A Three-Hierarchy Evaluation of Polarimetric Performance of GF-3, Compared with ALOS-2/PALSAR-2 and RADARSAT-2

GaoFen-3 (GF-3) is the first Chinese civilian multi-polarization synthetic aperture radar (SAR) satellite, launched on 10 August of 2016, and put into operation at the end of January 2017. The polarimetric SAR (PolSAR) system of GF-3 is able to provide quad-polarization (quad-pol) images in a variety of geophysical research and applications. However, this ability increases the complexity of maintaining image quality and calibration. As a result, to evaluate the quality of polarimetric data, polarimetric signatures are necessary to guarantee accuracy. Compared with some other operational space-borne PolSAR systems, such as ALOS-2/PALSAR-2 (ALOS-2) and RADARSAT-2, GF-3 has less reported calibration and image quality files, forcing users to validate the quality of polarimetric imagery of GF-3 before quantitative applications. In this study, without the validation data obtained from a calibration infrastructure, an innovative, three-hierarchy strategy was proposed to assess PolSAR data quality, in which the performance of GF-3 data was evaluated with ALOS-2 and RADARSAT-2 data as references. Experimental results suggested that: (1) PolSAR data of GF-3 satisfied backscatter reciprocity, similar with that of RADARSAT-2; (2) most of the GF-3 PolSAR images had no signs of polarimetric distortion affecting decomposition, and the system of GF-3 may have been improved around May 2017; and (3) the classification accuracy of GF-3 varied from 75.0% to 91.4% because of changing image-acquiring situations. In conclusion, the proposed three-hierarchy approach has the ability to evaluate polarimetric performance. It proved that the residual polarimetric distortion of calibrated GF-3 PolSAR data remained at an insignificant level, with reference to that of ALOS-2 and RADARSAT-2, and imposed no significant impact on the polarimetric decomposition components and classification accuracy.


Introduction
GaoFen-3 (GF-3) was launched on 10 August of 2016 and was put into operation at the end of January, 2017 [1]. It is the first Chinese space-borne multi-polarization co-/cross-imaging radar mission in C-band, with a fully polarimetric quad-polarization (quad-pol) mode [2]. The quad-pol mode provides data with at least 40 beams and ground range resolutions of about 8 m and 25 m [3]. These polarimetric data are expected to be substantially applied in sea and ocean monitoring, disaster reduction, water conservancy, and meteorology [4]. The performance of these applications depends extremely on the polarimetric fidelity. This arouses special concern in users of GF-3 polarimetric data about the operations of polarimetric calibration and quality of polarimetric signatures. Despite the introduction of the synthetic aperture radar (SAR) payload design and the report of in-orbit tests and

Histogram-Based Analysis
A histogram helps to clearly present the overall distribution of the digital image. In the present study, Bhattacharyya distance (Bd) was used for quantitative measurement of the similarities of the histogram [23]. In this paper, Bhattacharyya distance was selected for its simplicity and effectiveness in quantitative comparisons of images.
Bhattacharyya distance is defined as： where H1 and H2 are the frequency of normalized histograms, and n is the number of bins in the histogram. The range of Bd was [ ) 0 − +∞ and the same images presented the minimum value of 0.

Pixel-Based Analysis
The scattering matrix of each pixel in a PolSAR image contained full polarimetric information of the corresponding target, which can be used to detect and distinguish land-cover types. Target decomposition theory, making use of a scattering matrix or its second-order statistics, has been widely applied in expressing the scattering mechanisms that lead to the polarimetric signatures seen in a PolSAR image, such as surface scattering, double-bounce scattering, and volume scattering [24,25]. Based on target decomposition theory, this study obtained the scattering mechanism of sampling pixels performed in a PolSAR image and analyzed whether that appropriately reflected the backscatter property of the corresponding land-cover types. Two theories of target decomposition were categorized: coherent target decomposition (CTD) and incoherent target decomposition (ITD). CTD dealt with decomposition of the scattering matrix, whereas ITD made use of the second-order statistics such as coherency or the covariance matrix [26]. In ITD, two representative groups were distinguished: eigenvalue-based approaches (E-ITD) and mode-based approaches (M-ITD). One representative theory in CTD, E-ITD, and M-ITD, respectively, was selected here.

Histogram-Based Analysis
A histogram helps to clearly present the overall distribution of the digital image. In the present study, Bhattacharyya distance (Bd) was used for quantitative measurement of the similarities of the histogram [23]. In this paper, Bhattacharyya distance was selected for its simplicity and effectiveness in quantitative comparisons of images.
Bhattacharyya distance is defined as: where H 1 and H 2 are the frequency of normalized histograms, and n is the number of bins in the histogram. The range of Bd was [0 − +∞) and the same images presented the minimum value of 0.

