snaphu − phase unwrapping algorithm for SAR interferometry
snaphu [options] [infile] [linelength] [options]
snaphu is a statistical-cost network-flow algorithm for phase unwrapping. Given an input interferogram and other observable data, snaphu attempts to compute congruent phase-unwrapped solutions that are maximally probable in an approximate a posteriori sense. The algorithm’s solver routine is based on network optimization. By default, snaphu assumes that its input is a synthetic aperture radar (SAR) interferogram measuring surface topography. Deformation measurements are assumed if the −d option is given. Smooth, generic data are assumed if the −s option is given.
This man page documents only the syntax and usage of snaphu. Its theoretical foundations are discussed in the references cited below.
The most common input parameters may be given on the command line, while many other twiddle parameters are handled via the −f option and configuration files. At the very least, the name of a wrapped-phase input file and its line length must be specified. Range should increase towards the right in the interferogram, and the flat-earth phase ramp should be removed from the input interferogram before snaphu is run. For deformation interferograms, phase variations due to topography should be removed as well.
Except for the input file name and the line length, all input parameters take default values if not specified. However, these parameters should be customized whenever possible since the accuracy of the solution depends on how well the statistics of the estimation problem are modeled. To avoid poor-quality solutions, users are strongly encouraged to provide their best estimates of the relevant problem parameters. Parameters are set in the order in which they are given on the command line, so multiple configuration files or options may be given, with later values overriding earlier ones.
Allowable file formats are detailed below. The default format for the input file is COMPLEX_DATA, but any of the described formats may be used. If either of the ALT_LINE_DATA or ALT_SAMPLE_DATA formats are used, the magnitude and phase (in radians from 0 to 2pi) of the interferogram should be in the first and second channels of the file, respectively. If the FLOAT_DATA format is used, the input file should contain only the phase of the interferogram (in radians from 0 to 2pi); the magnitude may be passed with the −m option.
Read brightness data from the file ampfile. The file should contain the amplitudes (not powers) of the two individual SAR images forming the interferogram if the formats ALT_SAMPLE_DATA (default) or ALT_LINE_DATA are used. It should contain an average of those two images if the FLOAT_DATA format is used. If (1) the amplitudes of both images are available, (2) the interferogram magnitude is also available, and (3) the −c option is not used, then a coherence estimate is automatically formed from the available data. The number of looks used for this estimate can be set in a configuration file. If no amplitude or power data are specified, then the magnitude of the input interferogram is used as the average amplitude, and no coherence estimate is formed. Note that the magnitude of the interferogram is not equal to the average amplitude of the SAR images. The amplitude data should be in the same system of units used for the input interferogram, and also coregistered to it.
Similar to the −a option, except the data in the specified file is assumed to represent the powers of the two individual SAR images.
For topography mode, use Bperp (decimal value, in meters) as the value of the perpendicular component of the interferometric baseline. The sign is defined such that Bperp is negative if the unwrapped phase increases with the elevation. By default, repeat-pass or ping-pong mode is assumed; for single-antenna-transmit data, the value of Bperp should be halved, or the transmit mode should be set accordingly in a configuration file (see the −f option). The baseline value is only used in topography mode.
Read correlation data from the file corrfile. The correlation data should be the same size as, and registered to, the input interferogram. Consequently, a raw correlation estimate may need to be upsampled if it incorporates more looks than the interferogram. If the −c option is not given, a coherence estimate is formed from the available data if possible. Otherwise, a uniform default coherence is assumed for the entire interferogram. If the ALT_LINE_DATA (default) or ALT_SAMPLE_DATA formats are used, the correlation data should be in the second data channel of the file; the first channel is ignored. The FLOAT_DATA format may also be used. The correlation values should be between zero and one, inclusive.
Parse the string configstr as if it were a line from a configuration file containing a keyword-value pair (see the -f option). Configuration lines generally have whitespace between the keyword and the value, so configstr will usually need to be surrounded by quotation marks on a command line so that the shell does not split it into separate arguments (snaphu itself does not require or allow quotation marks, however). The syntax for how quotation marks are handled is defined by the shell. Multiple instances of the -C option may be used in order to specify multiple configuration inputs. Later instances of the -C option take precedence over both earlier instances of the -C option and the configurations specified by earlier instances of the -f option.
