oversampling ...

[Somnium]

With the recent discussion about sensors heating up, i was looking to get some thoughts on the importance of matching the scope to the pixel size. how do people feel about hard and fast rules around sampling? in particular what are the problems with oversampling. from my understanding, by oversampling you are dividing any given number of photons into smaller buckets without gaining any true resolution due to seeing restrictions. this results in an overall reduction in the sensitivity of the sensor.

say you were using 5 micron pixels, you would need 4 times as much time to gain the same signal from a 10 micron pixel. however, using the same scope and FL (i.e. no reducer), you are using 4, 5 micron pixels to capture the same detail as the 10 micron pixel. doesn't it all work out even in the wash? if say i am imaging at .5 arc seconds per pixel, am i going to have to image a particular object for longer (at a given magnification) to achieve the same SNR at 7.5 as/p?
is the higher sampling going to lead to blurrier images (at a given magnification) because it is not falling into larger discrete buckets?

[gregbradley]

In my experience sampling is not as vital as it is made out to be. The earlier 11002 9 micron pixel camera worked well on any scope widefield or long focal length.

Also with drizzle integration in PI (I haven't used it yet) it would seem some resolution loss can be regained if you are oversampled.

I think the main thing with oversampling is losing some sensitivity. I doubt its 4X as your example suggests. I have never experienced that.

You do see a more impinging image with correct sampling.

Sampling theory also is not 2X for a good sample, its at least 2x. So 3X seems to work so for example a 3 metre focal length gives you around .66 arc sec/pixel with a 9 micron pixelled camera. That seems to work out pretty well. Not too bad in poor seeing and excellent in good seeing.

More of an issue for me is smaller pixels have smaller well depth which can lead to weak stars that are easily damaged in processing and you have to be more careful to retain colour, shape and look.

As a very rough guide I am finding that even with 77% QE and 305mm aperture at a fast F3.8 I still need to accumulate around 10 hours of data to get a decently solid signal to process well. 15 hours would be better. 30 is probably overkill and perhaps less than 10 hours from a really dark site but even then around 10 seems a good target exposure length for that setup.

I am not sure what you are referring to about sensors heating up? Is this to do with DSLRs? Astro CCDs are cooled with regulated temperature control and I don't see any evidence of them heating up. With DSLRs I think its more the batteries that heat up as they expend their charge.

Greg.

[Slawomir]

I have been pondering about optimal sampling recently as well, and I thought that maybe it would also depend whether one is using an OSC camera, doing RGB imaging or narrowband. I feel that narrowband imaging could benefit more from higher sampling (less arcsec per pixel) as opposed to RGB or OSC imagining. Also, I think that imaging at under 1"/px with smaller telescopes (say 4 inch) would not show more detail even when seeing would allow for it (but it would work with larger telescopes).

[Somnium]

sorry Greg, this was in reference to the onsemi sensor thread heating up, not CCDs themselves ;)

[Somnium]

so the 4X sensitivity drop was just around the reduction in flux on the pixel due to the area which would be 25 microns squared as opposed to 100 for the 10 micron pixels

the point about well depth is a good one because stars are point sources then the majority of light will be concentrated on a few pixels regardless of the size and that is likely to become saturated, but larger wells will have this issue in longer exposures ...

[gregbradley]

I see, but the light is still hitting the sensor its just collected in the next pixel along. Read noise then comes into it where there would be a smaller signal for each pixel for the same exposure but then the counterplay is to simply expose for longer if the smaller pixels fill quicker and with less light than a larger pixel. Certainly some report that. Paul Haese advocates that strategy.

The KAF9000 is one of the larger pixelled sensors with 12 microns. Yet it is only slightly higher in QE than the 16803 but suffers badly from residual ghost images. I am not sure why but perhaps that is another factor to consider when comparing pixel size. Is there a relationship between pixel size and RBI (ghost images)?

Greg.

