Synthetic Aperture Radar Radiometric Cross Calibration Based on Distributed Targets

Synthetic aperture radar (SAR) radiometric calibration is an essential step in inferring meaningful physical information about the target. The majority of current calibration processes are directly based on artificial calibrators, but the number of calibrators is always restricted. Cross calibration, which has been widely used for optical and meteorological satellites, performs calibration by another calibrated satellite. Distributed targets are used as the calibration reference, allowing for frequent calibration. However, the potential of introducing cross calibration to SAR satellites has not been confirmed, and two issues must be resolved: 1) the number of calibration targets available for cross calibration is limited to a few known distributed targets; 2) the imaging parameters of calibrated and uncalibrated SARs are different, which may cause errors when cross calibration is applied. In this article, a method for selecting stable distributed targets using time-series stability analysis was presented first. The saline–alkali land and urban areas were selected as the calibration targets with low and high backscattering, respectively. Second, different stable pixel extraction methods were adopted based on the characteristics of different targets. Third, the Oh model was used to correct the scattering difference caused by the incidence angle difference, which greatly improved the accuracy of the backscattering coefficients of the calibration targets. The cross calibration experiments on the same series of satellite (Sentinel-1 A/B) data revealed that the accuracy of cross calibration was comparable to the accuracy of the artificial calibrator-based method, with a difference of less than 0.48 dB (1$\sigma$).

