i.sar.coregistration resamples a secondary
Sentinel-1 TOPS (IW, EW) Single Look Complex image onto the geometry of
a reference image of the same sub-swath and relative
orbit, so that the two can be combined pixel by pixel, e.g. by
i.sar.interferometry. Both images are complex maps imported by
r.in.s1slc (measure=complex), with their
metadata.
The processing follows ESA SNAP (microwave toolbox)
Back-Geocoding and Enhanced Spectral Diversity:
Geometric coregistration: every reference pixel is
located on the ground from the reference orbit (zero-Doppler time, slant
range and terrain height), and the ground point is projected into the
secondary image with the secondary orbit. Orbit state vectors are
interpolated with Lagrange polynomials over eight vectors. The terrain
height comes from the dem or, without it, from the
heights of the annotation geolocation grid. The positions are computed
on a grid of nodes (every 10 lines and 100 samples) for two heights
enclosing the DEM range and interpolated to every pixel, where the
height is solved from the DEM.
TOPS deramping: the azimuth spectrum of a TOPS
burst drifts along the burst (antenna steering). Before interpolation,
the secondary signal is deramped and demodulated with the burst Doppler
rate, reference time and Doppler centroid computed from the annotated
azimuth FM rate, Doppler centroid polynomials and steering rate; after
interpolation the ramp is restored at the interpolated position.
Interpolation: a Hann-windowed sinc over 6, 8 or 16
samples, or bicubic or bilinear interpolation
(interpolation). Taps on null samples are skipped and
the weights renormalized; a pixel whose nearest secondary sample is null
is null.
Enhanced spectral diversity (ESD): with burst maps
(see below), the residual azimuth misregistration is estimated in the
overlaps of consecutive bursts, where each ground point is seen twice
under looks whose Doppler frequencies differ by about kt × burst cycle
time. The phase of the double difference interferogram (mk
sk*) (mk+1
sk+1*)* is -2π Δf Δtaz ε for
a misregistration of ε lines. The phases are averaged with inverse
quadratic weights γ²/(1-γ²) of the local phase consistency γ of the
double difference (pixels with γ below esd_coherence
are ignored), and the shift is added to the secondary positions before
the final interpolation. Azimuth coregistration better than 0.01 line is
needed to avoid phase jumps at the burst seams of TOPS interferograms,
and orbits alone rarely give it.
Range refinement (-r flag):
residual range misregistration, e.g. from a slant range timing error, is
estimated as in SNAP Spectral Diversity. Each burst is first
coregistered in a window (xcorr_window, 512 × 512 by
default, clipped to the valid area) at its centre; the de-meaned
amplitudes of the reference and coregistered secondary windows are
cross-correlated (normalized correlation over shifts of up to 8 samples,
peak refined by FFT oversampling of the correlation by 128). The complex
signals are oversampled by 2 in range before detection, which removes
the bias of the amplitude correlation of speckle towards whole samples.
Bursts whose correlation peak is below xcorr_threshold
or whose shift exceeds one sample are discarded; the mean range shift of
the others is added to all the secondary range positions, before ESD and
the final interpolation. It also works on debursted images (one window
per burst segment).
Input images
The reference and secondary options
are basenames of images imported by r.in.s1slc, i.e.
{output}_{swath}_{pol}:
burst maps {basename}_bNN_i and _q,
imported with the -b flag of r.in.s1slc: each
burst is complete, including the lines it shares with its neighbours,
which ESD needs. This is the recommended input. Reference and secondary
bursts are paired by geometry (and their burst IDs are checked when both
products have them); reference bursts that no secondary burst covers are
skipped with a warning.
debursted maps {basename}_i and _q: the
coregistration uses the orbits only (no ESD). Around each burst seam,
the rows between the reference and secondary seams hold data of
different bursts in the two images, i.e. of different parts of the
azimuth spectrum, and are poorly coherent.
Both images must be of the same kind. The module fails if a basename
matches both kinds of maps.
Output
The coregistered secondary maps have the geometry of the reference
maps and their names, the reference basename being replaced by
output: {output}_i and
{output}_q, or {output}_bNN_i and
{output}_bNN_q for each reference burst. They are FCELL
maps (null where the secondary image has no data) with the units,
calibration, semantic label and timestamp of the secondary image.
The extra option adds maps in reference geometry:
the azimuth and range offsets (secondary minus reference position, in
lines of the secondary burst and in samples) and the ellipsoidal terrain
height used (elevation), which
i.sar.interferometry uses to remove the topographic phase.
Each output map carries a
cell_misc/<map>/description.json file with the
secondary product and annotation metadata (orbit, Doppler, bursts), the
reference raster geometry and a coregistration section:
reference and secondary maps and products, burst pairs, DEM,
interpolation, the ESD result (shift, and phase, Doppler separation,
weight and shift of each overlap) and the range refinement result
(shift, and window, azimuth and range offsets and correlation peak of
each burst).
DEM
The dem is any GDAL raster in any CRS covering the
scene; it is reprojected to a WGS84 longitude/latitude grid by bilinear
interpolation. Heights above the EGM96 geoid
(dem_height=geoid, e.g. SRTM, Copernicus DEM) are
converted to ellipsoidal heights with the PROJ grid
us_nga_egm96_15.tif or egm96_15.gtx, which
must be installed (e.g.
projsync --file us_nga_egm96_15.tif). Pixels where the DEM
has no data are null.
Processing device
The per-pixel work (height solution, deramping, interpolation) runs
in the compute library libsarcoreg, on an OpenCL device or
on the host CPU with OpenMP (device), the same code
being compiled for both. OpenCL devices need double precision
(cl_khr_fp64). The default, device=auto,
uses OpenCL whenever a device is usable: the OpenCL GPUs first, then the
OpenCL CPU devices; only without any, it falls back to the host OpenMP
code, with a warning giving the reason. platform
restricts OpenCL to the platforms whose name contains the given text.
nprocs sets the number of OpenMP threads of the host
path.
NOTES
The current project must be unprojected (XY), like the
r.in.s1slc maps. The computational region is not used: every
map is processed in full.
The range refinement estimates one constant shift for the whole
image. It corrects timing errors, not the range offsets of the terrain
relief: without a DEM over hilly terrain it only absorbs their mean.
Burst IDs of products before IPF 3.4 are absent; pairing is then
geometric only.
The module requires NumPy and the GDAL Python bindings, and a C
compiler with OpenMP and the OpenCL headers and ICD loader to build
libsarcoreg.
EXAMPLES
Import two acquisitions burst by burst and coregister the second one
on the first, with the Copernicus DEM:
N. Yagüe-Martínez, P. Prats-Iraola, F. Rodríguez González, R. Brcic,
R. Shau, D. Geudtner, M. Eineder and R. Bamler, Interferometric
Processing of Sentinel-1 TOPS Data, IEEE Transactions on Geoscience
and Remote Sensing, 54(4), 2220-2234, 2016,
doi:10.1109/TGRS.2015.2497902.
ESA, Sentinel-1 Product Specification,
S1-RS-MDA-52-7441.