 41


EnviSnow

4.2 Retrieval tests using measured data


In order to perform the evaluation in more realistic cases, a further comparison was made by using
experimental data derived from an ENVISAT ASAR image collected on a flat agricultural area in
the Po Valley (Scrivia watershed) in November 2003. Ground SMC measurements were carried out
in correspondence to the satellite overpass over a set of agricultural fields. From the available
measurements, 13 bare soil fields with comparable roughness have been extracted and used for the
test. Dataset needed to train the ANN was derived by the IFAC archive of SIR-C/X-SAR and
MAC-91 measurements acquired during various experiments carried out in several test sites. Data
collected at frequency and incidence angle close to ENVISAT ASAR over bare soils with
comparable features were extracted and used for training. After the ANN was trained, all the
inversion methods (Regression, ANN, Bayes and Nelder-Mead Iteration) were tested on the
selected fields.
The measured soil moisture ranged between 24 % and 35 %. This interval was divided in 3 classes
containing a comparable number of fields (4-5) and for each class, the mean values SMC were
compared with inversion results . The error of the tested algorithms in each class and the mean error
for all the three classes are represented in Fig. 4.3 (a and b). We can see that in the case the highest
accuracy was obtained by the Nelder iteration approach. However, also the algorithms based an NN
and Bayes theorem gave satisfactory results.

e rro r % fo r e a c h c la s s o f S MC
20
18
16
n e ld e r (a)
14 re g r
12 b a ye s
10
8 NN
6
4
2
0
24% - 27% 27% - 30% >30%
S MC c la s s e s

NN
(b)

bay es

regr

nelder

0 2 4 6 8 10 12 14
m ean error %

Figure: 4.3: (a) Error of the tested retrieval algorithms in each SMC class; (b)
mean SMC retrieval error for the tested algorithms.

Deliverable No. 8, D4-WP2 42


EnviSnow

4.3 General remarks on algorithms performance

A final comparison of the algorithms was carried out by considering other aspects, such as stability
and computation time. The results are summarized in Table 4.1, where the performances of each
algorithm have been synthetically classified with a number of asterisks.

Table 4.1: Performance comparison of the tested SMC retrieval algorithms


(for each aspect considered: * = lowest performance, **** = highest performance).

Algorithm Accuracy Stability Computational


time
Bayes (mean error ~11%) (needs only few “a priori” information) (0.1pixel/sec)

*** **** *
Neural Net (mean error ~9%) (depends strongly on the training set) (~1000pixel/sec
Once trained)
***
** ***
Iterative (mean error ~4%) (depends on the starting values) (~1pixel/sec)
(Nelder)
****
*** **

(mean error ~12%) (depends on the dataset) (~100.000pixel/sec)


Regression
**
** ****

We can conclude that the Iteration (Nelder-Mead) algorithm is the most accurate but slow, the
Bayes approach is the most stable but very slow, Regression is the fastest but less accurate, while
the algorithm based on the NN gives the best compromise. Therefore the selected algorithm for
SMC retrieval will be based on the NN approach.

Deliverable No. 8, D4-WP2 43


EnviSnow

5 THE ENVISNOW SOIL MOISTURE ALGORITHM

It should be noted that, according to the Description of Work, the algorithm should have regarded
the estimate of soil moisture of bare soils. However, since at the time of observations, most part of
the mountainous site was heavy vegetated, an effort has been made to include a correction for
vegetation in the algorithm.
The block diagram of the algorithm is shown in Fig. 5.1. The SAR image, either single look
complex (SLC) or multi-look detected, is first calibrated according to the procedure suggested by
the administrator of the sensor (ESA, RADARSAT, etc) and then geocoded by using a Digital
Elevation Model of the area and filtered to reduce the speckle. In case of SLC images, which are
delivered without cross-track radiometric corrections, the External Calibration Files available from
ESA are used to correct for the antenna elevation pattern. Looking for an operative application for
mountainous terrain, we have to consider that the physical size of the scattering area for each pixel
varies according to topography and this effect must be taken into account in the radiometric
calibration. Therefore in the calibration procedure the local incidence angle has to be used instead
of the nominal one, which results from the Earth ellipsoid model. The local incidence angle is
obtained by the use of a DEM and the knowledge of orbital and imaging parameters. In order to
correlate ground data and ASAR acquisition, calibrated data must be geocoded. The procedure is
realized through the coregistration of the calibrated ASAR image with a simulated SAR
backscattering image generated from the DEM using an empirical function relating backscattering
intensity to local incidence angle. Layover and shadow masks are also produced at this stage. In
order to reduce speckle effect, geocoded images have to be processed using gamma filter having a
3x3 pixel window. The image is then masked for the effect of shadow and layover (by using orbital
and DEM data), and classified by using a vegetation map obtained from ground data or optical
images. This classification makes it possible to exclude forest areas from the analysis and to take
into account the attenuation effect of grass. Auxiliary data useful to classify the image and to
correct for the effects of soil surface roughness and vegetation cover can be obtained from ground
measurement or from other remote sensing data. For example: arboreous vegetation can be
separated from SAR data acquired in VH polarization at 30-40 degrees incidence angle. A further
low pass filtering is then applied to the retrieved SMC maps and the achieved ground resolution is
of the order of 0.5 km.

