# -*- coding: utf-8 -*-
# AROSICS - Automated and Robust Open-Source Image Co-Registration Software
#
# Copyright (C) 2017-2024
# - Daniel Scheffler (GFZ Potsdam, daniel.scheffler@gfz-potsdam.de)
# - Helmholtz Centre Potsdam - GFZ German Research Centre for Geosciences Potsdam,
# Germany (https://www.gfz-potsdam.de/)
#
# This software was developed within the context of the GeoMultiSens project funded
# by the German Federal Ministry of Education and Research
# (project grant code: 01 IS 14 010 A-C).
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import time
import warnings
from copy import copy
from typing import Iterable, Union, Tuple, List, Optional # noqa F401
from multiprocessing import cpu_count
# custom
from osgeo import gdal # noqa
import numpy as np
from scipy.fft import fft2, ifft2, fftshift
from packaging.version import parse as parse_version
try:
import pyfftw
# pyfftw>=0.13.0 is currently not used due to https://github.com/pyFFTW/pyFFTW/issues/294
if parse_version(pyfftw.__version__) >= parse_version('0.13.0'):
pyfftw = None
except ImportError:
pyfftw = None
from shapely.geometry import Point, Polygon
# internal modules
from .DeShifter import DESHIFTER, _dict_rspAlg_rsp_Int
from . import geometry as GEO
from . import plotting as PLT
from geoarray import GeoArray
from py_tools_ds.convenience.object_oriented import alias_property
from py_tools_ds.geo.coord_calc import get_corner_coordinates
from py_tools_ds.geo.vector.topology import get_overlap_polygon, get_smallest_boxImYX_that_contains_boxMapYX
from py_tools_ds.geo.projection import prj_equal
from py_tools_ds.geo.vector.geometry import boxObj, round_shapelyPoly_coords
from py_tools_ds.geo.coord_grid import move_shapelyPoly_to_image_grid, is_coord_grid_equal
from py_tools_ds.geo.coord_trafo import reproject_shapelyGeometry, mapXY2imXY, imXY2mapXY
from py_tools_ds.geo.raster.reproject import warp_ndarray
from py_tools_ds.geo.map_info import geotransform2mapinfo
from py_tools_ds.io.vector.writer import write_shp
__author__ = 'Daniel Scheffler'
[docs]class GeoArray_CoReg(GeoArray):
def __init__(self,
CoReg_params: dict,
imID: str
) -> None:
assert imID in ['ref', 'shift']
# run GeoArray init
path_or_geoArr = CoReg_params['im_ref'] if imID == 'ref' else CoReg_params['im_tgt']
nodata = CoReg_params['nodata'][0 if imID == 'ref' else 1]
progress = CoReg_params['progress']
q = CoReg_params['q'] if not CoReg_params['v'] else False
super(GeoArray_CoReg, self).__init__(path_or_geoArr, nodata=nodata, progress=progress, q=q)
self.imID = imID
self.imName = 'reference image' if imID == 'ref' else 'image to be shifted'
self.v = CoReg_params['v']
assert isinstance(self, GeoArray), \
'Something went wrong with the creation of GeoArray instance for the %s. The created ' \
'instance does not seem to belong to the GeoArray class. If you are working in Jupyter Notebook, reset ' \
'the kernel and try again.' % self.imName
# set title to be used in plots
self.title = os.path.basename(self.filePath) if self.filePath else self.imName
# validate params
# assert self.prj, 'The %s has no projection.' % self.imName # TODO
# assert not re.search('LOCAL_CS', self.prj), 'The %s is not georeferenced.' % self.imName # TODO
assert self.gt, 'The %s has no map information.' % self.imName
# set band4match
self.band4match = (CoReg_params['r_b4match'] if imID == 'ref' else CoReg_params['s_b4match']) - 1
assert self.bands >= self.band4match + 1 >= 1, \
"The %s has %s %s. So its band number to match must be %s%s. Got %s." \
% (self.imName, self.bands, 'bands' if self.bands > 1 else
'band', 'between 1 and ' if self.bands > 1 else '', self.bands, self.band4match)
# compute nodata mask and validate that it is not completely filled with nodata
self.calc_mask_nodata(fromBand=self.band4match) # this avoids that all bands have to be read
if True not in self.mask_nodata[:]:
raise RuntimeError(f'The {self.imName} passed to AROSICS only contains nodata values.')
# set footprint_poly
given_footprint_poly = CoReg_params['footprint_poly_%s' % ('ref' if imID == 'ref' else 'tgt')]
given_corner_coord = CoReg_params['data_corners_%s' % ('ref' if imID == 'ref' else 'tgt')]
if given_footprint_poly:
self.footprint_poly = given_footprint_poly
elif given_corner_coord is not None:
self.footprint_poly = Polygon(given_corner_coord)
elif not CoReg_params['calc_corners']:
# use the image extent
self.footprint_poly = Polygon(get_corner_coordinates(gt=self.gt, cols=self.cols, rows=self.rows))
else:
# footprint_poly is calculated automatically by GeoArray
if not CoReg_params['q']:
print('Calculating footprint polygon and actual data corner coordinates for %s...' % self.imName)
with warnings.catch_warnings(record=True) as w:
_ = self.footprint_poly # execute getter
if len(w) > 0 and 'disjunct polygon(s) outside' in str(w[-1].message):
warnings.warn('The footprint of the %s contains multiple separate image parts. '
'AROSICS will only process the largest image part.' % self.imName)
# FIXME use a convex hull as footprint poly
# validate footprint poly
if not self.footprint_poly.is_valid:
self.footprint_poly = self.footprint_poly.buffer(0)
if not self.q:
print('Bounding box of calculated footprint for %s:\n\t%s' % (self.imName, self.footprint_poly.bounds))
# add bad data mask
given_mask = CoReg_params['mask_baddata_%s' % ('ref' if imID == 'ref' else 'tgt')]
if given_mask is not None:
self.mask_baddata = given_mask # runs GeoArray.mask_baddata.setter -> sets it to BadDataMask()
poly = alias_property('footprint_poly') # ensures that self.poly is updated if self.footprint_poly is updated
[docs]class COREG(object):
"""The COREG class detects and corrects global X/Y shifts between a target and reference image.
Geometric shifts are calculated at a specific (adjustable) image position. Correction performs a global shifting
in X- or Y direction.
See help(COREG) for documentation!
"""
def __init__(self,
im_ref: Union[GeoArray, str],
im_tgt: Union[GeoArray, str],
path_out: str = None,
fmt_out: str = 'ENVI',
out_crea_options: list = None,
r_b4match: int = 1,
s_b4match: int = 1,
wp: Tuple[float, float] = (None, None),
ws: Tuple[int, int] = (256, 256),
max_iter: int = 5,
max_shift: int = 5,
align_grids: bool = False,
match_gsd: bool = False,
out_gsd: Tuple[float] = None,
target_xyGrid: List[List] = None,
resamp_alg_deshift: str = 'cubic',
resamp_alg_calc: str = 'cubic',
footprint_poly_ref: str = None,
footprint_poly_tgt: str = None,
data_corners_ref: list = None,
data_corners_tgt: list = None,
nodata: Tuple = (None, None),
calc_corners: bool = True,
binary_ws: bool = True,
mask_baddata_ref: Union[GeoArray, str] = None,
mask_baddata_tgt: Union[GeoArray, str] = None,
CPUs: int = None,
force_quadratic_win: bool = True,
progress: bool = True,
v: bool = False,
path_verbose_out: str = None,
q: bool = False,
ignore_errors: bool = False
) -> None:
"""Get an instance of the COREG class.
:param im_ref:
source path or GeoArray instance of reference image
(any GDAL compatible image format is supported)
:param im_tgt:
source path or GeoArray instance of the target image, i.e., the image to be shifted
(any GDAL compatible image format is supported)
:param path_out:
target path of the coregistered image
- if None (default), the method correct_shifts() does not write to disk
- if 'auto': /dir/of/im1/<im1>__shifted_to__<im0>.bsq
:param fmt_out:
raster file format for output file. ignored if path_out is None. can be any GDAL compatible raster file
format (e.g. 'ENVI', 'GTIFF'; default: ENVI). Refer to https://gdal.org/drivers/raster/index.html to get a
full list of supported formats.
:param out_crea_options:
GDAL creation options for the output image, e.g. ["QUALITY=80", "REVERSIBLE=YES", "WRITE_METADATA=YES"]
:param r_b4match:
band of reference image to be used for matching (starts with 1; default: 1)
:param s_b4match:
band of shift image to be used for matching (starts with 1; default: 1)
:param wp:
custom matching window position as (X, Y) map coordinate in the same projection as the reference image
(default: central position of image overlap)
:param ws:
custom matching window size [pixels] as (X, Y) tuple (default: (256,256))
:param max_iter:
maximum number of iterations for matching (default: 5)
:param max_shift:
maximum shift distance in reference image pixel units (default: 5 px)
:param align_grids:
True: align the input coordinate grid to the reference (does not affect the output pixel size as long as
input and output pixel sizes are compatible (5:30 or 10:30 but not 4:30)), default = False
- NOTE: If this is set to False, the mis-registration in the target image is corrected by updating its
geocoding information only, i.e., without performing any resampling which preserves the original
image values (except that a resampling is needed due to other reasons, e.g., if 'match_gsd',
'out_gsd' or 'target_xyGrid' are given).
