Question regarding number of subs

[Paul2019]

Hi All,

Just getting my self back into the swing of things and getting back onto some widefield with my DSLR (350D) and a tripod, Using the kit lens at 50mm F/5.6. I can manage about 8secs before I get trails. 

Obviously I am aware that longer subs are better than shorter ones and that a larger number of shorter ones wont provide the same image as the same amount of exposure time in larger subs.

But is there a limit to the number of subs I can stack? I.e will 200 subs be too many? How about 2000? Where will I hit the limit of what is practical?

I am planning on collecting some data around Orion and although the nebula is visible in 8secs, I want to try and get a bit more from it.

I am awaiting this whole mess to be over so I can put and order in for a HEQ5 but for now it will have to do. But heres 3 minutes unprocessed from the other night,

Thanks in advance

Edited April 4, 2020 by Paul2019

[discardedastro]

Longer subs are better than shorter at some point - you need to have "enough" SNR in your subs, or you'll just be multiplying too much noise in each frame. But there's no real hard limit on the number of subs you can combine.

In practice, more subs means more memory and storage space, and your computation will be more difficult/complex/time-consuming. But shorter subs will of course be less prone to tracking errors, etc.

I've done 200 5s frames on a Z6 with a 24-70mm f/4.5 and stacking is fine in PI (this just on a tripod, no motion). I'd definitely be focused on getting tracking of some form though - you'll need longer exposures to get any nebulosity or detail preserved. Maybe a homemade tracker or a Star Adventurer etc?

[vlaiv]

29 minutes ago, Paul2019 said:

Obviously I am aware that longer subs are better than shorter ones and that a larger number of shorter ones wont provide the same image as the same amount of exposure time in larger subs.

Do you know why is that? It is solely down to read noise. If read noise was 0 - then there would be no difference. 100 subs each 1s long would give exactly the same image as 1 sub 100 seconds long (as long as total imaging time is the same).

In fact, there is very small difference between two approaches when read noise is non zero if read noise is not dominant noise source. If you have heat noise (dark current noise) or light pollution noise that swamps read noise - you'll be hard pressed to see the difference between the two.

All other noise sources depend on time, and only read noise is per exposure - for this reason there is point of diminishing returns (which depends on both read noise and other noise sources) - after which there simply are no gains in going longer exposure. If your read noise is low and you shoot in light pollution on a hot summers night (and your sensor on camera gets hotter than usual) - there is a chance that you won't gain anything by going longer than those 8 seconds.

On the other hand, if you use cooled camera, narrow band filters and you are shooting in dark location, high resolution image - then it's worth making single sub be half an hour or more.

31 minutes ago, Paul2019 said:

But is there a limit to the number of subs I can stack? I.e will 200 subs be too many? How about 2000? Where will I hit the limit of what is practical?

In principle there is no limit to number of frames you can stack. In practice, yes there is limit in terms of memory / size of each sub, and numeric precision of data format used. With 32bit floating point format and modern day cameras you'll run out of precision after stacking about couple of thousand of subs. 32bit integer format is better in this respect as you would be able to stack millions of subs before you run out of precision, but 32bit integer lacks dynamic range of 32bit float precision format.

In any case, as long as you stay in hundreds of subs (or maybe one to two thousand) and use 32bit precision (which ever version) - you will be fine.

Another important point to add to this is - you can use arbitrarily short subs and stack them and you will get result, but problem with very short subs is that you won't have alignment stars visible in your subs (regardless of that - if you had a means of aligning the subs - they would still produce image even if any single sub has no photons captured in particular pixel).

40 minutes ago, Paul2019 said:

I am planning on collecting some data around Orion and although the nebula is visible in 8secs, I want to try and get a bit more from it.

I am awaiting this whole mess to be over so I can put and order in for a HEQ5 but for now it will have to do. But heres 3 minutes unprocessed from the other night,

Very important aspect of astrophotograpy is processing. Try processing that image of yours and you will be surprised how much difference 3 minutes makes compared to 8 seconds.

Even just a slight tweak on that 8 bit version you posted changes image quite a bit:

[Paul2019]

31 minutes ago, vlaiv said:

Do you know why is that? It is solely down to read noise. If read noise was 0 - then there would be no difference. 100 subs each 1s long would give exactly the same image as 1 sub 100 seconds long (as long as total imaging time is the same).

In fact, there is very small difference between two approaches when read noise is non zero if read noise is not dominant noise source. If you have heat noise (dark current noise) or light pollution noise that swamps read noise - you'll be hard pressed to see the difference between the two.

All other noise sources depend on time, and only read noise is per exposure - for this reason there is point of diminishing returns (which depends on both read noise and other noise sources) - after which there simply are no gains in going longer exposure. If your read noise is low and you shoot in light pollution on a hot summers night (and your sensor on camera gets hotter than usual) - there is a chance that you won't gain anything by going longer than those 8 seconds.

