This is a cut-based classifier with binary output star / galaxy. It is used for a bright star catalog, SNVETO, to identify bright stars.
The SNVETO catalog has its purpose to provide the location and an exclusion radius around bright stars that we do not search for supernovae. The idea is that bright stars are often associated with subtraction artifacts that can show up in the difference imaging. The total area excluded in the SN fields is typically under 2% (compare with 10% gaps between CCDs). The SNVETO catalog will also contain a list of variable objects (stars, AGNs, etc), but this is currently work in progress.
As for star/galaxy separation the input quantities include variables available from standard query's of the DESDM database and relies primarily on spread_model and spreaderr_model as well as class_star.
This analysis studied the combination of spread_model + 3 * spreaderr_model beginning with DC6B simulation data. The main point is not this particular variable, but that there strong star / galaxy separation power in spreaderr_model in addition to spread_model.
I call it a "cut based decision tree" as a little play on multivariate analyses that should perform much better. Essentially, it is a decision tree because I make multiple passes selecting stars:
Prequalification: Require at least two bands (of g,r,i,z) that pass quality requirements: mag_psf>1 and mag_psf<99, magerr_psf<1, and flags<=3
Pass 1: Require spread_model + 3. * spreaderr_model < 0.003 AND class_star > 0.8 in at least two (g,r,i,z) bands
Pass 2: Require spread_model + 3. * spreaderr_model < 0.003 OR class_star > 0.8 in at least two (g,r,i,z) bands but then pass additional selection
- a_image > 1.4
- Ellipticity: 1 - b_image / a_image < 0.2
- In two of the (g,r,i,z) bands, require delta mag < 0.5 for mag_psf<21 and delta_mag < 0.5 + 0.5 x ( mag_psf - 21 ) where delta_mag is mag_psf - mag_aper_7 [better would be to use mag_model, which happens not to be in the SE catalogs, I think].