Wideband deconvolution

WSClean has a ‘wideband’ multi-frequency deconvolution mode, which allows cleaning channels joinedly. This means that peak finding is performed in the sum of all channels, allowing deep cleaning, and the PSF is subtracted from each channel & polarization individually, scaled to the value of the peak in that image, which takes care of spectral variation and deconvolves each channel with its own PSF.

Wideband options are available since WSClean version 1.1.

Usage

A typical run in multi-frequency deconvolution would look like:

wsclean -join-channels -channels-out 4 [other parameters] \
  <measurement set>

This outputs 4 deconvolved images at the different frequencies and the weighted average of those 4. This mode can be combined with Cotton-Schwab imaging (with -mgain ...) which would fill the MODEL_DATA column with frequency dependent model info, so that it is possible to perform self-cal with the proper frequency information. Joining channels can also be combined with joining polarizations (-join-polarizations, see polarimetric deconvolution) to clean polarizations and channels both joinedly. This is available from version 1.5.

Joined channel deconvolution is computationally more expensive than bandwidth-integrated cleaning. If the processing time (or memory) becomes untractable, the speed and memory usage can be improved by using channel interpolation (see description of -deconvolution-channels below) or by using parallel/subimage deconvolution.

Note

Something to be aware of is that the -join-channels parameter turns on MF weighting.

Multiple measurement sets and subbands

WSClean calculates the full available bandwidth (over all measurement sets) and splits that into chunks with the same amount of input channels (or 1 off if the nr of input channels is not divisable by the nr of output channels). For example, if you have four measurement sets where each contains data from a different subband, e.g. 100 MHz, 110 MHz, 200 MHz and 210 MHz, and you specify “-join-channels -channels-out 2”, the first images will be centred at 105 MHz and contain data from the first two MSes, and the other images will be centred at 205 MHz.

Example:

wsclean -join-channels -channels-out 3 -niter 10000 \
  -mgain 0.8 -auto-threshold 3 \
  band-100-MHz.ms band-110-MHz.ms \
  band-145-MHz.ms band-155-MHz.ms \
  band-190-MHz.ms band-200-MHz.ms

This will perform multi-frequency deconvolution and output images at 105 MHz, 150 MHz and 195 MHz. This works since WSClean version 1.9.

In case the measurement sets have a different amout of channels and have gaps between them, the normal division of WSClean might divide the channels somewhat undesirably. For example, the data of the MSSS survey observations are stored in 10 measurement sets, each containing one ‘band’, but one of those 10 has fewer channels. The default WSClean division divides those “missing” channels out over the bandwidth, producing this imaging table:

.      # Pol Ch JG ²G In Freq(MHz) ch
| Independent group:
+-+-J- 0  I   0  0  0  0  119-121 (39)
  +-J- 1  I   1  0  1  0  121-126 (40)
  +-J- 2  I   2  0  2  0  126-130 (39)
  +-J- 3  I   3  0  3  0  130-136 (40)
  +-J- 4  I   4  0  4  0  136-146 (39)
  +-J- 5  I   5  0  5  0  146-150 (40)
  +-J- 6  I   6  0  6  0  150-152 (39)
  +-J- 7  I   7  0  7  0  156-158 (40)

There are two ways to remedy this. The first, most automated way is to use an option called -gap-channel-division (available since WSClean 2.6), which calculates the gaps between channels, and splits the input channels into output channels by splitting the largest gap until the number of output channels has been reached. For the above situation, this gives the following table:

.      # Pol Ch JG ²G In Freq(MHz) ch
| Independent group:
+-+-J- 0  I   0  0  0  0  119-121 (40)
  +-J- 1  I   1  0  1  0  124-126 (40)
  +-J- 2  I   2  0  2  0  128-130 (40)
  +-J- 3  I   3  0  3  0  134-136 (40)
  +-J- 4  I   4  0  4  0  142-144 (36)
  +-J- 5  I   5  0  5  0  146-148 (40)
  +-J- 6  I   6  0  6  0  150-152 (40)
  +-J- 7  I   7  0  7  0  156-158 (40)

Notice the different bandwidth per channel and output channel index 4 which now contains 36 channels.

