Over sampling with a CMOS

[jjosefsen]

Hi,

Been reading some topics here and on CN about desired sampling rates, most of the threads seem to go by the same convention, that you need to be in the sweetspot of 1-3px/arcsec in "average seeing" conditions.

In one thread on CN a user proposed that oversampling wasn't as big a problem with low read noise CMOS sensors as with the old CCD sensors where read noise is harder to overcome.

I've been trying to wrap my head around it for the last few days and I am not sure I understand why over sampling is such a problem, can anyone give some practical examples of why it would be a problem on a CMOS based sensor?

 

Under sampling is bad because fotons can land on 1 pixel or a "block" of pixels and lead to "blocky" or square stars.

Over sampling is bad because ??

 

Looking forward to hearing your thoughts on this.

[RolandKol]

As I understand:

Oversampling is bad, as you will need much Longer exposures to get the the same amount of  Signal if you would be in a sweet spot.

And in this case CCD would suffer more, as they have higher noise in comparison to CMOS (the end result would be a bit noisier with CCD).... However, some CMOS sensors, do suffer from AMP glow during long exposures.

I would not bother my self with cam issues, but more about guiding, wind and "UF'os" which may visit your FOV during a 20min exposure.

[jjosefsen]

But isn't that more to do with pixel size of the camera? Why does pixel scale / focal length matter?

[coatesg]

Oversampling, when comparing two cameras with equivalent read noise and quantum efficiency on the same scope, would potentially result in lower signal to noise ratio when compared to a properly sampled image (SNR is the important thing, not just raw signal). CMOS sensors are more likely prone to this as they have quite small pixel sizes compared to classical CCD cameras (I would love a CMOS with low read, low thermal noise and ~6-7um pixels....)

You can of course bin in the same way that CCDs allow, but you won't get the same reduction in read noise that binning gives for CCDs - CCDs combine charge before readout so in 2x bin, whereas CMOS does it after readout (you could do it in software to the same effect). CCDs reduce the contribution of read noise by a quarter, whereas CMOS (with an average approach to binning) reduces the contribution by half.

You do, however, need sufficient sampling such that you are at least at around 1/3 to 1/2 of the seeing limit for long exposures to sample the image properly. This also helps processes like deconvolution where undersampling causes issues. Somewhere in the region of 1"/px seems reasonable to me. 

[RolandKol]

With smaller pixels, you need longer exposures to fill them all...

Long focal length also increases your exposures to gather more photons....

(not taking aperture into account)

You end up oversampled, which is not bad for your final image, but increases your exposure time drastically.

if you mix F12 with small pixel size sensor, -  CMOS is more handy as it has very low read noise at high GAIN, which you will automatically want to increase to reduce exposure time.

I hope someone will add and correct me if I am not right, but this is what I understood from my research before buying the cam.
 

 

[jjosefsen]

Thanks for your answers so far, definately food for thought.

I'm still not completely understanding why  the Focal Length matters when it comes to filling the pixel wells, particularly as aperture tends to go up with FL, at least for the reflectors I'm looking at.

 

I was looking at an 8"/208mm F3.9 Reflector, but currently it would be paired with an imx183 based sensor (2.4um pixels), leading to oversampling, hence why I have been reading up on sampling rates and trying to understand the potential problems.

My next scope would ideally sit somewhere around 800mm - 1000mm FL and be F4 to F5, but if that means getting another camera, well then I might be better off just improving my widefield setup and then at a later time getting a dedicated longer FL setup, maybe an RC paired with a CCD camera.

[vlaiv]

Oversampling affects SNR - because there is limited amount of light from target.

Probably easiest to explain if we take extended uniformly emitting target - one that emits 100 photons per arc second squared.

For the moment, let's forget about proper sampling / Nyquist / other stuff and concentrate on resolution solely.

Let's examine two different scenarios:

1. You are imaging at 1"/pixel

2. You are imaging at 0.5"/pixel

These units are linear, so actual area that matches one pixel will be in first case 1 arc second squared (1"x1") and in second case it will be 0.25 arc seconds squared (0.5" x 0.5")

If target emits on average 100 photons per arc second squared of surface (in duration of single exposure), in first case your pixel will get 100 photons, while in second case it will get 25 photons.

