Text Size:
Smaller Text Normal Text Larger Text

Facebook page

Twitter profile

YouTube channel

NOTCam Engineering Grade Array (SWIR1)

Content of this page:

Detector overview

The Engineering Grade Array was used in NOTCam from June 2001 to the 20th of October 2005, when the first Science Grade Array (SWIR2) was installed and commissioned (see also the Science Array commissioning report ). The first Science Array (SWIR2) ceased to work in April 2006. The engineering grade array was re-installed on the 5th of May 2006 and was used until 7th of December 2007 when the New Science Array (SWIR3) was installed.

Array Hawaii HgCdTe 1024 x 1024 x 18.5 micron
SWIR1 P/N: 8-344R#1 #131
Bad pixels 2%
Gain about 2.8 e-/ADU
Dark current TBD
Linearity ~ good to 99% up to 15 000 ADUs
Saturation starts at 40 000 ADUs
Read out noise 10-12 e-
Read out time 3.6 s
Pick-up noise removed by September 2003
Cross talk not seen
Memory effect present for strong saturation

Fig 1: Schematic drawing of the detector with the four quadrants marked in the numbering system used by CUO for NOTCam. (Note that Rockwell numbers from 1 to 4, but in the same order.) The arrows show the direction of the fast readout and the corners in which it starts. All four quadrants are read out simultaneously.

Left: Valid from 2001-2005. The lower left corner of quadrant #1 is the location of pixel (x=1,y=1).
Right: Valid since Jan 2006. The lower left corner of quadrant #0 is the location of pixel (x=1,y=1).

The x-flip was introduced together with the Multi Extension Fits (MEF) data format in Jan-2006 in order to adopt to standard for MEF. In the MEF quadrant numbering system, the right hand image is numbered as follows: LL=1, LR=2, UL=3, UR=4. See fits headers.

NOTCam is offered with two possible readout modes: the standard reset-read-read mode and a ramp-sampling mode (multiple non-destructive reads during the integration). Read more about this in NOTCam User's Guide.

Cosmetics - bad pixels

Fig 2: Two dark images obtained with the reset-read-read mode (command mdark t N). Integration time (t) is 0 seconds (left) and 50 seconds (right). The images are a median combination of 10 individual images. Note the amplifier glow and the higher number of hot pixels on the 50s long dark integration.

Dark-0s Dark-50s

Dark images were only obtained after Dec-2001 since before that, the shutter malfunctioned. The darks have several features. Most prominent are the hot rows that are the same in each quadrant. There is an apparent rise in brightness in the first few rows in each quadrant where the count level can be more than 3 times the average level in the remaining part of the quadrant. This reset ramp, as well as the hot rows, seem to subtract out well, although the reset ramp may take time to stabilize sufficiently that the subtraction is perfect. In dark images of longer integration times there is clear evidence of shift register glow along the edges. This effect also apparently subtracts out well.

There are two dead columns (x=1 and x=513 before x-flip, or x=511 and x=1024 after x-flip), but this is due to a bug in the controller and has nothing to do with the detector itself. Quadrant #1 has a bad column at x=247. Quadrant #3 has 3 bad columns at x=726, x=771, and x=920. The images shown here are in the old system, i.e. before x-flip in 2006.

There are individual bad pixels spread over the whole array as well as bad features (check quadrant #2). Pixels are here called bad when they deviate by more than 6 sigma from the mean level. Among the bad pixels we distinguish between hot and cold. The cold pixels include also the dead pixels which show no response at all. The hot pixels may be strongly non-linear and their number increases with integration time.

Percentage of bad pixels (hot and cold) in dark frames (Data obtained 27-Jan-2002.)
Quadrant Exptime Cold (%) Hot (%)
0 0 0.6 0.2
0 5 0.6 0.3
0 50 0.8 1.2
1 0 1.5 0.2
1 5 1.8 0.2
1 50 1.6 1.1
2 0 1.2 0.4
2 5 1.6 0.5
2 50 1.4 1.7
3 0 1.3 0.3
3 5 1.3 0.4
3 50 1.6 1.5
Average 50 1.4 1.4

The results above show that about 2.8% of the total pixels were bad at a typical exposure time of 50s in Jan-2002.

