# Licensed under a 3-clause BSD style license - see LICENSE.rst
"""
Module to remove cosmic rays from an astronomical image using the
L.A.Cosmic algorithm (van Dokkum 2001; PASP 113, 1420).
"""
import numpy as np
from astropy import log
from astropy.nddata import block_reduce, block_replicate
from scipy import ndimage
__all__ = ['lacosmic']
[docs]
def lacosmic(data, contrast, cr_threshold, neighbor_threshold,
error=None, mask=None, background=None, effective_gain=None,
readnoise=None, maxiter=4, border_mode='mirror'):
r"""
Remove cosmic rays from an astronomical image using the L.A.Cosmic
algorithm.
The `L.A.Cosmic algorithm
<http://www.astro.yale.edu/dokkum/lacosmic/>`_ is based on Laplacian
edge detection and is described in `van Dokkum (2001; PASP 113,
1420)
<https://ui.adsabs.harvard.edu/abs/2001PASP..113.1420V/abstract>`_.
Parameters
----------
data : array_like
The 2D array of the image.
contrast : float
Contrast threshold between the Laplacian image and the
fine-structure image. If your image is critically sampled, use
a value around 2. If your image is undersampled (e.g., HST
data), a value of 4 or 5 (or more) is more appropriate. If your
image is oversampled, use a value between 1 and 2. For details,
please see `PASP 113, 1420 (2001)
<https://ui.adsabs.harvard.edu/abs/2001PASP..113.1420V/abstract>`_,
which calls this parameter :math:`f_{\mbox{lim}}`. In
particular, Figure 4 shows the approximate relationship between
the ``contrast`` parameter and the full-width half-maximum (in
pixels) of stars in your image.
cr_threshold : float
The Laplacian signal-to-noise ratio threshold for cosmic-ray
detection.
neighbor_threshold : float
The Laplacian signal-to-noise ratio threshold for detection of
cosmic rays in pixels neighboring the initially-identified
cosmic rays.
error : array_like, optional
The 1-sigma errors of the input ``data``. If ``error`` is not
input, then ``effective_gain`` and ``readnoise`` will be used to
construct an approximate model of the ``error``. If ``error``
is input, it will override the ``effective_gain`` and
``readnoise`` parameters. ``error`` must have the same shape as
``data``.
mask : array_like (bool), optional
A boolean mask, with the same shape as ``data``, where a `True`
value indicates the corresponding element of ``data`` is masked.
Masked pixels are ignored when identifying cosmic rays. It is
highly recommended that saturated stars be included in ``mask``.
background : float or array_like, optional
The background level previously subtracted from the input
``data``. ``background`` may either be a scalar value or a 2D
image with the same shape as the input ``data``. If the input
``data`` has not been background subtracted, then set
``background=None`` (default).
effective_gain : float, array-like, optional
Ratio of counts (e.g., electrons or photons) to the units of
``data``. For example, if your input ``data`` are in units of
ADU, then ``effective_gain`` should represent electrons/ADU. If
your input ``data`` are in units of electrons/s then
``effective_gain`` should be the exposure time (or an exposure
time map). ``effective_gain`` and ``readnoise`` must be
specified if ``error`` is not input.
readnoise : float, optional
The read noise (in electrons) in the input ``data``.
``effective_gain`` and ``readnoise`` must be specified if
``error`` is not input.
maxiter : float, optional
The maximum number of iterations. The default is 4. The routine
will automatically exit if no additional cosmic rays are
identified in an iteration. If the routine is still identifying
cosmic rays after four iterations, then you are likely digging
into sources (e.g., saturated stars) and/or the noise. In
that case, try inputing a ``mask`` or increasing the value of
``cr_threshold``.
border_mode : {'reflect', 'constant', 'nearest', 'mirror', 'wrap'}, optional
The mode in which the array borders are handled during
convolution and median filtering. For 'constant', the fill
value is 0. The default is 'mirror', which matches the original
L.A.Cosmic algorithm.
Returns
-------
cleaned_image : `~numpy.ndarray`
The cosmic-ray cleaned image.
crmask : `~numpy.ndarray` (bool)
A mask image of the identified cosmic rays. Cosmic-ray pixels
have a value of `True`.