Let's also ignore read and thermal noise for the moment (they add into this, and make things even worse ) and just look at shot noise associated with target. Its magnitude is square root of signal level, and in first case it will be 10 while in second case it will be 5.

SNR for case 1 - 100 / 10 = 10

SNR for case 2 - 25 / 5 = 5

So SNR per pixel is twice lower in second case when using smaller pixel (for same aperture).

Resolution is tied to pixel size and focal length. So using higher "magnification" - or higher resolution means lower SNR per pixel for same imaging time, regardless if you are oversampling or undersampling or sampling just right (goldilocks? )

Same effect of higher resolution can be because using smaller pixels with same focal length. So focal length is not only factor that contributes to this it is actual resolution. Maybe better description of "speed" of telescope would be aperture at resolution than f/ratio. It describes better which scope will be "faster" at a given target resolution - one with bigger aperture.

Back to sampling. Oversampling just means that you will not capture any additional detail if using higher resolution. Why use higher resolution further if you are not going to capture any new detail (limited by seeing, guiding, aperture and pixel size - there is twist to the story, but a small one) if it will hurt your SNR and you need to spend more time capturing to achieve same SNR.

Binning to the rescue! Regardless of what sort of binning you use (CCD and hardware, or CMOS and software) - it increases SNR by at least x2 (if binning 2x2). Software binning will increase it x2 and hardware just a bit more - and that depends on ratio of magnitude of read noise to other noise sources in the image (general guideline for exposure dictates that we go "above read noise" - so in reality read noise is small compared to other noise sources, and software binning is close to hardware one).

There is a little twist to the story - detail in the image depends on pixel size - not because of sampling, but because of the fact that pixels have surface. Nyquist works for point sampling (imagine pixels that have no height and width, but are equally spaced and somehow sample light). Fact that pixel has surface adds additional blur to the story.

When binning normally - you are loosing some detail even if you oversampled in the first place. There is however way to "bin" your images - or reduce resolution to regain SNR when oversampled without introducing this blur.

Have a look here:

 

[RolandKol]

I would avoid F4 or faster ones reflectors.... they are extreamly sensitive to collimation errors... you will end buying expensive collimation tools and "playing" a lot before each session...

(not had one, but wanted to... read a lot and got scared a bit...)

F5 would be a good start, I have 130PDS and it takes time to set up (not always, but regularly, especially after moving imaging outside), with F4 it would be each time before imaging.

As per sensor, - if your cam worked with 80ED @F6.4, it will be good on F5 also... You will end up a bit over-sampled....

Before buying my cam, I checked these at the start... http://astronomy.tools/calculators/field_of_view/  and  http://astronomy.tools/calculators/ccd_suitability

but ended up simply looking for the Final Image result with the cam and scope I want at.
I found only few for you  now

https://www.flickr.com/photos/143795068@N07/30501479998/ and https://www.astrobin.com/342946

Your sensor is not yet very popular (I checked ASI183, which is  with your sensor, vs ASI1600 before buying my cam), - not much results for 183, so I bought ASi1600, but it will work in your case also.
I guess... 183 will compete with popular ASI1600 quite fast.

P.S.

I noticed, guys who image with Hyperstar or Rasa, -  almost all go for imx 183 sensor.

 

 

[jjosefsen]

1 hour ago, vlaiv said:

*Wisdom by the bucket load*

Thank you! Now I get it. I was not at all considering surface of target as the source of photons. This makes sense.. smaller surface = less photons over time.

I was actually reading that other thread as well, but was still lacking the understanding of the actual problem.

Thank you for the excellent explanation!

[jjosefsen]

1 hour ago, RolandKol said:

I would avoid F4 or faster ones reflectors.... they are extreamly sensitive to collimation errors... you will end buying expensive collimation tools and "playing" a lot before each session...

(not had one, but wanted to... read a lot and got scared a bit...)

F5 would be a good start, I have 130PDS and it takes time to set up (not always, but regularly, especially after moving imaging outside), with F4 it would be each time before imaging.

