Instructions for Using the Alpha-Release of the WFC3/UVIS
Pixel-based CTE Correction
Jay Anderson
STScI
February 19, 2013
1. Abstract
The current version of the pixel-based CTE correction for WFC3/UVIS is available only
as a stand-alone FORTRAN program. This routine is currently an “alpha” release,
meaning that we reserve the right to make major changes to it. The routine does work
reasonably well in the domain where it can: where the source is bright or where the
background is moderate. Faint sources on low backgrounds will always be very difficult
to correct for, as they tend to experience losses that are a large fraction of the initial
counts. It will always be impossible to reconstruct something from nothing.
This routine is not yet a part of the pipeline and it will probably be several months before
it is included in the pipeline. Also, at this point, there are also no additional reference
files (such as de-trailed darks) that can be used in conjunction with it.
2. Construction of the model
A comprehensive description of the construction of the model will eventually follow in a
separate ISR, but it is worthwhile to provide a brief history here. The model is generally
based on the pixel-based correction developed by Anderson & Bedin (2010) for ACS,
which itself was an extension of the model constructed by Massey et al. (2010) for their
reduction of the COSMOS data.
There are two aspects to the CTE model: (1) how charge is lost from one pixel and (2)
how it is released into the upstream pixels. The basic model assumes that charge “traps”
are distributed throughout the detector, and each trap is able to grab a particular electron
out of the packets that pass through it (either the first, second, fifth, five-hundredth, etc.).
Since it is impossible to know exactly what traps can be found in which pixels, we make
the assumption that each of the 2070 pixels up the column has an identical spectrum of
fractional traps. The current UVIS detector has, on average, about 500 traps in each
column. Most of these traps will affect only large pixel packets, but unfortunately there
are enough traps that grab the first few electrons to make CTE losses pathological when
sources are faint on a low background. An iterative approach was used to construct the
model.
2.1 The initial model The first iteration was similar to what was described in Anderson
& Bedin (2010) for ACS. Warm pixels (WPs) serve as convenient delta-function
sources for probing how charge is transferred down the detector. We examined the trails
behind the few relatively bright WPs that UVIS has. The area under each trail told us
the number of traps that impacted the WP during its journey (the losses), and the shape of
the trail told us how charge gets released. This provided the initial version of the model.
2.2 Probing losses in the smallest packets. Unfortunately, it is very hard to measure
the trails behind WPs that have less than 50 electrons. This is partly because the losses
are large (and no longer perturbations) and partly because the trail gets lost in the noise.
In an effort to pin the model down for the smaller charge packets — where CTE losses
are most critical — we took a set of short-dark and long-dark exposures. The short-dark
exposures were 100s and the long-dark exposures were the standard 900s. To study the
CTE losses from small packets, we first apply our initial CTE correction to the long-dark
stacks. This gives us a pretty good idea of how much flux started out in each of the
medium-bright WPs (say, those with 75 to 1,000 electrons). We can then multiply these
images by 0.1111 (100/900, the ratio of exposure times) to get the expected number of
counts in each short-dark exposure. We then compare the observed number of counts in
the short dark with the expected number to get a direct picture of CTE losses as a
function of the number of transfers each WP experienced. This allows us to study
electron packets from 10 to 125 electrons in size.
The above procedure allowed us to see how many electrons a charge packet loses on
(essentially) zero background. We found the losses to be even greater than expected. A
pixel packet that started out with 80 electrons at the top of the chip ended up with less
than 30 by the time it reached the readout register; and a packet with 50 ended up with
less than 10. We examined the trends for charge packets of different size and found that
packets with between 1 and 10 electrons continued to lose more and more electrons as the
packet got larger. But once packets reached the size of 12 e−, their losses appeared to be
largely independent of packet size. This was true until they reached the size of 50 e−or
so, at which point they slowly started losing more again.
These trends told us that there aren’t many traps that grab electrons #13 through #50.
Thus, if an image has a background of about 12 electrons, then perhaps that could keep
many of the traps filled and significantly mitigate the CTE losses.
2.3 The efficacy of some background. To test this optimistic hypothesis, we took a
series of observations of the center of globular cluster Omega Centauri, with pairs of
short and deep exposures (10s and 700s) through F336W. This is not unlike what was
done above to study WPs in the dark exposures. The reason for this target/filter choice
was that the field contains a nice, flat distribution of stars. In the short exposures, there
are not many stars brighter then S/N ~ 50, but below this (which is the location of the
turnoff) there is a relatively flat distribution with magnitude. These image pairs allowed
us to assess losses directly by comparing the observed counts in the short exposures
against the predictions from the scaled-down deep exposures. We also varied the
Figure 1: This figure shows the observed trends that were used to constrain the model, and the model itself.
