import os
import numpy as np
from polsartools.utils.proc_utils import process_chunks_parallel
from polsartools.utils.utils import conv2d,time_it
from polsartools.utils.convert_matrices import T3_C3_mat
from .fp_infiles import fp_c3t3files
[docs]
@time_it
def nnedfp(infolder, window_size=1, outType="tif", cog_flag=False,
cog_overviews = [2, 4, 8, 16], write_flag=True,
max_workers=None,block_size=(512, 512),
progress_callback=None, # for QGIS plugin
):
"""Perform Non-Negative Eigenvalue Decomposition (NNED) for full-pol SAR data.
This function implements the Non-Negative Eigenvalue Decomposition for
full-polarimetric SAR data, ensuring physically meaningful decomposition results
by constraining eigenvalues to be non-negative. The method decomposes the total
power into odd-bounce (surface), double-bounce, and volume scattering components.
Examples
--------
>>> # Basic usage with default parameters
>>> nnedfp("/path/to/fullpol_data")
>>> # Advanced usage with custom parameters
>>> nnedfp(
... infolder="/path/to/fullpol_data",
... window_size=5,
... outType="tif",
... cog_flag=True,
... block_size=(1024, 1024)
... )
Parameters
----------
infolder : str
Path to the input folder containing full-pol T3 or C3 matrix files.
window_size : int, default=1
Size of the spatial averaging window. Larger windows reduce speckle noise
but decrease spatial resolution.
outType : {'tif', 'bin'}, default='tif'
Output file format:
- 'tif': GeoTIFF format with georeferencing information
- 'bin': Raw binary format
cog_flag : bool, default=False
If True, creates Cloud Optimized GeoTIFF (COG) outputs with internal tiling
and overviews for efficient web access.
cog_overviews : list[int], default=[2, 4, 8, 16]
Overview levels for COG creation. Each number represents the
decimation factor for that overview level.
write_flag : bool, default=True
If True, writes results to disk. If False, only processes data in memory.
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
Writes three output files to disk:
1. NNED_odd: Odd-bounce (surface) scattering component
2. NNED_dbl: Double-bounce scattering component
3. NNED_vol: Volume scattering component
Notes
-----
The NNED method ensures physical realizability by:
1. Maintaining non-negative eigenvalues
2. Preserving total power conservation
3. Ensuring positive semi-definite coherency matrices
"""
input_filepaths = fp_c3t3files(infolder)
output_filepaths = []
if outType == "bin":
output_filepaths.append(os.path.join(infolder, "NNED_odd.bin"))
output_filepaths.append(os.path.join(infolder, "NNED_dbl.bin"))
output_filepaths.append(os.path.join(infolder, "NNED_vol.bin"))
else:
output_filepaths.append(os.path.join(infolder, "NNED_odd.tif"))
output_filepaths.append(os.path.join(infolder, "NNED_dbl.tif"))
output_filepaths.append(os.path.join(infolder, "NNED_vol.tif"))
process_chunks_parallel(input_filepaths, list(output_filepaths),
window_size=window_size, write_flag=write_flag,
processing_func=process_chunk_nnedfp,block_size=block_size,
max_workers=max_workers, num_outputs=len(output_filepaths),
cog_flag=cog_flag,
cog_overviews=cog_overviews,
progress_callback=progress_callback
)
def process_chunk_nnedfp(chunks, window_size, input_filepaths, *args):
# additional_arg1 = args[0] if len(args) > 0 else None
# additional_arg2 = args[1] if len(args) > 1 else None
if 'T11' in input_filepaths[0] and 'T22' in input_filepaths[5] and 'T33' in input_filepaths[8]:
t11_T1 = np.array(chunks[0])
t12_T1 = np.array(chunks[1])+1j*np.array(chunks[2])
t13_T1 = np.array(chunks[3])+1j*np.array(chunks[4])
t21_T1 = np.conj(t12_T1)
t22_T1 = np.array(chunks[5])
t23_T1 = np.array(chunks[6])+1j*np.array(chunks[7])
t31_T1 = np.conj(t13_T1)
t32_T1 = np.conj(t23_T1)
t33_T1 = np.array(chunks[8])
T3 = np.array([[t11_T1, t12_T1, t13_T1],
[t21_T1, t22_T1, t23_T1],
[t31_T1, t32_T1, t33_T1]])
T_T1 = T3_C3_mat(T3)
if 'C11' in input_filepaths[0] and 'C22' in input_filepaths[5] and 'C33' in input_filepaths[8]:
C11 = np.array(chunks[0])
C12 = np.array(chunks[1])+1j*np.array(chunks[2])
C13 = np.array(chunks[3])+1j*np.array(chunks[4])
C21 = np.conj(C12)
C22 = np.array(chunks[5])
C23 = np.array(chunks[6])+1j*np.array(chunks[7])
C31 = np.conj(C13)
C32 = np.conj(C23)
C33 = np.array(chunks[8])
T_T1 = np.array([[C11, C12, C13],
[C21, C22, C23],
[C31, C32, C33]])
if window_size>1:
kernel = np.ones((window_size,window_size),np.float32)/(window_size*window_size)
t11f = conv2d(T_T1[0,0,:,:],kernel)
t12f = conv2d(np.real(T_T1[0,1,:,:]),kernel)+1j*conv2d(np.imag(T_T1[0,1,:,:]),kernel)
t13f = conv2d(np.real(T_T1[0,2,:,:]),kernel)+1j*conv2d(np.imag(T_T1[0,2,:,:]),kernel)
t21f = np.conj(t12f)
t22f = conv2d(T_T1[1,1,:,:],kernel)
t23f = conv2d(np.real(T_T1[1,2,:,:]),kernel)+1j*conv2d(np.imag(T_T1[1,2,:,:]),kernel)
t31f = np.conj(t13f)
t32f = np.conj(t23f)
t33f = conv2d(T_T1[2,2,:,:],kernel)
T_T1 = np.array([[t11f, t12f, t13f], [t21f, t22f, t23f], [t31f, t32f, t33f]])
_,_,rows,cols = np.shape(T_T1)
T_T1 = T_T1.reshape(9, rows, cols)
# Indices for vectorized access
i, j = np.indices((rows, cols))
SpanMax = -float('inf')
SpanMin = float('inf')
Span_data = np.real(T_T1[0,:,:]) + np.real(T_T1[4,:,:]) + np.real(T_T1[8,:,:])
# Find the maximum and minimum values in one step
SpanMax = np.nanmax(Span_data)
SpanMin = np.nanmin(Span_data)
# Ensure SpanMin does not go below eps
SpanMin = np.nanmax([SpanMin, 1e-6])
# Set all values in Span_data below eps to eps
Span_data[Span_data < 1e-6] = 1e-6
epsilon = np.real(T_T1[0,:,:])
rho_re = np.real(T_T1[2,:,:])
rho_im = np.imag(T_T1[2,:,:])
nhu = np.real(T_T1[4,:,:])
gamma = np.real(T_T1[8,:,:])
# Pre-calculate epsilon_veg, rho_re_veg, rho_im_veg, nhu_veg, gamma_veg
fv = 3. * nhu / 2.
