Figure 1: Dark images showing the jailbar pattern, taken from a sequence of 10 DCEs. From left to right: DCE 1, DCE 2, DCE 10. Units are DN/s; slopes calculated from RAW data ramps.
A jailbar pattern is seen in 24 micron images, most obvious in darks. The pattern is due to an apparent low-flux nonlinearity in the data ramps of all pixels of one of the four readouts, resulting in systematically different slopes. The strength and sign of the pattern varies with DCE in a given sequence, though the "deviant" slopes always occur in the same readout. Analysis of RAW data shows that the effect is additive. The jailbar offset is sufficiently small that it should never be seen in science data, even at the lowest expected background levels. However, the dark calibration should probably be adjusted in order to avoid DCE-dependent variations induced by dark subtraction.
One of the pathological 24 micron detector characteristics is the appearance of a "jailbar" pattern, in which there is a systematic difference in signal level between readouts. The effect is most obvious in dark images; Figure 1 shows a typical case for several DCEs in a sequence. The fourth readout is consistently lower in signal than the other three in the first two DCEs, and slightly higher in the 10th DCE. This effect is quantified more clearly in Figure 2, which shows slopes averaged over all pixels in each readout as a function of DCE. The negative pattern lessens from DCE 1 to 2, and essentially disappears by DCE 3. In all subsequent DCEs, a positive jailbar pattern, where the signal in readout 4 is ~75% larger than the others, appears and remains stable.
Figure 1: Dark images showing the jailbar pattern, taken from a sequence
of 10 DCEs. From left to right: DCE 1, DCE 2, DCE 10. Units are DN/s; slopes
calculated from RAW data ramps.
Figure 2: Dark mean slopes as a function of DCE for each of the 4
readouts.
What is the cause of the jailbar pattern, and is the effect additive or multiplicative? Images with significant illumination have shown no evidence of a readout-dependent pattern (see below), suggesting it is additive, but to date there has not been a systematic analysis. The origin of the jailbars is revealed in Figure 3, which plots the mean data ramps over each readout, for 3 different DCEs in a dark integration. It is immediately obvious that the average ramp of readout 4 is quite different than that of the others. In DCE 1, all readouts show a negative ramp, probably the result of the bias boost frames. In subsequent DCEs, however, readouts 1-3 show relatively linear ramps (except for an odd discontinuity between reads 2 and 3), while readout 4 shows a significant nonlinearity where the ramp actually turns over. This nonlinearity results in much lower slopes, creating the apparent jailbar pattern. The origin of the nonlinearity is unclear, perhaps a small drift in the bias voltage?
Figure 3: Dark mean RAW data ramps each of the 4 readouts, for various
DCEs in a sequence.
How does the low flux nonlinearity vary with illumination level? Figure 4 shows mean ramps for an image with faint external illumination of about 100 DN/s. A linear fit to each ramp is also shown for comparison. Readouts 1-3 are linear to within 0.05%; readout 4 is slightly less linear, to within about 0.5%, but considerably more linear than what was seen in the darks. The mean slope of readout 4 is about 2% greater than for the other readouts in DCE 10 of this data set, but the difference between them is nearly identical to that seen in the darks. This result indicates that the low flux nonlinearity, and hence the jailbar pattern, is an additive effect. Clearer evidence of this is shown in Figure 5, which compares the difference in the mean slopes of readouts 2, 3, and 4 from readout 1, as a function of DCE and illumination level. At all 3 illumination levels plotted, readout 2 and 3 differences are close to zero at all DCEs, while readout 4 differences are about -5, -2, and 3 DN/s for DCEs 1, 2, and 3-10, respectively. Thus, there is a consistent offset in readout 4 from the other readouts which does not change with illumination level. The effect is probably repeatable; similear offsets are seen in a comparable data set taken with a different CE box (CE 1), as shown in Fig. 6. Repeatability can be further checked with IOC darks and standard observations at varying backgrounds.
Figure 4: Mean RAW data ramps each of the 4 readouts, for a faint
external stim illumination (~100 DN/s). The blue lines represent linear
fits to the data.
Figure 5: Difference in mean slopes from readout 1, as a function of DCE.
Colored symbols represent data sets at different illumination levels.
Figure 6: Same as Figure 5, but for different data set using the other
CE box.
The additive nature of the jailbars means that the higher the illumination on the array, the less significant the jailbar pattern becomes relative to photon noise. In stim images with half-well illumination (~3000 DN/s for 10 sec. exposures), no pattern is detectable, as expected since the photon noise ~50 DN/s is much larger than this jailbar offset.
These results are encouraging vis-a-vis the data calibration. Typical low-background regions should be ~16 MJy/sr (dominated by zodi), which should translate to approximately 200 DN/s on the array. The photon noise from this illumination is then at least 3-4 times larger than the expected jailbar offsets (keep in mind that the offsets derived above are readout averages, i.e. over 128x128/4 pixels, which greatly improves the effective S/N of the measurements). Therefore, we are very unlikely to see, and thus not have to correct for, the jailbar pattern in science images. However, the change in the offset as a function of DCE could conceivably lead to an artificially-induced jailbar pattern from dark subtraction using master darks averaged over all DCEs in a single sequence. In order to avoid this, we may want to change the dark subtraction strategy by doing separate calibrations for DCE 1, 2, 3, and 4-x for a sequence of x DCEs. In this scheme, the dark calibration IER should be changed so that many sequences (20-30) of ~10 DCEs each are taken, rather than one sequence of 50 DCEs, in order to improve statistics. (There are other reasons for a DCE-dependent scheme, such as the different nonlinearity behavior in the first DCE of every sequence.)