[ericwbenson]

Hi Aidan,
Oversampling by itself is not a problem, it just means longer sub-exposures to get past the read noise, and less FOV for the same resolution, the actual SNR is not affected. If you can live with these, and your camera can control the dark current for the longer subs, then by all means oversample.
BTW a 'benefit" of oversampling is that you could be less prone to saturating/blooming ;) (sorta, getting past the read noise might put you back to square one in terms of saturating, one would have to do the math)

Best,
EB

[Shiraz]

FWIW, that's how I see it as well. with the same fill factors, 10 micron pixels will get exactly 4x as many photons as 5 micron ones for a uniform illumination and that is the basis for considerations of sampling - you need to get the pixels just small enough to get all the detail - and not one jot smaller.

Granted you can bolt any cam onto any scope and get reasonable images. But if you want to get the best resolved image in the shortest possible time, you need to match your sampling to your seeing convolved with the scope psf (about 2.75 pixels per FWHM seems to be best). Getting it wrong can be dramatic - a 1.4x mismatch could make a 2x difference to the required imaging time, with no better detail.

If you oversample, the results will not be any better resolution than properly sampled, but the SNR will be worse and the image will look blurry because there will not be any abrupt edges and stars will look big (it won't actually be any blurrier than a well sampled image, but you will perceive it as blurrier unless you downsample it). With a bit of care, you could oversample and then use software binning after the event to get back some SNR and sharpness perception - that would give you true high res in exceptional conditions, with the ability to throttle back under average conditions. I guess this is what you were getting at with the "work out even in the wash" comment?

[Somnium]

true, read noise comes into play with over sampling. got to love this hobby, there is never a straight forward solution to anything ... so i am looking for the right sensor to put on the back of my iDK 14.5 (97" FL) my prime objective is galaxy and galaxy cluster imaging, its pretty much what i got into this game for. but i dont want to limit my FOV completely if i want to do survey work. i was thinking of the ML16200 but the pixels are 6 microns ... am i being crazy here ? or should i go to a 9 micron sensor

[Shiraz]

maybe 9 microns would be best unless you have found some super-good seeing

[Atmos]

With the sensor thread heating up recently I have been doing some dirty calculations and from what I can tell the KAF-16803 is probably one of the cleanest on the market at the moment (haven't compared it against anything without ABG but they should be even cleaner). Although the PL16803 has a read noise of 10e- and comparing it against the MLx814 which has a read noise of 2e-, you may at first consider the MLx814 to be far cleaner but in reality it isn't. This is where well depth becomes important as well, the 16803 has 100,000e- depth while the 814 has ~ 15,000e-, with a gain of 1.53e-/ADU and 0.23e-/ADU.

Ray can correct me if I am wrong here but that would ultimately make the 16803 ~ 33% cleaner? It may have a much higher read noise but the gain offsets it. That is my understanding of it anyway. This is meant irrespective of focal length, just in generalisation.

[Shiraz]

Hi Colin. My understanding is that the read noise determines how long the subs must be to reach the point where read noise is overwhelmed by sky noise. The 16803 requires >20x as long to get to "sky limited" as the 814, based on the figures you quote, so the 16803 needs much longer subs. Happily, it can handle longer subs without overloading, because it has very deep wells. As a generalisation, if you give the 814 lots of short subs and the 16803 a few long subs, you should end up at about the same point after roughly the same total integration time (ie, the final image will have about the same SNR and very roughly about the same dynamic range). So yes, I think that the 16803 could end up about as clean as the 814, but at the cost of requiring very much longer subs - it does seem to be a very nice chip.

In the situation where read noise dominates (eg narrowband), I think that the 814 will end up with an advantage, but of course it has small pixels, so you will be limited in how big a scope you can use - the 16803 can be used with much bigger aperture scopes, which is a major advantage in theory. However, the issue of scope size is also not clear cut. For illustration of this, consider that a 694 is well matched to ~ 1.2 m fl scopes and there are a few of these at f4 or thereabouts, including some low cost corrected Newtonians. 16803 pixels match to ~2.4 m scopes, but f4 scopes at these fls will be either hideously expensive or very unwieldy - so most scopes for the 9 micron chips end up around f8 (ie they have ~the same apertures as the fast scopes used for the smaller chips). Thus this potential performance advantage of the 9 micron chips is not fully realised at present. If GSO ever brings out an affordable 20inch f4 RH, the shouting will all be over - 9 micron will reign supreme if you have a strong enough mount to carry a large fast scope.