Run in deformation mode. The problem statistics and resulting cost functions are based on the assumption that the true unwrapped phase represents surface displacement rather than elevation.
Flatten using the unwrapped phase estimate in the file estimatefile. The estimate is subtracted from the input interferogram before unwrapping, and is inserted back into the solution just before the output is written. The estimate also affects the cost functions used, since subtracting a constant from a random variable shifts the probability density function of the random variable. If the formats ALT_LINE_DATA (default) or ALT_SAMPLE_DATA are used, the unwrapped estimate (in radians) should be in the second data channel of the file; the first channel is ignored. The FLOAT_DATA format may also be used.
Read configuration parameters from file configfile. The file is parsed line by line for key-value pairs. Template configuration files are included with the snaphu source code: snaphu.conf.full contains all valid key-value pairs; snaphu.conf.brief contains the most important parameters. Lines not beginning with alphanumeric characters are treated as comment lines. Command line options specified after −f will override parameters specified in the configfile and vice versa. The −f option may be given multiple times with different configuration files, with parameters in later-specified files overriding those in earlier ones.
Grow a connected component mask for the unwrapped solution and write the mask to the file maskfile. A connected component is a region of pixels in the solution that is believed to have been unwrapped in a relative, internally self-consistent manner according to the statistical costs used. Regions that are smaller than a preselected threshold are masked out. Parameters for this option can be set in the configuration file. The connected component file is composed of unsigned characters by default, with all pixels of the same value belonging to the same connected component and zero corresponding to masked pixels.
Grow a connected component mask (see the −g option) for the input data array, assuming that it is already unwrapped, and write the mask to the file maskfile. Statistical cost functions are computed for forming the mask, but a new unwrapped solution is not computed.
Print a help message summarizing command-line options and exit.
Run in initialize-only mode. Normally, snaphu uses either an approximate minimum spanning tree (MST) algorithm or a minimum cost flow (MCF) algorithm for generating the initialization to its iterative, modified network-simplex solver. If −i is given, the initialization is written to the output and the program exits without running the iterative solver.
Keep temporary tile outputs. If this option is specified when snaphu runs in tile mode, the temporary directory where tile outputs are stored will be left in place rather than deleted. The tile-mode initialization of the -S option will also be left in place rather than deleted.
Log all runtime parameters and some other environment information into the specified file. The log file is a text file in the same format as a configuration file.
Read interferogram magnitude data from the specified file. This option is useful mainly if the wrapped-phase input file is given as a set of real phase values rather than complex interferogram values. The interferogram magnitude is used to form a coherence estimate if appropriate amplitude data are given as well. The default file format is FLOAT_DATA. If the formats ALT_LINE_DATA or ALT_SAMPLE_DATA are used, the magnitude should be in the first data channel of the file; the second channel is ignored. If the COMPLEX_DATA format is used, the phase information is ignored. Areas where the magnitude is zero are treated as masked areas (see the −M option).
Read a byte mask from the specified file. The mask file should be the same size as the input array to be unwrapped. The mask should have the binary (not ASCII) value 0 where pixels of the input array are to be ignored during the primary optimization stage of the program. Values elsewhere should be binary 1. Masking is not applied until after the initialization stage of snaphu. Masked areas are treated as areas in which the solution phase value is irrelevant to the solution cost. The magnitude of the interferogram is set to zero in masked areas in the output file. Areas with zero magnitude in the input data are treated as masked areas as well. Areas near the edges of the input may also be masked via options in a configuration file.
Run in no-statistical-costs mode. If the −i or −p options are given, snaphu will not use statistical costs. Information from a weight file (−w option) will still be used if given.
Write the unwrapped output to a file called outfile. If the file formats ALT_LINE_DATA (default) or ALT_SAMPLE_DATA are used, the unwrapped phase is written into the second data channel, while the interferogram magnitude is written into the first channel. The format FLOAT_DATA may also be used.
Run in Lp-norm mode with p=value, where value is a nonnegative decimal. Instead of statistical cost functions, the program uses Lp cost functions with statistically based weights (unless −n is also given). Solutions are still always congruent. Moreover, congruence is enforced within the solver routine, not as a post-optimization processing step. Therefore, if p=2, for example, least-squares cost functions are used, but the solution will probably be more accurate than one generated from a transform-based least-squares algorithm.
