Temporal H/alpha Target Decomposition for Landslide Monitoring Using Ku-Band GB-SAR Time Series

We investigate the applicability of the entropy (<italic>H</italic>)/ alpha <inline-formula><tex-math notation="LaTeX">$(\bar{\alpha })$</tex-math></inline-formula> target decomposition realized by the temporally averaged coherency matrix, called temporal <inline-formula><tex-math notation="LaTeX">$H/(\bar{\alpha })$</tex-math></inline-formula>. We apply the temporal <inline-formula><tex-math notation="LaTeX">$H/\bar{\alpha }$</tex-math></inline-formula> to ground-based synthetic aperture radar (GB-SAR) continuous monitoring data to characterize the scattering mechanism temporal change. As a case study, this work demonstrates the application of the temporal <inline-formula><tex-math notation="LaTeX">$H/\bar{\alpha }$</tex-math></inline-formula> technique to landslide monitoring to detect and investigate the temporal scattering mechanism. The study acquired long-term GB-SAR polarimetric data over the postlandslide slope, Minami-Aso, Kumamoto, Japan. The study first investigated the property of the temporal <inline-formula><tex-math notation="LaTeX">$H/\bar{\alpha }$</tex-math></inline-formula> parameters over selected land cover types by comparing it with that derived by spatial averaging (spatial <inline-formula><tex-math notation="LaTeX">$H/\bar{\alpha }$</tex-math></inline-formula>) to explain the landslide monitoring results. Also, the rainfall effects on the temporal <italic>H</italic>/<inline-formula><tex-math notation="LaTeX">$\bar{\alpha }$</tex-math></inline-formula> parameters are demonstrated. The temporal <italic>H</italic> and <inline-formula><tex-math notation="LaTeX">$\bar{\alpha }$</tex-math></inline-formula> values increase up to 0.07<inline-formula><tex-math notation="LaTeX">$^\circ $</tex-math></inline-formula> and 13.54<inline-formula><tex-math notation="LaTeX">$^\circ $</tex-math></inline-formula>, respectively, when the rainfall rate is 52.5 mm/h. The time-series analysis of the temporal <inline-formula><tex-math notation="LaTeX">$H/\bar{\alpha }$</tex-math></inline-formula> indicates an obvious temporal transition of the scattering mechanism and a change of the backscattering stationarity when a landslide occurs. The applicability of the temporal <inline-formula><tex-math notation="LaTeX">$H/\bar{\alpha }$</tex-math></inline-formula> for the change-detection is discussed by comparing it with the classical spatial <inline-formula><tex-math notation="LaTeX">$H/\bar{\alpha }$</tex-math></inline-formula> parameters.

Multipolarimetric information can be used for the polarimetric target decomposition theorems [9]. Over naturally distributed areas, incoherent decompositions, such as eigenvectorand model-based decomposition based on polarimetric matrices (i.e., covariance or coherency matrix) provide the statistical description of the scattering mechanism variation. In the eigenvector-based decomposition, the theory is based on the derivation of the polarimetric parameters from the eigenvalues and eigenvectors of the coherency matrix proposed by Cloude [10] and Cloude and Pottier [11]. Speckle filtering must be applied before the physical interpretation of the polarimetric data realized by ensemble averaging. Therefore, the filtered coherency matrix shows second-order statistics. The derived parameters (i.e., polarimetric entropy H and mean alpha angleᾱ) account for the scattering randomness and the intrinsic scattering mechanism within an ensemble. The ensemble averaging of the coherency matrix is typically performed by spatial multilooking due to insufficient polarimetric data along the temporal axis in spaceborne SAR [12].
