Add IDW interpolation to rubbersheeting algorithm

python/packages/nisar/workflows/crossmul.py

@@ -14,7 +14,7 @@
 
 from isce3.io import HDF5OptimizedReader
 from nisar.products.readers import SLC
-from nisar.workflows import prepare_insar_hdf5
+from nisar.workflows import h5_prep
 from nisar.workflows.compute_stats import compute_stats_real_data
 from nisar.workflows.crossmul_runconfig import CrossmulRunConfig
 from nisar.workflows.helpers import (copy_raster,
@@ -220,6 +220,6 @@ def stats_offsets(h5_ds, freq, pol):
     # get a runconfig dict from command line args
     crossmul_runconfig = CrossmulRunConfig(args, resample_type)
     # prepare RIFG HDF5
-    out_paths = prepare_insar_hdf5.run(crossmul_runconfig.cfg)
+    _, out_paths = h5_prep.get_products_and_paths(crossmul_runconfig.cfg)
     # run crossmul
-    run(crossmul_runconfig.cfg, out_paths['RIFG'], resample_type)
+    run(crossmul_runconfig.cfg, out_paths['RIFG'], resample_type, dump_on_disk=True)

python/packages/nisar/workflows/rubbersheet.py

@@ -12,9 +12,10 @@
 import numpy as np
 from isce3.io import HDF5OptimizedReader
 from isce3.math import offsets_polyfit
+from isce3.unwrap.preprocess import interpret_subswath_mask
 from nisar.products.insar.product_paths import RIFGGroupsPaths
 from nisar.products.readers import SLC
-from nisar.workflows import prepare_insar_hdf5
+from nisar.workflows import h5_prep
 from nisar.workflows.compute_stats import compute_stats_real_hdf5_dataset
 from nisar.workflows.helpers import (get_cfg_freq_pols,
                                      get_ground_track_velocity_product,
@@ -23,6 +24,7 @@
 from nisar.workflows.yaml_argparse import YamlArgparse
 from osgeo import gdal
 from scipy import interpolate, ndimage, signal
+from scipy.spatial import cKDTree
 
 
 def run_rubbersheet_with_polyfit(cfg: dict, output_hdf5: str = None):
@@ -291,6 +293,14 @@ def run_rubbersheet_with_interpolation(cfg: dict, output_hdf5: str = None):
                                                                            zero_doppler_time,
                                                                            dem_file,
                                                                            rubbersheet_dir)
+            valid_mask = None
+            # Apply the subswath mask to the outlier detection
+            if rubbersheet_params['subswath_mask_apply_enabled']:
+                pixel_offsets_mask_path = f'{pixel_offsets_path}/mask'
+                subswath_mask = dst_h5[pixel_offsets_mask_path][()]
+                ref_valid, sec_valid, _ = interpret_subswath_mask(subswath_mask)
+                valid_mask = ref_valid & sec_valid
+
             for pol in pol_list:
                 # Create input and output directories for pol under processing
                 pol_group_path = f'{pixel_offsets_path}/{pol}'
@@ -305,7 +315,7 @@ def run_rubbersheet_with_interpolation(cfg: dict, output_hdf5: str = None):
                     # Identify outliers
                     offset_az_culled, offset_rg_culled = identify_outliers(
                         str(dense_offsets_dir),
-                        rubbersheet_params)
+                        rubbersheet_params, valid_mask)
                     # Fill outliers holes
                     offset_az = fill_outliers_holes(offset_az_culled,
                                                     rubbersheet_params)
@@ -325,7 +335,9 @@ def run_rubbersheet_with_interpolation(cfg: dict, output_hdf5: str = None):
                                   key.startswith('layer')]
                     # Apply offset blending
                     offset_az, offset_rg = _offset_blending(off_product_dir,
-                                                            rubbersheet_params, layer_keys)
+                                                            rubbersheet_params,
+                                                            layer_keys,
+                                                            valid_mask)
 
                     # Get correlation peak path for the first offset layer
                     corr_peak_path = str(f'{off_prod_dir}/{layer_keys[0]}/correlation_peak')
@@ -336,8 +348,16 @@ def run_rubbersheet_with_interpolation(cfg: dict, output_hdf5: str = None):
                     # If there are residual NaNs, use interpolation to fill residual holes
                     nan_count = np.count_nonzero(np.isnan(offset))
                     if nan_count > 0:
-                        offsets[k] = _interpolate_offsets(offset,
-                                                          rubbersheet_params['interpolation_method'])
+                        if rubbersheet_params['interpolation_method'] == 'idw':
+                            # Get the IDW parameters
+                            power, number, radius = ( rubbersheet_params['idw_interpolation'].get(i)
+                                                     for i in ('power', 'number', 'radius'))
+                            offsets[k] = _interpolate_offsets_by_idw(offset, power, number, radius)
+                            # Additional linear interpolation is required to fill the extra outliers beyond the radius
+                            offsets[k] =  _interpolate_offsets(offsets[k],'linear')
+                        else:
+                            offsets[k] = _interpolate_offsets(offset,
+                                                              rubbersheet_params['interpolation_method'])
                     # If required, filter offsets
                     offsets[k] = _filter_offsets(offsets[k], rubbersheet_params)
                     # Save offsets on disk for resampling
@@ -434,7 +454,7 @@ def _write_to_disk(outpath, array, format='ENVI',
     ds.GetRasterBand(1).WriteArray(array)
     ds.FlushCache()
 
