Extraction and Analysis of Radar Scatterer Attributes for PAZ SAR by Combining Time Series InSAR, PolSAR, and Land Use Measurements

Abstract

Extracting meaningful attributes of radar scatterers from SAR images, PAZ in our case, facilitates a better understanding of SAR data and physical interpretation of deformation processes. The attribute categories and attribute extraction method are not yet thoroughly investigated. Therefore, this study recognizes three attribute categories: geometric, physical, and land-use attributes, and aims to design a new scheme to extract these attributes of every coherent radar scatterer. Specifically, we propose to obtain geometric information and its dynamics over time of the radar scatterers using time series InSAR (interferometric SAR) techniques, with SAR images in HH and VV separately. As all InSAR observations are relative in time and space, we convert the radar scatterers in HH and VV to a common reference system by applying a spatial reference alignment method. Regarding the physical attributes of the radar scatterers, we first employ a Random Forest classification method to categorize scatterers in terms of scattering mechanisms (including surface, low-, high-volume, and double bounce scattering), and then assign the scattering mechanism to every radar scatterer. We propose using a land-use product (i.e., TOP10NL data for our case) to create reliable labeled samples for training and validation. In addition, the radar scatterers can inherit land-use attributes from the TOP10NL data. We demonstrate this new scheme with 30 Spanish PAZ SAR images in HH and VV acquired between 2019 and 2021, covering an area in the province of Friesland, the Netherlands, and analyze the extracted attributes for data and deformation interpretation.

1. Introduction

Synthetic Aperture Radar (SAR) transmits its electromagnetic wave and then records the echo signal reflected back after interacting with the earth’s surface [1]. As the electromagnetic wave is synthetic, the orientation, particularly of the electric field vector, namely polarization, is predefined. For linear polarization, VV and VH polarization measure the proportion of vertically transmitted waves that return vertically and horizontally to the sensor, respectively. HH and HV polarization measure the proportion of horizontally transmitted waves that return horizontally and vertically. Using single-polarization SAR data, detecting, mapping, and monitoring surface movements with Interferometric SAR (InSAR) [2] and Time Series InSAR (TSInSAR) [3] can be realized [4,5,6,7,8,9]. Several SAR satellite missions possess multi-polarization channels. For instance, the PAZ X-band satellite operates in both co-polarimetric channels HH and VV [10,11], the Sentinel-1 C-band mission has either HH and HV or VV and VH [12], and the Radarsat-2 C-band and ALOS-1 and 2 L-band satellites have all four polarimetric channels [13,14]. When dual-polarimetric (HH and VV, VV and VH, or HH and HV), or quad-polarimetric (VV, VH, HH, and HV) data are available, land use and land cover (LULC) classification, and scatterers’ characterization using Polarimetric SAR (PolSAR) [15], and InSAR measurement augmentation [16] can be accomplished [17,18,19]. It is evident that both TSInSAR and PolSAR techniques have the unique merit of delivering useful information on surface targets from different perspectives. Note that surface targets are observed as radar scatterers in SAR imagery. Yet, the study on combining all information from TSInSAR and PolSAR to extract meaningful attributes of radar scatterers, to obtain a better insight into SAR data and geo-processes such as surface movements, when having dual- or quad-pol SAR data, is still limited.

Theoretically, TSInSAR can circumvent spatial and temporal decorrelation and atmospheric heterogeneity, and primarily deliver the geometric measurements, i.e., position and deformation of constantly and temporarily coherent radar scatterers [9,20,21,22,23,24]. An issue in the TSInSAR post processing, when integrating the deformations from SAR data with different polarizations, is the spatial reference misalignment. Usually the spatial reference point or area can be selected based on its (amplitude) stability over time [25] or ground truth data [26], then all InSAR measurements are with respect to this selected reference. In the case that the external measurements and ground truth data are unavailable, the reference error is introduced to all the InSAR measurements if this selected reference point or area is unstable in reality. Therefore, a robust spatial reference selection method that can be well suited for error mitigation at a global scale, should be developed.

The radar scatterer classification in terms of scattering mechanisms is one of the standard outputs of PolSAR data analysis. The scattering mechanisms, such as surface, volume, and double bounce scattering, represent the physical attributes of radar scatterers. The polarimetric SAR decomposition [27] and machine/deep learning [28,29] are the methods for PolSAR image classification. The importance of the polarimetric features that are used in machine/deep learning differs, and the step of creating labeled samples (i.e., a group of radar scatterers) for training and validation is vital [30]. Manually annotating the reference data mainly based on visual interpretation of SAR images is time-consuming and inaccurate sometimes. To improve the quality and efficiency of adding labels to data, particularly when having an external land-use data source, the methods for incorporating such data into labeled sample creation and assigning land-use attributes to radar scatterers need to be designed.

This paper is dedicated to addressing the issues raised above. In a nutshell, we propose and demonstrate appropriate methods and schemes by combining TSInSAR and PolSAR and land-use observations in order to extract meaningful attributes (i.e., geometric, physical, and land-use attributes) of radar scatterers in multi-pol SAR images, HH and VV PAZ SAR images for our case. This paper is organized as follows. Section 2 presents the methods of coherent radar scatterer selection, spatial reference selection, and radar scatterer classification in terms of scattering mechanisms, and land use attribute extraction based on a land use product (i.e., TOP10NL for our study), to extract radar scatterer attributes. Section 3 presents all data used and geographic information of the test site. The results of the test site are presented in Section 4, followed by discussion and conclusions in Section 5 and Section 6, respectively.

2. Methods

This section presents the methods to extract radar scatterer attributes. These attributes can be described from the geometric, physical, and land-use aspects. In other words, radar scatterer attributes can be categorized into geometric, physical, and land-use attributes. The flowchart for using co-polarimetric SAR data processing to extract these attributes is depicted in Figure 1, and detailed in the following subsections.

2.1. Geometric Attribute Extraction

The geometric attributes of radar scatterers contain the information of radar scatterers’ position and position change over time (i.e., deformation time series), which can be obtained from TSInSAR. The processing procedure is shown by the red arrows in Figure 1. Since TSInSAR provides this information merely for coherent radar scatterers, recognizing coherent radar scatterers is the first step of radar scatterers’ geometric attribute extraction.

