import os
import numpy as np
from polsartools.utils.proc_utils import process_chunks_parallel
from polsartools.utils.utils import conv2d,time_it,eig22
from .dxp_infiles import dxpc2files, S_norm
[docs]
@time_it
def powers_dp(in_dir, method=1, win=1, fmt="tif", cog=False,
ovr = [2, 4, 8, 16], comp=False,
max_workers=None,block_size=(512, 512),
progress_callback=None, # for QGIS plugin
):
""" This function computes the scattering power components for dual-pol SAR C2 matrix data (decomposition/factorization based approach)
Examples
--------
>>> # Basic usage with default parameters
>>> powers_dp("/path/to/c2_data")
>>> # Advanced usage with custom parameters
>>> powers_dp(
... in_dir="/path/to/c2_data",
... method=2,
... win=3,
... fmt="tif",
... cog=True,
... block_size=(1024, 1024)
... )
Parameters
----------
in_dir : str
Path to the input folder containing C2 matrix files.
method : int
1: Decomposition based powers
2: Factorisation based powers
win : int, default=1
Size of the spatial averaging window. Larger windows reduce speckle noise
but decrease spatial resolution.
fmt : {'tif', 'bin'}, default='tif'
Output file format:
- 'tif': GeoTIFF format with georeferencing information
- 'bin': Raw binary format
cog : bool, default=False
If True, creates a Cloud Optimized GeoTIFF (COG) with internal tiling
and overviews for efficient web access.
ovr : list[int], default=[2, 4, 8, 16]
Overview levels for COG creation. Each number represents the
decimation factor for that overview level.
comp : bool, default=False
If True, applies LZW compression to the output GeoTIFF files.
max_workers : int | None, default=None
Maximum number of parallel processing workers. If None, uses
CPU count - 1 workers.
block_size : tuple[int, int], default=(512, 512)
Size of processing blocks (rows, cols) for parallel computation.
Larger blocks use more memory but may be more efficient.
Returns
-------
None
"""
write_flag=True
input_filepaths = dxpc2files(in_dir)
output_filepaths = []
if fmt == "bin":
if method==1:
output_filepaths.append(os.path.join(in_dir, "Alpha_dp.bin"))
output_filepaths.append(os.path.join(in_dir, "Pdl_dcmp.bin"))
output_filepaths.append(os.path.join(in_dir, "Psl_dcmp.bin"))
output_filepaths.append(os.path.join(in_dir, "Pu_dcmp.bin"))
else:
output_filepaths.append(os.path.join(in_dir, "Pdl_fact.bin"))
output_filepaths.append(os.path.join(in_dir, "Psl_fact.bin"))
output_filepaths.append(os.path.join(in_dir, "Pr_fact.bin"))
else:
if method==1:
output_filepaths.append(os.path.join(in_dir, "Alpha_dp.tif"))
output_filepaths.append(os.path.join(in_dir, "Pdl_dcmp.tif"))
output_filepaths.append(os.path.join(in_dir, "Psl_dcmp.tif"))
output_filepaths.append(os.path.join(in_dir, "Pu_dcmp.tif"))
else:
output_filepaths.append(os.path.join(in_dir, "Pdl_fact.tif"))
output_filepaths.append(os.path.join(in_dir, "Psl_fact.tif"))
output_filepaths.append(os.path.join(in_dir, "Pr_fact.tif"))
process_chunks_parallel(input_filepaths, list(output_filepaths), win, write_flag,
process_chunk_dp_pow,
*[method],
block_size=block_size, max_workers=max_workers, num_outputs=len(output_filepaths),
cog=cog,ovr=ovr, comp=comp,
progress_callback=progress_callback
)
def process_chunk_dp_pow(chunks, window_size,*args):
method = int(args[-1])
kernel = np.ones((window_size,window_size),np.float32)/(window_size*window_size)
c11_T1 = np.array(chunks[0])
c12_T1 = np.array(chunks[1])+1j*np.array(chunks[2])
# c21_T1 = np.conj(c12_T1)
c22_T1 = np.array(chunks[3])
##### Normalizing Stokes vector elements
# def S_norm(S_array):
# S_5 = np.nanpercentile(S_array, 2)
# S_95 = np.nanpercentile(S_array, 98)
# S_cln = np.where(S_array > S_95, S_95, S_array)
# S_cln = np.where(S_cln < S_5, S_5, S_cln)
# S_cln_max = np.nanmax(S_cln)
# S_norm_array = np.divide(S_cln,S_cln_max)
# return S_norm_array
if window_size>1:
c11s = conv2d(np.real(c11_T1),kernel)
c12s = conv2d(np.real(c12_T1),kernel)+1j*conv2d(np.imag(c12_T1),kernel)