[Somnium]

something like a 6303 sensor ?

[strongmanmike]

:clap: Ray, you always sum things up and explain things so well, you are indeed the IIS CCD functionality guru, in fact people really don't need to post a thread to ask technical questions regarding CCD imaging theory and mechanics, they just need to send you a PM :D..but luckily they don't, so everyone else gets the knowledge too.

Mike

[gregbradley]

I agree Ray and that matches my experience. If you want to shoot galaxies then the smaller sensored Sony cameras are very good.

Your 14.5 would do best with a reducer in that configuration.

With the 16803 you can crop or you can use subframing when taking the images to get smaller file sizes to speed up processing.

These Sony sensors have made more affordable scopes capable of deeper imaging than the exKodak ones.

You end up taking shorter exposures with the Sony's (although I take 10mins at 305 F3.8 usually without any issue of overexposure) but the file sizes are small and the processing simple (usually no darks,flats or bias just data rejection and combine - very fast). The Kodak's require darks, flats biases to clean them up and generally larger files so its slower processing.

Greg.

[Somnium]

but the 814, has pixels of 3.69 microns, so unless i have a ~0.3 reducer then i am going to have the same issue i would have with the 6 micron ML16200, correct?

[RickS]

Yes, that's correct. For the iDK 14.5 9um pixels would be the best match unless you manage to find a very powerful reducer.

[gregbradley]

I was referring to the 694 which has 4.54 micron pixels. Still small but not as much.

The Planewave .66X reducer would no doubt work on the iDK being both Dall Kirkhams. That would give you 1629mm focal length and the 694 pixels should match close enough for that giving .57arc secs/pixel that should still work fairly well despite being oversampled.

native 2468mm focal length .66X reducer 1629mm focal length

9 micron pixels .75 arc secs/pixel 1.14 arc secs/pixel

6 micron pixels .50 arc secs/pixel .76 arc secs/pixel

4.54 micron pixels .38 arc secs/pixel .57 arc secs/pixel

If your seeing averages 3 arc seconds and you aim to sample 3X then you aim for 1 arc second per pixel. So you can see from that the bigger pixels match the seeing better for ideal sampling. So we get the usual idea that smaller pixelled cameras are better suited to shorter focal lengths unless you are on a mountain top.

But the advantage of the 4.54 micron pixels is 2x2 binning gives 9 microns plus more than 77% QE as its now binned. You would not lose anything there but gain. On a night of good seeing you could go back to 1x1 binning.
You could also do that with the 16200 and get 12 micron pixels and higher QE and with 16mp you have plenty of pixels still. You would be bang on 1 arc second per pixel at 2x2 with the 16200.

I think there would still be some trial and error as 2x2 binning tends to fatten stars a touch and lose a bit of detail in my experience. But the above suggests it would be a perfect match.

Its certainly the direction modern digital cameras are going. They are getting smaller and smaller pixels, more megapixels yet still manage to get less noise at higher ISO's with each new model.

I should try this out myself as I tended to only use the Trius 694 on my CDK17 at 1x1.

If I were using the 2x2 strategy I might go for the Sony ICX814 as it has 9 megapixels versus 6 and at 2x2 it now becomes 2.25 megapixels versus 1.5 megapixels with the 694 which is not many.

But the usual strategy is to use 9 micron pixelled cameras at 1x1 binning and long exposures with no reducer at long focal lengths.

Greg.

[Atmos]

I guess it really depends on what systems you're using as to how a comparison between the ICX814 and KAF16803 really is. From a quick and dirty calculation, on the same telescope and FL the 16803 should be ~8x quicker to reach read noise limited at Ha than the 814; this is using a QE of 40% and 53% respectively.

I can go through the numbers when I am not at work :) This basically means that a MLx814 on a 10" F/4 would be read noise limited at the same time as a 10" F/6.8 with PL-16803. That is effectively a 10" RC with 0.85FR. In real world terms. Pixel scale of 0.76" & 1.1" respectively.

[loc46south]

Hi Aidan - I did the sums and I chose the 12.5 to go with the ST10XME - it gives a native resolution of 0.67 arc sec pixel which fits right in with my seeing. Plus the ST10 has excellent QE and well depth.

