def get_vel(fn, fill=True):
ds = gdal.Open(fn)
if fill:
import dem_downsample_fill
ds = dem_downsample_fill.gdalfill_ds(ds)
u_b = ds.GetRasterBand(1)
v_b = ds.GetRasterBand(2)
u = iolib.b_getma(u_b)
v = iolib.b_getma(v_b)
m = np.ma.sqrt(u*u + v*v)
return u, v, m
def apply_z_shift(ds, dz, createcopy=True):
print("Z shift: ", dz)
if createcopy:
ds_shift = iolib.mem_drv.CreateCopy('', ds, 1)
else:
ds_shift = ds
b = ds_shift.GetRasterBand(1)
a = iolib.b_getma(b)
a += dz
b.WriteArray(a.filled())
return ds_shift
def apply_z_shift(ds, dz, createcopy=True):
if isinstance(dz, np.ndarray):
print("Z shift offset array mean: ", dz.mean())
else:
print("Z shift offset: ", dz)
if createcopy:
ds_shift = iolib.mem_drv.CreateCopy('', ds, 0)
else:
ds_shift = ds
b = ds_shift.GetRasterBand(1)
a = iolib.b_getma(b)
a += dz
b.WriteArray(a.filled())
return ds_shift
def gdalfill_b(b, edgemask=True, prog_func=None):
# Create mask of exterior nodata
bma = iolib.b_getma(b)
# Check to make sure there are actually holes
# if bma.count_masked() > 0:
if np.any(bma.mask):
# Create 8-bit mask_ds, and add edgemask
mask_ds = memdrv.Create('', b.XSize, b.YSize, 1, gdal.GDT_Byte)
maskband = mask_ds.GetRasterBand(1)
# The - here inverts the mask so that holes are set to 0 (invalid) and
# filled by gdalFillNodata
maskband.WriteArray((-bma.mask).astype(int))
# Now fill holes in the output
print("Filling holes")
max_distance = 40
smoothing_iterations = 0
fill_opt = []
gdal.FillNodata(b,
maskband,
max_distance,
smoothing_iterations,
fill_opt,
callback=prog_func)
# Apply the original edgemask
# Note: need to implement convexhull option like geolib.get_outline
if edgemask:
print("Generating outer edgemask")
edgemask = malib.get_edgemask(bma)
out_ma = np.ma.array(b.ReadAsArray(), mask=edgemask)
else:
out_ma = np.ma.array(b.ReadAsArray())
# Set ndv to input ndv
out_ma.set_fill_value(bma.fill_value)
# Write the filled, masked array to our tmp dataset
b.WriteArray(out_ma.filled())
# Free mask_ds
mask_ds = None
else:
print("No holes found in input band")
return b
def warp(src_ds, res=None, extent=None, t_srs=None, r='cubic', driver=mem_drv, dst_fn=None, dst_ndv=None, verbose=True):
"""Warp an input dataset with predetermined arguments specifying output res/extent/srs
This is the function that actually calls gdal.ReprojectImage
Parameters
----------
src_ds : gdal.Dataset object
Dataset to be warped
res : float
Desired output resolution
extent : list of float
Desired output extent in t_srs coordinate system
t_srs : osr.SpatialReference()
Desired output spatial reference
r : str
Desired resampling algorithm
driver : GDAL Driver to use for warp
Either MEM or GTiff
dst_fn : str
Output filename (for disk warp)
dst_ndv : float
Desired output NoData Value
Returns
-------
dst_ds : gdal.Dataset object
Warped dataset (either in memory or on disk)
"""
src_srs = geolib.get_ds_srs(src_ds)
if t_srs is None:
t_srs = geolib.get_ds_srs(src_ds)
src_gt = src_ds.GetGeoTransform()
#Note: get_res returns [x_res, y_res]
#Could just use gt here and average x_res and y_res
src_res = geolib.get_res(src_ds, t_srs=t_srs, square=True)[0]
if res is None:
res = src_res
if extent is None:
extent = geolib.ds_geom_extent(src_ds, t_srs=t_srs)
#Note: GDAL Lanczos creates block artifacts