a prerequisite for qualitative and quantitative remote sensing applications [2], [3], [4]. The normalized radar cross section (NRCS) of distributed targets and the radar cross section (RCS) of a corner reflector are two examples of physically meaningful quantities that are converted from image digital numbers (DN) during the radiometric calibration process. [5], [6], [7].
The rapid development of spaceborne SAR technology has created an urgent need for operational high-frequency radiometric calibration [8], [9], [10], [11], [12], [13], [14]. The majority of currently used external radiometric calibration methods rely on the deployment of artificial calibrators as reference targets at the calibration site, as shown in Fig. 1(a). The artificial calibrators must be evenly distributed over tens of kilometers of the SAR image scene. Thus, it is usually very labor-intensive and time-consuming to deploy artificial calibrators in the field. In addition, the operating parameters of active calibrators are usually designed for a single SAR sensor. They are difficult to share with other SAR sensors that operate at drastically different incidence angles or operating frequencies. With the launch of more SAR satellites and constellation satellites, the implementation of this approach becomes unfeasible due to the difficulties in increasing the number of artificial calibrators, thus emphasizing the need for alternative SAR calibration methods [15].
The cross calibration method, which has been widely utilized in optical imaging and meteorological satellites [16], [17], [18], [19], has the ability to address the drawbacks of conventional SAR radiometric calibration, as shown in Fig. 1(b). The already calibrated SAR satellite data are used to perform radiometric calibration of the SAR data to be calibrated. In other words, the backscattering coefficients of the SAR satellite images to be calibrated can be obtained from the backscattering coefficients of the calibrated SAR satellite images. This approach can avoid the deployment of many artificial calibrators and is not constrained by the satellite revisit time.
However, due to the differences in imaging mechanisms between SAR and optical/meteorological sensors, cross calibration cannot be directly applied to SAR calibration, and there are two issues that need to be addressed as follows.
1) The number of reference targets available for cross calibration is limited to a few known distributed targets, such as the Amazon rainforest [20], [21]. From the basic process of SAR calibration, to increase the calibration frequency, the imaging frequency of the calibration targets must be significantly increased. However, the revisit time of the spaceborne SAR sensor is fixed when the orbit is determined. Therefore, it is necessary to increase the number of calibration targets significantly. The Amazon rainforest was proposed for the estimation of the SAR antenna pattern in the 1980 s and is still in use today. D'Aria et al. [22] proposed a radiometric stability monitoring method combining natural permanent scatterers (PS) with artificial calibrators. This PS-based calibration method is characterized by its intrinsic robustness due to the consensus of many natural targets. Yang et al. [23] demonstrate that the median center value of the backscattering coefficients in urban areas is a statistically stable feature. After filtering the urban dataset through the neural network, the final calibration accuracy is within 0.5 dB. These studies illustrate the feasibility of using natural stable targets for radiometric calibration. However, these studies are only for one kind of calibration target, and the number of stable calibration targets needs to be expanded. The rational selection of the targets that exist in nature can effectively increase the number of calibration targets. Stable targets in nature include point-like targets (e.g., parabolic satellite antennas) and uniformly distributed targets (e.g., rainforests, deserts, saline-alkali land). These targets are widely present in nature in large numbers. In this article, a method for selecting stable calibration targets is presented. The calibration targets are selected on the basis of the stability of the time series scattering. Depending on the characteristics of different targets, different stable pixel extraction methods are applied. By these methods, the number of stable pixels can be effectively increased for cross calibration.
2) The imaging parameters (e.g., incidence angle) of calibrated and uncalibrated SARs are different [24], which may result in calibration errors if cross calibration is applied directly. The radiometric calibration process is based on the correlation between the image digital number and the backscattering coefficient. So it is crucial to accurately obtain the backscattering coefficients of the data to be calibrated. Some scholars have studied other calibration targets that can replace the Amazon rainforest for site-free calibration, such as oil spill platforms, oceans, and deserts [25], [26], [27], [28]. However, when analyzing the scattering characteristics of these ground objects, it is found that the scattering characteristics of these ground objects depend on imaging parameters like incidence angle, polarization, and operation frequency. Calibration errors will occur if the backscattering coefficients from calibrated SAR imaging conditions are utilized as the backscattering coefficients from uncalibrated SAR imaging conditions. There are many scattering models, such as the Oh model [29], Dubois model [30], and the advanced integrated equation model (AIEM) [31] that can be used for surface parameter inversion [29], [32], [33]. Based on the calibrated SAR data, we can use the inversion models (such as the Oh model, Dubois model, and AIEM model) to obtain the surface parameters. Assuming the surface parameters remain unchanged, we can use the corres ponding forward models to obtain the backscattering coefficients under the uncalibrated SAR imaging parameters. In this article, based on the selected stable distribution targets, we propose a method based on the Oh model [29], [34] to correct the scattering difference caused by the incidence angle difference between calibrated SAR and uncalibrated SAR, which can effectively improve the accuracy of the cross calibration. In general, the main contributions of this article are as follows. 1) To address the problem of insufficient stable calibration targets, stability analysis of the scattering characteristics of various ground objects is carried out, and the salinealkali and urban areas are selected as stable reference distributed targets. The median value is selected to extract the backscattering coefficients of the nonuniform calibration targets, which effectively reduces the errors caused by extreme points or registration of nonuniform calibration targets.
2) Regarding the difference in incidence angle between calibrated and uncalibrated SARs, the Oh scattering model is introduced into the cross calibration process. The difference in incidence angle is corrected by the Oh model, which effectively improves the calibration accuracy. The rest of this article is organized as follows. Section II briefly reviews some related work. Section III presents detailed methods for stable target selection, pixel extraction, and incidence angle correction. Section IV presents the experimental results on the Sentinel-1 A/B dataset. Finally, Section V concludes this article.

A. Basic Principles of SAR Calibration
The purpose of SAR radiometric calibration is to obtain the scattering characteristics (radar cross section or backscattering coefficient) of the target. Through the relationship between the pixel power of the target and the backscattering coefficient, the total transfer function of the SAR system can be obtained. The formula is as follows: where P r is the target pixel power, H is the total transfer function, and σ is the backscattering coefficient. SAR radiation calibration is the process of accurately calculating the total transfer function H, but there are many sources of error throughout the SAR data acquisition chain. The radiometric calibration formula that accounts for changes in various system parameters and imaging geometry is as follows: where DN represents the pixel value in the ith row and jth column, K is the absolute calibration coefficient, which is a constant. θ is the incidence angle of the distance to each pixel in the SAR image. θ ref is the incidence angle of the reference point, such as artificial corner reflectors and active calibrators. R is the slant distance value of each pixel point in the range direction. R ref is the slant distance value of the reference point. G(θ) 2 is the gain value of the antenna. It is a function of the incidence angle in the distance direction. Artificial corner reflectors and active calibrators are traditionally used as calibration reference targets. However, the availability of these artificial targets is often restricted, which limits the calibration frequency. In recent years, natural targets, which are abundant in nature, have been gradually used as calibration reference targets.

