why stack?

[alacant]

Hi

I'm struggling to understand why a stack of NEF snaps is better than any of the individual components. I've cracked why a jpg is worse and have read a little about signal to noise ratios, beyond which one seems to require an advanced physics qualification to get any further! I wonder if there's a one or two liner which clarifies it?

TIA 

[Cornelius Varley]

One single image will contain x amount of signal (the data) and y amount of random noise, this is called the signal to noise ratio. Stacking a number of subs together will produce an image with a better signal to noise ratio (more data, less random noise).

[glennbech]

Since noise is both "positive" and "negative" it cancels itself out when you have many frames. 

[dph1nm]

There is nothing magic about stacking. Think of it simply as a way of getting a longer exposure.

Nigel

[StuartJPP]

Signal = Constant (positive only), Noise = Random (positive and negative). 

If you average a large enough collection of both positive and negative random numbers the net result is zero, which leaves only the signal (the photons from distant bodies).

[alacant]

Hi

Thanks everyone. So, 'large enough collection of positive and negative randome numbers': more photos to stack, no? What about light pollution? That's always positive and constant no?

 

[Pompey Monkey]

37 minutes ago, alacant said:

Hi

Thanks everyone. So, 'large enough collection of positive and negative randome numbers': more photos to stack, no? What about light pollution? That's always positive and constant no?

 

Noise being "positive and negative" is an example of "a lie told to children" : Wikipedia

Noise in AP is all positive, but it varies about a mean value with a certain distribution, depending on what the noise source is (LP, dark current, bias, cosmic rays etc.). The more samples, i.e. exposures captured, the more the average of the noise tends towards the mean value - it's "multiplicative". A smaller variation about the mean value allows mean to be subtracted from the stacked image, with a lower residual variation (noise). At the same time, the signal (stars etc) is "additive" and increases with the number of exposures and consequently the signal to noise ratio gets better*.

As for LP, this is just like other noise, in that it has a variation about a mean value. The constant part is easy to remove from the image, but, unfortunately, the noise component is proportional to the amount of LP - It is when the magnitude of the LP noise exceeds the magnitude of the object you are trying to image that you reached your limit.

Stick with it and you'll understand it

 

* There is also a noise component for the signal, but I'm not going there at breakfast time!

[StuartJPP]

1 hour ago, Pompey Monkey said:

Noise being "positive and negative" is an example of "a lie told to children" : Wikipedia

Noise in AP is all positive, but it varies about a mean value with a certain distribution, depending on what the noise source is (LP, dark current, bias, cosmic rays etc.). The more samples, i.e. exposures captured, the more the average of the noise tends towards the mean value - it's "multiplicative". A smaller variation about the mean value allows mean to be subtracted from the stacked image, with a lower residual variation (noise). At the same time, the signal (stars etc) is "additive" and increases with the number of exposures and consequently the signal to noise ratio gets better*.

As for LP, this is just like other noise, in that it has a variation about a mean value. The constant part is easy to remove from the image, but, unfortunately, the noise component is proportional to the amount of LP - It is when the magnitude of the LP noise exceeds the magnitude of the object you are trying to image that you reached your limit.

Stick with it and you'll understand it

 

* There is also a noise component for the signal, but I'm not going there at breakfast time!

 

Yes and the OP wanted it in"simple" terms...

 

[Physicist13]

The random noise reduces as square root of number of images stacked, and the "signal" part goes as N , so overall signal to noise goes as root N .  Basically a way of making say 10x1min exp=1x10min. Before anyone cherps up...read noise reduces this gain in SNR, which is why its important to get subs out of read noise if posisble. 

[John78]

If you look at the why combine section on dss website you can see pictures of what stacking does for you...

http://deepskystacker.free.fr/english/index.html

Stacking doesn't extract something from nothing you do need signal to be more than the noise so if you can't make out the target in a single frame, stacking them isn't going to help.

[sagramore]

I can also add (from my science & datacollection background) that generally there is rule of thumb about improving your signal to noise ratio:

• In order to double your signal to noise ratio you require FOUR times as much data collection (there is a "squared" term in the equations)

In AP terms that means that to half the amount of noise you have you need to take four times as many subs. It's a good thing to keep in mind when you are trying to decide if it's worth staying out for another 30 mins of subs. If you only took 30 minutes already then yes it is (your SNR improves by ~40%) but if you already took 3 hours of subs then no it probably isn't (SNR improvement by only ~8%).

EDIT: Just saw that this was mentioned by Physicist13 - not trying to steal your thunder! Will leave mine with the examples though

 

47 minutes ago, StuartJPP said:

 

Yes and the OP wanted it in"simple" terms...