-def identify_outliers(offsets_dir, rubbersheet_params):
+def identify_outliers(offsets_dir, rubbersheet_params, mask = None):
     '''
     Identify outliers in the offset fields.
     Outliers are identified by a thresholding
@@ -448,6 +468,8 @@ def identify_outliers(offsets_dir, rubbersheet_params):
         where pixel offsets are located
     rubbersheet_params: cfg
         Dictionary containing rubbersheet parameters
+    mask: np.ndarray
+        A mask indicating the interested region (default: None)
 
     Returns
     -------
@@ -468,6 +490,9 @@ def identify_outliers(offsets_dir, rubbersheet_params):
     elif metric == 'median_filter':
         offset_az = _open_raster(f'{offsets_dir}/dense_offsets', 1)
         offset_rg = _open_raster(f'{offsets_dir}/dense_offsets', 2)
+        if mask is not None:
+            offset_az[~mask] = np.nan
+            offset_rg[~mask] = np.nan
         mask_data = compute_mad_mask(offset_az, window_az, window_rg, threshold) | \
                     compute_mad_mask(offset_rg, window_az, window_rg, threshold)
     elif metric == 'covariance':
@@ -671,7 +696,7 @@ def _fill_nan_with_mean(arr_in, arr_out, neighborhood_size):
     return filled_arr
 
 
-def _offset_blending(off_product_dir, rubbersheet_params, layer_keys):
+def _offset_blending(off_product_dir, rubbersheet_params, layer_keys, mask = None):
     '''
     Blends offsets layers at different resolution. Implements a
     pyramidal filling algorithm using the offset layer at higher
@@ -689,6 +714,8 @@ def _offset_blending(off_product_dir, rubbersheet_params, layer_keys):
         Dictionary containing the user-defined rubbersheet options
     layer_keys: list
         List of layers within the offset product
+    mask: np.ndarray
+        A mask indicting the interested region, default: None
 
     Returns
     -------
@@ -700,7 +727,7 @@ def _offset_blending(off_product_dir, rubbersheet_params, layer_keys):
 
     # Filter outliers from layer one
     offset_az, offset_rg = identify_outliers(str(off_product_dir / layer_keys[0]),
-                                             rubbersheet_params)
+                                             rubbersheet_params, mask)
 
     # Replace the NaN locations in layer1 with the mean of pixels in layers
     # at lower resolution computed in a neighborhood centered at the NaN location
@@ -711,12 +738,12 @@ def _offset_blending(off_product_dir, rubbersheet_params, layer_keys):
 
         if nan_count_az > 0:
             offset_az_culled, _ = identify_outliers(str(off_product_dir / layer_key),
-                                                    rubbersheet_params)
+                                                    rubbersheet_params, mask)
             offset_az = _fill_nan_with_mean(offset_az, offset_az_culled, filter_size)
 
         if nan_count_rg > 0:
             _, offset_rg_culled = identify_outliers(str(off_product_dir / layer_key),
-                                                    rubbersheet_params)
+                                                    rubbersheet_params, mask)
             offset_rg = _fill_nan_with_mean(offset_rg, offset_rg_culled, filter_size)
 
     # Fill remaining holes by iteratively filling the output offset layer
@@ -727,11 +754,84 @@ def _offset_blending(off_product_dir, rubbersheet_params, layer_keys):
 
     return offset_az, offset_rg
 
+def _interpolate_offsets_by_idw(
+    Z: np.ndarray,
+    power: float = 2.0,
+    number: int | None = 100,
+    radius: float | None = 200.0):
+    '''
+    Performs Inverse Distance Weighting (IDW) interpolation on a 2D grid containing NaN values.
+    Missing pixels are filled using IDW interpolation based on nearby
+    valid samples. The interpolation can optionally
+    limit the number of nearest neighbors and the maximum search
+    radius to improve efficiency and control smoothing.
 