B. Radiometric Calibration Based on Natural Point Targets
Point targets in nature are Earth surface targets with strong radar scattering responses that are comparable to the responses of artificial calibrators in SAR images, both of which have a high signal cluster ratio. They are initially studied in the permanent scatterer interferometric synthetic aperture radar (PS-InSAR) technique and then introduced to the SAR radiometric calibration. In 1999, Usai et al. [35] found that exposed rocks and artificial buildings (e.g., roads and houses) in SAR images showed high coherence over a long time scale. Their coherence was studied over the entire time series and could reverse surface deformation. On this basis, Ferretti et al. [36] proposed the PS-InSAR technique. The basic idea is to filter the points on the time-series SAR images that are less affected by decoherence, extract and solve the differential phase model of these points, and extract the deformation phase. In 2010, D'Aria et al. [22] proposed an SAR radiometric calibration method based on an artificial calibrator and permanent scatterers. The permanent scatters enable continuous radiometric calibration and avoid the time-consuming and laborious problems of using solely artificial calibrators. In 2014, Guccione et al. [37] extracted point targets from natural permanent scatterers and applied them to SAR long-term relative radiometric calibration and antenna pointing tests. It achieved an accuracy of 1 dB. In 2018, Guccione et al. [38] also used stable point targets for L/P band SAR radiometric calibration and antenna pattern estimation. In 2019, Du et al. [39] verified the feasibility of using parabolic antennas for spaceborne P-band polarimetric SAR calibration.
Although radiometric calibration based on natural point targets can solve the problem of insufficient artificial calibrators, Biancardi et al. [40] point out two limitations of such methods. First, the image scene must contain point targets, such as buildings and rocks, requiring approximate imaging conditions, which are difficult to apply for cross calibration between different SAR satellites. Second, to extract permanent scatterers, long-time interferometric SAR data are needed, and the interferometric processing is very complex. In addition, the side lobes of point targets can also lead to radiometric errors and require additional corrections [15], [41], [42].

C. Radiometric Calibration Based on Natural Distributed Targets
Distributed targets in nature are large areas of the earth's surface with similar scattering. Typical distributed targets include farmland, grassland, bare land, etc. In 1985, Moore et al. proposed using the Amazon rainforest to estimate the antenna pattern of spaceborne SAR [43]. Its primary purpose was to determine the direction of the SIR-B two-way pitch antenna from the SIR-B digital image of the Amazon rainforest. The Amazon rainforest was used not only because it is the world's largest, flattest, most stable distributed target, but also because its backscattering coefficient hardly varies with the incidence angle due to the volume scattering properties of the rainforest. However, the number of tropical rainforests used for calibration is limited; for example, the Amazon rainforest and several rainforests in Canada. It is difficult to increase the frequency of SAR imaging with calibration targets. Other distributed uniform targets in nature (e.g., crops, bare land, and deserts) are suitable for calibration under specific conditions [44]. In 2018, Yang et al. [23] analyzed C-band SAR images and found that the median center of gravity of the backscattering coefficients in urban areas had good temporal stability. Its accuracy reaches 0.21 dB, which is equivalent to the nominal absolute radiometric accuracy. Zakharov et al. [45] proposed monitoring the radiometric stability of the ALOS/PALSAR and PALSAR-2 satellites using the Caspian Sea oil platform. The results show that the SAR sensor has a backscattering stability of 0.6 dB per year. The stability of distributed uniform targets, such as Chile's Atacama Desert, the Amazon rainforest, and the permanent ice sheet in Antarctica, is also found to be comparable to that of artificial corner reflectors. The stability of the radiometric calibration of Lake Vostok in Antarctica even reaches 0.19 dB in 3.5 years.
Although these studies show that radiometric calibration using distributed uniform targets is possible, the conclusions above were reached using the same SAR imaging parameters. The backscattering coefficients of most natural targets tend to vary with imaging parameters like the incidence angle. Therefore, the natural targets are difficult to directly apply for cross calibration, and their backscattering coefficients must be corrected according to different imaging parameters. For point targets, it is difficult to estimate and correct for this variation, but distributed targets can be estimated and corrected based on the target's backscattering model. In this article, we begin by selecting distributed uniform targets with stable backscattering coefficients. The stability of the backscattering coefficient is incorporated into the calibration requirements for the calibration targets to construct an SAR calibration target database that meets the needs of radiometric cross calibration.