Pixel-Based Analysis
The scattering matrix of each pixel in a PolSAR image contained full polarimetric information of the corresponding target, which can be used to detect and distinguish land-cover types. Target decomposition theory, making use of a scattering matrix or its second-order statistics, has been widely applied in expressing the scattering mechanisms that lead to the polarimetric signatures seen in a PolSAR image, such as surface scattering, double-bounce scattering, and volume scattering [24,25]. Based on target decomposition theory, this study obtained the scattering mechanism of sampling pixels performed in a PolSAR image and analyzed whether that appropriately reflected the backscatter property of the corresponding land-cover types. Two theories of target decomposition were categorized: coherent target decomposition (CTD) and incoherent target decomposition (ITD). CTD dealt with decomposition of the scattering matrix, whereas ITD made use of the second-order statistics such as coherency or the covariance matrix [26]. In ITD, two representative groups were distinguished: eigenvalue-based approaches (E-ITD) and mode-based approaches (M-ITD). One representative theory in CTD, E-ITD, and M-ITD, respectively, was selected here. Pauli decomposition, as one of the most known and applied CTD theories, decomposes a scattering matrix using a Pauli matrix where every base matrix is associated to a basic scattering mechanism [24]. This model is expressed as: where S is the 2 × 2 Sinclair matrix, and S hv is the scattering coefficient of horizontal transmitting (h) and vertical receiving polarization (v), and the other three coefficients are defined similarly. Pauli decomposition can be interpreted as the coherent decomposition of the Sinclair matrix into four physical mechanisms: (A) surface scattering, (B) double-bounce scattering of orthogonal dihedral corners, (C) cross-polarization components, and (D) all asymmetric components. Among four components for the Pauli decomposition result, only the first three components were chosen for further analysis because the fourth component was not associated with any specific physical scattering mechanism.

Eigenvalue-Based Decomposition
For development of target decomposition approaches based on the Huynen theory, there were three different forms of decomposition result, as there were three completely different matrixes with the first rank corresponding to the coherency matrix [27]. To obtain a unique form of the decomposition result, Cloude first proposed the eigenvalue-based decomposition method because of the invariability of eigenvalues, regardless of the change of base [28]. Further, three parameters (H, α and A) related to the eigenvalues and eigenvectors of the coherence matrix were developed to enterprise the scattering mechanisms of the target [29]. Then, H/ α/A decomposition theory with wide applications was used in this study. In this theory, scattering entropy (H) describes the randomness of eigenvalues of the coherency matrix, mean alpha angle (α) is one of the mean parameters of the dominant scattering mechanism from the coherency matrix, and anisotropy (A) presents the relationship between the second and the third eigenvalues, as the entropy (H) does not completely describe the ratio of the eigenvalues [29]. H, α and A were calculated as: where λ i is the eigenvalue of the coherency matrix, and P i is the discrete probability distribution of the eigenvalues. The coherency matrix was the second-order statistical matrix acquired from the Sinclair matrix. For reciprocal backscatter SAR, this could be obtained by: where T 3 is the coherency matrix used for ITD decomposition.

Model-Based Decomposition
Before the model-based decomposition theory was proposed, the existing decompositions were focused so much on mathematics that they could not be easily interpreted as physical scattering mechanisms [25]. Then, based on the physical model of some simple scattering mechanisms, many decomposition methods were developed to depict the backscattering properties of the target. The well-known four-component decomposition method was put forward by Yamaguchi et al., based on four physical models [30]. It was selected here because of its good performance in depicting the basic physical scattering mechanism of targets in urban areas. The four-component decomposition model is expressed as: (8) where T 3 is the coherency matrix given in (7). The four scattering components represent the surface scatter (P s ), double-bounce scatter (P d ), volume scatter (P v ), and helix scatter (P c ), respectively.

Land-Cover Classification
In this study, pixels in PolSAR images were categorized into three land-cover types, including built-up areas, vegetation, and water. In this paper, a support vector machine (SVM) was used as the classifier for its supervised process and better performance in classification accuracy compared with other popular classifiers, such as maximum likelihood and k-nearest neighbor [31][32][33]. The SVM algorithm is a binary, linear classifier that uses a set of training samples, each of which is marked as belonging to one or the other of two categories, to build a model that assigns each pixel to one category or the other [32]. To keep consistent with process in pixel-based analyses, ten decomposition components obtained by Pauli decomposition, eigenvalue-based decomposition, and model-based decomposition were used as the training features in the SVM classifier.

Study Areas
Current, area-wide information management in highly dynamic urban settings is critically required for future development. In this regard, PolSAR data offered the possibility of a fast and area-wide assessment of urban changes and developments. Hence, two study areas in Beijing and Wuhan, with rapid economic growth and urbanization in recent years, were selected to evaluate the performance of GF-3 PolSAR data. Beijing, the capital of China, is located in the North China Plain Although both Beijing and Wuhan are metropolises of China with large populations, there are some differences between them. Combining both traditional and modern architecture, Beijing is one of the oldest cities in the world, with many historical sites. However, Wuhan is a typical modern city with huge development in the past ten years. In addition, water area in Beijing is distinctly less than that in Wuhan, since Wuhan is developed along the Changjiang (Yangtze) River. As shown in Figure  2, the study areas of Beijing consisted of a large proportion of built-up areas and vegetation, but only a small proportion of water. By contrast, almost half of the study area in Wuhan was water; the rest was built-up areas and vegetation. For the histogram-based experiments, all the PolSAR images over Beijing were clipped into the subset as the left map in Figure 2, and all the PolSAR images over Wuhan were clipped into the subset as the right map in Figure 2. Further, the samples for pixel-based experiments and for training the classifier were selected over the region, demarcated by boxes with green, red, and blue lines.

