GNU Astronomy Utilities
7.2.1 NoiseChisel changes after publication
NoiseChisel was initially introduced in
Akhlaghi and Ichikawa . It is
thus strongly recommended to read this paper for a good understanding of
what it does and how each parameter influences the output. To help in
understanding how it works, that paper has a large number of figures
showing every step on multiple mock and real examples.
However, the paper cannot be updated anymore, but NoiseChisel has evolved
(and will continue to do so): better algorithms or steps have been found,
thus options will be added or removed. This book is thus the final and
definitive guide to NoiseChisel. The aim of this section is to make the
transition from the paper to the installed version on your system, as
smooth as possible with the list below. For a more detailed list of changes
in previous Gnuastro releases/versions, please see the NEWS
The most important change since the publication of that paper is that from
Gnuastro 0.6, NoiseChisel is only in charge on detection. Segmentation of
the detected signal was spun-off into a separate program:
Segment. This spin-off allows much greater creativity and is in the
spirit of Gnuastro’s modular design (see Program design philosophy).
Below you can see the major changes since that paper was published. First,
the removed options/features are discussed, then we review the new features
that have been added.
- --skysubtracted: This option was used to account for the extra
noise that is added if the Sky value has already been subtracted. However,
NoiseChisel estimates the Sky standard deviation based on the input data,
not exposure maps or other theoretical considerations. Therefore the
standard deviation of the undetected pixels also contains the errors due to
any previous sky subtraction. This option is therefore no longer present in
--dilate: In the paper, true detections were dilated for a final
dig into the noise. However, the simple 8-connected dilation produced boxy
results which were not realistic and could miss diffuse flux. The final dig
into the noise is now done by “grow”ing the true detections, similar to
how true clumps were grown, see the description of --detgrowquant
below and in Detection options for more on the new alternative.
- Segmentation has been completely moved to a new program: Segment.
- --widekernel: NoiseChisel uses the difference between the mode and
median to identify if a tile should be used for estimating the quantile
thresholds (see Quantifying signal in a tile). Until now, NoiseChisel
would convolve an image once and estimate the proper tiles for quantile
estimations on the convolved image. The same convolved image would later be
used for quantile estimation. A larger kernel does increase the skewness
(and thus difference between the mode and median, therefore helps in
detecting the presence signal), however, it disfigures the
shapes/morphology of the objects.
This new --widekernel option (and a corresponding --wkhdu
option to specify its HDU) option are added to solve such cases. When its
given, the input will be convolved with both the sharp (given through the
--kernel option) and wide kernels. The mode and median are
calculated on the dataset that is convolved with the wider kernel, then the
quantiles are estimated on the image convolved with the sharper kernel.
- --noerodequant: to specify a quantile threshold where erosion
will not apply. This is useful to detect sharper point-like sources that
will be missed due to too much erosion. To completely ignore this features
give this option a value of 1 (only the largest valued pixel in the input
will not be eroded).
- --qthreshtilequant: to manually remove the measured qthresh from
some tiles. This feature helps in detecting large and extended diffuse
(almost flat) signal when necessary, see Detection options.
- --detgrowquant: is used to grow the final true detections until a
given quantile in the same way that clumps are grown during segmentation
(compare columns 2 and 3 in Figure 10 of the paper). It replaces the old
--dilate option in the paper and older versions of
Gnuastro. Dilation is a blind growth method which causes objects to be boxy
or diamond shaped when too many layers are added. However, with the growth
method that is defined now, we can follow the signal into the noise with
any shape. The appropriate quantile depends on your dataset’s correlated
noise properties and how cleanly it was Sky subtracted. The new
--detgrowmaxholesize can also be used to specify the maximum hole
size to fill as part of this growth, see the description in Detection options for more details.
This new growth process can be much more successful in detecting diffuse
flux around true detections compared to dilation and give more realistic
results, but it can also increase the NoiseChisel run time (depending on
the given value and input size).
- --cleangrowndet: A process to further clean/remove the possibility
of false detections, see the descriptions under this option in
Read in other formats.
GNU Astronomy Utilities 0.7 manual, August 2018.