coadd that will be used for initial event detection. A full evaluation of DIFFIM can be found in Kessler et al. (2015), but we will summarize the differenced imaging pipeline here.
The first step in the difference imaging process is using codemakeweightto recalcu-
late the weight map for the single tile search images (not the coadded image). In addi- tion, it also masked image defects in order to prevent subtraction issues further down the
pipeline. SExtractor is then applied to the image in order to make a star catalog that
will be used both for zero point determination and withPSFex in order to calculate the
PSF for the search image.
Once the zero point and psf for the search image have been determined a code called
doFake is used to insert artificial SNe into the individual search images. For more in- formation on artificial SNe and their use in the survey, refer to the section below titled
"Artificial Supernovae." At this point in the processing, we use SWARP to do a median
coadd all of the images from one cutsfilter-epoch in order to get full depth from our z-
band and deep field imaging. Again, once the coadd is complete,SExtractorandPSFex
are run on the coadd, as they were on the individual images to create a start catalog and determine the PSF of the combined image.
With the coadds now available for image substraction, we distort our templates to
match the astrometry for both the combined and single exposure image usingSWARPbe-
cause that functionality is not included in the actual difference imaging code that we use.
hotpants, based on a method of image subtraction that is described in detail in Alard
(2000), is the code used for image subtraction. In general, hotpants finds a spatially
varying kernel that it will convolve with the template image in order to match the seeing of the search image prior to image subtraction. The template usually has better seeing than the search image, so most of the time, during Y1, the template was convolved to match
the new observation. However,DIFFIMknows which image actually has the smaller PSF
(which can happen when the search image is taken in particularly good conditions) and
forceshotpants to convolve that image to match the other. Without this functionality,
hotpantswould attempt to find a kernel to reduce the PSF size of the template in cases where the search image has particularly good seeing would produce differenced images of very poor quality.
Image subtraction completed, DIFFIM runs a final instance of SExtractor on the
newly created differenced images looking for things that exist above S/N of 3.0. If there
is a combined search image then SExtractor is run on that, otherwise it is applied to
the differenced image produced from single search images. The produced detections,
which we denotes-detections, and a variety of associated parameters are sent to a code
filterObj, which does most of the remaining work filtering useful detections into a catalog that is provided to DES scientists. A more detailed description of the work done byfilterObj can be found in § 2.2.4, where we describe our automated SN detection process.
AfterfilterObjcompletes processing,DIFFIMruns thefakeMatchalgorithm which
compares s-detections to the fakes inserted bydoFakeand uses them to measure realtime
pipelne efficiency. The output from fakeMatch is utilized later by our online pipeline
monitoring to help us determine both data quality and the success or failure of bothSNeSE
andDIFFIM. During SV, thefakeMatchoutput was produced to into a log file, but during Y1 and future seasons it will write quality assurance data to the DES Oracle database.
The final step ofDIFFIM, makeStamps creates 51 by 51 pixel “stamps" centered on
every detection thacutst passes throughfilterObj. In previous (and potentially future)
versions ofDIFFIM), this step came prior tofakeMatch, but was actually decoupled from
season (SV) and Season 1 (Y1), 25,000 artificial Type Ia light curves were simulated
using the nominal SALT2 model withinSNANA. snfakethen utilized the PSF model de-
termined in the detrending pipeline to add artificial PSFs to the SN field seach images. The S fields, which exist within SDSS Stripe 82, had fakes places in a realistic angular distribution around host galaxies, which themselves were chosen to represent a somewhat realistic host magnitude distribution. During SV and the first half of Y1, fakes were ran- domly distributed throughout the other 8 fields, without concern given for the location of potential “host" galaxies. This was done because those fields lacked a well calibrated cat- alog of field galaxy locations and photometry. By the second half of Y1 (starting October 21, 2013) a new set of fakes were generated and placed in all 10 fields in the same manner as those placed in the S fields. This did not have a major impact on image quality tests, as sky noise in the images dominated host noise in all but the brightest hosts, but it did provide more realistic differenced image stamps to be used later by our machine learning code.
Every observation from the SV season contained dozens, if not hundreds, of fake su- pernovae that could be used to assure the quality of the images. If too many “bright" fakes
were missing from theSExtractorcatalogs generated from the resulting differenced im-
ages, then the observed image could be flagged as a possible failure. An example plot used to determine the quality of an exposure can be found in Figure 2.5. During the first season of the survey proper, we placed ˜25,000 fake Type Ia supernovae light curves using galaxy catalogs developed from the SV data. Additionally, to improve our future image quality measurements, 8 artificial stars of magnitude 20 were placed on each search im- age, far away from any known stars and galaxies. At this magnitude all of these bright fakes should be detected on the difference image of a “good quality" observation with high S/N. We found during SV that, while the fake discovery rate was useful in determin- ing image quality, too often the number of bright fakes found on a single chip was simply too low to be used as a consistent image quality measurement proxy. For the deep fields the required S/N was 80 and for the shallow fields the requirement was 20 for an image
to be considered “good." That way, after the images have been processed the detection statistics can be displayed quickly to the whole working group. This was done via a web page so that the SNWG could provide fast, quantitative feedback to the observers. If a field wide exposure (as opposed to an individual CCD image) was determined to be of
poor quality through this process then the CTIOobstacdead-man countdown would not
be reset after the exposure so that more exposures of the same field could be made the next night.
Figure 2.5: An example supernova observation monitoring image which shows a distri- bution of fake supernovae placed on a field exposure and how many of those fakes were
then detected bySExtractor, the part ofDIFFIMthat finds detections. Also presented is
the percentage of fakes detected as a function of magnitude. If the 50% mark was mea- sured to be at too low a magnitude (labeled as “Bad") then the exposure was considered a failure.
Figure 2.6: An example supernova observation monitoring image which shows a distribu- tion of bright (magnitude=20) fake supernovae placed on a field exposure and how many