A few studies have discussed temporal averaging's applicability in the derivation of the covariance or coherency matrix. In the spaceborne SAR context, Weissgerber et al. estimated the H value by temporal averaging using a temporal stack of polarimetric data [13] called temporal H (H temp ). However, this technique requires sufficient temporal stack polarimetric data, which is impractical in current orbital SAR missions. Furthermore, the polarimetric response has an incidence-angle dependence. Hence, temporal observations are necessary in the same position to fix the incidence angle along with the time series, which is a difficult requirement for orbital SAR. In contrast, GB-SAR measurement easily fulfills all the abovementioned requirements. Two studies in [3] and [4] used temporal averaging for the polarimetric GB-SAR measurements: one for urban monitoring and the other for glacier monitoring. However, very few cases have been reported up to now, and all the above studies discussed only the applicability of polarimetric entropy out of the other decomposition parameters via temporally averaged coherency matrix.
Considering the limited analysis in existing studies regarding polarimetric GB-SAR monitoring, we demonstrate the polarimetric decomposition technique in the continuous GB-SAR monitoring framework. We particularly focus on applying H/ᾱ decomposition, derived by a temporally averaged coherency matrix referred to as the temporal H/ᾱ denoted as H temp /ᾱ temp to This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ exploit the advantage of continuous GB-SAR monitoring (i.e., a high acquisition rate and a stationary observation). Furthermore, we jointly discuss H temp andᾱ temp parameters through the case study of landslide monitoring. In the presented analysis, the derived H temp /ᾱ temp time series is further compared with conventional H/ᾱ values derived by a spatially averaged coherency matrix referred to as H spatial /ᾱ spatial . This study contributes to the novel GB-SAR continuous monitoring approach that gives a quantitative characterization of the terrain surface temporal change and provides additional key information to the commonly used InSAR technique for more informative monitoring.
The remainder of this article is organized as follows. Section II introduces the GB-SAR monitoring campaign performed by Tohoku University over a postlandslide mountainous area. Section III elaborates on H temp /ᾱ temp decomposition technique. Section IV presents the time series H temp /ᾱ temp of the landslide evolution with the discussions of the properties of the both H temp /ᾱ temp and H spatial /ᾱ spatial parameters over the observed slope and the rainfall effect on the decomposition parameters. Section V presents the interpretation of the results in Section IV and the discussion of the comparison between the H temp /ᾱ temp and H spatial /ᾱ spatial results; and finally, Section VI draws the conclusions.

II. TEST SITE AND GB-SAR SYSTEM
Several landslides occurred in Kumamoto, Japan, due to a series of earthquakes in April 2016. The most noticeable landslide in Tateno, Minami-Aso, located in the caldera's outer rim, destroyed a 200 m Aso large bridge and the railroad line [14]. After the landslide occurrence, recovery projects were initiated to remove the debris flow. Considering the potential for a secondary landslide that threatened the onsite operators involved with daily road construction and the restoration work over the postlandslide slope, Tohoku University deployed polarimetric GB-SAR to monitor the slope instability [15]. The near-real-time early warning system was achieved by the continuous operational mode that repeatedly acquired the data with a 5 min temporal interval. The long-term continuous GB-SAR observation began in March 2017 and continued until August 2020.
The deployed GB-SAR system (Metasensing Fast GB-SAR) is the FMCW radar operating at the Ku-band (17.2 GHz) center frequency with a 300-MHz frequency bandwidth (see Fig. 1). This bandwidth gives a 0.5 m range resolution. The cross-range resolution is varied according to the range and computed as 4.8 mrad (4.8 m at 1000 m range). Fig. 2 depicts the observed slope steepness and plots the relative topographic height regarding the GB-SAR location over the bare surface pixels projected on the ground range axis. The GB-SAR look angle θ look , slope orientation angle θ ori , and local incidence angle θ inc varied according to the slope location. For example, angles θ look , θ ori , and θ inc are around 18.1 • , 32.6 • , and 75.4 • at the 600 m ground-range, respectively.