Polarimetric SAR Data and Ground Reference Data
Many factors have impacts on the polarimetric performance of SAR imaging, including sensor parameters and image acquisition situations. To make a general cross-comparison, we made efforts to collect more PolSAR data from GF-3, ALOS-2, and RADARSAT-2 in the study areas as much as possible. A total of 13 PolSAR images in two areas were collected, including three of GF-3, three of ALOS-2, and one of RADARSAT-2 in Beijing, and three of GF-3, two of ALOS-2, and one of RADARSAT-2 in Wuhan. The 13 images had variable parameters, such as incidence angles and imaging times ( Table 1). The nominal resolutions of GF-3, RADARSAT-2, and ALOS-2 were similar, at about 8 m. Although the operating band of GF-3 (C-band) was different from ALOS-2 (L-band), we conducted a comparison between them. The first reason was that the comparison of GF-3 with ALOS-2 was under the hypothesis that if the distortion effect was insignificant, the GF-3 (C-band) should present more surface scattering phenomena and less double-bounce phenomena than ALOS-2 (L-band) in forest areas with a dense canopy. Another reason was that over each study site, we collected much more GF-3 and ALOS-2 data than RADARSAT-2 data. Thus, the data quantity of GF-3 and ALOS-2 made it possible to compare the stability of these sensors, to some extent. Although both Beijing and Wuhan are metropolises of China with large populations, there are some differences between them. Combining both traditional and modern architecture, Beijing is one of the oldest cities in the world, with many historical sites. However, Wuhan is a typical modern city with huge development in the past ten years. In addition, water area in Beijing is distinctly less than that in Wuhan, since Wuhan is developed along the Changjiang (Yangtze) River. As shown in Figure 2, the study areas of Beijing consisted of a large proportion of built-up areas and vegetation, but only a small proportion of water. By contrast, almost half of the study area in Wuhan was water; the rest was built-up areas and vegetation. For the histogram-based experiments, all the PolSAR images over Beijing were clipped into the subset as the left map in Figure 2, and all the PolSAR images over Wuhan were clipped into the subset as the right map in Figure 2. Further, the samples for pixel-based experiments and for training the classifier were selected over the region, demarcated by boxes with green, red, and blue lines.

Polarimetric SAR Data and Ground Reference Data
Many factors have impacts on the polarimetric performance of SAR imaging, including sensor parameters and image acquisition situations. To make a general cross-comparison, we made efforts to collect more PolSAR data from GF-3, ALOS-2, and RADARSAT-2 in the study areas as much as possible. A total of 13 PolSAR images in two areas were collected, including three of GF-3, three of ALOS-2, and one of RADARSAT-2 in Beijing, and three of GF-3, two of ALOS-2, and one of RADARSAT-2 in Wuhan. The 13 images had variable parameters, such as incidence angles and imaging times ( Table 1). The nominal resolutions of GF-3, RADARSAT-2, and ALOS-2 were similar, at about 8 m. Although the operating band of GF-3 (C-band) was different from ALOS-2 (L-band), we conducted a comparison between them. The first reason was that the comparison of GF-3 with ALOS-2 was under the hypothesis that if the distortion effect was insignificant, the GF-3 (C-band) should present more surface scattering phenomena and less double-bounce phenomena than ALOS-2 (L-band) in forest areas with a dense canopy. Another reason was that over each study site, we collected much more GF-3 and ALOS-2 data than RADARSAT-2 data. Thus, the data quantity of GF-3 and ALOS-2 made it possible to compare the stability of these sensors, to some extent. Ground truth data were collected through fieldwork from 2017 to 2018. Ground reference data were used to select samples for pixel-based analysis. Also, a total of 80% of the data were randomly selected to train the classifier, and the remaining 20% were used to validate classification accuracy. The locations of selected samples in optical images of Google Earth are presented in Figure 2. The samples of built-up areas included residential areas, commercial buildings, grounds, and roads. The building structures in the sample areas of Beijing and Wuhan were not completely the same. Vegetation samples in Beijing were mostly trees in forest parks, and the remaining parts were grasslands. In Wuhan, by contrast, all the vegetation samples were selected in mountain forests. In regards to water, the samples were selected in artificial lakes over Beijing but in natural lakes over Wuhan.

Image Processing
In data pre-processing, radiometric calibrations of GF-3, ALOS-2, and RADARSAT-2 were carried out using algorithms developed by the China Academy of Space Technology (CAST), the Japan Aerospace Exploration Agency (JAXA), and the Canadian Space Agency (CSA), respectively [2,16,34].
The GF-3 digital image was calibrated as: where σ o slc is the backscattering coefficient (dB); I and Q are the real and imagery parts of the complex image, respectively; Q v is the maximum value before image quantization; and K dB is the calibration constant. Q v and K dB are both supplied in the header file.
The ALOS-2 digital image was calculated as: where CF 1 and A are the calibration coefficients. CF 1 for PALSAR-2 JAXA standard product was obtained as −83 dB. A for PALSAR-2 JAXA standard SLC data was equal to −32 dB [16].
The RADARSAT-2 digital image was calibrated as: σ o slc = 10 log 10 where B and A are the offset and the gain, respectively, both supplied in the LUT (look-up-table) file [34]. As to the pixel-based analysis and classification, all using PolSAR images were processed with a 5 × 3 multilook (azimuth × range), and were georeferenced using the WGS84 reference ellipsoid.