#Wait for gdalwarp to support gaussian resampling
#Want to use Lanczos for downsampling
#if src_res < res:
# gra = gdal.GRA_Lanczos
#See http://blog.codinghorror.com/better-image-resizing/
# Suggests cubic for downsampling, bilinear for upsampling
# gra = gdal.GRA_Cubic
#Cubic for upsampling
#elif src_res >= res:
# gra = gdal.GRA_Bilinear
gra = parse_rs_alg(r)
#At this point, the resolution and extent values must be float
#Extent must be list
res = float(res)
extent = [float(i) for i in extent]
#Might want to move this to memwarp_multi, keep memwarp basic w/ gdal.GRA types
#Create progress function
prog_func = None
if verbose:
prog_func = gdal.TermProgress
if dst_fn is None:
#This is a dummy fn if only in mem, but can be accessed later via GetFileList()
#Actually, no, doesn't look like the filename survivies
dst_fn = ''
#Compute output image dimensions
dst_nl = int(round((extent[3] - extent[1])/res))
dst_ns = int(round((extent[2] - extent[0])/res))
#dst_nl = int(math.ceil((extent[3] - extent[1])/res))
#dst_ns = int(math.ceil((extent[2] - extent[0])/res))
#dst_nl = int(math.floor((extent[3] - extent[1])/res))
#dst_ns = int(math.floor((extent[2] - extent[0])/res))
if verbose:
print('nl: %i ns: %i res: %0.3f' % (dst_nl, dst_ns, res))
#Create output dataset
src_b = src_ds.GetRasterBand(1)
src_dt = src_b.DataType
src_nl = src_ds.RasterYSize
src_ns = src_ds.RasterXSize
dst_ds = driver.Create(dst_fn, dst_ns, dst_nl, src_ds.RasterCount, src_dt)
dst_ds.SetProjection(t_srs.ExportToWkt())
#Might be an issue to use src_gt rotation terms here with arbitrary extent/res
dst_gt = [extent[0], res, src_gt[2], extent[3], src_gt[4], -res]
dst_ds.SetGeoTransform(dst_gt)
#This will smooth the input before downsampling to prevent aliasing, fill gaps
#Pretty inefficent, as we need to create another intermediate dataset
gauss = False
for n in range(1, src_ds.RasterCount+1):
if dst_ndv is None:
src_b = src_ds.GetRasterBand(n)
src_ndv = iolib.get_ndv_b(src_b)
dst_ndv = src_ndv
b = dst_ds.GetRasterBand(n)
b.SetNoDataValue(dst_ndv)
b.Fill(dst_ndv)
if gauss:
from pygeotools.lib import filtlib
#src_a = src_b.GetVirtualMemArray()
#Compute resampling ratio to determine filter window size
res_ratio = float(res)/src_res
if verbose:
print("Resampling factor: %0.3f" % res_ratio)
#Might be more efficient to do iterative gauss filter with size 3, rather than larger windows
f_size = math.floor(res_ratio/2.)*2+1
#This is conservative to avoid filling holes with noise
#f_size = math.floor(res_ratio/2.)*2-1
if f_size <= 1:
continue
if verbose:
print("Smoothing window size: %i" % f_size)
#Create temp dataset to store filtered array - avoid overwriting original
temp_ds = driver.Create('', src_ns, src_nl, src_ds.RasterCount, src_dt)
temp_ds.SetProjection(src_srs.ExportToWkt())
temp_ds.SetGeoTransform(src_gt)
temp_b = temp_ds.GetRasterBand(n)
temp_b.SetNoDataValue(dst_ndv)
temp_b.Fill(dst_ndv)
src_a = iolib.b_getma(src_b)
src_a = filtlib.gauss_fltr_astropy(src_a, size=f_size)
#Want to run with maskfill, so only fills gaps, without expanding isolated points
temp_b.WriteArray(src_a)
src_ds = temp_ds
#In theory, NN should be fine since we already smoothed. In practice, cubic still provides slightly better results
#gra = gdal.GRA_NearestNeighbour
"""
