To Cull or not to Cull (for seeing; FWHM) - Forums

I was in Florida where we typically have good seeing, and I would cull a over about FWHM = 2, with much of my shots < 1.5.

Now in NC seeing sucks (this is a technical acronym for Sky Undergoing Continual Kinetics Sometimes). I’ve been building time on the Jellyfish, and found a great deal of my shots over 2.0, many over 3.0, some over 4.0. Price I pay for escaping from Florida I guess.

So I went to cull (this was really a second pass, I had culled many during acquisition). I decided to draw a line at 2.5. But this time I integrated first with all of them, and then with just the ones under 2.5. Going to show Ha only for simplicity — I started with 120 images, and culled to 96.

And you know what — I’m not sure I should be culling for FHWM.

This gif (I hope it shows animated here) is a 3:1 enlargement of the raw stack, preview stretch, from Pixinsight. It has no noise reduction and no sharpening. You probably have to click it to see it full size.

There’s some additional noise in the culled version (as expected), and to my eyes there are a few areas where there is a hint of a suggestion of better detail, but by and large all I can see is a VERY slight enlargement of the stars themselves.

So… to cull or not cull…

Oh… my weighting in stacking is PSF Scale SNR. I don’t know how strongly that favors FWHM in weighting, if it does it might explain this.

To cull or not cull….

Linwood

📷 DoesFWHMMatter.gif

For those that like numbers, PI says the noise evaluation is (not culled, and culled in that order)

Ch | noise | count(%) | layers |

0 | 6.870e-01 | 49.23 | 4 |

0 | 7.134e-01 | 50.80 | 4 |

I’ve never understood that measure but maybe someone else does.

Not culled: FWHM 2.986 px and eccentricity 0.2792 (1.86”)

Culled: FWHM 3.119 px and eccentricity 0.2807 (1.946”)

So PI thinks culling for FWHM yields WORSE FWHM than not culling. I can’t explain that one, but reran it twice.

Come over to the coast and you’ll get some of that good seeing again. I rarely see over 2.5” here during the summer. I’ve seen as low as 1.3” so far. I don’t think I have the capability of sampling much lower. Wintertime here is horrible though as the winds over the Appalachian’s seem to ruin it.

That being said, I generally don’t cull unless theres a huge spread or general outliers. Test the FWHM of your stacks with, and without the culled data. Most times you won’t see much of a difference. With things like galaxies, I will build two data sets… one with a FWHM rejection in place, and another the full dataset for SNR.

Every situation is different, but I’d avoid culling unless you absolutely have to.

Both Adam Block and John Hayes have spoken in favor of tossing them all in. There’ll be a marginal difference in resulting fwhm, but improved signal to work with going forward.

Not sure about PSF Scale SNR, but I found the PSF Signal Weight (the default and recommended) promoted stars with poor FWHM a bit much, so I use a 75/25 mix of PSFSW and FWHM. FWIW.

Now that you’ve joined us in the seeing trenches…..going LRGB and shooting L when seeing is best, and RGB the rest of the time, promotes best possible resolution and no downside to the poor seeing data.

Cheers,
Scott

Scott Badger · Oct 23, 2025, 10:38 PM
Now that you’ve joined us in the seeing trenches…..going LRGB and shooting L when seeing is best, and RGB the rest of the time, promotes best possible resolution and no downside to the poor seeing data.
Cheers,
Scott

100% agree here.

It depends a lot on the weighing during stacking. In the tests I’ve done in Siril, there is little advantage in aggressively culling if I’m favoring FWHM during stacking. What I do is look at the metrics of the lights, pull the few really out to lunch outliers if any and stack the rest, favoring FWHM.

Scott Badger · Oct 23, 2025, 10:38 PM
Now that you’ve joined us in the seeing trenches…..going LRGB and shooting L when seeing is best, and RGB the rest of the time, promotes best possible resolution and no downside to the poor seeing data.

Except I like narrowband.

I shot some RGB for this target for stars, it was barely visible (of course it was only an hour each but still).

In my Bortle 8 city sky in Oregon if I culled over 3, I would not stay with this hobby. Anything under 5 I keep and stack. I sometimes get a night under 2 for all my subs but that is rare. I stopped stacking in WBPP because it rejects too many. APP will stack anything you put in and I get decent results with it.