On the other hand, if you use cooled camera, narrow band filters and you are shooting in dark location, high resolution image - then it's worth making single sub be half an hour or more.

In principle there is no limit to number of frames you can stack. In practice, yes there is limit in terms of memory / size of each sub, and numeric precision of data format used. With 32bit floating point format and modern day cameras you'll run out of precision after stacking about couple of thousand of subs. 32bit integer format is better in this respect as you would be able to stack millions of subs before you run out of precision, but 32bit integer lacks dynamic range of 32bit float precision format.

In any case, as long as you stay in hundreds of subs (or maybe one to two thousand) and use 32bit precision (which ever version) - you will be fine.

Another important point to add to this is - you can use arbitrarily short subs and stack them and you will get result, but problem with very short subs is that you won't have alignment stars visible in your subs (regardless of that - if you had a means of aligning the subs - they would still produce image even if any single sub has no photons captured in particular pixel).

Very important aspect of astrophotograpy is processing. Try processing that image of yours and you will be surprised how much difference 3 minutes makes compared to 8 seconds.

Even just a slight tweak on that 8 bit version you posted changes image quite a bit:

Wow, thanks for the in depth reply. 

So my understanding is that the noise is going to limit my image and there will become a point where noise is too great?

So does that mean with each additional light the noise will increase? or have I got my wires crossed somewhere? Its all so confusing.

I was going to aim for around 3 hours or 1350 Lights @ 8 secs, I think it will be really interesting to compare short exposures on a short focal length compared to longer as I dont really know if there will be any difference at all. It would be amazing to image barnards loop but I feel thats not going to happen with 8 second subs.

Another question I had was would I see a significant increase in quality if I were to buy a fixed focal length lens? I was debating a 50mm F/2.8 but obviously my current priority is to get some form of tracking (Hopefully an HEQ5, but I'm torn between this and the NEQ6)

Edited April 4, 2020 by Paul2019

[vlaiv]

1 hour ago, Paul2019 said:

Wow, thanks for the in depth reply.  So my understanding is that the noise is going to limit my image and there will become a point where noise is too great? So does that mean with each additional light the noise will increase? or have I got my wires crossed somewhere? Its all so confusing

Here is very bare explanation of all key points:

Noise is random unwanted signal.

Next to that there is unwanted non random signal.

Finally there is signal that we want - target signal.

There are 4 main types of noise (random components) in AP and associated signal (either wanted or unwanted):

- read noise - it is Gaussian type noise. You have it each time you read out a sub and it is always same intensity/magnitude (that is why we say for example that read noise is 3.5e for particular camera). There is also "read signal" - or more often called bias signal. This is signal, it is unwanted signal (but not random) and it is removed in calibration. You can inspect that signal by taking set of bias subs and creating master bias

- thermal noise - it is Poisson type of noise and it is related to dark current signal. Their relationship is: intensity of thermal noise is equal to square root of intensity of dark current signal. Dark signal builds up linearly with time (for given temperature, if we change temperature things get messy - this is why it is important to shoot darks at same temperature as lights). Dark signal is removed in calibration.

- light pollution noise - it is Poisson type of noise and it is related to sky signal. Again their relationship is: LP noise is square root of intensity of Sky signal (same as with dark current and thermal noise). Sky signal increases linearly with time (same as dark current signal). Sky background is removed in processing (either setting black point or using advanced things like background removal if one has gradients in their images).

- Target noise - it is again Poisson type noise and it is related to target signal. Again - target noise is square root of target signal (like previous two), Again target signal increases linearly with time (as previous two).

Target signal is good signal and we don't want to remove it but rather capture it.

Finally - noise adds in "quadrature" - or linearly independent vectors - or square root of sum of squares.

One more thing - "stacking" (adding or averaging) of noise will produce value that is square root of number of subs larger or smaller (depending on whether we are adding or averaging) than individual noise level (this can be shown by above rule on noise addition)

If you put all of this together:

All signal except bias signal (not important as is removed in calibration) raises linearly with time - so does our target signal.

All noise except read noise is square root of one of above signals (thermal, sky or target) - so it raises like square root (less than linear).

Read noise is per sub - but when added it raises like square root of number of subs (rule about stacking) - while signal adds linearly.

Overall - signal will always "beat" noise given enough time - regardless how small signal is and how large noise is - because signal raises linearly and noise raises like square root

This is graph of square root function and I've added linear function (straight line).

At the beginning - root can be larger than linear value but there is a point where linear shoots off and beats square root and can be arbitrarily larger. This means that we can have any SNR we want - if we put in enough exposure time (number of subs) for any given signal strength and noise levels.