Run in quantify-only mode. The input data are assumed to be unwrapped already, and the total cost of this solution is calculated and printed. The unwrapped phase is wrapped assuming congruence for the cost calculation. Round-off errors may limit the precision of the quantified cost. See the −u option for allowable file formats.
Run in smooth-solution mode. The problem statistics and resulting cost functions are based on the assumption that the true unwrapped phase represents a generic surface with no discontinuities. This is the same as deformation mode with the DEFOMAX parameter set to zero.
Do single-tile re-optimization after tile-mode initialization. If this option is specified, snaphu will run in tile mode to generate an unwrapped solution, which is then used as the initialization to a single-tile optimization that produces the final unwrapped output. The tile-mode initialization may itself be initialized by the MST or MCF algorithms (or an input unwrapped phase file) as normal. This option is equivalent to running an instance of snaphu in tile mode, then running another instance of snaphu taking the tile-mode output as an unwrapped input via the -u option. Tile parameters must be specified when using this option. This approach is often faster than unwrapping an interferogram as a single tile from an MST initialization, especially if multiple processors are used.
Run in topography mode. The problem statistics and resulting cost functions are based on the assumption that the true unwrapped phase represents surface elevation. This is the default.
Assume that the input file is unwrapped rather than wrapped. The algorithm makes iterative improvements to this solution instead of using an initialization routine. The input file may be in the formats ALT_LINE_DATA (default) or ALT_SAMPLE_DATA; the interferogram magnitude should be in the first data channel and the unwrapped phase should be in the second data channel. The format FLOAT_DATA may also be used.
Run in verbose mode. Extra information on the algorithm’s progress is printed to the standard output.
Read external, scalar weights from file weightfile. The weights, which should be positive short integers, are applied to whichever cost functions are used. There is one weight value for each arc in the network, so weightfile should be the concatenation of raster horizontal-flow and vertical-flow arc weights. Thus, for an N row by M column interferogram, weightfile would consist of a rasterized (N-1) by M array followed by a rasterized N by (M-1) array of short integer data. This option is not well tested.
−−aa ampfile1 ampfile2
Amplitude data are read from the files specified. The data from the two individual SAR images forming the interferogram are assumed to be separately stored in files ampfile1 and ampfile2. These files should be in the format FLOAT_DATA. This option is similar to the −a option.
−−AA pwrfile1 pwrfile2
Similar to the −−aa option, but power data are read from the specified files.
Assemble the tile-mode temporary files from a previous tile-mode run of snaphu, possibly with different secondary optimization parameters, to produce a new unwrapped solution. The tile directory name must be specified with the −−tiledir option. Most configuration options (from the command line and any configuration files) must be specified similar to the previous run, including the output file name from which the names of temporary tile files are derived. The previous output file may hence be overwritten by the new output file. This option is useful if the user wishes to modify tile-assembly parameters without unwrapping the individual tiles over again.
Print the software copyright notice and bug report info, then exit.
Read statistical cost arrays from file costfile. This file should be in the format written by the −−costoutfile option. The cost file does not control whether snaphu runs in topography, deformation, or smooth-solution mode; the latter two must be specified explicitly even if costfile was generated while running in those modes.
Write statistical cost arrays to file costfile. This option can be used with the −−costinfile option to save the time of generating statistical costs if the same costs are used multiple times.
Dump all sorts of intermediate arrays to files.
Use a minimum spanning tree (MST) algorithm for the initialization. This is the default.
Use a minimum cost flow (MCF) algorithm for the initialization. The cs2 solver by Goldberg and Cherkassky is used. The modified network-simplex solver in L1 mode may give different results than the cs2 solver, though in principle both should be L1 optimal.
Use n parallel processes when in tile mode. The program forks a new process for each tile so that tiles can be unwrapped in parallel; at most n processes will run concurrently. Forking is done before data are read. The standard output streams of child processes are directed to log files in the temporary tile directory.