:param match_gsd:
True: match the input pixel size to the reference pixel size,
default = False
:param out_gsd:
output pixel size (X/Y) in units of the reference coordinate system (default = pixel size of the target
image), given values are overridden by match_gsd=True
:param target_xyGrid:
a list with a target x-grid and a target y-grid like [[15,45], [15,45]]
This overrides 'out_gsd', 'align_grids' and 'match_gsd'.
:param resamp_alg_deshift:
the resampling algorithm to be used for shift correction (if neccessary)
- valid algorithms: nearest, bilinear, cubic, cubic_spline, lanczos, average, mode, max, min, med, q1, q3
- default: cubic
:param resamp_alg_calc:
the resampling algorithm to be used for all warping processes during calculatio of spatial shift
- valid algorithms: nearest, bilinear, cubic, cubic_spline, lanczos, average, mode, max, min, med, q1, q3
- default: cubic (highly recommended)
:param footprint_poly_ref:
footprint polygon of the reference image (WKT string or shapely.geometry.Polygon),
e.g. 'POLYGON ((299999 6000000, 299999 5890200, 409799 5890200, 409799 6000000, 299999 6000000))'
:param footprint_poly_tgt:
footprint polygon of the image to be shifted (WKT string or shapely.geometry.Polygon)
e.g. 'POLYGON ((299999 6000000, 299999 5890200, 409799 5890200, 409799 6000000, 299999 6000000))'
:param data_corners_ref:
map coordinates of data corners within reference image. ignored if footprint_poly_ref is given.
:param data_corners_tgt:
map coordinates of data corners within image to be shifted. ignored if footprint_poly_tgt is given.
:param nodata:
no-data values for reference image and image to be shifted. The default is (None, None) which indicates
that there is no specific no-data value for both of the input images.
:param calc_corners:
calculate true positions of the dataset corners in order to get a useful matching window position within
the actual image overlap (default: True; deactivated if '-cor0' and '-cor1' are given)
:param binary_ws:
use binary X/Y dimensions for the matching window (default: True)
:param mask_baddata_ref:
path to a 2D boolean mask file (or an instance of GeoArray) for the reference image where all bad data
pixels (e.g. clouds) are marked with True and the remaining pixels with False. Must have the same
geographic extent and projection like 'im_ref'. The mask is used to check if the chosen matching window
position is valid in the sense of useful data. Otherwise, this window position is rejected.
:param mask_baddata_tgt:
path to a 2D boolean mask file (or an instance of GeoArray) for the image to be shifted where all bad data
pixels (e.g. clouds) are marked with True and the remaining pixels with False. Must have the same
geographic extent and projection like 'im_ref'. The mask is used to check if the chosen matching window
position is valid in the sense of useful data. Otherwise, this window position is rejected.
:param CPUs:
number of CPUs to use during pixel grid equalization (default: None, which means 'all CPUs available')
:param force_quadratic_win:
force a quadratic matching window (default: True)
:param progress:
show progress bars (default: True)
:param v:
verbose mode (default: False)
:param path_verbose_out:
an optional output directory for intermediate results
(if not given, no intermediate results are written to disk)
:param q:
quiet mode (default: False)
:param ignore_errors:
Useful for batch processing. (default: False)
In case of error COREG.success == False and COREG.x_shift_px/COREG.y_shift_px is None
"""
self.params = dict([x for x in locals().items() if x[0] != "self"])
# input validation
if gdal.GetDriverByName(fmt_out) is None:
raise ValueError(fmt_out, "'%s' is not a supported GDAL driver." % fmt_out)
if match_gsd and out_gsd:
warnings.warn("'-out_gsd' is ignored because '-match_gsd' is set.\n")
if out_gsd and (not isinstance(out_gsd, list) or len(out_gsd) != 2):
raise ValueError(out_gsd, 'out_gsd must be a list with two values.')
if data_corners_ref and not isinstance(data_corners_ref[0], list):
# group if not [[x,y],[x,y]..] but [x,y,x,y,]
data_corners_ref = [data_corners_ref[i:i + 2]
for i in range(0, len(data_corners_ref), 2)]
if data_corners_tgt and not isinstance(data_corners_tgt[0], list):
# group if not [[x,y],[x,y]..]
data_corners_tgt = [data_corners_tgt[i:i + 2]
for i in range(0, len(data_corners_tgt), 2)]
if nodata and (not isinstance(nodata, Iterable) or len(nodata) != 2):
raise ValueError(nodata, "'nodata' must be an iterable with two values. "
"Got %s with length %s." % (type(nodata), len(nodata)))
for rspAlg in [resamp_alg_deshift, resamp_alg_calc]:
if rspAlg not in _dict_rspAlg_rsp_Int.keys():
raise ValueError("'%s' is not a supported resampling algorithm." % rspAlg)
if resamp_alg_calc in ['average', 5] and (v or not q):
warnings.warn("The resampling algorithm 'average' causes sinus-shaped patterns in fft images that will "
"affect the precision of the calculated spatial shifts! It is highly recommended to "
"choose another resampling algorithm.")
self.path_out = path_out # updated by self.set_outpathes
self.fmt_out = fmt_out
self.out_creaOpt = out_crea_options
self.win_pos_XY = wp # updated by self.get_opt_winpos_winsize()
self.win_size_XY = ws # updated by self.get_opt_winpos_winsize()
self.max_iter = max_iter
self.max_shift = max_shift
self.align_grids = align_grids
self.match_gsd = match_gsd
self.out_gsd = out_gsd
self.target_xyGrid = target_xyGrid
self.rspAlg_DS = resamp_alg_deshift \
if isinstance(resamp_alg_deshift, str) else _dict_rspAlg_rsp_Int[resamp_alg_deshift]
self.rspAlg_calc = resamp_alg_calc \
if isinstance(resamp_alg_calc, str) else _dict_rspAlg_rsp_Int[resamp_alg_calc]
self.calc_corners = calc_corners
self.CPUs = CPUs if CPUs and CPUs <= cpu_count() else cpu_count()
self.bin_ws = binary_ws
self.force_quadratic_win = force_quadratic_win
self.v = v
self.path_verbose_out = path_verbose_out
self.q = q if not v else False # overridden by v
self.progress = progress if not q else False # overridden by q
self.ignErr = ignore_errors
self.max_win_sz_changes = 3 # TODO: änderung der window size, falls nach max_iter kein valider match gefunden
self.ref: Optional[GeoArray_CoReg] = None # set by self.get_image_params
self.shift: Optional[GeoArray_CoReg] = None # set by self.get_image_params
self.matchBox: Optional[boxObj] = None # set by self.get_clip_window_properties()
self.otherBox: Optional[boxObj] = None # set by self.get_clip_window_properties()
self.matchWin: Optional[GeoArray] = None # set by self._get_image_windows_to_match()
self.otherWin: Optional[GeoArray] = None # set by self._get_image_windows_to_match()
self.overlap_poly = None # set by self._get_overlap_properties()
self.overlap_percentage = None # set by self._get_overlap_properties()
self.overlap_area = None # set by self._get_overlap_properties()
self.imfft_xgsd = None # set by self.get_clip_window_properties()
self.imfft_ygsd = None # set by self.get_clip_window_properties()
self.fftw_works = None # set by self._calc_shifted_cross_power_spectrum()
self.fftw_win_size_YX = None # set by calc_shifted_cross_power_spectrum()
self.x_shift_px = None # always in shift image units (image coords) # set by calculate_spatial_shifts()
self.y_shift_px = None # always in shift image units (image coords) # set by calculate_spatial_shifts()
self.x_shift_map = None # set by self.get_updated_map_info()
self.y_shift_map = None # set by self.get_updated_map_info()
self.vec_length_map = None
self.vec_angle_deg = None
self.updated_map_info = None # set by self.get_updated_map_info()
self.ssim_orig = None # set by self._validate_ssim_improvement()
self.ssim_deshifted = None # set by self._validate_ssim_improvement()
self._ssim_improved = None # private attribute to be filled by self.ssim_improved
self.shift_reliability = None # set by self.calculate_spatial_shifts()
self.tracked_errors = [] # expanded each time an error occurs
self.success = None # default
self.deshift_results = None # set by self.correct_shifts()
self._check_and_handle_metaRotation()
self._get_image_params()
self._set_outpathes(im_ref, im_tgt)
self.grid2use = 'ref' if self.shift.xgsd <= self.ref.xgsd else 'shift'
if self.v:
print('resolutions: ', self.ref.xgsd, self.shift.xgsd)
self._get_overlap_properties()
if self.v and self.path_verbose_out:
write_shp(os.path.join(self.path_verbose_out, 'poly_imref.shp'), self.ref.poly, self.ref.prj)
write_shp(os.path.join(self.path_verbose_out, 'poly_im2shift.shp'), self.shift.poly, self.shift.prj)
write_shp(os.path.join(self.path_verbose_out, 'overlap_poly.shp'), self.overlap_poly, self.ref.prj)
# FIXME: transform_mapPt1_to_mapPt2(im2shift_center_map, ds_imref.GetProjection(), ds_im2shift.GetProjection())
# FIXME später basteln für den fall, dass projektionen nicht gleich sind
# get_clip_window_properties
self._get_opt_winpos_winsize()
if not self.q:
print('Matching window position (X,Y): %s/%s' % (self.win_pos_XY[0], self.win_pos_XY[1]))
self._get_clip_window_properties() # sets self.matchBox, self.otherBox and much more
if self.v and self.path_verbose_out and self.matchBox.mapPoly and self.success is not False:
write_shp(os.path.join(self.path_verbose_out, 'poly_matchWin.shp'),
self.matchBox.mapPoly, self.matchBox.prj)
self.success = False if self.success is False or not self.matchBox.boxMapYX else None
self._coreg_info = None # private attribute to be filled by self.coreg_info property
[docs] def _handle_error(self, error, warn=False, warnMsg=None):
"""Append the given error to self.tracked_errors.