The second option is to manually add splits by specifying frequencies. For this, the option -channel-division-frequencies can be used. The list doesn’t necessarily contain all splits necessary. If fewer splits are given than required, WSClean will further subdivide the remaining parts as necessary. Be aware that splits are in Hz, but can be in scientific notation (e.g. 1.4e9 is 1.4 GHz). An example:

wsclean -size 512 512 -scale 1amin -channels-out 8 \
  -channel-division-frequencies 140e6,141e6,142e6 \
  observation.ms
[..]
=== IMAGING TABLE ===
      # Pol Ch JG ²G In Freq(MHz)
+-+-J- 0  I   0  0  0  0  134-137 (39)
+-+-J- 1  I   1  1  1  0  137-140 (39)
+-+-J- 2  I   2  2  2  0  140-140 (6)
+-+-J- 3  I   3  3  3  0  140-141 (7)
+-+-J- 4  I   4  4  4  0  141-142 (6)
+-+-J- 5  I   5  5  5  0  142-142 (6)
+-+-J- 6  I   6  6  6  0  142-153 (140)
+-+-J- 7  I   7  7  7  0  153-164 (141)

Fitting smooth spectra

The joined channel deconvolution method discussed above does not enforce a smooth spectra; each channel gets a separate solution. If one wants to image many spectral channels separately, while it is known that the sources have smooth behaviour, it is possible to enforce this during cleaning. Currently, WSClean supports fitting a polynomial and fitting a double-logarithmic polynomial. The command line parameters for this are -fit-spectral-pol and -fit-spectral-log-pol. Both require an extra parameter specifying the number of terms (degrees of freedom). For example, -fit-spectral-log-pol 2 will fit a power law through all the output channels.

A simple – but very slow – example to perform cleaning with spectral fitting:

wsclean -multiscale -join-channels -channels-out 128 -niter 10000 \
  -mgain 0.8 -auto-threshold 3 -fit-spectral-pol 4 observations.ms

This will fit a polynomial with 4 terms (i.e., a third-order polynomial). This would be similar to CASA multi-term deconvolution with nterms=4. During each minor clean cycle, the 128 images at different frequencies will be added together, the pixel with the highest summed brightness is selected, the brightness for that pixel is found for each image, a 3rd order polynomial is fitted through those measurements and the smoothed “model” component is added to the model, as well as convolved with the PSF and subtracted from the residual dirty image. Spectral fitting works in all joined-channel modes (i.e., hogbom, multi-scale, iuwt, moresane).

As you might imagine, doing a clean with 128 images in memory is expensive, both in terms of memory and computing. Since it is not necessary to have that many images in memory when fitting only a few terms, it is also possible to decrease the number of output channels just during deconvolution. This is done with the -deconvolution-channels parameter, for example:

wsclean -join-channels -channels-out 128 -niter 10000 \
  -mgain 0.8 -auto-threshold 3 -fit-spectral-pol 4 \
  -deconvolution-channels 8 observations.ms

This will decrease the number of images from 128 to 8 before starting the deconvolution by averaging groups of 16 channels together. Cleaning is then performed with just 8 images. After cleaning, the requested function (3rd order polynomial in this case) is fitted to the model, and the model is interpolated using that function. This is much faster than the previous command, and equally precise. Setting the deconvolution channels is supported in all modes since WSClean 2.2. The spectral-fitting features were added in WSClean version 1.11.

Forced spectrum fitting

Spectral fitting is useful to reduce the degrees of freedom in deconvolution. It does not always produce accurate, physical spectral indices. This can be for various reasons, e.g. because the bandwidth is too small, sources are complex causing degeneracies, low signal-to-noise, or a combination. WSClean therefore has a method called ‘forced spectrum fitting’. In this mode, a pre-existing spectral index map is used during the deconvolution, and the resulting spectra of the model components are forced onto this spectral index map. Together with multiscale cleaning and source list output, the forced spectrum mode allows building (text) models of sources with accurate spectral index information.

The spectral index map may be the result of earlier runs or from fitting between data from other bandwidths / telescopes. An added advantage is that those maps can be smoothed or otherwise processed to limit the effect of noise and imaging artefacts. The mode is enabled by combining -force-spectrum <fits filename> with -fit-log-pol <N>. The fits file should have the same dimensions and coordinate system as used for the imaging, and each pixel should hold a spectral index value. This mode should currently not be used together with the -deconvolution-channels option.