Deliverable No. 8, D4-WP2 44


EnviSnow

SAR Data Orbital and


HH/VV/HV pol. SAR image calibration
calibration parameters

Geocoding
DEM

Speckle Simulated
filtering SAR image

Vegetation/ Classified Layover and


roughness data image shadow mask

Backscattering
extraction

RT,IEM Vegetation and


roughness
correction

Training Neural Local incidence


data network angles map

Experimental Pixel
aggregation SMC map
SMC data

Figure5.1. Block diagram of the soil moisture algorithm

Deliverable No. 8, D4-WP2 45


EnviSnow

6 VALIDATION OF THE ALGORITHM WITH EO DATA

As results from the previous analysis of experimental data and model simulations, the best SAR
parameters for soil moisture retrieval are L-band (or lower frequency), HH polarization and small
(i.e. < 30o ) incidence angles. Of these requirements the last one only has been matched in collecting
data to validate the algorithm. Indeed, only C- band data are available from the present in orbit
satellite, moreover, due to conflicts with other experimenters, and especially with commercial users,
who have absolute priority, only very few HH polarized images have been obtained from
ENVISAT ASAR C-band system.
Validation has been carried out on two sites, both in North Italy: the Scrivia watershed, a flat area
in the Po valley, and the Cordevole watershed, a mountainous area on the Dolomites (Fig. 6.1).

CORDEVOLE

SCRIVIA

Figure 6.1: The location of the two test sites used for validating the algorithms

Deliverable No. 8, D4-WP2 46


EnviSnow

6.1 SMC retrieval in flat areas


The first site selected for validating the algorithm was the watershed of Scrivia located in North-
west Italy. The area is a flat alluvial plain of about 300 Km2 situated close to the confluence of the
Scrivia river with the Po river. It is characterized by large, homogeneous, agricultural fields of
wheat, corn, sugarbeet, and some minor fodder crops (Fig. 6.2).

Castelnuovo
Scrivia

Figure 6.2. Aerial photo of the Scrivia site superimposed to the area of
maximum flooding extent

The area was monitored by the ENVISAT ASAR on November 7, 2003, April 30, 2004, and June
4, 2004. All the images were collected at HH polarization and θ = 23o incidence angle. Weather was
cloudy in November, rainy in April and sunny in June. At each date soil moisture was rather
uniform on the area and its average value increased from November to April and then decreased
again in June. Simultaneously to satellite passes over the area, local measurements were carried out
in a “training area” on some selected fields, chosen for their characteristics representative of the
whole area , dimensions and homogeneity. The measurements included: soil moisture (SMC) of the
first 5 cm layer, surface roughness (s) and vegetation biomass (PWC or total biomass).

Fig. 6.3 shows a multi-temporal composition of the SAR images collected in November 2003
(Red), April 2004(Green) and June 2004 (Blue). The image shows the capability of SAR
backscattering in pointing out the rivers and the structure of agricultural fields. Since the
backscattering increases with soil moisture, the dominant color is green, correspondent to the
wettest situation in April (heavy and persistent rainfalls). However, some fields in the northern part
of the area (red in the image) show a marked low value of backscattering. Indeed, in April their soil
was over-saturated with the formation of a film of water, which caused a scattering mainly forward.
So in the composite image the total backscattering from these fields is dominated by the
measurement in November (red). The enlargement on the right of the image represents the portion

Deliverable No. 8, D4-WP2 47


EnviSnow

30 Km

Test site

Vogh
Castelnuovo Scrivia

10 Km

Tanaro river

Figure. 6.3: Composition of SAR backscattering images (C-band, HH pol., θ=23°) collected on
November 2003 (Red), April 2004 (Green), and June 2004 (Blue)
of the area, close to Castelnuovo Scrivia, with the fields (marked by white polygons) where ground
measurements were carried out.

