Jump to content

  •  

CNers have asked about a donation box for Cloudy Nights over the years, so here you go. Donation is not required by any means, so please enjoy your stay.

Photo

Nikon D810A Review

dslr astrophotography solar equipment
  • Please log in to reply
148 replies to this topic

#101 Herra Kuulapaa

Herra Kuulapaa

    Viking 1

  • -----
  • Posts: 509
  • Joined: 10 Dec 2014
  • Loc: Vantaa, Finland

Posted 08 July 2015 - 03:34 PM

There's something wrong with the updated temperature range link I guess. First series was pretty limited, but the second covers much larger temperature range.

#102 whwang

whwang

    Mercury-Atlas

  • *****
  • Posts: 2622
  • Joined: 20 Mar 2013

Posted 08 July 2015 - 03:52 PM

Thanks for your note, Wei-Hao!

 

I use PixInsight for data analysis. PixInsight uses DCRAW to load NEF files. 
In PixInsight's options for DSLR_RAW files I use "Pure RAW" options set, it should "Create raw Bayer CFA image" without "Automatic color balance" and "Camera white balance", this options turned OFF.

 

With this, I believe that I get the original raw data from the sensor :)

Specifically, DCRAW is called by PixInsight with following options

      dcraw -4 -o 0 -D -t 0 -k 0 -H 1

where -D means "Show the raw data as a grayscale image with no interpolation, with the original unscaled pixel values."
 

In other words, I'm pretty confident that I work with raw data from the sensor.

And gains are different for different colors naturally, by electronic design of camera amplifier, without relation to any additional "color balance" that could be applied later.

I think that if you will make CFA splitting in 4 channels (R,G1,G2,B) and measure gain separately for them you will see results exact or very close to mine.

 

...with this I can be wrong somewhere, even if I've tried to do my best...

 

BTW, Wei-Hao, could you please give some details (or link?) on how do you estimate a dark current having in mind that raw files are preprocessed and their mean values are shifted toward 600ADU point at any exposure duration?

 

Yours,

    Yuriy

 

Thanks, Yuriy.  This is my bad.  I take a look at the data again.  D810A indeed has different gains in the four channels.  On my D800, the gains in the four channels are indistinguishable.  So I naively assumed this should also be the case on D810A without actually checking it.  Obviously I was wrong.   This is an interesting new thing to learn.

 

I never directly measure dark current, as this is always cursed by whatever raw cooking Nikon/Canon/Sony does.  What I do is to measure dark noise.  I collect many dark frames after the camera warms up.  Then on each pixel, I measure the standard deviation of its values across the many exposures.  This gives dark noise of that pixel, after removing the contribution from read noise (measured in a similar way).  I repeat this calculation on all the pixels, to construct dark noise of the whole sensor.  With the measured dark noise, I can convert it to "equivalent" dark current" by assuming that dark noise follows Poisson statistics.  

 

On an non-hacked D800, the measurement of standard deviation in the above procedure doesn't quite work, because D800's data is severely clipped.  We lose more than half of the noise distribution and standard deviation gives very skewed results.  This is no longer the issue with the firmware hack.  In principle, the hack should also enable direct measurement of dark current without using dark noise, but my personal preference is still to measure noise, since noise is what we ultimately care.

 

On D810A, its data are also clipped, but the clipping is far less severe than D800.  At least at ISO 800, simple standard deviation gives very good estimate of noise.  

 

Finally, although I am not a fan of the two image subtraction method, my guess is that it should work OK on D810A.  You should be able to measure noise on a subtracted dark pair from D810A, just like what you did with the gain measurement.

 

Cheers,

Wei-Hao


  • Jon Rista likes this

#103 sharkmelley

sharkmelley

    Vanguard

  • *****
  • Posts: 2144
  • Joined: 19 Feb 2013

Posted 08 July 2015 - 04:07 PM

If you put the raw files into the PixInsight script I wrote, it should report the percentage of pixels clipped to zero.

http://www.cloudynig...tics/?p=6670169

 

The D810A would be a good test case for it.

