5.1. GF-3 C-Band PolSAR Dataset
Firstly, the algorithm proposed in this article is validated using the GF-3 [
8] C-band PolSAR data. The color-coded images are shown in
Figure 2.
The GF-3 [
8] C-band PolSAR data is over Xi’an, the fully polarimetric data were acquired on 11 July 2017, the image size is
pixels, and the spatial resolution is about
.
Figure 2a–c show the color-coded images of the Y4R, S4R, and G4U methods, while the color-coded image of the proposed four-component decomposition method is shown in
Figure 2d. In this demonstration, the volume scattering power
is colored green, while the double-bounce scattering power
and odd-bounce scattering power
are colored red and blue, respectively. During data processing, according to (45), the data accuracy is set to
, and 99% of the data meets the requirement. The data used above is mainly from urban areas, so there are many buildings, which also means that double-bounce scattering should be dominant. At the same time, it can be intuitively seen from
Figure 2 that the red color of the urban areas is enhanced in
Figure 2d as compared with
Figure 2a–c. This is because the method proposed in this article concentrates the energy of off-diagonal terms in the coherency matrix that are not used for decomposition on the terms participating in the decomposition, so that the energy of the coherency matrix is fully utilized. At the same time, this enhanced red color in the building area is beneficial for easier identification of artificial structures in vegetation areas.
In order to verify the ability to maintain total energy, the summation
of each pixel along the distance direction is compared with the total SPAN of each pixel along the same distance direction, as shown in
Figure 3. At the same time, as several methods have compressed the energy of the cross-polarization
, the results of several methods can be seen in
Figure 4, where UV represents the total energy of the unprocessed coherency matrix cross-polarization
.
Figure 3a is the original image of the
results along the distance direction by Y4R, S4R, G4U, and the proposed method. In order to present the results more clearly,
Figure 3b selected a portion of
Figure 3a for magnification. From both
Figure 3a,b, it can be seen that the
results of the proposed method and existing methods are basically consistent with the total energy SPAN along the distance direction, which is similar to the conclusion in [
4] that Y4R, S4R, and G4U are consistent with the total energy SPAN.
Due to the fact that Y4R, S4R, G4U, and the proposed methods all reduce the overestimation of volume scattering by reducing cross-polarization energy,
Figure 4 shows the total energy of the cross-polarization term
as an unprocessed coherency matrix processed by different methods. From
Figure 4, it can be seen that, compared with the other three methods, the proposed method has the most significant reduction in total energy of cross-polarization. Compared to methods such as Y4R, S4R, and G4U, the total energy of cross-polarization proposed in this paper has been reduced by approximately 27%.
When
is consistent with SPAN, the reduction of cross-polarization power will cause changes in the power of volume scattering, as can be seen from the analysis in
Section 4. Therefore, to further verify the scattering results of the newly developed four-component decomposition method, the composition power profile results of different methods along the distance direction for the entire image are shown in
Figure 5,
Figure 6 and
Figure 7.
Figure 5 and
Figure 6 represent the power profiles for helix scattering and volume scattering along the distance direction, respectively, and
Figure 7 shows the sum of the power distribution along the distance direction for double-bounce scattering and odd-bounce scattering. Similarly, in order to make the details clearer,
Figure 7 is divided into the original image in
Figure 7a and a partially enlarged image in
Figure 7b for display.
Figure 5 shows the power of helix scattering by different methods along the distance direction. From
Figure 5, it can be seen that, for the entire image, compared with existing methods, the newly developed four-component decomposition method has lower helix scattering power than the Y4R, S4R, and G4U methods. At the same time,
Figure 6 demonstrates the volume scattering power of different methods along the distance direction. The results are consistent with the helix scattering power results, indicating that the proposed method has a lower volume scattering power. In other words, the helix scattering and volume scattering powers of the proposed method have been suppressed.
Figure 7 shows the sum of double-bounce scattering and odd-bounce scattering along the distance direction using different methods.
Figure 7a is the original image of the
results along the distance direction by Y4R, S4R, G4U, and the proposed method.
Figure 7b is an enlarged view of a segment in
Figure 7a, which is for the purpose of presenting the results more clearly. From the figures, it can be seen that for the entire graph, the sum of double-bounce scattering power and odd-bounce scattering power along the distance direction is higher than the other methods; in other words, the sum of double-bounce scattering and odd-bounce scattering has been improved.