# c21s = conv2d(np.real(c21_T1),kernel)+1j*conv2d(np.imag(c21_T1),kernel)
c22s = conv2d(np.real(c22_T1),kernel)
else:
c11s = c11_T1
c12s = c12_T1
c22s = c22_T1
C11_av_db = 10*np.log10(c11s)
s0 = c11s + c22s
s1 = c11s - c22s
s2 = 2*c12s.real
s3 = 2*c12s.imag
##### Calculate Entropy
## Here eigen values are calculated using Stokes vector elements
tpp = np.sqrt(np.square(s1) + np.square(s2) + np.square(s3))
lmbd1 = (s0 + tpp)/2
lmbd2 = (s0 - tpp)/2
prob1 = lmbd1/(lmbd1 + lmbd2)
prob2 = lmbd2/(lmbd1 + lmbd2)
ent = -prob1*np.log2(prob1) - prob2*np.log2(prob2)
dop = (lmbd1 - lmbd2)/(lmbd1 + lmbd2)
beta = lmbd1/(lmbd1 + lmbd2)
##### Taking abs of Stokes vector elements
s0 = np.abs(s0)
s1 = np.abs(s1)
s2 = np.abs(s2)
s3 = np.abs(s3)
s1_s_norm = S_norm(s1) #This is S1 normalzied for DpRSI, does not include slope mask
s1_norm = S_norm(s1)
s2_norm = S_norm(s2)
s3_norm = S_norm(s3)
if method==1:
##### Power Calculation
dprbi = np.sqrt(np.square(s1_norm) + np.square(s2_norm) + np.square(s3_norm))/np.sqrt(3)
dprsi_con1 = (1 - ent)*np.sqrt(1 - np.square(s1_s_norm)) # For Valid pixels
dprsi_con2 = np.sqrt(1 - np.square(s1_s_norm)) # For Noise pixels
NESZ = -16 ## For Sentinel-1
dprsi = np.where(C11_av_db > NESZ, dprsi_con1, dprsi_con2)
shp = np.shape(dprbi)
alpha1 = np.arctan2(dprbi, 1 - dprbi)
alpha1 = np.degrees(alpha1)
alpha2 = np.arctan2(1-dprsi, dprsi)
alpha2 = np.degrees(alpha2)
alpha_dp = (alpha1 + alpha2)/2; #Dual-pol target characteristic parameter proposed in Verma et al. 2024
## Alpha as geomteric factor
alpha_dp_rad = np.radians(2*alpha_dp)
cos_a = np.cos(alpha_dp_rad)
## Power components for valid pixels (VV > NESZ)
Pu_v = (1 - dop)*s0
Pd_v = (1/2)*dop*s0*(1 - cos_a)
Ps_v = (1/2)*dop*s0*(1 + cos_a)
## Power components for noise pixels (VV < NESZ)
Pu_n = (1 - beta)*s0
Pd_n = (1/2)*beta*s0*(1 - cos_a)
Ps_n = (1/2)*beta*s0*(1 + cos_a)
## Dual-pol scattering power
Pu = np.where(C11_av_db > NESZ, Pu_v, Pu_n) # Unpolized power
Pd = np.where(C11_av_db > NESZ, Pd_v, Pd_n) # "Dihedral-like" power
Ps = np.where(C11_av_db > NESZ, Ps_v, Ps_n) # "Surface-like" power
return alpha_dp.astype(np.float32), Pd.astype(np.float32), Ps.astype(np.float32), Pu.astype(np.float32)
else:
##### Power Calculation
dprbi = np.sqrt(np.square(s1_norm) + np.square(s2_norm) + np.square(s3_norm))/np.sqrt(3)
dprsi_con1 = (1 - ent)*np.sqrt(1 - np.square(s1_s_norm)) # For Valid pixels
dprsi_con2 = np.sqrt(1 - np.square(s1_s_norm)) # For Noise pixels
NESZ = -16 ## For Sentinel-1
dprsi = np.where(C11_av_db > NESZ, dprsi_con1, dprsi_con2)
shp = np.shape(dprbi)
dprbi_flt = dprbi.flatten()
dprsi_flt = dprsi.flatten()
shp_flt = np.shape(dprbi_flt)
indices_vec = np.array([dprsi_flt, dprbi_flt]).transpose()
indices_vec_sort = np.array([[max(row), min(row)] for row in indices_vec])
y1 = indices_vec_sort[:,0] #First dominant
y2 = (1 - indices_vec_sort[:,0])*indices_vec_sort[:,1] #Second dominant
residue = 1 - (y1 + y2)
dprsi_dom = np.where(dprsi_flt > dprbi_flt)[0] #Keeps the tuple where dprsi is dominant
dprbi_dom = np.where(dprsi_flt < dprbi_flt)[0] #Keeps the tuple where dprbi is dominant
## Surface-like power component
Ps = np.zeros(shp_flt)
##dprsi_dom and dprbi_dom are not dprsi and dprbi values, they just indicate tuples (pixel) for which they are greater
Ps[dprsi_dom] = y1[dprsi_dom] #In these tuples dprsi was dominant, hence taking y1 (first dominant)
Ps[dprbi_dom] = y2[dprbi_dom] #In these tuples dprbi was dominant, hence taking y2 (Second domiant)
Ps = Ps.reshape(shp[0],shp[1])
Ps = np.multiply(s0,Ps)
## Dihedral-like power component
Pd = np.zeros(shp_flt)
Pd[dprbi_dom] = y1[dprbi_dom]
Pd[dprsi_dom] = y2[dprsi_dom]
Pd = Pd.reshape(shp[0],shp[1])
Pd = np.multiply(s0,Pd)
## Residue (diffused) power component
Pr = residue.reshape(shp[0],shp[1])
Pr = np.multiply(s0,Pr)
return Pd.astype(np.float32), Ps.astype(np.float32), Pr.astype(np.float32)
