Revised 1999Dec23: Previous version (1999Feb23) is below.
I have computed medians and probable errors (converted here to standard deviation assuming a gaussian distribution) for edge pixels of CD5 images. Numbering the complete image X1 0..2042, X2 0..2036, the actual image area is X1 6..2037, X2 2..2033. The corners are odd so the edge pixels used here are X1 0..5, 2038..2042, X2 2..2033 referred to as columns 0..5, 2038..2042 and X2 0..1, 2034..2036, X1 6..2037 referred to as rows 0..1, 2034..2036. Yes: I have forgotten to add 32768.
VDARK15 sd IDARK15 sd Type
Image area -23237 24 -22977 41
Columns
X1 = 0 -23168 36 -22545 118 C?
1 -24900 1 -24733 1 B
2 -23250 5 -23006 9 D*
3 -24899 1 -24733 1 B
4 -23126 42 -22929 31 C
5 -24898 1 -24731 1 B
2038 -23223 25 -22799 73 C
2039 -18143 50 -17553 129 A
2040 -23212 24 -22790 71 C*
2041 -4791 3073 -22788 73 A, =2042 for I
2042 -23224 24 -22788 73 C
Rows
X2 = 0 -18237 485 -18374 497 A
1 -23164 50 -23007 57 C
2034 -22734 105 -21837 527 A
2035 -23268 9 -23026 15 D
2036 -23269 7 -23022 12 D
A No idea
B Constant; I guess this is the -ve end-stop
C "Dark". Increases with dark current
D "Unused". Rather constant; try it as a reference value
* I arbitrarily selected X1 = 2040 for a "Dark" estimate and X1 = 2 for "Unused" for use as reference. See MK IV Dark images, hot pixels and cosmic rays
I have no idea how this can be squared with the results for the original 2063x2063 format. One is tempted to match column 2 here with one of the "Unused" columns 0..15; probably one in the range 9..14 noted as being consistent. Column 4 would then be one of the "Dark" columns 16..23. But then, how does one match column 0? Columns 1, 3 and 5 have so low a standard deviation - less than the estimated noise of the read-out electronics - that I suspect they represent the negative full scale saturation value. Columns 2038, 2040 and 2042 (plus column 2041 on the I images, which as Herb Johnson pointed out is identical to column 2042) match the "Dark" columns found at 2056..2062 in the 2063x2063 format.
Revised 1999Feb3.
The MKIV data supplied on CD ROM by Tom Droege includes a band
of pixels around the edge which are variously referred to as
"Dark" and "Unused". The full format
includes a strip all around the actual image varying from 7 to
24 pixels in width but only a small subset of this appears to be
useful. The bits I have chosen to look at in greater detail are
Columns 9 -14: "Unused" or UN
Columns 19 - 21: "Dark" or BOT
Columns 2058 - 2060: "Dark" or TOP
Yes - I am labelling them upside down/sideways to match the way I am plotting the images.
All the remaining edge pixels gave higher variance when I tried to do anything with them. I give it as my professional opinion - for what that's worth - that they are entirely useless and could be thrown away without loss.
The "Dark" pixels show marked trends at both edges of the image. As already discussed by Tom on several occasions, these trends are the result of instrumental time-constants. It looks as though the selected "Dark" pixels could be used rather simply to eliminate this effect. As we can already get rid of the effect perfectly well using regular Dark processing, I have put investigating this on the back burner. The middle of the band is not affected so I have restricted analysis to rows 400 - 1800. Enough of this chatter - let's look at some pictures.
This shows a negative temperature coefficient for both UN (red crosses) and BOT (black squares) (TOP is similar and is not shown). Only the later files, starting with d141916, had CCD temperature logged. I have arbitrarily taken the temperature for the earlier files as -23.5C as logged for files 11ba etc and have filled in the remaining gaps with the nearest logged temperature.
I have also subtracted different mean values for files x0 - dark1123 (Nov22 - 23), fz3512d - fz3520 (Nov 26) and the d series (Dec 14 - 15).
Ordinary temperature coefficients within the CCD are removed by the method of measurement so the observed effect is some sort of residual nastiness. The effect is big enough to give a useful measure of the temperature (0.7C).
Here is all the data corrected for the temperature coefficient and plotted against exposure. I have added 50 seconds to Tom's logged exposure times to allow for the read-out time and normalized to -20C using an exponential factor with scale 8C (dark current doubling every 5 to 6C).
Um! Not a very impressive correlation. Let's try something else.
The "Unused" values (UN: red crosses) decrease with exposure. This is very odd. Even odder when looked at in detail. A possible fit is with a constant negative leakage, independent of temperature. This may just be statistics since the temperature range is so small. An alternative fit may be obtained assuming zero leakage and a (negative) term proportional to the mean illumination (see next picture). Go figure!
Here is a plot of the temperature corrected data against light level. I am defining "Light" as mean value of image pixels minus mean value of "Dark" pixels.
The "Dark" values (black squares) show at least as good a positive correlation with "Light" as they did with exposure.This suggests that there might be some residual illumination effect here. The masking may not be not as opaque as it should have been. Examination of the standard deviation (below) suggests this is a real effect.
The "Unused" pixels have an apparent negative response to illumination. Very odd.
Here is what is left after subtracting off the "Light" effect.
This really ought to include the effect of dark current so I have
plotted it against normalized exposure as above. The thin lines
show the best fit. The "Dark" pixels show a positive correlation.
The "Unused" pixels (red crosses) show no remaining correlation with exposure.
sd: UN BOT
25.1 42.6 raw (after subtracting the means)
13.9 29.2 after subtracting temperature coefficients
6.7 16.8 after subtracting "Light" effect
6.7 15.4 after subtracting Dark current
3.6 2.8 sd of replicates
So there is a good deal of unexplained variance left over!
The standard deviation of the "Dark" pixels (black squares) shows a lot of scatter.
The "Unused" pixels (red crosses) show a lower standard deviation which is independent of temperature. Nice to find something that is solidly in agreement with expectation!
The increasing trend with exposure time for the standard deviation of the "Dark" pixels is again not a very convincing correlation.
The standard deviation of the "Dark" pixels increases more or less in proportion to "Light" level (as defined above). That this is a result of pixel-by-pixel variation is shown by comparing pixel values for two images. For example, the BOT pixels for d151707 and d151710 have a correlation of 0.85, showing that most of the variance is consistent between the two images.
The "Unused" pixels show no variation with illumination. This, of course, makes the apparent correlation between UN and "Light" found above extremely unlikely.
The clean behaviour of the "Unused" pixel values and their consistent negative temperature coefficient suggests their use as a thermometer for those cases where the temperature was not logged. The results of this excercise are shown in the plot. Additional corrections have been applied for exposure length and illumination level. The standard deviation of the difference from the measured temperature is 0.7C which is usefully small. Or, at least, better than nothing.
The red crosses in this plot are the points where I guessed the temperature.
Maybe it is not really that good! There's a fair amount of wish-fulfilment built in to the fitting process. The first fit I tried put all the -23.5C points at around -8C ...