This sets self.success to False and raises the error in case self.ignore_errors = True.
:param error: instance of an error
:param warn: whether to give a warning in case error would be ignored otherwise
:param warnMsg: a custom message for the warning
:return:
"""
warn = warn or warnMsg is not None or self.v
self.tracked_errors.append(error)
self.success = False
if self.ignErr and warn:
warnMsg = repr(error) if not warnMsg else warnMsg
print('\nWARNING: ' + warnMsg)
if not self.ignErr:
raise error
[docs] def _set_outpathes(self,
im_ref: Union[(GeoArray, str)],
im_tgt: Union[(GeoArray, str)]
) -> None:
def get_baseN(path):
return os.path.splitext(os.path.basename(path))[0]
# get input paths
def get_input_path(im):
path = im.filePath if isinstance(im, GeoArray) else im
if isinstance(im, GeoArray) and im.filePath is None and self.path_out == 'auto':
raise ValueError(self.path_out, "The output path must be explicitly set in case the input "
"reference or target image is in-memory (without a reference to a "
"physical file on disk). Received path_out='%s'." % self.path_out)
return path
path_im_ref = get_input_path(im_ref)
path_im_tgt = get_input_path(im_tgt)
if self.path_out: # this also applies to self.path_out='auto'
if self.path_out == 'auto':
dir_out, fName_out = os.path.dirname(path_im_tgt), ''
else:
dir_out, fName_out = os.path.split(self.path_out)
if dir_out and fName_out:
# a valid output path is given => do nothing
pass
else:
# automatically create an output directory and filename if not given
if not dir_out:
if not path_im_ref:
dir_out = os.path.abspath(os.path.curdir)
else:
dir_out = os.path.dirname(path_im_ref)
if not fName_out:
ext = 'bsq' if self.fmt_out == 'ENVI' else \
gdal.GetDriverByName(self.fmt_out).GetMetadataItem(gdal.DMD_EXTENSION)
fName_out = fName_out if fName_out not in ['.', ''] else \
'%s__shifted_to__%s' % (get_baseN(path_im_tgt), get_baseN(path_im_ref))
fName_out = fName_out + '.%s' % ext if ext else fName_out
self.path_out = os.path.abspath(os.path.join(dir_out, fName_out))
assert ' ' not in self.path_out, \
"The path of the output image contains whitespaces. This is not supported by GDAL."
else:
# this only happens if COREG is not instanced from within Python and self.path_out is explicitly set to None
# => DESHIFTER will return an array
pass
if self.v:
if self.path_verbose_out:
dir_out, dirname_out = os.path.split(self.path_verbose_out)
if not dir_out:
if self.path_out:
self.path_verbose_out = os.path.dirname(self.path_out)
else:
self.path_verbose_out = \
os.path.abspath(os.path.join(os.path.curdir, 'CoReg_verboseOut__%s__shifted_to__%s'
% (get_baseN(path_im_tgt), get_baseN(path_im_ref))))
elif dirname_out and not dir_out:
self.path_verbose_out = os.path.abspath(os.path.join(os.path.curdir, dirname_out))
assert ' ' not in self.path_verbose_out, \
"'path_verbose_out' contains whitespaces. This is not supported by GDAL."
else:
self.path_verbose_out = None
if self.path_verbose_out and not os.path.isdir(self.path_verbose_out):
os.makedirs(self.path_verbose_out)
[docs] def _get_image_params(self) -> None:
self.ref = GeoArray_CoReg(self.params, 'ref')
self.shift = GeoArray_CoReg(self.params, 'shift')
if not prj_equal(self.ref.prj, self.shift.prj):
from pyproj import CRS
crs_ref = CRS.from_user_input(self.ref.prj)
crs_shift = CRS.from_user_input(self.shift.prj)
name_ref, name_shift = \
(crs_ref.name, crs_shift.name) if not crs_ref.name == crs_shift.name else (crs_ref.srs, crs_shift.srs)
raise RuntimeError(
'Input projections are not equal. Different projections are currently not supported. '
'Got %s vs. %s.' % (name_ref, name_shift))
[docs] def _get_overlap_properties(self) -> None:
with warnings.catch_warnings():
# already warned in GeoArray_CoReg.__init__()
warnings.filterwarnings("ignore", category=RuntimeWarning, message=".*disjunct.*")
overlap_tmp = get_overlap_polygon(self.ref.poly, self.shift.poly, self.v)
self.overlap_poly = overlap_tmp['overlap poly'] # has to be in reference projection
self.overlap_percentage = overlap_tmp['overlap percentage']
self.overlap_area = overlap_tmp['overlap area']
assert self.overlap_poly, 'The input images have no spatial overlap.'
# overlap are must at least cover 16*16 pixels
px_area = self.ref.xgsd * self.ref.ygsd if self.grid2use == 'ref' else self.shift.xgsd * self.shift.ygsd
px_covered = self.overlap_area / px_area
assert px_covered > 16 * 16, \
'Overlap area covers only %s pixels. At least 16*16 pixels are needed.' % px_covered
@property
def are_pixGrids_equal(self):
return prj_equal(self.ref.prj, self.shift.prj) and \
is_coord_grid_equal(self.ref.gt, *self.shift.xygrid_specs, tolerance=1e-8)
[docs] def equalize_pixGrids(self) -> None:
"""Equalize image grids and projections of reference and target image (align target to reference).
NOTE: This method is only called by COREG_LOCAL to speed up processing during detection of displacements.
"""
if not self.are_pixGrids_equal or GEO.has_metaRotation(self.ref) or GEO.has_metaRotation(self.shift):
if not self.q:
print("Equalizing pixel grids and projections of reference and target image...")
def equalize(gA_from: GeoArray, gA_to: GeoArray) -> GeoArray:
if gA_from.bands > 1:
gA_from = gA_from.get_subset(zslice=slice(gA_from.band4match, gA_from.band4match + 1))
gA_from.reproject_to_new_grid(prototype=gA_to, CPUs=self.CPUs)
gA_from.band4match = 0 # after resampling there is only one band in the GeoArray
return gA_from
if self.grid2use == 'ref':
# resample target to reference image
self.shift = equalize(gA_from=self.shift, gA_to=self.ref)
else:
# resample reference to target image
self.ref = equalize(gA_from=self.ref, gA_to=self.shift)
# self.ref.gt = (self.ref.gt[0], 1, self.ref.gt[2], self.ref.gt[3], self.ref.gt[4], -1)
# self.shift.gt = (self.shift.gt[0], 1, self.shift.gt[2], self.shift.gt[3], self.shift.gt[4], -1)
[docs] def show_matchWin(self,
figsize: tuple = (15, 15),
interactive: bool = True,
after_correction: bool = None,
pmin=2,
pmax=98):
"""Show the image content within the matching window.
