-------------------------------------------------------------------
The Hubble Ultra Deep Field ACS/WFC Combined Images - V1.0
==========================================================
Space Telescope Science Institute
Date: Tuesday, 9th March 2004
Initial Document written by: Anton M. Koekemoer
Incorporating Revisions by: Steven V. W. Beckwith
1. Introduction
The Hubble Ultra Deep Field (PI: Steven V. W. Beckwith) is a
400-orbit Cycle 12 program to image a single field of the Wide
Field Camera (WFC) of the Advanced Camera for Surveys (ACS) in
four filters: F435W (B), F606W (V), F775W (i), and F850LP (z).
The observations took place over 4 months from September 2003 to
January 2004 under two program IDs: 9978 and 10086.
The observations consist of half-orbit exposures, cycling through
each of the filters in a 4-point dither pattern to provide sub-pixel
sampling, as well as a larger-scale 3-point line pattern to cover
the 2 second of arc gap between the two ACS/WFC chips. The total
exposure times are summarized below, with typical exposure times
of 1200s for individual images. The AB magnitude zeropoints for ACS
are current as of March 2004.
Number of Number of Total Exp. AB mag.
Orbits Exposures Time (s) zeropoint
B (F435W): 56 112 134880 25.673
V (F606W): 56 112 135320 26.486
i (F775W): 144 288 347110 25.654
z (F850LP): 144 288 346620 24.862
The observations used two primary roll angles separated by
90 degrees in each of two epochs. Each of the two epochs contained
observations at two roll angles differing by 4 degrees. Thus, there
were four roll angles in all. A roll angle, PA, refers to the
orientation on the sky of the -V3 axis of the spacecraft (thus
the +ve y-axis of the ACS chips). For each filter, the number of
orbits for a particular date range / PA combination are:
Observation Dates PA B V i z Tot. orbits
09/24/2003 - 10/02/2003 310 6 8 18 18 50
10/03/2003 - 10/28/2003 314 22 20 58 56 156
12/04/2003 - 12/22/2003 40 6 8 18 20 52
12/11/2003 - 01/15/2004 44 22 20 50 50 142
Total orbits: 56 56 144 144 400
The data were initially calibrated using the "calacs" software
(ACS Data Handbook, Mack et al. 2002), and subsequently combined
using a pipeline version of the MultiDrizzle software (Koekemoer
et al. 2002). Here we provide more details on all these steps.
2. Data Processing and Calibration
Each of the 800 ACS/WFC exposures was processed through the standard
HST data pipeline, OPUS. Each exposure was first calibrated using
the "calacs" software within STSDAS. Calacs removes the bias level,
subtracts the dark current, corrects the flat field and gain, flags
known bad pixels, and does an initial photometric calibration.
Special calibration files were created for these observations:
* Hyperdarks were created using all the dark-current frames from
the 6-month period encompassing the UDF observations. These files
(provided by M. Mutchler of the ACS team) provide a higher S/N
than the usual dark reference files that are subtracted during
standard calibration and give better overall correction at the
depths of the UDF.
* New flatfield images were made using the "L-flat" technique based
on stellar photometry of 47 Tucanae (provided by J. Mack of the
ACS group). These flatfield images produced a more uniform sky
level across the images than the standard pipeline products down
to ~1 - 2% of the sky brightness.
* Sky-flats were generated to correct the ~1-2% residuals that
remain after calibrating the data with the L-flats. For each
band, the sky flats resulted from six data processing steps:
1) An image of the field was created from the pipeline data by
combining all exposures.
2) This image was used to create a mask identifying the spatial
extent of all the astronomical objects in the image; this
process also provided cosmic ray masks.
3) A median image was created from the original calibrated
exposures using the combined object masks and cosmic ray +
bad pixel masks to exclude all pixels affected by celestial
sources, cosmic rays or bad pixels.
4) The median image was averaged over 100 pixels (convolved with
a 100 pixel smoothing function), corresponding to the spatial
frequency of the flat-field residuals.
5) The calibrated files were divided by this smoothed median
image to produce an image of the sky.
6) This sky image was subtracted from each individual calibrated
image.
* A bad-pixel file, containing a number of additional bad columns
and other defects not present in the standard data quality arrays
when the data were processed. The bad pixel file for each band
was created after correcting the images with the improved sky
values in the previous step, then subtracting an image made by
combining all exposures. The resulting images made it easy to
identify bad pixels and cosmic rays that were significantly above
the noise level. A difference image was then created for each
exposure, expressing the difference in terms of the underlying
r.m.s. of each pixel (including Poisson noise from any objects).