The comparison between retrieved and measured SMC on the test fields is shown in Fig. 6.4.

40

30
SMC% retrieved

20

10

0
0 10 20 30 40
SMC% measured

Figure 6.4: Comparison of measured and retrieved SMC on the


Scrivia area

Deliverable No. 8, D4-WP2 48


EnviSnow

The retrieval of soil moisture at the three dates was performed for the whole area by using the NN
trained with data collected on 13 fields in November 2003. The results of “per pixel” classification
are represented in Fig. 6.5. We can see that the average colours correspond to the average soil
moisture values measured on the area: average SMC was 25-30% in November, higher than 30-
35% in April, and of the order of 10% in June.

(a)

(b)

(c)

10 20 35 %
Figure 6.5: SMC maps of the flat test area of Scrivia for different dates: (a) 7/11/2003
(b) 29/04/2004 (c) 4/06/2004. (ENVISAT ASAR, HH pol., 23° incidence angle).

Deliverable No. 8, D4-WP2 49


EnviSnow

6.2 SMC retrieval in mountain areas

6.2.1 The test area and the ground measurements


After the testing of the algorithm over the flat agricultural area in Po Valley, a second validation
exercise was carried out on the mountain areas in the Cordevole watershed. This watershed, located
at the foothill of Mount Sella in Northern Italy (Fig. 6. 6) was selected because of its relatively
smooth topography and the availability of historical and topographic data. The area covers 95 km2
at a mean altitude of 1948 m asl with a mean slope inclination of 51 %. In the middle of the basin,
at the top of Mount Cherz (2000 m asl), there is a fairly large plateau free of forests. Spruce forests
cover about 30 km2, pasture 43 km2, and crops 6.31 km2. Only 0.71 km2 are occupied by urban
areas. The length of rivers is 188 km. The site was equipped with a corner reflector and with a
meteorological station.

During the ENVISAT passes, three ground campaigns were carried out on 14 June, 19 July, and 27
September 2004. The measurements included photographic surveys and measurements of soil
moisture (with a TDR probe) and fresh biomass of vegetation (gravimetric method) in different
parts of the site. In June and July a significant portion of the area was covered by a rather dense
grass, which in September was almost dry. The mean value of soil moisture was close to 40% in
June, to 45% in July and to 35% in September.
15 km

Cherz
plateau

14-06-2004 19-07-2004 27-09-2004

Figure 6.6 Orthophoto of the test site and images of Cherz plateau on different dates

Deliverable No. 8, D4-WP2 50


EnviSnow

6.2.2 Preliminary analysis


The selection of available Envisat images for the Cordevole was performed, looking for snow free
conditions and low incidence angles. Unfortunately, due to the well-known user conflicts for the
northern part of Italy, it was possible to acquire three images only in summer 2004, and all these
images were in VV polarization. This polarization was demonstrated to be less sensitive than the
HH to the moisture of soil, due to the masking effects of vegetation. For these reasons, a
preliminary sensitivity analysis was performed by comparing backscattering values and
corresponding SMC ground measurements, which were collected on a portion of the site (training
area) during the experimental campaigns. The obtained regression was then compared with
historical data collected on bare soils with comparable roughness at the same frequency,
polarization and incidence angle. These data were selected from the IFAC archive of SIR-C/X-
SAR, MAC-91 and ERS1-2 data acquired on different dates on various test sites. The obtained
regressions pointed out a strong attenuation effect due to the vegetation cover, although the
sensitivity to the SMC remained quite unchanged, as shown in Fig. 6.7.

0 20 40 60 80
10

0
σ C-band VV

archive bare
-5 soils
Cordevole
0

-10

-15

-20
SMC%

Figure. 6.7: C-band backscattering (VV pol, θ=23°) as a function of SMC for the two datasets:
historical acquisitions on bare soils (blue) and Envisat data on Cordevole test site (pink)

The obtained regression equations for the two datasets are:

σ0VV = 0.2631SMC% - 11.704 for the historical data


σ0VV = 0.2625SMC% - 21.605 for the Cordevole test site

The above equations clearly show that the dense vegetation present at the date of acquisitions
induced a strong attenuation in the measured backscattering but it did not affect the sensitivity to
SMC too much. Indeed, the slope of the two regressions is quite similar.