To force the spectral index to a spectral index map, the option -fit-log-pol 2 should be used and the provided fits file should contain a single image. To additionally constrain higher order terms (curvature, etc.), the -fit-log-pol parameter should be adjusted accordingly and the fits file provided to the -force-spectrum option should contain multiple images (so it should be a cube). Any 3rd dimension will be interpreted by WSClean as the “terms”-dimension, and thus the images will in sequence be interpreted as the spectral index map, curvature map, 3rd order log-polynomial map, etc. In Python, it is simple to create such a file from a standard image. For example, the following code will open an existing image my-example-image.fits and use it as template for writing a 3-dimensional forced spectrum map, for which the first term (the spectral index) is set to -0.85 and the second term (the curvature) is set to 0.5:

from astropy.io import fits
import numpy

with fits.open('my-example-image.fits') as hdu:
  hdu[0].data = numpy.ndarray([1, 2, height, width])
  hdu[0].data[:, 0, :, :] = -0.85
  hdu[0].data[:, 1, :, :] = 0.5
  hdu[0].writeto('forced-spectrum-map.fits')

This resulting map can be used in wsclean with parameters -fit-log-pol 3 -force-spectrum forced-spectrum-map.fits. Remember that the -fit-log-pol option takes as parameter the number of terms being fitted, and not the order of the polynomial.

A description and further testing of the forced spectrum method can be found in Ceccotti et al. (2023). Basic spectral index forcing is available since WSClean version 3.0. In version 3.3 this was generalized to constraining any number of terms.

Fit normal or logarithmic polynomials?

In general, fitting polynomials (fit-spectral-pol) works better than fitting logarithmic functions (fit-spectral-log-pol). The latter option fits every component to a power law or higher-termed logarithmic function. This works fine on strong sources, but when cleaning picks up some partly-negative artefacts, it can create bad results and fail. As an example, when one spectral value is negative and the others are positive, a power-law fit that minimizes the squared error, might create extremely high values at the edge channels.

As a side note, I believe that CASA also doesn’t fit logarithmic polynomials, but rather fits normal polynomials (I’m not 100% sure from the papers about the technique, but from inspecting the CASA code, it seems to use normal polynomials). Hence, getting CASA’s “nterms” behaviour is closest to -fit-spectral-pol.

Avoiding “steps” in the model data visibilities

When splitting up bandwidth with the -channels-out option, the output MODEL_DATA visibilities will have a step function. For example, when splitting up the bandwidth of a 256 channel set in 4 parts (-channels-out 4), there will be a step each 64 channels. When the MODEL_DATA is important, for example when removing the continuuum data from a HI set, these steps might be undesirable. Using -fit-spectral-pol does not remove these steps; that option only forces the individual images to obey a polynomial. To remove the steps, one would have to use -fit-spectral-pol and specify as many channels to -channels-out as that are present in the measurement set. This is of course very computationally and memory expensive, in particular during deconvolution. To make that possible in some cases, the above option -deconvolution-channels can be used in combination with the other options.

Another option to solve these step functions is to interpolate in the frequency during the prediction (i.e. when calculating the MODEL_DATA). WSClean can currently not do this, but will hopefully soon be implemented. Because of the already-available features mentioned above it has not been of very high priority, since the quality improvement on continuum imaging is rather small and the computational savings might only exist in some more exotic cases.

Relation to CASA’s multi-term deconvolution

To avoid confusion, WSClean’s wide-band mode is not the same as the multi-frequency deconvolution algorithm in CASA, i.e., the algorithm described in Sault & Wieringa (1994), further enhanced in Rau and Cornwell (2011). The multi-frequency deconvolution algorithm in WSClean is called “joined-channel deconvolution” and is described in Offringa and Smirnov (2017). As shown there, the multi-frequency implementation of WSClean is computationally orders of magnitude faster than MSMFS, which is the most important difference between the two algorithms. Accuracy wise, they produce similar results.

Although WSClean’s mode and CASA’s mode both mitigate the problem of spectral variation over the imaged bandwidth, the underlying algorithms are different and have different settings. Both CASA and WSClean can fit smooth functions over frequency. WSClean has additional functionality for the imaging of cubes that can have unsmooth spectral functions, such as for spectral lines or off-axis imaging (where the beam causes large/steep fluctuations over frequency). Smooth function fitting is implemented in WSClean version 1.10.