All 800 images were subsequently combined using both median and
averaging, to produce an image that showed the median or average
deviation of each pixel. Sigma-clipping was used to identify
additional bad pixels. This technique was especially valuable in
identifying charge traps not seen in the original data quality
file.
In addition, the bias level correction performed by "calacs" (see
ACS Data Handbook 2002, Mack et al.) did not fully remove the bias
levels in the 4 amplifier quadrants, but left slight residual
offsets of a few tenths of an ADU. These were corrected by means of
a special script (provided by M. Sirianni, ACS group) which first
reversed the multiplicative flatfield correction, then solved
iteratively for the residual bias differences between the quadrants
and removes these, before re-applying the flatfield.
The last remaining effect in the data is an apparent amplifier
cross-talk between the four amplifiers of the ACS/WFC chips, which
produces very low level dark "ghosts" evident as mirror images of
bright objects in the other quadrants. Some efforts have been made
to understand this effect (M. Giavalisco et al., in preparation)
and the ACS group is currently in the process of constructing a
well-understood physical model for it. However, its amplitude is
very low, and detailed examination of objects in the affected
regions showed that its effect on the data is well below any
scientifically significant level.
3. Astrometric Correction
Astrometric corrections were applied to the individual exposures
after removing the instrumental signatures from individual exposures
as described above.
We used a set of distortion correction images (Anderson et al. 2003)
to remove the residual distortion not accounted for by the 4th-order
polynomial that is used for astrometric calibration in the standard
ACS pipeline (Cox, 2002, ACS-ISR-02-02). R. Hook created a version
of drizzle, called wdrizzle, that used the information from these
distortion residual images when drizzling each input exposure onto
an astrometric grid. The output images were aligned to accuracies
better than 0.05 - 0.1 ACS pixels (2.5 - 5 milliseconds of arc)
across the entire field of view.
Once image distortion had been removed by this procedure, the
separate exposures were aligned on a common astrometric grid by
using an external reference frame derived from the GOODS dataset
(Giavalisco et al. 2003.) This method is more accurate than the
standard use of guide stars to align the separate exposures.
The UDF is projected onto the same tangent point as the GOODS
reference frame defined by the pointing center of the Spitzer IRAC
and MIPS ultra-deep imaging that is being obtained in this part of
the sky. A set of GOODS images was produced that overlapped the UDF
(Koekemoer et al) and catalogs were generated based on these images.
The catalogs from the GOODS ACS/WFC images were used as the
astrometric baseline for registering each UDF exposure. Each
exposure was cleaned of cosmic rays using a single-image cosmic ray
rejection algorithm, then is was compared with a catalog to match
objects within 2 seconds of arc of GOODS sources. Typically 200 -
800 objects matched, depending upon the band. The resulting
differences between the object positions - GOODS vs. UDF - were
used to determine a shift and rotation correction for each exposure,
typically yielding overall accuracies of about 0.0002 degrees for
the rotation and 0.1 pixels for the shift. This accuracy was
sufficient to continue with the cosmic ray rejection and produce a
first-pass UDF image.
After a first-pass combined UDF image had been produced with cosmic
ray masking, each of the input exposures was modified by replacing
each of pixels affected by cosmic rays with the values from the
combined UDF image. This procedure yielded a clean set of input
exposures to compare with the GOODS catalog for improved astrometric
registration. After this step, the residual error in the rotation
was less than 0.0001 degrees, and the residual translation errors
were less than 0.01 pixels r.m.s. (500 microseconds of arc.) These
astrometric shifts and rotations appear in the WCS header keywords
for each input exposure. Finally, the cosmic ray and final drizzle
combination were repeated.
4. Cosmic Ray Rejection
Each exposure was transformed onto a separate set of rectified,
registered output images, using the "drizzle" task to remove the
geometric distortion, and using the new updated astrometric keywords
to ensure accurate registration. The rectified image for each band
were then combined to create a clean median image using
sigma-clipping to remove pixels with values many standard deviations
from the mean.
The clean median image was then transformed back to the distorted
frame of each input image using the WCS-based "wblot" program
provided by R. Hook, effectively applying the inverse transformation
to "wdrizzle". The input image was then compared with this clean
image, as well as with its derivative, to reject pixels deviating
from the clean image using the following algorithm:
abs(Input - Clean) > scale*deriv + SNR*sqrt( rn*rn +
gain*abs(Clean + background) ) / gain
where "scale" was set to 1.5, "deriv" is the derivative image,
"SNR" was set to 3.5 (corresponding to 3.5 sigma), "rn" was the
read-out noise, and background was the sky value measured in the
original input image.