D. SAR Cross Calibration
Cross calibration of remote sensing sensors refers to using the data from calibrated sensors to calibrate the data to be calibrated. Cross calibrations of optical imaging satellites and meteorological satellites have been extensively studied [16], [17], [18], [19], while there are few studies similar to or related to SAR cross calibration in the past decades.
Zink et al. [46] reported a cross-calibration flight experiment with airborne E-SAR/DC-8 SAR sensors in 1989. Forty-two trihedral corner reflectors, four dihedral corner reflectors, an active calibrator, and a receiver were deployed as calibration targets. The experimental results showed that cross calibration was feasible when the reflector pointing was precisely adjusted, the antenna pattern was known, and the flight trajectory and platform attitude data were known. The backscattering coefficients of the distributed targets are relatively consistent, with differences of less than 2 dB. The authors noted that this difference was due to differences in incidence angle and soil moisture. In 2014, Zakharov et al. [47] used a parabolic antenna as a calibration target to cross-compare the radiometric stability of the ERS AMI scatterometer and the ENVISAT ASAR. The radiometric stability of the former is 1 dB higher than the latter's. It is worth mentioning that this work uses an artificial calibrator precisely pointed at the SAR sensor. In practice, it is challenging to implement cross calibration for natural calibration targets whose orientation cannot be controlled and whose directional scattering characteristics are unknown. III. METHODOLOGY Fig. 3 shows the overall flow chart of the proposed SAR cross radiometric calibration method based on distributed targets. It mainly consists of four parts, namely, the selection of stable targets, the extraction of stable pixels, the correction of SAR imaging parameter differences, and the calculation of the calibration coefficient.

A. Selection of Stable Distributed Targets
Cross calibration is based on imaging the same target scene with calibrated and uncalibrated SAR. In this approach, it is assumed that the target scene maintains the same physical characteristics. In other words, the calibration targets for cross calibration need to meet the stability requirements for the backscattering coefficients.
The backscattering coefficient is related to soil moisture, surface roughness, dielectric constant, incidence angle, etc. It can be expressed as where θ is the incidence angle related to the SAR imaging geometry, λ is wavelength, and P represents the polarization related to the SAR system parameters. m v , s, and represent soil moisture, surface root mean square height, and dielectric constant, respectively. They are associated with the physical characteristics of the target. Cross calibration requires that the physical characteristics of the selected calibration targets remain relatively stable. In other words, the soil moisture (m v ), surface root mean square height (s), and dielectric constant ( ) should remain stable during the cross calibration. They are essential for selecting stable targets and improving the cross calibration accuracy. The stability of the physical characteristics corresponds to the time-series stability of the backscattering coefficients of the calibration target when the incidence angle (θ) and the polarization of the sensor (P ) remain constant (same satellite and same incidence angle). The backscattering coefficient is determined by two factors, i.e., the satellite system and the observed ground objects. If the system's parameters, such as its operating frequency, the incidence angle, and other parameters, are the same, the only factor affecting the changes in backscattering coefficient will be the characteristics of the ground objects [such as soil moisture (m v ), surface root mean square height (s), and dielectric constant ( )]. In this instance, the higher the stability of the backscattering coefficient, the more stable the characteristics of the ground objects.
In this article, the selection of stable distributed targets is carried out by investigating the time-series stability of the backscattering coefficients of different candidate distributed target scene categories. First, two sets of SAR data from the same satellite are obtained, with different imaging times but the same flight direction and incidence angle. Then the standard deviation STD of the backscattering coefficients between the two sets of data is calculated, since the standard deviation measures the dispersion of the mean of a set of data. The smaller the standard deviation, the more stable the backscattering coefficients are. A threshold is set in order to select the target. The target scene category should not be utilized as cross calibration reference targets if the standard deviation is higher than the threshold, but they can be utilized as reference targets if the standard deviation is lower than the threshold. The threshold set in this experiment is 1 dB.
The detailed process of selecting a stable distributed target is shown in Algorithm 1.