Each of the seven panels corresponds to a set of WPs that the long darks tell us should contain 10, 20, … 80
electrons in the short darks. This WP is then observed after a different number of transfers (250, 750, 1250, and
1750) and the total number of electrons (background + WP) is shown. The black trends show the behavior on
very low background, the green on a background of ~2, and the blue on a background of 12. The data were
taken in September 2012.
background using the post-flash option. This gave us a direct assessment of how various
background levels shield sources from CTE losses. This study confirmed the efficacy of
12 electrons background. (See Anderson et al 2013).
Finally, we took a series of short-long darks with various backgrounds to help us pin
down the model in the context of background mitigation. Again, we constructed a model
of the dark current in each pixel from the de-trailed long dark exposures (using our initial
readout model). Then we examined how the various WPs lost flux as a function of the
number of transfers and the background level. Figure 1 shows the WP trends that were
observed in the short-dark exposures taken with three different levels of background (0
e−, 2e−, and 12e−). We actually fit the model to 10 different levels of background, but for
clarity we show only three here.
The black curves in the panels from left to right show what happens to WPs that start out
with 10, 20, 30, 40, 50, 60 and 80 electrons on a detector with no background. Losses
are 70% or greater for stars that are transferred all the way down the chip, even for WPs
that start with 80 electrons. The green curve shows the results for backgrounds of about
2 electrons. Losses are still large, but are down by perhaps a factor of two from the
pathological zero-background case. The blue curve shows the losses for WPs on a
background of 12. Losses are less than 20% for all WPs.
The model, indicated by the dashed curves, does a nice job describing most of these
trends. At the very lowest end (a small WP on a low background), the model overpredicts the losses. The model has a monotonic function with one degree of freedom
(the number of traps encountered per 2048 transfers as a function of packet size) to
constrain this two-dimensional distribution of losses versus WP intensity and
background. The errors at the very low end indicate some inadequacy in the model
algorithm, but in fairness, it is very hard to reconstruct sources in this regime anyway.
2.4 The current UVIS model. The current UVIS model algorithm is somewhat different
from the algorithm that is operating in the ACS pipeline. Whereas the ACS model
allowed traps to affect fractional pixel levels and work on real-number pixel arrays, the
new model explicitly deals only with integer numbers of electrons. The new model is
also specified in a somewhat simpler manner than the ACS model. Whereas the ACS
model specified the “trap density” at various electron levels, the current model is simply
specified by the cumulative number of traps as a function of packet-size in electrons. We
show the cumulative trap distribution in Figure 2, below.
Figure 2: From the current model: the cumulative number of accessible traps per 2048 pixels-up-the-column,
as a function of packet size. The left plot shows the small-packet region with linear scaling, and the right plot
shows the full range of packet sizes with log-log scaling. The marginal losses are essentially the slope of the
curve.
It turns out that the trail profile for WFC3/UVIS does not appear to be a perceptible
function of packet size. We found that 20% of the trapped electrons are released after the
first transfer, 8.5% in the second transfer, 6.75% in the third about 1.5% in the tenth,
0.1% in the fiftieth, and the trail goes to zero after about 60 pixels. There are not a lot of
bright warm pixels in the UVIS detector, so it is not trivial to follow the trail out to large
distances.
2.5 Dealing with “readout CRs”. It often happens that cosmic rays (CRs) strike the
detector during readout. A full-chip readout takes about 90s, so an exposure that is 810s
will have 10% of its CRs hit during readout. CRs that hit during readout do not undergo
the same number of transfers as the electron packet they arrive at the readout-register
with; they undergo fewer. For instance, if a CR hits the detector in pixel [200,200] at the
time when pixel [200,2000] is passing through it, then the CR-added electrons will
undergo one tenth the implied number of parallel transfers. It is clear that if we treat all
electrons as having undergone the number of transfers implied by their vertical pixel
location, then we will overestimate the amount of trailing suffered for these CRs, and
hence will over-subtract the trails.
Like the ACS model, the current UVIS model also searches for over-subtracted trails.
We define an over-subtracted trail as either: a single pixel value below −10 e−, two
consecutive pixels totaling −12 e−, or three totaling −15 e−. When we detect such an
over-subtracted trail, we iteratively reduce the local CTE scaling by 25% until the trail is
no longer negative. This does not identify all readout-CRs, but it does deal with many of
them. For images that have backgrounds greater than 10 or so, this will still end up oversubtracting CRs a bit, since we allow their trails to be subtracted down to −10, rather than
to 0. It would be possible to have the algorithm use the background sky value rather than
zero, as the baseline, below which it looks for over-subtracted trails.
