To drizzle, or not to drizzle: that is the question.

[Slawomir]

Hi all,

Since it is cloudy tonight, I thought of experimenting with the little data I have. It seems that drizzle integration works nicely even with 12 subs, as long as signal in each sub is not too weak, but more subs is always better of course. May drizzle all three channels for the current project, once I collect enough data.

Suavi

[DJScotty]

Hmmmmmm.
That's an appreciable difference

[Atmos]

According to NASA (JPL), 9 subs is enough statistically for drizzle to be effective :)
Drizzling is good though, even if it doesn't make your detail appear it does do a better job with noise sampling and rejection plus making deconvolution more effective.

[sil]

doesnt make any improvement when i try with PixInsight, nothing like your result, maybe cos I use 200-500 dslr frames usually? scaling my integration file gives me same result as drizzle to same scale.

eg I usually integrate 3 or 5 frames to make a cleaner temp integration, then crop this around my target and register/integrate/drizzle with this as my target but I got no improvement between the two final integration images I could notice, no extra detail, no smaller stars appearing, not noise difference, But I did notice some bright stars obtained verticle artifacts (not visible in frames or anywhere else, not lense artifacts or star streaks but look like a 1 pixel line 20pixels long drawn in difference mode, definitely a processing artifact. Admittetdly its been a while since I tested since drizzle integration was added to PI.

[gregbradley]

Quite a dramatic improvement. What software did you use for that? What settings did you apply?

Greg.

[sil]

Thats PixInsight in the screenshots.

[Slawomir]

Yes, I only use PixInsight for all data processing these days. Standard settings for x2 drizzle integration. Scaling does give similar results to drizzle, but drizzle is more sophisticated (and more CPU involved) than scaling and because of this can recover extra detail from dithered subs and when optics and seeing are both decent, guiding is good and of course when one is undersampling.

[multiweb]

+1 Big fan of drizzling. :thumbsup:

[Slawomir]

Last night added more OIII data and now have 7 hours in total (3nm filter at f/4.5 and 1.6 arcsec pp). To my eye, drizzle has done a better job at pixel rejection than a few algorithms I tried in PI. I have not been dithering enough, only once every 3 or so subs.

Throughout the night, as the telescope was moving away from the CBD, mean ADU value dropped by 600 between the first and last sub. That's huge! Getting decent signal in OIII is a real challenge from where I live.

The image on the most right is a single 15 minute sub. The next one up is integration of 28 15-minute subs without rejection, then with rejection and the most left one is drizzled data. Looks okay to me.

[Atmos]

Drizzling also helps with better noise sampling. I find I get smoother results from drizzling and then 50% down sampling than just simple stacking.

[Octane]

Yep, I drizzle everything now. At 530mm at f/5, you end up with crunchy stars otherwise.

It takes a while, especially when you work with the number of sub-exposures that I often take for my subjects, but, the end result is well worth the time.

H

[alpal]

Hi Everyone,
The question remains about oversampling & undersampling.

I have a 10" f4 scope using a KAF8300 sensor.
It has 5.4 um pixels.

My arc-sec/pixel = 1.11
http://www.wilmslowastro.com/software/formulae.htm#ARCSEC_PIXEL

Therefore if say - my seeing conditions are very good
I will get 2 arc seconds FWHM.

Using the Nyquist sampling theorem you need double the
arc-sec/pixel ratio - follows that to sample 2 arc seconds you need a minimum of 1 arc second per pixel.


Therefore I am slightly under sampled at 1.11.
In theory I can gain something from using drizzle.


I think that the above theory is wrong & you really need 3 x to be correctly sampled
as the pixels are square & not round like stars:
https://astronomy.tools/calculators/ccd_suitability

I certainly noticed a difference - an improvement - using 3 x drizzle
on my last image.
I got that little bit more detail when zooming in on tiny structures.

I was always interested if Mike Sidonio could gain anything by using drizzle.
I checked & he's oversampled so in theory he can't but
I would like to see the results - if only I could twist his arm. :)


cheers
Allan

[troypiggo]

Allan, instead of having questions unanswered and getting lost in the maths, why don't you follow Slawomir's example and try processing some images with and without drizzle, and see if it works for your setup/combo? Who cares about definitions, it's the result that counts, no?

