Roger N Clark said:
In all long exposure photography of more than a few seconds of exposure, noise from dark current is a factor unless extremely cold, that is until the 7D2 came along. If you are doing 30-second night scenes of a city for example, unless it is very cold, noise from dark current will be greater than read noise.
I believe the dark current in Exmor sensors is some of the lowest out there outside of CCD sensors (and Sony CCD sensors have an order of magnitude lower dark current than the well established KAF CCD sensors, 0.003e-/px/s @ -10C vs. 0.02e-/px/s @ -15 to -20C). Current Exmor designs (i.e. A7s) seem to use both analog and digital CDS, and in testing they come fairly close to CCD level quality as far as overall noise levels and dark current noise levels (that was based on testing with a D800 that had been hacked with the black point hack.)
It would be very nice if you could test some Exmor cameras yourself. I much prefer your testing and reports over anything else out there, including DXO. I've been bummed that you have continued to stick to only testing Canon cameras despite the prolific expansion of Exmor sensors into an increasing percentage of the market over the last few years. It would be nice to have a broader and more diverse source of test data from you specifically, something that can be more readily compared with your existing Canon tests, rather than having to make assumptions about how other sources of data may relate to yours.
Roger N Clark said:
Some in this thread advocated doing astrophotography at low iso to avoid blowing out bright stars. A fine strategy if you don't care about the low end. The problem at low ISOs is that the digitization steps are too large to record the very faintest stuff, meaning quantization noise becomes a factor. These are analog systems and to record the very faintest stuff (the usual desire in astrophotos), one should work at an iso that is about 0.3 or lower gain (electrons/data number, or e/DN). For the 7D2, than means ISO 800 or more. (Note the e/DN is really an inverse gain--that's the way electronics engineers started defining it years ago). DN is one integer interval in the analog to digital converter.
I agree that quantization noise could be an issue. However read noise and dark current are so low in cameras with Exmor sensors, which makes it easier to expose the darker part of the signal well above the read noise floor. At lower ISOs, you can expose well above the read noise floor and still not clip the signal, improving per-sub SNR along the way, which gives you plenty of room to push the signal around without data burnout when stretching, reducing noise, and otherwise processing.
In my practical experience, even though you do encounter some issues with quantization noise, those issues pale in comparison to the issues with overall Canon read noise, banding in particular but the blotchy color artifacts as well, both of which Canon RAW images are famous for. It isn't just the amount of noise that can be a problem, it's how it presents, it's characteristic. Canon noise is far less clean and has non-random aspects than the noise that comes from Exmor based cameras, and even the noise from the D5200 (which uses the Toshiba sensor). I am currently working with some Samsung NX1 data, and the quality of the noise is phenomenal and lower than the 7D II, whereas the 7D II data which I am also working with is still heavily color blotched and has some rather strong horizontal banding. (I'll see if I can share the data sometime soon here...but in the grand scheme of things, the 7D II data does not appear to have improved much over past Canon cameras.)
Roger N Clark said:
So to detect faint signals and record the best detail in that faint signal, whether low light astro photo, or shadow detail in a very dark shadow, it is best to work at an iso that adequately digitizes that low end, and that is NOT at low iso, whether canon, nikon, sony, or whomever. It has nothing to do with read noise.
To detect faint signals, expose longer.
I've recently been in some very lengthy and detailed discussions about noise in general, read noise and dark current specifically, in astrophotography with a bunch of other (some very skilled) astrophotographers. Given all the math (Lee Jay posted a link to one of the handy pages we've been referencing), longer exposures are by far the best way to solve the problem of recording faint detail. More specifically, exposing to the point where photon shot noise of the faintest details just barely swamp read noise. Most of the experimental data we've worked with is actually from CCD cameras, where read noise levels are usually between 5e- and 10e-, which is actually higher than testing often indicates DSLRs have at ISO settings above 400 or 800.
