STT discontinued ?

Hi All,

I am seeing the STT series is discontinued at the dealership pages. Sorry to see this list of cameras go away. Esp the 3200 going back to ST10 days is a loss. Where (other than BSI) does one get this QE at the full well size of 50k ? Nowhere else does one get a combination of KAF3200, guiding filter wheel and AO unit. :-(

Any replacement in sight?

Cheers,
Gert

 

Thanks, Doug.

Much appreciating that you are going to support this sensor.
The STF doesn't seems to support the self guiding filter wheel and AO? (please confirm)

Any knowledge if existing STT stock is on closeout sale anywhere?

Best,
Gert

 

Thank you, Doug.

Much appreciate the push for new sensors.

Please keep the 3200 in your lineup. I'm wondering why it's getting left behind. I see no other sensor combining QE, full well depth, coming close to 3200 noise outside of BSI. Wonder why Kodak / Truesense / On-Semi never copied the pixel from 3200 into the 6303 / 16803 die size.

Cheers,
Gert

 

I started my search for a STT-3200 or -1603 just days after the line was discontinued this past summer -- imagine my dismay to discover that they were impossible to find! I even PM'd Doug to ask if there was any way to produce just one more. I knew the STF line remains available and the additional cooling of the STT may not be necessary, given the spec noise performance; however, the lack of guiding filter wheel and AO in the STF series kept me searching, against the odds. Yesterday, I happened to find a new STT-3200, perhaps the very last one.

I am generating a photon transfer curve that I will post. I cannot believe how uniform the array is; the bias frames are unbelievable -- they look like something one would naively simulate without the usual CCD defects and noise. I am looking forward to first light under the stars.

 

@Jason,

The KAF3200 sensor exhibits one of the best characteristics in the field. Kodak/Truesense/On-Semi should use the pixel technology with larger sensors. I can't figure out why they would not use this technology in a full-frame or even 16803 die size. Equivalent specs are only achieved with BSI sensors i.e. by E2V. Please report your findings on the camera.

Meanwhile I hope Doug will pull through with the announcement of supporting the 3200 in the Aluma series.

Clear Skies,
Gert

 

Initial measurements with ~800 data points and no RBI pre flash are:

• Dark current doubling rate: 6.5°C (spec is 6° nominal)
• J = 1.92 pA/cm**2 at 25°C (spec is < 7 nominal)
• This yields 0.39 e-/pix/sec at 0°C (spec is 0.5 typical) and 0.05 e-/pix/sec at -20°C
• ADC = 0.61 e-/DN (spec is 0.77 for 1x1 binning, will have to investigate as this is the only value that appears out of line)
• Saturation signal = 64,450 DN = 39,300 e- (spec is 2**16 <=> ~50 ke-, need to check ADC value)
• R = 11 DN rms = 6.7 e- rms (spec is 10 e- rms, need to check ADC value)
• Dynamic range = 75 dB (spec is 72 min to > 77 nominal)
• PRNU (FPN) = 1/205 = 0.49% (spec is 1% nominal)
• DSNU = 8 e-/pix/sec (spec is 15 nominal)
• Linearity < 0.5% (spec is 1%)

Some time in the next week, I will repeat with pre flash and I will look more closely at the ADC. I did observe it (correctly) change gain when binning 2x2. It is possible I am picking up another noise source or have not correctly measured the ADC (it's late...!).

All in all, the camera exceeds its specification, both the Kodak data sheet and the SBIG/Diffraction Limited nominal values. On bias frames, it is difficult to differentiate the overscan bias lines from the rest of the array, which is quite remarkable. The FFT of an area of the bias frame is essentially entirely concentrated in the central pixel.

 

Ah, interesting but not surprising in retrospect that the ADC is tuned. Thanks for the info, Doug.

Depending on what I was doing, I measured the ADC as low as 0.46 e-/DN, which is probably much closer to where it really is, as that agrees with the linear fit of the PTC shot noise plot. RBI and/or thermal noise likely led to the ADC result I posted earlier and RBI is likely responsible for the little wiggle in the FPN plot, as I was ramping up and down in integration time at one point, finding it easy to saturate with the high QE. After throwing out some outliers, I've come up with this initial PTC plot. I'll spend a bit more time characterizing the camera with and without RBI pre flash and update the thread in a few weeks.

 

Hi Jason,

Could you quickly just refresh my memory on the PTC creation? How do you separate shot noise frpm FPN? So I recall that you can get SN from the poisson distribution of the signal at given exposure. SN=sqrt(DN*ADC) But where do you extract FPN?

Thanks for publishing the excellent data!

Best,
Gert

 

The simplest method I'm aware of is to take pairs of flat frames with the overscan region at differing signal levels.

A spreadsheet program is very helpful for managing all of the data.

Create columns of the mean and standard deviation of the overscan bias lines only as well as for an evenly illuminated area (say 128 x 128 up to 512 x 512 near the center) of the CCD that doesn't contain many "bad" (dead, nonlinear, warm, hot, etc.) pixels. Use the same pixels in all subsequent frames. Vary the signal level, either by changing the flat field source intensity, varying the integration time, or both. I tend to choose powers of 2 for the sizes of the region of interest because I can then also perform FFT analyses on those pixels to look for other information in the frequency domain.