Zero pixels

The number of dead pixels can be expected to increase with every thermal cycle of the array, because the different thermal expansion properties of the layers of an infrared array may cause detachment of the bump bonds upon repeated thermal cycles. In 2003 NOTCam had to be opened six times, and we have monitored the evolution of the dead pixels. It turned out that the percentage of the array that had zero-valued pixels increased from 1% in Jan-2003 to 5% in Dec-2003. Mainly the left edge of the detector was affected.

Fig 3: The percentage of "zero" pixels apparently increasing for each thermal cycle. The asterisks show the percentage of zero valued pixels measured on different dates, and the horizontal lines shows the periods when NOTCam has been opened and therefore has been thermally cycled. As explained below, an adjustment of the dc-offset voltages of the array cured the problem, and the true number of "zero" pixels (i.e. dead pixels) is below 1%.

Fig 4: The spatial distribution of the zero pixels (black) in Jan-03 (left) and Dec-03 (right).

Jan-03 Dec-03

Eventually it was found that the large number of "zero pixels" are not dead pixels. Investigation of the reset image (i.e. the pre-read) showed that its peak histogram level had drifted from about 4000 ADUs towards zero. This caused the apparent presence of dead pixels, and a slight modifications to the dc-offset voltages were needed to adjust the reset levels of each quadrant back to the original values of around 4000 ADUs. Thus, the bottom line is that the actual number of dead pixels in this array is around 1%, and that this number has been all the time the same.

We suspect the origin of the problem is some drift in the array itself. It should be noted that since the adjustment of the voltages in 2005, the array has behaved well, and no further adjustments have been required.

Check the NOTCam detector monitoring results for an updated graph of the zero pixel evolution:

Reset-read-read mode
Ramp-sampling mode

Hot pixels

The table below shows the increase of hot pixels with longer integration times for a selected region of the array. The percentage of hot pixels in darks of different exposure times within the image section [150:400,550:800] has been estimated. Pixels are in this case considered bad when they differ by more than 110 ADU from expected, where 110 ADU = 10 sigma with sigma being the worst-case noise level (i.e. taken from the dark image with 729s exposure time). These data were obtained Sep-2003 in reset-read-read mode.

Percentage of hot pixels in a dark image section.
Quadrant Exptime [s] Cold [%] Hot [%]
2 0 0.2 0.3
2 3 <0.1 0.4
2 9 <0.1 0.6
2 27 <0.1 1.0
2 81 <0.1 1.9
2243 <0.1 3.4
2729 <0.1 7.4

Dark level

From several sets of dark frames with exposure times 0s, 1s, 3s, 5s, 25s, and 50s we have plotted the mean levels per quadrant as a function of integration time. There is no clear repetitive dependence on integration time, however. Either the dark current is too small to be measurable or it is not linear with time. A rough indication of the dark current is thus that it is less than 1 e-/s per pixel. The peak of the histogram varies from about 30 to 60 ADU whether the integration time is 0s or 50s. The level seems not to be dependent on what level the detector has seen just before.

Because the dark current is so small (but variable) we currently do not recommend obtaining dark images at a different time than the target frames for use in the data reduction of background limited images. The dark is probably best eliminated by the automatic subtraction of it together with the sky subtraction. If you plan to make flatfields out of the target frames, however, you'll need to correct for the dark. The investigation of the dark images is on-going.

The pick-up noise described below with amplitudes as large as 20 ADUs have a strong effect on the very low level darks. Because the images are stored as unsigned integers, negative values will wrap around to very high values. This produces very nasty looking images, which need special "treatment" before they can be used.

Pick-up noise

Fig 5: Example of the pick-up noise situation in May 2003 (left) and in Sep 2003 (right), before and after the improvement, respectively. Each of the images below are difference images of two independent dark frames. The readout noise was reduced from 24 electrons to 10-12 electrons with the removal of the pick-up noise in 2003.


Before September 2003 all data suffer to some extent from the interference pattern visible on the left image above. As can be seen, the effect is strongest in the quadrants #2 and #3 (the upper ones), while the lower two (#1 and #0) are looking much nicer. The interference pattern is due to pick-up noise and seems to vary with time. Usually it is eliminated completely in the combination of several images that are well within BLIP (background limited performance), but it can be persistent, especially in images of very low background, and the peak to peak variation was generally around 20 ADUs. In order to beat this noise, it was usually recommended to try to select integration times such that the background was around 1200 ADUs minimum, and to have at least 10-20 images to median combine. Also, results were generally better using the ramp-sampling mode with many readouts.

As a consequence of the pick-up noise combined with the fact that the data is stored as unsigned integers, there were frequently pixels which got negative values in the reset frame subtraction of very low count level images (e.g. darks). These negative values then wrapped around to very high positive numbers and such images would have to be treated for this effect before they could be used.

Until its removal, the pick-up noise was always present, even in the first tests in the lab. After having tried several solutions without success, it was finally the physical separation between the power supply for the array electronics and the electronics itself which gave results. The power supply unit was moved step by step out of the electronics box, and the effect decreased correspondingly.

Readout noise

The readout noise in [e-] for the reset-read-read mode.
Date Quad 0 Quad 1 Quad 2 Quad 3 Median overall
28-Jan-2002 17 21 26 24 24
31-Mar-2002 20 9 22 22 19
8-Jun-2002 16 20 27 25 21
16-Sep-2002 16 19 30 21 21
05-Jul-2003 12 12 14 13 12

The readout noise is calculated in a representative area within each quadrant, and the same areas are used for all dates. The overall median is the median of the value found in 81 boxes over the whole array.

Note the improvement from July 2003. This is when the solution to the problem of the pick-up noise was found and remedied.

Since January 2003 the readout noise has been monitored regularly for both readout modes. Check the monitoring results:

reset-read-read mode
ramp-sampling mode

Please, check the NOTCam User's Guide for a description of the two different readout modes available with NOTCam.


The gain in [e-/ADU] for the reset-read-read mode.
Date Quad 0 Quad 1 Quad 2 Quad 3 Median overall
28-Jan-2002 3.3 3.2 3.4 3.2 3.2
31-Mar-2002 2.8 2.2 2.8 2.7 2.8
8-Jun-2002 3.1 2.9 3.2 2.9 2.9
16-Sep-2002 3.2 3.1 3.5 2.4 3.1
05-Jul-2003 3.1 3.1 3.2 3.0 3.1

The gain is calculated in a representative area in each quadrant, and the same area is used for all dates. The overall median is the median gain calculated from values obtained in 81 boxes over the whole array.

Since January 2003 the gain has been monitored regularly for both readout modes. Check the monitoring results:

reset-read-read mode
ramp-sampling mode


Non-linearity is an inherent feature of infrared arrays which distinguishes them clearly from the linear CCDs. While the saturation of the detector starts at 40000 ADUs the array is found to be linear to 1% accuracy only up to about 15000 ADU on the average.

For each readout mode you can check the non-linear behaviour for each of the four quadrants from the monitoring data:

reset-read-read mode
ramp-sampling mode

Detector flat field

Fig 6: Processed flat field obtained from 10 differential twilight flats. The differential method (pair-wise subtraction of bright minus faint) is used to eliminate the thermal contribution from the master flat. No bad pixel correction was attempted, instead the final master flat was median smothed by a 3 pixel box, which almost eliminates the bad columns.


There are several cosmetic features on the engineering grade detector. Most prominent is the bright band of higher quantum efficiency, a number of larger areas with bad pixels, and some hair-like features. The standard deviation in small boxes of 20 x 20 pixels is 3-4%. The deviation over the whole field is +20% in the bright band and about -15% in the darkest corner.

Please, check our archive of sky flats.

Example of raw and processed image

Fig 7: V361 Cep in the K band. A single raw image of 50 seconds integration (left) and the sky-subtracted and flatfielded combined image of 6 dithered images (right). The observations were done in beam-switch mode using ramp-sampling readout. North left and East up. No bad-pixel removal or correction of bad columns. Only dedicated sky subtraction using off-target fields and flatfielding.

Raw-im Proc-im

Comments to Anlaug Amanda Djupvik
Back to top Last modified: May 31 2023