 

I think this is an unfair comment. I would argue that Pompey Monkey's response was relatively simple and well written without blurring the truth.

[Martin Meredith]

2 hours ago, John78 said:

If you look at the why combine section on dss website you can see pictures of what stacking does for you...

http://deepskystacker.free.fr/english/index.html

Stacking doesn't extract something from nothing you do need signal to be more than the noise so if you can't make out the target in a single frame, stacking them isn't going to help.

It isn't necessary to be able to see the target in a single sub, actually. As Nigel says above, stacking works for the same reason that longer exposures work. And just as you wouldn't expect to see a faint target in a 1s exposure but might see it in say 100s, the same goes for stacking. The target might not appear in any single sub yet still appear in the stacked result.

He's an animation of stacking based on 10s subs. In the initial sub few (if any) of the 32 faint (mag 18+) galaxies in the galaxy group (Shakhbazian 10 in Bootes) at the centre are visible. By the time 64 are stacked they are all visible (due to the longer exposure!). 

cheers

Martin

 

 

[John78]

2 minutes ago, Martin Meredith said:

It isn't necessary to be able to see the target in a single sub, actually. As Nigel says above, stacking works for the same reason that longer exposures work. And just as you wouldn't expect to see a faint target in a 1s exposure but might see it in say 100s, the same goes for stacking. The target might not appear in any single sub yet still appear in the stacked result.

He's an animation of stacking based on 10s subs. In the initial sub few (if any) of the 32 faint (mag 18+) galaxies in the galaxy group (Shakhbazian 10 in Bootes) at the centre are visible. By the time 64 are stacked they are all visible (due to the longer exposure!). 

cheers

Martin

 

 

I agree in part - but - 100 x 1sec != 1 x 100sec if the number of photons is too low to excite the detector above the noise level - in your example image you can see everything in the 2x its just above the noise level.  If I shoot at Andromeda with my DSLR using 100 stacked 1 secs I get nothing at all, if I shoot 1 x 100 sec its obvious, live bulb on olympus' is a great feature for this.

[Martin Meredith]

What is required is for the target photon plus noise to be above 0 when rounded to the nearest integer. The plus noise bit is important. Here noise is helping us to ensure that the target photon is not quantised away to zero on every sub. The SNR in a given frame can be really low (far less than 1) and stacking (=long exposure) will still bring out the target eventually.

This article explains it in more detail.

cheers

Martin

 

[sagramore]

This direct quote from the DeepSkyStacker help page also sheds some light (no pun intended!!) on the debate between 100 x 1s vs 1 x 100s:

Quote

Are 100 x 1 minute and 10x10 minutes giving the same result?
Yes when considering the SNR but definitely No when considering the final result.
the difference between a 10 minutes exposure and a 1 minute exposure is that the SNR in the 10 minutes exposure is 3.16 higher than in 1 minute exposure.

Thus you will get the same SNR if you combine 10 light frames of 10 minutes or 100 light frames of 1 minute. However you will probably not have the same signal (the interesting part). Simply put you will only get a signal if your exposure is long enough to catch some photons on most of the light frames so that the signal is not considered as noise.

For example for a very faint nebula you might get a few photons every 10 minutes. If you are using 10 minutes exposures, you will have captured photons on each of your light frames and when combined the signal will be strong. 
If you are using 1 minute exposures you will capture photons only for some of your light frames and when combined the photons will be considered as noise since they are not in most of the light frames.

 

[dph1nm]

4 hours ago, Physicist13 said:

The random noise reduces as square root of number of images stacked, and the "signal" part goes as N , so overall signal to noise goes as root N .

To be fair, this is only if the exposures are the same length. There is nothing stopping you stacking exposures of different lengths. That's why I prefer to describe the improvement in terms of total exposure.

NigelM

[dph1nm]

15 minutes ago, sagramore said:

If you are using 1 minute exposures you will capture photons only for some of your light frames and when combined the photons will be considered as noise since they are not in most of the light frames.

Unfortunately for the author of DSS, this is nonsense. Even if you do some sort of sigma or median clipping  (which DSS recommends) object photons are very unlikely to be considered as noise.

NigelM

[alacant]

Phew! OK. Yeah. Thanks for the gif. That really hits home, especially the 64 exposures. A picture is worth... Dunno though. I'm manual only here so I'd have to sit through 64 lots of 10s. But hey, why not?

I really do think I'm getting somewhere with this. So much better to understand that just do. Thanks a million everyone.