Collect as many frame pairs as practical, all the way from no apparent illumination to saturation. The more frames one collects, the more statistically meaningful your results. Collecting a dozen or so of frame pairs is useful for a quick check and is easily enough analyzed manually with a program like Maxim DL, which has very handy tools like pixel math, the Graph popup, and the Info popup; for extensive characterization, I have written some C code to perform the analysis. If you can install the CFITSIO library and can compile C code, I am happy to provide the source.

One thing I do is take n, say n = 6, images at the same illumination level instead of just a pair. This allows me to reject an image if necessary; e.g., a cosmic ray in my region of interest. If they're all ok, then I have n*(n-1)/2 unique pairs; e.g., for n = 6, there are 15 unique pairs: 1-2, 1-3, 1-4, 1-5, 1-6, 2-3, 2-4, 2-5, 2-6, 3-4, 3-5, 3-6, 4-5, 4-6, and 5-6.

The mean and standard deviation of the overscan bias should not vary significantly across all of the frames at all illumination levels. This is because this is not a physical pixel, rather a "dummy" pixel that is generated only by the readout circuit. Be certain not to collect statistics on active or other lines in the overscan region. The overscan bias standard deviation is the read noise in [DN rms]. Read noise is not quite Gaussian, but it is generally close enough that the standard deviation of the bias is a meaningful estimate. For best results, take the mean of all of the standard deviations to obtain a better estimate of the read noise in [DN rms]. We will have to wait to convert to [e- rms] until after we can estimate the gain in [e-/DN].

The signal level in a single flat frame is the mean of the region of interest minus the average of the bias for that frame, assuming dark current is negligible (one reason to perform PTC cold and with short integrations). So, take the average of the signal level for each frame pair as an estimate of the signal level.

For each flat frame pair, you have a standard deviation for the illuminated pixels for each image. Take the average of these two values as an estimate of the total noise.

For each flat field frame pair, subtract the images and collect the mean and standard deviation of the subtracted image. The mean is not strictly necessary, ideally it's nearly zero, but if it's more than say 10% of the signal level, you might wish to reject that frame pair. This removes constant FPN (I describe how to estimate it further below). If performing this with Maxim DL pixel math, you need to add an offset to avoid the subtracted pixel values going negative that are then set to zero. Adding the offset is ok because for a set of numbers, x, mean(x)+c = mean(x+c) and the standard deviation is not affected by the offset, since by definition the mean value is subtracted out when computing the standard deviation. What would potentially change is the rms value, because the mean is not removed in its computation. The standard deviation is a measure of the variation, while the rms is a type of average.

OK, we have all the data we need...!

The total noise, measured above, is the rss of shot noise, read noise, and fpn.

The shot+read noise is the standard deviation of the difference frame divided by the sqrt(2); recall that constant FPN is subtracted away in the difference frame. The sqrt(2) arises because even when subtracting images, noise always adds!

The FPN (actually the photo-response non-uniformity, PRNU, since these are light frames) is the square root of ((total noise)**2 - (shot+read noise)**2). Dark signal non-uniformity, DSNU, is computed in a similar manner.

The read noise was already computed.

Thus, the shot noise alone is the square root of ((shot+read noise)**2 - (read noise)**2); alternatively, the square root of ((total noise)**2 - (read noise)**2 - (FPN)**2).

Now plot on a log-log graph, as a function of signal, the total noise, shot noise, FPN, and read noise. Ideally, the shot noise will have a slope of 1/2 and the FPN will have a slope of 1.

This is because the simplified SNR equation for an uncalibrated light frame is SNR = (S + bias) / sqrt(S + bias + B + D + R**2 + FPN**2). The bias is not a true noise source and once its value is known, it should be subtracted from the signal. The signal S, background B (assumed zero), dark current D (assumed negligible), and fixed pattern noise FPN are integration time-dependent while the read noise is the only time-independent term. Further, the FPN = (S+B) * PRNU or, since B = 0, FPN = S * PRNU.

Since the PTC plots signal vs. noise, let's focus on the denominator of the SNR equation.

N = sqrt(S + R**2 + FPN**2)

Thus, in the read noise limit, N ≈ R, in the signal shot noise limit, the noise N ≈ sqrt(S), and in the FPN limit, N ≈ S * PRNU. The three regimes have an exponent of zero, 1/2, and 1 on the signal, which manifest as slopes of zero, 1/2, and 1 on the log-log PTC plot.

The gain [e-/DN] is the x-intercept of a linear fit, y = m x + b, to the shot noise. If the fit is done in log-log space, you will need to use a power law fit, y = b * x**m, since log(y) = m log(x) + log(b). My initial data has a fit of y = 1.4418 x**0.4746. If using a power law fit, be sure to set y = 1 (since log(1) = 0) when solving for x. The exponent 0.4746 is expected to be 1/2 -- pretty close.

