ul 2004
in using
lifferent
copolar
ntial for
relation
for low
m. This
en bare
'ameters
r (1997)
ane into
der to
h of the
jattering
; data in
trix data
>s shows
entropy.
rameters
entropy
roups of
cattering
the a -
y values
ring and
volume
dy It is
es in the
1 roads:
5). For
ng is the
ond less
teraction
present.
ample, is
ar wave
ation 1s
are soils
pond for
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part BI. Istanbul 2004
90
B Dare soil
80 D Road
20 A l'arest
A Wheat
60 + Trihedral
“| + Dihedral
X Apricot
a-angle
X Pear
O Peach
® Lawn
* House
% Low building
= High building
0 9.1 02 €0574 94^ 05 * 05 07 —Og **09 I
Entropy
9 Bare soil
0.9 = - DO Road
0.8 2 $ A Forest
A Wheat
0.7
+ n * lrihedral
0.6 » + Dihedral
X Apricot
E a n X Pear
m 5 Oo 084% AR JY O Peach
9 Lawn
Anisotropy
e.
>
0.2 * House
0.1 * Low building
35 High building
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 |
Entropy
(b)
Figure 3. Entropy/ @ -angle and entropy/anisotropy scatterplots
for training sites of various surface types. Each point represents
the mean value of a training field.
Ferro-Famil and Pottier (2001) noted that the introduction of the
anisotropy information improves the discrimination between
targets. Each area in the H/a plane is thus divided into two
classes according to the pixels anisotropy value being greater
than 0.5 or less than 0.5. For our targets, anisotropy is not a
discriminating parameter at X-band. Its value is approximately
the same for bare soils, roads, orchards, forest, lawn, wheat and
low building (0.3<A<0.4). For the other targets (trihedral,
dihedral, house and high building), the anisotropy is superior to
0.5:
The results obtained in this study are in agreement with the
results obtained by Ferro-Famil and Pottier (2001). They
demonstrated that in P-band, it is possible to separate the tree
parcels in 3 categories; small, moderate and old trees. The
results obtained in L-band indicated that different types of trees
couldn’t be separated precisely. In C-band no discrimination
between different tree ages is possible.
This signature study demonstrates the effectiveness of
polarimetric X-band for mapping specific land use classes. A
decision tree model can be used to classify the scene by
combining the different parameters of the radar signal
(backscattering coefficients, copolarization and depolarization
ratios, correlation coefficients, entropy, @ -angle, and
anisotropy). The classification rules can be easily determined
from training data. The following list proposes the parameters
to be used in such a classification process. By applying these
parameters in order, ambiguity between classes is eliminated:
Trihedral: VV or HV/VV or Entropy.
- A Dihedral: correlation coefficient between cross and copolar
or a -angle.
- House: depolarization ratios or Entropy or Anisotropy.
- High building: HH or HV or Anisotropy.
- Low building: HV.
- Bare soils: Entropy or « -angle.
- Roads: HV.
- Forest: H/@ plane (zone 4).
- Orchards, wheat and lawn: difficult to classify as the
confusion between the classes will be large.
4.3 Relationship between surface roughness and the radar
signal
For bare soils, the backscattering coefficients decrease by about
8dB in HH and VV and 10dB in HV between the two bare soil
fields R4 and R5 (rms surface height of 0.85 and 0.80cm,
respectively ; Figure 4). This decrease is due to the increase of
both incidence angle and correlation length, and the decrease of
soil moisture (cf. Table 1). The radar signal is therefore
dependent on the incidence angle. This relationship is given
approximately by the function cos"0 (Baghdadi et al., 2000).
The parameter o is dependent on the dominant scattering
mechanism and sensor parameters (Shi et al., 1994). In general,
for a radar frequency given, the parameter a is calculated for
each surface type and each polarization. Next, the mean a is
computed for each polarization by calculating the mean of a for
all classes. However, for our image, only the wheat field class is
present at different incidence angles. Because of this, we have
used the parameter o. estimated from the wheat field class data
to reduce the angular dependence of the radar signal. In Table 2
the angular dependence is listed for various polarizations. The
coefficient of determination R? is 0.95 for the HH and VV
polarizations, and 0.88 for the HV polarization. Our training
sites of various classes are located between incidence angles of
26? and 32?. This variation in the incidence angle could
potentially cause variations in signal power of up to 1.9dB in
the HH polarization, 1.6dB in the HV polarization, and 2.9dB in
the VV polarization.
HH CROSS VN
Cos 77g Cos? 778g Cost? 79
Table 2: The angular dependence of radar signal at different
polarizations.
The difference in incidence angle between the two bare soil
fields R4 and R5 (rms of 0.85 and 0.80cm, respectively) could
cause a decrease in the radar signal of about 1.6dB in the HH
polarization, 1.4dB in the HV polarization, and 2.4dB in the VV
polarization. As mentioned in the literature, an increase in the
surface moisture of about 5% leads to an increase of the
retrodiffusion coefficient of approximately 1dB when the
incidence angle is smaller than 20° (Le Toan et al., 1994).
Thus, the decrease in the radar signal caused only by the
decrease in soil moisture (about 1496) is approximately 3dB.
The slaking crust observed in the field R5 (high correlation
length comparatively to the field R4) increases the specular
scattering and leads to a decrease of the radar signal of 3.4dB
for the HH polarization, 5.6dB for the HV polarization, and
2.6dB for the VV polarization. It is thus possible to track the
surface degradation due to the slaking process and to
distinguish the freshly tilled fields (R1 to R4 as compared to
R5).