Backscatter Reciprocity
To evaluate the backscatter reciprocity (S hv = S vh ) of PolSAR images, three images with the same acquiring season were selected. The difference of backscatter coefficients between HV and VH of GF-1703, A2-1603, and R2-0903 was computed, and their histograms are presented in Figure 3. represented the difference between HV and VH used to evaluate backscatter reciprocity. The histogram of GF-1703 showed that most of the pixels were concentrated on zero, similar with A2-1603 (Bd = 0.01) and R2-0903 (Bd = 0.02). As indicated by the statistical results, GF-3 had similar percentages of pixels, with σ o HV − σ o V H lower than any specific values as ALOS-2, e.g., 1 dB, 2 dB, 3 dB, 5 dB, and 10 dB, and the percentages were lower than that of RADARSAT-2 ( Table 2). where B and A are the offset and the gain, respectively, both supplied in the LUT (look-up-table) file [34]. As to the pixel-based analysis and classification, all using PolSAR images were processed with a 5 × 3 multilook (azimuth × range), and were georeferenced using the WGS84 reference ellipsoid.

Distribution of Backscattering Coefficients
Considering the impact of season and incidence angle, one group of images (GF-1703, A2-1603, and R2-0903) obtained in the same season were selected over Beijing, and another group of images (GF-1702, A2-1601, and R2-1607) with a similar incidence angle around 37° were selected over Wuhan to compare the distribution of backscatter coefficients. As shown in Figure 4, GF-1703 presented more similar characteristics with R2-0903 because it operated at the same frequency as C-band compared to A2-1603. Histograms of showed that GF-1703 and R2-0903 had the same highest frequency (14%), but A2-1603 had the highest frequency at 17%. As to the histograms of , GF-1703 and R2-0903 also displayed the same highest frequencies (18%), but A2-1603 reached the highest frequency of 24%. In addition, both GF-1703 and R2-0903 had the highest frequency at −20dB , but A2-1603 reached the highest frequency at −17 dB . The similarity analysis verified the observation of

Distribution of Backscattering Coefficients
Considering the impact of season and incidence angle, one group of images (GF-1703, A2-1603, and R2-0903) obtained in the same season were selected over Beijing, and another group of images (GF-1702, A2-1601, and R2-1607) with a similar incidence angle around 37 • were selected over Wuhan to compare the distribution of backscatter coefficients. As shown in Figure 4, GF-1703 presented more similar characteristics with R2-0903 because it operated at the same frequency as C-band compared to A2-1603. Histograms of σ o HH showed that GF-1703 and R2-0903 had the same highest frequency (14%), but A2-1603 had the highest frequency at 17%. As to the histograms of σ o VV , GF-1703 and R2-0903 also displayed the same highest frequencies (18%), but A2-1603 reached the highest frequency of 24%. values than those between GF-3 and ALOS-2 for all σ o HH , σ o VV and σ o HV . In particular, GF-1710 and R2-0903 obtained high similarities of σ o HH and σ o VV (Bd < 0.05). This was the expected performance for GF-3, compared with RADARSAT-2 operating at C-band and ALOS-2 operating at L-band (Table 3).
Sensors 2019, 19, 1493 9 of 21 histograms, with the Bhattacharyya distances between GF-3 and RADARSAT-2 presenting significantly lower values than those between GF-3 and ALOS-2 for all , and . In particular, GF-1710 and R2-0903 obtained high similarities of and (Bd < 0.05). This was the expected performance for GF-3, compared with RADARSAT-2 operating at C-band and ALOS-2 operating at L-band (Table 3).  In Wuhan, GF-1702, A2-1601, and R2-1607 were selected for their similar incidence angles, but they were obtained in different seasons. Notably, all histograms showed double peaks ( Figure 5) because there were large parts of water in the study area. The Bhattacharyya distances between GF-1702, A2-1601, and R2-16037 ranged from 0.10 to 0.69, and were generally larger than that in Beijing. GF-1702 and A2-1601 reached a high similarity in (Bd = 0.1); while GF-1710 and R2-0903 had a  In Wuhan, GF-1702, A2-1601, and R2-1607 were selected for their similar incidence angles, but they were obtained in different seasons. Notably, all histograms showed double peaks ( Figure 5) because there were large parts of water in the study area. The Bhattacharyya distances between GF-1702, A2-1601, and R2-16037 ranged from 0.10 to 0.69, and were generally larger than that in Beijing. GF-1702 and A2-1601 reached a high similarity in σ o HH (Bd = 0.1); while GF-1710 and R2-0903 had a medium similarity in σ o VV , with a Bd around 0.5. This meant that the observed backscattering coefficients did not only depend on frequency, but were impacted by the season. The histogram-based analysis indicated that GF-3 had a similar histogram with the data of other sensors when they had the same operating band and image-obtaining season.
Sensors 2019, 19,1493 10 of 21 medium similarity in , with a Bd around 0.5. This meant that the observed backscattering coefficients did not only depend on frequency, but were impacted by the season. The histogrambased analysis indicated that GF-3 had a similar histogram with the data of other sensors when they had the same operating band and image-obtaining season.