[Slawomir]

I have read somewhere that sampling at 1/3 of your usual FWHM is optimal for getting the most detail from data, so I believe your observations are spot on Allan. Drizzle x3 requires more subs/better data than x2, but if it works then why not use it :) I like Troy's suggestion of directly comparing different methods visually and by measuring noise/SNR.

As for twisting Mike's arm - good luck with that! :lol:

[alpal]

I have posted 2 pictures on flickr.
The first one is the full frame of Eta:
https://www.flickr.com/photos/24719437@N03/31782635104/in/photostream

The 2nd one is a crop using 3 x drizzle:
https://www.flickr.com/photos/24719437@N03/32759728116/in/photostream

I reckon there is more detail in the close up 2nd crop using drizzle.

[alpal]

Yes - I think the article I pointed to below is a rebuttal
of Nyquist's sampling theorem that we've all been using.
3 x is considered better than twice when considering sampling.
As for twisting Mikes arm - that was a joke -
I'm not a strongman.:)

cheers
Allan

[troypiggo]

Too hard to compare apples and oranges (processed images, processed differently, different crops). Why not be a little more scientific about it, just process lum, and apply identical processing to a drizzled vs undrizzled. Compare same crop at 100%.

[RobF]

Really well presented investigation Slawomir, and agree the extra work for drizzle appears worth it for your set up.

Rick's posts #3 and 6 in this thread changed my thinking on drizzle:
http://www.iceinspace.com.au/forum/showthread.php?t=138224


At approx 2 arcsecs/px on my current set up, I then took the time to compare images more carefully in PI (based on some quality data from Astrofest) and realised I was almost always happy to be (quote) "trading off resolution against SNR".


An example zoomable with/without pic for my FSQ/QHY9 rig here (but not as well prepared
perhaps as your examples :)):
https://goo.gl/photos/QgRWtAKxtqBFbnuM9

What I find fascinating is that the background noise often also appears more pleasing, even though you know you theoretically have less SNR (in addition to smooth stars of course). Believe that is shown in that last example.


(Dammit - Rick covered and described that better too: "Drizzle introduces correlated noise that isn't as visually intrusive.")

[Slawomir]

Nice illustration Rob, stars look nicer in drizzled data for sure.

I'm wondering if purposely putting together a system that reasonably under-samples and then drizzling data is the ultimate solution for amateur astrophotographer.

As I understand it, at the same f-ratio (and same QE+read noise), undersampling system will be going deeper in the same time as opposed to a well-sampling system (say 2"pp at f/6 vs 1"pp at f/6), thus in the end one will have a higher SNR and a similar detail with undersampled and drizzled data comparing to a well-sampled data, given equal time of exposure. Or am I missing something here?

[RickS]

There's nothing wrong with Nyquist sampling at 2x but the conventional use (sampling a single analog signal) isn't what we're doing.

Stan Moore makes a pretty good case that you should sample at about 3.5x for imaging: http://www.stanmooreastro.com/pixel_size.htm



That's what we're doing with the C300 at SRO (image scale 1.26 arcsec/pixel with seeing often down to 1 arcsec) and what I had in mind when I matched the AP140EDF with 9um pixels (2.4 arcsec/pixel with typical poor SEQ seeing.)

I like this approach because it gives you better resolution than the image scale, a big FOV and images that you can print really big.



Alas, it's not magic, Suavi. The increase in resolution does come at the expected cost in reduced real SNR. It's just that the apparent SNR is better because of the way that Drizzle correlates noise. It effectively does noise reduction. You can get the same effect by applying noise reduction to the data from a well-sampled system.

Cheers,
Rick.

[Slawomir]

Thank you Rick for clarifying that.

So it looks like one day I should invest in a camera with about 6micron pixels to give me just under 2"pp with my telescope at f/6, I think something like KAF16200 would be nice. I was hoping that I can get a wider FOV with my tiny ICX814 at f/4.5, but such fast f-ratio on a short refractor is a challenge to get spot on in terms of alignment and spacing. And on top of that frequent refocus is needed...I much prefer riding at f/6. Time to start slowly putting some dollars aside for a camera upgrade! LOL

[RickS]

That would be a good combination, Suavi. We have a FLI ML16200 at DSW and it looks like a really nice camera. Read noise is only 5e- which is great for a KAF sensor.