B. Stable Pixel Extraction
Not all pixels within the selected distributed calibration target are used for cross calibration. Stable pixels are extracted in the following two cases: 1) If the selected stable target is a uniform area with a single type of ground feature, the mean value within a block is directly taken as the stable DN/backscattering coefficient. 2) If the environment is complicated and there are many extreme points in the selected stable target, the extreme points need to be removed. For example, some scholars use urban areas as a source to obtain stable high backscattering coefficients [48]. However, urban areas have a complex environment with the coexistence of roads, trees, bridges, and houses. To remove the extreme points, this article adopts three methods for DN/backscattering coefficient extraction: 1) the median method, 2) the mean method, and 3) the high-frequency mean method. The high-frequency mean is calculated as follows: Divide the minimum value to the maximum value of the backscattering coefficients into ten intervals, calculate the frequency of the backscattering coefficients in each interval, and calculate the average value of the backscattering coefficients in the interval, where the frequency is greater than 10%. The data extraction by the method can also achieve a certain level of noise filtering to avoid errors caused by noise.
In cross calibration, to ensure the stability of the backscattering coefficient, it is necessary to ensure that the physical parameters, including soil moisture (m v ), do not vary much. Since soil moisture (m v ) is more susceptible during the rainy season compared to other physical parameters, the stability of the selected target is ensured by analyzing the time-series stability of the backscattering coefficient. In this article, if the target is not bare ground (such as urban), the effect of the rainy season on the backscattering coefficient is not analyzed. If the target is bare (such as saline-alkali land), we can examine whether the rainy season affects the backscattering coefficient by adding the time series variation of the backscattering coefficient with and without the rainy season data. If the backscattering coefficient varies significantly between the rainy season and other seasons, the data from the rainy season data can be removed from the data of this calibration target. Whether the rainy season data need to be deleted depends on the acquisition time of the two datasets. If one dataset is in the rainy season and the other is not, calibration should not be performed. This operation avoids the effect of changes in soil moisture on the backscattering coefficient.
The detailed process of extracting stable pixels is shown in Algorithm 2.