:param figsize: figure size
:param interactive: whether to return an interactive figure based on 'holoviews' library
:param after_correction: True/False: show the image content AFTER shift correction or before
None: show both states - before and after correction (default)
:param pmin: percentage to be used for excluding the darkest pixels from stretching (default: 2)
:param pmax: percentage to be used for excluding the brightest pixels from stretching
(default: 98)
:return:
"""
if not self.success and after_correction in [True, None]:
warnings.warn('It is only possible to show the matching window before correction of spatial displacements '
'because no valid displacement has been calculated yet.')
after_correction = False
if interactive:
# use Holoviews
try:
import holoviews as hv
except ImportError:
hv = None
if not hv:
raise ImportError(
"This method requires the library 'holoviews'. It can be installed for Conda with "
"the shell command 'conda install -c conda-forge holoviews bokeh'.")
hv.notebook_extension('matplotlib')
hv.Store.add_style_opts(hv.Image, ['vmin', 'vmax'])
# hv.Store.option_setters.options().Image = hv.Options('style', cmap='gnuplot2')
# hv.Store.add_style_opts(hv.Image, ['cmap'])
# renderer = hv.Store.renderers['matplotlib'].instance(fig='svg', holomap='gif')
# RasterPlot = renderer.plotting_class(hv.Image)
# RasterPlot.cmap = 'gray'
otherWin_corr = self._get_deshifted_otherWin() if after_correction in [True, None] else None
xmin, xmax, ymin, ymax = self.matchBox.boundsMap
def get_hv_image(geoArr: GeoArray):
from skimage.exposure import rescale_intensity # import here to avoid static TLS ImportError
arr_masked = np.ma.masked_equal(geoArr[:], geoArr.nodata)
vmin = np.nanpercentile(arr_masked.compressed(), pmin)
vmax = np.nanpercentile(arr_masked.compressed(), pmax)
arr2plot = rescale_intensity(arr_masked, in_range=(vmin, vmax), out_range='int8')
return (
hv.Image(
arr2plot,
bounds=(xmin, ymin, xmax, ymax)
).options(
cmap='gray',
vmin=vmin,
vmax=vmax,
interpolation='none',
fig_inches=figsize,
show_grid=True
)
)
hvIm_matchWin = get_hv_image(self.matchWin)
hvIm_otherWin_orig = get_hv_image(self.otherWin)
hvIm_otherWin_corr = get_hv_image(otherWin_corr) if after_correction in [True, None] else None
if after_correction is None:
# view both states
print('Matching window before and after correction (left and right): ')
# get layouts (docs on options: https://holoviews.org/user_guide)
layout_before = (
hvIm_matchWin.options(title='BEFORE co-registration') +
hvIm_matchWin.options(title='AFTER co-registration')
).options(
fig_inches=figsize
)
layout_after = (
hvIm_otherWin_orig.options(title='BEFORE co-registration') +
hvIm_otherWin_corr.options(title='AFTER co-registration')
).options(
fig_inches=figsize
)
# plot!
imgs = {1: layout_before, 2: layout_after}
hmap = hv.HoloMap(imgs, kdims=['image']).collate().cols(2)
else:
# view state before or after correction
imgs = {1: hvIm_matchWin, 2: hvIm_otherWin_corr if after_correction else hvIm_otherWin_orig}
hmap = hv.HoloMap(imgs, kdims=['image'])
# Construct a HoloMap by evaluating the function over all the keys
# hmap = hv.HoloMap(imgs_corr, kdims=['image']) + hv.HoloMap(imgs_corr, kdims=['image'])
# Construct a HoloMap by defining the sampling on the Dimension
# dmap = hv.DynamicMap(image_slice, kdims=[hv.Dimension('z_axis', values=keys)])
return hmap
else:
# TODO add titles
# TODO handle after_correction=None here
self.matchWin.show(figsize=figsize)
if after_correction:
self._get_deshifted_otherWin().show(figsize=figsize, pmin=pmin, pmax=pmax)
else:
self.otherWin.show(figsize=figsize, pmin=pmin, pmax=pmax)
[docs] def show_cross_power_spectrum(self, interactive: bool = False) -> None:
"""Show a 3D surface of the cross power spectrum.
NOTE: The cross power spectrum is the result from phase correlating the reference and target
image within the matching window.
:param interactive: whether to return an interactice 3D surface plot based on 'plotly' library
:return:
"""
if interactive:
# create plotly 3D surface
# import plotly.plotly as py # online mode -> every plot is uploaded into online plotly account
from plotly.offline import iplot, init_notebook_mode
import plotly.graph_objs as go
init_notebook_mode(connected=True)
z_data = self._calc_shifted_cross_power_spectrum()
data = [go.Surface(z=z_data)]
layout = go.Layout(
title='cross power spectrum',
autosize=False,
width=1000,
height=1000,
margin={'l': 65, 'r': 50, 'b': 65, 't': 90})
fig = go.Figure(data=data, layout=layout)
return iplot(fig, filename='SCPS')
else:
# use matplotlib
scps = self._calc_shifted_cross_power_spectrum()
PLT.subplot_3dsurface(scps.astype(np.float32))
[docs] def _get_opt_winpos_winsize(self) -> None:
"""Calculate optimal window position and size in reference image units.
NOTE: The returned values are computed according to DGM, cloud_mask and trueCornerLonLat.
"""
# dummy algorithm: get center position of overlap instead of searching ideal window position in whole overlap
# TODO automatischer Algorithmus zur Bestimmung der optimalen Window Position
wp = tuple(self.win_pos_XY)
assert type(self.win_pos_XY) in [tuple, list, np.ndarray], \
'The window position must be a tuple of two elements. Got %s with %s elements.' % (type(wp), len(wp))
wp = tuple(wp)
if None in wp:
# use centroid point if possible
overlap_center_pos_x, overlap_center_pos_y = self.overlap_poly.centroid.coords.xy
wp = (wp[0] if wp[0] else overlap_center_pos_x[0]), (wp[1] if wp[1] else overlap_center_pos_y[0])
# validate window position
if not self.overlap_poly.buffer(1e-5).contains(Point(wp)):
# in case the centroid point is not within overlap area
if not self.q:
warnings.warn("The centroid point of the two input images could not be used as matching window "
"position since it is outside of the overlap area. Instead the so called "
"'representative point' is used. Alternatively you can provide your own window "
"position as input parameter.")
# -> use representative point: a point that is garanteed to be within overlap polygon
overlap_center_pos_x, overlap_center_pos_y = self.overlap_poly.representative_point().coords.xy
wp = overlap_center_pos_x[0], overlap_center_pos_y[0]
assert self.overlap_poly.buffer(1e-5).contains(Point(wp))
else:
# validate window position
if not self.overlap_poly.buffer(1e-5).contains(Point(wp)):
self._handle_error(ValueError('The provided window position %s/%s is outside of the overlap '
'area of the two input images. Check the coordinates.' % wp))
# check if window position is within bad data area if a respective mask has been provided
for im in [self.ref, self.shift]:
if im.mask_baddata is not None:
imX, imY = mapXY2imXY(wp, im.mask_baddata.gt)
if im.mask_baddata[int(imY), int(imX)] is True:
self._handle_error(
RuntimeError('According to the provided bad data mask for the %s the chosen window position '
'%s / %s is within a bad data area. Using this window position for coregistration '
'is not reasonable. Please provide a better window position!'
% (im.imName, wp[0], wp[1])))
self.win_pos_XY = wp
self.win_size_XY = (int(self.win_size_XY[0]), int(self.win_size_XY[1])) if self.win_size_XY else (512, 512)
[docs] def _get_clip_window_properties(self) -> None:
"""Calculate all properties of the matching window and the other window.
These windows are used to read the corresponding image positions in the reference and the target image.
NOTE: Even if X- and Y-dimension of the target window is equal, the output window can be NON-quadratic!