The PRNU is 1 divided by x-intercept of a linear fit to the FPN plot. My initial data has a fit y = 0.0053 x**0.9846, so the x-intercept occurs at x = 205. The exponent 0.9848 is expected to be 1 -- again, pretty close.

The saturation signal is where the noise terms no longer increase as signal increases; this occurs near 64,400 DN.

FPN will exceed shot noise when the (S+B+D) > 1/PRNU**2. This occurs at 205**2 = 42,025 DN, about 20,000 DN lower than saturation. So, one thing that looks very promising for this device is that the FPN is not a factor for ~2/3 of the dynamic range and remains essentially negligible until near saturation.

The dynamic range is full well / read noise rms (both in the same units, either DN or e-); to report in decibels, take 20*log10(full well / read noise rms).

I can recommend Richard Crisp's talk on PTC, http://www.narrowbandimaging.com/images/ptc_talk_wsp_2009_crisp_final_comments_web.pdf

Two books that I recommend are Photon Transfer: DN to lambda (aimed at a wide audience) and the phone book-sized Scientific Charge-Coupled Devices (aimed at engineers), both by Jim Janesick.

Hope this is helpful. If I've made any mistakes, I hope someone corrects me!

Jason

 

Ah, yes doing the pair analysis was the trick. Now I remember.
There is actually another nice summary by David Gardner.
http://www.couriertronics.com/docs/...ton_Transfer_Curve_Charactrization_Method.pdf
Please keep your excellent results coming.
Thanks,
Gert

 

The PTC also tells us a very valuable piece of information: what is the minimum integration time I need to be in the shot noise-limited regime?

The single-frame, single-pixel SNR is proportional to integration time t in the read noise-limit, necessitating longer integration to improve SNR; coaddition is of less utility in this regime.

The single-frame, single-pixel SNR is proportional to sqrt(t) in the shot noise-limit; coaddition has maximum benefit. A single 10-min image has the same single-frame, single-pixel SNR as a 2x5-min or 5x2-min frame stack, if the individual images are shot noise-limited. Among the many benefits of coadding is the removal of sensor noise from the background over spatial scales in the final image -- that is to say, the background looks smooth and uniform in an image that consists of many coadded frames. Individual stars and extended sources like nebulae will be just as "bright" in any of the individual frames as well as in the final coadded frame.

At the left end of the PTC plot, we see the total noise curve is dominated by read noise; eventually, the shot noise comes up and dominates the total noise. My rule of thumb is that at a signal level of 1.7 * R**2, which would be 1.7 * 14.6**2 = 360 DN = 165 e-, I can stop integrating and instead consider coadding. If only coadding a very few number of frames, one might want to use 2 or 3 instead of 1.7 to get well away from the shot noise or of course consult your PTC!

For my 4" refractor at f/5.6, I estimate the following G or R filter signal rates with no pixel binning:

• 51 e-/sec, 18 mag/arcsec**2, very bright nebulae, e.g., Ring, integrate for ~3.3 sec
• 8.1 e-/sec, 20 mag/arcsec**2, bright nebulae, e.g., Orion and Lagoon, integrate for ~21 sec
• 1.3 e-/sec, 22 mag/arcsec**2, most galaxies (*), integrate for 280 sec = 4:40 min
• 0.21 e-/sec, 24 mag/arcsec**2, faint nebulae integrate for 13 min
  
• 0.032 e-/sec, 26 mag/arcsec**2, integrated flux nebulae (IFN), galactic halos, integrate for 87 min (not practical with this pixel etendue!; however, 10 minutes at 3x3 is feasible)

Within reason, one can scale these rates for their own aperture, pixel IFOV, QE, and other factors.

Compared to my STT-8300, which is a fine camera (see attached image at f/7 from last fall), the above values are about 3x shorter integration time due to increased QE, improved dynamic range, better uniformity, and lower read noise. Given how infrequently I am able to observe lately, this will be most welcome!

The sky iteslf is 21 - 22 mag/arcsec**2 in most dark sites, yielding rates of 3.2 - 1.3 e-/sec for my refractor. This is additive to the faint object of interest: obtaining enough frames to generate contrast above the mean background (sufficient spatial SNR in the final coadded image) is crucial to detecting extended sources much fainter than the sky itself. One can also see why imaging under a full moon, which yields a sky brightness of about 18 mag/arcsec**2, is so much more challenging, except for narrowband or the very brightest objects.

(*) Interesting fact: most galaxies are ~22 mag/arcsec**2, regardless of distance! This is for the same reason that walls do not get brighter as you walk towards them: the change in solid angle cancels out the change in distance**2, yielding constant radiance regardless of distance. This is why our images even with small telescopes can be littered with background galaxies.

Amazingly, it appears the weather is clearing, so I might get some on-sky images this weekend.

Jason

 

There are no surprises in a quick look at the PTC with RBI pre-flash. I've also attached two bias frames with overscan, with and without RBI pre-flash. - Jason