Moonlight is a much bigger problem, I shoot NB only when the Moon is out. No moon nights I can get decent LRBG even in my skies.

Dan Kearl · Oct 23, 2025, 11:24 PM
In my Bortle 8 city sky in Oregon if I culled over 3, I would not stay with this hobby. Anything under 5 I keep and stack. I sometimes get a night under 2 for all my subs but that is rare. I stopped stacking in WBPP because it rejects too many. APP will stack anything you put in and I get decent results with it.

Yeah… not trying to compare seeing, some are luckier than others.

But if your seeing averages X, is it worth throwing out subs over 1.5*X? 2*X? Or never throw them out just for seeing (there are clouds and gradients and eccentricity and other reasons of course).

Not everyone agrees with me but I believe that the way low photon count details are effected by momentary effects of seeing is different than stars. Stars are so bright that the photon flux per second is much higher then the faint, low photon count details we are most interested in. I think that can lead to situations when short time scale errors (seeing or guiding) can be very visible with stars and yet have hardly any effect on low count details. As such I believe that stars can be a little bit misleading in terms of overall quality and resolution. This is probably part of the reason why not culling too deeply works.

At my home site in Vermont seeing is very bad. I only culled over 4”, as those subs just add glow around the stars in the final integration. Under 4” and I would keep it. This is for widefield imaging of course, if you are most interested in fine details at longer focal length then the culling process might be a little more aggressive.

Chris White- Overcast Observatory · Oct 24, 2025, 12:29 AM
At my home site in Vermont seeing is very bad. I only culled over 4”, as those subs just add glow around the stars in the final integration. Under 4” and I would keep it. This is for widefield imaging of course, if you are most interested in fine details at longer focal length then the culling process might be a little more aggressive.

That is what I’m most interested in, but that’s what surprised me here — pixel peeping after culling poor FWHM did not yield more detail.

Linwood Ferguson · Oct 24, 2025, 12:31 AM
Chris White- Overcast Observatory · Oct 24, 2025, 12:29 AM
At my home site in Vermont seeing is very bad. I only culled over 4”, as those subs just add glow around the stars in the final integration. Under 4” and I would keep it. This is for widefield imaging of course, if you are most interested in fine details at longer focal length then the culling process might be a little more aggressive.
That is what I’m most interested in, but that’s what surprised me here — pixel peeping after culling poor FWHM did not yield more detail.

Your experience matches my own. Like I said, I only cull the really bad ones that I notice add some glow around stars, like 4” or more. At my remote site I dont cull anything for FWHM. I dont even look at them. To be honest, my culling method is very simple. I just look at everything through blink and dump subs that have obvious flaws, like clouds or the occasional elongation from whatever…. Everything else, I use without discrimination. I’m imaging at 650mm FL and 875mm FL.

Linwood Ferguson:
Scott Badger · Oct 23, 2025, 10:38 PM

Now that you’ve joined us in the seeing trenches…..going LRGB and shooting L when seeing is best, and RGB the rest of the time, promotes best possible resolution and no downside to the poor seeing data.

Except I like narrowband.

I shot some RGB for this target for stars, it was barely visible (of course it was only an hour each but still).

then image H-alpha during the times of good seeing. Luminance is not something that is independent of RGB. In LRGB combination, luminance is effectively being distributed cleverly between R, G and B, most heavily to the G channel because that’s what our eyes are most sensitive to. In the case of narrowband, the information mostly exists in H-alpha, which, in SHO, is mapped to green. You could get a similar effect in pure RGB imaging by taking green during times of good seeing.

I cull pretty aggressively. I dont have an absolute FWHM threshold that I go by. Instead I either cull with inspection using Blink, or with SubFrameSelector with thresholds set depending on the dataset. I tend to prefer chasing detail versus bringing out faint stuff, so I am okay with my compromise.

Arun H · Oct 24, 2025, 12:51 AM

then image H-alpha during the times of good seeing. Luminance is not something that is independent of RGB. In LRGB combination, luminance is effectively being distributed cleverly between R, G and B, most heavily to the G channel because that’s what our eyes are most sensitive to. In the case of narrowband, the information mostly exists in H-alpha, which, in SHO, is mapped to green. You could get a similar effect in pure RGB imaging by taking green during times of good seeing.