Deliverable No. 8, D4-WP2 51


EnviSnow

6.2.3 Retrieval of soil moisture: supervised approach

A first step of the retrieval procedure was to train the Neural Network by using the radar
backscattering (as input) and the corresponding SMC (as output) measured on 11 points of the
training area of Cordevole. The trained ANN was then tested by using the remaining part of the
available dataset composed by 82 samples obtained at different dates. The high SMC values
measured in the three campaigns, due to the rainy summer season, did not allow us to test the ANN
over a wider range of SMC, since the lower measured value was of close to 25%. The retrieval
performed for all data collected on the three dates of observations made is possible to separate 5
classes of soil moisture (Fig. 6.8). The first four classes of SMC were retrieved with a mean error
lower than 10%, and only for the higher values of moisture the error was over 20%, (Fig. 6.9). This
result is reasonable considering the relatively low sensitivity of backscattering coefficient to the
higher values of SMC.

60

50
S M C% m eas

40 A NN output
SMC%

30

20

10

0
25% - 30% 30% - 35% 35% - 40% 40% - 45 % >45%

S M C c las s es

Figure 6.8. Comparison of measured (blue) and retrieved (dark red) soil moisture on
Cordevole with the supervised approach

error %

>45%
40% - 45 %
35% - 40%
30% - 35%
25% - 30%

0 5 10 15 20 25

Figure 6.9. The mean error for each class of retrieved soil moisture in the supervised
approach

Deliverable No. 8, D4-WP2 52


EnviSnow

6.2.4 Correction for vegetation effects and unsupervised retrieval of soil moisture

The obtained result, although rather satisfactory, requires the availability of a considerable amount
of ground data collected on the area under study, needed to train the ANN with a robust dataset.
Such approach might have some limitations for an operational use, especially for areas not easy
accessible and in case of an extensive monitoring in both space and time.

For heavy vegetated soils, as in the case of Cordevole, the attenuation introduced by vegetation
cover leads to a different backscattering level, and requires a correction for this attenuation,
especially for VV polarized data. The contribution of vegetation cover can be simulated by using a
discrete elements model based on the radiative transfer theory as reported in section 2.7 (Macelloni
et al., 2001).

According to the ground measurements of vegetation carried out on Cordevole, the backscattering
coefficient was computed for a vegetation covered soil with a plant a density of 200 blades per
square meter and a mean height of 30 cm. The obtained simulated data were comparable to the
measured ones. The extinction coefficient ke of the vegetation layer was derived from the model
outputs and used to correct for the vegetation effects, assuming that the dominant effect of
vegetation is attenuation. The complete data set showed a good continuity between the archive bare
soil data and data collected on Cordevole and corrected for vegetation effects (Fig 6.10).

0 10 20 30 40 50 60 70 80
8
6
4 archive data
2 for bare soils

0
sigma VV

-2
-4 Cordevole dataset
-6 corrected for
vegetation
-8
-10
-12
-14
SMC%

Fig. 6.10: Backscattering coefficient (VV pol, θ=23°) as a function of SMC for bare soil (archive) data
and the dataset of Cordevole corrected for vegetation.

The obtained regression for the whole dataset is log type, which well describes the loss of
sensitivity of the radar signal to the higher SMC values. The equation and correlation coefficient are
listed below:

Deliverable No. 8, D4-WP2 53


EnviSnow

σ0= 6.9029*Ln(SMC) - 26.326

R2 = 0.7837

On the basis of these results, the ANN was finally trained with the archive including all data and
tested using the backscattering coefficients corrected for vegetation effects. The comparison
between retrieved and measured SMC showed only a small decrease in the retrieval accuracy
respect to the supervised case (Fig. 6.11). The advantage of this approach is that it does not require
the collection of data on the site, but only some auxiliary information on the vegetation cover. The
comparison between mean errors computed for each class by using the two approaches ANN1 and
ANN2 is shown in Fig. 6.12

60

50 SMC% meas
ANN 1
40 ANN2
SMC%

30

20

10

0
25% - 30% 30% - 35% 35% - 40% 40% - 45 % >45%

Figure 6.11. Retrieval of soil moisture on Cordevole site: Comparison of results obtained with
the supervised (ANN1) and unsupervised (ANN2) approaches and ground measurements