Furthermore, in order to quantitatively compare the decomposition results of different methods, the relevant methods include Y4R, S4R, and G4U mentioned earlier, as well as recent literature in the model-based category, such as Maurya1 [
20], Maurya2 [
23], and An [
24]. Maurya1, Maurya2, and An are three-component decomposition methods, where Maurya2 and An use hybrid decomposition techniques; that is, they determine the volume scattering power through eigenvalue decomposition, followed by model-based decomposition using unitary transformation. Different patches are selected, as shown in
Figure 2a. Here, building areas are given special attention: patches A1 and A2 are flyovers, and patches A3 and A4 are buildings. At the same time, consideration was also given to the garden area in the city, as shown in patch A5. The quantitative comparison results for different regions based on the normalized average values of decomposed scattering power are provided in
Table 1, and the total power of different regions is shown in (3).
Table 1 analyzes the normalized average values of the four scattering mechanisms, different types of terrain are selected as shown in
Figure 2a. Due to the fact that Maurya1, Maurya2, and An are three-component decompositions, their decomposition results do not include helix scattering
. Patch A1 and patch A2 are flyovers, as can be seen from
Figure 2a. For flyovers, their architecture is similar to urban constructions, and scattering should be mainly double-bounce scattering and odd-bounce scattering. For patch A1, the percentages of volume scattering
contribution are 28.98%, 24.64%, 25.75%, and 19.68% for Y4R, S4R, G4U, and the proposed method, while the results of Maurya1, Maurya2, and An are 33.08%, 22.22%, and 22.22%, respectively. At the same time, for patch A2, the results of the volume scattering
contribution of several methods are as follows: 34.34%, 29.85%, 30.71%, 33.32%, 18.94%, 18.94% and 24.75%, respectively. In other words, for model-based decomposition methods Y4R, S4R, G4U, and Maurya1, the method proposed in this paper can reduce the volume scattering of Patch A1 and patch A2 by up to 13.4% and 8.59%, respectively. For the hybrid three-decomposition methods Maurya2 and An, the volume scattering power
is solved by
, so
of these two methods are the same. From the equation, three eigenvalues can be obtained,
, then
, which leads to better suppression of volume scattering compared to fully model-based decomposition and improves the proportion of double-bounce scattering and odd-bounce scattering. For the model-based decomposition method, the odd-bounce scattering
of the two flyovers has been improved by 6.57% and 6.16% by the proposed method, respectively, while the hybrid three-decomposition methods can further increase the odd-bounce scattering contribution of the two flyovers by up to 5.7% and 13.86% on the basis of the method proposed in this article. When the total power is consistent, as shown in
Figure 3, the increase in the proportion of double-bounce scattering
and odd-bounce scattering
is due to the decrease in the proportion of volume scattering
. For the four-component decomposition methods Y4R, S4R, G4U, and the proposed method, the proportion of helix scattering
decreased by 1.71% and 1.21%, respectively.
Patch A3 and patch A4 are built-up areas. Based on
Figure 2a, they are dominated by double-bounce scattering. For patch A3, the percentages of double-bounce scattering contribution are 80.22%, 80.80%, 80.89%, 80.04%, and 82.75% for Y4R, S4R, G4U, Maurya1, and proposed method, respectively. At the same time, for patch A4, the results of several methods are as follows: 71.90%, 72.99%, 73.11%, 71.12%, and 76.27%. For odd-bounce scattering, the results of the two built-up areas have increased by 1.41% and 1.76%, respectively. At the same time, the proportion of volume scattering is reduced by 4.26% and 8.03%, respectively, while the proportion of helix scattering decreases by 0.35% and 0.68%, respectively. At the same time, for the hybrid three-component decomposition methods Maurya2 and An, the proportions of volume scattering are 4.95% and 9.24%, which are higher than the proportion of volume scattering in the proposed method. In other words, in patch A3 and patch A4, the proposed method can achieve higher proportions of double-bounce scattering and odd-bounce scattering than Maurya2 and An.
Patch A5 is a garden area in the city; both the color in
Figure 2a and the data results in
Table 1 indicate that the volume scattering in this area is enhanced relative to the built-up areas. The proposed four-component decomposition method also enhances the proportion of double-bounce scattering and odd-bounce scattering in the region, the percentages of volume scattering contribution are 39.84%, 37.09%, 37.56%, 32.71%, and 32.16%, for Y4R, S4R, G4U, Maurya1, and the proposed method, respectively. The proportions of double-bounce scattering and odd-bounce scattering increased by 4.35% and 4.25%, respectively. However, in this garden area, the percentages of the volume scattering contribution of Maurya2 and An are 14.72%, which is much lower than other methods. Furthermore, by observing the results of Maurya2, it can be seen that the sum of normalized powers is not equal to 100%, which means that this method did not fully decompose the energy.