"""
# FIXME image sizes like 10000*256 are still possible
wpX, wpY = self.win_pos_XY
wsX, wsY = self.win_size_XY
# image units -> map units
ref_wsX = wsX * self.ref.xgsd
ref_wsY = wsY * self.ref.ygsd
shift_wsX = wsX * self.shift.xgsd
shift_wsY = wsY * self.shift.ygsd
ref_box_kwargs = \
dict(wp=(wpX, wpY),
ws=(ref_wsX, ref_wsY),
gt=self.ref.gt)
shift_box_kwargs = \
dict(wp=(wpX, wpY),
ws=(shift_wsX, shift_wsY),
gt=self.shift.gt)
matchBox =\
boxObj(**ref_box_kwargs) if self.grid2use == 'ref' else \
boxObj(**shift_box_kwargs)
otherBox = \
boxObj(**shift_box_kwargs) if self.grid2use == 'ref' else \
boxObj(**ref_box_kwargs)
overlapWin = \
boxObj(mapPoly=self.overlap_poly,
gt=self.ref.gt)
# clip matching window to overlap area
matchBox.mapPoly = matchBox.mapPoly.intersection(overlapWin.mapPoly)
# check if matchBox extent touches no data area of the image -> if yes: shrink it
overlapPoly_within_matchWin = matchBox.mapPoly.intersection(self.overlap_poly)
if overlapPoly_within_matchWin.area < matchBox.mapPoly.area:
wsX_start, wsY_start = \
1 if wsX >= wsY else \
wsX / wsY, 1 if wsY >= wsX else \
wsY / wsX
box = boxObj(**dict(wp=(wpX, wpY),
ws=(wsX_start, wsY_start),
gt=matchBox.gt))
while True:
box.buffer_imXY(1, 1)
if not box.mapPoly.within(overlapPoly_within_matchWin):
box.buffer_imXY(-1, -1)
matchBox = box
break
# move matching window to imref grid or im2shift grid
mW_rows, mW_cols = \
(self.ref.rows, self.ref.cols) if self.grid2use == 'ref' else \
(self.shift.rows, self.shift.cols)
matchBox.mapPoly = move_shapelyPoly_to_image_grid(matchBox.mapPoly,
matchBox.gt, mW_rows,
mW_cols,
'NW')
# check, if matchBox was moved outside of overlap_poly when moving it to the image grid
if not matchBox.mapPoly.within(overlapWin.mapPoly):
# further shrink matchPoly (1 px buffer is enough because the window was only moved to the grid)
xLarger, yLarger = matchBox.is_larger_DimXY(overlapWin.boundsIm)
matchBox.buffer_imXY(-1 if xLarger else 0,
-1 if yLarger else 0)
# matching_win directly on grid2use (fix rounding error through coordinate transformation)
matchBox.imPoly = round_shapelyPoly_coords(matchBox.imPoly, precision=0)
# check if matching window larger than the other one or equal
if not (matchBox.mapPoly.within(otherBox.mapPoly) or
matchBox.mapPoly == otherBox.mapPoly):
# if yes, find the smallest 'other window' that encloses the matching window
otherBox.boxImYX = \
get_smallest_boxImYX_that_contains_boxMapYX(
matchBox.boxMapYX,
otherBox.gt,
tolerance_ndigits=5 # avoids float coordinate rounding issues
)
# in case after enlarging the 'other window', it gets too large for the overlap area
# -> shrink match window and recompute smallest possible other window until everything is fine
t_start = time.time()
while not otherBox.mapPoly.within(overlapWin.mapPoly):
xLarger, yLarger = otherBox.is_larger_DimXY(overlapWin.boundsIm)
matchBox.buffer_imXY(-1 if xLarger else 0,
-1 if yLarger else 0)
previous_area = otherBox.mapPoly.area
otherBox.boxImYX = \
get_smallest_boxImYX_that_contains_boxMapYX(
matchBox.boxMapYX,
otherBox.gt,
tolerance_ndigits=5 # avoids float coordinate rounding issues
)
if previous_area == otherBox.mapPoly.area or \
time.time() - t_start > 1.5:
# happens, e.g, in case of a triangular footprint
# NOTE: first condition is not always fulfilled -> therefore added timeout of 1.5 sec
self._handle_error(
RuntimeError('Matching window in target image is larger than overlap area but further shrinking '
'the matching window is not possible. Check if the footprints of the input data have '
'been computed correctly.' +
(' Matching window shrinking timed out.' if time.time() - t_start > 5 else '')))
break # break out of while loop in order to avoid that code gets stuck here
# output validation
for winBox in [matchBox, otherBox]:
if winBox.imDimsYX[0] < 16 or \
winBox.imDimsYX[1] < 16:
self._handle_error(
RuntimeError("One of the input images does not have sufficient gray value information "
"(non-no-data values) for placing a matching window at the position %s. "
"Matching failed." % str((wpX, wpY))))
if self.success is not False:
# check result -> ProgrammingError if not fulfilled
def within_equal(inner, outer):
return inner.within(outer.buffer(1e-5)) or \
inner.equals(outer)
assert within_equal(matchBox.mapPoly,
otherBox.mapPoly)
assert within_equal(otherBox.mapPoly,
overlapWin.mapPoly)
if self.grid2use == 'ref':
self.imfft_xgsd = self.ref.xgsd
self.imfft_ygsd = self.ref.ygsd
self.ref.win = matchBox
self.shift.win = otherBox
else:
self.imfft_xgsd = self.shift.xgsd
self.imfft_ygsd = self.shift.ygsd
self.ref.win = otherBox
self.shift.win = matchBox
self.matchBox = matchBox
self.otherBox = otherBox
self.ref.win.size_YX = tuple([int(i) for i in self.ref.win.imDimsYX])
self.shift.win.size_YX = tuple([int(i) for i in self.shift.win.imDimsYX])
match_win_size_XY = tuple(reversed([int(i) for i in matchBox.imDimsYX]))
if not self.q and \
match_win_size_XY != self.win_size_XY:
print('Target window size %s not possible due to too small overlap area or window position too close '
'to an image edge. New matching window size: %s.' % (self.win_size_XY, match_win_size_XY))
# write_shp('matchMapPoly.shp', matchBox.mapPoly,matchBox.prj)
# write_shp('otherMapPoly.shp', otherBox.mapPoly,otherBox.prj)
[docs] def _get_image_windows_to_match(self) -> None:
"""Read the matching window and the other window as subsets.
The other window is resampled to the resolution and the pixel grid of the matching window.
The result consists of two images with the same dimensions and exactly the same corner coordinates.
"""
match_fullGeoArr = self.ref if self.grid2use == 'ref' else self.shift
other_fullGeoArr = self.shift if self.grid2use == 'ref' else self.ref
# read matchWin via subset-read -> self.matchWin.data
rS, rE, cS, cE = GEO.get_GeoArrayPosition_from_boxImYX(self.matchBox.boxImYX)
assert np.array_equal(np.abs(np.array([rS, rE, cS, cE])), np.array([rS, rE, cS, cE])) and \
rE <= match_fullGeoArr.rows and cE <= match_fullGeoArr.cols, \
'Requested area is not completely within the input array for %s.' % match_fullGeoArr.imName
self.matchWin = GeoArray(match_fullGeoArr[rS:rE + 1, cS:cE + 1, match_fullGeoArr.band4match],
geotransform=GEO.get_subset_GeoTransform(match_fullGeoArr.gt, self.matchBox.boxImYX),
projection=copy(match_fullGeoArr.prj),
nodata=copy(match_fullGeoArr.nodata))
self.matchWin.imID = match_fullGeoArr.imID
# read otherWin via subset-read
rS, rE, cS, cE = GEO.get_GeoArrayPosition_from_boxImYX(self.otherBox.boxImYX)
assert np.array_equal(np.abs(np.array([rS, rE, cS, cE])), np.array([rS, rE, cS, cE])) and \
rE <= other_fullGeoArr.rows and cE <= other_fullGeoArr.cols, \
'Requested area is not completely within the input array for %s.' % other_fullGeoArr.imName
self.otherWin = GeoArray(other_fullGeoArr[rS:rE + 1, cS:cE + 1, other_fullGeoArr.band4match],
geotransform=GEO.get_subset_GeoTransform(other_fullGeoArr.gt, self.otherBox.boxImYX),
projection=copy(other_fullGeoArr.prj),
nodata=copy(other_fullGeoArr.nodata))
self.otherWin.imID = other_fullGeoArr.imID
# self.matchWin.deepcopy_array()
# self.otherWin.deepcopy_array()
if self.v:
print('Original matching windows:')
ref_data, shift_data = (self.matchWin[:], self.otherWin[:]) if self.grid2use == 'ref' else \
(self.otherWin[:], self.matchWin[:])
PLT.subplot_imshow([ref_data, shift_data], [self.ref.title, self.shift.title], grid=True)
# resample otherWin.arr to the resolution of matchWin AND make sure the pixel edges are identical
# (in order to make each image show the same window with the same coordinates)
# TODO replace cubic resampling by PSF resampling - average resampling leads to sinus like distortions in the
# TODO fft image that make a precise coregistration impossible. Thats why there is currently no way around
# TODO cubic resampling.
tgt_xmin, tgt_xmax, tgt_ymin, tgt_ymax = self.matchBox.boundsMap
# equalize pixel grids and projection of matchWin and otherWin (ONLY if grids are really different)
if not (self.matchWin.xygrid_specs == self.otherWin.xygrid_specs and
prj_equal(self.matchWin.prj, self.otherWin.prj)):
self.otherWin.arr, self.otherWin.gt = warp_ndarray(self.otherWin.arr,
self.otherWin.gt,
self.otherWin.prj,
self.matchWin.prj,
out_gsd=(self.imfft_xgsd, abs(self.imfft_ygsd)),
out_bounds=([tgt_xmin, tgt_ymin, tgt_xmax, tgt_ymax]),
rspAlg=_dict_rspAlg_rsp_Int[self.rspAlg_calc],
in_nodata=self.otherWin.nodata,
CPUs=self.CPUs,
progress=False)[:2]
if self.matchWin.shape != self.otherWin.shape:
self._handle_error(
RuntimeError('Caught a possible ProgrammingError at window position %s: Bad output of '
'get_image_windows_to_match. Reference image shape is %s whereas shift '
'image shape is %s.' % (str(self.matchBox.wp), self.matchWin.shape, self.otherWin.shape)),
warn=True)
# check of odd dimensions of output images
rows, cols = [i if i % 2 == 0 else i - 1 for i in self.matchWin.shape]
self.matchWin.arr, self.otherWin.arr = self.matchWin.arr[:rows, :cols], self.otherWin.arr[:rows, :cols]
if self.matchWin.box.imDimsYX != self.matchBox.imDimsYX:
self.matchBox = self.matchWin.box # update matchBox
self.otherBox = self.otherWin.box # update otherBox
assert self.matchWin.arr is not None and self.otherWin.arr is not None, 'Creation of matching windows failed.'
[docs] @staticmethod
def _shrink_winsize_to_binarySize(win_shape_YX: tuple,
target_size: tuple = None
) -> Optional[Tuple[int, int]]:
"""Shrink a given window size to the closest binary window size (a power of 2).
NOTE: X- and Y-dimension are handled separately.
