How many people drizzle their data and, if so, why and how?

Another question regarding this topic.
cfa drizzle osc images for better color? 
seems a few people do that. Is there really a benefit?

I don't, because my sampling is good for my current setup. And results have been mixed when drizzling was applied, so.. I don't bother, I don't feel I get a lot of extra detail, and it does take a lot more time and space to do all of it. 

But, I'm working on a rig, that would have under sampling.  (basically, the pixel size is too big to capture fine details)
I will look how the data turns out, but will most likely be drizzling that data to combat the under-sampling. 


And yes, Drizzling requires dithering. it uses the shift of pixels and the image that dithering does, to calculate what data "should" be there and then fill it in. So without dithering, you're not providing the software with enough data to do the drizzle and you might not get good results.
Automatic dithering is best. Or, do it manually with a little manual shift after 4-5 shots, or even some drift in guiding might already be enough.


But, looking at your setup, the 1600MM and the 200PDS, your sampling is good.
So there might not be a "requirement" for you to drizzle. But, others, with experience, might be able to provide better suggestions.

Another question regarding this topic.
cfa drizzle osc images for better color? 
seems a few people do that. Is there really a benefit?

In general is advisable to drizzle CFA images at 1x to have better details but at the expense of sligthtly lower SNR. Natural (i.e. drift) or imposed dithering is obviously a requirement. It is also the case if you're undersampling your seeing by a meaningful factor.

David Koslicki:
@Ian Dixon I actually use APP too! In case it helps, I've found that the type of kernel you pick (topHat, point, square, Gauss) doesn't matter too much, as long as you don't pick point. I've tried varying the droplet size and found that smaller droplets = sharper but more fine scale noise. I haven't checked the theory though to see if this is just particular to my setup, or a general fact.

Thanks!

@Ian Dixon I doubt that drizzle will gain you much when sampling at 0.38"/pixel. Ball parking it: I imagine your tracking would need to be under ~0.4 arc seconds RMS and excellent seeing for you to gain anything from drizzle, and then only if you are imaging something really small that you want to bring out some detail. My rational is that if your tracking accuracy is any worse than this, then the photons emitted from one point source of light are already falling on multiple pixels in each exposure. From what I've looked into, drizzle is for the opposite problem: multiple point sources of light falling on a single pixel.

My usual strategy, however, is to always test theory with data. Since it only costs hard drive space and CPU cycles, it never hurts to try and see what the results are! When I do this, it often helps improve my understanding of the underlying processing algorithms.

@Ruediger From reading the drizzle paper (https://iopscience.iop.org/article/10.1086/338393/pdf), it's more complicated than a simple anti-aliasing algorithm. While that's the most visible feature (smoothing out the jagged edges), since the shift of the image (dither) is being used to infer what smaller pixels (the drizzle "drops") would have measured, you really are dividing the input from a single pixel between several output pixels (see Section 7 of that paper). So you really are teasing out extra information. To see this in action, find a pair of stars/objects that are really close to each other and notice how the resolution changes. Using a different crop of the image I posted above, note how the faint smudge on the lower right of this star is better resolved when drizzled (hence, more than just smoothing edges):
No drizzle

Drizzle


Also, drizzle will definitely not work if images are not dithered or somehow moved from sub to sub. The algorithm would have nothing to work with then.

Happy to share some experience here with 1.000mm FL and an ASI295mc.
This combination is wether an over- nor an undersampling.
I did it in APP, mostly with 2x drizzling and dropplet 0.5.
I found out that it sometimes helps a little bit in resolution but with the cost of noise.
As some others stated here the processing time and file sizes are another disadvantage.
I assume that you just really benefit from drizzling with undersampled data as many others said here before.
You will find nice explanations about pros and cons from Mabula and other guys in the web which I highly recommend.

But if you have the time it's worth it to push some buttons compare the results with your own eyes😉

Good luck
Mike

Hopefully I can clarify a few points with respect to dithering and drizzle, my comments apply to both mono and color CMOS and CCD cameras.

Dithering: a random (ideally) movement of the camera sensor pixels relative to the image target, done between subexposures. For deepsky targets you should be doing this whether or not you drizzle. If you don't dither then you will see background noise artifacts due to pattern noise from your sensor. If you are guiding since image moves very little you will see the pattern noise burned in your final image. If not guiding, where there will be some drift between frames, you need to dither at least occasionally, otherwise you end up with a pattern noise called walking noise, since non-guided drift is mostly in a defined non-random direction. Again, all this is independent of whether you want to drizzle in processing. It is important you dither.