2.1.1. Coherent Radar Scatterer Selection in Time Series InSAR

A standard hybrid time series InSAR (TSInSAR) approach provides the temporal measurements of InSAR measurement points (IMPs), i.e., constantly coherent radar scatterers (CCSs) and temporarily coherent radar scatterers (TCSs), embracing both pointwise and distributed scatterers, abbreviated to PSs and DSs. In practice, it can first parallelly run two TSInSAR processing schemes, the single-master single-look (i.e., PSI, persistent scatterer interferometry [3]) and equivalent single-master small-baseline with an adaptive filter, and then integrate all measurements from both schemes. Here, master and slave are two SAR images used for interferogram generation. The slave is co-registered to the master reference grid. A widely used method to select CCS and TCS candidates is defined as [3,21]

$$D_A=\frac{{\sigma }_A}{{\mu }_a},A\in [a,\Delta a],$$

(1)

where $D_A$ is referred to as amplitude dispersion when $A=a$ for pointwise CCS selection and amplitude difference dispersion when $A=\Delta a$ for TCS selection. $a,{\mu }_a,{\sigma }_a,\Delta a,{\sigma }_{\Delta a}$ separately represent the SLC amplitude, mean value of amplitude time series, standard deviation of amplitude time series, amplitude difference between masters and slaves, and standard deviation of amplitude difference. After the pointwise CCS selection, the equivalent small-baseline phases for these pointwise CCSs are extracted from single-master single-look interferograms. Then, the small-baseline phases for CCSs and TCSs are aggregated and unwrapped using a conventional SBAS (small baseline subset) algorithm [31,32,33].

2.1.2. Spatial Reference Selection

InSAR measurements, i.e., deformations, are usually relative w.r.t. predefined spatial reference (point). Only when the spatial reference is aligned to an absolute geographic coordinate system using ground truth data such as GPS and leveling, can we obtain absolute InSAR measurements. The spatial reference can be either a single point or an area. A radar scatterer can be considered as a spatial reference point only when it is a constantly coherent radar scatterer with extremely low variation in phase and amplitude time series, implicitly meaning a zero-deformation time series, cf. DePSI [34] and SARScape [35]. To determine a spatial reference area, we opt for a locally stable region or entire study site, in which case average deformation time series within this reference area is considered as the zero-deformation reference, cf. StaMPS [21]. The deformations of the spatial reference area can be calculated by

$${\widehat{y}}_{\mathrm{ref}}^k=\frac{1}{N}\sum _{q=1}^N{\widehat{y}}_q^k,$$

(2)

where ${\widehat{y}}_{\mathrm{ref}}^k$ is the average deformation at kth acquisition, N is the total number of InSAR measurement points (IMPs, CCSs, and TCSs ∈ IMPs), and ${\widehat{y}}_q^k$ is the measurement of point q within the selected reference area, at kth acquisition.

Considering the difference in scattering mechanisms, spatial density, and distribution of IMPs for HH and VV, neither of these two ways to determine a common reference is optimal. Therefore, we propose to take the average deformation time series of a group of nearly stable InSAR measurement siblings for HH and VV. Specifically, when having $N_1$ and $N_2$ IMPs, respectively, for HH and VV, the size of such a group of siblings is $N_s\le \min [N_1,N_2]$. Every sibling pair is recognized based on the geometric and geophysical matching criteria as follows:

• Spatial closeness and akin scattering mechanism [25];
• Nuance in deformation time series;
• Nearly stable temporal behavior.

Then, the common reference for HH and VV has ${\widehat{y}}_{\mathrm{ref}}^k=\frac{1}{N_s}{\sum }_{j=1}^{N_s}{\widehat{y}}_j^k,(\forall k)$ deformation over time, which is subtracted from all IMP deformations. In doing so, the IMP temporal measurements in HH and VV can be blended. The advantage of combining and aligning IMPs in both HH and VV is that we can increase the IMP spatial density, and use the blended deformation time series of IMPs in HH and VV to better describe the geometric dynamics over time of any detected ground targets. Note that we identify and exclude IMPs with potential phase unwrapping errors by setting the threshold of the deformation gradient.

2.2. Physical and Land-Use Attribute Extraction

Physical attributes are described by scattering mechanisms mainly covering surface, volume, and double bounce scattering. Benefiting from distinct polarimetric features in co-polarimetric SAR (HH and VV), we can classify IMPs in terms of different scattering mechanisms and then assign the scattering mechanism to every IMP. The land-use attributes of IMPs can be inherited from land-use products. The procedure for obtaining physical and land-use attributes is shown by the blue and green arrows separately in Figure 1.

2.2.1. Speckle Noise Removal

To improve the quality of the polarimetric features, a spatio-temporal filter method is recommended to mitigate speckle noise in SAR images. The filter method we employ is customized upon a multi-temporal speckle filter [36]. In the case of applying this filter on SAR amplitude, for a pixel at position ($rg,az$), with the amplitude value $A^k(rg,az)$, at the kth acquisition, its resultant amplitude value after the filter, denoted by ${\tilde{A}}^k(rg,az)$, can be expressed as

$${\tilde{A}}^k(rg,az)={[{[A^k(rg,az)]}_{\mathrm{filter}\_\mathrm{space}}]}_{\mathrm{filter}\_\mathrm{time}}=S^k(rg,az)T(rg,az),$$

(3)

where the spatial filter output is $S^k(rg,az)={\left[A^k(rg,az)\right]}_{\mathrm{filter}\_\mathrm{space}}$. The Boxcar filter, Lee-sigma filter [37], and IDAN (Intensity Driven Adaptive Neighborhood) filter [38] are examples of spatial filter methods. The subscript $\mathrm{filter}\_\mathrm{space}$ and $\mathrm{filter}\_\mathrm{time}$ indicate the spatial and temporal filter processes, respectively. The temporal filter output $T(rg,az)$ is defined as

$$T(rg,az)=\frac{1}{m+1}\sum _{i=1}^{m+1}\frac{A^i(rg,az)}{{\left[A^i(rg,az)\right]}_{\mathrm{filter}\_\mathrm{space}}},$$

(4)

where $m+1$ is the total number of SAR image acquisitions, $A^i(rg,az)$ represents SAR amplitude at position ($rg,az$) of ith acquisition, $i\in [1,m+1]$. Having the spatial filter output as a divisor in Equation (4), we can normalize the amplitude values. As an example, ref. [39] used this filter to generate the filtered normalized radar cross-section maps prior to the Cloude–Pottier polarimetric decomposition for Sentinel-1 SAR datasets.

2.2.2. Polarimetric Features

The representative (incoherent) polarimetric features, i.e., real and imaginary parts of the co-polarization cross product, summation, difference, and ratio of the co-polarization intensities, sigma naughts in HH and VV, and entropy, can be employed for IMP physical characterization.