:param win_shape_YX: source window shape as pixel units (rows,colums)
:param target_size: source window shape as pixel units (rows,colums)
"""
binarySizes = [2 ** i for i in range(3, 14)] # [8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192]
possibSizes_X = [i for i in binarySizes if i <= win_shape_YX[1]]
possibSizes_Y = [i for i in binarySizes if i <= win_shape_YX[0]]
if possibSizes_X and possibSizes_Y:
tgt_size_X, tgt_size_Y = target_size if target_size else (max(possibSizes_X), max(possibSizes_Y))
closest_to_target_X = int(min(possibSizes_X, key=lambda x: abs(x - tgt_size_X)))
closest_to_target_Y = int(min(possibSizes_Y, key=lambda y: abs(y - tgt_size_Y)))
return closest_to_target_Y, closest_to_target_X
else:
return None
[docs] def _calc_shifted_cross_power_spectrum(self,
im0=None,
im1=None,
precision=np.complex64
) -> np.ndarray:
"""Calculate the shifted cross power spectrum for quantifying X/Y-shifts.
:param im0: reference image
:param im1: subject image to shift
:param precision: to be quantified as a datatype
:return: 2D-numpy-array of the shifted cross power spectrum
"""
im0 = im0 if im0 is not None else self.matchWin[:] if self.matchWin.imID == 'ref' else self.otherWin[:]
im1 = im1 if im1 is not None else self.otherWin[:] if self.otherWin.imID == 'shift' else self.matchWin[:]
assert im0.shape == im1.shape, 'The reference and the target image must have the same dimensions.'
if im0.shape[0] % 2 != 0:
warnings.warn('Odd row count in one of the match images!')
if im1.shape[1] % 2 != 0:
warnings.warn('Odd column count in one of the match images!')
wsYX = self._shrink_winsize_to_binarySize(im0.shape) if self.bin_ws else im0.shape
wsYX = ((min(wsYX),) * 2 if self.force_quadratic_win else wsYX) if wsYX else None
if wsYX not in [None, (0, 0)]:
time0 = time.time()
if self.v:
print('final window size: %s/%s (X/Y)' % (wsYX[1], wsYX[0]))
# FIXME size of self.matchWin is not updated
# FIXME CoRegPoints_grid.WIN_SZ is taken from self.matchBox.imDimsYX but this is not updated
center_YX = np.array(im0.shape) / 2
xmin, xmax = int(center_YX[1] - wsYX[1] / 2), int(center_YX[1] + wsYX[1] / 2)
ymin, ymax = int(center_YX[0] - wsYX[0] / 2), int(center_YX[0] + wsYX[0] / 2)
in_arr0 = im0[ymin:ymax, xmin:xmax].astype(precision)
in_arr1 = im1[ymin:ymax, xmin:xmax].astype(precision)
if self.v:
PLT.subplot_imshow([np.real(in_arr0).astype(np.float32), np.real(in_arr1).astype(np.float32)],
['FFTin ' + self.ref.title, 'FFTin ' + self.shift.title], grid=True)
fft_arr0, fft_arr1 = None, None
if pyfftw and self.fftw_works is not False: # if module is installed and working
try:
fft_arr0 = pyfftw.FFTW(in_arr0, np.empty_like(in_arr0), axes=(0, 1))()
fft_arr1 = pyfftw.FFTW(in_arr1, np.empty_like(in_arr1), axes=(0, 1))()
# catch empty output arrays (for some reason this happens sometimes) -> use numpy fft
# => this is caused by the call of pyfftw.FFTW. Exactly at that moment the input array
# in_arr0 is overwritten with zeros (maybe this is a bug in pyFFTW?)
if np.std(fft_arr0) == 0 or np.std(fft_arr1) == 0:
raise RuntimeError('FFTW result is unexpectedly empty.')
self.fftw_works = True
except RuntimeError:
self.fftw_works = False
# recreate input arrays and use numpy fft as fallback
in_arr0 = im0[ymin:ymax, xmin:xmax].astype(precision)
in_arr1 = im1[ymin:ymax, xmin:xmax].astype(precision)
if self.fftw_works is False or fft_arr0 is None or fft_arr1 is None:
fft_arr0 = fft2(in_arr0)
fft_arr1 = fft2(in_arr1)
# GeoArray(fft_arr0.astype(np.float32)).show(figsize=(15,15))
# GeoArray(fft_arr1.astype(np.float32)).show(figsize=(15,15))
if self.v:
print('forward FFTW: %.2fs' % (time.time() - time0))
eps = np.abs(fft_arr1).max() * 1e-15
# cps == cross-power spectrum of im0 and im2
with np.errstate(invalid='ignore'): # ignore RuntimeWarning: invalid value encountered in true_divide
temp = np.array(fft_arr0 * fft_arr1.conjugate()) / (np.abs(fft_arr0) * np.abs(fft_arr1) + eps)
time0 = time.time()
if 'pyfft' in globals():
ifft_arr = pyfftw.FFTW(temp, np.empty_like(temp), axes=(0, 1), direction='FFTW_BACKWARD')()
else:
ifft_arr = ifft2(temp)
if self.v:
print('backward FFTW: %.2fs' % (time.time() - time0))
cps = np.abs(ifft_arr)
# scps = shifted cps => shift the zero-frequency component to the center of the spectrum
scps = fftshift(cps)
if self.v:
PLT.subplot_imshow([np.real(in_arr0).astype(np.uint16), np.real(in_arr1).astype(np.uint16),
np.real(fft_arr0).astype(np.uint8), np.real(fft_arr1).astype(np.uint8), scps],
titles=['matching window im0', 'matching window im1',
"fft result im0", "fft result im1", "cross power spectrum"], grid=True)
PLT.subplot_3dsurface(np.real(scps).astype(np.float32))
else:
scps = None
self._handle_error(
RuntimeError('The matching window became too small for calculating a reliable match. Matching failed.'))
self.fftw_win_size_YX = wsYX
return scps
[docs] @staticmethod
def _get_peakpos(scps: np.ndarray) -> np.ndarray:
"""Return the row/column position of the peak within the given cross power spectrum.
:param scps: shifted cross power spectrum
:return: [row, column]
"""
max_flat_idx = np.argmax(scps)
return np.array(np.unravel_index(max_flat_idx, scps.shape))
[docs] @staticmethod
def _get_shifts_from_peakpos(peakpos: np.array,
arr_shape: tuple
) -> (float, float):
y_shift = peakpos[0] - arr_shape[0] // 2
x_shift = peakpos[1] - arr_shape[1] // 2
return x_shift, y_shift
[docs] @staticmethod
def _clip_image(im, center_YX, winSzYX): # TODO this is also implemented in GeoArray
def get_bounds(YX: tuple, wsY: float, wsX: float):
return (
int(YX[1] - (wsX / 2)),
int(YX[1] + (wsX / 2)),
int(YX[0] - (wsY / 2)),
int(YX[0] + (wsY / 2))
)
wsY, wsX = winSzYX
xmin, xmax, ymin, ymax = get_bounds(center_YX, wsY, wsX)
return im[ymin:ymax, xmin:xmax]
[docs] def _get_grossly_deshifted_images(self, im0, im1, x_intshift, y_intshift):
# TODO this is also implemented in GeoArray # this should update ref.win.data and shift.win.data
# FIXME avoid that matching window gets smaller although shifting it with the previous win_size would not move
# it into nodata-area
# get_grossly_deshifted_im0
old_center_YX = np.array(im0.shape) / 2
new_center_YX = [old_center_YX[0] + y_intshift,
old_center_YX[1] + x_intshift]
x_left = new_center_YX[1]
x_right = im0.shape[1] - new_center_YX[1]
y_above = new_center_YX[0]
y_below = im0.shape[0] - new_center_YX[0]
maxposs_winsz_x = 2 * min(x_left, x_right)
maxposs_winsz_y = 2 * min(y_above, y_below)
if self.force_quadratic_win:
maxposs_winsz_x = maxposs_winsz_y = min([maxposs_winsz_x, maxposs_winsz_y])
gdsh_im0 = self._clip_image(im0, new_center_YX, [maxposs_winsz_y, maxposs_winsz_x])
# get_corresponding_im1_clip
crsp_im1 = self._clip_image(im1, np.array(im1.shape) / 2, gdsh_im0.shape)
if self.v:
PLT.subplot_imshow([self._clip_image(im0, old_center_YX, gdsh_im0.shape), crsp_im1],
titles=['reference original', 'target'],
grid=True)
PLT.subplot_imshow([gdsh_im0, crsp_im1],
titles=['reference virtually shifted', 'target'],
grid=True)
return gdsh_im0, crsp_im1
[docs] def _find_side_maximum(self, scps):
# peakR, peakC = self._get_peakpos(scps)
peakR, peakC = scps.shape[0] // 2, scps.shape[1] // 2
profileX = scps[peakR, :].flatten() # row profile with values from left to right
profileY = scps[:, peakC].flatten() # column profile with values from top to bottom
# get scps values of side maxima
sm_left, sm_right = profileX[peakC - 1], profileX[peakC + 1]
sm_above, sm_below = profileY[peakR - 1], profileY[peakR + 1]
sidemax_lr = {'value': max([sm_left, sm_right]),
'side': 'left' if sm_left > sm_right else 'right',
'direction_factor': -1 if sm_left > sm_right else 1}
sidemax_ab = {'value': max([sm_above, sm_below]),
'side': 'above' if sm_above > sm_below else 'below',
'direction_factor': -1 if sm_above > sm_below else 1}
if self.v:
print('Horizontal side maximum found %s. value: %s' % (sidemax_lr['side'], sidemax_lr['value']))
print('Vertical side maximum found %s. value: %s' % (sidemax_ab['side'], sidemax_ab['value']))
PLT.subplot_2dline([[range(profileX.size), profileX], [range(profileY.size), profileY]],
titles=['X-Profile', 'Y-Profile'], shapetuple=(1, 2), grid=True)
return sidemax_lr, sidemax_ab
[docs] def _calc_integer_shifts(self, scps: np.ndarray) -> (int, int):
peakpos = self._get_peakpos(scps)
x_intshift, y_intshift = self._get_shifts_from_peakpos(peakpos, scps.shape)
return x_intshift, y_intshift
[docs] def _calc_shift_reliability(self, scps: np.ndarray):
"""Calculate a confidence percentage to be used as an assessment for reliability of the calculated shifts.