[gregbradley]

Yes that's right Suavi. Nyquist theorem for sampling is a minimum of 2. Minimum being the main word there. So practical use is to go for 3X.

Greg.

[gregbradley]

Roland Christen of AP fame always recommends 1 arc sec/pixel for those with average seeing of around 3 arc secs so 2 arc sec is still not quite there. I have noticed this with the various scopes and camera combos I have used over time. When you are close to that 1 arc sec the images look nicer.

I would also add though that full well depth is a very important and underrated aspect of a sensor. Too low and you lose a lot of dynamic range if you overexpose. That's why I tend to keep coming back to my 16803 camera. Its still the best out there for overall performance.

Maybe these Sony CMOS full frame sensors if someone can get their hands on a mono version of the Sony A7r2 sensor it would be pretty sensational and may leave CCDs for dead. At higher ISO this thing's read noise is .58 of an electron. Its backside illuminated etc etc. QE would be very high and 42mp would give super high resolution.

Greg.

[Slawomir]

Sounds good in theory Greg, but with my scope I would need to get even smaller pixels - ICX834 with 3.1 micron pixels would give me exactly 1"pp, but that means even shallower wells. So big fat pixels on a big scope but that also means heavy (end very pricey!) rig, and I need to set up/down for every session. That's why I need to stick with a light scope = drizzling :)

FLI 16200 sounds VERY good, maybe in a few years... :-)

[Atmos]

When it comes down to trying to get perfect arcsec/pixel and trying to be seeing limited is what you do when you're imaging with something larger than a 4" scope :P
Start worrying about it when you get an ASA 10" F/3.6 newt, that way you can use your QSI690 and have a perfect thing going on, albeit a narrow FOV.

[Slawomir]

Something similar to Mike's setup, so it must be very good :thumbsup:

[ErwinL]

Suavi,

from the appearance of the undrizzled images you've shown in your posts, I'd guess that the subs have been adjusted by full pixels prior to stacking.
If my guess is right: Did you ever compare drizzling against sub-pixel alignment?

Erwin

[Slawomir]

Hi Erwin,

To be honest I do not know how exactly PixInsight creates a drizzled image. I just followed a regular path of Star Alignment that also creates drizzle data while aligning all subs, and then Image Integration allows for refining pixel rejection algorithms (it also updates individual drizzle files for each sub), and finally I used Drizzle Integration tool with standard settings to create drizzled masters.

[RickS]

Hi Suavi,

I think that Erwin is asking if you created the original (non-drizzled) images using nearest neighbour interpolation as opposed to one of the other modes that does sub-pixel alignment. Nearest neighbour can give a blocky appearance.

Erwin: I'll let Suavi answer about his data but my experience with undersampled data is that Drizzle still does a much better job of rounding up stars than sub-pixel interpolation without Drizzle.

Cheers,
Rick.

[Slawomir]

Got it, thank you for clarifying that Rick :thumbsup:

The non-drizzled image is a stack of original subs without upscaling, at about 1.6"pp. My system is a bit under-sampling when I use a reducer, thus there are blocky stars in the non-drizzled image. I rarely upscale final images as I am confident that drizzle will yield a superior result to upscaling a final image, but perhaps one day I could actually try to measure if there are any differences.

Suavi

[ErwinL]

Ah, ok, I see things are a bit different in PixInsight (which I never used). What I'm doing regularly and meant in my question is sub-pixel alignment and interpolation without upscaling. The basic calculations are comparable to drizzling but the image size remains unchanged.



Rick: Do you think this is also valid for this 'in-place' drizzling-like interpolation? (Unfortunately, my equipment and the seeing at my location do not allow undersampling, so I can't do real tests.)

Erwin

[RickS]

Hi Erwin, the default interpolation done by PixInsight image registration is sub-pixel interpolation without upscaling. You can also set the interpolation to be nearest neighbour but I doubt that many people do.



With a FSQ-106ED and KAF-16803 sensor I have an image scale of 3.5 arcsec/pixel. Even with sub-pixel registration smaller stars don't cover many pixels and they look blocky. Drizzle x2 is a big improvement in this case.

Cheers,
Rick.