III. THEORY OF TEMPORAL H/ALPHA DECOMPOSITION
Cloude and Pottier introduced a scheme for parameterizing polarimetric scattering problems based on the eigenvalues and eigenvectors of the coherency matrix [10], [11], [16].  Relative topographic height with reference to the GB-SAR location over the bare surface pixels projected on the ground range axis. The GB-SAR look angle θ look , orientation angle θ ori , and local incidence angle θ inc are defined. The 3 × 3 coherency matrix T in the monostatic case is generated by ensemble averaging of the outer product of the 3 × 1 Pauli scattering vector k p , realizing the second-order moments of the fluctuations as where · denotes ensemble averaging assuming the random medium's homogeneity, and k * T p is the complex conjugate and transpose. The Pauli scattering vector in linear polarization basis is defined as Decomposing the coherency matrix into the eigenvectors u i and the eigenvalues λ i , which are types of scattering processes and their relative magnitudes (used as statistical weight), respectively, we obtain The T matrix is defined as the positive semidefinite Hermitian matrix; thus, the derived eigenvalues are real and positive and may have zero values. The eigenvectors in (3) are parameterized as u i = cos α i sin α i cos β i e jδ i sin α i cos β i e jγ i T , (4) where phase δ i expresses the relative phase between S HH − S VV and S HH + S VV . Similarly, phase γ i denotes the relative phase between S HV and S HH + S VV . α i represents the intrinsic scattering type. β i is related to the polarimetric orientation angle of the radar line-of-sight (LOS). The probabilistic model gives the polarimetric H andᾱ as where scattering probabilities are derived by H represents the degree of the statistical disorder of each distinct scatter type within the ensemble [12]. The higher H indicates high scattering randomness.ᾱ identifies the dominant scattering mechanism for random media. The three dominant scattering states ofᾱ are single-bounce scattering (ᾱ → 0), dipole-like scattering (ᾱ → π / 4 ), and double-bounce scattering or helix scattering (ᾱ → π / 2 ). A spatial multilooking operation practically realizes ensemble averaging in (1), assuming ergodicity and spatial stationarity of the corresponding samples, because of the lack of temporal polarimetric data stack in spaceborne and airborne SAR. This process generally degrades the spatial resolution and has a risk of mixing different distributions, leading to a biased estimate of the decomposition values [13], [17].
Nonetheless, the above issues can be solved by temporal averaging. The GB-SAR observation, which enables dense temporal monitoring, makes this implementation possible. The stationary observation in GB-SAR monitoring can also fix the incidence angle along the time sequence. This configuration is suited for temporal averaging because the scattering mechanism has the incidence angle dependence [18].
The ensemble-averaged coherency matrix is then realized by temporal averaging, assuming temporal stationarity as where n is the acquisition time indices and N is the number of polarimetric data along time. The temporal averaged coherency matrix is decomposed into eigenvectors and eigenvalues, followed by a derivation of the temporal entropy H temp and temporal mean alpha angleᾱ temp values with (5) and (6). Accordingly, the derived H temp value indicates a descriptor that accounts for the backscattering process's degree of temporal stationarity. α temp value shows a dominant scattering mechanism within the defined time window.

IV. APPLICATION TO LANDSLIDE MONITORING
Through the restoration work period over the postlandslide slope, a small-scale secondary debris flow (hereafter referred to as "landslide") occurred at the middle part of the monitored slope triggered by heavy rain on June 30, 2019. Fig. 3 shows the observed postlandslide mountainous slope and the enlarged view of the landslide. Two regions of interest (ROI) were defined within the landslide affected zone for further processing: ROI-I in the avalanche region and ROI-II in the debris cover region [see Fig. 3(d)]. The land cover in ROI-I and II before the landslide showed mainly sparse vegetated cover [see Fig. 3(a)]. After the landslide occurrence, a part of the vegetation in ROI-I flowed away due to the avalanche. The vegetation in ROI-II was then covered with debris. The landslide's exact occurrence time was unknown because the installed optical camera failed to visualize the landslide due to the dark (pre-dawn morning) and foggy conditions.

A. Processing Chain
The number of samples in the averaging operation affects the polarimetric H estimation. As demonstrated in [13] and [17], the bias in H was pronounced in the case of the insufficient number of averaging samples, leading to a wrong interpretation of the target character. Therefore, a tradeoff between bias and temporal resolution must be considered according to the system operational mode. In this analysis, the processing chain defined the 1 h temporal averaging window to derive H temp andᾱ temp values. GB-SAR acquired polarimetric images each 5 min; thus, 12 polarimetric data were temporally averaged within 1 h. Sliding this temporal window by 30 min realized the time series of H temp andᾱ temp values. Therefore, the temporal window was slid with overwrapping, providing a smoother time-series outcome.