The co-polarization cross product is the complex multiplication of two SAR images in HH and VV ($S_{HH}$ and $S_{VV}$), expressed as

$$S_c=S_{HH}·S_{VV}^{*}=\Re (S_{HH}·S_{VV}^{*})+j\Im (S_{HH}·S_{VV}^{*}),$$

(5)

where $S_{HH}$ and $S_{VV}$ represent the complex number of SAR data in HH and VV. The superscript ${}^{*}$ is the complex conjugate operator, and j is the imaginary unit, $\Re (.)\mathrm{and}\Im (.)$ are the operators to separately compute the real and imaginary parts of a complex number. The real and imaginary parts of $S_c$ separately reflects the difference between even and odd bounce scattering, and the inter-channel correlation.

The intensity of SAR in HH and VV can be computed by $|S_{HH}{|}^2$ and $|S_{VV}{|}^2$, respectively. The co-polarization summation, intensity difference, and intensity ratio are separately defined as $|S_{HH}{|}^2+{\left|S_{VV}\right|}^2$, $|S_{HH}{|}^2-{\left|S_{VV}\right|}^2$ and $|S_{HH}{|}^2/{\left|S_{VV}\right|}^2$, which shows the relation of the dipoles in HH and VV orientation.

The sigma naughts in HH and VV, ${\sigma }_{HH}^{°},{\sigma }_{VV}^{°}$, are the normalized radar cross section, which are directly related to the power of the ground targets returned to SAR satellite antenna [40]. ${\sigma }_{HH}^{°}$ and ${\sigma }_{VV}^{°}$ can be theoretically defined as

$${\sigma }_{HH}^{°}=(k_s|S_{HH}{|}^2-\mathrm{NEBN})sin{\theta }_{\mathrm{loc}},$$

(6)

$${\sigma }_{VV}^{°}=(k_s|S_{VV}{|}^2-\mathrm{NEBN})sin{\theta }_{\mathrm{loc}},$$

(7)

where $k_s$ is the calibration and processor scaling factor. NEBN is short for Noise Equivalent Beta Naught, and it represents the influence of different noise contributions to the signal. ${\theta }_{\mathrm{loc}}$ is the local incidence angle.

The entropy is derived from a Cloude–Pottier eigenvalue/eigenvector decomposition [41,42], which presents the proportional importance of the dominant scattering type. It is defined as

$$H=-\sum _{\mathrm{i}=1}^{\mathrm{I}}{p}_{\mathrm{i}}{log}_{\mathrm{I}}{p}_{\mathrm{i}},$$

(8)

where $p_{\mathrm{i}}=\frac{{\lambda }_{\mathrm{i}}}{{\sum }_{\mathrm{i}=1}^{\mathrm{I}}{\lambda }_{\mathrm{i}}},\mathrm{i}\in [1,\mathrm{I}]$. ${\lambda }_{\mathrm{i}}(\forall \mathrm{i})$ are the real eigenvalues of the $\mathrm{I}\times \mathrm{I}$ coherency matrix. It is worth noting that besides the entropy, the Cloude–Pottier eigenvalue/eigenvector decomposition has two other parameters: anisotropy and mean scattering angle $\alpha$. For dual polarized data, anisotropy cannot be estimated, as the depolarization information is contained in the two minor eigenvalues in the quad-polarized scheme, but merely one minor eigenvalue can be obtained. Furthermore, the estimation of $\alpha$ is inferior due to the lack of the third eigenvalue. Details refer to [41,42]. Hence, when having HH and VV, I equals 2, and anisotropy and $\alpha$ are not included in the analysis of radar scatterer classification.

2.2.3. IMP Classification in Terms of Scattering Mechanisms

By aggregating the aforementioned polarimetric features, the IMP classification in terms of scattering mechanisms can be carried out with a machine learning method such as Random Forest [43]. Here, the scattering mechanisms include surface, low-volume (e.g., grass and farmlands), high-volume (e.g., tall (canopy) trees), and double bounce scattering. Such a classification process can be applied to every SAR acquisition and their temporally averaged SAR image. Samples for the double bounce (especially dihedral) and surface scattering classes are chosen using an external topographic base map with land use information, and TOP10NL data for our study. In doing so, we are able to improve the quality and quantity of training samples for double bounce and surface scattering classes, since TOP10NL offers four main land use classes: road, building, water body, and railway class, and roads and water bodies, are examples of surface scatterers, and buildings and railway tracks are examples of dihedral scatterers. Based on the visual interpretation of SAR amplitude maps and auxiliary optical images, the samples for the volume scattering classes are drawn. The volume scattering class can be divided into two sub-classes, i.e., low- and high-volume scattering class, mainly based on the basis of the SAR intensity in VV and HH, cf. [44]. For instance, a meadow is the sample for the low-volume scattering class, and the areas with tall trees are the samples for the high-volume scattering class. These samples with reliable labels are used to create the training samples.

It is worth noting that the use of TOP10NL for labeled sample creation in radar coordinates is not straightforward, as TOP10NL is stored as points, lines, and polygons in a local coordinate reference, i.e., RD (Rijksdriehoekscoordinaten) coordinates. Hence, the TOP10NL data have to be first rasterized, then converted into radar coordinates. As the elevation information is missing in TOP10NL, the projection error from the cross-range to the azimuth-range plane is inevitable, and should be taken into account when manually selecting and labeling data. Note that associating TOP10NL information with individual radar scatterers is resolution-cell-wise.

In principle, when running Random Forest, an ensemble of decision trees is used for IMP classification [45,46,47,48]. Each tree in the forest is trained with a subset of the total training samples. These subsets are extracted from the training samples using a bagging approach where samples are drawn with replacement [49]. Due to the ’with-replacement’ criteria, some samples are left ’out of the bag’ (OOB). During the training phase, the decision trees are grown by thresholding a feature value. This feature and its threshold value are chosen at each node using the Gini criteria [48]. At the post-training step, the OOB samples act as test data and facilitate an unbiased estimation of the classification accuracy of each decision tree, as well as for the whole forest. The classification accuracy can be measured using the OOB error, as well as Cohen’s kappa coefficient [50] and $F_1$-score [51]. To assess the feature importance, mean decrease accuracy (MDA) [52,53] can be employed. The final Random Forest-based IMP classification results suggest the proper physical attribute for every IMP. Furthermore, using the TOP10NL-derived land use information, land-use attributes can be readily associated with IMP.