"""
block_size = 2.0
kernel = np.array([[0.0, -1.0, 0.0], [-1.0, 4.0, -1.0], [0.0, -1.0, 0.0]])
clean_data = data.copy()
if background is not None:
clean_data += background
final_crmask = np.zeros(data.shape, dtype=bool)
if error is not None:
if data.shape != error.shape:
raise ValueError('error and data must have the same shape')
clean_error_image = error
ncosmics, ncosmics_tot = 0, 0
for iteration in range(maxiter):
sampled_img = block_replicate(clean_data, block_size)
convolved_img = ndimage.convolve(sampled_img, kernel,
mode=border_mode).clip(min=0.0)
laplacian_img = block_reduce(convolved_img, block_size)
if clean_error_image is None:
if effective_gain is None or readnoise is None:
raise ValueError('effective_gain and readnoise must be '
'input if error is not input')
med5_img = ndimage.median_filter(clean_data, size=5,
mode=border_mode).clip(min=1.e-5)
error_image = (np.sqrt(effective_gain * med5_img + readnoise ** 2)
/ effective_gain)
else:
error_image = clean_error_image
snr_img = laplacian_img / (block_size * error_image)
# this is used to remove extended structures (larger than ~5x5)
snr_img -= ndimage.median_filter(snr_img, size=5, mode=border_mode)
# used to remove compact bright objects
med3_img = ndimage.median_filter(clean_data, size=3, mode=border_mode)
med7_img = ndimage.median_filter(med3_img, size=7, mode=border_mode)
finestruct_img = ((med3_img - med7_img) / error_image).clip(min=0.01)
cr_mask1 = snr_img > cr_threshold
# NOTE: to follow the paper exactly, this condition should be
# "> contrast * block_size". "lacos_im.cl" uses simply "> contrast"
cr_mask2 = (snr_img / finestruct_img) > contrast
cr_mask = cr_mask1 * cr_mask2
if mask is not None:
cr_mask = np.logical_and(cr_mask, ~mask)
# grow cosmic rays by one pixel and check in snr_img
selem = np.ones((3, 3))
neigh_mask = ndimage.binary_dilation(cr_mask, selem)
cr_mask = cr_mask1 * neigh_mask
# now grow one more pixel and lower the detection threshold
neigh_mask = ndimage.binary_dilation(cr_mask, selem)
cr_mask = (snr_img > neighbor_threshold) * neigh_mask
# previously unknown cosmic rays found in this iteration
crmask_new = np.logical_and(~final_crmask, cr_mask)
ncosmics = np.count_nonzero(crmask_new)
final_crmask = np.logical_or(final_crmask, cr_mask)
ncosmics_tot += ncosmics
log.info(f'Iteration {iteration + 1}: Found {ncosmics} cosmic-ray '
f'pixels, Total: {ncosmics_tot}')
if ncosmics == 0:
if background is not None:
clean_data -= background
return clean_data, final_crmask
clean_data = _clean_masked_pixels(clean_data, final_crmask, size=5,
exclude_mask=mask)
if background is not None:
clean_data -= background
return clean_data, final_crmask
def _clean_masked_pixels(data, mask, size=5, exclude_mask=None):
"""
Clean masked pixels in an image. Each masked pixel is replaced by
the median of unmasked pixels in a 2D window of ``size`` centered on
it. If all pixels in the window are masked, then the window is
increased in size until unmasked pixels are found.
Pixels in ``exclude_mask`` are not cleaned, but they are excluded
when calculating the local median.
"""
if (size % 2) != 1: # pragma: no cover
raise ValueError('size must be an odd integer')
if data.shape != mask.shape: # pragma: no cover
raise ValueError('mask must have the same shape as data')
ny, nx = data.shape
mask_coords = np.argwhere(mask)
if exclude_mask is not None:
if data.shape != exclude_mask.shape: # pragma: no cover
raise ValueError('exclude_mask must have the same shape as data')
maskall = np.logical_or(mask, exclude_mask)
else:
maskall = mask
mask_idx = maskall.nonzero()
data_nanmask = data.copy()
data_nanmask[mask_idx] = np.nan
nexpanded = 0
for coord in mask_coords:
y, x = coord
median_val, expanded = _local_median(data_nanmask, x, y, nx, ny,
size=size)
data[y, x] = median_val
if expanded:
nexpanded += 1
if nexpanded > 0:
log.info(f' Found {nexpanded} {size}x{size} masked regions while '
'cleaning.')
return data
def _local_median(data_nanmask, x, y, nx, ny, size=5, expanded=False):
"""
Compute the local median in a 2D window, excluding NaN.
"""
hy, hx = size // 2, size // 2
x0, x1 = np.array([x - hx, x + hx + 1]).clip(0, nx)
y0, y1 = np.array([y - hy, y + hy + 1]).clip(0, ny)
region = data_nanmask[y0:y1, x0:x1].ravel()
goodpixels = region[np.isfinite(region)]
if len(goodpixels) > 0:
median_val = np.median(goodpixels)
else:
newsize = size + 2 # keep size odd
median_val, expanded = _local_median(data_nanmask, x, y, nx, ny,
size=newsize, expanded=True)
return median_val, expanded