-import journal
-import numpy as np
-from scipy import interpolate
+    Parameters
+    ---------
+    Z: np.ndarray
+        2D array of shape (H, W) containing NaN values to be filled
+    power: float
+        Power parameter p controlling the distance weighting
+    number: int or None
+        Number of nearest neighbors used for interpolation.
+        If None, all valid points are used (default: 100)
+    radius: float or None
+        Maximum search radius in pixels. If None, no radius
+        limitation is applied (default: 200)
+    eps: float
+        Small value added to distance to avoid division by zero
+
+    Returns
+    -------
+    filled: np.ndarray
+        2D array with the same shape as Z where NaN values
+        have been filled using IDW interpolation
+    '''
+
+    valid = np.isfinite(Z)
+    yy, xx = np.nonzero(valid)
+    if yy.size == 0:
+        raise ValueError("No valid points found to interpolate from.")
+
+    points = np.column_stack([xx, yy]).astype(float)
+    values = Z[yy, xx].astype(float)
+
+    qmask = ~valid
+    yq, xq = np.nonzero(qmask)
+    if yq.size == 0:
+        return Z.copy()
+
+    qpoints = np.column_stack([xq, yq]).astype(float)
+
+    tree = cKDTree(points)
+    k = points.shape[0] if number is None else min(number, points.shape[0])
+
+    dists, idxs = tree.query(
+        qpoints, k=k, workers=-1,
+        **({'distance_upper_bound': radius} if radius is not None else {}))
+
+    # Ensure 2D — tree.query returns 1D when k=1
+    dists = np.reshape(dists, (qpoints.shape[0], -1))
+    idxs  = np.reshape(idxs,  (qpoints.shape[0], -1))
+
+    use_mask = np.isfinite(dists) & (idxs < points.shape[0])
+
+    # Clip out-of-bounds sentinel indices before indexing into values
+    safe_idxs = np.clip(idxs, 0, points.shape[0] - 1)
+
+    w = np.where(use_mask, 1.0 / dists**power, 0.0)
+    v = np.where(use_mask, values[safe_idxs], 0.0)
+
+    ws = w.sum(axis=1)
+    ok = ws > 0
+
+    filled_vals = np.full(qpoints.shape[0], np.nan, dtype=float)
+    filled_vals[ok] = (w[ok] * v[ok]).sum(axis=1) / ws[ok]
+
+    out = Z.copy()
+    out[yq, xq] = filled_vals
 
+    return out
 
 def _interpolate_offsets(offset, interp_method):
     '''
@@ -851,5 +951,5 @@ def run(cfg: dict, output_hdf5: str = None):
 
     # Prepare RIFG. Culled offsets will be
     # allocated in RIFG product
-    out_paths = prepare_insar_hdf5.run(rubbersheet_runconfig.cfg)
+    _, out_paths = h5_prep.get_products_and_paths(rubbersheet_runconfig.cfg)
     run(rubbersheet_runconfig.cfg, out_paths['RIFG'])

share/nisar/defaults/insar.yaml

@@ -306,7 +306,7 @@ runconfig:
                     # When `min_cluster_pixels` is 0, this pre-removal step is
                     # skipped entirely.
                     min_cluster_pixels: 2
-                    
+
                     bridge_algorithm_enabled: True
                     # Maximum radius used when bridging disconnected
                     # regions.
@@ -545,6 +545,19 @@ runconfig:
                 # Interpolation method. Interpolation is used to fill residual outlier
                 # holes if present
                 interpolation_method: linear
+
+                # Flag to enable the use of subswath mask for the outlier detection
+                subswath_mask_apply_enabled: True
+
+                # IDW interpolation method parameters
+                idw_interpolation:
+                    # Power parameter controlling the distance weighting
+                    power: 2
+                    # Number of nearest neighbors used for interpolation
+                    number: 100
+                    # Maximum search radius in pixels
+                    radius: 200
+
                 # It is good practice to filter the offsets prior to interferogram
                 # formation to reduce noise. We expose: no_filter: do not filter the offsets/
                 # degrade offsets resolution. median_filter, boxcar_filter (moving average),
@@ -780,7 +793,7 @@ runconfig:
                     # During deramping, uniform sampling is applied
                     # when number of samples is larger than maximum pixel
                     bridge_ramp_maximum_pixel: 1e6
-    
+
             correction_luts:
                 # Boolean flag to activate/deactivate model-based solid earth tide corrections
                 solid_earth_tides_enabled: True

share/nisar/schemas/insar.yaml

@@ -548,7 +548,13 @@ rubbersheet_options:
 
     # Interpolation method. Interpolation is used to fill residual
     # outlier holes (if present)
-    interpolation_method: enum('no_interpolation', 'nearest', 'linear', 'cubic', required=False)
+    interpolation_method: enum('no_interpolation', 'nearest', 'linear', 'cubic', 'idw', required=False)
+
+    # Inverse distance weighting (IDW) interpolation method
+    idw_interpolation: include('idw_interpolation_options', required=False)
+
+    # Flag to enable the use of the subswath mask for the outlier detection
+    subswath_mask_apply_enabled: bool(required=False)
 
     # It is good practice to filter the offsets prior to interferogram
     # formation to reduce noise. We expose: no_filter: do not filter the offsets/
@@ -595,6 +601,17 @@ polyfitting_options:
     # Number of samples  along the azimuth in pixels for polyfitting
     samples_along_azimuth: num(min=0, required=False)
 
+idw_interpolation_options:
+    # Power parameter controlling the distance weighting;
+    # If the power is 0, then it is equally weighted
+    power: int(min=0, required=False)
+    # Number of nearest neighbors used for interpolation
+    number: int(min=1, required=False)
+    # Maximum search radius in pixels;
+    # If unspecified, all valid pixels are used,
+    # which can significantly increase runtime and is not recommended
+    radius: num(min=1, required=False)
+
 crossmul_options:
     # Path to HDF5 file or directory with coregistered SLCs
     # If directory then the following directory tree is required: