CFHT STIS2 Detector Information

STIS2: review on the speed optimization

October 1996

Based on the memorandum by J.-C. Cuillandre (download version).

[Postscript file]  (Full document)


  • Abstract
  • 1. Summary of the STIS2 fast readout characterization
  •   1.1 CDS integration time
  •   1.2 Serial register timing
  •   1.3 Gain, linearity and full well capacity
  •   1.4 Charge transfer efficiency
  •     1.4.1 Extended pixel edge response
  •     1.4.2 Impact on the PSF of the pinhole
  •   1.5 The fake overscan in raster readout mode
  •   1.6 Saturated signal response
  • 2 STIS2 data acquisition and reduction
  •   2.1 Overscan
  •   2.2 Bias
  •   2.3 Dark
  •   2.4 Flat-fields
  •   2.5 Cosmetics
  •   2.6 Fringing
  •   2.7 Cosmic rays
  •   2.8 Noise equation
  • 3 Miscellaneous figures
  • 4 Detector summary sheet


      Here follows a short summary of the Stis2 chip characteristics in the ``fast'' readout mode (this mode can be easily selected within the Pegasus session). This mode allows a gain of a factor of more than two (about 2mn30sec instead of 5mn15sec for a full 2Kx2K readout) on the readout time with minor consequences on the data quality. The CTE is (less than) slightly degraded and all the other characteristics are the same as in the standard readout mode (noise, gain,...).

      These results have been obtained with the detector mounted on the telescope and using the MOS internal calibration lamps. This is the best I could do with the test tools available with this setup. I use to do this kind of telescope setup test with the detectors I use for observations in order to confirm the good behavior of the detector and cross-check the results obtained in the laboratory. The results obtained that way have always been very consistent with the laboratory results. Presently, the characteristics measured with the Stis2 standard mode match with the previous measurements obtained by D. McKenna in the Waimea laboratory and confirmed by T. Abbott. The dark current appears actually to be negligible on long exposures.

      According to the set of data collected on the telescope on the 7th and 19th of October (and presented in this memo), the level of confidence is very high for spectroscopic and imaging applications. But new tests in the Waimea laboratory would be welcome to confirm these results and implement new features like the dynamic anti-blooming which would be of great interest when imaging in the I band, or test a very fast readout mode...

      This memo also contains some useful information regarding the data quality and data reduction strategy for the Stis2 data.

      The last pages present various illustrations (images acquired to check the correct behavior of the fast code) and informations about the Stis2 fast readout mode.

    Summary of the STIS2 fast readout characterization

    CDS integration time

    Prior any change in the DSP code, I checked if the CDS (Correlated Double Sampling) integration time (current value 8 microseconds or 160 units as coded in the DSP code) could be reduced (the major time spent during a pixel cycle is due to the CDS operation). I measured the readout noise versus this CDS integration time. The plot of the figure 1 shows that the current value is at the beginning of the plateau, hence its value of 8 microseconds fits well the requirements for low noise readout in spectroscopic applications. Dan McKenna actually found that the readout was increasing when increasing the CDS integration time above 8 microseconds so this is the optimal value for low noise applications.

    Figure 1 Readout noise vs. CDS integration time (160 = 8 microseconds).

    Serial register timing

    The serial timing DSP parameters (STIME, RTIME and CTIME) have been decreased without affecting the noise characteristics. The readout noise is still the same (1.9 ADUs or 9 e-, variance=4.0, the bias level is now 1193 ADUs). The readout speed could be even more increased but then there is an increase (however low) of the readout noise. The present pixel cycle time is 31 microseconds and the figure 1 indicates what readout time could be achieved while keeping a constraint on the readout noise. For example, reducing the CDS integration time by a factor of two microseconds (80 DSP units), the readout noise would be equal to 13 e- and the readout time for the whole 2Kx2K chip would be: (31 - 8)*2048*2088 = 98sec = 1mn40sec.

    Figure 2 Photon transfer curves, the gain is 4.5 e-/ADU.

    Gain, linearity and full well capacity

    A transfer curve was built using the data acquired in the higher part of the dynamic (0-20000 ADUs) and in the lower part of the dynamic (0-6000 ADUs) (see figure 2) The gain is 4.5 e-/ADU, exactly the same as in the standard mode.

    The linearity residual was computed using these data, it is better than 0.6% over the whole ADC dynamic (0-30000 ADUs) and more particularly I checked that the linearity residual is still very low at low flux: it is less than 0.2% below 6000 ADUs. The figure 3 shows the linearity residual plots in different parts of the dynamic.

    There is no shutter balistic correction on these data (raster 200x200 pixels in the center of the chip) as the fitting of the linearity response removes the shutter systematic error. The error introduced by the shutter balistic is an extra exposition delay of about 10ms at the center of the chip.

    The full well capacity matches with the highest values allowed with the controller (about 32000 ADUs), it is about 140000 electrons.

    Figure 3 Linearity residual in different parts of the dynamic.

    Charge transfer efficiency

    1.4.1 Extended pixel edge response

    The CTE was measured using the Extended Pixel Edge Response (EPER) on low level flat-field exposures. The first pixel of the overscan (this overscan information is real for a 2048x2048 pixels readout mode but not in a raster readout mode where pixels between the image area and the overscan are skipped.

    The CTE can be easily computed with the following equation: CTE=1- I_{2049}/(I_{2048} x 2048), where I(n) is the intensity (corrected for the bias value) of the pixel n. This equation is valid only for a 2048x2048 pixels image, the readout adding 40 overscanned pixels. Actually one must add 20 to n because there is 20 prescan pixels in this chip. Note that a lot of pixels along columns must be averaged to reduce the photon noise while measuring I(n).

    The CTE is better than 0.999997 in the fast mode and better than 0.999998 in the standard mode. This difference is negligible and has no effect on the data quality.

    1.4.2 Impact on the PSF of the pinhole

    I checked that the PSF of the pinhole (peak intensity = 14000 ADUs) was the same with the fast and slow modes (using the CFHT IQE tool):

    Fast DSP code:

    Slow DSP code:

    The fake overscan in raster readout mode

    In raster readout mode, a large amount of time can be saved if one skips quickly the pixels between the raster and the overscan. But this results in a change of the video chain behavior and then the overscan looks bad (there is a strong gradient along the lines, see figure 4). In full frame mode (2048x2048 pixels) not one pixel is skipped and then the overscan is acquired in exactly the same conditions as the last pixel of the imaging area, so this overscan can be used to calibrate the camera response during the readout (I insist, only in full frame mode).

    Figure 1 High: overscan region in slow mode. Low: overscan region in fast mode, note the gradient.

    Saturated signal response

    As mentioned before, I didn't detect any real degradation of the CTE. But when checking the video chain edge response with a saturated pinhole bleeding along columns, I saw that it does not go back to its normal low illumination level as quickly as in the standard mode. This effect can be seen on the two images of the figure 5 This should not be a problem as all astrophysic programs use the CCD in the lower part of the dynamic (say N < 20000 ADUs). Compared to the initial code, the degradation of the signal in fast mode due to this effect on bright objects still remains minor.

    Figure 5 Edge response of the video chain. The black lines are saturated columns. High: initial code. Low: fast code. The disturbance is larger when reading fast: it extends on 7 pixels in the slow mode and 20 pixels in the fast mode.

    STIS2 data acquisition and reduction


    In raster readout mode, the overscan presents a strong gradient due to the fast skipping from the end of the raster to the overscan region. Hence, the information contained in the overscan in raster mode should be used with care. This is not a problem with the full frame mode (2048x2048 pixels). As a bias correction is required, the overscan should be simply kept away from the data reduction process.


    The dark current is negligible (less than 1 ADU per 30mn). However during the readout, some noise is injected in the image, probably due to the CCD clocking itself (spurious noise). The structure of the bias is shown on figure 6 proving that a bias correction is required even if the largest amplitude does not exceed 3 ADUs. The overscan won't be of any help to solve this problem. An averaged bias removes efficiently this pattern so a few bias exposures is enough to built a clean bias frame with no cosmic events. Take about 9 bias exposures (at least...) at the beginning and the end of each night. The averaged bias could also be smoothed (function boxcar in Iraf with a 1*n pattern to avoid spreading effects on hot columns) to keep only the large scale structures. The bias has proven to be very stable in time so all bias from an observing run can be combined together. The bias level is 1193 ADUs.

    Figure 6 High: structure of the 2Kx2K bias frame, the largest difference does not exceed 3 ADUs. Low: histogram of a 2Kx2K bias frame, the standard deviation is 1.9 ADUs.


    The Stis2 chip does not suffer from dark current at the operating temperature -110 degrees Celcius. The dark current is uniform on the whole chip (no pattern) and lower than 1 ADU on a 30mn exposure.


    The linearity domain (0.6%) of this camera is 0-30000 ADUs, so dome flats and twilight flats level should not exceed 20000 ADUs. 15000 ADUs would be a good compromise to get high S/N flat-fields while being in the perfect linearity domain of the camera.


    The major defect is a group of bad columns (from X=858 to X=861 from the line Y=793 to Y=2048) due to a cluster of bad pixels. So use telescope shifts along lines larger than this size (6 pixels) when imaging with the shift-and-add technique to efficiently remove this defect while combining your images (on MOS, with a 0.44''/pixel scale, this technique requires at least 3'' shifts toward East or West). Probably due to some interactions between the CCD and cosmic rays, some hot spots can appear during an observing run: during the last run (October 1996) a hot pixel appeared at (1817,43) adding about 300 ADUs to the pixels passing over this site. Also the summing well on the opposite side of the output amplifier started generating charges but as it appears in the 20th column of the overscan, this is not a concern.


    Due to bright and narrow emission lines in the night sky spectrum, Stis2, as most of the thinned chip, suffers from fringing when imaging in the red part of the optical spectrum (above 7000Å). Figure 7 (high) shows a ``sky'' image in the I band (8300Å), obtained by stacking together several fields to remove the objects. The fringing pattern amplitude is about 0.3% of the sky background. A careful correction is necessary but is complicated by the fact that the fringing pattern and amplitude may change during the night. Twilight flat fields (in band I) should be used with high care as the sky spectrum differs strongly from the night sky spectrum, resulting in a different fringing pattern.

    Figure 7 Left: fringing due to the backgroung sky in the I band (1000x700 pixels central region of the chip). Right: a 256x256 pixels region of a 40mn spectroscopic exposure. Note the cosmic ray events.

    Cosmic rays

    Most of the cosmic rays interactions with the CCD are located within a pixel or two (figure 7 right) and can be easily detected and removed by some dedicated algorithm (Cf software developed for the HST data in STSDAS) or simply by an optimal combination of several frames. The rate at the summit is about 4 events/cm2/mn, or equivalently 80 events/mn on the whole 2048x2048 pixels CCD (the pixel size is 21 microns).

    Noise equation

    With a readout noise of 1.9 ADUs, a pixel to pixel non uniformity of about erqe=0.4% (standard va lue for a non-inverted integration mode CCD) and a gain of g=4.5e-/ADU, the noise in a raw image with a flux level of SADUs is: (with a bias correction but no flat-field correction):

    Miscellaneous figures

    Figure 8 High: pinhole image, the peak intensity is 15000 ADUs and the background level is zero. Low: idem with the fast code. The pattern is maybe due to some optic effects related to the pinhole (diffraction) as it does not appear with larger holes as showed on figure 9.

    Figure 9 High: image of the center of the grid ( 2 pixels large square holes) in equalized histogram display mode. The effect seen on the pinhole image is not present. Low: same image in linear display mode, the central pixels are saturated (32000 ADUs).

    Figure 10 Short exposures on a star field in the R filter, the sky background is almost zero. High: 2sec exposure time. Low: 5sec exposure time. Even with a zero level background (the worst situation for CTE) no CTE lost are present. The bright star actually saturates on the right image and the effect seen on this image is the same kind of pattern as seen on figure 8.

    Figure 11 Globular cluster M79. 2mn exposure in the B band. The field of view is 9'x9'.

    Detector summary sheet