C++ API

I strongly advise to use the Python bindings instead of calling the C++ code directly. Primarily because this is the way I call and test the code all the time.

In any case, just include despeckcl.h and call the following functions for denoising a SAR interferogram, or, in the case of NL-SAR, any SAR covariance matrix.

Boxcar

int boxcar(float *ampl_master, float *ampl_slave, float *phase, float *ref_filt, float *phase_filt, float *coh_filt, const int height, const int width, const int window_width, std::vector<std::string> enabled_log_levels)

Filters the input with a boxcar filter

Parameters:
  • ampl_master – the amplitude of the master image
  • ampl_slave – the amplitude of the slave image
  • phase – the interferometric phase of the master and slave images
  • ref_filt – the filtered reflectivity estimate
  • phase_filt – the interferometric phase estimate
  • coh_filt – the coherence estimate
  • height – height of the images
  • width – width of the images
  • window_width – the window width of the boxcar window, has to be an odd number
  • enabled_log_levels – enabled log levels, log levels are: error, fatal, warning, debug, info

NL-InSAR

int nlinsar(float *ampl_master, float *ampl_slave, float *phase, float *ref_filt, float *phase_filt, float *coh_filt, const int height, const int width, const int search_window_size, const int patch_size, const int niter, const int lmin, std::vector<std::string> enabled_log_levels)

Filters the input with the NL-InSAR filter

Parameters:
  • ampl_master – the amplitude of the master image
  • ampl_slave – the amplitude of the slave image
  • phase – the interferometric phase of the master and slave images
  • ref_filt – the filtered reflectivity estimate
  • phase_filt – the interferometric phase estimate
  • coh_filt – the coherence estimate
  • height – height of the images
  • width – width of the images
  • search_window_size – width of the search window, has to be an odd number
  • patch_size – width of the patch, has to be an odd number
  • niter – number of iterations
  • lmin – minimum number of looks for the smoothing step
  • enabled_log_levels – enabled log levels, log levels are: error, fatal, warning, debug, info

NL-SAR

The functions for NL-SAR are overloaded and can be called either passing in the master and slave acquisitions’ amplitudes and their interferometric phase for InSAR, or point to the covariance matrix for InSAR and PolSAR. There is slight changes to Charles Deledalle’s original implementation (https://www.math.u-bordeaux.fr/~cdeledal/nlsar.php). This implementation only uses a quadratic search window of fixed size.

using nlsar_stats_collection = std::map<nlsar::params, nlsar::stats>

Stores for each parameter set, i.e. the combinations of scale size and patch size, the corresponding statistics calculated on a homogeneous area.

class nlsar::params

Stores one parameter set

int patch_size
int scale_size
class nlsar::stats

Computes and stores the statitics gathered over a homogenous terrain

nlsar::stats::stats(std::vector<float> dissims, unsigned int lut_size)
Parameters:
  • dissims – computed dissimilarities on the homogenous area
  • lut_size – size of the quantilles’ look-up-table
nlsar_stats_collection nlsar_training(float *ampl_master, float *ampl_slave, float *phase, const int height, const int width, const std::vector<int> patch_sizes, const std::vector<int> scale_sizes)

Gathers the statistics for NL-SAR’s weighting kernel on a homogeneous area. The area should be small. Around 30 by 30 pixels.

Parameters:
  • ampl_master – the amplitude of the master image
  • ampl_slave – the amplitude of the slave image
  • phase – the interferometric phase of the master and slave images
  • patch_sizes – widths of the patches, have to be odd numbers
  • scale_sizes – widths of the scales, have to be odd numbers
nlsar_stats_collection nlsar_training(float *covmat_raw, const int height, const int width, const std::vector<int> patch_sizes, const std::vector<int> scale_sizes)

Gathers the statistics for NL-SAR’s weighting kernel on a homogeneous area. The area should be small. Around 30 by 30 pixels.

Parameters:
  • covmat_raw – the unfiltered covariance matrix
  • patch_sizes – widths of the patches, have to be odd numbers
  • scale_sizes – widths of the scales, have to be odd numbers
void store_nlsar_stats_collection(nlsar_stats_collection nsc, std::string filename)

Stores the computed statistics in a file for later use.

nlsar_stats_collection load_nlsar_stats_collection(std::string filename)

Loads previously computed statistics from a file.

int nlsar(float *ampl_master, float *ampl_slave, float *phase, float *ref_filt, float *phase_filt, float *coh_filt, const int height, const int width, const int search_window_size, const std::vector<int> patch_sizes, const std::vector<int> scale_sizes, std::map<nlsar::params, nlsar::stats> nlsar_stats, const float h_param, const float c_param, std::vector<std::string> enabled_log_levels)

Filters the input with the NL-SAR filter

Parameters:
  • ampl_master – the amplitude of the master image
  • ampl_slave – the amplitude of the slave image
  • phase – the interferometric phase of the master and slave images
  • ref_filt – the filtered reflectivity estimate
  • phase_filt – the interferometric phase estimate
  • coh_filt – the coherence estimate
  • height – height of the images
  • width – width of the images
  • search_window_size – width of the search window, has to be an odd number
  • patch_sizes – widths of the patches, have to be odd numbers
  • scale_sizes – widths of the scales, have to be odd numbers
  • std::map<nlsar::params, nlsar::stats> nlsar_stats – statistics computed on homogenous training area for all parameters
  • h_param – parameter h of the weighting kernel. In my experience 3 is a good default value.
  • c_param – parameter c of the weighting kernel. Can most probably always be set to 49, as in the original paper.
  • enabled_log_levels – enabled log levels, log levels are: error, fatal, warning, debug, info
int nlsar(float *covmat_raw, float *covmat_filt, const int height, const int width, const int search_window_size, const std::vector<int> patch_sizes, const std::vector<int> scale_sizes, std::map<nlsar::params, nlsar::stats> nlsar_stats, const float h_param, const float c_param, std::vector<std::string> enabled_log_levels)

Filters the input with the NL-SAR filter

Parameters:
  • covmat_raw – the unfilterered covariance matrix
  • covmat_filt – the filtered covariance matrix
  • height – height of the image
  • width – width of the image
  • search_window_size – width of the search window, has to be an odd number
  • patch_sizes – widths of the patches, have to be odd numbers
  • scale_sizes – widths of the scales, have to be odd numbers
  • std::map<nlsar::params, nlsar::stats> nlsar_stats – statistics computed on homogenous training area for all parameters
  • h_param – parameter h of the weighting kernel. In my experience 3 is a good default value.
  • c_param – parameter c of the weighting kernel. Can most probably always be set to 49, as in the original paper.
  • enabled_log_levels – enabled log levels, log levels are: error, fatal, warning, debug, info

Goldstein

int goldstein(float *ampl_master, float *ampl_slave, float *phase, float *ref_filt, float *phase_filt, float *coh_filt, const unsigned int height, const unsigned int width, const unsigned int patch_size, const unsigned int overlap, const float alpha, std::vector<std::string> enabled_log_levels)

Filters the input with the Goldstein filter

Parameters:
  • ampl_master – the amplitude of the master image
  • ampl_slave – the amplitude of the slave image
  • phase – the interferometric phase of the master and slave images
  • ref_filt – the filtered reflectivity estimate
  • phase_filt – the interferometric phase estimate
  • coh_filt – the coherence estimate
  • patch_size – width of the patch for each 2D FFT
  • overlap – overlap of the patches
  • alpha – strength of filtering
  • enabled_log_levels – enabled log levels, log levels are: error, fatal, warning, debug, info
Returns:

a tuple containing the reflectivy, phase and coherence estimates
Rtype:

tuple of ndarrays