if not verbose:
#Suppress GDAL progress bar
orig_stdout = sys.stdout
sys.stdout = open(os.devnull, 'w')
"""
#Note: default maxerror=0.0, second 0.0 argument
gdal.ReprojectImage(src_ds, dst_ds, src_srs.ExportToWkt(), t_srs.ExportToWkt(), gra, 0.0, 0.0, prog_func)
"""
if not verbose:
sys.stdout.close()
sys.stdout = orig_stdout
"""
#Note: this is now done in diskwarp
#Write out to disk
#if driver != mem_drv:
# dst_ds.FlushCache()
#Return GDAL dataset object in memory
return dst_ds
def main():
parser = getparser()
args = parser.parse_args()
hs_overlay = args.hs_overlay
kmz = args.kmz
opacity = args.alpha
cmap = args.cmap
fn = args.fn
print(fn)
ds = gdal.Open(fn)
b = ds.GetRasterBand(1)
ndv = iolib.get_ndv_b(b)
print("Loading input raster")
a = iolib.b_getma(b)
clim = args.clim
if clim is None:
clim = malib.calcperc(a, (2, 98))
print("Generating color ramp")
cramp_fn = os.path.splitext(fn)[0]+'_ramp.txt'
ncolors = 21
csteps = np.linspace(0, 1, ncolors)
cm = plt.get_cmap(cmap)
#Compute raster values between specified min/max
vals = np.linspace(clim[0], clim[1], ncolors)
#Compute rgba for these values on the given color ramp
cvals = cm(csteps, bytes=True)
#Combine into single array
cramp = np.vstack((vals, cvals.T)).T
#Set alpha to desired transparency
cramp[:,-1] = opacity * 255
header = '#val r g b a'
footer = 'nv %s %s %s 0' % (ndv, ndv, ndv)
np.savetxt(cramp_fn, cramp, fmt='%f %i %i %i %i', header=header, footer=footer, comments='')
print("Generating gdaldem color-relief tif")
color_fn = os.path.splitext(fn)[0]+'_color.tif'
if not os.path.exists(color_fn):
#cmd = 'gdaldem color-relief -nearest_color_entry -alpha %s %s %s' % (fn, cramp_fn, color_fn)
cmd = ['gdaldem', 'color-relief', '-alpha']
cmd.extend(iolib.gdal_opt_co)
cmd.extend([fn, cramp_fn, color_fn])
print(' '.join(cmd))
subprocess.call(cmd, shell=False)
if kmz:
make_kmz(color_fn)
if hs_overlay:
print("Generating shaded relief")
hs_fn = os.path.splitext(fn)[0]+'_hs_az315.tif'
#Check to see if file exists, or if provided as input
if not os.path.exists(hs_fn):
cmd = ['gdaldem', 'hillshade']
#cmd.extend('-compute_edges')
cmd.extend(iolib.gdal_opt_co)
cmd.extend([fn, hs_fn])
print(' '.join(cmd))
subprocess.call(cmd, shell=False)
print("Loading shaded relief and calculating percentile stretch")
hs = iolib.fn_getma(hs_fn)
hs_clim = malib.calcperc(hs, (1, 99))
#Since imagemagick was compiled with quantum depth 16, need to scale levels
hs_clim = (hs_clim[0]*65535/255., hs_clim[1]*65535/255.)
print("Generating color composite shaded relief")
overlay_fn = os.path.splitext(color_fn)[0]+'_hs.tif'
if not os.path.exists(overlay_fn):
#Can also try hsvmerge.py
#cmd = 'composite %s %s -dissolve "%i" %s' % (color_fn, hs_fn, opacity*100, overlay_fn)
#This uses imagemagick composite function
#For some reason, this level adjustment is not working
#cmd = ['convert', hs_fn, color_fn, '-compose', 'dissolve', \
cmd = ['convert', hs_fn, '-level', '%i,%i' % hs_clim, color_fn, '-compose', 'dissolve', \
'-define', 'compose:args=%i' % int(opacity*100), '-composite', '-compress', 'LZW', overlay_fn]
#cmd = ['composite', color_fn, hs_fn, '-dissolve', str(int(opacity*100)), '-compress', 'LZW', overlay_fn]
print(' '.join(cmd))
subprocess.call(cmd, shell=False)
print("Updating georeferencing metadata")