 

Although the gain, read noise and dark current calculations will be affected, the percentage of clipping should give some idea of how reliable or unreliable the results are.

 

Mark


  • Jon Rista likes this

#104 YuriyT

YuriyT

    Sputnik

  • -----
  • Posts: 46
  • Joined: 22 Jul 2006
  • Loc: Moscow, Russia

Posted 08 July 2015 - 04:18 PM

... What I do is to measure dark noise.  I collect many dark frames after the camera warms up.  Then on each pixel, I measure ...

Wei-Hao, thanks a lot for an extensive comment! 
Will read through it again and again.
I had the same idea to go for measurement of the dark current through its noise, will try to figure out how to do it in practical way.
Thanks a lot,

    Yuriy



#105 Jon Rista

Jon Rista

    ISS

  • *****
  • Posts: 22885
  • Joined: 10 Jan 2014
  • Loc: Colorado

Posted 08 July 2015 - 05:06 PM

If you put the raw files into the PixInsight script I wrote, it should report the percentage of pixels clipped to zero.

http://www.cloudynig...tics/?p=6670169

 

The D810A would be a good test case for it.

 

Although the gain, read noise and dark current calculations will be affected, the percentage of clipping should give some idea of how reliable or unreliable the results are.

 

Mark

 

Mark, is there any chance you could integrate an option to evaluate dark current noise via Wei-Hao's approach? Where he evaluates per-pixel dark current NOISE by comparing the same pixel across many dark frames? I like that approach, in that it compares changes in the same pixel (and does that comparison for each pixel...or maybe a sampling of pixels if you don't want to spend a huge amount of time processing a billion pixels. :p



#106 sharkmelley

sharkmelley

    Vanguard

  • *****
  • Posts: 2144
  • Joined: 19 Feb 2013

Posted 08 July 2015 - 05:37 PM

 

If you put the raw files into the PixInsight script I wrote, it should report the percentage of pixels clipped to zero.

http://www.cloudynig...tics/?p=6670169

 

The D810A would be a good test case for it.

 

Although the gain, read noise and dark current calculations will be affected, the percentage of clipping should give some idea of how reliable or unreliable the results are.

 

Mark

 

Mark, is there any chance you could integrate an option to evaluate dark current noise via Wei-Hao's approach? Where he evaluates per-pixel dark current NOISE by comparing the same pixel across many dark frames? I like that approach, in that it compares changes in the same pixel (and does that comparison for each pixel...or maybe a sampling of pixels if you don't want to spend a huge amount of time processing a billion pixels. :p

 

 

Yes, that would be fairly straightforward to implement and I quite like the idea of that approach. However, I'm a bit tied up for the next couple of weeks and probably won't get a chance to look at it until the end of the month.

 

I'm certainly interested to see the difference between the "Wei-Hao" method and the "2 frame subtraction" method the script  currently uses for calculating noise.

 

Mark


Edited by sharkmelley, 08 July 2015 - 05:47 PM.


#107 whwang

whwang

    Mercury-Atlas

  • *****
  • Posts: 2622
  • Joined: 20 Mar 2013

Posted 08 July 2015 - 07:26 PM

Hi,

 

The attached graph can show how much D810A's dark is clipped.  I sent this graph to Jerry a few days ago.

 

What you see is the histogram of pixel values of 5-minute ISO 800 darks from a hacked D800 (top, 25 deg C) and D810A (bottom, 30 deg C), both unsubstracted for bias. Here you clearly see the clipping of D810A's data at about 570 ADU.  There are also gaps in the data.  Here I also did not convert ADU into electrons.  Since the 4 channels of D810A have different gains, converting the data into electron unit may give cleaner histograms.

 

The mean value is at about 600 ADU in D810A, as oppose to 220 ADU in the hacked D800.  This is artificial, as D810A's mean is shifted by the camera's raw processing.  If you look at the width of the distribution, which is an indicator of dark noise, you will realize D810A's dark noise is much lower than D800's (even at a higher temperature).