−−piece firstrow firstcol nrow ncol
Read and unwrap only a subset or part of the input interferogram. The read piece is the nrow by ncol rectangle whose upper left corner is the pixel at row firstrow and column firstcol (indexed from 1). All input files (such as amplitude, coherence, etc.) are assumed to be the same size as the input phase file. All output files are nrow by ncol.
−−tile ntilerow ntilecol rowovrlp colovrlp
Unwrap the interferogram in tile mode. The interferogram is partitioned into ntilerow by ntilecol tiles, each of which is unwrapped independently. Tiles overlap by rowovrlp and colovrlp pixels in the row and column directions. The tiles are then segmented into reliable regions based on the cost functions, and the regions are reassembled. The program creates a subdirectory for temporary files in the directory of the eventual output file (see the −−tiledir and -k options). Tiles can be unwrapped in parallel (see the −−nproc option).
Use dirname as the name of the directory in which temporary tile-model outputs are written and/or read. The directory is created if it does not exist, and it is removed at the end of the run unless the -k or −−assemble options are specified.
The formats of
input files may be specified in a configuration file. All of
these formats are composed of raster, single-precision
(float, real*4, or complex*8) floating-point data types in
the platform’s native byte order. Data are read line
by line in row-major order (across then down, with the
column index varying faster than the row index). Regardless
of the file format, all input data arrays should have the
same number of samples in width and depth and should be
coregistered to one another. Note that weight files and cost
files have their own formats. The allowable formats for
other data files are described below.
Alternating floats correspond to the real (in-phase) and imaginary (quadrature) components of complex data samples. The specified line length should be the number of complex samples (pairs of real and imaginary samples) per line.
Alternating lines (rows) of data correspond to lines of purely real data from two separate arrays. The first array is often the magnitude of the interferogram, and the second may be unwrapped phase, coherence, etc. This is also sometimes called hgt, rmg, or line-interleaved format.
Alternating samples correspond to purely real samples from two separate arrays. This format is sometimes used for the amplitudes of the two SAR images.
The file contains data for only one channel or array, and the data are purely real.
Unwrap a wrapped topographic interferogram called ‘‘wrappedfile’’ whose line length is 1024 complex samples (output will be written to a file whose name is compiled into the program):
snaphu wrappedfile 1024
Unwrap the same file as above, but use brightness information from the file ‘‘ampfile,’’ set the perpendicular baseline to -165 m at midswath, and place the output in a file called ‘‘unwrappedfile’’ (coherence data are generated automatically if ‘‘wrappedfile’’ contains complex data and ‘‘ampfile’’ contains amplitude data from both SAR images):
wrappedfile 1024 -a ampfile \
-b -165 -o unwrappedfile
Unwrap the interferogram as above, but read correlation information from the file ‘‘corrfile’’ instead of generating it from the interferogram and amplitude data:
wrappedfile 1024 -a ampfile -c corrfile \
-b -165 -o unwrappedfile
The following is equivalent to the previous example, but input parameters are read from a configuration file, and verbose output is displayed:
# This is a comment line which will be ignored
EOF (ie, Ctrl-D)
snaphu -v -f configfile wrappedfile 1024
Unwrap the same interferogram, but use only the MST initialization (with scalar statistical weights) and write the output to ‘‘mstfile’’:
snaphu -f configfile -i wrappedfile 1024 -o mstfile
Read the unwrapped data in ‘‘mstfile’’ and use that as the initialization to the modified network-simplex solver:
snaphu -f configfile -u mstfile 1024 -o unwrappedfile
Note that in the previous two examples, the output file name in the configuration file is overrided by the one given on the command line. The previous two commands together are in principle equivalent to the preceding one, although round-off errors in flow-to-phase conversions may cause minor differences
Unwrap the interferogram as above, but use the MCF algorithm for initialization:
snaphu -f configfile wrappedfile 1024 --mcf
Unwrap the interferogram once again, but first flatten it with the unwrapped data in ‘‘estfile,’’ then reinsert the subtracted phase after unwrapping:
snaphu -f configfile wrappedfile 1024 -e estfile
The following assumes that the wrapped input interferogram measures deformation, not topography. Unwrap the interferogram with the given correlation data:
snaphu -d wrappedfile 1024 -c corrfile
Unwrap the input interferogram by minimizing the unweighted congruent L2 norm:
snaphu -p 2 -n wrappedfile 1024
Unwrap the interferogram as a three-by-four set of tiles that overlap by 30 pixels, with the specified configuration file, using two processors:
wrappedfile 1024 -f configfile \
--tile 3 4 30 30 --nproc 2
The program may print a warning message about costs being clipped to avoid overflow. If too many costs are clipped, the value of COSTSCALE may need to be decreased in a configuration file (via the −f option). If the program prints a warning message about an unexpected increase in the total solution cost, this is an indication that too many costs are clipped. It is usually okay if just a few costs are clipped.