The polarimetric calibration was preliminarily applied to all the polarimetric data. The Appendix demonstrates the practical polarimetric calibration methodology using the trihedral corner refractor (CR) adopted in this analysis.

B. Properties of H temp /ᾱ temp and H spatial /ᾱ spatial Parameters Over the Observed Terrain
Several types of land cover exist in the observed postlandslide slope. For example, the slope recovery project covered the metal wire mesh over the observed slope's upper part for restoration. The white dashed line represents the spatial coverage of the metal wire mesh, as shown in Fig. 3(a). Understanding the properties of H temp /ᾱ temp and H spatial /ᾱ spatial parameters over such regions is significant for later landslide monitoring.
We address herein an investigation of three land cover types, namely metal wire mesh cover (ROI-W), bare surface terrain (ROI-G), and vegetation cover (ROI-V). In this investigation, the polarimetric data acquired on August 25, 2018, were employed. There was no rainfall on this day. Fig. 4 presents the digital elevation model (DEM) acquired in July 2018, which shows almost the same land cover as August 2018, indicating the selected ROI location. Note that the land cover had changed In this analysis, we derived both H temp /ᾱ temp and H spatial /ᾱ spatial values using the 12 polarimetric data acquired from 11:02 A.M. to 11:57 A.M. of August 25, 2018. H spatial and α spatial values were derived by the polarimetric data acquired at 11:57 A.M. with a 4 × 3 rectangular spatial window to average 12 pixels equivalent to the number of temporal samples in H temp /ᾱ temp . Note that the same number of averaging samples kept a fair comparison because the averaging number influenced the bias in the H/ᾱ parameters, as demonstrated in Section IV-A. Fig. 5 presents the derived decomposition images. Fig. 6 shows distributions of both H temp /ᾱ temp and H spatial /ᾱ spatial over the three ROIs plotted on the H/ᾱ 2D space. Table I lists the mean decomposition values averaged over each ROI. Nine zones were specified on H/ᾱ 2D space related to specific scattering characteristics (right side, Fig. 6 [12]). These bounds are based on the scattering mechanism's properties and are not dependent on a particular dataset [19]. Although the ground-truth of the investigated vegetation and the terrain surface morphology was not available, the analysis was made to perform a comparative evaluation of the ROIs and between two methods for later landslide monitoring. Consequently, the performed analysis reveals different H/ᾱ results between temporal and spatial averaging. In the following, we provide the interpretation of the decomposition results for each ROI.
1) ROI-W: Similar ROI-averaged values betweenᾱ temp and α spatial were found for ROI-W (ᾱ temp = 49.49 • ,ᾱ spatial = 49.02 • in Table I). These values corresponded to the dipole-like scattering mechanism. The results revealed that the metal wire  mesh produced a dipole-like scattering in Ku-band. This was plausible because the length of one side of the wire mesh is ≈81 mm, which is 4.7 times longer than the Ku-band wavelength (17.4 mm) with a wire thickness of 3.2 mm in diameter [see Fig. 3(c)]. Therefore, metal wire mesh can be assumed to be modeled by a cloud of a randomly oriented dipole-like object in the Ku-band plotted on zone-8 or zone-5 of H/ᾱ 2D space. Furthermore,ᾱ temp values in ROI-W had a wider distribution thanᾱ spatial andᾱ temp of the other two regions [see Fig. 6(a)]. This indicated the existence of a variety of scattering mechanisms over ROI-W, varying pixel-to-pixel. On the contrary, a large difference of 0.52 was observed between the ROI-averaged H temp and H spatial (H temp = 0.32, H spatial = 0.84 in Table I). The spatial multilooking of heterogeneous pixels caused the higher H spatial in terms of the scattering mechanism [4], [13], which is understood by the broader distribution ofᾱ temp . Meanwhile, a lower H temp value revealed the stable variation of the polarimetric response along the temporal axis.