3. Data and Test Site Description

We used 30 PAZ SAR images in HH and VV acquired between September 2019 and October 2021, at 5:18 PM. PAZ is the Spanish SAR satellite mission that was launched by Hisdesat in February 2018 [54]. It offers X-band SAR images in co-polarimetric modes, nominally on a weekly basis. In our case, most of the time interval between adjacent acquisitions is 22 days. The maximum time interval is 66 days and occurs between 1 August 2020 and 6 October 2020. We selected the SAR image acquired on 28 October 2020 as the master, indicated by the star in Figure 2a, because there was no rain (see Figure 2c) in blue) and its SAR stack coherence using Equation (A1) [34] is the highest. The perpendicular baseline of all 29 slave acquisitions shown in the black circle w.r.t. the master ranges [$-173,203$] m. TOP10NL, with a scale of 1:5000 (equivalently $0.5$ m spatial resolution) in 2017, was collected. Figure 3 shows the TOP10NL-based classification map over the test site. The classification categories include buildings in yellow, roads in grey, water bodies in cyan, and main tracks of the railway in blue. The hourly air temperature and cumulative precipitation data from weather station 270 were obtained from https://www.knmi.nl/home (accessed on 14 January 2022). Note that the temperature in black and precipitation data in blue were downsampled in order to associate with the 30 PAZ SAR acquisitions; see Figure 2c). It means that we only used the temperature recorded at 5 PM and cumulative precipitation between 1 AM and 5 PM at every SAR acquisition date.

As the test site is rather flat, covers coastal regions, and has complex and dynamic weather situations [55], the traditional linear tropospheric correction does not suffice. Then, to properly estimate the atmospheric phase delay and eliminate it from all interferograms, the GACOS (Generic Atmospheric Correction Online Service) product was collected and utilized. This product is generated by using numerical weather modeling based on the data combination of the high-resolution ECMWF weather model at $0.1°$ and 6 hr resolutions [56], SRTM DEM (90 m) [57] and ASTER GDEM (90 m) [58]. This product offers zenith total delay maps, and stratified and turbulent mixing information using the Iterative Tropospheric Decomposition (ITD) model [59]. In addition, assuming a linear relation between interferometric phase and elevation over non-deforming area, the linear tropospheric model can be applied independent of external weather data [60].

The test site covers the parts of Leeuwarden and Sneek, their outskirts and villages, in the province of Friesland, the Netherlands, as shown in red in Figure 3. The spatial extent of the test site is about $12\times 37$ km${}^2$. The soil over this site mainly consists of loam, clay, and peat [61]. Due to, e.g., soil shrinkage, consolidation, peat oxidation, and water extraction, continuous land subsidence is recorded (1 to 5 mm y${}^{-1}$ between 2015 and 2020, cf. https://bodemdalingskaart.nl/, accessed on 5 May 2022).

4. Results

4.1. IMP Geometric Attribute Extraction and Analysis

4.1.1. Coherence Scatterer Selection and Deformation Time Series Generation

We separately generated 29 interferograms in HH and VV for PSI processing. We set the maximum temporal baseline, $Critical\left(B_T\right)$, as 500 days, the maximum perpendicular baseline, $Critical\left(B_{\perp }\right)$, as 300 m, and the minimum coherence $(1-B_{\perp }/Critical\left(B_{\perp }\right))(1-B_T/Critical\left(B_T\right))$ as $0.35$, to form small baseline subsets. Here $B_T$ and $B_{\perp }$ represent the temporal baseline and perpendicular baseline of the interferograms. The value of $0.35$ was determined based on the experimental analysis, during which we tested the coherence values ranging between $0.25$ and $0.55$. Consequently, 181 interferograms in HH and VV were created, respectively, for SBAS; see Figure 2b). The acquisition information of these 181 interferograms is listed in

Table A1. Acquisition information for 181 interferogram generation. ID indicates the interferogram index, and the format of the master and slave acquisition date is yyyymmdd.