In topography mode, if the unwrapped result contains too many discontinuities, try increasing the value of LAYMINEI or decreasing the value of LAYCONST. The former determines the normalized intensity threshold for layover, and the latter is the relative layover probability. If there are too many discontinuities running in azimuth, try decreasing the value of AZDZFACTOR, which affects the ratio of azimuth to range costs. If the baseline is not known, take a guess at it and be sure its sign is correct. Specify the SAR imaging geometry parameters as well as possible. The defaults assume ERS data with five looks taken in azimuth.
In deformation mode, if the unwrapped result contains too many discontinuities, try increasing the value of DEFOTHRESHFACTOR or decreasing the value of DEFOCONST. If the surface displacement varies slowly and true discontinuities are not expected at all, DEFOMAX_CYCLE can be set to zero. This behavior is also invoked with the −s option. The resulting cost functions will be similar to correlation-weighted L2 cost functions, though the former are not necessarily centered on the wrapped gradients. Congruence is still enforced during rather than after optimization.
The program can be run in initialize-only (−i) mode for quick down-and-dirty MST or MCF solutions.
Once the iterative solver has started, snaphu traps the interrupt (INT) and hangup (HUP) signals. Upon receiving an interrupt, for example if the user types Ctrl-C, the program finishes a minor iteration, dumps its current solution to the output, and exits. If a second interrupt is given after the first (caught) interrupt, the program exits immediately. If a hangup signal is received, the program dumps its current solution then continues to execute normally.
Upon successful termination, the program exits with code 0. Errors result in exit code 1.
files may be useful for reference, but are not required.
They are included in the program source distribution and may
be installed somewhere on the system.
Template configuration file setting all valid input parameters (though some may be commented out).
General-purpose template configuration file setting the most important or commonly modified input parameters.
In addition to parameters read from configuration files specified on the command line, default parameters may be read from a system-wide configuration file if such a file is named when the program is compiled.
The −w option has not been tested exhaustively.
Extreme shadow discontinuities (i.e., abrupt elevation drops in increasing range due to cliffs facing away from the radar) are not modeled that well in the cost functions for topography mode.
Abrupt changes in surface reflectivity, such as those of coastlines between bright land and dark water, might be misinterpreted as layover and assigned inappropriate costs.
The algorithm’s behavior may be unpredictable if the costs are badly scaled and excessively clipped to fit into their short-integer data types.
There is no error checking that ensures that the network node potentials (incost and outcost) do not overflow their integer data types.
Automatic flow clipping is built into the MST initialization, but it can give erratic results and may loop infinitely for certain input data sets. It is consequently turned off by default.
Dedicated programs for specific Lp objective functions may work better than snaphu in Lp mode. Note that snaphu enforces congruence as part of the problem formulation, however, not as a post-optimization processing step.
A tree pruning capability is built into the code and can be enabled from a configuration file, but this functionality is experimental and not well tested.
C. W. Chen and H. A. Zebker, ‘‘Two-dimensional phase unwrapping with use of statistical models for cost functions in nonlinear optimization,’’ Journal of the Optical Society of America A, 18, 338-351 (2001).
C. W. Chen and H. A. Zebker, ‘‘Network approaches to two-dimensional phase unwrapping: intractability and two new algorithms,’’ Journal of the Optical Society of America A, 17, 401-414 (2000).
C. W. Chen and H. A. Zebker, ‘‘Phase unwrapping for large SAR interferograms: Statistical segmentation and generalized network models,’’ IEEE Transactions on Geoscience and Remote Sensing, 40, 1709-1719 (2002).