Strange pattern after using "cgcurs"

[taianmoon]

Hi,

I'm trying to clean a 5500MHz image using Miriad functions "mfclean" and then "cgcurs". However, after the dirty map is cleaned using "cgcurs", in which several regions were specified (tring with two brightest sources in the image now), there appeared a strange pattern. As can be seen in the attachment, the left panel shows the restored image after "mfclean" is used on the dirty map (which looks okay), and the right panel shows the restored image after "cgcurs" is used on the dirty map with selected regions. There seems to be some patterns near the edge of the image, especially on the top of the image.

I've tried to adjust the image size (tried 3, 4, and 5 times of the primary beam) and tried to use several different weighting parameters, but failed to remove this pattern. So not sure what happened...

Thanks a lot!

Tai-An

[Mark.Wieringa]

Hi Tai-An,

the 'strange pattern' is just the part of the image that hasn't been cleaned. Because you specified clean areas with cgcurs, mfclean limits the map areas that are searched for components. If some parts of the image are further than half the beam size away they don't get cleaned. Specify options=double in invert to increase the beam size and allow mfclean to clean a larger area.

Cheers,

Mark

[taianmoon]

Hi Mark,

I've doubled the beamsize and problems in both images have been solved. However, there are still evident ring-like artefacts for the bright compact source in the lower part of this image, so I'm trying on the "peeling" process. The attachment shows the cleaned & self-calibrated map on the left, and a preliminary trial on "peeling" on the right. I don't think peeling has improved the map much, so thinking I might have done it wrong. Below is a quick list of procedures I've gone through via Miriad:

1. Construct on-axis sources model: use the restored map (attachment left) and "cgcurs" to select on-axis sources (all sources except the bright compact source in the lower part) and then "mfclean" with niter=10,000
2. Subtract on-axis sources model from the original visibility data: use "uvmodel"
3. Construct off-axis source model: use the restored map (attachment left) and "cgcurs" to select off-axis source (the bright compact source in the lower part) and then "mfclean" with niter=10,000
4. Perform amplitude & phase self-calibration on on-axis-model-subtracted visibility data: use "selfcal" with model=off-axis sources model. --> "model gains" obtained
5. Apply "model gains" to the original visibility data: use "gpcopy" with mode=apply, vis=(on-axis-model-subtracted visibility data), out=(original visibility data)
6. Subtract off-axis source model from the original visibility data, which has gone through (5): use "uvmodel"
7. Use the off-axis-model-subtracted visibility data after step (6) as the final "peeled" visbility data: use "invert", "mfclean", "restor" to make image (attachment right).

I wonder if I misunderstood any step/parameter setting according to the paper text... Thanks a lot for help again!

Tai-An

[Mark.Wieringa]

Hi Tai-An,

I think you are getting closer, but I agree your final image still has some problems. Did you use uvaver to apply the selfcal gains before uvmodel? Otherwise they get ignored (a warning message appears).

You could try the following, rather long, procedure:
1. Image/mfclean the whole field,
2. Selfcal with full field model and image/mfclean again - you can use nfbin=2 (or 8) in selfcal
3. Use cgcurs to create a small box region around the off-axis source, use immask to 'mask' it out in the clean model
4. Use uvaver to apply the selfcal correction to the uvdata (output uv2) then use uvmodel to subtract the masked model (be sure to use the mfs option, output uv3)
5. Image/mfclean/selfcal uv3, using the offset box region for cleaning;
6. uvaver to apply the offset selfcal corrections (output uv4), then uvmodel to subtract the offset source (output uv5) - make an image to see if the source is gone now, if not try more selfcal iterations
7. Use gpedit, options=invert to invert the gain solutions determined in step 5, then gpcopy then to uv5 and apply them with uvaver (output uv6)
8. Use uvmodel to add the masked model from step 3 back in (output uv7)
9. Image/mfclean/selfcal the uv7 data

If you find you can't get rid of the offset source in step 6, there may be other problems - bad data (interference?) may limit the dynamic range you can reach.

Cheers,

Mark

[taianmoon]

Hi Mark,

Thanks for the info. I'm tring now but seems not working...

In step (3), the input (in) of "immask" should be the full-field model or the restored map from step (2)? Because it seems that I need to subtract the masked model in step (4) but "immask" seems need to input an image..

In step (5), why do we use the offset box region for cleaning, because uv3 seems to have the visibility data in which the offset box is excluded, so should we clean the whole field except the offset box here? Also, is it needed to image/mfclean again after selfcal here?

