Multitemporal multitrack InSAR
The SAR data sets spanning from 13 July 2007 to 17 October 2010 include 32, 24, and 19 images in descending (incidence angle = 23°, heading angle = 193°) and ascending (incidence angle = 23°, heading angle = 350°) orbits of the Envisat C-band satellite and ascending (incidence angle = 34.5°, heading angle = 350°) orbit of ALOS L-band satellite, respectively (
Fig. 1 and table S2). This period was chosen because data sets from three different geometries were available. Using these data sets, we generated 428, 151, and 77 interferograms, respectively. For the Envisat interferometric data set, the maximum temporal and perpendicular baselines are set to be 3 years and 300 m. The corresponding numbers are 3 years and 1500 m for the ALOS data set.
Each data set was processed separately to generate three accurate and high-resolution time series of surface deformation. To this end, we applied the wavelet-based InSAR (WabInSAR) algorithm (
33,
34). The geometrical phase was calculated and removed using a 30-m DEM provided by the SRTM (
35) (
www2.jpl.nasa.gov/srtm/) and the satellite precise ephemeris data. The time series of complex interferometric phase noise was then calculated in the wavelet domain and analyzed in a statistical framework to identify elite (that is, less noisy) pixels. The absolute estimate of the phase change for elite pixels was obtained via an iterative 2D sparse phase unwrapping algorithm. Each unwrapped interferogram was corrected for the effect of orbital error (
36) and the topography-correlated component of atmospheric delay (
37). Using a robust regression, the phase changes from the large set of interferograms were inverted. This approach reduced the effects of outliers due to improper phase unwrapping. A high-pass filter, using continuous wavelet transforms, was applied to reduce the temporal component of the atmospheric delay. Lastly, we geocoded all data sets to obtain precise locations of elite pixels in a geographic reference frame. Figure S1 shows the line-of-sight (LOS) velocity associated with each data set.
Assume that
and
are the displacement time series and the associated variances for the data sets, where
N is the number of acquisitions in each data set and
i = 1,2,3 refers to Envisat descending, Envisat ascending, and ALOS ascending data sets, respectively. In the space domain and using a nearest neighbor algorithm, we oversampled data sets 2 and 3 over data set 1. In the time domain, the displacement and noise time series of one track were interpolated on the other. Assuming that, at time
ta,
da and
are the displacement and associated SD, respectively, and that the respective values at time
tb are
db and
, then, at time step
tc, the interpolated displacement and variance are given as
To set up the interpolation in time domain, we first generated a vector with length M including acquisition dates of all three SAR data sets. This vector is the base for temporal oversampling, namely, all three data sets were oversampled on this new vector.
Assume {
y0,
y1, …,
yM}
i = 1,2,3 and {σ
20, …, σ
2M}
i = 1,2,3as the interpolated time series of displacement and variance for a given pixel. Below, we set up a Kalman filter–based framework to combine these three InSAR time series with horizontal velocities of GNSS data sets (E-W and N-S components) to generate a seamless, high-resolution, and accurate time series of east (E), west (W), and U (vertical) motions. The GNSS velocities were obtained from Bürgmann
et al. (
26) and represented a 200-station subset of the regional BAVU velocity field (
38) that relied on data from both continuous and campaign GNSS measurements. We split this data set into two parts: One group of stations was used to combine with the InSAR data sets (so-called tie points), and the other part was used to validate the results (so-called checkpoints) (
Fig. 1). The E and N displacements of the GNSS tie points were interpolated on the location of the elite pixels from data set 1, using a Kriging interpolation approach and inverse distance weighting (
39).
The measurement model at any time step
tk relates the LOS observations to the corresponding E, N, and U displacements
where
C represents the unit vectors projecting the 3D displacements onto the LOS and is a function of heading and incidence angles and
e is the measurement noise equal to σ, which is assumed to be normally distributed.
The dynamic model is used to integrate GNSS and InSAR data and applies a statistical smoothing in time (
40) and is given as
where
V represents the interpolated
E or
N velocities of the GNSS tie points and ε is the system dynamic noise estimated via
Eq. 2. Grewal and Andrews (
40) have provided a recursive solution to
Eqs. 2 and
3. Because this approach is an extension to an existing InSAR time series algorithm (that is, WabInSAR), hereafter, we shall call it WabInSAR-3D.
To validate the WabInSAR-3D approach, we used a synthetic test. To this end, we simulated 3D deformation time series as a combination of a periodic function with a period of 1 year, a Heaviside function with step size of 5 and a linear trend with slopes of 3, 2, and 1 in E, N, and U directions, respectively (fig. S1). Using the geometrical parameters of the three SAR data sets, we projected the simulated 3D displacement field onto the LOS of Envisat descending and ascending and ALOS ascending orbit track, similar to our data set over the SFBA. The simulated LOS time series were inverted to recover the 3D displacement field. We first applied the method presented by Ozawa and Ueda (
41), which jointly inverted data from different look angles and solved for E and U time series. The top row in fig. S1 shows the results from implementing the Ozawa and Ueda (
41) approach together with the simulated E, N, and U time series. Using this approach, no N displacement was retrieved because they assumed that its contribution was negligible due to the near-polar orbiting satellites. For the E component, the periodicity, trend, and jump were recovered accurately, owing to multiple look and heading angles. However, the estimated U component is relatively uncertain, the mismatch between simulated and recovered signal increases near the jump, and the error propagates throughout the signal. The bottom row in fig. S1 presents results from applying the WabInSAR-3D. The estimated E component is accurate and matches the simulated signal slightly better than that of Ozawa and Ueda (
41). For the N component, WabInSAR-3D is able to at least retrieve the trend and slight periodicity, but the jump is lost. The estimated U component here shows a significant improvement over the other approach and retrieves the periodicity, trend, and jump accurately. Now that we have validated WabInSAR-3D using a simulated data set, we applied it to the SAR and GNSS data sets over the SFBA.
The E, N, and U velocities relative to GNSS station LUTZ are shown in
Fig. 2 (A to C), which are obtained from applying the WabInSAR-3D approach to 68 SAR images acquired by Envisat and ALOS satellites, together with horizontal velocities of 26 GNSS tie points, which are randomly selected. We obtained an SD of 2.19 and 2.34 mm/year, comparing the estimated E and N velocities, with horizontal velocities of 29 GNSS checkpoints not used in the combination. Using observations of four continuous GNSS stations of the BARD network relative to station LUTZ, the time series of E, N, and U displacement were also validated (fig. S3).
To transfer the vertical velocities from the central local reference frame relative to GNSS station LUTZ into a stable North America reference frame (NA12), we used vertical velocities of four permanent GNSS stations of the Plate Boundary Observatory (PBO) network that lie in the area of InSAR coverage (
Fig. 1A), provided by Blewitt
et al. (
24), and applied a conformal translation. We found that a scale factor of 1 and uniform shift of 0.25 mm/year were sufficient to firmly tie the results from the combined InSAR and GNSS measurements to the continental reference frame. We obtained an SD of 0.52 mm/year for the transferred U velocity map in the NA12 reference frame to the four PBO stations.