> 45%

40% - 45 % e% A NN2
e% A NN1
35% - 40%

30% - 35%

25% - 30%

0 5 10 15 20

Figure 6.12: Comparison between mean errors computed


for each class by using the two approaches ANN1 and ANN2

Deliverable No. 8, D4-WP2 54


EnviSnow

Lastly, the algorithm was employed to generate SMC maps at about 0.5 km x 0.5 km resolution,
from ENVISAT ASAR images. Some examples obtained by overlapping SMC maps for the three
dates and the DEM of the area are shown below (Fig. 6.13a,b,c). Unfortunately, the rather constant
SMC values recorded during the three field campaigns did not allow us to obtain a wide temporal
variation, because the mean value of the measurements carried out for each campaign ranged
between 35% in September and 45% in July. However, 4 classes of SMC could be separated in the
images, with the lowest class almost empty in the image of July.

( Fig. 6.13 a )

Deliverable No. 8, D4-WP2 55


EnviSnow

( Fig. 6.13 b )

( Fig. 6.13 c )

Figure 6.13: Soil moisture maps obtained from ENVISAT ASAR images (at VV polarization, 25° incidence
angle) taken over Cordevole on 14 June (a), 19 July (b), and 27 September (c).

Deliverable No. 8, D4-WP2 56


EnviSnow

7 CONCLUDING REMARKS

The sensitivity of radar backscattering to soil moisture has been reviewed with particular regard to
backscattering data available from ASAR-ENVISAT. It has been shown that, although the best
configuration parameters for this application (L-band, HH pol, 25° incidence angle) are not
available from the present satellite sensors, the retrieval of soil moisture on bare and grass-covered
soils is possible at C-band too, by compensating the attenuation effect of herbaceous cover, and by
using adequate inversion algorithms. Among the implemented and tested algorithms: semi-
empirical, Bayes, Nelder-Mead and Neural Networks, the latter was selected as the most reliable for
operational uses. The algorithm was tested on a flat agricultural site (Scrivia watershed) and in a
mountainous area (Cordevole watershed). On the first site, five levels of soil moisture between 10-
15% and 45% were identified, with a mean error of the order of 10%. On the mountainous
watershed, soil was covered by a thick grass layer and, since only VV polarized ASAR data were
available, a correction for vegetation effects was necessary. In the latter case, at least four levels of
soil moisture between 25% and 45% could be identified, with a mean error of the order of 10%.
However, it should be noted that before this approach can be considered fully operational, test of
more data in a wider moisture and environmental conditions, and further analyses will be necessary.

Deliverable No. 8, D4-WP2 57


EnviSnow

REFERENCES

Baquero M. et al., “Perpendicular, monostatic measurements of the dielectric constant of a layer of


