2. Methods
Our InSAR/GNSS integration approach is an extension of standard
published methods (e.g., Tong et al. , 2013; Weiss et al.,2020) although in addition to secular velocity, we also calculate
line-of-sight (LOS) displacement time series [e.g., Neely et
al., 2019]. The GNSS weekly displacements were derived by means of a
median filter [Klein et al. , 2019] of daily time series
estimated as part of a NASA MEaSUREs project [Bock et al. ,
2016]. Moreover, we use the secular velocity from a GNSS-only
interseismic model [Zeng & Shen , 2017] to create
semi-vertical vector InSAR time series from the LOS displacements. A
brief description of the method follows:
- Gather Sentinel-1 Terrain Observations with Progressive Scans (TOPS)
data from multiple tracks and re-assemble into common re-defined
frames, typically 250 km by 500 km.
- Geometrically co-register all SAR acquisitions and construct all
interferograms with perpendicular baseline < 150 m and
temporal separation < 90 days [Xu et al. , 2017;Sandwell et al. , 2016b].
- Mask bodies of water and areas of persistent low coherence regions and
replace them with nearest-neighbors [Shanker & Zebker ,
2009]. This step improves the phase unwrapping accuracy which is
done with Statistical-Cost, Network-Flow Algorithm for Phase
Unwrapping (SNAPHU) [Chen & Zebker , 2002].
- Perform elevation dependent atmospheric phase correction
[Elliott et al. , 2008]. Compute the difference between the
remaining InSAR phase and projected GNSS weekly solutions
[Klein et al. , 2019], interpolate this difference, filter
at 80-km wavelength and remove this difference from each
interferogram.
- Construct time-series using a coherence-based SBAS approach integrated
with atmospheric phase correction using common-scene stacking
[Tymofyeyeva & Fialko , 2015; Tong & Schimdt , 2016;Xu et al. , 2017].
- Subtract a horizontal GNSS velocity model [e.g., Zeng &
Shen , 2017] from the time-series to create semi-vertical InSAR
time-series.
Since Sentinel-1 TOPS data is acquired under burst acquisition mode and
there is occasional inconsistency in data coverage, especially in the
early days of the mission, the frame boundaries in step 1) are a
compromise between spatial coverage and acquisition numbers. The total
number of interferograms generated here is 5230, connecting acquisitions
from 910 dates over 9 tracks. Enhanced spectral diversity
[Prats-Iraola et al. , 2012] is not performed in step 2, since
it will remove an expected tectonic signal that will eventually supply a
third InSAR component [Li et al., 2021]. The estimated
mis-registration could be up to 2/1000 pixel/yr along SAFS and spread
across the scenes, where a constant shift from ESD is inadequate, while
the performance of bivariate approach [Wang et al., 2017] is
yet to be evaluated. Moreover, the common scene stacking time series
approach (step 5) is capable of mitigating along-track orbital errors by
absorbing burst discontinuities, that are random in time, into
atmospheric phase screens [Xu et al. , 2017]. The
nearest-neighbor interpolation in step 3) is implemented so phases are
allowed to vary properly along very long coastlines, and stay connected
through snowy Sierras and heavy vegetations in northern California.
Elevation dependency in step 4) is assumed as a bivariate quadratic
polynomial thus spatial variations in atmospheric contribution are
accounted for. The relatively large, 80-km wavelength filter, that is
applied to the GNSS correction for each interferogram, is sufficient to
absorb the large-scale atmospheric and orbital errors affecting the
InSAR displacements and also accommodate areas such as the Central
Valley having sparse GNSS coverage. A remove-restore approach
[Tong et al. , 2013] using a purely horizontal secular
velocity model is not used because it is incompatible with the
significant vertical signal in Central Valley. Not only is this vertical
deformation distributed over hundreds of kilometers, but it also has a
sharp transition around the edges of the sedimentary basins. The
interpolation of discrepancies between GNSS and InSAR is adopted here,
taking advantage that though the vertical deformation changes
dramatically over a large area, the differences from the two type of
observations maybe systematic and vary slow enough in space to be well
evaluated. When the final velocity is computed, the first and last four
records are not used, mainly because the atmospheric correction approach
gains less constraints when acquisitions are non-evenly distributed.