2 Detailed comparisons

2.1 The Datasets

In a sense we are quite lucky as the observations that were mostly affected by the flare turn out to have almost identical counterparts done in more normal conditions during revolutions 222 and 224. I have thus extracted these data from the archive and used it to assess the quality of CAM data on revolution 722 and 723. Table 1 lists the relevant parameters of the set of observations considered here.

2.2 Raw data

Since CAM is an imager, and since the event we are expecting is a particle shower, the easiest way to get a rapid idea of the event's impact is to look at CAM's images (here corrected for dark with the CALG dark). This is done on figure 1 where I show 64 images of 22202606, i.e. normal 5.04 integrations, and 64 images of 72200108, identical configuration though it is clear that the number of particles has ``somewhat'' increased. The change in aspect from one image to the other is not due to a rapid background variation but to the fact that the lookup table is adjusted each time to the full intensity of the image, which in turn is totally dominated by the glitches.

The second CAM observation on revolution 722 is identical to the first one, both in setup and in degradation, thus for the rest of this report I will not consider it. The 5.04s observation that was done on revolution 723 has the same aspect as that of revolution 222, thus to save space I will not display it. Similarly, there are no visible differences between the images taken at 10.08s in revolution 224 and those from 723, apart from the change in background level (a factor of 4). But there is more here than meets the eye.

  
Figure 1: Dark corrected images in LW3, 5.04s integration, 6''/pix for a ``normal'' revolution on the left and during revolution 722 on the right. The raw 10.08s data taken either on revolution 224 or 723 show much less glitches than what can be seen here during revolution 722! The display actually does not do justice to the glitch fireworks of revolution 722...

2.3 (Not so well) deglitched data

For deglitching, I have only used the multi-resolution median transform algorithm (also called mm or mr1d_deglitch) as it is the best adapted to that sort of data: constant background, isolated point sources, glitches leaving long lasting imprints on the detector. Furthermore as the sources are extremely weak I can relax the constraint on the number of scales inspected for glitches.

For the LW3 observations, with 20 exposures per raster positions, the scale parameter can be raised to 4. As for the sigma parameter, experience shows that at 5.04s integration time, a value of 6 is a good choice (efficient deglitching and few noise spikes removed). Applying these parameters on revolution 222 indeed gives excellent results. 5% of the array is affected by glitches.

The same parameter set on revolution 723 also produces quite satisfatory results with again about 5% of the detector being masked.

This is no longer the case for revolution 722. This choice of parameters removes part of the glitches, but only to show that more are still remaining on the array. In fact to clean the array to an acceptable (see later in what sense) level I had to perform 2 successive rounds of deglitching up to scale number 4 and down to 3- deviations. It is worth mentionning here that such a processing would very likely affect source signals in the field, except maybe the faintest ones, and is therefore not recommended when one has science in mind. At the end of this process, 36% of the array has been rejected due to glitches and inspection of the data shows (see e.g. fig. 2) that glitches still remain on the detector.

Therefore we can conclude from these numbers that we have seen during revolution 722 an increase in the rate of glitches by a factor of at least 7, and probably around 10. We were back to a normal rate of glitch impact at the start of revolution 723.

For the LW2 data, taken at 10.08s, the small number of exposures taken would restrict the number of scales to be used for deglitching to 3. With such a small number, it is impossible to properly deglitch either the revolution 224 or 723 data with the multi-resolution method. Thus I have assumed that the sources are faint enough so that they do not show on individual frames and thus I can increase the number of scales up to 4. I also have had to decrease the sigma parameter to 3 to produce efficient deglitching for both revolutions. Thus the final setup was Nscale=4 and Nsigma=3 for both revolutions.

It produces very similar results on both revolutions, i.e. most of the glitches are gone, though long-lasting ones are obviously still there, and approximately 10% of the array is flagged as glitched, indicating once again a normal behavior in terms of glitch rate for revolution 723.