My only NB option at the moment is Ha, but if I’m imaging an emission nebula, like the little rosette (recent target), I don’t shoot L at all and apply the best seeing strategy to Ha (as you suggest). For a target like this, the RGB is so dominated by red anyhow, I usually don’t bother adding Ha to the color space, but using it as the luminance adds a LOT more detail.

Cheers,
Scott

I cull obviously bad subs, as in elongated stars from got know what, but bad seeing is normal in southern Ontario. If I culled for bad seeing I wouldn’t have much to stack and my scopes would stay indoors most clear nights. lol

Another bit of data for those interested, and since BlurXterminator has become such a part of this hobby I involved it.

I found some more culled data, so this is a comparison of Ha with 96 frames (8 hours) vs 155 frames (13 hours). The FWHM extended up to near 5”, and now there are about 40% of the frames over my prior cull target of 2.5”.

On the right side is the unmodified stack. On the left side is a run of BlurXTerminator with mild settings, the question in my mind is whether the AI-recovered detail from Bx sees more detail with more-but-less-focused data, or less-but-better-data.

To me there is much more spread of the stars now (even post BX even with identical parameters), but for me the detail is harder to tell — it’s different, but is it just brighter from more data, or more blurred? There’s clearly more noise in the culled set.

But there’s so little detail difference it certainly argues toward the “no cull”.

PS. I had no bad eccentricity (thanks Astro-Physics), and I did cull the one frame with a cloud that came by.

📷 FWHM_Again.gif

I would think this all comes down to just your preference and willingness to throw in the data or not. I mean does a 4” FWHM star make you sick or repulsive to you? lol kidding… but I would lean towards including all of it and getting the signal as much as you can. I do look at those values and like you are saying, if values creep up to 2x that of some of my other images, I would get an indication of what is happening but mainly with narrowband I would favor more signal and if needing to deal with FWHM or stars, do a good seeing evening just for those perhaps? I usually stack over in APP, not Pix, but I’ve done a lot of comparisons on stacking some vs all data and the SNR is really improved with the more data I do, even when they are ‘questionable’.

Linwood Ferguson · Oct 24, 2025, 03:45 PM
Another bit of data for those interested, and since BlurXterminator has become such a part of this hobby I involved it.
I found some more culled data, so this is a comparison of Ha with 96 frames (8 hours) vs 155 frames (13 hours). The FWHM extended up to near 5”, and now there are about 40% of the frames over my prior cull target of 2.5”.
On the right side is the unmodified stack. On the left side is a run of BlurXTerminator with mild settings, the question in my mind is whether the AI-recovered detail from Bx sees more detail with more-but-less-focused data, or less-but-better-data.
To me there is much more spread of the stars now (even post BX even with identical parameters), but for me the detail is harder to tell — it’s different, but is it just brighter from more data, or more blurred? There’s clearly more noise in the culled set.
But there’s so little detail difference it certainly argues toward the “no cull”.
PS. I had no bad eccentricity (thanks Astro-Physics), and I did cull the one frame with a cloud that came by.
📷 FWHM_Again.gif

The culled data is showing improved detail. To me it looks like the clear winner in this case.

Tony Gondola · Oct 24, 2025, 04:28 PM
data is showing improved detail. To me it looks like the clear winner in this case.

Not as bright and more contrast, but I if the culled versions are actually sharper, it’s pretty marginal to my eyes….

Cheers,
Scott

This is a game of inches for sure…

Bear in mind that image was at 0.62” scale and then enlarged 4:1, so it’s beyond pixel peeping.

To be fair, if fwhm is large, then the contribution is not really helping. Details are just as blurred. Just because snr improves, doesn’t mean its meaningful signal. Test test test. Everyone’s milage varies.

Chris White- Overcast Observatory · Oct 24, 2025, 07:04 PM
To be fair, if fwhm is large, then the contribution is not really helping. Details are just as blurred. Just because snr improves, doesn’t mean its meaningful signal. Test test test. Everyone’s milage varies.

I agree here, but I do feel those details easily sharpen if one needs vs trying to bring detail forward. If one has 600 images it is pretty easy to throw out 150 terrible images, but if you only have 20 images, would hate to throw out 5 images. This is why I really like APP’s normalization grading and graphing for expelling bad images, so easy to do a side by side test. stack both versions, load back in as lights and analyze the results.