unknown material” , Jan. 1996, Rep. EUR 16 370 EN.
Baronti, S., Del Frate, F., Ferrazzoli, P., Paloscia, S., Pampaloni, P., Schiavon, G., 1995, “SAR
polarimetric features of agricultural areas”, International Journal of Remote Sensing, 16, 2639-2356
Beaudoin, A., Le Toan, T. and Gwyn, Q., 1990, SAR observations and modelling of the C-band
backscatter variability, due to multi-scale geometry and soil moisture. IEEE Transactions on
Geoscience and Remote Sensing, 28, 886- 895.
Bennet J.M. and Mattsson L., “Introduction to surface roughness and scattering. Washington D.
C., 1989.
Bennet J.M., "Recent developments in surface roughness characterization," Meas. Sci. Technol.,
vol. 3, pp. 1119-1127, 1992.
Brown S.R. and C.H. Scholz, "Broad bandwidth study of the topography of natural rock surfaces",
J. Gephys. Res., 90, B14, pp. 11575-11582, 1985.
Champion and R. Faivre , 1977, Sensitivity of the Radar Signal to Soil Moisture: Variation with
Incidence Angle, Frequency, and Polarization IEEE Transactions on Geoscience and Remote
Sensing, vol. 35, no. 3, May 1997, pp. 781-783
Church E.L., "Fractal surface finish, "Applied Optics, vol. 27, pp. 1518-1526,1988.
Coppo P., D. Tarchi, and M. Weinberger, “An experimental model for surface scattering models
validation at the European Microwave Signature Laboratory,” in Proc. Int. Geoscience and Remote
Sensing Symp. (IGARSS’95), Florence, Italy, July 10–14, 1995, pp. 930–932.
Coppo P., G. Macelloni, P. Pampaloni, S. Paloscia and S. Sigismondi, 1996, The SIR-C/X-SAR
experiment: the sensitivity of microwave backscattering to surface roughness of bare soils, Proc.
Int. Geosci Remote Sensing Symp. (IGARSS'96), Lincoln, Nebraska, 2131-2133.
Dawson M.S., Fung A.K. and Manry M.T. “A robust statistical-based estimator for soil moisture
retrieval from radar measurements”, IEEE Transactions on Geoscience and Remote Sensing, 35, pp.
57-67, 1997.
De Carolis G., F.Mattia, G.Pasquariello, F.Posa, and P. Smacchia, 1994, "A sirnple model for
retrieval bare soil moisture from radar-scattering coefficients," Journal of Electromagnetic Waves
and Applications, vol. 8, pp. 1017-1041.
De Loor, G.P., 1982,The dielectric properties of wet materials. International Geoscience and
Remote Sensing Symposium. Digest, edited by W.Keydel, (IGARSS'82, Munchen), 82CH14723-6,
TP-1, 1-5.
Deroin J.P., Company A. and Simonin A. “An empirical model for interpreting the relationship
between backscattering and arid land surface roughness as seen with the SAR”, IEEE Transactions
on Geoscience and Remote Sensing, 35, pp.86-92, 1997.
Dobson C., Pierce L., Sarabandi K., Ulaby F., Sharik T., 1992, Preliminary analysis of ERS-1 SAR
for forest ecosystem studies, IEEE Transactions on Geoscience and Remote Sensing, 30, 2, 203-
211
Dobson, M.C., F.T. Ulaby, M. Hallikainen and M. El Rayes, 1985, Microwave dielectric behavior
of wet soil - part II: Four component dielectric mixing models. IEEE Transactions on Geoscience
and Remote Sensing, 23, 35-46

Deliverable No. 8, D4-WP2 58


EnviSnow

Dobson, M.C., F.T.Ulaby, L.E. Pierce, T.L. Sharik, K.M. Bergen, J.Kellindorfer, J.R. Kendra, E.
Li, Y.C.Lin, A. Nashshibi, K. Sarabandi, and P. Siqueira, 1995, “Estimation of Forest Biophysical
Characteristics in Northern Michigan with SIR-C/X-SAR”, IEEE Transactions on Geoscience and
Remote Sensing, 33, 877 – 895
Dong W.P., P.J.Sullivan, K.J.Stout, 1992, "Comprehensive study of parameters for characterizing
three-dimensional surface topography I: Some inherent properties of parameter variation," Wear,
vol. 159, pp. 161-171.
Dong W.P., P.J.Sullivan, K.J.Stout, 1993, "Comprehensive study of parameters for characterizing
three-dimensional surface topography II: Statistical properties of parameter variation," Wear, vol.
167, pp. 9-21.
Dong W.P., P.J.Sullivan, K.J.Stout, 1994, "Comprehensive study of parameters for characterizing
three-dimensional surface topography III: Parameters for characterizing amplitude and some
functional properties," Wear, vol. 178, pp. 29-43.
Dubois P.C., Van Zyl J.J. and Engman T. “Measuring soil moisture with imaging radars”. IEEE
Transactions on Geoscience and Remote Sensing, 33, pp. 896-904, 1995.
Dubuc B. et al, "Evaluating the fractal dimension of profiles," Physical Review A, vol. 39, pp. 1500-
1512, 1989.
Ferrazzoli, P., S.Paloscia, P. Pampaloni, G.Schiavon, S.Sigismondi and D. Solimini, 1997, “The
potential of multifrequency polarimetric SAR in assessing agricultural and arboreous biomass”,
IEEE Transactions on Geoscience and Remote Sensing, 35, 5-17
Fung A.K. “Microwave scattering and emission models and their applications”. Boston: Artech
House Inc., 1994.
Fung A.K., M.S. Dawson, K.S.Chen, A.Y Hsu, E.T.Engman, P.O’Neil and J.Wang,, 1996, "A
modified model for scattering from soil surfaces with application to soil moisture sensing,"
presented at IGARSS96, Lincoln, Nebraska.
Haddad, Z.S., Dubois, P., and van Zyl, J.J., 1996, “Bayesian estimation of soil parameters from
radar backscatter data”, IEEE Transactions on Geoscience and Remote Sensing, 34, 1, pp. 76 -82
Hallikainen M.T., F. Ulaby, M.C. Dobson., M.A. El-Rayes, Lin Ku Wu, 1985, Microwave
dielectric behavior of wet soils - Part I: Empirical models and experimental observations. IEEE
Transactions on Geoscience and Remote Sensing, 23, 25-33
Hornik K., Stincombe M. and White H., 1989. “Multilayer feed-forward networks are universal
approximators,” Neural Networks, Vol. II, 359-366.
Hunt B.R., 1977, "Bayesian Methods in Nonlinear Digital Image Restoration," IEEE Transactions
on computers, vol. 26.
Huppi, R., and Schanda, E., 1986 , L to X band scatter and emission measurements of vegetation.
International Geoscience and Remote Sensing Symposium. Moving Towards the 21st Century,
edited by T.D. Guyenne and J.J. Hunt (European Space Agency, Paris), SP - 284, 1113-1118.
Ishimaru A. and Chen J.S., 1990, "Scattering from very rough surfaces based on the modified
second order Kirchhoff approximation with angular and propagation shadowing," Journal of the
Acoustical Society of America, vol. 88, pp. 1877-1883.
Karam M. A., “Bridging the quasi static and the physical optics approximations: an elliptic disk
case,” Applied Optics, vol 37,9, pp. 1666-1673, 1998.