ID	Master, Slave Date	ID	Master, Slave Date	ID	Master, Slave Date	ID	Master, Slave Date
1	20190928, 20191111	47	20200116, 20200618	92	20200618, 20210124	137	20210215, 20210422
2	20190928, 20191203	48	20200116, 20200801	93	20200618, 20210215	138	20210215, 20210514
3	20190928, 20191225	49	20200116, 20201028	94	20200618, 20210309	139	20210215, 20210605
4	20190928, 20200116	50	20200207, 20200229	95	20200618, 20210422	140	20210215, 20210719
5	20190928, 20200207	51	20200207, 20200413	96	20200801, 20201028	141	20210215, 20210810
6	20190928, 20200527	52	20200207, 20200505	97	20200801, 20201211	142	20210215, 20210923
7	20191020, 20191111	53	20200207, 20200527	98	20200801, 20210102	143	20210309, 20210331
8	20191020, 20191203	54	20200207, 20200618	99	20200801, 20210124	144	20210309, 20210422
9	20191020, 20200116	55	20200207, 20200801	100	20200801, 20210215	145	20210309, 20210514
10	20191020, 20200229	56	20200207, 20201211	101	20200801, 20210309	146	20210309, 20210605
11	20191020, 20200322	57	20200229, 20200322	102	20200801, 20210422	147	20210309, 20210719
12	20191020, 20200413	58	20200229, 20200413	103	20201006, 20201028	148	20210309, 20210810
13	20191111, 20191203	59	20200229, 20200505	104	20201006, 20210124	149	20210309, 20210901
14	20191111, 20191225	60	20200229, 20200527	105	20201006, 20210331	150	20210309, 20210923
15	20191111, 20200116	61	20200229, 20200618	106	20201028, 20201211	151	20210309, 20211015
16	20191111, 20200207	62	20200229, 20200801	107	20201028, 20210124	152	20210331, 20210422
17	20191111, 20200229	63	20200229, 20201028	108	20201028, 20210215	153	20210331, 20210605
18	20191111, 20200322	64	20200322, 20200413	109	20201028, 20210309	154	20210331, 20210901
19	20191111, 20200413	65	20200322, 20200505	110	20201028, 20210331	155	20210331, 20211015
20	20191111, 20200505	66	20200322, 20200618	111	20201028, 20210422	156	20210422, 20210514
21	20191111, 20200527	67	20200322, 20201006	112	20201028, 20210605	157	20210422, 20210605
22	20191111, 20200618	68	20200322, 20201028	113	20201211, 20210102	158	20210422, 20210719
23	20191111, 20200801	69	20200413, 20200505	114	20201211, 20210124	159	20210422, 20210810
24	20191203, 20191225	70	20200413, 20200527	115	20201211, 20210215	160	20210422, 20210901
25	20191203, 20200116	71	20200413, 20200618	116	20201211, 20210309	161	20210422, 20210923
26	20191203, 20200207	72	20200413, 20200801	117	20201211, 20210422	162	20210422, 20211015
27	20191203, 20200229	73	20200413, 20201028	118	20201211, 20210514	163	20210514, 20210605
28	20191203, 20200322	74	20200413, 20210124	119	20201211, 20210605	164	20210514, 20210719
29	20191203, 20200413	75	20200505, 20200527	120	20201211, 20210719	165	20210514, 20210810
30	20191203, 20200505	76	20200505, 20200618	121	20201211, 20210810	166	20210514, 20210923
31	20191203, 20200527	77	20200505, 20200801	122	20210102, 20210215	167	20210605, 20210719
32	20191203, 20200618	78	20200505, 20201028	123	20210102, 20210309	168	20210605, 20210810
33	20191203, 20200801	79	20200505, 20201211	124	20210102, 20210422	169	20210605, 20210901
34	20191225, 20200116	80	20200505, 20210124	125	20210102, 20210514	170	20210605, 20210923
35	20191225, 20200207	81	20200505, 20210309	126	20210102, 20210719	171	20210605, 20211015
36	20191225, 20200229	82	20200527, 20200618	127	20210102, 20210810	172	20210719, 20210810
37	20191225, 20200505	83	20200527, 20200801	128	20210124, 20210215	173	20210719, 20210901
38	20191225, 20200527	84	20200527, 20201028	129	20210124, 20210309	174	20210719, 20210923
39	20191225, 20200618	85	20200527, 20201211	130	20210124, 20210331	175	20210719, 20211015
40	20191225, 20200801	86	20200527, 20210102	131	20210124, 20210422	176	20210810, 20210901
41	20200116, 20200207	87	20200527, 20210215	132	20210124, 20210605	177	20210810, 20210923
42	20200116, 20200229	88	20200618, 20200801	133	20210124, 20210901	178	20210810, 20211015
43	20200116, 20200322	89	20200618, 20201028	134	20210124, 20210923	179	20210901, 20210923
44	20200116, 20200413	90	20200618, 20201211	135	20210124, 20211015	180	20210901, 20211015
45	20200116, 20200505	91	20200618, 20210102	136	20210215, 20210309	181	20210923, 20211015
46	20200116, 20200527

. We realized the interferogram stack generation by using Doris [62], and we applied the standard PSI and SBAS using StaMPS [63]. $D_a$ was defined as $0.3$ for pointwise CCS selection, and $D_{\Delta a}$$0.55$ for TCS selection (using Equation (1)). The GACOS data were used to estimate the atmospheric phase delay in all interferograms and eliminate it from deformation estimation. The GACOS-based atmospheric phase values were assigned to all CCSs and TCSs, and assisted in phase unwrapping and parameter estimation. Figure 4a,b show the GACOS-based atmospheric phase distribution of all 181 interferograms for IMPs in HH and VV, respectively. Every column in Figure 4a,b shows the IMP atmospheric phase distribution (normally distributed) in the designated interferogram, albeit showing the difference in the IMP registry in HH and VV, the expectation of the GACOS-based atmospheric phase in HH and VV both ranges between $-13$ and 6 radians, and the dispersion limits’ $[0.2,6]$ radians. It means that the atmospheric phase delay is at the level of a few centimeters, i.e., 32 mm at most, and varies in space with the maximum atmospheric phase difference of 14 mm. As a result, $43,551$ pointwise CCSs and $494,435$ TCSs in HH ($537,986$ IMPs in total), and $36,287$ pointwise CCSs and $449,387$ TCSs in VV ($485,674$ IMPs in total) were finally obtained. There are $160,635$ IMPs in HH and VV sharing the same locations, among which there are $16,079$ CCSs.

4.1.2. Spatial Reference Selection and Alignment

According to the criteria of selecting the spatial reference (see Section 2.1.2), we defined that every IMP sibling pair is selected from the CCS registry, situated at the same location, having 0 mm y${}^{-1}$ linear deformation velocity with a $0.1$ mm y${}^{-1}$ tolerance (provided that InSAR deformation measurements (X-band) empirically have ${\sigma }_y=3$ mm uncertainty cf. [64]), less than 7 mm deformation gradient (approx. 31/4 mm, to exclude the temporal unwrapping errors), less than $2{\sigma }_y=6$ mm deformation difference between HH and VV. Then, 16 sibling pairs were identified and indicated by the blue pentagram in Figure 5a. As the identified 16 reference sibling pairs were stable over time and dispersedly distributed over the whole test site, by averaging their deformation time series and subtracting these values from all IMPs in HH and VV, we mitigated the reference error at a global scale. Figure 6a,b separately show the cumulative deformation time series of these pairs in HH and VV. Here, each row depicts the deformation time series of the reference point from the pairs. The average deformation time series of the reference points in HH (indicated by the square) and VV (circle) are similar, all having less than 1 mm difference; see Figure 6c). As proposed in Section 2.1.2, the deformation time series of the common reference is the mean value of the average deformation time series shown in the dot line, which was used to align all IMPs in HH and VV into a common coordinate reference system by subtracting these mean values from all IMP deformation time series.

The deformation time series of IMPs in HH and VV w.r.t. the common reference are part of IMP geometric attributes. The linear deformation velocity in the line-of-sight direction of IMPs in HH and VV is separately shown in Figure 5a,b. The deformation time series of the IMPs in HH and VV at the same spots were averaged between HH and VV, and the corresponding deformation velocities were re-calculated as well. We found that $75\%$ and $80\%$ of IMPs in HH and VV are nearly stable, with the linear deformation velocity [$-2,2$] mm y${}^{-1}$. A total of $2\%$ and $1\%$ of IMPs in HH and VV present the pronounced subsiding trend, having more than $-5$ mm y${}^{-1}$ linear deformation velocity, and are mainly observed on the outskirts of the cities, coastal line, and some transport corridors.

