With the ASI1600, more subs is better than few subs, because it's 12-bit. We need to stack a good number of subs in order to recover bits, so using the longest subs possible can actually be a detriment. This is especially true at Gain 0, which has ~1.4e- quantization noise alone, which is pretty significant (more than any other camera I have ever used or processed data from.)
well, you never had a 8bit camera ? Give you an example with a TIS/DMK at low gain. Shows a similar gain readnosie dependence like the ASI but at a level 20-30 times higher. Compared a single exposure of 200 msec with a stack of 2000 taken at 1/10000 msec. Single Frame of a stack looks totally dominated by A/D noise alone and shows dramatic posterization. Result is still dominated by readnoise, but it is of course an extreme example.
stacking extreme comments.jpg
Why do I show this ? Actually I should add some more stacks with slightly more exposure time. What becomes obvious is that there is no discrete step for a minimum exposure time, where everything is bad if you stay below. If one fails to bring the histogram to the "recomended" range, it might take longer for the total stack, but it could still lead to a decent result with slightly more exposure time.
On the other hand, dividing a non-saturated exposure into more subs, will not give any benefit in terms of "more bits = more precision". The upper bits are not "filled". Of course it makes sense as long as a readnoise drops significantly with higher gain which can be applied for the shorter subs.
(again: the discussion is a little bit academic (or useless), since for some people shorter subs are simply essential for other reasons like guiding Errors, seeing and OTA seeing and so on.
The primary issue with the stacked image is not posterization. It is FPN. By far, FPN is the single largest issue with that stacked image. Quantization error will average out with stacking...however with undithered subs, FPN will NOT average out, and it will become more visible the more subs you stack. Your 2000 sub stack is likely fully limited by FPN, which probably kicked in well before the 2000 sub count and is limiting SNR. I would offer that if you dithered those 2000 subs, even sparsely, that once stacked you would no longer see FPN, or at the very least, it would not be nearly as clear and would not limit your SNR as much.
FTR, FPN is not just bands and glows. FPN is fundamentally due to non-uniform response of the pixels to photonic and dark signals. The posterized look in this image is most likely due to the per-pixel FPN that arises due to that non-uniform response. Since no dithering was used, that spatially random pattern, combined with the vertical banding, was effectively "frozen" into the image. You might be surprised what even dithering every 100 frames does for a stack like that, let alone every 10 frames.
Oh, and the notion that "upper bits are not filled". That is very false. Signal combines linearly in a stack, while noise combines logarithmically. When you stack, you are ADDING signal. Let's say that the signal of your subs is half a bit. You stack 2000 subs, you end up with 2000*0.5e-, which is 1000e-! That is a very NON-trivial signal. That is a HUGE signal. Let's say you had a 1e- quantization error:
SNR = (2000*0.5e-)/SQRT(2000 * (0.5e- + 1e-)) = 1000e-/SQRT(2000 * 1.5e-) = 1000e-/SQRT(3000e-) = 1000e-/54.8e- = 18.3:1
That is a good SNR. Assuming you did not have FPN, that would be a perfectly usable SNR. Lets say you had 5% PRNU. Photon transfer theory states that once your signal tops 1/PRNU^2, you are FPN limited. Well, 1/0.05^2 = 400e-, which would mean the signal in your 2000 frame stack was certainly fully limited by FPN. At 4% PRNU, you would be able to get to 625e- before FPN limited you. At 3% PRNU, you would be good until 1111e-, however at 1000e-, you are still going to be largely limited by FPN, certainly more than by any other noise. Given how strong the pattern is in the image above, I would say the PRNU is probably higher than 3%.
Oh, and one final thing. Comparing a single 200 MILLIsecond sub to a stack of 2000 0.1 MILIsecond subs is FAR from what I've been talking about. I have been talking about exposures that are deep enough to properly swamp read noise. A single 200ms sub is about enough to capture a slightly blurred speckle pattern of a star, and nowhere remotely close enough to capture any faint DSO details. Let alone a 0.1ms sub. You were working with a photographic test chart, which is clearly significantly brighter than a DSO. Now, your 200ms sub appears to properly swamp the read noise...however, your 0.1ms subs are not. My guess is you did NOT change the amount of light illuminating your chart between the two exposures. That does not in any way model what I've been talking about.
I have been talking about the difference between dark skies and bright skies. Let's say a longish L exposure of 300 seconds under dark skies of 21.3mag/sq" gives you a signal of 70e-. Now lets say you move back home, where your skies are 18.5mag/sq". The difference in sky brightness between those two sites is ~15x. There are 15 TIMES more photons reaching the sensor in your back yard, than at the dark site. That means instead of a 300s exposure, you now only need a 20s to get THE SAME 70e- signal. That 70e- signal is going to swamp the read noise, which is still only 3.5e- (or broken down into read noise and quantization noise, 3.2e- & 1.4e-). At the dark site, you have a background SNR of 70e-/SQRT(70e- + 3.5^2), and in your back yard you have a background SNR of 70e-/SQRT(70e- + 3.5^2).
The above example with the photographic test chart is not modeling this scenario, because when you switch to 0.1ms exposures, you are not compensating for a concurrent INCREASE in the photon flux. Your photon flux remains the same, and therefor the 0.1ms exposures are most assuredly not swamping the read noise.
Well...I just dropped some math in a no-math thread. So I'll stop here.
Edited by Jon Rista, 15 December 2016 - 11:57 AM.