The most common recommendations you are likely to find as far as exposing narrow band images is about 20 minutes. In actual testing, to increase the SNR enough to reveal very faint detail above the read noise floor, exposures as high as 55-90 minutes have been used (depending on the filter, with a 3-5nm Ha filter you can obviously get away with shorter exposures than a 5nm OIII filter, simply given the overall photon flux for those bands.)
Similar testing revealed that much longer exposures than are commonly recommended for OSC/DSLR cameras should be used to increase SNR and reveal faint detail above the read+dark current noise floor, assuming you have actually reached the read noise floor in a DSLR (in the case of Canon cameras, that is usually ISO 400 for APS-C, ISO 800 for FF.) Most beginners will expose a mere 2-3 minutes at most, and then wonder why their images are so noisy (regardless of camera brand used.) When you push exposure much farther, beyond five minutes, even beyond 10 minutes, then you start to see the signal strength itself, including those fainter details, begin to reveal themselves...even with a DSLR at non-cold temperatures with the higher dark current. (Side note: Depending on light pollution levels and use (or more accurately lack thereof) of light pollution filtration, 2-3 minutes may be all you CAN expose...which changes the restrictions, skyfog/LP levels, and therefor goals, and thus changes the whole discussion.)
My current project, wide field Orion's Sword with the 5D III+EF 600/4 II+Astronomik CLS-XL, using 480s, 300s, 90s and 15s exposures at ISO 400 (dithered, calibrated with a 200-frame master bias and 30-frame master flat along with the cosmetic corrections for hot and cold pixels in DSS), has revealed a lot of the very faint outer dust regions. This image below is just 1 hours worth of 300s exposures (out of my minimum goal of 6 hours), with sensor temperatures around 12-16°C gathered during the dark late waning/new moon phases:
The outer detail is rather noisy, as it's still buried in the noise floor (I presume primarily dark current/dark current noise, given the sensor temperatures). I had to perform some extensive color noise reduction and debanding to get it as clean as it is. Hence the reason for the newer 8 minute exposures...however all of the brighter stars and many of the medium sized stars are clipped already in the 5 minute exposures (or so close to clipping that they burn out when processing).
I also grabbed a handful of 10 minute exposures, however that did not seem to reveal any additional faint detail, although it did slightly improve the SNR of that faint detail. It primarily increased the amount of clipping in my brighter stars even more, to the point I found it completely unacceptable. Given the amount of noise overall, and the amount of color noise, I would much prefer to take the quantization noise and use an Exmor at ISO 100 or 200, avoid clipping the stars entirely, and get even longer exposures...say 12 minutes. I think in the end, based on my experience with a couple integrations from D5100s, the results would be far cleaner in the outer dark nebula regions. (Granted...that's anecdotal, and I don't really have the right to share the data without the consent of the owners...but as far as my own experiences influencing my own choices, there you have it.)
As a side note, since I did bring up the use of ISO 100/200 on Exmor-based cameras as a means of avoiding clipped stars. I recently started using the HDRCompostion tool in PixInsight, along with the MaskedStretch tool after linear processing. The image above I believe is actually a composition of my 300s and 90s exposures. HDRComposition with a set of decreasing exposure-length integrations, and MaskedStretch, along with very high precision 32-bit IEEE float FITS data, should help preserve the star detail and avoid blowing them out. Masked stretch can have an odd effect on stars...giving them a somewhat unnatural falloff into the background, but I guess I'll have to see if that is preferential to heavily blown out larger stars from 480-600 second exposures. As is usually the case, there are options to deal with camera limitations in post. If I had the option, I'd still use a D800 at ISO 100 or 200, take even longer exposures, and maybe even still do HDRComposition...just with fewer sets of exposures. (Why? Because it takes a really freaking LOOOONG time to get all that data in the first place!
Many days, sometimes spanning a couple of months, depending on the weather. Anything I can do to lessen the amount of time I have to spend pointing my camera at the sky gathering data when your talking about getting dozens if not hundreds of many-minute long exposures is extremely valuable, IMO.)