 

At this ISO and fairly high temperature, D810A's dark is not clipped severely.  I fitted a Gaussian to the distribution, the result is not very different from directly running a standard deviation of the clipped data.  From this point of view, directly running standard deviation on a subtracted image pair should still work.

 

I am happy to distribute my source code for calculating noise pixel-wise on several images.  Unfortunately, my calculation is made in a commercial package called IDL.  This package is probably too expansive for most of us to afford.  So giving people my source code will not help much.  Sorry about this.  You guys will need to wait for Mark to implement this in his PixInsight script.

 

Cheers,

Wei-Hao

Attached Thumbnails

  • ISO800_dark.png


#108 whwang

whwang

    Mercury-Atlas

  • *****
  • Posts: 2622
  • Joined: 20 Mar 2013

Posted 08 July 2015 - 08:02 PM

Hi,

 

After learning that D810A has different gains in R, G, and B, I re-calculated the gains for ISO 200, 400, 800, 1600, and 3200.  The results can be well described by the following formula:

  gain_R = ISO * 0.0016427

  gain_G = ISO * 0.0027025
  gain_B = ISO * 0.0034504

 

These are in ADU/e-.

 

The read noise I measured at ISO 200, 400, 800, 1600, 3200 are 

4.477,   3.997,   3.207,   2.813,   2.397  electrons.

 

For comparison, the gain of D800 can be described as

  gain = ISO * 0.002964

Read noise at ISO 400, 800, 1600 are

2.379,  2.165,  2.104 electrons.

 

D800 has significantly better read noise, which helps short subs under dark sky.

 

More interesting things to come.  Stay tuned.

 

Cheers,

Wei-Hao


  • sharkmelley likes this

#109 Plane

Plane

    Explorer 1

  • -----
  • Posts: 97
  • Joined: 09 Dec 2009

Posted 08 July 2015 - 08:11 PM

 

What you see is the histogram of pixel values of 5-minute ISO 800 darks from a hacked D800 (top, 25 deg C) and D810A (bottom, 30 deg C), both unsubstracted for bias. Here you clearly see the clipping of D810A's data at about 570 ADU.  There are also gaps in the data.  Here I also did not convert ADU into electrons.  Since the 4 channels of D810A have different gains, converting the data into electron unit may give cleaner histograms.

 

The mean value is at about 600 ADU in D810A, as oppose to 220 ADU in the hacked D800.  This is artificial, as D810A's mean is shifted by the camera's raw processing.  If you look at the width of the distribution, which is an indicator of dark noise, you will realize D810A's dark noise is much lower than D800's (even at a higher temperature).

 

At this ISO and fairly high temperature, D810A's dark is not clipped severely.  I fitted a Gaussian to the distribution, the result is not very different from directly running a standard deviation of the clipped data.  From this point of view, directly running standard deviation on a subtracted image pair should still work.

 

 

Hi Wei-Hao,

 

Thanks for the information. I was looking at the your D800 dark current. (220 - 128) / 300 = 0.30667 ADU/ps. 128ADU was the bias level I put in the firmware. 

Base on the gain at ISO800 2.4466 ADU/e- you sent me a while ago, this will give you a dark current of 0.125eps at 25 C. This is very close to my 0.14eps for IMX071 at room temperature given temperature inaccuracy.

 

As for D810A, I would still tend to hold back on saying it has superior dark signal performance as we are comparing an processed image vs a RAW sensor data of D800. Based on some report done by ccs_hello, it appears the internal preprocessing will suppress the deviance in long exposure not just the hot pixels. 

 

Best regards,

astronomer



#110 whwang

whwang

    Mercury-Atlas

  • *****
  • Posts: 2622
  • Joined: 20 Mar 2013

Posted 08 July 2015 - 08:42 PM

 

Hi Wei-Hao,

 

Thanks for the information. I was looking at the your D800 dark current. (220 - 128) / 300 = 0.30667 ADU/ps. 128ADU was the bias level I put in the firmware. 