This procedure is essentially sigma-clipping, with the addition of
the derivative image which helps to soften the rejection and prevent
pixels being incorrectly identified as cosmic rays in regions of
extremely sharp gradients, such as near bright stars. A second
iteration was carried out using somewhat more stringent cuts,
masking additional pixels around those masked in the first pass.
This algorithm is the standard "driz_cr" that was initially
implemented for the Hubble Deep Fields (Fruchter and Mutchler); the
parameters were tuned for the UDF data by a process of iteratively
testing the effects of different combinations to ensure robust
rejection of all cosmic rays whlie at the same time avoiding
over-rejection of good pixels in bright objects.
5. Image Combination
The cosmic ray masks were then used as input to the final drizzle
combination of all the images. The "drizzle" program (Fruchter and
Hook 2003) essentially performs a weighted sum of the input images,
and allows input pixels to be shrunk by a specific amount before
being mapped onto the output plane. The final pixel scale was set
to 0.6 of the input ACS pixels or 30 mas/pixel. Since this scale
provides Nyquist-limited sampling of the PSF in all 4 bands, it is
the optimum pixel size to use.
The images were weighted before combination. The weights were the
inverse variance of each exposure. The inverse variance has the
appealing property that the final summed image is properly weighted
with a resulting output inverse variance image that is simply the
sum of all the input inverse variance images. An input variance
image was calculated separately for each exposure, taking into
account the flatfield variation, the sky level, and all the bad
pixel information.
Since the UDF has such a large number of pointings in each band
with a good sampling of the sub-pixel space, the final images
were drizzled using a point kernel (i.e. setting "pixfrac" = 0),
corresponding to pure interlacing, so that each input pixel maps
onto only a single output pixel. This choice has the advantage of
minimizing the amount by which the output image has been processed
(convolved with a smoothing function), thus providing the sharpest
possible image.
The images are oriented with North toward the top (increasing y-axis
values.) The tangent point for the UDF, defined as the point at
which the reference pixel of the image intersects the spherical sky,
is located at the tangent point for GOODS:
RA = 53.122751, Dec = -27.805089 (J2000).
This tangent point defines the astrometric grid for the UDF and
also for the GOODS multi-waveband dataset in this portion of the
sky. It is recorded in the CRPIX and CRVAL keywords for the images,
as follows:
CRVAL1 = 53.122751
CRVAL2 = -27.805089
CRPIX1 = 9470.5
CRPIX2 = 3610.5
Note that the reference pixel values are set to half-integer values.
As a result, the corner, rather than the center, of the input pixels
were mapped to the reference pixel when drizzle was run, thus better
facilitating down-sampling or block-averaging of the image. In this
reference frame, the image center is at the UDF target coordinates
of RA = 03:32:39.0, Dec = -27:47:29.1 (J2000) while the reference
pixel in the image corresponds to the GOODS tangent point.
The final ACS/WFC data products consist of 2 images for each of
the four bands:
Filename Description
h_udf_wfc_b_drz_img.fits F435W drizzled image (electrons/s)
h_udf_wfc_b_wht_img.fits F435W inverse-variance weight image
h_udf_wfc_v_drz_img.fits F606W drizzled image (electrons/s)
h_udf_wfc_v_wht_img.fits F606W inverse-variance weight image
h_udf_wfc_i_drz_img.fits F775W drizzled image (electrons/s)
h_udf_wfc_i_wht_img.fits F775W inverse-variance weight image
h_udf_wfc_z_drz_img.fits F850LP drizzled image (electrons/s)
h_udf_wfc_z_wht_img.fits F850LP inverse-variance weight image
Each image is 10500x10500 pixels (total size 430 Mb each), with a
pixel scale of 0.03"/pixel, and oriented with North toward the top.
6. Cataloging
We used the SExtractor program (Bertin & Arnouts 1996) to produce
catalogs from the final drizzled ACS/WFC images, using the inverse
variance images as input for calculating the r.m.s. uncertainties
associated with the flux measurements for each source. The catalogs
were initially run on the i-band image, which is our deepest image
and thus provides the most complete sample of sources. Subsequent
catalogs were produced in "dual-image" mode, using the i-band
isophotes of each source to measure its photometry in the other
bands, thereby producing isophotally matched magnitudes that can
be directly compared to produce colors for each source.