4.2. IMP Physical Attribute Extraction and Analysis

4.2.1. Scattering Mechanism Classification and Accuracy Analysis

As stated in Section 2.2.3, the IMPs can be classified in terms of scattering mechanisms, and categorized into surface, low- and high-volume, and double bounce scattering mechanism classes. We first removed the speckle noise using the multi-temporal speckle filter method (Equation (3)), the step in which the Boxcar filter with the $3\times 3$ window size in space and temporal filter using Equation (4) were applied. Then, we applied the scattering mechanism classification to the temporally average SAR images in HH and VV. The Random Forest method was trained using $1074$ samples with eight representative features mentioned in Section 2.2.2. These samples, covering four types, surface, low- and high-volume, and double bounce scattering, were crafted using the assistance of the TOP10NL data and visual interpretation of SAR amplitude and optical images. Acknowledging the classification accuracy can be biased when dealing with data whose classes are unbalanced [65], we used 320 surface, 320 low-volume and 272 high-volume, and 162 double bounce scattering samples. Here, we assumed that the double bounce scattering samples are highly separable thanks to their high signal-to-noise ratio. The samples that can cause errors are surface scattering and low- and high-volume scattering, which were equally and substantially represented in the training samples. In addition, this Random Forest classifier was tested with the samples over the test set that were not taken from the training dataset; see Figure A1. The number of samples in the test set was kept at around 650 per class; see

Table A2. Confusion matrix of the samples.

		Reference Data
		Double Bounce	Low-Volume	High-Volume	Surface
Classified data	Double bounce	651	8	10	4
	Low-volume	0	641	15	19
	High-volume	4	19	652	0
	Surface	0	0	0	649

. We found that the classifier performs well throughout the test set, as the kappa value equals $0.96$ and the weighted $F_1$-score is $0.97$.

The accuracy in the classification calculated using the unbiased OOB error is $0.96$. Out of all the features, we found that the summation of co-polarization intensities, $|S_{HH}{|}^2+{\left|S_{VV}\right|}^2$, is the most important feature. This can be seen in the feature importance plot shown in Figure 7, with an average decrease of $11\%$ when randomly permuted. This is likely due to the increased dynamic range of the feature values, which would assist in the classification. Furthermore, ${\sigma }_{HH}^{°}$ turned out to be the second most important feature, with an average decrease of $7\%$, probably because of the higher sensitivity of this polarization mode to surface and volume scatterers. The real part of the co-polarization cross product, $\Re (S_{HH}·S_{VV}^{*})$, also shows significant values of mean decrease in accuracy of $\sim 4\%$. The entropy H feature and the imaginary part of the co-polarization cross product, $\Im (S_{HH}·S_{VV}^{*})$, seem less important. The least important features are the ratio and difference of the co-pol intensities, $|S_{HH}{|}^2/{\left|S_{VV}\right|}^2$ and $|S_{HH}{|}^2-{\left|S_{VV}\right|}^2$. This could be caused by the reduced dynamic range between the values, and lead to more mixing of classes, thereby reducing classification accuracy.

Figure 8a shows the scattering mechanism map over the entire test site. The scattering mechanisms of all pixels, irrespective of coherent or non-coherent scatterers, are categorized into double bounce (in red), low-volume (in yellow), high-volume (in green), and surface scattering (in blue). As IMPs’ geometric attributes (IMP positions) were obtained, we extracted the physical attributes (i.e., scattering mechanisms) for all IMPs based on their positions. Figure 8b–e show the spatial distribution of 537,986 IMPs in HH categorized as double bounce, low- and high-volume, and surface scattering class, respectively, along with the corresponding linear deformation velocity estimations, while Figure 8f–i show the counterpart of 485,674 IMPs in VV. The results show that per class, the spatial distribution and temporal behavior of IMPs in HH and VV were distinct. The HH channel observed a bit more IMPs with large subsiding rates (e.g., [$-10,-5$] mm y${}^{-1}$) shown in the orange and red color in Figure 8b–e. Note that water bodies were also labeled as surface scatterers. Theoretically, no IMPs can be observed over (calm) water areas. However, for instance a handful of IMPs in the Wadden sea conservation area were observed; see Figure 8e,i, which are probably reflected by the sandbanks. IMPs detected on the canals could be reflected by human-made structures, such as bridges, river banks and stakes, over the canal areas. Limited by the scale of Figure 8b–i and high spatial density of IMPs with the four different classes in HH and VV separately, the subtle difference in IMP spatial distribution is hard to be recognized using Figure 8b–i; therefore, in the following, we illustrate three small areas in Figure 9 and further investigate the two most important polarimetric features over time of five IMPs from these areas; see Figure 10.

Table 1. Total number of IMPs in HH and VV with double bounce, low-volume, high-volume, and surface scattering mechanism (SM).

SM	Double Bounce	Low-Volume	High-Volume	Surface
HH	302,586	64,482	120,112	50,806
VV	245,056	71,504	91,102	78,012

lists the total number of IMPs in HH and VV with these four scattering mechanisms. Most IMPs have double bounce scattering characteristics; $56\%$ for HH, and $50\%$ for VV. Remarkably, the number of IMPs in HH with double bounce and high-volume scattering is separately more than the one in VV, while the number of IMPs in HH with low-volume and surface scattering is $7\%$ and $35\%$ less than the number in VV. It implies SAR in the VV channel is more sensitive to detect low-volume and surface scatterers as IMPs. It is counterintuitive that the low- and high-volume scatterers were recognized as IMPs, as such scatterers rarely (constantly) keep temporal coherence. We infer this is attributed to (1) double bounce at stems or trunks of low- and high-volume scatterers, (2) invariance of certain low- and high-volume targets, and/or (3) scattering mechanism mislabeling. As radar signal usually has specular reflection over smooth surfaces and diffusion over rough surfaces, it implies IMPs with surface scattering do not directly represent a smooth surface itself, but may represent natural reflectors, such as bare rocks, rough roads, and spatially homogenous fields.

4.2.2. IMP Land-Use Attribute Extraction and Association with Physical Attributes

Using the TOP10NL data, we categorized all IMPs in HH and VV in accordance with land use types, encompassing building, road, water, railway, and uncharted classes, and assigned these land-use attributes to the corresponding IMPs.

Table 2. Total number of IMPs in HH and VV labeled as building, road, water, railway, and uncharted class.