2.6 CTE reconstruction with very low backgrounds. The fact that many of UVIS’s
observations have been taken with low background makes CTE reconstruction
particularly challenging. This is a regime that ACS has not had to face, since even “bias”
exposures with no integration time collect ~5 electrons dark-current in their top-row
pixels by the time that pixel completes its 2048 parallel transfers. UVIS pixels currently
get about 0.5 electron dark current during readout; they got even less in the past.
We mentioned above that a source with only a few electrons loses almost all of them
during the 2000 parallel transfers down the detector. This makes it very hard to
reconstruct faint sources on low backgrounds, as there is not much evidence in the readout image that there was anything there. This is particularly true when we fold in the
contribution of the 3 e- readnoise. A source that loses so many electrons that it cannot
stand out above the readnoise will be impossible to reconstruct. A corollary of this is that
if we ignore the fact that our observed images have readnoise and try to determine what
original image could get pushed through the read-out model to produce the observed
pixel distribution, we will end up with an image that is perhaps 10× noisier than the
original observation1. It is clear that readnoise-mitigation will be even more important
for UVIS than it was for ACS.
A further complication with low backgrounds is that the exact number of electrons in the
background can make a big difference in the CTE losses. A 25 e- WP on no background
will lose more than 20 electrons (80%), while the same WP on a background of 3 will
lose perhaps 10 electrons (40%). It is therefore critical to accurately estimate the
background so that sources can be reconstructed as accurately as possible. It is clear that
this estimate of the background must be much more accurate than the readnoise (±3
electrons) allows us to know the number of electrons in any given pixel.
For all the above reasons, the new model includes an improved readnoise-mitigation
algorithm. In brief, the goal of the algorithm is to identify the smoothest possible image
that is consistent with being the observed image plus readnoise. The reconstruction
algorithm acts on this smooth image to make a conservative estimate of how charge may
have been transferred from one pixel to another in the real image during readout. The
negative of this transfer is then added to the original image.
3. Download
The routine can be downloaded by visiting the following website:
http://www.stsci.edu/~jayander/X/EXPORT_WFC3UV_CTE
There are two FORTRAN programs available for download from this directory.
Instructions on compiling the programs and on their parameters are given in the
README file and in comments at the top of the files. We will just give a very brief
description here.
The first program is the correction routine itself, named wfc3uv_ctereverse. It
takes a _raw.fits file and generates what is being called a _rac.fits file. This
output file is as similar as possible to the original raw file, except that it is real*4 instead
of an unsigned integer*2, and has had its electrons re-arranged in accordance with a
model of how charge likely got redistributed during readout. These rac files should be
able to be run through the CALWFC3 pipeline as if they were normal raw files. The
routine also has the option (use “FLC+”) of taking a raw file and an flt file. It will
determine from the raw file how the electrons need to be redistributed and will apply this
1
It is easy to see where this comes from. If an observed image is full of empty pixels and
has one pixel far from the readout has one electron, then in order to read-out such a
distribution, we would need to start with (say) 10 electrons, 9 of which would get lost
along the way. This would allow us to read out a value of “1” for the target pixel, but the
model would also have to account for the fact that (say) “2” electrons would have been
added to the upstream pixel, and single electrons to seven other upstream pixels. The
original image would have to look something like: (0 +10 -2 -1 -1 -1 -1 -1 -1 -1) to be
read out as (0 1 0 0 0 0 0 0 0 0). Generalizing this to an image full of readnoise, where
each pixel can vary by ±3, it is clear that the de-blurred result will be very noisy.
to the flt image, producing an flc file, which can be used for drizzle or other
standard exposure-level image analysis. The routine has several flags and the comments
at the beginning of it should help you use it. It is worth noting that the routine cannot
operate directly on the flt file, since that file may have had post-flash electrons
subtracted, and the reconstruction routine needs to know about them, since they lessen
the CTE blurring. Depending on the background, it can take 15 minutes to an hour to
construct the correction for a raw image.
The second program is named wfc3uv_forward, and it applies the forward CTE
modeling. It takes a file in what I call my “z” format (8412×2070, real*4, with each
2103×2070 amplifier arranged with its parallel readout direction down and its serial
readout direction to the left) as an original distribution of pixels. The routine then
simulates the image what would result if this distribution were read out (with or without
readnoise added afterward). Instead of starting with a pre-readout image, this routine is
also able to take a list of sources (x,y and total flux) and a sky background and simulate
the initial image and the image that would be read-out. There are comments at the
beginning of the program to provide more information on how to run it. We should note
that Figure 1 shows us that the current model over-predicts losses when the background
is below 5 electrons and the source has less than 20 electrons per pixel. As such, this
forward model should not be used to predict in detail what happens to very faint sources
on very low backgrounds. We are working on ways to adjust the model to accommodate
this inadequacy. It is not simply a modification of the parameters, but rather will require
some fine-tuning of the algorithm itself.