C. Correction of Imaging Parameter Differences
SAR cross calibration is using the calibrated data to calibrate the uncalibrated data. It is necessary to consider the variation of the backscattering coefficients of the calibration targets caused by differences in the incidence angle, radar wavelength, polarization, terrain slope, etc. The problem is mathematically described as where f denotes the backscattering model of the calibration target; σ 1 and σ 2 are the backscattering coefficients of the target when illuminated by the calibrated SAR and the SAR to be calibrated, respectively. When stable reference targets and pixels are selected and extracted using the methods in Section III-A and III-B, it can be considered that s 1 = s 2 , m v1 = m v2 , 1 = 2 . Then, we can use the calibrated SAR data to obtain the surface parameters by the inversion models, assuming that the surface parameters remain unchanged, and then obtain the backscattering coefficients under the uncalibrated SAR imaging parameters by the corresponding forward models arg min s,m, The process for obtaining the backscattering coefficient σ 2 is shown in Fig. 4. The Oh model [29] is used as the surface scattering model f in this article. It is an empirical model accounting for the SAR data of bare ground with different surface roughness, soil moisture, wave bands, incidence angles, and polarizations where m v represents soil moisture, θ represents incidence angle, k represents wave number, s represents surface root mean square height, p represents copolarization ratio, and q represents crosspolarization ratio. The Oh model is chosen because it can be applied to SAR data with a wide range of surface roughness, X-, C-, and L-bands, and an incidence angle between 10 and 70. If all parameters except the incidence angle are the same or approximate, the Oh model can be used to derive the backscattering coefficients to be calibrated at different incidence angles. According to (6), the relationship between the calibrated SAR data and the SAR data to be calibrated in VH polarization can be expressed as On the other hand, according to (8), the calibrated SAR data and the SAR data to be calibrated can be expressed as By dividing (10) with (11) and combining (9), we can obtain Rewriting (12), the relationship between the calibrated SAR data and the SAR data to be calibrated in VV polarization can be expressed as The derived backscattering coefficients in (9) and (13) can then be used to calibrate the SAR satellite data to be calibrated.

D. Calculation of Calibration Coefficient
In this article, calibration targets with stable high and low backscattering coefficients are used for radiometric calibration. The model corrects the backscattering coefficient of the calibrated SAR satellite to obtain the backscattering coefficient of the SAR satellite to be calibrated. Two points are used to determine a straight line. The high and low backscattering coefficients are combined to fit the calibration coefficients in the radiometric calibration process. The calibration equation is as follows: where σ is the backscattering coefficient of the SAR satellite data to be calibrated obtained by this method. The DN value is the digital quantization value of the corresponding pixel in the SAR satellite data to be calibrated. m is the calibration coefficient of the SAR satellite data to be calibrated. n is the calibration offset value of the SAR satellite data to be calibrated.

IV. EXPERIMENT AND DISCUSSION
To quantitatively analyze the accuracy and effectiveness of the proposed SAR cross calibration method based on distributed targets, the Sentinel-1 (S-1) SAR dataset is used. The time series stability of different reference targets, the stability of the urban target under different extraction methods, the error of the backscattering coefficient caused by the incidence angle, and the stability of the backscattering coefficient of urban and saline-alkali land in the 12-month SAR data are presented. Finally, based on the above research, in the Dunhuang area, Sentinel-1B data are calibrated based on Sentinel-1 A data.

A. Study Area and Experiment Dataset
The study area of this article includes the urban area of Beijing, China; the saline-alkali land and urban area of Dunhuang, China; the salt lake area of Railroad Valley Playa, USA; the desert area of Hotan, China; and the grassland area of Gongger, China. The locations of the datasets are shown in Fig. 5. The selection of these study areas is according to the well-known optical calibration sites. The CEOS Working Group on Calibration and Validation (WGCV) provides calibration sites with good uniformity, rich types, and clear latitude and longitude data. They are good candidate sites for SAR cross-calibration. But not all optical calibration sites can be used for SAR cross calibration. We need to select from the optical calibration sites.
The Sentinel-1 SAR data are then obtained for these study areas. Sentinel-1 (S-1) is the European flagship SAR mission, initiated by the European Space Agency and the European Commission. It is a constellation of two C-band twin satellites that provide backscattering on a global scale [49]. The Sentinel-1 Cband-based imaging system uses four imaging modes, including stripes (SM), wide interference (IW), ultra-wide (EW), and wave mode (WV). Sentinel-1 has a very high radiometric accuracy of 1 dB. This article uses Sentinel-1 data in dual polarization IW imaging mode.
These data are used to analyze the scattering stability and extract the backscattering coefficients. The initial data need to be preprocessed. Data preprocessing includes orbit correction, radiometric correction, and terrain correction. The state information of the orbit is refined, and the distance distortion is improved. It establishes the relationship between the digital values of the images and the values of the backscattering coefficients. Data preprocessing is performed using the sentinel application platform (SNAP). The sentinel application platform is the basic platform of all Sentinel toolboxes and is a desktop CS platform. It has extensibility, portability, and modular interface.