out_ndv = 0
overlay_ds = gdal.Open(overlay_fn, gdal.GA_Update)
overlay_ds.SetProjection(ds.GetProjection())
overlay_ds.SetGeoTransform(ds.GetGeoTransform())
for n in range(overlay_ds.RasterCount):
overlay_ds.GetRasterBand(n+1).SetNoDataValue(out_ndv)
overlay_ds = None
#Rewrite with blocks and LZW-compression
print("Creating tiled and compressed version")
tmp_fn = '/tmp/temp_%s.tif' % os.getpid()
cmd = ['gdal_translate',]
cmd.extend(iolib.gdal_opt_co)
cmd.extend((overlay_fn, tmp_fn))
print(' '.join(cmd))
subprocess.call(cmd, shell=False)
shutil.move(tmp_fn, overlay_fn)
if not os.path.exists(overlay_fn+'.ovr'):
print("Generating overviews")
cmd = ['gdaladdo', '-ro', '-r', 'average', '--config', \
'COMPRESS_OVERVIEW', 'LZW', '--config', 'BIGTIFF_OVERVIEW', 'YES', \
overlay_fn, '2', '4', '8', '16', '32', '64']
print(' '.join(cmd))
subprocess.call(cmd, shell=False)
if kmz:
make_kmz(overlay_fn)
def main():
parser = getparser()
args = parser.parse_args()
if args.seedmode == 'existing_velocity':
if args.vx_fn is None or args.vy_fn is None:
parser.error('"-seedmode existing_velocity" requires "-vx_fn" and "-vy_fn"')
print('\n%s' % datetime.now())
print('%s UTC\n' % datetime.utcnow())
align = args.align
seedmode = args.seedmode
spr = args.refinement
erode = args.erode
#Correlator tile timeout
#With proper seeding, correlation should be very fast
#timeout = 360
timeout = 1200
threads = args.threads
kernel = (args.kernel, args.kernel)
#SGM correlator
if spr > 3:
#kernel = (7,7)
kernel = (11,11)
erode = 0
#Smooth the output F.tif
smoothF = args.filter
res = args.tr
#Resample input to something easier to work with
#res = 4.0
#Open input files
fn1 = args.fn1
fn2 = args.fn2
if not iolib.fn_check(fn1) or not iolib.fn_check(fn2):
sys.exit("Unable to locate input files")
if args.outdir is not None:
outdir = args.outdir
else:
outdir = '%s__%s_vmap_%sm_%ipx_spm%i' % (os.path.splitext(os.path.split(fn1)[1])[0], \
os.path.splitext(os.path.split(fn2)[1])[0], res, kernel[0], spr)
#Note, can encounter filename length issues in boost, just use vmap prefix
outprefix = '%s/vmap' % (outdir)
if not os.path.exists(outdir):
os.makedirs(outdir)
#Check to see if inputs have geolocation and projection information
ds1 = iolib.fn_getds(fn1)
ds2 = iolib.fn_getds(fn2)
if geolib.srs_check(ds1) and geolib.srs_check(ds2):
ds1_clip_fn = os.path.join(outdir, os.path.splitext(os.path.basename(fn1))[0]+'_warp.tif')
ds2_clip_fn = os.path.join(outdir, os.path.splitext(os.path.basename(fn2))[0]+'_warp.tif')
if not os.path.exists(ds1_clip_fn) or not os.path.exists(ds2_clip_fn):
#This should write out files to new subdir
ds1_clip, ds2_clip = warplib.diskwarp_multi_fn([fn1, fn2], extent='intersection', res=res, r='average', outdir=outdir)
ds1_clip = None
ds2_clip = None
#However, if inputs have identical extent/res/proj, then link to original files
if not os.path.exists(ds1_clip_fn):
os.symlink(os.path.abspath(fn1), ds1_clip_fn)
if not os.path.exists(ds2_clip_fn):
os.symlink(os.path.abspath(fn2), ds2_clip_fn)
align = 'None'
#Mask support - limit correlation only to rock/ice surfaces, no water/veg
#This masks input images - guarantee we won't waste time correlating over vegetation
#TODO: Add support to load arbitrary raster or shp mask