Base on the gain at ISO800 2.4466 ADU/e- you sent me a while ago, this will give you a dark current of 0.125eps at 25 C. This is very close to my 0.14eps for IMX071 at room temperature given temperature inaccuracy.

 

As for D810A, I would still tend to hold back on saying it has superior dark signal performance as we are comparing an processed image vs a RAW sensor data of D800. Based on some report done by ccs_hello, it appears the internal preprocessing will suppress the deviance in long exposure not just the hot pixels. 

 

Best regards,

astronomer

 

 

 

Do you mean there is some spatial filtering in D810A?  If so, we do not only need to worry about dark, even the on-sky exposures will be damaged.  I personally do not care about hot pixels.  They are sparse, and can be easily removed with dithering plus sigma rejection in the stacking, even without dark subtraction.  What I care is the dark current in the majority of pixels.  It is hard to imagine Nikon can come up with some scheme to suppress the structure in those cooler pixels.

 

The files I used for D800's dark measurement come from the USB hack, not the firmware hack.  Is the 128 ADU for the firmware hack?

 

Cheers,

Wei-Hao



#111 Plane

Plane

    Explorer 1

  • -----
  • Posts: 97
  • Joined: 09 Dec 2009

Posted 08 July 2015 - 10:05 PM

 

Do you mean there is some spatial filtering in D810A?  If so, we do not only need to worry about dark, even the on-sky exposures will be damaged.  I personally do not care about hot pixels.  They are sparse, and can be easily removed with dithering plus sigma rejection in the stacking, even without dark subtraction.  What I care is the dark current in the majority of pixels.  It is hard to imagine Nikon can come up with some scheme to suppress the structure in those cooler pixels.

 

The files I used for D800's dark measurement come from the USB hack, not the firmware hack.  Is the 128 ADU for the firmware hack?

 

Cheers,

Wei-Hao

 

 

There's always spatial filter on the processioning part, not just filtering the hot pixels. It is apparent in dark frame with preprocessing on/off. It's hard to prove how much it differs but we are getting close to extracting the high speed signal out from the image sensor. Thus we can compare with the RAW array on file simultaneously.

 

If you only used the USB hack, the bias should be 600ADU. Did you already subtracted the master bias so 220 is the actual dark counts?

 

Best,

astronomer


Edited by Plane, 08 July 2015 - 10:11 PM.


#112 whwang

whwang

    Mercury-Atlas

  • *****
  • Posts: 2622
  • Joined: 20 Mar 2013

Posted 08 July 2015 - 10:23 PM

OK.  It is official.  D810A has the same H alpha sensitivity as a modified D800.

 

Last night I took images of M57 using my modified D800 and a D810A from Maunakea.  With each of them, I took 21 10-sec exposures, and somewhat larger numbers of darks.  I placed them as close to the image center as possible (using Vega to sync the goto).  Given that they are both very close to the image center, the images are not flat-field corrected.

 

The attached image shows the stacked R channel from D800, D810A, and the ratio.  Before the ratio image is created, both images are subtracted for sky background using background regions around M57.  The image brightness scales are also corrected using their gains, so the final images are in electron unit.  The weather was good, and there were no clouds.  The D800 images were taken right before M57 passed the meridian and the D810A images right after, so they were taken under conditions that are practically identical.

 

The color bar in the bottom of the image shows the scale in the ratio map.  You can see that in the M57 region, the ratio is very close to 1.0.  Indeed, when I move my cursor around the image with the M57 ring, all I can read is 1.0x or 0.9x.  This means that the H alpha quantum efficiency of D810A and modified D800 are basically the same.

 

Precisely speaking, without using narrow-band filters, we are not seeing just H alpha.  There are also [NII] and [SII] lines coming in.  So what we are seeing here are averaged QE at the locations of these three lines, weighted by the relative line intensities in M57.  I know that the filter in my modified D800 has a transmission of 48% at the wavelength of [SII] relative to that at H alpha. So in an extreme scenario, D810A may transmits more light from [SII] to compensate its lower H alpha transmission.  However, [SII] is not that strong even in planetary nebulas like M57.  So even if there is such a compensation effect, it should be small.  Practically, we can still conclude that D810A is as good as a modified D800 at H alpha.