Polarimetric Performance of Target
In this paper, we assumed the first component (PauliA) of the Pauli decomposition represented surface scattering power, and that the sum (PauliB + PauliC) of the second and third components represented compound scattering based on dihedral structures and dipoles [25,30]. The physical meaning of PauliB + PauliC corresponded to the scatterers. When HH was superior to VV in the radar of built-up areas, the main contribution of PauliB was made by orthogonal ground-wall structures, and the main source of the cross-pol component (PauliC) was the radar from rotated dihedral structures. Thus, PauliB + PauliC could be seen as the double-bounce scattering power of the compound [35,36]. When HH and VV were almost equal in the radar of forest canopy, the contribution of double-bounce scattering (PauliB) was small, and the main source of PauliB + PauliC could be seen as volume scattering from dipoles [36]. When the backscattering intensities were low over water, PauliB + PauliC had little physical meaning under the noise effect. It was also noticed that water and vegetation were always recognized as distributed targets (or incoherent targets), and over those areas it was difficult to give a practical, physically-based interpretation of the components of the Pauli decomposition. Since the targets in urban areas seem to be coherent with slight speckle noise, the results of the Pauli decomposition in urban areas presented expected double-bounce scattering phenomena for all the three sensors with higher PauliB + PauliC values than PauliA. It was interesting that many dots in Figure 6 were aligned with the curve x = y for all the three sensors. This could be explained that most of the selected samples were incoherent targets so that the components

Polarimetric Performance of Target
In this paper, we assumed the first component (Pauli A ) of the Pauli decomposition represented surface scattering power, and that the sum (Pauli B + Pauli C ) of the second and third components represented compound scattering based on dihedral structures and dipoles [25,30]. The physical meaning of Pauli B + Pauli C corresponded to the scatterers. When HH was superior to VV in the radar of built-up areas, the main contribution of Pauli B was made by orthogonal ground-wall structures, and the main source of the cross-pol component (Pauli C ) was the radar from rotated dihedral structures. Thus, Pauli B + Pauli C could be seen as the double-bounce scattering power of the compound [35,36]. When HH and VV were almost equal in the radar of forest canopy, the contribution of double-bounce scattering (Pauli B ) was small, and the main source of Pauli B + Pauli C could be seen as volume scattering from dipoles [36]. When the backscattering intensities were low over water, Pauli B + Pauli C had little physical meaning under the noise effect. It was also noticed that water and vegetation were always recognized as distributed targets (or incoherent targets), and over those areas it was difficult to give a practical, physically-based interpretation of the components of the Pauli decomposition. Since the targets in urban areas seem to be coherent with slight speckle noise, the results of the Pauli decomposition in urban areas presented expected double-bounce scattering phenomena for all the three sensors with higher Pauli B + Pauli C values than Pauli A . It was interesting that many dots in Figure 6 were aligned with the curve x = y for all the three sensors. This could be explained that most of the selected samples were incoherent targets so that the components acquired from coherent decomposition (Pauli decomposition) may be insufficient to enterprise the scattering mechanism, i.e., Pauli A and Pauli B + Pauli C had little physical meaning. For example, in water and vegetation areas, most samples were incoherent targets, and they obtained almost equal values in both Pauli B + Pauli C and Pauli A . However, built-up areas, vegetation, and water were clearly discriminated by the intensity of Pauli A and Pauli B + Pauli C . As shown in Figure 6, both Pauli A and Pauli B + Pauli C showed much higher values for built-up areas, medium values for vegetation, and much lower values for water. acquired from coherent decomposition (Pauli decomposition) may be insufficient to enterprise the scattering mechanism, i.e., PauliA and PauliB + PauliC had little physical meaning. For example, in water and vegetation areas, most samples were incoherent targets, and they obtained almost equal values in both PauliB + PauliC and PauliA. However, built-up areas, vegetation, and water were clearly discriminated by the intensity of PauliA and PauliB + PauliC. As shown in Figure 6, both PauliA and PauliB + PauliC showed much higher values for built-up areas, medium values for vegetation, and much lower values for water.  Providing that GF-1703 contained more CTs, the decreased entropy was explicable [19].