2) ROI-G: The decomposition parameters from ROI-G revealed lowᾱ temp andᾱ spatial results (ROI-averagedᾱ temp = 30 • andᾱ spatial = 21 • ), indicating the surface scattering mechanism typically presented over the bare surface terrain. Theᾱ temp distribution was wider than that ofᾱ spatial , which might be explained by the spatial variation of the terrain surface morphology. In contrast, H spatial results appeared higher than H spatial values, and half of H spatial /ᾱ spatial plots were located on zone-6, which corresponded to the medium entropy surface scattering in Fig. 6(e). The spatial averaging again influenced the increased H values over the terrain due to the variety of scattering mechanisms revealed byᾱ temp distribution.
3) ROI-V: Unlike ROI-W and ROI-G, ROI-V showed a similar distribution between H temp /ᾱ temp and H spatial /ᾱ spatial on the 2-D space. Both distributions were located around zone-5, which corresponded to medium entropy vegetation scattering. H temp over ROI-V depicted the highest values among H temp values of the three ROIs, which was likely caused by the winddriven vegetation temporal fluctuation. The medium H spatial (ROI-averaged H spatial = 0.76) over the ROI-V resulted from the spatial multilooking of random media (heterogeneous pixels).

C. Influence of Rainfall in H temp /ᾱ temp Decomposition
The propagation delay caused by rainfall along a propagation path affects the polarimetric data [20], [21]. Rainfall causes a polarization-dependent phase shift through the propagation in the medium, which leads to the copolarization phase difference (CPD). Polarization-dependent signal attenuation and crosspol interference may also appear. The rain polarization effects mainly occur due to the spheroid shape raindrops driven by gravity and the raindrop alignment (canting angle) [22]. Hence, such parameters on the anisotropic raindrop shape, raindrop size distributions, and raindrop's orientation angle along the propagation path determine co-pol phase shift and attenuation. Heavy rain over a propagation path will severely limit measurement reliability. Therefore, a proper understanding of the rainfall effect in the measured GB-SAR backscattering data is significant.
An 11-day time series of a trihedral CR response was measured and plotted in Fig. 7 to assess rainfall's influence on the polarimetric dataset. Fig. 8 depicts the employed trihedral CR. Fig. 7(a) shows the rainfall rate measured at the observed area by the installed weather sensor. Fig. 7(b) and (c) illustrate the copolarization channel imbalance of the phase and the amplitude measured at a single pixel corresponding to the CR. Fig. 7(d) and (e) show H temp andᾱ temp values derived by the processing chain demonstrated in Section IV-A. At a glance, the channel imbalances of the phase and the amplitude correlated well with the rainfall rate and indicated the same peak time. The CPD and the amplitude ratio reached ≈38 • and −4.9 dB, respectively, at the peak time. Accordingly, the copolarization channel imbalance affected the polarimetric decomposition results. Both H temp andᾱ temp results increased as rainfall began, where both increased up to 0.07°and 13.54 • in the peak time. We also confirmed that H temp time series' peak time (5:30 A.M. June 30) was almost in accordance with the rainfall rate's peak time (6:00 A.M. June 30).
The backscattering from natural terrains, such as soil surface and vegetation, contains an additional rainfall effect when the dielectric constant (moisture) and roughness over the terrain surface turn into change. These surface parameter changes modify the scattering matrix and the polarimetric decomposition parameters. The dielectric constant is one of the parameters modifying the copolarization amplitude imbalance, while the surface roughness increases the cross-polarization term [23]. A H temp simulation demonstrated in [3] revealed that if the copolarization amplitude imbalance becomes 3 dB without any phase difference, H temp reaches nearly 0.1 in the worst case.
According to the abovementioned facts, multiple sources of rainfall effects must be considered depending on the terrain situations. Such effects lead to a difficult polarimetric data interpretation because the separation of rainfall effects and the actual terrain surface change (i.e., change induced by a landslide) is challenging. To separate both phenomena as much as possible, we also analyze the time-series data of a CR and the region unaffected by a landslide.