In step (6), what is the input visibility data of "uvaver" (vis)? Now I'm using uv3 following the previous step, but it might be uv2, which contains the whole field, because uvmodel is then used to exclude the offset source...

Thanks again!

Tai-An

[Mark.Wieringa]

Hi Tai-An,

- the input of immask should be the full field model. Clean models are images too in Miriad, they just have clean components in them, so they are hard to display.

- uv3 is uv data with all components from outside the offset box subtracted, so when you image it, you should see the offset source and some low level residuals outside the box.
You generally run selfcal more than once - until you see no more improvement, sometimes with shorter time intervals for later iterations. After the last one you could skip the imaging step if you're doing this from a script and are not looking at the images.

- In step 6 the input for uvaver is uv3, it has the selfcal solution for the offset source we want to remove from the data.

I agree the process is rather complicated and I may have overlooked something so it would be best to check at each step if things make sense.

Cheers,

Mark

[taianmoon]

Hi Mark,

Ah...I seem to misunderstand what "masked model" means.

Now it seems that the selfcal process in step 5 is problematic. Attached is the restored map after 2 selfcal's on uv3. Presumably the offset source should be cleaned and gone, but it's not. Below is my parameter setting of selfcal:

Task: selfcal
vis = uv3
select =
model = uv3_2.imodel (model established in the previous mfclean)
clip =
interval =
nfbin = 2
options = amplitude, mfs
minants =
refant =
flux =
offset =
line =

And the restored map after the first and second selfcal look similar, so I'm doubting that some setting might be wrong in selfcal?

Thanks again!

Tai-An

[Mark.Wieringa]

Hi Tai-An,

It's hard to tell from the image what is going wrong.
A few things to check:
- The clean cutoff, if this is set too low, you may pick up too much of the calibration error sidelobes. Check the model after clean to see if most of the components are at the peak of the strong source.
- In the same way, you could set a small clean box around the peak of the source, avoiding the sidelobes, and run selfcal with a clip level just above the highest sidelobe. If the image improves you can lower the clip level in the following iteration
- Try phase only for the first selfcal, then add amplitude in the second iteration if things improved.
- You could shorten the interval to 1 minute or less, look at the phase rms selfcal prints out: if it starts going up (>30 or 40 degrees) your interval is too short (data too noisy)

If none of this helps, do another check for bad data / telescope problems, or let me know the project and observing dates and I can try to have a closer look at the data (if it's not freely available on ATOA yet you'll have to send me an access link).

Cheers,

Mark

[Mark.Wieringa]

Hi Tai-An,

I had a look at the data and tried imaging and selfcal. I arrived at close to the same image you did.
The strong source to the south lies close to the edge of the primary beam, this causes two problems:
- the source varies in flux during the observation due to the asymmetry of the beam: this can only be fixed with peeling
- the source has a very steep and curved spectrum: this gives problems with mfclean, since it can only solve for a simple linear slope

You can see both effects in the image: the rings are due to mfclean failing to match the spectrum and the asymmetric 'spoke' pattern is due to the beam variation.

In your case it looks like the first effect is worse, so I tried splitting the 2 GHz band into 3 sub-bands for imaging. This improved the image greatly, see below, and it may be good enough for your science in the centre of the field.
I tried peeling the sub-bands, but the results were not a great improvement. In fact the 'strong' source was rather weak at all but the lowest frequency and selfcal didn't really have enough S/N to do a good job.

peeled, subband1, combined subbands

Screen Shot 2016-11-30 at 15.58.47.png (1.74 MiB) Viewed 3578 times

The images show the lowest freq sub band in panel 1, after peeling in panel 2 and all sub-bands added together (without fixing beam sizes...)

Cheers,

Mark

[taianmoon]

Hi Mark,

Thanks a lot for this and it sounds like good news. I'm trying 3 sub-bands now, but it turns out that the map seems a bit noisy, as the attached figure shows, which is the lowest-freq sub-band. I'm wondering if another weighting or mfclean cutoff is needed (currently I'm using robust=0.5 and mfclean cutoff=4e-5). Or clean is needed instead of mfclean?

Also, how to add the sub-bands together (add models/restored maps/uv data)? I think uvmodel might not do the job... And presumably the field of view changes for each sub-band so how can we achieve that? (If we simply add the respective pixels together the positions just don't match...)

Thanks again!

Tai-An