Stacking, law of diminishing returns?

[SilverAstro]

I am a bit puzzled about what is going on with stacking very many subs to gain more signal (brightness, depth of magnitude)

I follow the bit about random noise evens out but coherent signal builds up, however in my humble little experiments with camera on tripod
( https://stargazerslounge.com/topic/286191-pointshoot-vesta/ )
I see very little or no improvement beyond about 10 subs. Yet I see incredible DSO images on the forum that are made with great numbers of subs ?help!

For some reason DSS nolonger stacks my un-tracked frames (it used to ! )
 so I am manually aligning my Vesta frames in Gimp and stacking them with Opacity varying as (100/n)% where n is the layer number, to give each sub equal weight in the final. Am I wrong in that?

Has anyone got (or seen on the interweb) an example of say 10, 20, 50, 100 subs etc or similar ?  and the resulting progression ?? I have seen tutorials where a single is compared with a several stack but not found anything yet on any law of diminishing returns,
thanks.

[furrysocks2]

Any help to you? http://www.narrowbandimaging.com/incoming/combining_images.pdf

[SilverAstro]

9 minutes ago, furrysocks2 said:

Any help to you? http://www.narrowbandimaging.com/incoming/combining_images.pdf

Thanks, that looks good, will study in a minute,

I came back to add that it looks like I may have to do 100more (or someting silly like that!)  over my 10 to see a significant improvement ( I found this interesting post on CN :  http://www.cloudynights.com/topic/542879-law-of-diminishing-returns-project/#entry7314831 )

but since my camera needs 3 button presses per shot I think I am on a goosechase !

perhaps I must think about a Canon with auto control if I am to persue this imaging lark

 

[Martin Meredith]

Here's an example I made from 1-64 in multiples of 2

There is no question that stacking continues to help. What matters for SNR (leaving aside read noise) is the total number of photons you collect. There is a law of diminishing returns but it is actually more gradual than is sometimes stated.

cheers

Martin

 

 

 

[Tiki]

Gaussian noise cancels out quite nicely with more subs. Less well behaved noise though doesn't necessarily cancel at all and can even have the opposite effect.

 

[SilverAstro]

18 minutes ago, Martin Meredith said:

Here's an example I made from 1-64 in multiples of 2

Thanks Martin, that is a brill. example ( downloaded for future ref! ); and that topic is yet another impressive example of your art. It was thinking about your recent Arp104 that set me to revisiting this stacking thought, but I had a brain failure and forgot about exponentials and squares !   

[zicklurky]

There's a very good video in this thread: 

Around about 7 minutes Craig talks about stacking more images and the diminishing returns. But the full video is well worth a watch  

[abhoriel]

The signal to noise ratio increases proportionally to the square root of the number of lights. So yes, in that regard, there is diminishing returns in that each extra light that you take has less of a positive impact than the light you took before. But extra lights still always equates to better signal to noise ratio!

Taking extra subs won't really get you more detail per se, but because the stacked image has less noise, it lets you perform a more aggressive stretch on it during processing.

I'd recommend reading this page (especially the heading "How many frames?"): http://deepskystacker.free.fr/english/theory.htm

 

 

 

[Gina]

I have noticed a considerable difference between 200 odd subs and 400 odd, with 60s exposures on my ASI1600MM-Cool camera, cooled to -30C and at high gain, on faint DSOs.

[sharkmelley]

At he beginning of Chapter 6 (Image Integration) of Warren Keller's book "Inside PixInsight" he states:

Quote

There is, however, a law of diminishing returns.  Statistically the limit from which you can obtain significant benefit is thirty images.

Now I don't know where this idea comes from but luckily he is demonstrably wrong.  Otherwise I would be wasting my time taking 500x30sec exposures for a typical image and Gina (previous comment) would be wasting her time taking 200-400 exposures.  Also, planetary imagers would be wasting their time taking a few thousand frames of video.

Statistically, each successive doubling of the number of exposures will increase the signal-to-noise ratio by a factor of 1.4 i.e. the square root of 2.  So yes, there is a law of diminishing returns since each successive exposure has less and less impact on the final stacked image.  However, I think the limit is actually measured in hours, not on the number of exposures and also it depends on individual circumstance.  Whereas one person may think that it is not worth the effort to spend 1 extra hour of imaging beyond the first hour, another imager may think it is not worth spending another 10 hours beyond the first 10 hours.  For a deep field Hubble image they might decide another 100 hours beyond the first 100 hours is not worth the extra effort.  It all depends on goals, patience and maybe on the availability of clear nights.  The limit for one target may be completely different to another.  For instance, I might think that a star cluster is adequately imaged after 1 hour whereas a dust cloud may take 10 hours.

To my mind, it makes no sense to propose a limit in terms of number of exposures and especially it makes no sense to propose a limit as low as 30.

Mark

 

 

[ollypenrice]

I think that there are a number of factors in play. Firstly the read noise of the camera comes into it: from personal experience there is not the slightest possibility that 100x1 minute gives an equivalment to 10x10 minutes but that's with a 'traditional' CCD. New technology cameras with low read noise chips do benefit from multiple short subs and I'm sure Mark is right with regard to these cameras. He's proved as much with his images.

Although I don't have a mathematical justification for this, I suspect that Warren Keller won't be far out if talking about deep sets of old style CCD subs. If I took a set of 30x15 minute CCD subs from one of my cameras I would expect that to be about that as far as far as quality goes. (I'll also admit that this is not something I'm particularly interested in putting to the test!) Because our guests frequently want to image things we've imaged before, I sometimes have the chance to combine pretty decent datasets into more than decent ones. I suppose I've been in the 'thirty x 15 minutes' ballpark on occasion and there is a benefit. But of what kind?

If you just did a log stretch of 15x15 and compared it with the same log stretch of 30x15 the difference would be trivial and a bit of judicious NR on the faint parts would have the shorter set on a par with the longer. However, the deep set would obviously allow for far more aggressive processing, so faint detail can be lifted into view and contrast enhancing on various scales can make new structures visible.

Then there's the funny side, humans being funny creatures: I posted one of our combined dataset images on the French forum and was told off for havng used too much noise reduction. As you can guess, I hadn't used any at all.

Olly