if args.mask_input:
ds1_masked_fn = os.path.splitext(ds1_clip_fn)[0]+'_masked.tif'
ds2_masked_fn = os.path.splitext(ds2_clip_fn)[0]+'_masked.tif'
if not os.path.exists(ds1_masked_fn) or not os.path.exists(ds2_masked_fn):
#Load NLCD or bareground mask
from demcoreg.dem_mask import get_lulc_mask
ds1_clip = iolib.fn_getds(ds1_clip_fn)
lulc_mask_fn = os.path.join(outdir, 'lulc_mask.tif')
#if not os.path.exists(nlcd_mask_fn):
lulc_mask = get_lulc_mask(ds1_clip, mask_glaciers=False, filter='not_forest')
iolib.writeGTiff(lulc_mask, lulc_mask_fn, ds1_clip)
ds1_clip = None
#Now apply to original images
#This could be problematic for huge inputs, see apply_mask.py
#lulc_mask = lulc_mask.astype(int)
for fn in (ds1_clip_fn, ds2_clip_fn):
ds = iolib.fn_getds(fn)
a = iolib.ds_getma(ds)
a = np.ma.array(a, mask=~(lulc_mask))
if a.count() > 0:
out_fn = os.path.splitext(fn)[0]+'_masked.tif'
iolib.writeGTiff(a,out_fn,ds)
a = None
else:
sys.exit("No unmasked pixels over bare earth")
ds1_clip_fn = ds1_masked_fn
ds2_clip_fn = ds2_masked_fn
else:
ds1_clip_fn = fn1
ds2_clip_fn = fn2
#Now let user specify alignment methods as option - don't hardcode
#align = 'Homography'
#align = 'AffineEpipolar'
ds1 = None
ds2 = None
#Should have extra kwargs option here
stereo_opt = get_stereo_opt(threads=threads, kernel=kernel, timeout=timeout, \
erode=erode, spr=spr, align=align)
#Stereo arguments
#Latest version of ASP should accept tif without camera models
#stereo_args = [ds1_clip_fn, ds2_clip_fn, outprefix]
#Nope - still need to provide dummy camera models, and they must be unique files
#Use the dummy.tsai file bundled in the vmap repo
dummy_tsai = os.path.join(os.path.split(os.path.realpath(__file__))[0], 'dummy.tsai')
dummy_tsai2 = os.path.splitext(dummy_tsai)[0]+'2.tsai'
if not os.path.exists(dummy_tsai2):
dummy_tsai2 = os.symlink(dummy_tsai, os.path.splitext(dummy_tsai)[0]+'2.tsai')
stereo_args = [ds1_clip_fn, ds2_clip_fn, dummy_tsai, dummy_tsai2, outprefix]
#Run stereo_pprc
if not os.path.exists(outprefix+'-R_sub.tif'):
run_cmd('stereo_pprc', stereo_opt+stereo_args, msg='0: Preprocessing')
#Copy proj info to outputs, this should happen automatically now?
for ext in ('L', 'R', 'L_sub', 'R_sub', 'lMask', 'rMask', 'lMask_sub', 'rMask_sub'):
geolib.copyproj(ds1_clip_fn, '%s-%s.tif' % (outprefix,ext))
#Prepare seeding for stereo_corr
#TODO: these are untested after refactoring
if not os.path.exists(outprefix+'_D_sub.tif'):
#Don't need to do anything for default seed-mode 1
if seedmode == 'sparse_disp':
#Sparse correlation of full-res images
stereo_opt.extend(['--corr-seed-mode', '3'])
sparse_disp_opt = []
sparse_disp_opt.extend(['--Debug', '--coarse', '512', '--fine', '256', '--no_epipolar_fltr'])
sparse_disp_opt.extend(['-P', str(threads)])
sparse_disp_args = [outprefix+'-L.tif', outprefix+'-R.tif', outprefix]
run_cmd('sparse_disp', sparse_disp_opt+sparse_disp_args, msg='0.5: D_sub generation')
elif seedmode == 'existing_velocity':
#User-input low-res velocity maps for seeding
#TODO: Add functions that fetch best available velocities for Ant/GrIS or user-defined low-res velocities
#Automatically query GoLive velocities here
vx_fn = args.vx_fn
vy_fn = args.vy_fn
#Check for existence
#HMA seeding
vdir = '/nobackup/deshean/rpcdem/hma/velocity_jpl_amaury_2013-2015'