For vegetation, GF-3 had similar pixels with RADARSAT-2 and ALOS-2 distributed in medium and high entropy portions adjacent to the curve, representing the bound of the minimum observable α value as a function of entropy. GF-1702 was an exception, where most pixels were concentrated in the medium entropy portion. For water, GF-3 also had similar pixels to RADARSAT-2, distributed in the medium right-hand portion of the H-α map. The high entropy of water can be explained as water had a low backscattering power for both C-band or L-band sensors and, thus, the observed coherency matrix of water had approximate low eigenvalues. Also, the waves in water surfaces can increase the randomness of scattering. In general, high entropy means that there is a high scattering order or random scattering with approximate eigenvalues [29]. However, the pixels in water of ALOS-2 exhibited some differences between Beijing and Wuhan. explicable [19]. For vegetation, GF-3 had similar pixels with RADARSAT-2 and ALOS-2 distributed in medium and high entropy portions adjacent to the curve, representing the bound of the minimum observable value as a function of entropy. GF-1702 was an exception, where most pixels were concentrated in the medium entropy portion. For water, GF-3 also had similar pixels to RADARSAT-2, distributed in the medium right-hand portion of the H-map. The high entropy of water can be explained as water had a low backscattering power for both C-band or L-band sensors and, thus, the observed coherency matrix of water had approximate low eigenvalues. Also, the waves in water surfaces can increase the randomness of scattering. In general, high entropy means that there is a high scattering order or random scattering with approximate eigenvalues [29]. However, the pixels in water of ALOS-2 exhibited some differences between Beijing and Wuhan. As shown in Figure 8, both Ps and Pd showed much higher values for built-up areas than vegetation and water. Further, most of the pixels in built-up areas had larger double-bounce scattering powers than surface scattering, except GF-1712 with a lower incidence angle. However, the distributions of pixels in vegetation and water were variable for 13 images. Most pixels in GF-3 data (Figure 8a-c,h-f) displayed dominant surface scattering over forest areas, especially for (b) As shown in Figure 8, both P s and P d showed much higher values for built-up areas than vegetation and water. Further, most of the pixels in built-up areas had larger double-bounce scattering powers than surface scattering, except GF-1712 with a lower incidence angle. However, the distributions of pixels in vegetation and water were variable for 13 images. Most pixels in GF-3 data (Figure 8a-c,h-f) displayed dominant surface scattering over forest areas, especially for (b) GF1710 and (h) GF-1702, where over 90% of pixels presented higher surface scattering power than double-bounce scattering power. By contrast, a large proportion of pixels with dominant double-bounce scattering appeared in ALOS-2 images (Figure 8e,l,m) over forest areas, especially for (m) A2-1601. This meant that the variation between GF-3 and ALOS-2 achieved expected results corresponding to the increased ability at longer wavelengths to penetrate vegetation canopies. In Beijing (Figure 8a-g), only GF-1710 and R2-0903 displayed good discrimination between vegetation and water, while the pixels in vegetation and water were mixed in other maps. By contrast, Figure 8h-m exhibited that vegetation and water could be separated in the P s -P d map because of the larger surface or double-bounce scattering power of the pixels in vegetation than that in water. GF1710 and (h) GF-1702, where over 90% of pixels presented higher surface scattering power than double-bounce scattering power. By contrast, a large proportion of pixels with dominant doublebounce scattering appeared in ALOS-2 images (Figure 8e,l,m) over forest areas, especially for (m) A2-1601. This meant that the variation between GF-3 and ALOS-2 achieved expected results corresponding to the increased ability at longer wavelengths to penetrate vegetation canopies. In Beijing (Figure 8a-g), only GF-1710 and R2-0903 displayed good discrimination between vegetation and water, while the pixels in vegetation and water were mixed in other maps. By contrast, Figure  8h-m exhibited that vegetation and water could be separated in the Ps-Pd map because of the larger surface or double-bounce scattering power of the pixels in vegetation than that in water. As presented in Figure 9, GF-3 and RADARSAT-2 shared similar Pc-Pv diagrams, where most of the pixels in built-up areas and vegetation were mixed, but they had much higher volumes and helix scattering powers than that in water. Compared with GF-3 and RADARSAT-2, most of the pixels in ALOS-2 obtained better helix scattering power in vegetation and water. In general, the pixel-based analysis indicated that GF-3 had similar polarimetric decomposition results with that of ALOS-2 and RADARSAT-2, and that different types of samples were significantly separated for all three sensors. As presented in Figure 9, GF-3 and RADARSAT-2 shared similar P c -P v diagrams, where most of the pixels in built-up areas and vegetation were mixed, but they had much higher volumes and helix scattering powers than that in water. Compared with GF-3 and RADARSAT-2, most of the pixels in ALOS-2 obtained better helix scattering power in vegetation and water. In general, the pixel-based analysis indicated that GF-3 had similar polarimetric decomposition results with that of ALOS-2 and RADARSAT-2, and that different types of samples were significantly separated for all three sensors.