As per sensor, - if your cam worked with 80ED @F6.4, it will be good on F5 also... You will end up a bit over-sampled....

Before buying my cam, I checked these at the start... http://astronomy.tools/calculators/field_of_view/  and  http://astronomy.tools/calculators/ccd_suitability

but ended up simply looking for the Final Image result with the cam and scope I want at.
I found only few for you  now

https://www.flickr.com/photos/143795068@N07/30501479998/ and https://www.astrobin.com/342946

Your sensor is not yet very popular (I checked ASI183, which is  with your sensor, vs ASI1600 before buying my cam), - not much results for 183, so I bought ASi1600, but it will work in your case also.
I guess... 183 will compete with popular ASI1600 quite fast.

P.S.

I noticed, guys who image with Hyperstar or Rasa, -  almost all go for imx 183 sensor.

 

 

Thank you for the reply. The camera was bought used, so it was a way for me to get into mono imaging, which I am still very much learning.

I have read several reviews of various F4 reflectors, and in general if you go for some of the better ones (TS UNC comes to mind) they are not to bad as long as you stay at F4 or above. When you start pushing F3 things start to get real "interesting". I don't mind tinkering to be honest, it would be a 2nd scope, so it wouldn't necessarily stop me from imaging if it was acting up. But as I don't have a permanent setup, your point is very valid, and something I need to think about before purchasing.

Unfortunately it appears my lil SW 80ED might not be one of the "good ones". Which has lead me to think about replacing it. What ever I do I think I'm done with cheap SW scopes. ?

[vlaiv]

I think that actually 8" F/3.9 scope paired with ASI183 gives good combination.

While it might look like it is oversampling, you don't necessarily need to look at it that way.

It depends on your mount/guiding performance. If you can get your guiding to be 0.6" RMS or less then camera / scope pair is a good match.

Sure it looks "oversampled" on the paper - with resolution of 0.62"/pixel. But ASI183 is high resolution (in this context I mean lots of mega pixels ) and you can easily afford to bin your images - either classical bin x2 in software or maybe utilize approach I outlined in above post (I'll explain how to do it below).

If you bin it x2 you will get resolution ~1.24"/pixel - and that is quite good resolution for 8" provided you can guide well (with 0.6" or below RMS).

Think of it this way, can you find CCD on market that has 2748x1836 with pixel size of 4.8um, 16K Full well, that will give you with that scope resolution of 1.24"/pixel and have read noise of 4.4e?

I think you will be hard pressed to find it, and certainly at that price point. If you bin your subs with ASI183 (unity gain) and that scope - that is exactly what you will get.

On topic of binning - bin your images 2x2 in software after you calibrate them and before stacking. Make sure you use 32bit precision for both calibration and binning. After that stack as normal.

If you want to try out "binning" explained in above mentioned thread - simplest way to do it is to "split" each calibrated sub into four "sub-subs". Then stack resulting subs as you normally would (you will have 4 times as many subs, but each will be 4 times smaller in size - 2x height, 2x width).

Split is done simply by taking every second pixel in X and Y and putting it together (first sub-sub will have X=0,2,4,6,... and Y=0,2,4,6,... ; second sub-sub will have odd X indices and even Y like X=1,3,5,7, ... and Y=0,2,4,6, ...., third will have Y odd, X even, and fourth will have both X and Y with odd indices).

Maybe image is worth 1000 words ... :

(yellow pixels marked with 1 will end up in first sub-sub, blue marked with 2 will be in second sub-sub, red marked 3 and green marked 4 in their respective sub-subs).

If you use PI I believe you can do pixel math to split images like that, and if not, we can do simple ImageJ plugin that will do that for you (I intend to write both simple split, and split and combine).

In the end, use stacking that will do Lanczos interpolation when aligning / registering images for best results (I think PI has it if you use that. Not sure if DSS has it - I think it uses bilinear or bicubic filtering, but not sure).

[geordie85]

48 minutes ago, vlaiv said:

I think that actually 8" F/3.9 scope paired with ASI183 gives good combination.

While it might look like it is oversampling, you don't necessarily need to look at it that way.