B. Stable Low Backscattering Target Selection and Pixel Extraction
The process of selecting stable targets with low backscattering coefficients is performed among candidate distributed targets of grassland, salt lake, desert, and saline-alkali land. All four calibration targets have backscattering coefficients below −10 dBm 2 /m 2 . Their time series stability is analyzed using data from Sentinel-1 A. Sentinel-1 A calibrated data in March and June of the same year are used to analyze the variation of the backscattering coefficient. The SAR imaging time difference is three months (June and March) because it spans a season.
The specific information of the Sentinel-1 A dataset is shown in Table I. For more precise registration, the data sizes are initially cropped differently. After the registration is completed, all data sizes become 90×90 pixels. Fig. 6 shows the variation of the backscattering coefficients for the different calibration targets between March and June. It is clear that the desert and saline-alkali land are stable because their data distributions are consistent with the y = x line. Table II shows a comparison of the standard deviation of the backscattering coefficients between March and June. The smaller the standard deviation, the more minor the variation of the backscattering coefficients with time, and the more stable the scattering characteristics of the calibration target.
According to Fig. 6 and Table II, the saline-alkali land has the most stable backscattering coefficients, and its standard deviations the VV and VH polarization are lower than those of other targets. The saline-alkali land is a type of salt accumulation, which affects crop growth normally. According to incomplete statistics from the United Nations Educational, Scientific and Cultural Organization (UNESCO) and the Food and Agriculture Organization of the United Nations (FAO), the area of saline-alkali land in the world is 954.38 million hectares, widely dispersed across more than 30 countries on six continents. It is mainly distributed in India, Pakistan, Russia, South Africa, and other countries. It can be noticed from Table II that the backscattering coefficient in the VH polarization is too low, close to the noise equivalent backscattering coefficient (NESZ) of the Sentinel-1 satellite. In this article, considering the influence of noise on the experimental results, all subsequent experiments are not carried out on the VH polarization but only on the VV polarization. Thereafter, saline-alkali land is used as a stable target with a low backscattering coefficient.
Since the selected saline-alkali area is uniform and has no other surface vegetation cover, taking the average of 10×10 pixel blocks is directly used to extract the backscattering coefficients to avoid errors caused by image registration during cross calibration. Because the saline-alkali land is bare, the backscattering coefficients are easily affected during the rainy season. Fig. 7 shows the time series variation of the mean backscattering coefficients under VV polarization for Dunhuang saline-alkali soil slices in 2019. The average data error for 12 scenes throughout the year is 1.53 dB. Fig. 8 shows the same data, but excludes the rainy season. After excluding the influence of the rainy season, the average error of the data for nine scenes is reduced to 0.62 dB.

C. Stable High Backscattering Target Selection and Pixel Extraction
In this article, the urban area is chosen directly as the stable target for a high backscattering coefficient. Due to the complex distribution of features in the actual urban area, the backscattering coefficients are not uniform. This article studies the influence of different extraction methods on the stability of the backscattering coefficients. It includes the stability of the median, high-frequency mean, and mean method of extracting backscattering coefficients in time series. Then we reselect the data to experiment with the extraction method. Differences between different extraction methods are shown in Fig. 9. The error is 0.82 dB when extracted by the mean method, 0.25 dB  when extracted by the high-frequency mean method, and 0.15 dB when extracted by the median method. Therefore, the median method has been adopted.

D. Correction of Incidence Angle Difference
To verify the effect of incidence angle changes on the scattering characteristics of urban areas, this article selects two sets  of data from three positions in the three cities of San Francisco, Chengdu, and Beijing. The incidence angle between every two sets of data differs by approximately 10.
The variation of scattering characteristics with incidence angle is slight in urban areas, and the error (standard deviation) is 0.4 dB, and no correction for the difference in incidence angle is applied to the urban data in the experiment. The specific experimental results are shown in Fig. 10.  The variation of scattering characteristics with incidence angle is significant in the saline-alkali area. The error (standard deviation) is 1.82 dB before correction and 0.81 dB after the Oh model-based correction using (13), as shown in Fig. 11.