vx_fn = os.path.join(vdir, 'PKH_WRS2_B8_2013_2015_snr5_n1_r170_res12.x_vel.TIF')
vy_fn = os.path.join(vdir, 'PKH_WRS2_B8_2013_2015_snr5_n1_r170_res12.y_vel.TIF')
if os.path.exists(vx_fn) and os.path.exists(vy_fn):
ds1_clip = iolib.fn_getds(ds1_clip_fn)
ds1_res = geolib.get_res(ds1_clip, square=True)[0]
#Compute L_sub res - use this for output dimensions
L_sub_fn = outprefix+'-L_sub.tif'
L_sub_ds = gdal.Open(L_sub_fn)
L_sub_x_scale = float(ds1_clip.RasterXSize) / L_sub_ds.RasterXSize
L_sub_y_scale = float(ds1_clip.RasterYSize) / L_sub_ds.RasterYSize
L_sub_scale = np.max([L_sub_x_scale, L_sub_y_scale])
L_sub_res = ds1_res * L_sub_scale
#Since we are likely upsampling here, use cubicspline
vx_ds_clip, vy_ds_clip = warplib.memwarp_multi_fn([vx_fn, vy_fn], extent=ds1_clip, \
t_srs=ds1_clip, res=L_sub_res, r='cubicspline')
ds1_clip = None
#Get vx and vy arrays
vx = iolib.ds_getma(vx_ds_clip)
vy = iolib.ds_getma(vy_ds_clip)
#Determine time interval between inputs
#Use to scaling of known low-res velocities
t_factor = get_t_factor_fn(ds1_clip_fn, ds2_clip_fn, ds=vx_ds_clip)
if t_factor is not None:
#Compute expected offset in scaled pixels
dx = (vx*t_factor)/L_sub_res
dy = (vy*t_factor)/L_sub_res
#Note: Joughin and Rignot's values are positive y up!
#ASP is positive y down, so need to multiply these values by -1
#dy = -(vy*t_factor)/L_sub_res
#Should smooth/fill dx and dy
#If absolute search window is only 30x30
#Don't seed, just use fixed search window
#search_window_area_thresh = 900
search_window_area_thresh = 0
search_window = np.array([dx.min(), dy.min(), dx.max(), dy.max()])
dx_p = calcperc(dx, perc=(0.5, 99.5))
dy_p = calcperc(dy, perc=(0.5, 99.5))
search_window = np.array([dx_p[0], dy_p[0], dx_p[1], dy_p[1]])
search_window_area = (search_window[2]-search_window[0]) * (search_window[3]-search_window[1])
if search_window_area < search_window_area_thresh:
stereo_opt.extend(['--corr-seed-mode', '0'])
stereo_opt.append('--corr-search')
stereo_opt.extend([str(x) for x in search_window])
#pad_perc=0.1
#stereo_opt.extend(['--corr-sub-seed-percent', str(pad_perc)]
#Otherwise, generate a D_sub map from low-res velocity
else:
stereo_opt.extend(['--corr-seed-mode', '3'])
#This is relative to the D_sub scaled disparities
d_sub_fn = L_sub_fn.split('-L_sub')[0]+'-D_sub.tif'
gen_d_sub(d_sub_fn, dx, dy)
#If the above didn't generate a D_sub.tif for seeding, run stereo_corr to generate Low-res D_sub.tif
if not os.path.exists(outprefix+'-D_sub.tif'):
newopt = ['--compute-low-res-disparity-only',]
run_cmd('stereo_corr', newopt+stereo_opt+stereo_args, msg='1.1: Low-res Correlation')
#Copy projection info to D_sub
geolib.copyproj(outprefix+'-L_sub.tif', outprefix+'-D_sub.tif')
#Mask D_sub to limit correlation over bare earth surfaces
#This _should_ be a better approach than masking input images, but stereo_corr doesn't honor D_sub
#Still need to mask input images before stereo_pprc
#Left this in here for reference, or if this changes in ASP
if False:
D_sub_ds = gdal.Open(outprefix+'-D_sub.tif', gdal.GA_Update)
#Mask support - limit correlation only to rock/ice surfaces, no water/veg
from demcoreg.dem_mask import get_nlcd, mask_nlcd
nlcd_fn = get_nlcd()