Comparison of Classification Results
A comparison of classification results in Beijing among GF-3, ALOS-2, and RADARSAT-2 are summarized in Table 4. GF-1703, R2-0903, and A2-1603 had similar performances in land-cover classification. They obtained lower overall classification accuracies (CAs, < 80%) and lower overall Kappa coefficients (KC, < 0.70), as well as lower product accuracies (PA, < 80%) in built-up areas and vegetation areas. GF-1712 and A2-1612 performed better, with about 83% CA and 0.70 KC. GF-1710 achieved the best performance with 91% CA and 0.83 KC. In general, the results of classification were good for GF-3 data, except the water in GF-1712 (PA, < 40%; UA, < 30%). For GF-1712 obtained in winter, the water frozen into ice changed the backscattering power and scattering mechanism that lead to the mixture of water and other land-cover types. Moreover, the accuracy of water was easily impacted and changed as there was only a small proportion of water in the study area over Beijing. Table 4. Comparison of land-cover classification results in Beijing using images from GF-3, ALOS-2, and RADARSAT-2 data.

Comparison of Classification Results
A comparison of classification results in Beijing among GF-3, ALOS-2, and RADARSAT-2 are summarized in Table 4. GF-1703, R2-0903, and A2-1603 had similar performances in land-cover classification. They obtained lower overall classification accuracies (CAs, <80%) and lower overall Kappa coefficients (KC, <0.70), as well as lower product accuracies (PA, <80%) in built-up areas and vegetation areas. GF-1712 and A2-1612 performed better, with about 83% CA and 0.70 KC. GF-1710 achieved the best performance with 91% CA and 0.83 KC. In general, the results of classification were good for GF-3 data, except the water in GF-1712 (PA, <40%; UA, <30%). For GF-1712 obtained in winter, the water frozen into ice changed the backscattering power and scattering mechanism that lead to the mixture of water and other land-cover types. Moreover, the accuracy of water was easily impacted and changed as there was only a small proportion of water in the study area over Beijing. As shown in Figure 10, variable classification results of GF-3 were obtained. In GF-1703, some pixels in residential buildings were incorrectly assigned to vegetation, which also happened to R2-0903 and A2-1603. In GF-1712, some trees in forest parks, grasslands in golf courses, and commercial buildings and the ground were incorrectly assigned to water, which was also presented in R2-0903, A2-1603, and AL1610. GF-1710 performed the best, where the artificial lake and forest park were almost perfectly detected from built-up areas.
A comparison of classification results in Wuhan among GF-3, ALOS-2, and RADARSAT-2 are summarized in Table 5. GF-3 and ALOS-2 were stable and similar, around 87% CA and 0.80 KC. By contrast RADARSAT-2 performed better with 92% OCA and 0.89 OKC. All three GF-3 images acquired lower product accuracies (PA < 75%) in built-up areas. Table 5. Comparison of land-cover classification results in Wuhan using images from GF-3, ALOS-2, and RADARSAT-2 data. As shown in Figure 11, the classification results of GF-3 were found to be stable. However, GF-1702 demonstrated that some pixels in residential buildings under construction were assigned to vegetation. In contrast, RADARSAT-2 had a better performance in built-up areas, while ALOS-2 incorrectly assigned some pixels in built-up areas and vegetation to water. For all classification results, with the changes of the image-acquiring situations, the classification accuracy of GF-3 experienced a variation from 75.0% to 91.4%, similar to that of ALOS-2 and RADARSAT-2. As shown in Figure 10, variable classification results of GF-3 were obtained. In GF-1703, some pixels in residential buildings were incorrectly assigned to vegetation, which also happened to R2-0903 and A2-1603. In GF-1712, some trees in forest parks, grasslands in golf courses, and commercial buildings and the ground were incorrectly assigned to water, which was also presented in R2-0903, A2-1603, and AL1610. GF-1710 performed the best, where the artificial lake and forest park were almost perfectly detected from built-up areas. A comparison of classification results in Wuhan among GF-3, ALOS-2, and RADARSAT-2 are summarized in Table 5. GF-3 and ALOS-2 were stable and similar, around 87% CA and 0.80 KC. By contrast RADARSAT-2 performed better with 92% OCA and 0.89 OKC. All three GF-3 images acquired lower product accuracies (PA, < 75%) in built-up areas. As shown in Figure 11, the classification results of GF-3 were found to be stable. However, GF-1702 demonstrated that some pixels in residential buildings under construction were assigned to vegetation. In contrast, RADARSAT-2 had a better performance in built-up areas, while ALOS-2 incorrectly assigned some pixels in built-up areas and vegetation to water. For all classification results, with the changes of the image-acquiring situations, the classification accuracy of GF-3 experienced a variation from 75.0% to 91.4%, similar to that of ALOS-2 and RADARSAT-2.

Difference between C-Band and L-Band
Differences exist in the performance of individual C-band or L-band data in its application [33,[37][38][39]. Given the increased ability at longer wavelengths to penetrate vegetation canopies, the

Difference between C-Band and L-Band
Differences exist in the performance of individual C-band or L-band data in its application [33,[37][38][39]. Given the increased ability at longer wavelengths to penetrate vegetation canopies, the pixels in vegetation should be more concentrated at high entropy for the C-band data because of the predominating canopy volume-scattering mechanisms [25]. Table 3 indicated that C-band PolSAR images (GF-1703 and R2-0903) outperformed the L-band image (A2-1603) regarding similarity of backscatter coefficients. Also, as shown in Figure 8, the pixels in vegetation of GF-3 had a similar performance to RADARSAT-2, but different from ALOS-2. In addition, the classification results indicated that vegetation in Wuhan obtained a generally higher product accuracy than that in Beijing (Tables 4 and 5). Since effective surface roughness of a scattering boundary is relative to the wavelength of the incident microwaves, there may be a difference in backscatter levels of C-band and L-band data [40]. Nevertheless, the discrimination between the pixels of built-up areas, vegetation, and water of GF-3 and ALOS-2 was generally similar in Pauli B + Pauli C − Pauli A , H-α, P s -P d , and P v -P c maps.

Incidence Angle Effects
The use of images acquired from different incidence angles sometimes leads to undesired variations in performance [41]. Among the 13 used images, incidence angles changed from 20 • to 45 • . For GF-3, GF-1712 possessed the lowest incidence angle (< 22 • ); GF-1702, GF-1705, GF-1708, and GF-1710 had medium angles (around 37 • ) similar to A2-1504, A2-1601, A2-1604, and R2-0903; and GF-1703 had the highest incidence angle (> 45 • ) similar to R2-1607. As shown in Figure 7c, the pixels in built-up areas with lower incidence angles obtained lower α than other images, and presented a dominant dipole-type scattering mechanism, rather than a double-bounce type. Also, as presented in Figure 8c, the pixels in built-up area did not clearly perform expected double-bounce scattering phenomena. The pixels in GF-3 with higher incidence angles had similar performances as others.