2.4 Data quality

Assuming one can deglitch the data obtained during such a storm, one may wish to know whether the resulting data can be used scientifically, i.e. is CAM in a state where the calibration measurements apply? The answer here is quite certainly no. The global increase in the number of glitches imply a similar increase in the number of glitches affecting the gain of the detector. As a result the latter varies quite significantly and stochastically during the observations, on scales as small as 1 minute. Therefore the data will be quite difficult, if not impossible, to calibrate.

This rapid variation is illustrated in figure 2 for the 5.04s integrations, and figure 3 for the 10.08s integrations. On these figures I have plotted the dark-corrected, deglitched as indicated above, signal from a representative pixel near the center of the array. For the 5.04s integrations the solid line is for revolution 722, the dash-dotted one for revolution 723 and the dotted one for revolution 222. For the 10.08s integrations, we have no data on revolution 722, so the solid line is for revolution 723 and the dotted one for revolution 224. Offsets have been used to put all these signals on the same plots.

  
Figure 2: For the 5.04s readouts, signal from a representative pixel (16,20) in the three revolutions used here: 722, solid line, 723 dot-dashed line and 222, dotted line. 15 ADU/g.s have been subtracted from rev 723's output to place it on the same graph. Variation of the mean flux level between revolutions is due to the fact that the zodiacal light contribution depends on the date of the observation and pointing direction.

On figure 2, where I remind that no multiplicative scaling has been applied, the degradation of data quality is obvious. More precisely, signal from revolution 722 is highly unstable, probably because glitches with positive tails (e.g. the pseudo-sources near readout number 1100 and 1200) pile up in time with glitches with negative tails (e.g. at readout number 400). It is doubtful that any photometry can be done with such an unstable array. The contrast is stricking with signal from revolution 222 or even with that of revolution 723.

It is also apparent on this graph, that, though the improvement is dramatic from revolution 722 to 723, we still are not in ideal conditions: there is a larger number of glitches with negative tails. In principle, faint source analysis is able to deal with these glitches.

A more quantitative assessment of the data quality can be made by comparing the ratio of mean intensity in the cube over rms in the data cube, a sort of Signal to Noise (S/N) ratio, for these three observations. To do that, I have flat-fielded the data with the CALG flat-field, and, since it is not very adequate on the edges of the detector, I have used only the area (10:21,10:21). Table 2 summarizes the results obtained this way.

For the 5.04s observations, the setup is strictly identical in the three observations, and the background only changes by 10%, which should have a minimal impact on the resulting S/N. Yet one can clearly see that this ratio has decreased by a factor of 3 from revolution 222 to revolution 722. This reinforces the previous statement that is quite useless to try and do science under such conditions. As was also hinted at before, we can see that revolution 723 represents a noticeable step toward full recovery but we are not quite there yet: the S/N is still significantly lower, reflecting the higher number of glitches with negative tails.

  
Figure 3: For the 10.08s readouts, signal from a representative pixel (19,16) in the two revolutions used here: 723, solid line, and 224, dotted line. 4 ADU/g.s have been added to rev 224's output to place it on the same graph. Variation of the mean flux level between revolutions is mostly due here to the two different pixel field-of-views used.

For the 10.08s integration time, we have no data taken on revolution 722. However, comparison of 10.08s data from revolution 224 and 723 with their corresponding 5.04s data (i.e revolution 222 and 723) shows that the trends should be quite similar. As can be seen on figure 3, data from revolution 723 are still quite noisy, indicating that the array still sees a higher number of glitches with tails compared to normal operating conditions. This is confirmed by the comparison of S/N in table 2: as the flux increases by a factor of 4 due to the use of the 6''/pix lens, we expect and increase of the S/N by a factor of 2. Such is not the case, the increase is only a factor 1.22. The data from revolution 723 appears significantly more chaotic in figure 3 than in figure 2. That is probably due to two facts: (1) the 10.08s observation is done before the 5.04s one (both last approximately 2 hrs) and (2) the x axis is not equivalent on both graphs, that of figure 3 should be multiplied by 2 to be brought on the same scale as that of figure 2.