	Buildings	Roads	Water	Railways	Uncharted
HH	343,967	81,701	1133	32,975	78,210
VV	273,612	109,944	1093	32,291	68,734

lists the total amount of IMPs in HH and VV per class. For building, water, and railway classes, the number of IMPs in HH is $26\%$, $3.7\%$, and $2\%$ more than the counterpart in VV. For road class, SAR in the VV channel was observed more by $35\%$, compared with the observed 81,701 IMPs in HH. As IMPs in HH and VV over water account for $0.2\%$ of all IMPs, the inaccuracy of this land-use attribute extraction result is allowable. As TOP10NL data were observed and sorted in 2017 and the SAR images were acquired between 2019 and 2021, such a time difference, as well as the class absence of the agricultural fields, leads to the uncharted class. The uncharted class includes the IMPs that reflect the new targets that appeared after 2017 and targets in agricultural fields.

By comparing the physical and land-use attributes of IMPs, we recognized that $64\%$ of IMPs from buildings in both HH and VV are double bounce scatterers. IMPs on roads have $37\%$ and $23\%$ double bounce scatterers (for instance, from lamp posts and embankments), about $20\%$ low-volume scatterers, $18\%$ and $11\%$ high-volume scatterers, and $25\%$ and $44\%$ of surface scatterers in HH and VV. For railways, $55\%$ and $52\%$ double bounce scatterers, $35\%$ and $34\%$ high-volume scatterers, $9\%$ and $13\%$ low-volume scatterers, and both $1\%$ surface scatterers in HH and VV were labeled for IMPs on railways.

Table 3. Scattering mechanism class percentage of IMPs in HH and VV from buildings, roads, and railways.

		Double Bounce	Low-Volume	High-Volume	Surface
Buildings	HH	64%	9%	21%	6%
	VV	64%	11%	19%	6%
Roads	HH	37%	20%	18%	25%
	VV	23%	22%	11%	44%
Railways	HH	55%	9%	35%	1%
	VV	52%	13%	34%	1%

lists this information. As IMPs over water may not be reliable and there is no further analysis on IMPs in the uncharted class, the scattering mechanism class percentage is not discussed for the water and uncharted class.

Taking a building located at $53.2036°$N, $5.8335°$E, namely ROC Friese Poort Leeuwarden (with four stories), as an example, we found 39 and 21 IMPs with double bounce scattering feature separately in HH and VV reflected from the building roof. Figure 9a presents the line-of-sight deformation map over this building outlined in yellow. The IMPs in HH and VV are indicated by the dot and cross, respectively. The corresponding deformation time series is shown in Figure 9(a.1) for IMPs in HH and Figure 9(a.2) for IMPs in VV. The result shows four pairs of IMPs in HH and VV are located at the same places and possess almost the same temporal behavior, with a $0.2$ mm y${}^{-1}$ average difference. The mean deformation rate of all IMPs is $-0.8$ mm y${}^{-1}$, and the temporal behavior of all IMPs can be well interpreted by a linear function of time, according to deformation modeling with multiple hypothesis testing [64]. Here, we assumed 3 mm observation error from the X-band PAZ data to initiate the multiple hypothesis testing analysis.

We took a ∼1.5 km segment of highway N383, at $53.2160°$N, $5.7087°$E, as a representative of the IMPs from the road class. A total of 38 and 71 IMPs in HH and VV were observed over this road segment. A total of 31 out of 38 IMPs in HH and 67 out of 71 IMPs in VV were labeled as surface scatterers; see Figure 9b, shown in blue. Seven and four IMPs in HH and VV were determined as low-volume scatterers, shown in yellow. The deformation map and deformation time series of all IMPs are shown in Figure 9(b.1,b.3,b.4). The mean deformation velocity is $-1.3$ mm y${}^{-1}$, and the maximum subsidence rate is $4.7$ mm y${}^{-1}$. Only one pair of surface scattering IMPs in HH and VV appeared at the same location ($53.2117°$N, $5.7138°$E), performing linearly over time with $-1.1$ mm y${}^{-1}$. Five IMPs with surface scattering features in HH and VV had temperature-related movement over time (using Equation (A3)), and the temperature-related parameter $\eta$ [mm${/}^{°}\mathrm{C}$] ranges between $0.12$ and $0.49$; see Figure 9(b.2). As the $\eta$ values are all positive, this implies, in addition to the linear temporal behavior, that those surface scatterers expanded when temperature increased and vice versa.

We also selected a 400 m railway segment, located at $53.1828°$N, $5.7756°$E, as the third example. A total of 29 and 27 IMPs in HH and VV were detected, among which there are 20 and 16 double bounce scatterers in HH and VV separately; see Figure 9c in red. The remaining IMPs were labeled as high-volume scatterers in green. Three pairs of double bounce IMPs in HH and VV were observed at the same spots, with an average subsiding rate of $2.2$ mm y${}^{-1}$. The deformation map of all IMPs is shown in Figure 9(c.1), with a mean deformation rate of $-1.7$ mm y${}^{-1}$ and a maximum subsidence rate of $4.3$ mm y${}^{-1}$. Assuming all double bounce scatterers are reflected from railway infrastructure such as rail tracks, we show the deformation time series of the double bounce IMPs in HH and VV in Figure 9(c.2,c.3). We found that all IMPs in HH and VV follow the linear deformation model, and the deformation trends of them in HH and VV are almost in concert. It implies that irregular settlement over this railway segment is less likely.

