Why Didn’t My SNR Improve Much from 3h to 5.5h on M81/M82?

Tony Gondola · Apr 9, 2026, 06:28 PM

When shooting in heavy LP you really want your subs to be just long enough to swamp the read noise. Anymore just makes the LP and associated gradients a bigger issue to have to deal with. Under B8, I rarely go longer then 60 sec. even for narrowband.

I’m sorry about the filetype difference. I’ve changed it so both of them will be in TIFF. I’ve never heard about this. Is there an accurate way to calculate/see the optimal exposure time to ensure you’re swamping the read noise?

Dante Hunter · Apr 9, 2026, 06:38 PM

Tony Gondola · Apr 9, 2026, 06:28 PM

When shooting in heavy LP you really want your subs to be just long enough to swamp the read noise. Anymore just makes the LP and associated gradients a bigger issue to have to deal with. Under B8, I rarely go longer then 60 sec. even for narrowband.

I’m sorry about the filetype difference. I’ve changed it so both of them will be in TIFF. I’ve never heard about this. Is there an accurate way to calculate/see the optimal exposure time to ensure you’re swamping the read noise?

Here’s a talk by Dr Robin Glover (the creator of SharpCap) that goes into all that.

https://www.youtube.com/watch?v=3RH93UvP358

Doubling your integration (especially at such a low integration time) will do very little for an image that is predominantly sky background.

One of my longest integration times to date was 31h on Corona Australis… Of the 31h, I actually culled out nearly 10h of integration time due to sky background, slight gradients from partial moon glow etc… Dropping out 33% of the total integration actually improved the image drastically by getting rid of the skyglow/gradients, but even ignoring them, there was barely any discernable difference in noise going from 21h to 31h… I assume 80h would be significantly cleaner, and I know that dropping back to a stack of only the best 10h also significantly reduces the SNR in the dimmer areas of the target…

For what it’s worth,

I think the expected improvement depends on how one defines noise and SNR. When the sub-frames are independent and identically distributed (i.i.d), and the noise is random (i.e shot noise and read noise), then the “square root” rule predicts that the noise of an ideal pixel is expected to reduce and it’s SNR is expected to increase by a factor \(\lambda = \sqrt{k}\), where \(k = (t + \Delta t) / t\) is the ratio of the image integration times. In terms of percentages, noise is expected to reduce by \((1 - 1 / \lambda) \cdot 100 \%\) and SNR to increase by \((\lambda - 1) \cdot 100 \%\). In your example, \(\lambda = \sqrt{5.5/3} = 1.354\), which represents a theoretical noise reduction of ~26.1% and an SNR gain of ~35.4%, rather than 25%.

One method of quantifying pixel noise in an image is to estimate the standard deviation of the (neutral) background when the data is still linear. Using a couple of sample background patches, free of obvious gradients, noise in the 5.5h image is reduced by 28.09%, which represents a relative SNR gain of 39.06% over the 3h image. The slightly larger than expected SNR improvement may be explained by better observing conditions during the second session. Keep in mind that the simple square root rule does not strictly hold when the subs are not i.i.d or dominant systematic effects, like fixed pattern noise or light pollution, are also part of the noise budget.

Your example seems to be performing as expected in terms of quantitative SNR gain. However, our visual response to light and our perception of noise is complex and not linear, as Andrew already indicated above. Significantly more time on the target might simply be required to achieve meaningful perceptual improvements solely through integration.