E. Calculation of Calibration Coefficient
The actual data from Dunhuang is calibrated. This article calibrates Sentinel-1B data (Data 1) through Sentinel-1 A data (Data2, Data3). The backscattering coefficients of Data 1 are obtained from the backscattering coefficients of Data 2 and Data 3. The calibration dataset information is shown in Table III. Since the stable backscattering coefficients are extracted by the median method in the urban area, they are not sensitive to incidence angle, so the incidence angle is not corrected. In the saline-alkali area, the incidence angle of the calibrated Sentinel-1 A data and the Sentinel-1B data to be calibrated differs by about 10, and the incidence angle difference is corrected by the Oh model. After the data are prepared, the actual cross-calibration experiment is performed, and the backscattering coefficient of Sentinel-1B to be calibrated is obtained from the calibrated Sentinel-1 A data.
In the experiment, the high and low backscattering coefficients are combined to fit the calibration coefficients in the radiometric calibration formula. A total of 18 data points in urban areas and saline-alkali land were used, with 9 data points for each target.
For comparison, the median, mean, and high-frequency mean methods are used to extract urban data for radiometric calibration. The fitted calibration curves are shown in Fig. 12, where the calibration formula by the median extraction method is as follows: σ = 3.715 × 10 −6 · DN 2 − 0.00187.

F. Verification of Calibration Accuracy
The calibration curves (Fig. 12) fitted by different extraction methods are used to calibrate other positions' pixels in Data 1. The specific locations of the four calibration regions selected for this article are shown in Fig. 13. The data size of each region is 90×90 pixels. These four calibration regions are mainly deserts and are distributed in four different directions. They are selected following these requirements: 1) The surface must be uniform; 2) the backscattering coefficient must be greater than the noise equivalent backscattering coefficient; and 3) they must be as dispersed as possible.
In this article, the comparison between the backscattering coefficients calculated by the cross-calibration method and the backscattering coefficients actually obtained by the lookup table (LUT) [50] in the four selected regions is shown in Fig. 14. The actual backscattering coefficient errors (standard deviation) of the four regions selected in this article are shown in Table IV.
The final accuracy of different extraction methods is obtained from the average of the actual backscattering coefficient errors of these four regions. The final calibration accuracy of the urban data extracted by the median method is 0.48 dB. The final calibration accuracy of the urban data extracted by the mean value method is 1.25 dB. The final calibration accuracy of the urban data extracted by the high-frequency mean method is 0.99 dB. The median method is the most stable for extracting urban data. To further verify our experimental results, we performed the same calibration experiments on Sentinel-1 A data. Dataset information is shown in Table V.
The same satellite calibration results are shown in Table VI. Fourteen scenes of saline-alkali land were used in our analytical experiments. Four saline-alkali land scenes were used for crosscalibration of the same satellite and the same series of satellites. The above experiments verify that the method in this article has good applicability in the cross-calibration of the same satellite and a series of satellites.

V. CONCLUSION
In this article, we propose a distributed target-based SAR cross radiometric calibration method to solve the problem of a limited number of artificial calibrators. The scattering in the VV polarization mode is studied. The relatively stable backscattering coefficients are obtained through the urban median extraction method and the Oh model calibration of saline-alkali land. Cross calibration is carried out through urban areas and salinealkali land. According to the findings, the accuracy of cross calibration could reach a similar accuracy to that of the artificial calibrator-based method, and the difference was within 0.48 dB (1σ). However, the method in this article is only used for the same series of SAR satellites. For different series of satellites, their satellite parameters are quite different, which may cause significant errors in the experimental results. In future work, we will study the generality of this method and its applicability to areas other than the Dunhuang site and Sentinel-1 satellite.