Figure 2. Schematic diagram of data time series. (Top)
Long-term changes are derived from ALOS2 acquired on orange dots.
Wildfire occurred from July to August 2014, and JAXA modified the center
frequency of PALSAR-2 Beam No. F2-6 data in June 2015 shown with the
cross. (Bottom) Short-term deformation during 2017-2018 as examined by
Sentinel-1 images. We compare the ALOS2 and Sentinel-1 deformation maps
during the five periods, (a)—(e) and (f)—(j).
While ALOS2 has only imaged the area since 2014 its data acquisition
interval is much longer than that of Sentinel-1 (Figure 2). Previous
studies demonstrate that it is not possible to infer the total
subsidence using pre- and post-wildfire SAR images, as the drastic
changes in land cover cause low interferometric coherence (Liu et al.,
2014; Molan et al., 2018). Additionally, JAXA changed the carrier
frequency of PALSAR-2 in June 2015 (Figure 2). Hence, monitoring
long-term deformation using ALOS2 InSAR is possible only since October
2015. Conversely, frequent data acquisition in Sentinel-1 started only
in 2017. Thus, we first performed an inter-comparison between ALOS2 and
Sentinel-1 InSAR, focusing on the seasonal changes in 2017. We stacked
Sentinel-1 interferograms to set the temporal coverages to nearly
identical with those of ALOS2 (Figure 2). Stacking was necessary because
we failed to derive long-term Sentinel-1 interferograms, as the burned
areas quickly lost coherence. During the temporal interval of ALOS2
images Sentinel-1 had more cycles. Therefore, the number of Sentinel-1
stacks varied from three to eight (Figure 2).
Although L-band SAR is known to have better interferometric coherence
than C-band SAR (e.g., Rosen et al., 1996) our study indicated that
Sentinel-1 could maintain a comparable interferometric coherence with
L-band ALOS2 even during winter season. This is likely due to the short
acquisition period of 12 days as well as the somewhat drier snow in the
area in winter that allows microwave to reach the ground. A dry snow
cover of depth less than 1 m is undetectable to microwave radiation,
whereas over wet snow surface scattering dominates (Rees, 2001). The
frequent data acquisition of Sentinel-1 since 2017 allowed us to examine
detailed seasonal changes in surface deformation (Figure 2). Some
Sentinel-1 InSAR pairs in early summer, however, did not show good
coherence, possibly due to snow wetness.
In order to infer long-term temporal changes and cumulative
displacements, we performed SBAS (Small Baseline Subset)-type
time-series analysis (Berardino et al., 2002; Schmidt and Bürgmann,
2003), using 50 high-quality ALOS2 interferograms that included
one-year- as well as short-term interferograms (Figure 3). We could
estimate the average LOS-change rates between each acquisition epoch
without assuming any temporal change models. In contrast to the original
SBAS approach, we did not estimate DEM errors because the
well-controlled orbit, as well as the precise TanDEM-X DEM, have no
sensitivities to those errors.
In order to estimate the errors of the derived time series, we assumed
each original SAR scene contained 0.2 cm errors, and made InSAR data
covariance matrix, following the method of Biggs et al. (2007). The
errors are relatively smaller than those in previous studies of SBAS
analysis (e.g. 0.4 cm in Schmidt et al., 2003; 0.75 cm in Biggs et al.,
2007), because, as noted earlier, we took out the long-wavelength phase
trend form each InSAR image, and the analysis area is smaller (12\(\mathbf{\times\ }\)12 km) than previous studies.