nlcd_ds = warplib.diskwarp_multi_fn([nlcd_fn,], extent=D_sub_ds, res=D_sub_ds, t_srs=D_sub_ds, r='near', outdir=outdir)[0]
#validmask = mask_nlcd(nlcd_ds, valid='rock+ice')
validmask = mask_nlcd(nlcd_ds, valid='not_forest', mask_glaciers=False)
nlcd_mask_fn = os.path.join(outdir, 'nlcd_validmask.tif')
iolib.writeGTiff(validmask, nlcd_mask_fn, nlcd_ds)
#Now apply to D_sub (band 3 is valid mask)
#validmask = validmask.astype(int)
for b in (1,2,3):
dsub = iolib.ds_getma(D_sub_ds, b)
dsub = np.ma.array(dsub, mask=~(validmask))
D_sub_ds.GetRasterBand(b).WriteArray(dsub.filled())
D_sub_ds = None
#OK, finally run stereo_corr full-res integer correlation with appropriate seeding
if not os.path.exists(outprefix+'-D.tif'):
run_cmd('stereo_corr', stereo_opt+stereo_args, msg='1: Correlation')
geolib.copyproj(ds1_clip_fn, outprefix+'-D.tif')
#Run stereo_rfne
if spr > 0:
if not os.path.exists(outprefix+'-RD.tif'):
run_cmd('stereo_rfne', stereo_opt+stereo_args, msg='2: Refinement')
geolib.copyproj(ds1_clip_fn, outprefix+'-RD.tif')
d_fn = make_ln(outdir, outprefix, '-RD.tif')
else:
ln_fn = outprefix+'-RD.tif'
if os.path.lexists(ln_fn):
os.remove(ln_fn)
os.symlink(os.path.split(outprefix)[1]+'-D.tif', ln_fn)
#Run stereo_fltr
if not os.path.exists(outprefix+'-F.tif'):
run_cmd('stereo_fltr', stereo_opt+stereo_args, msg='3: Filtering')
geolib.copyproj(ds1_clip_fn, outprefix+'-F.tif')
d_fn = make_ln(outdir, outprefix, '-F.tif')
if smoothF and not os.path.exists(outprefix+'-F_smooth.tif'):
print('Smoothing F.tif')
from pygeotools.lib import filtlib
#Fill holes and smooth F
F_fill_fn = outprefix+'-F_smooth.tif'
F_ds = gdal.Open(outprefix+'-F.tif', gdal.GA_ReadOnly)
#import dem_downsample_fill
#F_fill_ds = dem_downsample_fill.gdalfill_ds(F_fill_ds)
print('Creating F_smooth.tif')
F_fill_ds = iolib.gtif_drv.CreateCopy(F_fill_fn, F_ds, 0, options=iolib.gdal_opt)
F_ds = None
for n in (1, 2):
print('Smoothing band %i' % n)
b = F_fill_ds.GetRasterBand(n)
b_fill_bma = iolib.b_getma(b)
#b_fill_bma = iolib.b_getma(dem_downsample_fill.gdalfill(b))
#Filter extreme values (careful, could lose areas of valid data with fastest v)
#b_fill_bma = filtlib.perc_fltr(b_fill_bma, perc=(0.01, 99.99))
#These filters remove extreme values and fill data gaps
#b_fill_bma = filtlib.median_fltr_skimage(b_fill_bma, radius=7, erode=0)
#b_fill_bma = filtlib.median_fltr(b_fill_bma, fsize=7, origmask=True)
#Gaussian filter
b_fill_bma = filtlib.gauss_fltr_astropy(b_fill_bma, size=9)
b.WriteArray(b_fill_bma)
F_fill_ds = None
d_fn = make_ln(outdir, outprefix, '-F_smooth.tif')
print('\n%s' % datetime.now())
print('%s UTC\n' % datetime.utcnow())
#If time interval is specified, convert pixel displacements to rates
if args.dt != 'none':
#Check if vm.tif already exists
#Should probably just overwrite by default
#if os.path.exists(os.path.splitext(d_fn)[0]+'_vm.tif'):
# print("\nFound existing velocity magnitude map!\n"
#else:
#Generate output velocity products and figure
#Requires that vmap repo is in PATH
cmd = ['disp2v.py', d_fn]
#Note: this will attempt to automatically determine control surfaces
#disp2v.py will accept arbitrary mask, could pass through here
if args.remove_offsets:
cmd.append('-remove_offsets')
cmd.extend(['-dt', args.dt])
print("Converting disparities to velocities")
print(cmd)
subprocess.call(cmd)