Cheers
Geof

[Somnium]

yeah that was my thoughts too, by having the 16200, you have enough pixels to play with binning. i might go down this path and get the 16200, maybe look into the reducer down the track. ultimately my other options are the 16803 (a lot more pricey) and the 6303 (problems with blooming). the 16200 might be a bit more versatile.

[Somnium]

Thanks Geof

the 16200 at .5 is probably pushing, unless anyone wants to sell their second hand FL 16803 then i might have see how that goes. i cant justify the extra $7K +

[Atmos]

I just did a very quick calculation comparing the 964 against the 16200, they come in near identical in at 700nm (red). Just to give some of my methodology for scrutiny, I have used some numbers that I have taken from my own imaging system last full moon.

A month ago I took 23x900s frames of Ha from my suburban backyard, these ranged with me being sky limited from 60-120 seconds, 90 being the average stacked from all 23. I am sky limited at ~1860 ADU when imaging at 1.5"/pix. My QHY9 has a gain of 0.389 and a read noise of 8.5e-. The methodology is as follows.

1860 ADU for 90 minutes equates to 20.67 ADU per minute or 53.114e-/min.
The 8300 has a QE of 30% at Ha so theoretically there should be 177e- that were actually there. At 1.5"/pix I can therefore say that I am through my 3nm Ha filter I have a sky background of 78.71e-/arcsec^2/min.

Taking all of the figures from the FLI Microline stats page, the 964 chip would be seeing 125e- but only detecting 75.22e-. With a gain of 0.282 it would get 21.21 ADU per minute and it is sky background limited at 319 ADU so it would take 15 (903s) minutes to be sky limited under these conditions (full moon in suburban Melbourne).
Following the same calculations the 16200 would detect 87.58e- or 52.11225 ADU. It is sky limited at 823.5 ADU or 15.8 minutes (948s).
As a comparison, the 16803 would only need 130s!! Larger pixels (massive pixel surface area) and deep wells so high gain, even with a higher read noise.
The MLx814 comes in at 17.5 minutes (1049s). Although only having a read noise of 2e- it has a very small pixel surface area, lower QE than the 964 and even smaller wells (gain of 0.228). Given the extremely low read noise it only need 175 ADU to be sky limited BUT due to tiny pixels and very low gain, it takes a lot more time to get to that small ADU.

The importance of all this comes down to subframe length, the shorter your frames need to be to become sky limited the more you can take. The more subs you get the better your rejection becomes and higher SNR. But in short, the 16803 is the best of the lot, followed by the 16200. Although the 16200 took marginally longer, you're getting a MUCH larger FOV.

[Shiraz]

Colin, thanks for posting your methodology. Couple of quick questions, have you used inverse gain (electrons/ADU) as is commonly (perversely) called "gain" by the chip makers, or have you taken the quoted figure to be a gain in ADU/electron? were your subs 15 minutes or 90 minutes?

regards Ray

[Atmos]

Trying to actually track down all of the actual gain figures takes far too long so I am doing the straight calculation as (well depth)/65535. With my QHY9 I got my gain setting to the point where full saturation (ADU) happens when the well fills so it is actually at 0.389. I have been using e-/ADU as the conversion so the 16803 is the only one in that list that has above 1e-/ADU.

My subs were 15 minutes long but from that I could calculate that they wouldn't be sky limited until I hit 90 minutes (as an average of the night).

[Shiraz]

thanks for that Colin.

you quoted "1860 ADU for 90 minutes equates to 20.67 ADU per minute or 53.114e-/min" - shouldn't that be "20.67 ADU per minute or 8.04 e/min" (ie each electron generates more than 2 ADU if the inverse gain is 0.389e/ADU, so you should have fewer electrons than ADU), or have I completely misunderstood what you are doing?

[Atmos]

Nope, you are correct, I am just a sleep deprived idiot living on coffee today :question: It made perfect sense while I was driving from job to job today, now it just looks stupid :lol:

[Shiraz]

happens a lot to me as well :rofl:, but I am old, so that's my excuse.

[Atmos]

Well it makes more sense in that regards.
33 min for the 16893, 7ish for 964 and 6 for the 814 and 36 min for the 16200.