4. How accurate are the corrections under various circumstances?
We have run the pixel-reconstruction algorithm on the short-dark pairs of Omega Cen
images (from §2.3 above) and done aperture photometry and fit PSFs for positions. The
results are shown in Figure 3 for photometry and Figure 4 for astrometry.
Figure 3: Results of simple 5×5-pixel aperture photometry on the short and long images of Omega Cen, for
stars more than 1500 pixels from the readout register. We used the long exposures to predict the number of
counts in the short exposures. The pairs of exposures were taken at the same pointing, so that no aperture
correction is needed to compare the numbers of counts. Each point represents a single star. The median trends
are shown in the middle line, and the inter quartiles are shown above and below. The top rows of panels show
the losses in the uncorrected exposures. The magnitudes are given in instrumental units, −2.5 log10(number of
electrons), such that −10 is S/N ~ 100 and −5 is S/N ~ 10. When the sky background is zero, losses go from 0.1
magnitude at −10 to more than 0.5 magnitude at −7. When the background is 12 electrons, losses are never
more than 0.15 magnitude. The photometry on the corrected images is shown in the bottom panels.
Figure 3 shows that the pixel-based correction does a very nice job on images with some
amount of sky, but when the sky goes to zero, the correction becomes increasingly
inadequate as the source gets fainter. This is not surprising. In order to avoid readnoise
amplification, we had to be conservative in terms of what could be a source and what
could be noise. Even when the cores of faint sources could be distinguished from readnoise — say, for sources with 100 total counts (instrumental magnitude of −5) should
have 20 counts in their central pixels, which is 6 times the readnoise — the surrounding
pixels would not necessarily be identified as having significant flux, and as such would
not have the full correction applied to them.
Figure 4 shows the same results, but for astrometry. Again, the correction is quite good.
This gives us the hope that if the correction restores total flux and position for most stars,
it might also do a good job preserving shape. This, however, remains to be tested.
Figure 4: Same as the previous figure, but for astrometry. Positions were fit with an empirical library PSF for
F336W. Astrometry is clearly restored for most stars on most backgrounds.
5. Remaining issues
One issue that has come to light during our post-flash investigation is that
there are quite a few pixels that appear to have several traps in them. We
know this because the electrons that get added during the post-flash (or
background electrons in a science exposure) get trapped and do not make it
out of the pixel during the first transfer. For some reason, these pixels often
appear to be paired with WPs below them, which makes it somewhat hard to
interpret the trails behind WPs on images with background. The UVIS team
is investigating this phenomenon.
Figure 5: This is a close up of a post-flashed dark stack with a background of ~12 electrons. The whitest pixels
have values of ~2 electrons. The left panel shows a part of the image near the readout register, and on the right
shows a part far from the readout register. Black corresponds to higher values. Clear the WPs are CTEblurred on the right but not on the left. The white spots are pixels significantly below the background. These
also are sharp on the left and more blurred on the right.
The fact that these “holes” appear capable of holding several electrons may
hint to us that traps may not occur one at a time randomly on the detector,
but they may well be bunched up.
6. Anticipated improvements
The correction described here treats every pixel as having exactly the same
distribution of traps. The ACS algorithm is a bit more sophisticated than
this. Ogaz et al (in prep) studied the parallel overscan pixels for each
ACS/WFC column to determine a rough scaling for how many traps each
column had relative to the average.
We plan to do a similar exercise for WFC3/UVIS, but we can go even
farther. We can use the charge-injection procedure to add ~15,000 electrons
to lines separated by 10, 17, or 25 pixels. Examining the trails behind these
lines will allow us to estimate not only how many traps can be found in each
column, but where along the column the traps can be found. Of course, with
such a high injection level, we are not able to probe exactly which electrons
(first, hundredth, etc) each trap impacts, we can only estimate the total
number of traps. It might be possible to use the scan mode to move stars of
different brightness across the detector at a variety of rows to estimate the
loss in each column. The WFC3/UVIS team is exploring these options for
specifying the CTE model more locally.
References
Anderson et al 2013, “The Efficacy of Post-Flash for Mitigating CTE Losses in
WFC3/UVIS images”
http://www.stsci.edu/hst/wfc3/ins_performance/CTE/ANDERSON_UVIS_POST
FLASH_EFFICACY.pdf
Anderson, J. & Bedin, L. R. 2010 PASP 122 1035 “An Empirical Pixel-Based Corretion
for Imperfect CTE. I. HST’s Advanced Camera for Surveys”
Massey, R. et al. 2010 MNRAS 401 371
MacKenty & Smith 2013. “CTE White Paper”
http://www.stsci.edu/hst/wfc3/ins_performance/CTE/CTE_White_Paper.pdf
Ogaz, Sara et al. ISR in prep.