International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B1. Istanbul 2004
4.2 Signature Study
Image analysis was carried out to study the behaviour of the
principal surface types on the study site (bare soil fields with
varying surface roughness, wheat fields, orchards, forest areas,
buildings, houses and roads) and to investigate the polarimetric
parameters extracted from SAR data in order to discriminate the
observed classes. The parameters chosen correspond to
parameters frequently used in the literature: backscattering
coefficients, ^ copolarization and depolarization ratios,
magnitude of the correlation coefficients, entropy, « -angle,
and anisotropy. For each surface type, several training sites
based on field observations were selected. Statistical analysis of
various parameters for various targets are shown in Figure 2.
Each point plotted represents the parameter mean value for :
given training site (e.g. a field). This is calculated by averaging
the values of all the pixels from the site.
Results show that the backscattering coefficients are ineffective
in discriminating the different classes. We observe a poor
separation between the areas of vegetation (forest, wheat, ...)
for all polarizations, a good separation between different
building types especially in HV polarization, and a moderate
separation between bare soils and the other natural classes in
HH and VV polarizations. Roads can be easily differentiated
from the surface types using the HV polarization.
NN M Bare soil
+ 10 Road
A Forest
A Wheat
| 9 Trihedral
| + Dihedral
| XApricot
|X Pear
{ OPeach
x cos | S9 Lawn
x
> con)
10
NS
y
bUDo
| XC House
Backscattering coefficient HH (dB)
p»
+
oo
| x Low building
3$ High building
. Wi Bare soil
| ORoad
i A Forest
3 4A Wheat
pd ! @ Irihedral
+ Dihedral
| XApricot
i X Pear
xx | O Peach
@ lawn
i House
> pp
+».
x
x
oo
®
VIE
Backscattering coefficient HV (dB)
X3 Low building
DS ee ee
= 2: High building
rn HW Dare sot
+ | |D Road
15 j A Forest
| $ A Wheat
+ Trihedral
LE Dihedral
X Apricot
X Pear
sd A i O Peach
E i a a x x O e ® Lawn
A
i House
Backscattering coefficient VV (dB)
54 Low building
ish m
22 High building
Figure 2 . Behaviour of different parameters of the radar signal
calculated from X-band polarimetric SAR data in function of
various cover types.
The copolarization and depolarization ratios show poor
separability between classes. Only the house class can be
extracted without ambiguity using depolarization ratios.
However, the polarization ratios HV/HH and VV/HH show
good discrimination between forest and wheat. Bare soils and
roads are clearly distinguishable from the other classes in using
the ratio HV/VV but are not themselves separable.
Concerning the degree of coherence between different
polarizations, results show that the correlation between copolar
and cross polar is low for natural areas and the potential for
discrimination between different classes is poor. The correlation
between HH and VV is high for all classes except for low
buildings and forest where the correlation is medium. This
correlation parameter allows a good separability between bare
soils and the other classes.
Concerning the interpretation of polarimetric parameters
(entropy, & -angle and anisotropy), Cloude and Pottier (1997)
have proposed a division of the entropy and a -angle plane into
eight zones of different scattering behaviour, in order to
separate the data into basic scattering mechanisms. Each of the
eight zones corresponds to very specific physical scattering
characteristics:
Zone 1: High entropy multiple scattering
Zone 2: High entropy volume scattering
Zone 3: Medium entropy multiple scattering
Zone 4: Medium entropy volume scattering
Zone 5: Medium entropy surface scattering
Zone 6: Low entropy multiple scattering
Zone 7: Low entropy volume scattering
Zone 8: Low entropy surface scattering
Figure 3a shows the distribution of airborne RAMSES data in
the H/ 4 plane with the valid region for coherency matrix data
shown with a dotted line. The distribution of H/a values shows
that the distribution is concentrated at low to medium entropy.
Also, we observe overlapping of the polarimetric parameters
within H/a plane for different classes. The analysis of entropy
and @ -angle enable us to discriminate five principal groups of
clusters:
- Houses and dihedral: low entropy multiple scattering
(double bounce scattering) and high values for the a -
angle (zone 6).
- . Trihedrals: low entropy surface scattering and low values
for the a -angle (zone 8).
- High building: medium entropy multiple scattering and
high values for a (zone 3).
- Low building and forest: medium entropy volume
scattering and medium values for a (zone 4). It is
impossible to distinguish between these two classes in the
H/a plane.
- Wheat fields, lawns, orchards, bare soils and roads:
medium entropy and low values for a (zone 5). For
wheat fields, lawns and orchards, surface scattering is the
dominant process at X-band, however a second less
significant scattering process resulting from the interaction
with the vegetation layer (volume scattering) is present.
The medium entropy observed for wheat, for example, is
most probably due to penetration of the radar wave
through the vegetation canopy. This penetration is
however weak in X-band. Over roads and bare soils
mainly surface scattering appear.
In addition, targets occurring in zones 1 and 2 correspond for
the most part to vegetation and forest stands.
International
90
80
70
60
2.
"eb 50
$ 40
5
30
20
+
10 $
0 7
0 0.1
(a)
1 —
0.9
+
0.8
+
0.7 «
0.6
*
Anisotropy
e
(b)
Figure 3. En
for training «
the mean val
Ferro-Famil
anisotropy i
targets. Eacl
classes acco
than 0.5 or
discriminatii
the same for
low buildin
dihedral, hoi
0.5.
The results
results obta
demonstrate
parcels in :
results obtai
couldn't be
between difl
This signa
polarimetric
decision tre
combining
(backscatter
ratios, cor
anisotropy ).
from trainin
to be used |
parameters i
- Trihedi