D. Landslide Monitoring
1) H temp /ᾱ temp Results: Fig. 9(a) Fig. 9). In addition to two defined regions (i.e., ROI-I and ROI-II), the time series of the region unaffected by the landslide is also plotted in Fig. 9 and denoted as ROI-III. The land cover in ROI-III included both metal wire mesh and sparse vegetation. Fig. 3(d) depicts the ROI-III location. After the peak, H temp time series started to decrease until the end of investigation time with significant fluctuations. Consequently, the gap between the landslide regions (ROI-I and II) and ROI-III became smaller.
In Fig. 9(b), all ROIs revealedᾱ temp values close to 45 • , which corresponded to a dipole-like scattering before June 30 because of the wire mesh cover and vegetation. The ROI-I revealed a scattering mechanism transition at around 3:  α temp were apparent over ROI-II before and after the landslide, as shown in Fig. 10(a)-(d). Note that the high H temp region, appearing from 600 to 650 of a slant range corresponded to the vegetation cover, as confirmed by Fig. 3(b). The wind-driven vegetation temporal fluctuation caused high H temp , demonstrated in Section IV-B.
The hourly values in Fig. 9(a) and (b) are plotted on the 2-D H/ᾱ space to jointly describe the temporal transition of H temp andᾱ temp values, as presented in Figs. 11(a) and (b). The plot color represents the observation time from June 26 to July 1, 2019. A total of 144 H temp /ᾱ temp plots are drawn in Fig. 11. Both ROI-I and II showed the temporal transition from zone-8 to zone-5 before the landslide due to the rainfall. After the landslide, the plots on ROI-I moved toward zone-7, while those of ROI-II moved toward zone-9.
2) H spatial /ᾱ spatial Results: Six days of landslide transition were investigated using H spatial /ᾱ spatial parameters in the three ROIs. Fig. 9(c) and (d) displays H spatial andᾱ spatial time series from June 26 until July 1, 2019. A 4× 3 rectangular spatial window was used for the multilooking in H spatial /ᾱ spatial to average 12 pixels. Fully polarimetric data every 30 min were used to derive the time series in Fig. 9(c) and (d).
The H spatial time-series results of the three ROIs appeared higher than H temp values. Before June 30, ROI-II and ROI-III showed similar H spatial values (H spatial ≈ 0.81), while ROI-I depicted a higher H spatial than the other ROIs (H spatial ≈ 0.84). Slight H spatial transitions are shown in Fig. 9(c). From 1:00 A.M. to 1:00 P.M. of June 30, H spatial time series of ROI-I showed an upward trend with a 0.02 increase, whereas ROI-II showed a downward trend with a 0.05 decrease. The ROI-III also exhibited a slight upward trend with a 0.04 increase from 1:00 A.M. to 1:00 P.M.
Theᾱ spatial time-series results over ROI-I and ROII were similar to those ofᾱ temp time series. The ROI-I revealed scattering mechanism transition at around 3:30 A.M. of June 30, while ROI-II revealed one at around 2:00 A.M. of June 30. From the transition time,ᾱ spatial time series in ROI-I showed an upward trend with a 2.3 • increase (from 1:00 A.M. to 1:00 P.M. of June 30), while that of ROI-II dropped with a 2.79 • decrease. ROI-III showed a 2.06 • increase from 1:00 A.M. to 1:00 P.M. on June 30. The amount ofᾱ spatial change in ROI-II was smaller than that ofᾱ temp , whereas ROI-I presented a comparable amount of change betweenᾱ temp andᾱ spatial time series. Fig. 10(e)-(h) shows the processed H spatial andᾱ spatial images observed at 8:00 P.M. of June 29 and 11:00 A.M. of June 30 [similar observation time as in Fig. 10(a)-(d)], showing the derived images before and after the landslide. At a glance, visual H spatial andᾱ spatial changes were hardly seen from the comparison images over the landslide region. H spatial results higher than H temp images can be confirmed over the upper part of the slope due to the metal wire mesh cover.