It depends on your mount/guiding performance. If you can get your guiding to be 0.6" RMS or less then camera / scope pair is a good match.

Sure it looks "oversampled" on the paper - with resolution of 0.62"/pixel. But ASI183 is high resolution (in this context I mean lots of mega pixels ) and you can easily afford to bin your images - either classical bin x2 in software or maybe utilize approach I outlined in above post (I'll explain how to do it below).

If you bin it x2 you will get resolution ~1.24"/pixel - and that is quite good resolution for 8" provided you can guide well (with 0.6" or below RMS).

Think of it this way, can you find CCD on market that has 2748x1836 with pixel size of 4.8um, 16K Full well, that will give you with that scope resolution of 1.24"/pixel and have read noise of 4.4e?

I think you will be hard pressed to find it, and certainly at that price point. If you bin your subs with ASI183 (unity gain) and that scope - that is exactly what you will get.

On topic of binning - bin your images 2x2 in software after you calibrate them and before stacking. Make sure you use 32bit precision for both calibration and binning. After that stack as normal.

If you want to try out "binning" explained in above mentioned thread - simplest way to do it is to "split" each calibrated sub into four "sub-subs". Then stack resulting subs as you normally would (you will have 4 times as many subs, but each will be 4 times smaller in size - 2x height, 2x width).

Split is done simply by taking every second pixel in X and Y and putting it together (first sub-sub will have X=0,2,4,6,... and Y=0,2,4,6,... ; second sub-sub will have odd X indices and even Y like X=1,3,5,7, ... and Y=0,2,4,6, ...., third will have Y odd, X even, and fourth will have both X and Y with odd indices).

Maybe image is worth 1000 words ... :

(yellow pixels marked with 1 will end up in first sub-sub, blue marked with 2 will be in second sub-sub, red marked 3 and green marked 4 in their respective sub-subs).

If you use PI I believe you can do pixel math to split images like that, and if not, we can do simple ImageJ plugin that will do that for you (I intend to write both simple split, and split and combine).

In the end, use stacking that will do Lanczos interpolation when aligning / registering images for best results (I think PI has it if you use that. Not sure if DSS has it - I think it uses bilinear or bicubic filtering, but not sure).

I've been reading this thread with great interest and I'd very much like to try this, but I must admit I'm confused. 

I use astroart 6 and I'm confused how to calibrate subs and split them before stacking?? 

Are you able to explain this in simpleton terms? (I'm not very tech smart) 

[vlaiv]

55 minutes ago, geordie85 said:

I've been reading this thread with great interest and I'd very much like to try this, but I must admit I'm confused. 

I use astroart 6 and I'm confused how to calibrate subs and split them before stacking?? 

Are you able to explain this in simpleton terms? (I'm not very tech smart) 

I usually forget that most people are using ready made stacking solutions. It is simple matter of calibrating your images your self and then just stack them in your favorite software but without specifying calibration files (tell software not to do calibration - it might complain - DSS for example is warning you that you should include darks, bias, etc ... but it will stack without those).

Calibration can be done in ImageJ (or Fiji - both are the same thing, open source java based scientific image processing suite) for example. I'll create separate thread (tomorrow, as it's getting late to do it all now) in image processing and explain basic calibration methods for CMOS fits files there (darks, flats, flat darks, ....) as well as some "advanced" techniques - like this binning. Those will require custom plugins, so I'll upload those for anyone interested to use, and explain how to use them.

I'll post link to new thread here as well.

[jjosefsen]

2 hours ago, vlaiv said:

I think that actually 8" F/3.9 scope paired with ASI183 gives good combination.

While it might look like it is oversampling, you don't necessarily need to look at it that way.

It depends on your mount/guiding performance. If you can get your guiding to be 0.6" RMS or less then camera / scope pair is a good match.

Sure it looks "oversampled" on the paper - with resolution of 0.62"/pixel. But ASI183 is high resolution (in this context I mean lots of mega pixels ) and you can easily afford to bin your images - either classical bin x2 in software or maybe utilize approach I outlined in above post (I'll explain how to do it below).

If you bin it x2 you will get resolution ~1.24"/pixel - and that is quite good resolution for 8" provided you can guide well (with 0.6" or below RMS).