5.2. AIRSAR L-Band PolSAR Dataset
In order to further verify the introduced four-component decomposition scheme, another color-coded image is provided in
Figure 8. The PolSAR dataset was acquired by using AIRSAR L-band. The color-coded images are shown in
Figure 8.
The fully polarimetric data were acquired on 11 May 1999 over San Francisco. The image size is pixels, and the spatial resolution is about .
Similar to the validation of the GF-3 dataset, the color-coded images of Y4R, S4R, G4U, and the proposed method are shown in
Figure 8a–d, where the volume scattering power
is colored green, while the double-bounce scattering power
and odd-bounce scattering power
are colored red and blue, respectively. At the same time, the
, total power of cross-polarization,
,
, and
results of different methods are shown in
Figure 9,
Figure 10,
Figure 11,
Figure 12 and
Figure 13.
From
Figure 8, it can be seen that the mountainous areas covered with forests and built-up areas are mainly dominated by volume scattering and double-bounce scattering mechanisms, respectively, while the water area is dominated by odd-bounce scattering. At the same time, it can be intuitively seen that the red color of the urban areas is enhanced in
Figure 8d as compared with
Figure 8a–c. This is because the method in this article concentrates the energy of off-diagonal terms in the coherency matrix that are not used for decomposition on the terms participating in the decomposition, so that the energy of the coherency matrix is fully utilized.
The analysis method is consistent with the GF-3 dataset.
Figure 9a is the original image of the
results along the distance direction by Y4R, S4R, G4U, and the proposed method. In order to present the results more clearly,
Figure 9b selected a portion of
Figure 9a for magnification. From
Figure 3a,b, the
results of the newly developed four-component decomposition method are basically consistent with the existing methods and total energy SPAN along the distance direction.
For the power of cross-polarization terms
,
Figure 10 shows the results processed by Y4R, S4R, G4U, and the newly developed four-component decomposition method. Compared with the unprocessed coherency matrix, the total power of cross-polarization after being processed by different methods, such as Y4R, S4R, G4U, and the proposed method, has been reduced. At the same time, compared to Y4R, S4R and G4U, the proposed method has the most significant reduction in total energy of cross-polarization, which can be reduced by approximately 20%.
Figure 11 shows the power of helix scattering by different methods along the distance direction, and
Figure 12 demonstrates the volume scattering power of different methods along the distance direction. From the results in
Figure 11 and
Figure 12, it is evident that the newly developed method has a lower helix scattering power and volume scattering power than the Y4R, S4R, and G4U methods, which means that the newly developed four-component decomposition method further overcomes the overestimation of volume scattering and helix scattering. This result is similar to the results of the GF-3 dataset.
Like the GF-3 dataset, the sum of double-bounce scattering and odd-bounce scattering
was also analyzed, as shown in
Figure 13. Where
Figure 13a is the original image of the
results along the distance direction by Y4R, S4R, G4U, and the proposed method.
Figure 13b is an enlarged image. From the figures, the result of the newly developed four-component decomposition method is higher than the other methods, which reduces the impact of volume scattering.
Similar to the analysis in the previous section, different patches are selected as shown in
Figure 8a. The Y4R, S4R, G4U, Maurya1 [
20], Maurya2 [
23], and An [
24] decomposition methods are compared with the method proposed in this paper. At the same time, the percentage of the normalized mean of scattering powers by different methods for these selected patches is calculated in
Table 2. Here, urban areas are given special attention (patches A1 to A4); among these urban patches, patch A4 is an urban area with vegetation, while water areas were also studied (patch A5), as shown in
Figure 8a.
