Apply wavdetect to locate sources
This is applied to a 2n portion of the image because FFTs are involved. Wavdetect
may take 1 CPU hour of Ultra 60 time for a 512x512 region. We have found that
wavdetect performs impressively in locating sources with nearly-uniform sensitivity
across the ACIS-I field and without producing spurious ghost sources around
off-axis sources. We have not evaluated its performance on extended structures,
though it definitely can find compact sources within diffuse emission.
wavdetect
infile=Merge2.fits[energy=500:2000][bin x=3800:4311:1,y=3800:4311:1]
scellfile=Region_wav_srcimg.fits
imagefile=Region_wav_img.fits
defnbkgfile=Region_wav_bkg.fits
outfile=Region_wav_srclst.fits
The scellfile image plots the source positions with integer values increasing
with right ascension. The imagefile is more interesting, with sources
plotted with values scaled to intensity and size scaled to wavelet scale. The defnbkgfile
shows the background variations not included in the wavlet sources,
shown as little circles. These can all be examined using ds9. The
outfile is a FITS table file giving many parameters associated with each source. It
can be examined using fv or fdump, and is used in later analyis.
Note 1. For unknown reason, the command has failed unless all
output files are explicitly defined.
Note 2. Three variations on the parameters have been used within
the ACIS Team. First, we generally follow the recommendation of the CIAO
manual to increase the density of wavelet scales as follows (excerpt from the
wavdetect parameter file):
scales,s,h,"1 1.414 2 2.828 4 5.66 8 11.311 16",,,"wavelet scales (pixels)"
Of course, even larger scales are needed to locate and model diffuse features.
Second, sometimes wavdetect is run on two or three bands separately (e.g. 0.5-2
keV, 2-8 keV) and sometime on a unified band. No quantitative study of
the efficacy has been made. Third, the detection threshold has been run
at 10^-5, 10^-6 and 10^-7 in different circumstances. Recall that is it
wise to make a local copy of the wavdetect.par parameter file if parameters
are changed, so that future runs are not affected.
Note 3. Wavdetect is remarkably successful in discovering pointlike sources
in a sensitive and uniform fashion across the ACIS-I field. It does not
divide strong off-axis sources into multiple sources, as celldetect does. Nonetheless,
occasional failures occur and the scientist must check the results
visually. The following problems are noted by E. Feigelson from study
of the Orion Nebula field with ~1000 sources having stellar identifications. False
negatives include: missing some sources if they lie within 2.5" of another
(particularly brighter) source; missing some sources around its sensitivity
limit. False positives include: detecting spurious sources on the readout
trails of bright sources; and detecting noise, particularly off-axis when the
detection threshold is set below 10-6.
This last problem deserves some discussion. The faint spurious sources located
with wdetect are characterized by wrecon to have virtually no source counts; e.g.
NET_COUNTS + background = 2 or 3 events when the overall source limit is ~7 counts.
Examination shows this can happen under three circumstances: the background map
has unusually high values (>1 count); there is a chance nearby void in the distribution
of background events; or when the photons are distributed over an extremely large
area (NPIXSOU > 200). False positive detection may improve when the option of
a constant, rather than variable, background map is implemented.
Note 4: The "exposure map" in units s*cm2 map (recall this is map
produced by mkexpmap times the CCD-dependent
EXPOSUR? time in the header; see the exposure Recipe
page) may be passed to wavdetect under the "expfile" parameter, so that the
NET_RATE column of the wavdetect source catalog has units events/s/cm2 (despite
the fact that the TUNIT keyword of that column is "count/s").
Note 5: The wavdetect output unfortunately does not state which sources where
detected at which spatial scales. However, we believe one can infer this from the
parameter NPIXSOU in the Region_wav_srclst.fits file. NPIXSOU values around 10-20
are typical for unresolved sources near the field center, while values can rise to
102-103 for large extended features.
Note 6: Beware running more than one wavdetect simultaneously on the same
event file, even if they run on different CPUs. Bad interactions can ensue.
Note 7: There appears to be a bug in wrecon: if one runs the program with
very large wavelet scales to find diffuse structures, the reconstructed image
frequently shows truncated large wavelets which produce oddly shaped arclike structure
that are clearly spurious.
---ooo---
To apply wavdetect to an entire chip or array, T. Maeda has written a Perl
script Large_detect.pl to automatically
conduct a series of 1024x1024 wavdetect runs that scan across a chip or any
array of chips, producing a consolidated source list. It also runs celldetect
across the array, though we do not recommend using celldetect results. The command
is:
Large_detect.pl input.evt outroot binfactor
where input.evt is the input event list, outroot is used in the names of output
files, and binfactor is the pixel binning used by the detect programs
(usually = 1). The user may reset the parameter $imgsize setting
the size of regions for each run of wavdetect. It is currently set at
1024 and must be a power of 2. The parameter $margin, currently
set at 128, may also be changed. It causes adjacent wavdetect segments
to overlap so sources are not missed at the edges. (Duplicate sources
from overlapping regions are removed later in the script.) The output
file of interest is the merged wavelet source list with source characteristics:
outroot_wavdetect_srclist.fits
Many other files are produced, and it is recommended this be run in a separate directory.
This program typically takes >10 CPU hours to run on a Sun
Ultra 60 workstation for the ACIS-I array.
Note 1: This program has occasionally crashed with error messages from
celldetect about a missing PSF file. Cause and cure are unknown.
Note 2: The off-axis point-spread function seems to be too wide to find
with the standard scale of 1-16 and the 1 bin size. Hence, it may be wise to
run it separately with binsize=1 near-axis and binsize=4 far off-axis.
Note 3: The background level between FI and BI could be very different,
and it is recommended to run wavdetect separately for these types of chips.