In a survey like the EIS, where a large variety of astronomical -- and non-astronomical! -- objects of all kinds can be detected and measured over wide areas, one cannot avoid making choices. In the case of EIS, the priority is the detection of objects such as faint stars and galaxies. Brighter objects are generally saturated and/or already cataloged. The source extraction is performed with a new version of the SExtractor software (Bertin & Arnouts 1996) that can be retrieved from "http:/www.eso.org/eis". SExtractor is optimized for large scale imaging survey fields with low to moderate source density, and is therefore perfectly suited to EIS. The processing is done in 3 steps: detection, measurement, and classification, which are briefly described for the single image process.
Background-subtracted images are then filtered before being thresholded, to reduce the contribution of noise on spatial scales of the image where it is dominant. The median seeing (FWHM) of EIS images as a whole is about 0.9 arcsec, a little more than 3 pixels. The data are filtered by convolving with a slightly larger, constant, Gaussian profile with FWHM = 4 pixels. Although the choice of a convolution kernel with constant FWHM may not always be optimum (the seeing may vary by as much as a factor of 3), the impact on detectability is, however, fairly small (see Irwin 1985). On the other hand, it has the advantage of requiring no change of the relative detection threshold. It also simplifies the comparison with the coadded-image catalog, for which the convolution kernel is also fixed.
The detection threshold, , used in SExtractor is expressed in
units of the standard deviation
of the background noise. For
single images k = 0.6 is used, which corresponds to a typical
limiting surface brightness
mag/arcsec2. The new SExtractor
allows this noise-level to be set independently for each pixel i,
using a weight-map wi (Sect. 3.7), which is internally
converted to a relative variance:
. The
variable detection threshold is also used for deciding if a faint
detection lying close to a bright object is likely to be spurious or
not.
Some pixels are assigned a null weight by Weight Watcher, because they are unreliable: gain too low, charge bleeding, cosmic-ray, etc. The detection routine cannot simply ignore such pixels, because some objects, like those falling on bad columns or charge bleeding features, would be either truncated or split into two. A crude interpolation of bad pixels overcomes this problem. Unfortunately, interpolation creates correlated patterns which are sometimes detected at the very low thresholds applied in the EIS, but as these zones are flagged, they are easily filtered out in the final catalog.
The standard SExtractor star/galaxy classifier is a multilayered back-propagation neural-network fed with isophotal areas and the peak intensity of the profile. The classifier was trained with simulated ground-based, seeing-dominated, optical images. It will therefore perform well on images close to the conditions met in the original simulations. This is so for EIS images in patch A, but it is no longer the case for other patches, where very good seeing and strong optical distortions yield significantly elongated and skewed stellar profiles, varying over the frame. A new, more general, star/galaxy separation scheme is therefore needed for these fields, and is currently being implemented in SExtractor.
The current classifier returns a "stellarity index''
between 0 and 1. A value close to 0 means the object is extended
(galaxy), while a value close to 1 indicates a point-source (star). It
can be shown that the neural network output is approximately the
probability that an object is a point-source. This
is only valid for a sample of profiles which would be drawn from the
same parent population as the training set. Because the neural
classifier is a finely tuned system, these conditions are almost never
met with real images, and care has to be taken when interpreting the
stellarity index. Nevertheless, it is fair to adopt a stellarity index
value of 0.5 as a default limit between point-sources and elongated
objects. At faint levels (), star/galaxy separation begins
to break down for frames obtained under the least favorable seeing
conditions in patch A. A clump begins to form around a stellarity
index of 0.5, indicating that the algorithm cannot provide a reliable
classification for most objects.
In the discussion below two values of the stellarity index are adopted to separate stars and galaxies: the conservative value of 0.5, which tends to favor more complete star catalogs, and a value of 0.75, which assumes that beyond the classification limit galaxies largely outnumber stars.
During the extraction SExtractor sets several flags to describe any
anomalies encountered. The meanings of these flags, , are
summarized in Table 2. Information available in the
flag-maps generated by the Weight Watcher program are also propagated
to the catalog. Flags are set to indicate that a given object is
affected by bad pixels in the CCD-chip or by artifacts in the image
that have been marked either by the artificial retina or by the
polygon-masking during visual inspection
(Sect. 5.4). Table 3 summarizes
the meaning of the flags in the catalog set from the information
contained in the flag-maps,
.
The contents of the catalogs include: J2000.0 right ascension and
declination, x and y coordinates in the chip; total magnitude
(MAG_AUTO) and error; major and minor axes; position angle;
stellarity index; SExtractor flag (see Table 2);
Weight Watcher flag
(see Table 3); total
number of pixels above the analysis threshold (npix); total number of
pixels that are flagged by Weight Watcher (nflag). Further information
can be found at "http:/www.eso.org/eis".
During the processing of a patch through the pipeline the single frame catalogs are merged together into a "patch'' catalog which contains information of all objects identified in the individual frames. Note that objects may have multiple entries if they are in overlapping frames. From this patch catalog several single entry catalogs may be derived, for instance, the even/odd catalogs containing all objects detected in the even/odd frames. Objects detected in more than one frame are identified to produce a single-entry in the final catalog, choosing the parameters as determined from the best-seeing image. Objects in regions of overlap are paired whenever the barycenter of the smallest falls within an ellipse twice the size of the object ellipse of the larger one. Details on this procedure will be presented elsewhere (Deul et al. 1999).
From the flag information available in the single-entry catalog,
filtered catalogs can be produced for analysis purposes (see
Sect. 6). The filtering is required to eliminate
truncated objects and objects with a significant number of pixels
affected by cosmics and/or other artifacts. Objects with the following
characteristics are discarded: or nflag/npix
,where
is the SExtractor flag, npix the number of pixels above
the analysis threshold and nflag the number of pixels flagged by
Weight Watcher. The two-dimensional distribution of stars and galaxies
from the resulting catalog are shown in Figs. 15 and
16, for different limiting magnitudes.