4.2.3. Temporal Behavior of IMP Physical Attributes

The physical attribute of every IMP may vary over time; therefore, this section illustrates the scattering mechanism dynamics of five IMPs, with surface, low- and high-volume, and double bounce scattering from the building, road, and railway examples shown in Figure 9. These are one IMP in HH over the building at $53.2034°$N, $5.8325°$E, two IMPs over the road segment at $53.2204°$N, $5.7043°$E in HH and at $53.2215°$N, $5.7032°$E in VV, and two IMPs in HH over the railway segment at $53.1837°$N, $5.7749°$E and at $53.1834°$N, $5.7751°$E. As every single SAR acquisition is smeared by speckle noise and merely spatial filtering failed to dramatically reduce the noise impact on scattering mechanism classification per acquisition, we then focused on analyzing the temporal evolution and relation of the first two most important features $|S_{HH}{|}^2+{\left|S_{VV}\right|}^2$ and ${\sigma }_{HH}^{°}$ in dB for performing the classification using the Random Forest classification method (see Figure 7). Figure 10a–c show the feature spaces plotted between $|S_{HH}{|}^2+{\left|S_{VV}\right|}^2$ and ${\sigma }_{HH}^{°}$. The 2D feature spaces were partitioned using the first 50 trees of the trained Random Forest classification method. The bottom-left area of these three subfigures is classified as surface scattering in blue, and the top-right area is classified as double bounce scattering in red. The space in between is classified as low-volume in yellow and high-volume scattering in green. There are also a few instances of double bounce and surface scattering near the boundaries of volume scattering classes. These instances result in overlapping multi-colored lines in the middle of the figures that are due to diversity in the classification prediction results from 50 successive trees. The black outlined colored dots in Figure 10a–c show the temporal evolution of feature values for these five IMPs. Here, the evolution is represented using size and transparency. Specifically, as the epochs increase, both the size and transparency of IMPs increases. The temporally average SAR values that were used for the IMP scattering mechanism classification (see Section 4.2) are shown as the largest dots with the same class color but with a white border. We found that the range of values for both features decrease as IMPs move from double bounce to volume then to surface scattering class. The dispersion of the feature values increases if IMPs follow the same class sequence. The temporally average IMP values separately fall in their respective class partition in the feature space. Moreover, the temporal variation of these two feature values seldom takes the feature values to the the partition of other classes, which can be seen in Figure 10d,e as well. For these instances, we also found that as the precipitation dramatically increased to $42.4$ mm on 5 June 2021 from 0 mm on 14 May 2021, pointed out by the black arrow in Figure 2c, the values of these two features for IMPs with surface scattering character indicated by the blue square-line rose accordingly, by $16.6$ dB for $|S_{HH}{|}^2+{\left|S_{VV}\right|}^2$ and $9.6$ dB for ${\sigma }_{HH}^{°}$ indicated by the black arrow in Figure 10d,e. We suspect that surface soil reflection of such surface IMPs is enhanced, possibly related to the increase in soil moisture, and temporal evolution of (some) surface IMPs resonates with the changes in precipitation cf. [66].

5. Discussion

This study used the GACOS weather model to remove the atmospheric phase, which is at the level of a few centimeters for our case (see Figure 4 in Section 4.1.1). Due to the lack of ground truth data, it is not practically feasible to assess the accuracy of the GACOS-based atmospheric phase result. As for the precision evaluation, we found that the atmospheric phase residuals in all interferograms were well eliminated, and the spatial and temporal deformation smoothness of the IMPs was improved after the GACOS-based atmospheric phase removal. It implies that GACOS is able to denoise the unwrapping phase estimation and improve the deformation estimation of IMPs. In the case of having ground truth data from, e.g., GPS stations, the difference in the double-difference measurements between IMPs and ground truth counterparts is indicative of the performance of the GACOS-based atmospheric phase removal cf. [67].

Regarding the use of the spatial reference selection criteria (Equation (2)) for our case, we implicitly assumed that the TSInSAR-derived deformation estimations have the same uncertainty in HH and VV, resulting in an equal weight of all deformation estimations. If we have the complete stochastic structure of deformation estimations in HH and VV, we can define the appropriate weight values. In this scenario, we can introduce the weight parameters into Equation (2), namely ${\widehat{y}}_{\mathrm{ref}}^k=\frac{{\sum }_{q=1}^N{\widehat{y}}_q^k{\omega }_q^k}{{\sum }_{q=1}^N{\omega }_q^k}$ (${\omega }_q^k$ is the weight of the point q’s deformation, at $kth$ acquisition). In short, we believe that our spatial reference selection criteria are useful when aligning multi-polarimetric SAR observations.

As for the scattering mechanism classification, we crafted eight incoherent polarimetric features for the Random Forest classifier (Section 4.2.1). The accuracy of the scattering mechanism classification using the Random Forest method and eight representative features was convincing, up to $0.96$ (using the OOB error) and $0.91$ for the kappa coefficient (see Section 4.2.1). As refs. [25,68] showed, there are coherent polarimetric features such as co-polarimetric phase difference (CPD) that are promising for classification; the coherent polarimetric features will be investigated for future work.

As the result shows that more than half of IMPs in HH and half of IMPs in VV are double bounce scatterers (Table 1), we think it resonates with the nature of constantly/temporarily coherent radar scatterers. By visually comparing IMP spatial distribution (see Figure 8) with the corresponding Landsat-8 optical image acquired in 2021 and TOP10NL, we found that double bounce scatterers mainly reflected from, e.g., buildings and lamp posts along, e.g., roads and railway tracks. Hence, the deformation time series analysis of the double bounce IMPs can be contributed to structural health monitoring and infrastructure damage and risk assessment. The result also shows that some IMPs were labeled as low-volume and high-volume scatterers, and almost all of them kept a linear deformation trend over time. Especially over the railways, the high-volume IMPs account for more than $30\%$ in HH and VV, which is quite significant. This could be led by the not-high-enough spatial resolution of PAZ SAR data, and relatively small dimension of the railways (with $1.235$ m track gauge). To improve this, we recommend using higher resolution SAR data in multi-polarimetric channels including dual- and/or quad- polarization, and other geospatial data for data labeling, training, and testing, or investigating many more features for the scattering mechanism classification.

As we found, some surface scatterers have temperature-related behavior over time, which suggests that these surface scatterers are more susceptible to the variation in temperature than low/high-volume and double bounce scatterers. In addition, significant change in precipitation could lead to a pronounced change in $|S_{HH}{|}^2+{\left|S_{VV}\right|}^2$ and ${\sigma }_{HH}^{°}$ feature time series for the surface IMPs, as we illustrated in Section 4.2.3. We therefore suggest considering the precipitation measures as contextual information when analyzing and modeling the temporal behavior of surface scatterers.

6. Conclusions

In this study, we presented and demonstrated a new scheme to extract meaningful radar scatterer attributes from co-polarimetric SAR images by synthesizing co-polarimetric SAR output, i.e., radar scatterers’ position, deformation, and scattering mechanisms, and assimilating topographic base map information. The analysis of these attributes encompassing geometric, physical, and land-use aspects of radar scatterers contributes to significant improvements in multi-polarimetric SAR data and geo-process interpretation. Radar scatterers’ deformation time series interpretation can now be performed more in-depth for individual ground targets by having physical and land-use attributes. In addition, our scheme can be straightforwardly applied to other dual-pol (with both co- and cross-pol information) and quad-pol SAR data. According to the results showing that more double bounce scatterers can be captured by SAR in HH, and more surface scatterers by SAR in VV, we can maximize the information for land surface mapping and monitoring by integrating the radar scatterers from SAR in HH and VV. Particularly, for monitoring infrastructure such as buildings, SAR in HH is preferred, while for monitoring roads such as runways and highways, SAR in VV is recommended.