:param scps: shifted cross power spectrum
:return:
"""
# calculate mean power at peak
peakR, peakC = self._get_peakpos(scps)
power_at_peak = np.mean(scps[peakR - 1:peakR + 2, peakC - 1:peakC + 2])
# calculate mean power without peak + 3* standard deviation
scps_masked = scps
scps_masked[peakR - 1:peakR + 2, peakC - 1:peakC + 2] = -9999
scps_masked = np.ma.masked_equal(scps_masked, -9999)
power_without_peak = np.mean(scps_masked) + 2 * np.std(scps_masked)
# calculate confidence
confid = 100 - ((power_without_peak / power_at_peak) * 100)
confid = 100 if confid > 100 else 0 if confid < 0 else confid
if not self.q:
print('Estimated reliability of the calculated shifts: %.1f' % confid, '%')
return confid
[docs] def _validate_integer_shifts(self, im0, im1, x_intshift, y_intshift):
if (x_intshift, y_intshift) != (0, 0):
# temporalily deshift images on the basis of calculated integer shifts
gdsh_im0, crsp_im1 = self._get_grossly_deshifted_images(im0, im1, x_intshift, y_intshift)
# check if integer shifts are now gone (0/0)
scps = self._calc_shifted_cross_power_spectrum(gdsh_im0, crsp_im1)
if scps is not None:
if scps.shape[0] < 3 or scps.shape[1] < 3:
self._handle_error(RuntimeError('Shifted cross power spectrum became too small for computing the '
'point of registration. Matching failed.'))
return 'invalid', None, None, scps
peakpos = self._get_peakpos(scps)
x_shift, y_shift = self._get_shifts_from_peakpos(peakpos, scps.shape)
if (x_shift, y_shift) == (0, 0):
return 'valid', 0, 0, scps
else:
return 'invalid', x_shift, y_shift, scps
else:
return 'invalid', None, None, scps
else:
return 'valid', 0, 0, None
[docs] def _calc_subpixel_shifts(self, scps: np.ndarray):
sidemax_lr, sidemax_ab = self._find_side_maximum(scps)
x_subshift = (sidemax_lr['direction_factor'] * sidemax_lr['value']) / (np.max(scps) + sidemax_lr['value'])
y_subshift = (sidemax_ab['direction_factor'] * sidemax_ab['value']) / (np.max(scps) + sidemax_ab['value'])
return x_subshift, y_subshift
[docs] @staticmethod
def _get_total_shifts(x_intshift: int,
y_intshift: int,
x_subshift: float,
y_subshift: float):
return x_intshift + x_subshift, y_intshift + y_subshift
[docs] def _get_deshifted_otherWin(self) -> GeoArray:
"""Return a de-shifted version of self.otherWin as a GeoArray instance.
The output dimensions and geographic bounds are equal to those of self.matchWin and geometric shifts are
corrected according to the previously computed X/Y shifts within the matching window. This allows direct
application of algorithms e.g. measuring image similarity.
The image subset that is resampled in this function is always the same that has been resampled during
computation of geometric shifts (usually the image with the higher geometric resolution).
:returns: GeoArray instance of de-shifted self.otherWin
"""
# shift vectors have been calculated to fit target image onto reference image
# -> so the shift vectors have to be inverted if shifts are applied to reference image
coreg_info = self._get_inverted_coreg_info() if self.otherWin.imID == 'ref' else self.coreg_info
matchFull = self.ref if self.matchWin.imID == 'ref' else self.shift
otherFull = self.ref if self.otherWin.imID == 'ref' else self.shift
ds_results = DESHIFTER(otherFull, coreg_info,
band2process=otherFull.band4match + 1,
clipextent=self.matchBox.mapPoly.bounds,
target_xyGrid=matchFull.xygrid_specs,
q=True
).correct_shifts()
return ds_results['GeoArray_shifted']
[docs] def _validate_ssim_improvement(self,
v: bool = False
) -> (float, float):
"""Compute mean structural similarity index between reference and target image before and after co-registration.
:param v: verbose mode: shows images of the matchWin, otherWin and shifted version of otherWin
:return: SSIM before an after shift correction
"""
from skimage.metrics import structural_similarity as ssim # import here to avoid static TLS import error
assert self.success is not None, \
'Calculate geometric shifts first before trying to measure image similarity improvement!'
assert self.success in [True, None], \
'Since calculation of geometric shifts failed, no image similarity improvement can be measured.'
def normalize(array: np.ndarray) -> np.ndarray:
minval = np.min(array)
maxval = np.max(array)
# avoid zerodivision
if maxval == minval:
maxval += 1e-5
return ((array - minval) / (maxval - minval)).astype(np.float64)
# compute SSIM BEFORE shift correction #
########################################
# using gaussian weights could lead to value errors in case of small images when the automatically calculated
# window size exceeds the image size
self.ssim_orig = ssim(normalize(np.ma.masked_equal(self.matchWin[:],
self.matchWin.nodata)),
normalize(np.ma.masked_equal(self.otherWin[:],
self.otherWin.nodata)),
data_range=1)
# compute SSIM AFTER shift correction #
#######################################
# resample otherWin while correcting detected shifts and match geographic bounds of matchWin
otherWin_deshift_geoArr = self._get_deshifted_otherWin()
# get the corresponding matchWin data
matchWinData = self.matchWin[:]
# check if shapes of two images are unequal (due to bug (?), in some cases otherWin_deshift_geoArr does not have
# the exact same dimensions as self.matchWin -> maybe bounds are handled differently by gdal.Warp)
if not self.matchWin.shape == otherWin_deshift_geoArr.shape: # FIXME this seems to be already fixed
warnings.warn('SSIM input array shapes are not equal. This should not happen. '
'If it happens with your data, please report it here so that the issue can be fixed: '
'https://git.gfz-potsdam.de/danschef/arosics/-/issues/85')
matchFull = \
self.ref if self.matchWin.imID == 'ref' else\
self.shift
matchWinData, _, _ = matchFull.get_mapPos(self.matchBox.mapPoly.bounds,
self.matchWin.prj,
rspAlg='cubic',
band2get=matchFull.band4match)
self.matchWin.clip_to_poly(self.matchBox.mapPoly)
# at the image edges it is possible that the size of the matchBox must be reduced
# in order to make array shapes match
if not matchWinData.shape == otherWin_deshift_geoArr.shape: # FIXME this seems to be already fixed
self.matchBox.buffer_imXY(float(-np.ceil(abs(self.x_shift_px))),
float(-np.ceil(abs(self.y_shift_px))))
otherWin_deshift_geoArr = self._get_deshifted_otherWin()
matchWinData, _, _ = matchFull.get_mapPos(self.matchBox.mapPoly.bounds,
self.matchWin.prj,
rspAlg='cubic',
band2get=matchFull.band4match)
# output validation
if not matchWinData.shape == otherWin_deshift_geoArr.shape:
warnings.warn('SSIM input array shapes could not be equalized. SSIM calculation failed. '
'SSIM of the de-shifted target image is set to 0.')
self.ssim_deshifted = 0
return self.ssim_orig, self.ssim_deshifted
self.ssim_deshifted = ssim(normalize(np.ma.masked_equal(otherWin_deshift_geoArr[:],
otherWin_deshift_geoArr.nodata)),
normalize(np.ma.masked_equal(matchWinData,
self.matchWin.nodata)),
data_range=1)
if v:
GeoArray(matchWinData).show()
self.otherWin.show()
otherWin_deshift_geoArr.show()
if not self.q:
print('Image similarity within the matching window (SSIM before/after correction): %.4f => %.4f'
% (self.ssim_orig, self.ssim_deshifted))
self.ssim_improved = self.ssim_orig <= self.ssim_deshifted
# write win data to disk
# outDir = '/home/gfz-fe/scheffler/temp/ssim_debugging/'
# GeoArray(matchWinData, matchWinGt, matchWinPrj).save(outDir+'matchWinData.bsq')
# otherWinGt = (self.otherWin.boundsMap[0], self.matchWin.xgsd, 0,
# self.otherWin.boundsMap[3], 0, -self.matchWin.imParams.ygsd)
# GeoArray(self.otherWin.data, therWinGt, self.otherWin.prj).save(outDir+'otherWin.data.bsq')
# otherWin_deshift_geoArr.save(outDir+''shifted.bsq')
return self.ssim_orig, self.ssim_deshifted
@property
def ssim_improved(self) -> bool:
"""Return True if image similarity within the matching window has been improved by co-registration."""
if self.success is True:
if self._ssim_improved is None:
ssim_orig, ssim_deshifted = self._validate_ssim_improvement()
self._ssim_improved = ssim_orig <= ssim_deshifted
return self._ssim_improved
@ssim_improved.setter
def ssim_improved(self, has_improved: bool):
self._ssim_improved = has_improved
[docs] def calculate_spatial_shifts(self) -> str:
"""Compute the global X/Y shift between reference and the target image within the matching window.