epsilon_veg = fv
rho_re_veg = fv / 3.
rho_im_veg = 0.
nhu_veg = 2. * fv / 3.
gamma_veg = fv
# Calculate z, a, b
z = epsilon * gamma_veg + epsilon_veg * gamma - 2. * rho_re * rho_re_veg
a = epsilon_veg * gamma_veg - rho_re_veg * rho_re_veg
b = epsilon * gamma - rho_re * rho_re - rho_im * rho_im
# Calculate x1
x1 = nhu / nhu_veg
# Handle the case where a == 0 using a mask
mask_a_zero = a == 0.
x2 = np.where(mask_a_zero, b / z, (z - np.sqrt(z * z - 4. * a * b)) / (2. * a))
xmax = np.minimum(x1, x2)
# make sure the array is writable
epsilon.setflags(write=True)
rho_re.setflags(write=True)
rho_im.setflags(write=True)
nhu.setflags(write=True)
gamma.setflags(write=True)
# Update epsilon, rho_re, rho_im, nhu, gamma
epsilon -= xmax * epsilon_veg
rho_re -= xmax * rho_re_veg
rho_im -= xmax * rho_im_veg
nhu -= xmax * nhu_veg
gamma -= xmax * gamma_veg
# Calculate delta, lambda1, lambda2
delta = (epsilon - gamma) ** 2 + 4. * (rho_re * rho_re + rho_im * rho_im)
lambda1 = 0.5 * (epsilon + gamma + np.sqrt(delta))
lambda2 = 0.5 * (epsilon + gamma - np.sqrt(delta))
# Calculate OMEGA1, OMEGA2
OMEGA1 = lambda1 * (gamma - epsilon + np.sqrt(delta)) ** 2
OMEGA1 /= ( (gamma - epsilon + np.sqrt(delta)) ** 2 + 4. * (rho_re * rho_re + rho_im * rho_im))
OMEGA2 = lambda2 * (gamma - epsilon - np.sqrt(delta)) ** 2
OMEGA2 /= ((gamma - epsilon - np.sqrt(delta)) ** 2 + 4. * (rho_re * rho_re + rho_im * rho_im))
# Calculate hh1_re, hh1_im, hh2_re, hh2_im
hh1_re = 2. * rho_re / (gamma - epsilon + np.sqrt(delta))
hh1_im = 2. * rho_im / (gamma - epsilon + np.sqrt(delta))
hh2_re = 2. * rho_re / (gamma - epsilon - np.sqrt(delta))
hh2_im = 2. * rho_im / (gamma - epsilon - np.sqrt(delta))
# Define A0A0 and B0pB
A0A0 = (hh1_re + 1.) ** 2 + hh1_im ** 2
B0pB = (hh1_re - 1.) ** 2 + hh1_im ** 2
# Apply conditions for ALPre, ALPim, OMEGAodd, BETre, BETim, OMEGAdbl
mask_A0A0_greater = A0A0 > B0pB
ALPre = np.where(mask_A0A0_greater, hh1_re, hh2_re)
ALPim = np.where(mask_A0A0_greater, hh1_im, hh2_im)
OMEGAodd = np.where(mask_A0A0_greater, OMEGA1, OMEGA2)
BETre = np.where(mask_A0A0_greater, hh2_re, hh1_re)
BETim = np.where(mask_A0A0_greater, hh2_im, hh1_im)
OMEGAdbl = np.where(mask_A0A0_greater, OMEGA2, OMEGA1)
# Calculate NNED_odd, NNED_dbl, and NNED_vol
NNED_odd = OMEGAodd * (1 + ALPre**2 + ALPim**2)
NNED_dbl = OMEGAdbl * (1 + BETre**2 + BETim**2)
NNED_vol = xmax * (epsilon_veg + nhu_veg + gamma_veg)
# Apply limits to M_odd, M_dbl, M_vol
NNED_odd = np.clip(NNED_odd, 0., SpanMax)
NNED_dbl = np.clip(NNED_dbl, 0., SpanMax)
NNED_vol = np.clip(NNED_vol, 0., SpanMax)
return NNED_odd.astype(np.float32),NNED_dbl.astype(np.float32),NNED_vol.astype(np.float32)