The observations made for the GF-3 image also apply to the AIRSAR image. In the AIRSAR image, patches A1 to A3 are all urban areas, and the difference between patch A1, patch A2, and patch A3 lies in their different orientation angles. Patch A1 and patch A3 have similar orientation angles but different orientation angles from patch A2, as can be seen from
Figure 8a. For Y4R, S4R, G4U, Maurya1, and the proposed method, patch A1 and patch A3 with similar orientation angles, an additional increment of 5.92% and 5.21% by the newly four-component decomposition method of double-bounce scattering
compared with the Y4R method, 3.61% and 3.41% compared with the S4R method, 3.06% and 3.10% compared with the G4U method, and 6.41% and 5.63% compared with the Maurya1 method, respectively. For urban patch A2, there is a similar result as for patch A1 and patch A3. The proportion of double-bounce scattering
by the newly developed method has increased compared to the Y4R, S4R, G4U, and Maurya1 decomposition methods; the result can be 10%. The increase in the proportion of double-bounce scattering means a more accurate interpretation of urban areas. In the vegetation patch A4 case, the contribution of the volume scattering component is 40.50% by the newly developed four-component decomposition method. As can be seen from
Figure 8a, the selected vegetation area lies in the urban region; hence, this vegetation patch also expects significant odd-bounce scattering and double-bounce scattering mechanisms. The water area in patch A5 exhibits blue, so the odd-bounce scattering contribution is dominant. The percentages of odd-bounce scattering contribution are 92.72%, 92.72%, 92.69%, 93.76, and 94.75% for Y4R, S4R, G4U, Maurya1, and the proposed method, respectively, which is consistent with reality. For the hybrid three-component decomposition methods Maurya2 and An, the experimental results are similar to those of GF-3.
In the hybrid decomposition methods Maurya2 and An, since the volume scattering power is first determined based on the minimum that satisfies , and then model-based decomposition is performed through different unitary transformations, Maurya2 and An have the same volume scattering power. From the results of the GF-3 and AIRSAR datasets, it can be seen that the hybrid decomposition methods exhibit significant advantages in built-up areas, especially in highly oriented urban structures (patch A4 of the AIRSAR dataset), which can increase the proportion of double-bounce scattering and odd-bounce scattering. However, in vegetation areas, as is the minimum value that satisfies , this hybrid method will underestimate the volume scattering power, as can be seen in patch A5 of the GF-3 dataset. Compared with other methods, the proportion of volume scattering in the hybrid method can be reduced by up to 25.12% in these vegetation areas. The method proposed in this paper is based on four basic scattering models, which can avoid underestimation of volume scattering by hybrid methods in vegetation areas to some extent. At the same time, compared to other model-based methods, the method proposed in this paper can achieve better performance.
5.3. Iteration Number and Accuracy Analysis
For an iterative approach, the iteration number is very important, given the fact that cannot be infinite in practice. For this paper, we take the GF-3 C-band PolSAR dataset ( pixels) as an example to analyze the impact of iteration number from different aspects.
The number of iterations is closely related to the accuracy requirements. Therefore, this paper first analyzes the number of iterations required for data to meet different accuracies. Here, the iteration numbers required for 96%, 97%, 98%, 99%, and 100% of the GF-3 dataset to meet different accuracies are analyzed. The termination conditions
shown in (45) have accuracies of
,
,
and
,respectively. The iteration number
under different accuracies and data proportions is calculated in
Table 3.
From the above results, it can be seen that when the accuracy requirement is low, a simple iteration can meet the requirements. However, when the accuracy requirement is higher, the number of iterations increases, but it can be achieved through a limited number of operations. Due to the different pixels, the number of iterations that satisfy both
and
close to zero is also different. Therefore, in reality, this paper fixes the maximum iteration number
in (45). Here, we take
and analyze the proportion of data that satisfies different data accuracies under the maximum iteration number; the results are shown in
Table 4.
Furthermore, this paper achieves minimization of the cross-polarization term
by decoupling the off-diagonal terms that do not participate in the four-component decomposition. Therefore, this paper also analyzes the impact of data accuracy. The total power of cross-polarization under different accuracies is shown in
Table 5 below. Assuming 99% of the data meets the accuracy requirement, the number of iterations under different accuracies is shown in
Table 3.
The results in
Table 5 indicate that through the processing of the method proposed in this manuscript, as the data accuracy improves, the total power of cross-polarization moves to a stable result, which is consistent with the results of previous theoretical analyses.
Finally, computational time is considered in this manuscript. Here, the time cost of different methods was calculated by taking the patch A3 area in the GF-3 data as an example. The computational time of the method proposed was conducted with
and
. In the experiment, a local PC with an Intel Core i7 CPU and 16 GB of memory was used to process the data set. The results are shown in
Table 6 below.
From
Table 6, it can be seen that the model-based decomposition methods, including Y4R, S4R, G4U, and Maurya1, consume less computational time than other investigated methods. The proposed method in this paper involves iterative operations and takes relatively more computational time than those compared with the model-based methods shown in
Table 6. However, the computational time of the model-based methods is at the same level. The hybrid decomposition methods, including Maurya2 and An, have the longest computational times among all the methods. The reason is that the hybrid decomposition method involves solving matrix eigenvalues, which means the procedure of eigendecomposing the covariance matrix with the computational complexity of
(
is the dimension of the covariance matrix).