How accurate are noise/SNR evalutions in ImageIntegration and SubframeSelector?

[astrovienna]

Lately I've been trying to make more use of SubframeSelector.  Right now I'm processing a very faint PN at long focal length, and my eyeballing of the frames in Blink doesn't match the readings I'm getting in SubframeSelector.  Some frames that SS sees as outliers in FWHM, SNRweight, or other stats, whether good or bad, look quite average to me in Blink.

So I'm wondering how accurate the SS measurements are. That also got me to wondering how accurate the noise evaluation is in the ImageIntegration process.  How much should I rely on these tools, if the outlier frames look pretty ordinary in Blink?  Is it just that the STF applied by Blink isn't giving me enough insight into the quality of my frames?

Kevin

[Robert Q. Kimball]

I'm no expert, so this may not be relevant. 

It make a big difference in Blink if you use the ScreenTransferFunction on each of the images or you use the ScreenTransferFunction on one image then use the same stretch for all the other images.

To see this in action, use the SubFrameSelector to identify your best and worst image based on SNR .  Then open both of them and stretch them both individually.  Then, stretch one of the images and apply the SAME stretch to the second image.   I think you'll see a big difference.

Again I'm not an expert.

[ngc1535]

Just a quick comment (perhaps it is useful).

From image to image the stars PSF (FWHM) can change due to the seeing. In fact the blurring due to the seeing affects the best images (sharpest stars) very strongly compared to fuzzier ones when comparing metrics such as S/N.

However when you are blinking the images and using an automatic STF- these differences are difficult to see because you are displaying the stars in a non-linear way. The fuzzy stars and the sharp stars look more like each other. This is also true for variations in transparency/attenuation from image to image- again with a non-linear automatic STF... it is difficult by eye to see the differences.

So, the variation in S/N you see by virtue of a (computed) measurement (aperture photometry of some sort) when using the SubFrameSelector is likely a much better evaluation than what you are going to guess by visual inspection through blinking.

-adam

[msmythers]

Kevin

I like to think of Blink being useful for large scale issues. Clouds, trees, haze, gradients and so on. So both tools are important if you know you need them. If you know your environment and don't have large scale issues then you would not need to use a purely visual tool such as Blink.


Mike   

[astrovienna]

Mike, Adam and Robert, thanks very much for the helpful guidance.  It sounds like I should be using Blink to discard obviously flawed images, but otherwise relying on weightings in SubframeSelector. 

One related question:  can I use what I see in Blink to discard images that appear to have low contrast?  I often notice that one or two frames from a given night appear to have significantly lower contrast than others.  Maybe haze blew through for that frame, but I don't know for sure, because I'm usually asleep.  As long as I use the Blink option to apply the same STF to all the images, is it safe to assume that those frames really do have low contrast, and thus should be excluded?  Or am I  better off letting SubframeSelector do that work - ie, giving them lower weights?

Kevin

[RickS]

One related question:  can I use what I see in Blink to discard images that appear to have low contrast?  I often notice that one or two frames from a given night appear to have significantly lower contrast than others.  Maybe haze blew through for that frame, but I don't know for sure, because I'm usually asleep.  As long as I use the Blink option to apply the same STF to all the images, is it safe to assume that those frames really do have low contrast, and thus should be excluded?  Or am I  better off letting SubframeSelector do that work - ie, giving them lower weights?

I use Blink for this.  Sometimes subs affected by cloud or haze have stats that look OK, especially if there is some light pollution (reflected light contributes to signal and gives better SNR.)

Cheers,
Rick.

[astrovienna]

Thanks Rick.  I can usually spot clouds and haze without much problem in LRGB subs, because they tend to shift the vignetting pattern.  But I find it much harder in NB subs.  Vignetting patterns don't shift, they just have brighter backgrounds.

Back to the original point on accuracy.  I just compared ImageIntegration with noise evaluation weightings vs. all weights=1.  Below are the reports from process explorer.  As expected, the noise-weighted stack has higher noise reduction from both the reference and the median.  But, if I understand correctly, it has higher gaussian noise.  How can it have both higher noise reduction and higher noise?  I've been following Jordi Gallego's walkthrough on integration techniques (http://www.astrosurf.com/jordigallego/articles/Image_integration_JGallego.ppt), so I think I'm doing it right.  To my eye,  the STF-d results are identical, but these results are so similar that I doubt I'd be able to see the difference.

Noise weighted:
Total : 845927 0.518% ( 46866 + 799061 = 0.029% + 0.489%)
Gaussian noise estimates : 1.2566e-04
Scale estimates : 1.8514e-04
Location estimates : 1.7677e-03
SNR estimates : 8.0826e+02
Reference noise reduction : 1.4495
Median noise reduction : 1.6147

Unweighted:
Total : 845927 0.518% ( 46866 + 799061 = 0.029% + 0.489%)
Gaussian noise estimates : 1.2532e-04
Scale estimates : 1.6704e-04
Location estimates : 1.7601e-03
SNR estimates : 8.1228e+02
Reference noise reduction : 1.3112
Median noise reduction : 1.4608

[RickS]

Hi Kevin,

You can't directly compare the Gaussian noise estimates as the scale isn't the same.

Cheers,
Rick.

[astrovienna]

Thanks again, Rick.  I'm still at that stage where a little knowledge is a dangerous thing.  But with help I'm slowly learning.

Kevin