Weights Comparison NSG vs Non NSG

Roger Smith

Hi Adam,

I know you like to know things make sense in your understanding of PI processes and scripts, and go to lengths to do that in your video instruction. It is one of the reasons I really like, and beginning to understand what I am doing.

Now I am working to compare weights showing in NWeight Image Integration vs weights showing in Noise Evaluation Image Integration of my own 100 images of M101 color camera data. Data is from my very urban site with tons of varying light pollution over several nights of imaging.

I watched your Horizons NSG A Second Example video comparison video as a starting point, looking at the Process Console outputs.

Attached is an Excel file of the Process Console outputs and weights. I have  cooled CMOS color camera data.

My problem is there are too many places to find weight data and I don't understand which data is comparable....

NSG Script data Integration:

1. FITS header keyword - NWeight of the NSG created files. One NWeight for all 3 channels. 

2. Weight calculated is showing in the process console after Image Integration of the NSG images with no Normalization. All 3 channels have same weight for an image, and are different than NWeight. But why is there any weight because it should be using NWeight?

The above  item 1 & 2 weights are different.

Non NSG integration of same data set (registered, not the NSG data).

1. Weight calculated is showing in the process console after Image Integration of the Non - NSG images with with Normalization.  This is the same place you used in the example in your comparison video. 

2. There are 3 channels of data, and the weight in each channels is different, as expected. 

Can you explain how I can continue my comparison with so many different generated weights? I want to show NWeight is better than noise evaluation.

Why would one NWeight be used for all 3 channels.

The data leads me around in circles without logical conclusion, especially after watching your using it in your Horizons NSG Second Example.

At the bottom of each Process Console listing you can see the Image Integration settings.

Thanks for your advice and review.

Roger

BlockHead

There are two parts of your question- one part I do not know the answer to...so I will be coming back with that. BUT, I am pretty certain you may have missed something when trying to replicate what I did. You need to assign the *same* file as a reference when you run Image Integration for both methods in order to compare. In your Excel file if you pick the exact same file the Non-NSG images combined in ImageINtegration will show that file having a weight of 1.0. But you do not show that...that means when you did ImageIntegration with the non-NSG files and the normal noise-evaluation way- you did not assign a reference (or the same one).

-the Blockhead

BlockHead

From John Murphy (the developer of the script0:

"NSG will calculate and apply the scale and gradient correction independently for each color channel. So the target frames should not only match the gradient and scale of the reference frame, the color balance should also match the reference frame.

The NWEIGHT is calculated from channel 0. Hence, for a color image, the weight is entirely based on the red channel. As far as I am aware, ImageIntegration is only able to use a single weight per image. I had to choose between using a single channel, or devise some strategy to combine them. I chose the red channel partly for convenience (same code for both mono and color), and partly because red is a particularly important color in astrophotography. Since the weights are relative, it does not really matter which color has the most noise. The choice only really matters if the color noise varies from one color to another in a non linear way."

So indeed, the weight calculated from a single channel- but  it is the same sensor and conditions- so you wouldn't expect there to be a difference in weights between the channels. 

-the Blockhead

Roger Smith

Hi Adam,

Yes, I did overlook to make the same images as reference. Now I rerun with same image number for NSG script, NSG data image integration, and normal (non NSG) image integration. 

Remember, my data is from urban site with lots of light pollution, and cooled color CMOS camera.

Please see attached spreadsheet with data filled in (from Process Console and Fits Header). I have 100 images, but only plotted about 40 of them as very tedious to copy and paste each weight from Process Console.. 

On Results Tab, I made scatter plots. As you suspected, there is poor correlation between NWEIGHT, and the weight (red channel) calculated during normal integration.

I find all these have good correlation with NWEIGHT:

--- Stars Detected (from Subframe Selector)

--- Median (from Subframe Selector)

--- MAD (from Batch Statistics)

Maybe other factors showing in Batch Statistics Tab have good correlation with NWEIGHT.

Kindly review the spreadsheet (and the yellow-highlighted settings showing in Process Console Tabs)  to see if things look correct. 

If OK, then I will continue the effort for all 100 images.

How should I judge if the output of Image Integration after NSG Script running vs Normal Integration is actually better? 

Thanks,

    Roger

BlockHead

Hi Roger,

I believe your information supports my assertions. If you compare the columns- the images given the highest weights in your data set are the worst images according to NSG. Furthermore, don't take NSG's word for me. Go look at your normal registered images and compare them... are the images that Noise Evaluation actually calculated were your best images...ARE THEY?? I predict they will not be. This is the point of the exercise. 

By weighting cappier images higher..the combined image IS worse (whether you can see it or not.. but generally you can).

-the Blockhead

Roger Smith

Hi Adam,

Yes, some images in my set with low NWEIGHT had high noise weight (normal integration) and were visually 'cappier' than others. Thus the image integration noise weight, as used in image integration, was not 100% reliable measure of image quality. This was most prevalent in my worst images, not the ones with lower signal backgrounds.

I fully agree that highly weighting a poor image vs the better ones will result in a less than optimal integrated image. But visually we may not see the difference. Depends on the quantity (%) of images that are not good but are highly weighted by image integration.

If I have a set of images with uniform background and choose not to use the NSG script, I would weight the images based on number of stars, and lowest median values in subframe selector. Both these subframe selector parameters (in integration without NWEIGHT) had good linear fit (.94+) R^2 correlation to NWEIGHT.

So if 'weight' is not a good measure of image quality, then signal to noise must not be reliable way to judge (measure) an integrated image's quality. 

Everyone talks (writes) about a good images have high signal to noise. It is an easy qualitative concept. From the quantitative side, I cannot find anywhere in PI that signal to noise is calculated. Same when I search Rogelio Bernal Andreo's excellent book Mastering Pixinsight. Maybe is should only be a qualitative thing and judged visually as no one is going to post an image and say SNR=xxx for their image. 

Your comments appreciated if you like to provide. All this is my learning experience.

Thanks,

     Roger

John Murphy

NSG v1.1 fixed a problem with the NWEIGHT calculation. The noise ratio
needed to be squared. Alas, it would therefore not be reliable to use results
obtained from pre NSG v1.1 to find relationships between NWEIGHT and subframe
selector parameters.

Trying to find relationships between NWEIGHT and subframe selector
parameters is going to be fraught with difficulties because the subframe
parameters are affected by too many variables. 

Using subframe selector parameters to determine image weights can be
tricky. For example, the number of detected stars depends not just on transmission
and light pollution, but also on how sharp the star profiles are - due to
seeing, guiding and focus shift. The median level tells you nothing about transmission (how much of the useful signal was absorbed by the atmosphere / clouds).

NWEIGHT should depend on the transmission (how many useful photons
reached your sensor) and how much noise was introduced by light pollution / sky
glow / the sensor. It is a relative measurement, only relevant to files from
the same NSG run. It also depends on the PixInsights MRS noise calculation which, if used correctly, I believe to be quite good, although it is unlikely to be perfect.

Regards, John Murphy

Roger Smith

Hi John,

Thanks for your reply, and explanations. 

I will stick to the script and use it as it was designed and not guess what it could be used for.

Thanks,

     Roger