parameters) selected for the land cover classifications us-
ing the different combinations of
TerraSAR-X
,
RADARSAT-2
, and
PALSAR
data. The detailed descriptions of the polarimetric
parameters in Table 3 can be found in Qi
et al.
(2012).
Land Cover Classification with SVMs
SVMs
were used to perform land cover classification with the
selected features.
SVMs
are powerful classifiers that usually
produce higher classification accuracy than
DTs
(Qi
et al.
2012). Moreover,
SVMs
do not require any prior assumption on
the input data, so they are suitable for land cover classifica-
tion using multi-source
SAR
data (Lardeux
et al.
2009).
LibSVM
(Chang and Lin 2011) was used to implement land cover clas-
sification based on the selected features. Radial basis func-
tion (
RBF
) was selected as the mapping kernel of
SVM
. A grid
search strategy was used to derive the optimal parameters of
the
RBF
kernel
SVM
(Chang and Lin 2011). A comparison was
performed between the random forest algorithm, which is a
state-of-the-art classifier, and the
SVM
based on the features
selected with the
DT
. The
SVM
improved the overall accuracy
and kappa value by 1.50% and 1.83, respectively, compared
with the random forest algorithm.
Scattering Mechanism Interpretation of PolSAR Images
Land cover classification results of the different combinations
of L-, C-, and X-band data can be obtained through the above
steps. However, the analysis of land cover classification results
provides limited physical insight into how the frequency
variation affects land cover classification. This study investi-
gated this effect mechanism by using polarimetric decompo-
sition theorems to analyze scattering mechanism variations
caused by the frequency variation. Polarimetric decomposition
theorems can be divided into two categories: coherent and
incoherent decomposition. Coherent decomposition theorems
are proposed to investigate coherent targets, such as man-
made objects, whereas incoherent decomposition theorems are
suitable for describing areas that are dominated by distributed
scatterers (e.g., natural targets). The Pauli decomposition is a
commonly used coherent decomposition technique (Cloude
and Pottier 1996), while the Freeman–Durden decomposition
(Freeman and Durden 1998) and the Van Zyl decomposition
(Van Zyl 1992) are widely used incoherent decomposition
methods. All these three decomposition
radar backscatter into three canonical sc
single or odd-bounce (surface), double o
dral), and volume scattering (Table 4). Gi
incoherent polarimetric decomposition methods are suitable
for different land cover types, this study employed the Pauli,
Freeman–Durden, and Van Zyl decompositions to interpret
the scattering mechanisms in the
PolSAR
images to investigate
how the frequency variation affects land cover classification.
Table 4. Polarimetric parameters of the Pauli, Freeman–
Durden, and Van Zyl decompositions.
Decomposition
Method
Polarimetric Parameter
Odd-bounce
scattering
Double-bounce
scattering
Volume
scattering
Pauli
Pauli_Odd Pauli_Dbl
Pauli_Vol
Freeman–Durden Freeman_Odd Freeman_Dbl
Freeman_Vol
Van Zyl
Van Zyl_Odd Van Zyl_Dbl
Van Zyl_Vol
Results and Discussion
Land cover classification results were obtained by using the
different combinations of
TerraSAR-X
,
RADARSAT-2
, and
PALSAR
data (Figure 5). Figures 6 and 7 demonstrate some details of
the differences between these classification results. The overall
accuracies and kappa values were calculated for these clas-
sification results (Figure 8). Among all the classifications, the
combination of X-, C-, and L-band
PolSAR
images attained the
highest overall accuracy and kappa value, which were 98.31%
and 97.93, respectively. However, the addition of X-band data
improved the overall accuracy and kappa value by only 0.68%
and 0.84, respectively, in comparison with the classification
that used the C- and L-band data. The X-band image also made
limited contribution to the classification when combined with
the C- or L-band data. The combination of X- and L-band im-
ages increased the overall accuracy and kappa value by only
0.25% and 0.30 compared with the sole use of L-band data,
respectively. The use of X- and C-band data together even
decreased the overall accuracy and kappa value unlike the
sole use of C-band data. Among all the classifications that used
a single scene, the X-band image yielded the lowest overall
accuracy and kappa value. The results indicate that X-band
HH
data provide a limited contribution to land cover classification
when combined with C- or L-band fully
PolSAR
data.
Among all the combinations of two frequency bands, the
use of L- and C-band data attained the highest overall accura-
cy and kappa value, which were 97.63% and 97.09, respec-
tively (Figure 8). As shown in Figure 8, compared with the
sole use of C-band data, the addition of L-band data improved
the overall accuracy and kappa value by 8.05% and 9.90,
respectively. Compared with the sole use of L-band data, the
incorporation of C-band data increased the overall accuracy
and kappa value by 4.55% and 5.60, respectively. The results
showed that both C- and L-band data contributed to the
achievement made by their combination.
The producer’s and user’s accuracies were calculated for
each land cover class to discover the detailed contributions
of C- and L-band data (Figure 9). The combination of C- and
L-band improved the producer’s and user’s accuracies of bare/
sparsely vegetated land by 29.40% and 20.28%, respectively
and increased the producer’s and user’s accuracies of water by
7.48% and 10.04%, respectively, in comparison with the sole
use of C-band data. Using the validation samples as the refer-
ence data, we evaluated the confusion between different land
cover classes. Figure 10 shows the main confusions in the differ-
ent land cover classification results. The confusion between two
igure 10 refers to the validation samples
at were misclassified as each other. As
e improvements created by the combina-
ata were due mainly to the decrease in
confusion between water and bare/sparsely vegetated land.
The scattering mechanism differences between bare/
sparsely vegetated land and water in the L- and C-band
images were compared to investigate how the C-band im-
age reduced the confusion between water and bare/sparsely
vegetated land (Figure 11). The Pauli, Freeman–Durden, and
Van Zyl decomposition methods were employed to interpret
the scattering mechanisms in the C- and L-band images. The
values of polarimetric parameters in Figure 11 were the aver-
age values calculated based on the training samples of water
and bare/sparsely vegetated land. The C-band data increased
the volume scattering difference between water and bare/
sparsely vegetated land by as much as 138.65% unlike the
L-band image (Figure 11). The sample sites of bare/sparsely
vegetated land with different surface roughness, namely,
smooth, medium, and rough surfaces, were selected based on
the field observation (Figure 12a). By examining the scatter-
ing mechanism of bare land with different surface roughness
in the C-band
PolSAR
image, we found that the increase in
surface roughness caused a greater increase in volume scatter-
ing than in single- or double-bounce scattering (Figure 12b).
This finding indicates that the increase in the surface rough-
ness of bare/sparsely vegetated land leads to the growth of the
804
November 2019
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