 

So, never use Canon's failures to judge D810A.  It's not 20Da nor 60Da.  It's as good as a modified D800.

 

Cheers,

Wei-Hao

Attached Thumbnails

  • ds9.jpg

  • Phil Jones, Ohan Smit, sharkmelley and 1 other like this

#113 whwang

whwang

    Mercury-Atlas

  • *****
  • Posts: 2622
  • Joined: 20 Mar 2013

Posted 08 July 2015 - 10:29 PM

 

If you only used the USB hack, the bias should be 600ADU. Did you already subtracted the master bias so 220 is the actual dark counts?

 

 

Apologize for this.  After looking at my code again, now I believe the D800 one has bias subtracted.  I did this quite a while ago, so I kind of forget the details.


Edited by whwang, 08 July 2015 - 10:33 PM.

  • Plane likes this

#114 sharkmelley

sharkmelley

    Vanguard

  • *****
  • Posts: 2144
  • Joined: 19 Feb 2013

Posted 09 July 2015 - 12:26 AM

OK.  It is official.  D810A has the same H alpha sensitivity as a modified D800.

 

 

Excellent news!  That's a very useful analysis.

 

Mark



#115 whwang

whwang

    Mercury-Atlas

  • *****
  • Posts: 2622
  • Joined: 20 Mar 2013

Posted 10 July 2015 - 07:00 PM

Hi,

 

This is another test, to see if D810A's files are linear enough to allow for accurate flat-field correction.

 

In the attached image, the left two are from D800 without the firmware hack.  The right two are from D810A.  All are flattened and strongly stretched linearly.  In each case, a pass of background subtraction is made with PixInsight's ABE function with a simple 1st order function, to remove the sky gradient.

 

You can see that in the D800 ones, the vignetting in the corners and also the bottom (caused by the mirror) of the images are not perfectly corrected.  Additional to vignetting, the imperfect flat field plus the sky gradient cause very strange color pattern in the image background.  The two images here basically cannot be mosaicked without further correction.  It is very painful to deal with such images.

 

On the other hand, in the D810A images, the background is much flatter.  There are hints of dark corners, but not as severe as that in D800.  In each of the D810A image, you can see a large circular region in the center that is probably darker than what it should be, especially in the G channel (so that region appears more magenta).  This is likely flat field artifact caused by nonlinear data.

 

None of the above flat field artifact show up in a hacked D800.  In terms of data linearity, I would say that:

hacked D800 > D810A >>> D800

 

In short, D810A's files only show slight hint of nonliearity (likely caused by the data clipping), but it is far better than an unhacked D800.  For those who are not going after insanely faint clouds or not doing mosaics, D810A should be good enough.

 

Cheers,

Wei-Hao

Attached Thumbnails

  • D800_vs_D810A_flat.jpg


#116 DickLocke

DickLocke

    Lift Off

  • -----
  • Posts: 22
  • Joined: 05 Sep 2010

Posted 15 August 2015 - 04:58 PM

Hi all,

 

Very interesting...  Now, where did we end up an the best ISO setting for the D810A?  I saw ISO 400 mentioned.  Is that still the suggestion?  What are people's real-world experiences?

 

Thanks!

 

-Dick



#117 whwang

whwang

    Mercury-Atlas

  • *****
  • Posts: 2622
  • Joined: 20 Mar 2013

Posted 15 August 2015 - 09:10 PM

With my imaging condition (F2.8, dark sky), I like ISO 800.  However, I see no reason why ISO 400 or 1600 does not work for most people.  For slow optics under a dark sky, ISO 1600 should work slightly better.  On the other hand, if you have a fast optics or a bright sky, ISO 400 should work better.