Deliverable No. 8, D4-WP2 59


EnviSnow

Le Toan, T., 1982, Active microwave signatures of soil and crops: significant results of three years
of experiments. International Geoscience and Remote Sensing Symposium. Digest, edited by
W.Keydel, (IGARSS, Munchen), 82CH14723-6, TP2, 3.1-3.5.
Macelloni G., S. Paloscia, P. Pampaloni, S. Sigismondi, P. de Matthæis, P.Ferrazzoli, G.Schiavon,
D.Solimini, 1999, “The SIR-C/X-SAR experiment on Montespertoli: sensitivity to hydrological
parameters”, International Journal of Remote Sensing, vol.20, n.13, pp. 2597-2612
Macelloni G., Nesti G., Pampaloni P., Sigismondi S., Tarchi D. and Lolli S., 2000 “Experimental
validation of surface scattering and emission models”. IEEE Transactions on Geoscience and
Remote Sensing, vol. 38, n. 1, pp. 459-469.
Macelloni G., S. Paloscia, P. Pampaloni, F. Marliani and M. Gai, 2001, “The relationship between
the backscattering coefficient and the biomass of narrow and broad leaf crops,” IEEE Transactions
on Geoscience and Remote Sensing, vol. 39, n.4, pp. 873-884.
Macelloni G., S. Paloscia, P. Pampaloni, R. Ruisi, M. Dechambre, R. Valentin, A. Chanzy, L.
Prévot and N. Bruguire, “Modelling Radar Backscatter From Crops During The Growth Cycle”,
Agronomie, 22, 575-579, 2002a.
Macelloni G., S. Paloscia, P. Pampaloni, R. Ruisi, M. Dechambre, R.Valentin, A. Chanzy , J.-P.
Wigneron, 2002b, “Active and passive microwave measurements for the characterization of soils
and crops”, Agronomie, 22, 581-586.
Nelder J. A. and Mead R., “A Simplex Method for Function Minimization,” Computer Journal,
Vol. 7, p. 308-313, 1965.
Nesti G., P. Pampaloni, P. Coppo, M. Hallikainen, M. Mancini, P. Troch, and M. von Shonermark,
1995, “Experimental research at EMSL on scattering properties of non vegetated terrains,” in Proc.
Int. Geoscience and Remote Sensing Symp., (IGARSS’95), T. Stein, Ed., Firenze, Italy, pp. 2020–
2022.
Notarnicola C., F. Posa, ’Bayesian Fusion of Active and Passive Microwave Data for Estimating
Bare Soil Water Content’, in Proc. Int. Geoscience and Remote Sensing Symp.,IGARSS 2001, 9-13
July 2001.
Nowicki, "Multi-parameter representation of surface roughness," Wear, vol. 102, pp. 161-176,
1985.
Oh Y., Sarabandi K. and Ulaby F.T., 1992, “An empirical model and an inversion technique for
radar scattering from bare soil surfaces,” IEEE Transactions on Geoscience and Remote Sensing,
vol. 30, pp. 370-381.
Oh Y., 2000, “Retrieval of the effective soil moisture contents as a ground truth from natural soil
surfaces” Proceedings of the Geoscience and Remote Sensing Symposium, IGARSS 2000, IEEE
2000 International , vol. 4, pp. 1702 -1704
Oh Y., Sarabandi K. and Ulaby F.T., 2002, “Semi-empirical model of the ensemble-averaged
differential Mueller matrix for microwave backscattering from bare soil surfaces,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 40, pp. 1348 –1355.
S. Paloscia, P. Pampaloni, E. Santi, “The Use of Sar Data in Measuring Soil Moisture,” Envisnow
Project Contract EVG1-CT-2001-00052 Report WP2-3/2002, November 2002.

