Astrophotography by Dennis Isaacs

Contact me

 

Combination Tests

It seems that everywhere you look in astrophotography there are people debating the relative merits of different software and the ‘right’ length of exposure to get the best result. Having taken several exposures they need to be combined into a master, usually red, green, blue or H alpha in order to make a colour composite. Such combination leads to a master that exhibits more detail and less noise than the individual subs.

Detail needs to be assessed visually but the noise can (and should) be measured in an objective manner. Some of the combining algorithms can get rid of the normal interference in the form of satellite and aircraft trails, hot pixels and cosmic ray hits. In order to show exactly what this all means I did the following:-


I took ten cropped frames of the Rosette. Originally taken with the Tak FCT 100 and Artemis 11002 they were cropped from 11MB down to 0.3MB just to make them more manageable. All were ten minute subs through an Ha filter.

One frame was deliberately adulterated with an ‘X’ to simulate a satellite trail. One frame had been subjected to some kind of movement during exposure so it had badly misshapen stars. Both of these subs have been combined in a stack of up to ten to see what the final effect is in using different combination methods.



Frames 1 and 3 as used in the tests. Frame 1 is noisier for some reason as shown in the individual

noise measurements tabulated below.

 

 

All were combined in different ways using Maxim DL and the frames were all kept as 16 bit Integer FITS throughout. There is an obvious problem with summing ten frames where the average max pixel value is around 65,000 in each frame. The total value in the combined set of ten was about 643,000. This makes a nonsense of 16bit processing which has a maximum of 65,536. No attempt was made to accommodate these large values as it can be seen from the outset that the ever increasing noise and the inability to discriminate between legitimate signal and such things as satellite trails makes this form of combination a non-starter.

The Standard Deviation (strictly Sigma when dealing with the population but written here as ‘s’ for simplicity’s sake) was measured at the point p = 120, 340 using a cursor aperture of 8 pixels radius thus computing s over about 200 pixels. This measurement was used to determine the background noise throughout. (p = 120, 340 is just to the left of the 'X').
Frame
1
2
3
4
5
6
7
8
9
10
s
29.9
15.4
14.1
14.9
16.2
14.3
16.2
14.2
13.4
14.8

Noise, s, measured as the population standard deviation at point 120, 340 in each sub frame.

 

The cropped frames were combined in groups of 3, 5, 7 and 10 using Sum, Average, Median, Sigma Clip and SD Mask. Noise measurements were made from all 20 master frames and tabulated below.

Number of frames
3
5
7
10
Sum
36.6
47.1
60.7
80.4
Average
12.2
9.4
8.7
8.0
Median
12.7
11
10.5
9.3
Sigma Clip
14.8
11.1
10.5
9.1
SD Mask
10.8
8.5
7.7
7.1

 

The noise measurements give a clear indication of improvements with a bigger stack of frames and also show the difference between the different combination methods.

Notable is the fact that the noise measurement, s, for the Summed stack diverges. In other words, the more you add the worse the noise. This is the only such result and taken with the other points above makes a Summed output singularly useless if you want to clean up frames at the same time as combining them. In practice the Sum would be weighted by some divisor which means it is much the same as average.

Combining as an Average is widely regarded as the ‘quietest’ method. My results show it to be quieter than Median and Sigma Clip but not as quiet as SD Mask. Averaging is unable to separate proper signal from satellite trails and such unless there is a large number of clean frames. Something unlikely to happen most nights. If you want to use Average you should be prepared to throw away any subs that have been spoilt by satellite and aircraft trails, cosmic ray hits and jumps in guiding. That could mean a lot of hard earned data is useless. For this reason I would not recommend Average output.

Median output does the job quietly without any big surprises. Noise is respectable but this output does not always pick up hot pixels. If the inter frame alignment is good each frame might have a hot pixel in the same place and it will not be discriminated against. If there is a slight movement between frames you could end up with a smeared hot pixel which would probably not respond to a hot pixel filter. If this happens in RGB imaging the result is very unsightly with bi- or tri- coloured groups of pixels all over the place. Dithering is essential.

Sigma Clip is similar to SD Mask but as can be seen the noise is higher. Both these algorithms have parameters that can be varied by the user. In both cases I used the default settings.

SD Mask gave a result that was quieter than anything else which was something of a surprise. My previous tests showed it to be on a par with Average but not quieter. This is the combination method I have been using for some time now. I use it with my ST10 without dithering and my Artemis with dithering and both give immaculately clean results. If too many frames for the computer to handle need to be combined I combine them in groups of 8-12 and then combine the sub groups as an Average.

Using File – Combine Files from the main menu allows you to combine a set of files without opening them. Needless to say they need to be aligned first but this method reduces the RAM overhead and can speed things up. SD Mask is the slowest of all the above methods.


The test frames for each stack of 3, 5, 7 and 10 subs and each output are shown below.






The above results show that for three or more sub frames the fake satellite trail and the misshapen stars disappear if using

Median, Sigma Clip or SD Mask. Sum and Average are entirely unsuited to cleaning up this type of unwanted signal. In a

previous series of similar tests using a genuine satellite trail against a black background it needed at least seven frames to

clean it up using SD Mask. I believe this was due to the much higher contrast inherent in a sat trail against a black background.

In the example above the fake trail seems to have disappeared much more quickly because of the lower contrast between the

fake trail and the grey of the nebula.

 

If you look at the fine threads of nebulosity near the bottom centre of the picture you will see that it gradually becomes more visible

as it steps out of the noise. The small stars just continue to get sharper and more prominent. I do not believe there is any loss

of detail anywhere in the picture as a result of using sigma reject algorithms.

 

Conclusions are simple. If you do not like the idea of throwing away hard earned data using a Sum or Average output is not the

way to go. Median works well but I believe it is essential to dither between exposures to help clear hot pixels. The sigma reject

outputs work well at cleaning up unwanted noise but you should still use 7-10 or more sub frames for the best result. Residual

background noise is not much different from the Average combination and in this particular case SD Mask is actually lower.

Dithering is recommended for all exposures regardless of combination type.

 

Here is the un-cropped version.