:return: 'success' or 'fail'
"""
if self.success is False:
return 'fail'
# set self.matchWin and self.otherWin (GeoArray instances)
self._get_image_windows_to_match() # 45-90ms
im0 = self.matchWin[:] if self.matchWin.imID == 'ref' else self.otherWin[:]
im1 = self.otherWin[:] if self.otherWin.imID == 'shift' else self.matchWin[:]
xgsd_factor = self.imfft_xgsd / self.shift.xgsd
ygsd_factor = self.imfft_ygsd / self.shift.ygsd
if self.v:
print('Matching windows with equalized spatial resolution:')
PLT.subplot_imshow([im0, im1], [self.ref.title, self.shift.title], grid=True)
print('xgsd_factor', xgsd_factor)
print('ygsd_factor', ygsd_factor)
print('imfft_xgsd_mapvalues', self.imfft_xgsd)
print('imfft_ygsd_mapvalues', self.imfft_ygsd)
# calculate cross power spectrum without any de-shifting applied
scps = self._calc_shifted_cross_power_spectrum() # 8-18ms
if scps is None:
self.success = False
return 'fail'
# calculate spatial shifts
# calculate integer shifts
count_iter = 1
x_intshift, y_intshift = self._calc_integer_shifts(scps)
if (x_intshift, y_intshift) == (0, 0):
# in case integer shifts are zero
self.success = True
else:
valid_invalid, x_val_shift, y_val_shift, scps = \
self._validate_integer_shifts(im0, im1, x_intshift, y_intshift)
while valid_invalid != 'valid':
count_iter += 1
if count_iter > self.max_iter:
self._handle_error(RuntimeError('No match found in the given window.'))
break
if valid_invalid == 'invalid' and (x_val_shift, y_val_shift) == (None, None):
# this happens if matching window became too small
self.success = False
break
if not self.q:
print('No clear match found yet. Jumping to iteration %s...' % count_iter)
print('input shifts: ', x_val_shift, y_val_shift)
valid_invalid, x_val_shift, y_val_shift, scps = \
self._validate_integer_shifts(im0, im1, x_val_shift, y_val_shift)
# overwrite previous integer shifts if a valid match has been found
if valid_invalid == 'valid':
self.success = True
x_intshift, y_intshift = x_val_shift, y_val_shift
# calculate sub-pixel shifts
if self.success or self.success is None:
# get total pixel shifts
x_subshift, y_subshift = self._calc_subpixel_shifts(scps)
x_totalshift, y_totalshift = self._get_total_shifts(x_intshift, y_intshift, x_subshift, y_subshift)
if max([abs(x_totalshift), abs(y_totalshift)]) > self.max_shift:
self._handle_error(
RuntimeError("The calculated shift (X: %s px / Y: %s px) is recognized as too large to "
"be valid. If you know that it is valid, just set the '-max_shift' "
"parameter to an appropriate value. Otherwise try to use a different window "
"size for matching via the '-ws' parameter or define the spectral bands "
"to be used for matching manually ('-br' and '-bs')."
% (x_totalshift, y_totalshift)))
else:
self.success = True
self.x_shift_px, self.y_shift_px = x_totalshift * xgsd_factor, y_totalshift * ygsd_factor
# get map shifts
new_originX, new_originY = imXY2mapXY((self.x_shift_px, self.y_shift_px), self.shift.gt)
self.x_shift_map, self.y_shift_map = new_originX - self.shift.gt[0], new_originY - self.shift.gt[3]
# get length of shift vector in map units
self.vec_length_map = float(np.sqrt(self.x_shift_map ** 2 + self.y_shift_map ** 2))
# get angle of shift vector
self.vec_angle_deg = GEO.angle_to_north((self.x_shift_px, self.y_shift_px)).tolist()[0]
# print results
if not self.q:
print('Detected integer shifts (X/Y): %s/%s' % (x_intshift, y_intshift))
print('Detected subpixel shifts (X/Y): %s/%s' % (x_subshift, y_subshift))
print('Calculated total shifts in fft pixel units (X/Y): %s/%s'
% (x_totalshift, y_totalshift))
print('Calculated total shifts in reference pixel units (X/Y): %s/%s'
% (x_totalshift, y_totalshift))
print('Calculated total shifts in target pixel units (X/Y): %s/%s'
% (self.x_shift_px, self.y_shift_px))
print('Calculated map shifts (X,Y):\t\t\t\t %s/%s' % (self.x_shift_map, self.y_shift_map))
print('Calculated absolute shift vector length in map units: %s' % self.vec_length_map)
print('Calculated angle of shift vector in degrees from North: %s' % self.vec_angle_deg)
if self.x_shift_px or self.y_shift_px:
self._get_updated_map_info()
# set self.ssim_before and ssim_after
self._validate_ssim_improvement() # FIXME uses the not updated matchWin size
self.shift_reliability = self._calc_shift_reliability(scps)
return 'success'
[docs] def _get_updated_map_info(self) -> None:
original_map_info = geotransform2mapinfo(self.shift.gt, self.shift.prj)
self.updated_map_info = copy(original_map_info)
self.updated_map_info[3] = str(float(original_map_info[3]) + self.x_shift_map)
self.updated_map_info[4] = str(float(original_map_info[4]) + self.y_shift_map)
if not self.q:
print('Original map info:', original_map_info)
print('Updated map info: ', self.updated_map_info)
@property
def coreg_info(self) -> dict:
"""Return a dictionary containing everything to correct the detected global X/Y shift of the target image."""
if self._coreg_info:
return self._coreg_info
else:
if self.success is None:
self.calculate_spatial_shifts()
self._coreg_info = {
'corrected_shifts_px': {
'x': self.x_shift_px,
'y': self.y_shift_px
},
'corrected_shifts_map': {
'x': self.x_shift_map,
'y': self.y_shift_map
},
'original map info':
geotransform2mapinfo(self.shift.gt, self.shift.prj),
'updated map info':
self.updated_map_info,
'reference projection':
self.ref.prj,
'reference geotransform':
self.ref.gt,
'reference grid': [
[self.ref.gt[0], self.ref.gt[0] + self.ref.gt[1]],
[self.ref.gt[3], self.ref.gt[3] + self.ref.gt[5]]
],
'reference extent': {
'cols': self.ref.xgsd,
'rows': self.ref.ygsd
}, # FIXME not needed anymore
'success':
self.success}
return self._coreg_info
[docs] def _get_inverted_coreg_info(self) -> dict:
"""Return an inverted dictionary of coreg_info.
This dictionary can be passed to DESHIFTER in order to fit the REFERENCE image onto the TARGET image.
"""
inv_coreg_info = copy(self.coreg_info)
inv_coreg_info['corrected_shifts_px']['x'] *= -1
inv_coreg_info['corrected_shifts_px']['y'] *= -1
inv_coreg_info['corrected_shifts_map']['x'] *= -1
inv_coreg_info['corrected_shifts_map']['y'] *= -1
inv_coreg_info['original map info'] = geotransform2mapinfo(self.ref.gt, self.ref.prj)
inv_coreg_info['reference geotransform'] = self.shift.gt
inv_coreg_info['reference grid'] = self.shift.xygrid_specs
inv_coreg_info['reference projection'] = self.shift.prj
if inv_coreg_info['updated map info']:
updated_map_info = copy(inv_coreg_info['original map info'])
updated_map_info[3] = str(float(inv_coreg_info['original map info'][3]) - self.x_shift_map)
updated_map_info[4] = str(float(inv_coreg_info['original map info'][4]) - self.y_shift_map)
inv_coreg_info['updated map info'] = updated_map_info
return inv_coreg_info
[docs] def correct_shifts(self) -> dict:
"""Correct the already calculated X/Y shift of the target image.
:return: COREG.deshift_results (dictionary)
"""
DS = DESHIFTER(self.shift, self.coreg_info,
path_out=self.path_out,
fmt_out=self.fmt_out,
out_crea_options=self.out_creaOpt,
out_gsd=self.out_gsd,
resamp_alg=self.rspAlg_DS,
align_grids=self.align_grids,
match_gsd=self.match_gsd,
target_xyGrid=self.target_xyGrid,
nodata=self.shift.nodata,
CPUs=self.CPUs,
progress=self.progress,
v=self.v,
q=self.q)
self.deshift_results = DS.correct_shifts()
return self.deshift_results