where
C
is the calibrated value,
DN
is the digital number val-
ues, and
A
is the gain value that was supplied in the LUT file
of the
RADARSAT-2
product.
The calibrated values (
C
) of the
TerraSAR-X
and
RADARSAT-2
images were then converted to the normalized backscattering
coefficients (
σ
0
) expressed in decibel (dB) via the following
calculation:
σ
0
= 10×log(
C
)
(3)
The radiometric calibration for the
PALSAR
images was done
according to the following equation (Rosenqvist
et al.
2007):
σ
0
= 10×
log
10
(
DN
)
2
+
CF
(4)
where
σ
0
is the normalized backscattering coefficients ex-
pressed in dB,
DN
is the digital number values, and
CF
is
the calibration coefficient for
PALSAR
standard products and
equals −83 dB.
The radiometric calibration of the
TerraSAR-X
image was
conducted by using
NEST
_v5.1 software (Engdahl
et al.
2012),
and that of the
PALSAR
and
RADARSAT-2
PolSAR
images was
implemented using PolSARPro_v5.0 software (Pottier and
Ferro-Famil 2012).
A 7 × 7 refined Lee filter was applied to the
TerraSAR-X
,
PAL-
SAR
, and
RADARSAT-2
images to suppress the speckle noise (Lee
et al.
1999). The
TerraSAR-X
,
PALSAR
and
RADARSAT-2
images
were geometrically corrected using the range Doppler terrain
correction method embedded in
NEST
_v5.1 software (Engdahl
et al.
2012). In addition, the de-skewing and multi-looking
of
PALSAR
data were performed with
NEST
_v5.1 software. The
TerraSAR-X
and
RADARSAT-2
images were co-registered perfectly
after the geometric correction, but a slight misregistration
was observed between the
PALSAR
and the other two images.
The co-registration of
PALSAR
and
RADARSAT-2
was conducted
based on the ground control points that were selected manu-
ally. The root-mean-square accuracy of the co-registration was
better than 1 pixel of the
PALSAR
image.
Land Cover Classification of PolSAR Images
Many classification methods have been proposed for
PolSAR
data (Chen
et al.
1996; Qi
et al.
2012; Du
et al.
2015; Samat
et al.
2015). In the present study, a meth
polarimetric decomposition,
OBIA
,
DTs
, a
for
PolSAR
image classification (Qi
et al.
of polarimetric parameters were generated from the
PolSAR
data by using different polarimetric decomposition methods.
Second,
OBIA
was used to delineate image objects and extract
various features from these polarimetric parameters. Third,
the optimal features for land cover classification were selected
using
DTs
. Finally,
SVMs
were used to implement land cover
classification based on the selected features.
Polarimetric Decomposition of PolSAR Data
Compared with traditional single-polarization
SAR
systems,
the main advantage of
PolSAR
systems is the utilization of po-
larized waves, which are sensitive to the differences in scatter-
ing mechanisms between various ground targets. Polarimetric
decomposition aims at separating the
PolSAR
backscatter from
ground targets into independent elements (i.e., polarimetric
parameters) that are associated with various physical scatter-
ing mechanisms that are easily interpreted (Cloude and Pottier
1996). Numerous polarimetric decomposition methods have
been developed for separating the measured backscattering
matrix as a combination of the scattering responses of simpler
objects or to separate coherency matrix or covariance matrix
as the combination of second-order descriptors corresponding
to simpler or canonical objects presented as an easier physical
interpretation (Lee and Pottier 2009). Polarimetric parameters
extracted with polarimetric decomposition methods are re-
lated to the geometrical and geophysical properties of ground
targets and have been proven to be important in advancing the
accuracies of
PolSAR
image classification (Cloude and Pottier
1997; Freeman and Durden 1998; Pottier and Lee 2000). This
study used most of the polarimetric decomposition methods
embedded in PolSARPro_v5.0 software to derive polarimet-
ric parameters (Pottier and Ferro-Famil 2012). Similar to the
process in the previous study (Qi
et al.
2012), 72 polarimetric
parameters were generated from each
PolSAR
image and were
incorporated into land cover classification.
PolSAR
Image Segmentation and Feature Extraction
OBIA
can improve
PolSAR
image classification accuracies by
suppressing the speckle effect and exploiting texture features
to aid in the classification (Qi
et al.
2012). This study incor-
porated
OBIA
into
PolSAR
image classification. The
PALSAR
and
RADARSAT-2
Pauli composition images and the TerraSAR
HH
image were segmented together using the multi-resolution
segmentation technique (Benz
et al.
2004) embedded in
eCognition software (Baatz
et al.
2004). The joint segmenta-
tion aimed to generate an image object layer that can be used
with the different combinations of the images for land cover
classification. Given that the high color contrast among dif-
ferent land parcels results in improved segmentation results
(Benz
et al.
2004), the images used for segmentation were
linearly stretched and converted into 8-bit images to enhance
the color contrast among different land cover types. The scale
parameter controls the size of the resultant image objects, and
the optimal scale parameter is commonly determined by a
heuristic process (Benz
et al.
2004). Different scale parameters
were tested in the joint segmentation of the original 32-bit
images and 8-bit linearly stretched images (Figure 4). The
segmentation results were assessed based on visual interpreta-
tion, which is a widely used method for assessing segmenta-
tion results of remotely sensed images (Benz
et al.
2004; Qi
et
al.
2015; Costa
et al.
2018). The segmentation results showed
that the scale parameters of 20 and 60 were optimal for the 32-
bit and 8-bit images in delineating accurate land parcels and
avoiding fragmental image objects, respectively. In addition,
the segmentation results indicated that the 8-bit stretched
images were better than the original 32-bit images for the ac-
f land parcels with a subtle color differ-
s in the red circle area in Figure 4.
parameters were derived from a single
PolSAR
image, a total of 72 features (i.e., the mean values of
image objects) were extracted from a single
PolSAR
image by
using eCognition software for land cover classification (Baatz
et al.
2004). This study did not extract and employ any tex-
tural and spatial features for land cover classification because
of the differences in the spatial resolution among the
PALSAR
,
RADARSAT-2
, and
TerraSAR-X
data and the limited contribution
of textural and spatial information to
PolSAR
image classifica-
tion (Qi
et al.
2012).
Feature Selection Using DTs
A total of 72 features were extracted from the
SAR
data for
land cover classification. However, incorporating all these
features into classification is inappropriate because redundant
and irrelevant features could contaminate the performance
of the classifiers and lead to intensive computation (Qi
et al.
2012).
DTs
are effective in feature selection for classification
(Lawrence and Wright 2001).
DTs
can handle different types of
features and select the most optimum features that attain the
best classification result by assessing the effects of every fea-
ture to determine every split in the final tree (Qi
et al.
2012).
QUEST, a binary-split
DT
for classification and data mining
(Loh and Shih 1997), was used in this study to implement
feature selection. Table 3 lists the features (i.e., polarimetric
802
November 2019
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
