Completeness and false positive rates for the 11/19 reprocessing of MD09 (Eddie Schlafly)

I promised I'd look at the star/galaxy separation at the DRAVG phonecon. As a zeroth step, I looked at the completeness and false positive rates, then split these into the rates for stars and galaxies, and finally looked at the performance of the current star/galaxy separator (at least, the flag PM_SOURCE_MODE_EXT_LIMIT).

I have taken the SDSS Stripe 82 coadd images as truth.

In the histograms of source counts, black lines represent the reference SDSS counts; red lines the PS1 counts, and green line the counts of the matches between the two.

For most plots, I have not filtered on any flags. In one set of plots, marked (trim), I have tried to mask to a clean subset using the following flags: PM_SOURCE_MODE_FAIL PM_SOURCE_MODE_POOR PM_SOURCE_MODE_SATSTAR PM_SOURCE_MODE_BLEND PM_SOURCE_MODE_EXTERNAL PM_SOURCE_MODE_BADPSF PM_SOURCE_MODE_DEFECT PM_SOURCE_MODE_SATURATED PM_SOURCE_MODE_CR_LIMIT PM_SOURCE_MODE_EXT_LIMIT PM_SOURCE_MODE_MOMENTS_FAILURE PM_SOURCE_MODE_SIZE_SKIPPED PM_SOURCE_MODE_BIG_RADIUS

These flags evidently are evidently over-restrictive; they hurt completeness badly.

The overall normalization of the completeness and false positive curves should not be taken seriously. There are two missing frames in the SDSS coadd images, and PS1 detections in these frames register as false positives. Likewise, I've approximated the PS1 footprint as a disc of 1.5 degrees, and detections that occur beyond this register as false positives, and missing areas within this circle hurt the completeness. However, the trends in completeness and false positive rate with magnitude are significant.

I'm defining the completeness rate as N_{match}/N_{ref}; the false positive rate as (N_{detect}-N_{match})/N_{detect}. The star and galaxy completess and false positive rates are N_{match ref star/gal}/N_{ref star gal}; (N_{detected as star/gal} - N_{matched ref star/gal})/N_{detected as star/gal}.

Finally, the last couple plots give the performance of the star/galaxy separator. These give the fraction of reference stars and galaxies that are classified as stars and galaxies.

Below are plots for griz. The attached pdf has more plots; the plots I've displayed here were chosen because they looked relatively deep, whereas the pdf contains randomly 5 randomly chosen exposures for each filter.

The star/galaxy separator isn't doing a very good job right now, but there are at least some obvious things I want to try to improve it. In z-band the PS1 images are about as deep as the SDSS coadd images, and so things aren't very meaningful.

Attachments