 

Cheers,

Wei-Hao



#118 SunBlack

SunBlack

    Surveyor 1

  • -----
  • Posts: 1769
  • Joined: 05 May 2011
  • Loc: Rome (IT)

Posted 08 September 2015 - 05:37 AM

Read on the attachment

Attached Thumbnails

  • IMG_20150908_123424.jpg


#119 whwang

whwang

    Mercury-Atlas

  • *****
  • Posts: 2622
  • Joined: 20 Mar 2013

Posted 08 September 2015 - 09:04 AM

Based on real pictures I have seen, I have great reservation about the Halpha sensitivity reported in that article. I would love to see the details about the spectrograph and how they calibrate/process the spectra, in order to judge how much I can trust their results.



#120 SunBlack

SunBlack

    Surveyor 1

  • -----
  • Posts: 1769
  • Joined: 05 May 2011
  • Loc: Rome (IT)

Posted 08 September 2015 - 01:47 PM

You expect more Or less than 32%?

#121 sharkmelley

sharkmelley

    Vanguard

  • *****
  • Posts: 2144
  • Joined: 19 Feb 2013

Posted 08 September 2015 - 02:24 PM

I think the figures quoted are sensitivity of H-alpha relative to its peak sensitivity - which is at a wavelength down in the blue.  It's not really a meaningful figure.  It's not a measure of Quantum Efficiency.

 

Mark


  • whwang and Jon Rista like this

#122 CCD1024

CCD1024

    Explorer 1

  • -----
  • Posts: 50
  • Joined: 03 May 2013

Posted 08 September 2015 - 03:54 PM

if I read the article, QE is lower on D810A in red than 60Da but the readout noise is 1.5 e- vs 6 e- so a factor 4x

 

let say 32% 1.5e- vs 46% 6e-

the winner is the D810A which is 3x more sensitive than 60Da !!!



#123 whwang

whwang

    Mercury-Atlas

  • *****
  • Posts: 2622
  • Joined: 20 Mar 2013

Posted 08 September 2015 - 04:03 PM

if I read the article, QE is lower on D810A in red than 60Da but the readout noise is 1.5 e- vs 6 e- so a factor 4x

 

let say 32% 1.5e- vs 46% 6e-

the winner is the D810A which is 3x more sensitive than 60Da !!!

 

Please read the comment from Mark that's just above yours.

 

Also, when the subs are long enough (usually the case when H-alpha is of interest), read noise doesn't impact your image S/N.  The comparison you made is not the right way to compare H-alpha sensitivity.

 

The high read noise of 60Da mainly impact astro-landscape, where subs are hardly long enough.  This problem can be reduced by going to high ISO.  Canon's read noise decreases toward high ISO more dramatically than Sony's.

 

Cheers,

Wei-Hao



#124 Jon Rista

Jon Rista

    ISS

  • *****
  • Posts: 22885
  • Joined: 10 Jan 2014
  • Loc: Colorado

Posted 08 September 2015 - 04:51 PM

Read on the attachment

 

Why is the red channel so under exposed in those tests? It's significantly less exposed than the other two channels...I suspect that has an impact on his Ha sensitivity measures. 



#125 whwang

whwang

    Mercury-Atlas

  • *****
  • Posts: 2622
  • Joined: 20 Mar 2013

Posted 08 September 2015 - 04:55 PM

 

Read on the attachment

 

Why is the red channel so under exposed in those tests? It's significantly less exposed than the other two channels...I suspect that has an impact on his Ha sensitivity measures. 

 

 

 

It all depend on how they normalize the intensity of the spectra in R, G, and B channels.  As Mark pointed out, it is also normalized to the peak of the RGB combined spectra.  Without knowing the intrinsic spectra of their light source and knowing how the spectra are processed/calibrated, plots like these hardly have any values.

 

Cheers,

Wei-Hao




CNers have asked about a donation box for Cloudy Nights over the years, so here you go. Donation is not required by any means, so please enjoy your stay.


Recent Topics





Also tagged with one or more of these keywords: dslr, astrophotography, solar, equipment



Cloudy Nights LLC
Cloudy Nights Sponsor: Astronomics