The SExtractor parameters were optimized for the characteristics of
the UDF ACS/WFC images, specifically our pixel scale and PSF. We
used a seeing FWHM of 0.09 seconds of arc, and required a minimum
of 9 contiguous pixels with a detection threshold above 0.61 sigma,
with a total of 32 deblending sub-thresholds and a contrast
parameter of 0.03. More details on these and other parameters will
be presented in Beckwith et al. (2004, in preparation). Our output
parameters are:
- source ID number
- X and Y pixel location of the source on the UDF images
- Right Ascension and Declination (J2000)
- source orientation "Theta" (counterclockwise from the X-axis)
- ellipticity (1 - B/A, where A is the semi-major axis)
- R50, the half-light radius (in pixels)
- FWHM of a Gaussian fit to each source (in pixels)
- Stellarity (1 for point sources, 0 for fully resolved sources)
- AB magnitudes for each band (B,V,i,z), isophotally matched
- formal AB magnitude errors (from rms map + Poisson statistics)
- S/N (dividing the measured flux by the r.m.s., which includes
all the information from the inverse variance weight map).
The formal i-band catalog contains a total of 10,040 sources. A
visual inspection of all the sources revealed an additional 5
spurious sources (which do not form part of the catalog). Moreover,
the deblending algorithms in SExtractor caused an additional 100
sources to be missed, owing to their proximity to brighter
sources. These sources were identified manually, and formally
added by doing another SExtractor run with considerably different
deblending parameters, in order to detect them all. An initial
list of 208 sources was produced, which was then reduced to a
total of 100 sources after visual inspection and rejection of
sources that were clearly part of previously identified sources.
These additional sources are denoted by ID numbers 20001 - 20208.
Although the i-band image is our deepest image, there remain
additional sources that were not detected in i-band, even
though they may be detected in one of the other bands. Therefore
we produced a second catalog based on detection in the z-band
image, and we include from this catalog an additional 39 sources
that are detected at > 10 sigma in the z-band image, but are not
in the catalog that was run using the i-band image for detection.
These additional sources are denoted by ID numbers above 30000.
This catalog is accessible as the file:
h_udf_wfc_V1_i_cat.txt
We also present a z-detected catalog which has the same format as
the i-detected catalog, namely with magnitudes, errors and S/N
values in the other 3 bands (in this case B, V, i) that are
measured within the same isophotes as those detected in the z-band.
This catalog is accessible as the file:
h_udf_wfc_V1_z_cat.txt
Finally, we present segmentation maps for both the i-detected and
the z-detected catalogs, which are called:
h_udf_wfc_i_seg_img.fits.gz
h_udf_wfc_z_seg_img.fits.gz
These segmentation maps are integer FITS files, on the same scale
as the full-size UDF images (30 milliseconds of arc / pixel), and
with pixel values set to 0 where there are no objects, otherwise
set to an integer corresponding to the ID number of the object in
the catalog, thereby mapping out on the image the spatial extent
of each object in the catalog.
We caution that these catalogs are highly preliminary; they are
likely to be representative of most of the sources across the
field, but we have not had the opportunity to perform detailed
statistical analyses or simulations of their completeness and
related properties, particularly toward fainter magnitudes.
Similarly, the S/N values are likely correct to no more than
about 5-10%, depending upon the detailed structure of the weight
map underlying each object. Therefore we urge the community to
consider performing independent cataloging analyses if these
effects are critical to the science that the data are used for.
References:
Anderson, J. 2002, HST Calibration Workshop, Eds. S. Arribas,
A. M. Koekemoer, B. Whitmore (STScI: Baltimore), p. 13
Beckwith, S. V. W. et al. 2004, in preparation
Bertin, E. & Arnouts, S. 1996, A&AS 117, 393
Cox, C., 2002, Instrument Science Report ACS-ISR-02-02
Fruchter, A. S. & Hook, R. N. 2002, PASP 114, 144
Giavalisco et al. 2003, astro-ph/0309105
Koekemoer, A. M., Fruchter, A. S., Hook, R. N., & Hack, W. 2002,
HST Calibration Workshop, Ed. S. Arribas, A. M. Koekemoer,
B. Whitmore (STScI: Baltimore), p. 337
Mack, J. et al., 2002, ACS Data Handbook
-------------------------------------------------------------------