Think of it this way, can you find CCD on market that has 2748x1836 with pixel size of 4.8um, 16K Full well, that will give you with that scope resolution of 1.24"/pixel and have read noise of 4.4e?

I think you will be hard pressed to find it, and certainly at that price point. If you bin your subs with ASI183 (unity gain) and that scope - that is exactly what you will get.

On topic of binning - bin your images 2x2 in software after you calibrate them and before stacking. Make sure you use 32bit precision for both calibration and binning. After that stack as normal.

If you want to try out "binning" explained in above mentioned thread - simplest way to do it is to "split" each calibrated sub into four "sub-subs". Then stack resulting subs as you normally would (you will have 4 times as many subs, but each will be 4 times smaller in size - 2x height, 2x width).

Split is done simply by taking every second pixel in X and Y and putting it together (first sub-sub will have X=0,2,4,6,... and Y=0,2,4,6,... ; second sub-sub will have odd X indices and even Y like X=1,3,5,7, ... and Y=0,2,4,6, ...., third will have Y odd, X even, and fourth will have both X and Y with odd indices).

Maybe image is worth 1000 words ... :

(yellow pixels marked with 1 will end up in first sub-sub, blue marked with 2 will be in second sub-sub, red marked 3 and green marked 4 in their respective sub-subs).

If you use PI I believe you can do pixel math to split images like that, and if not, we can do simple ImageJ plugin that will do that for you (I intend to write both simple split, and split and combine).

In the end, use stacking that will do Lanczos interpolation when aligning / registering images for best results (I think PI has it if you use that. Not sure if DSS has it - I think it uses bilinear or bicubic filtering, but not sure).

Excellent points about the camera's, sometimes I forget that the AstroCam makers are using sensors meant for other applications, so sensor specs aren't tailored to our needs as such.

The guiding might be tough, I can get as low as 0.5" on my EQ6 sometimes, but that is with a vastly smaller scope riding on it.. Ususally it is somewhere around 0.7" RMS.

If I decided to bin 2x2 in acquisition (for simplicity's sake) would my guiding need to be around 0.6" RMS, If I understand it, I wouldn't have to in that case, as my effective pixel scale is 1.24"/pixel as you mention. I am still wrapping mu head around the post acquisition software binning you propose, so not sure if the same would apply.. Im guessing not..?

[vlaiv]

1 minute ago, jjosefsen said:

Excellent points about the camera's, sometimes I forget that the AstroCam makers are using sensors meant for other applications, so sensor specs aren't tailored to our needs as such.

The guiding might be tough, I can get as low as 0.5" on my EQ6 sometimes, but that is with a vastly smaller scope riding on it.. Ususally it is somewhere around 0.7" RMS.

If I decided to bin 2x2 in acquisition (for simplicity's sake) would my guiding need to be around 0.6" RMS, If I understand it, I wouldn't have to in that case, as my effective pixel scale is 1.24"/pixel as you mention. I am still wrapping mu head around the post acquisition software binning you propose, so not sure if the same would apply.. Im guessing not..?

Guide RMS being half of imaging resolution is just a guideline - you want your guide RMS to be as small as possible

If it is a bit over that value just go with that - it will not change much. It is not up to guiding only. Resolution needed to capture all details in image available depends on "total blur" introduced by seeing and guiding/tracking errors and aperture size. Another good guideline to tell you if you should be binning is to measure star FWHM on your subs - you want your resolution to be about 0.622 of star FWHM. Don't worry too much if it's slightly below or above that value.

Here is an example - you image at 0.62"/pixel and you take your sub, and measure average star FWHM on it to be 2.4" (or 3.87 pixels). What would be good resolution for such star FWHM? According to above guideline 2.4" * 0.622 = 1.4928"/pixel - so with bin x2 you will still be slightly oversampling - but not as bad as with 0.62"/pixel - and by binning you will increase your SNR by factor of 2. What about case if your measure FWHM is 3.2"? Then appropriate resolution would be 2"/pixel (3.2*0.622), so you can bin x3 in this case to have 1.86"/pixel. Image will be smaller but it will still contain all the information that could be captured under given circumstances (poor seeing for example). Bonus is SNR increase by factor of 3.