The temporal transition of H spatial /ᾱ spatial time series was plotted on the 2-D space in Fig. 11(c) and (d). Although slight H spatial /ᾱ spatial transitions were confirmed in both ROI-I and ROI-II, most of the plots remained in zone-5, which corresponded to medium entropy vegetation scattering. After the landslide, the plots on ROI-I moved toward zone-4, while those on ROI-II moved toward zone-6.

V. DISCUSSION
This section presents an interpretation of H temp /ᾱ temp timeseries results in the previous section concerning the landslide event. The applicability of H temp /ᾱ temp target decomposition parameters is also discussed by comparing it with that of the H spatial /ᾱ spatial results.

A. Interpretation
Rainfall makes the interpretation difficult; hence, CR and ROI-III were analyzed. Here, the time-series fluctuations of the H temp /ᾱ temp results in ROI-III were assumed to be caused only by the rainfall effect through the propagation and rainfall-driven soil-moisture and surface roughness changes.
According to the time series of H temp in the three ROIs and the CR, H temp time series' peak time in ROI-I was 1.5 h earlier than the other two regions and 2 h earlier than the CR. Only ROI-I showed an abrupt H temp increase of 0.12 from 3:00 A.M. (H temp = 0.58) to 3:30 A.M. (H temp = 0.70), while ROI-II, ROI-III, and CR showed slight H temp changes of 0.03, 0.03, and 0.01, respectively. Therefore, the abrupt increase of H temp in ROI-I can be assumed to be caused by the landslide because the CR and ROI-III revealed fewer rainfall effects (small changes of H temp ) at the same time. Theᾱ temp time-series results supported this assumption where those of ROI-I and ROI-II revealed the changes of the temporal trend. Theᾱ temp time series in ROI-I started to increase from 3:30 A.M. In contrast, ROI-II showed a decrease from 2:00 A.M. From the ROI-II results, we expect that debris flowing over ROI-II started around 2:00 A.M. On the contrary, the main avalanche in ROI-I was expected to occur around 3:30 A.M. because both H temp andᾱ temp in ROI-I changed at this time. Fig. 9(b) reveals that ROI-I and ROI-II have different temporal transitions of the scattering mechanisms after the landslide. According to Fig. 11(a), H temp /ᾱ temp in ROI-I are plotted on  zone-7, which corresponds to multiple-scattering events after the landslide. The rupture surface can be expected to cause such scattering events after the loss of mass due to the avalanche in ROI-I. On the contrary, ROI-II indicated the surface scattering mechanism after the landslide caused by the debris accumulation over the ROI-II.

B. Comparison
Between H temp /ᾱ temp and H spatial /ᾱ spatial H temp deals with the temporal stationarity of the scattering mechanism; thus, it is only sensitive to temporal scattering variation. Accordingly, H temp becomes an index of the temporal scattering mechanism change. This H temp property is especially appropriate to change detection of the corresponding scatterers. Meanwhile, H spatial reflects the spatial variation of the scattering mechanism. Therefore, temporal change detection by H spatial deals with the temporal difference of H spatial . If H spatial values are similar before and after the scatterer change, change detection by H spatial becomes challenging. This information difference of H temp and H spatial leads to quite different time-series results between Fig. 9(a) and (c).
Importantly, H temp /ᾱ temp can attain the full spatial resolution without any spatial averaging or mixing. When the scattering mechanism changes pixel-to-pixel, such as the presented Ku-band dataset, spatial averaging causes the mixing of heterogeneous pixels, which increases H spatial values [24]. The mixing of different distributions potentially leads to a wrong interpretation of the nature of scatterers and less spatial contrast (or, in other words, remove the spatial signature). Hence, the smaller decrease ofᾱ spatial compared toᾱ temp over ROI-II after the landslide can be explained by the mixing of the debris cover pixels and the neighboring vegetation pixels. In this case, α spatial deals with the contributions of both surface and volume scattering whileᾱ temp measures only surface scattering. In addition, the high H spatial increases the inability to distinguish between the scattering mechanisms because a feasible range of α gets smaller when the underlying H increases [19].
