Michael Wood-Vasey and Joel Welling have been working on an algorithm to sort the genuine transients associated with asteroids from image differences found by ppSub. The warped images corresponding to a given skycell are prepared as described in MWV Megacam MOPS Subtractions. Subtracted images are then produced via

ppSub ${warpDir}/${new}.fits ${warpDir}/${ref}.fits ${warpDir}/${outName} \

-inweight ${warpDir}/${new}.wt.fits -inmask ${warpDir}/${new}.mk.fits \ -refweight ${warpDir}/${ref}.wt.fits -refmask ${warpDir}/${ref}.mk.fits \ -sources ${warpDir}/${new}.cmf -type ISIS -photometry \ -threads 4 >& $logName

where ${warpDir} is obviously the directory holding the output files of pswarp and ${new} and ${ref} are the new and reference observations. This produces a catalog file for the subtracted image; catalogs for the new and reference images are available from the pswarp stage. The catalogs provide the input for the rest of the algorithm.

The algorithm uses a specific scale factor S to distinguish false displacements due to astrometry or warping errors from actual transients. We're currently using S=7 pixels, but that value will have to be tuned once the algorithm becomes stable.

For each source found in the subtracted image, the closest sources in the new and reference images are found using a tree search algorithm (specifically a 2D metric tree). Call a source in the subtracted image $P_{sub}$, and denote the source in the new image which is closest to $P_{sub}$ by

$P_{new} = Closest(new, P_{sub})$

and similarly for the reference image. Denote the separation between two points $P_1$ and $P_2$ by:

$ | P_1 - P_2 | $

For each source in the subtracted image, the following tests are performed.

• if $| P_{new} - P_{sub} | < S$ and $| P_{ref}-P_{sub} | < S$ the source is dropped as a subtraction artifact.

• if $| P_{new} - P_{sub} | > S$ the source is dropped as a transient in the reference image (which is ignored) or a subtraction artifact.

• if $| P_{ref} - P_{sub} | < 1.05 S $ and $P_{new}=Closest(new,P_{ref})$ and $P_{ref}=Closest(ref,P_{new})$ the source is dropped as a remainder from a mis-aligned subtraction.

The following tests are then performed on the FLAGS field of $P_{new}$ and $P_{ref}:

• if the bit 0x4000 (PM_SOURCE_MODE_EXT_LIMIT) is set in either $P_{new}$ or $P_{ref}$, that source is saturated so the the source is dropped as a probable saturation artifact.

• if $P_{new}$ bit 0x0800 (PM_SOURCE_MODE_DEFECT) the source is dropped as a defect.

Unfortunately the other flags have not proved reliable enough to use to eliminate candidates. We had expected that for good candidates PM_SOURCE_MODE_PSFMODEL would be set and PM_SOURCE_MODE_FAIL, PM_SOURCE_MODE_POOR, PM_SOURCE_MODE_SATURATED, PM_SOURCE_MODE_CR_LIMIT and PM_SOURCE_MODE_EXT_LIMIT would be unset, but visually very reasonable asteroid candidates violate each of these settings.

This selection process produces a list of candidate asteroids which are then ranked based on the product of the PSF_INST_MAG_SIG and PSF_CHISQ for $P_{new}$. PSF_INST_MAG_SIG is small if the source has a high signal-to-noise ration, and PSF_CHISQ is small if the profile of the source closely matches the PSF. Thus this product, which we call the Quality Factor, is small for clean, well-shaped sources.

The major problem with this algorithm is that it does not handle masked regions properly. False sources often arise at the boundaries of masks and pass the tests above very well. We plan to address the masking issue as soon as possible.

We have scripts to perform these processing steps and to extract stamp images of the candidates. The nearest neighbor algorithm used so far is external to IPP, and we hope to replace it with an existing algorithm in the near future.

Another goal is to produce a meaningful statistical P score in liu of the Quality Factor, so that a false discovery rate cutoff can be used to select a good subset of candidates to treat as probable detections.

--Welling 22:16, 24 November 2008 (HST)