Peitgen H.O. and Saupe D, 1988, “The science of fractal images,” Berlin: Springer Verlag.
Shi J., Van Zyl, J.J., Soares J.V. and Engman E.T., 1992 “Development of soil moisture retrieval
algorithm for L-band SAR measurements”. In Proceedings of the International Geoscience and
Remote Sensing Symposium-IGARSS ’92, Houston: IEEE, pp. 495-497.

Deliverable No. 8, D4-WP2 60


EnviSnow

Souyris J.C., L.Wang, C.Hsu, J.A.Kong, T.Le Toan, N.Boudier, S.H.Yueh, U.Wegmuller, and
C.Matzler, 1994, “Inversion of Soil Moisture with Radar Backscattering Data,” presented at
IGARSS’94, Pasadena, California.
Stofan E.R., D.L. Evans, C. Schmullius, B. Holt, J.J. Plaut, J.J. van Zyl, S.D.Wall and J.Way, 1995,
Overview of the results of spaceborne imaging radar-C,X-band Synthetic aperture radar (SIR-C/X-
SAR), IEEE Transactions on Geoscience and Remote Sensing,vol. 33, 817 – 828
Thomas T.R., Rough surface. London: Longman, 1982.
Tsang L., Kong J. and Shin R. “Theory of Microwave Remote Sensing”, J. Wiley & Sons, 1985.
Tsang L., M.C. Kubacsi and J.A. Kong, “Radiative transfer theory for active remote sensing of a
layer of small ellipsoidal scatterers,” Radio Sci., Vol. 16, 3, pp. 321-329, 1981.
Tsang L., C.H. Chan, J.A. Kong, and J. Joseph, “Polarimetric signatures of a canopy of dielectric
cylinders based on first and second order vector radiative transfer theory,” J. Electromagnetic
Waves and Applications, vol. 6, pp. 19-51, 1992.
Ulaby F.T., Moore R.K. and Fung A.K. “Microwave Remote Sensing: Active and Passive”. Vol.II -
Artech House, Dedham, MA, 1982.
Ulaby F.T., R.K. Moore, and A.K. Fung, 1986, Microwave Remote Sensing: Active and Passive.
Vol.III - Surface Scattering and Emission Theory, Artech House, Dedham, MA.
Ulaby F.T., R.K.Moore, A.K. Fung, 1990, From theory to applications, vol. 3. Norwood, MA:
Artech House.
Walters F. H., Parker, L. R. Jr., Morgan S. L., and Deming S. N., “Sequential Simplex
Optimization: A Technique for Improving Quality and Productivity in Research, Development, and
Manufacturing,” CRC Press, Boca Raton, FL, 1991.
Wang L., R.T.Shin, J.A.Kong, S.H. Yueh, 1993, “Application of neural network to remote sensing
of soil moisture using theoretical polarimetric backscattering coefficients,” presented at PIERS 93.
Winebrenner D., and Ishimaru A., "Investigation of a surface field phase perturbation technique for
scattering from rough surfaces," Radio Science, vol. 20, pp. 161-170, 1985.
Zebker, H.A., J.J.Van Zyl, S.L. Durdeen and L. Norikane, 1991, Calibrated imaging radar
polarimetry: technique, examples and applications. IEEE Transactions on Geoscience and Remote
Sensing, vol. 29, 942-961.

Deliverable No. 8, D4-WP2 61

View publication stats

You might also like