Seasonal Effects
The images used were acquired in different seasons that may lead to variations in the classification results [42]. In Beijing, GF-1710, acquired in the mid-autumn before trees began to shed their leaves, was found to have the best classification results (91.4% OCA, 0.837 OKC). In contrast, GF-1703 and GF-1712 acquired in winter before trees became green achieved lower product accuracies and user accuracies of vegetation, similar with A2-1603 and A2-1612 (Table 4). In Wuhan, little impact from the season on the product accuracy and user accuracy of vegetation was observed (Table 5), resulting from the fact that most of the mountain trees in Wuhan were evergreen.

Difference between Beijing and Wuhan
The difference in land-cover structure between Beijing and Wuhan give rise to the varied histograms of the backscattering coefficients in each polarimetric channel. As almost half of the study area in Wuhan was water, all of the histograms in Wuhan presented double peaks because water has a distinctly lower backscattering power than other land-cover types ( Figures 5 and 6). Because the histograms in Beijing just have a single peak, they had higher similarities between GF-3 and RADARSAT-2 with a shorter Bhattacharyya distance than that in Wuhan (Table 3). In addition, the building samples in urban areas of Beijing and Wuhan exhibited different characteristics, such as size, shape, orientation, and space interval. Nevertheless, most of the pixels in built-up areas of Beijing and Wuhan both presented a larger double-bounce scattering power than surface scattering for all three sensors. The classification results also displayed similar and lower product accuracy in built-up areas of both Beijing and Wuhan (Tables 4 and 5). Due to the mixture and coexistence of built-up structures, vegetation water areas and the heterogeneity of the objects (e.g., residential with gardens) resulted in different backscattering variations within these areas of homogenous land-cover classes [24]. Despite the difference in the vegetation species in Beijing and Wuhan, the two sample areas had similar distribution characteristics in the pixel-based analysis. However, the product accuracy of vegetation in Wuhan was found generally higher than that in Beijing for all three sensors (Tables 4 and 5). The pixel-based analysis of water in GF-3 and RADARSAT-2 obtained similar results between Wuhan and Beijing. However, water in ALOS-2 presented equal surface and double-bounce scattering power in Beijing, while it performed larger surface scattering power than double-bounce in Wuhan (Figures 6 and 8). For all three sensors, the product and user accuracy of water in Beijing was lower than that in Wuhan.

Polarimetric Distortion
Considering the impact of operational bands, incidence angles, and image acquiring situations, the results of the histogram-based, pixel-based, and classification analyses indirectly reflected the polarimetric fidelity of PolSAR data. In general, the histogram-based and classification analyses did not show any signal of distortion impacting GF-3 data. However, the pixel-based analysis indicated that GF-1702 and GF-1703 might suffer polarimetric distortion. Hence, it impacted the decomposition result of lower performance of polarimetric entropy as well as strange model-based decomposition results compared with other images (Figures 7-9). However, the remaining four GF-3 PolSAR images obtained later had no signs showing that any polarimetric distortion imposed significant impacts on the decomposition results. It may be inferred that the sensor of GF-3 operated unsteadily before May 2017 and has been improved. Overall, the experimental results based on a three-hierarchy framework indicated that polarimetric distortions of most GF-3 PolSAR images were similar to ALOS-2 and RADARSAT-2. ALOS-2 and RADARSAT-2 have been widely applied in Earth observations for many years, and their quality is confirmed to meet the users' requirements [11,12,15,33,37,[43][44][45]. The crosstalk accuracy of RADARSAT-2 of −30 dB, the channel imbalance of 0.5 dB in amplitude, and 5 degrees in phase are reported [46]. The accuracy requirement of ALOS-2 is −30 dB in crosstalk, 0.4 dB in channel imbalance amplitude, and 5 degrees in the channel imbalance phase [6]. Generally, previous studies have documented that GF-3 has a similar polarimetric performance to RADARSAT-2 and ALOS-2 using scattering properties and corner reflectors [3,18]. In this study, the polarimetric fidelity of GF-3 PolSAR data was proved at a similar level with that of RADARSAT-2 and ALOS-2, e.g., CTs <−30 dB and CIs <0.5 dB.

Conclusions
In this paper, an innovative, three-hierarchy strategy was proposed to evaluate PolSAR data quality based on the images themselves and their applications, with the support of validation information acquired from ground infrastructure. Its evaluation ability of polarimetric performance was demonstrated by GF-3 experiments using RADARSAT-2 and ALOS-2 as references. The experiments indicated that most of the calibrated GF-3 PolSAR data remained as insignificant polarimetric distortions. However, the performance of GF-3 data obtained before May 2017 showed some differences compared to data obtained after May 2017. This suggests that the system of GF-3 may have been improved around May 2017. Moreover, the results of the present study also proved that the backscattering properties of the target could be reasonably interpreted by decomposition theory using PolSAR images of GF-3; similar performances with that of RADARSAT-2 and ALOS-2 were found. Further, the polarization information of targets included in the pixels of GF-3 is applicable to detecting and distinguishing different land-cover types. Similar abilities of GF-3, ALOS-2, and RADARSAT-2 in land-cover classifications were also found. However, considering the image acquiring situations, incidence angles, operating bands, and many other factors, GF-3 had variable results in the pixel-based analysis and classification, as well as RADARSAT-2 and ALOS-2. Hence, when using PolSAR images in a specific study, the specifications of the data should be cautiously considered to ensure appropriateness.