The H temp /ᾱ temp decomposition sacrifices temporal resolution. Nevertheless, such a drawback can be practically solved by increasing the data acquisition rate, which GB-SAR easily realizes.

VI. CONCLUSION
We applied herein the temporal H/ᾱ decomposition approach to landslide monitoring as a GB-SAR continuous data analysis. We employed the dataset observed in the framework of a near-real-time polarimetric GB-SAR monitoring campaign in Minami-Aso, Kumamoto, Japan.
First, the properties of H temp /ᾱ temp and H spatial /ᾱ spatial parameters on the selected regions were investigated. We selected three land cover types over the observed slope, that are metal wire mesh cover (ROI-W), bare surface terrain (ROI-G), and vegetation cover (ROI-G). Theᾱ temp results revealed a broader distribution thanᾱ spatial , especially for ROI-W and ROI-G. These broaderᾱ temp results indicated a high scattering spatial diversity. ROI-W and ROI-G showed low H temp , indicating the scattering mechanism's temporal stationarity and a high H spatial caused by the spatial multilooking of heterogeneous pixels. In contrast, the ROI-V showed a similar distribution between H temp /ᾱ temp and H spatial /ᾱ spatial on the 2-D space.
For a better interpretation of H temp /ᾱ temp in the GB-SAR dataset, the impact of rainfall on the polarimetric parameters was investigated using the acquired time-series data. The trihedral CR polarimetric response during rainfall revealed the copolarization phase difference and the polarization-dependent signal attenuation. Accordingly, the derived H temp andᾱ temp parameters were correlated with the rainfall rate, indicating a significant impact of the rainfall effect on the decomposition results.
Triggered by the heavy rain, a small-scale landslide occurred over the observed post-landslide slope on June 30, 2019. The time series H temp /ᾱ temp parameters were analyzed to investigate the feasibility of the detection and temporal scattering mechanism transition over the landslide regions. Three regions were investigated, namely ROI-I in the avalanche region (upper part of landslide), ROI-II in the debris cover region (middle and lower part of landslide), and ROI-III in the region unaffected by the landslide.
The time-series analysis of H temp /ᾱ temp over ROI-I revealed a gentle increase ofᾱ temp and an abrupt increase of H temp presumably caused by the avalanche. ROI-II also showed a clear temporal transition ofᾱ temp because of the debris cover after the landslide.
Furthermore, H temp /ᾱ temp target decomposition parameters' applicability was discussed by comparing it with the classical H spatial /ᾱ spatial parameters. In summary, the following are the characteristics of temporal averaging compared to spatial averaging supported by the time-series results: 1) The H temp parameter can be an index of the temporal scattering mechanism change. This H temp property is especially appropriate to the change detection of the corresponding scatterers. 2) A full spatial resolution can be attained. Hence, temporal averaging can avoid the spatial mixing of heterogeneous pixels, which causes an increase of H and the mixing of the polarimetric information. In particular, GB-SAR can receive the abovementioned benefits because many polarimetric data are available in its monitoring with a fixed position. This is a great advantage compared to spaceborne SAR. Although the InSAR technique's displacement estimate is a major objective in GB-SAR campaigns so far, H temp /ᾱ temp parameters can give us additional key information regarding when and how the terrain surface has changed, as demonstrated in our analysis.

Polarimetric Calibration
Polarimetric calibration is an important step that is required to derive reliable polarimetric parameters. The system model is defined as follows using the transmitter and receiver distortion matrices [T] and [R] where [M] and [S] are the measured and desired scattering matrices, respectively. Neglecting the channel imbalance between HV and VH (cross-pol) and assuming reciprocity for the quasi-monostatic radar system, we can simplify the distortion matrices as [25], [26] M HH M HV M VH M VH = 1 δ y δ x f · S HH S HV S VH S VH · 1 δ x δ y f , where δ x and δ y account for the cross-talk of the horizontal and vertical polarizations, respectively, f represents a channel imbalance (amplitude and phase) between the HH and VV com-