If you don't feel comfortable with binning in software - then bin while capturing. I don't like that idea particularly, but that is just because I don't know exact process of binning in capture - what software/firmware does to bin values. Is it done in camera hardware prior to ADC (bad thing, as you end up with 12bit value, and it should be 14 bit value when binning 2x2), or does driver do it after downloading it from camera (then only drawback that I see is maybe some very small error for flat calibration - although not sure 100% it will happen - need to check the math). Maybe it is in firmware after ADC so you end up with 14bit value after all - just don't know what happens and that is the reason I'm not fond of that idea very much.

[jjosefsen]

4 minutes ago, vlaiv said:

Guide RMS being half of imaging resolution is just a guideline - you want your guide RMS to be as small as possible

If it is a bit over that value just go with that - it will not change much. It is not up to guiding only. Resolution needed to capture all details in image available depends on "total blur" introduced by seeing and guiding/tracking errors and aperture size. Another good guideline to tell you if you should be binning is to measure star FWHM on your subs - you want your resolution to be about 0.622 of star FWHM. Don't worry too much if it's slightly below or above that value.

Here is an example - you image at 0.62"/pixel and you take your sub, and measure average star FWHM on it to be 2.4" (or 3.87 pixels). What would be good resolution for such star FWHM? According to above guideline 2.4" * 0.622 = 1.4928"/pixel - so with bin x2 you will still be slightly oversampling - but not as bad as with 0.62"/pixel - and by binning you will increase your SNR by factor of 2. What about case if your measure FWHM is 3.2"? Then appropriate resolution would be 2"/pixel (3.2*0.622), so you can bin x3 in this case to have 1.86"/pixel. Image will be smaller but it will still contain all the information that could be captured under given circumstances (poor seeing for example). Bonus is SNR increase by factor of 3.

If you don't feel comfortable with binning in software - then bin while capturing. I don't like that idea particularly, but that is just because I don't know exact process of binning in capture - what software/firmware does to bin values. Is it done in camera hardware prior to ADC (bad thing, as you end up with 12bit value, and it should be 14 bit value when binning 2x2), or does driver do it after downloading it from camera (then only drawback that I see is maybe some very small error for flat calibration - although not sure 100% it will happen - need to check the math). Maybe it is in firmware after ADC so you end up with 14bit value after all - just don't know what happens and that is the reason I'm not fond of that idea very much.

Thanks for explaining it so well!

Im afraid I am very limited by my seeing, my FWHM is usually around 2.8 - 3.2 (that I can remember from memory atleast).. I often have quite high humitiy, possible because of my proximity to the coast.

Bookmarking this for future reference.

[Stub Mandrel]

With slightly oversampled  DSLR images you have two options:

1. Bin 2x2
2. Use super-pixel mode where each pixel is derived from a Bayer unit and gets red from one pixel, blue from one pixel and green as the average of the other two pixels.

Which is likely to give the best results?

[vlaiv]

9 minutes ago, Stub Mandrel said:

With slightly oversampled  DSLR images you have two options:

1. Bin 2x2
2. Use super-pixel mode where each pixel is derived from a Bayer unit and gets red from one pixel, blue from one pixel and green as the average of the other two pixels.

Which is likely to give the best results?

IMO, best approach with oversampled DSLR is to use sort of super-pixel mode. Split bayer pattern into 4 images rather than 3. Red, Blue and two frames of Green (instead of averaging pixels - you assign them to two different images, same as you would do with red or blue - just call it green1 and green2), and then stack and process each channel in similar way to regular RGB imaging. This will give you two times more green subs in the end, but when you stack them, resulting SNR will be same as averaging each 2 pixels of green in one single pixel (times sqrt(2)), but without artifacts related to binning (slight loss of resolution).

[vlaiv]

As I promised, here is small tutorial for calibration and software binning of fits files

(works on OSC as well - everything apart from binning - I'll add more "tutlets" - just a name I coined for small tutorials - later, as well as debayer plugin like discussed above, this new form of binning, and some more goodies ... )