![]() |
Figure 16: Same as in previous figure showing the distribution of 9006 (upper panel) and 23129 (lower panel) galaxies for the same two limiting magnitudes as in Fig. 15. Again note the empty regions in the upper part of the patch caused by bad frames |
The visual inspection of the catalogs was done using the new version of ESO SkyCat which also provides the possibility of accessing the EIS catalogs through the on-line server. Further information on the SkyCat setup can be found at "http:/www.eso.org/eis". This setup interprets the parameters and flags available in the EIS catalogs. To distinguish between the different object classes and flags, the following plot symbols and colors have been used:
This tool has been extensively used to fine-tune the configuration parameters used by SExtractor and Weight Watcher as well as to inspect the performance of the filtering of the catalogs (see Sect. 5.3). Users of the catalogs should be aware of the following features:
The visual inspection shows that, by adopting the filtering criteria described in the previous section, most of the spurious objects are appropriately removed.
As a first check on the quality of the object catalogs produced by the pipeline it is important to examine the uniformity of the detections across the effective area (excluding the regions masked out) of the EMMI-frame. This is shown in Figs. 17 and 18, where the normalized average counts of stars and galaxies as a function of the east-west (Fig. 17) and north-south position (Fig. 18) on the chip are displayed. The upper panels show the star counts brighter than I = 21, which is the limiting magnitude for reliable classification in patch A as a whole. The lower panels show the galaxy counts to the same limiting magnitudes.
The overall uniformity of the detections at magnitudes is
good. A small decrease in the number of stars is seen at the upper edge
of the chip and is almost compensated by an increase in the galaxy
counts. This behavior is likely to be due to misclassifications caused
by the increase in size of the PSF as shown in Sect. 3.5.
![]() |
Figure 18: As Fig. 17 but showing the detections in the north-south direction. Again at the "upper" edge of the chip we see a dip in the star counts and a corresponding increase in galaxy counts, which is due to the image distortions, see text for details |
To verify the pointing of the telescope, a reference field has been
observed before the start of each row (150 s) and, in some cases,
the start of sub-rows (50 s). These exposures, which for patch A
total 2250 s, were used to determine the offset required to
compensate for the problems detected with the NTT pointing
model. Using the EIS pipeline these images have been coadded and an
object catalog was produced extending to fainter magnitudes. This
catalog has been used to empirically determine the completeness of the
detections in typical single-frame EIS catalogs. This was done by
comparing the catalog produced from the coadded image of the reference
field to the individual catalogs derived for the various exposures
of that field. Since the limiting magnitude of the coadded image is
much fainter than that of the single frames, one can assume that the
coadded catalog is complete and that it is not significantly
contaminated with false objects at least to the limiting magnitude of
the single-frame catalogs. Keeping this in mind a match was made
between all the objects in the coadded catalog and those found in the
single-frame catalogs. The ratio between the number of paired objects
and the total number of objects in the coadded catalog provides a
measure of the differential completeness as a function of magnitude,
which is shown in Fig. 19. The completeness defined
in this way only measures the fraction of objects actually found, but
does not tell anything about the reliability of their properties. It is seen from the
figure that for objects of magnitude the completeness is
80%. At this magnitude the integrated completeness of the
catalog is 94%. The completeness does not vary for seeing between 0.7
and 1.3 arcsec. For a seeing of 1.5 arcsec the 80% differential
completeness limit is at
.
The number of false detections can be estimated in a similar
way. Figure 20 shows the ratio between objects that
were found in a single-frame catalog but not in the coadded one and
the total number of objects in the single-frame catalog. The figure
is based on a comparison obtained using a single frame with a seeing
of 1.07 arcsec, which is close to the median seeing of the
observations for the patch. It is seen that at 10% of
the objects are false detections and 20% of the objects with
magnitude
are spurious. The integral fraction of spurious
objects up to the limiting magnitude of I=23 is
6%.
A comparison between even and odd catalogs provides further useful information on the accuracy of the magnitudes and on the robustness of the classification as a function of magnitude. This comparison was done using a test region of 0.6 square degrees, with a median seeing of 0.95 arcsec. Using the same pairing procedure previously discussed, a catalog of paired objects in the test region was produced.
A lower limit estimate of the photometric errors can be obtained from
the repeatability of the magnitudes of the paired
objects. Figure 21 shows the magnitude difference of
these objects as a function of magnitude. The standard deviation of
the magnitude differences in the interval 16 < I < 20.5 ranges
between 0.02 and 0.1, reaching 0.3 at .
Figure 22 shows a comparison between the errors determined
from the magnitude difference shown above (divided by ) and
the SExtractor error estimates based on photon statistics. SExtractor
provides reasonable error estimates over the interval of interest. At
bright magnitudes photometric errors are dominated by effects such as
flatfield errors, image quality, intrinsic stability of the MAG_AUTO
estimator and relative photometry.
![]() |
Figure 22: Comparison between the standard deviation of the EIS-magnitude differences (filled squares) and the magnitude errors estimated by SExtractor (open squares) as function of magnitude |
For objects in the magnitude range 16 < I < 21 and adopting a
stellarity index of 0.75 to separate stars and galaxies, about 5% of
the objects
have different classifications in the even and the odd catalogs.
For magnitudes most objects are saturated and may be
classified as galaxies. However, they can be found as having the flag
, which has been used to exclude them from subsequent analysis.
Copyright The European Southern Observatory (ESO)