Drizzling: Processing step that basically maps the original coarser subexposure grid with its pixel values to a finer pixel grid that has 4X more pixels (for 2x drizzle), or 9X more pixels (for 3x drizzle). In this way you actually do potentially capture more image detail. The method was specifically developed for the case of undersampling, when your pixels are too large for your plate scale and seeing. See Fruchter and Hook for the math behind this transformation. In a nutshell this trick can recover detail that was lost in each subframe, but can be recovered by adding the contributions of all the shifted subframes. Note for lucky imaging for planetary work images are often oversampled to improve the image resolution and even there drizzle can help increase the detail captured.  So a case can be made for deepsky work with shorter lucky imaging that drizzle should help even if you are oversampled. My own personal experience, non-guided lucky imaging (less than about 20 to 30 seconds), is that 2X drizzle reduces my FWHM even though I am in theory at the sweet spot for sampling without drizzle (plate scale at 0.5 arc-sec, so x 2 to 3 for Nyquist sampling gives me 1 to 1.5 arc-sec, good enough for all but the absolute best seeing). Drizzle is very helpful to bring out fine detail, such as in planetary nebula. In my tests I see no impact on S/N of 2X drizzle, and did not find 3X drizzle improved my resolution over 2x. So I do use drizzle when I can. I use DeepSkyStacker for drizzle. The only reason I don't drizzle is if I am using the full frame of my 50 mb raw images, then DSS crashes on my computer. The issue is the 4X increase in pixels, so a 50 mb image becomes 200mb in the stacking! So for drizzle I either capture a smaller frame or set a smaller frame size in DSS.

In short. Dither. Try with and without drizzle processing on different targets and different nights and compare. Everyones setup and conditions are different, I really encourage you to experiment with it, and learn if and when to use it.

Clear skies and fast CPUs
Rick

I usually drizzle, particularly for smaller galaxies and for the stars that I will replace the narrowband ones with.  Usually I drizzle x2; I have a native 1.41"/pixel so that takes it to about 0.7"/pixel which gives me visible benefits. I also find the drizzled noise easier to deal with as it is finer.  Occasionally I drizzle x3 for tiny stuff - e.g. the waterbug galaxy and friends showed improvement at 3x over 2x :   https://www.astrobin.com/itma6w/?nc=user

The algorithm was designed to recover resolution from a set of dithered, undersampled images. 
I think the example star image shown above by David is already well sampled and not suited for Drizzle processing. 

Dithering works really well when I work with short-focal length setups such as camera lenses.
The images below are from a 50 mm f/1.8 Canon lens with ASI1600MM (effective 15"/pixel resolution). This type of data is significantly undersampled and stars are typically just one or two pixels wide and appear very blocky/square. With 2X Drizzle, the stars appear much more round and you gain a small amount of resolution. Upscaling with the widely-used Lanczos3 algorithm results in dark artefacts around the blocky stars and no improvement in resolution.



Dithered image files are huge and post-processing these files is very taxing on most systems. That's why I typically downsample the 2X Drizzle master file to the original resolution. You get much nicer star shapes without an increase in file size. 

Ruediger:

David Koslicki:
@Ruediger Fair enough: my interpretation of "information" is the Shannon entropy-esque definition. Hence a single (undrizzled) pixel with normalized value 1 will have entropy 0 while four (drizzled) pixels with, say, normalized values {1/2, 1/3, 2/3, 1/4} will have entropy Log[4], hence more information

Hi David,
I am argumenting based on the same definition based on Shannon Theorem: Since the information comes from a predictable algorithm, the  probability=1. Hence information equals zero. You cannot generate more information than contained in the raw data. That would violate the information theory that an information sink contains more information than the source. 

But forgive me if I am wrong, my studies of information theory are 30 years in the past. 🫣
Maybe we can agree on an empiric approach? just try it out and do what looks better 🤔

There is no information created by drizzling, information is only recovered. The information at the sub pixel level is contained in the total set of data, the collection of subexposures that are randomly moved with respect to each other. By simply adding them pixel by pixel we loose the extra information that the random motion produced. In drizzling we have the opportunity to recover this lost information. A too simple analogy would be sieving a sample of two different sized powders--if we use a large sieve so all the particle go through it looks like all the particles are the same size. If we use a smaller sieve we see some go through, some stay behind. We have more information on the sample, but did not create information.