Project

General

Profile

Additional Stellar and Bad Region Masking

The masked region defined for SVA1 Gold 1.0.2 is insufficiently conservative. There are issues...and some of them can be fixed easily.

-Eli.

The Issues

Erin's ngmix multiepoch fitting finds a bunch of low surface-brightness junk (objects that are really really large). Full details are on this page , and linked is a shortcut to a tile showing where these objects are: in the halos of bright stars.

Erics H & S and Peter M have noted via Balrog tests that there is a population of galaxies that have very large offsets between the windowed positions in g and i (and other bands!). This is expected at the faint end (see below) but bright objects? These are junk detections. (And largely the same bad objects as in ngmix!)

Finally, it is clear that many of these bright objects are highly correlated with bright star positions. These problems are all interrelated, so I will tackle them together.

Selecting Bad Objects with Windowed Positions

In SExtractor in dual-image mode, the object is detected on the detection image and then measured on the measurement image. While the object position used for photometry is fixed in the detection image (to get forced photometry at the faint end), there are still independent estimates of the positions in each individual band that are recorded as ALPHAWIN_J2000_[GRIZY]/DELTAWIN_J2000_[GRIZY] in the database. To first order, you would expect a consistent position measured in (Eg) the g band and the i band, though these are independent (since the g band is not used in the r+i+z detection image). However, there are some objects that are large outliers (even degrees sometimes!) However, some of these are expected: g band dropouts don't have any flux and so the windowed position is essentially unconstrained. But that's not the full story. Here is a plot of the fraction of objects that have large (>1") offsets between g and i band positions as a function of g band signal to noise (as estimated by flux_auto_g/fluxerr_auto_g):

There is a population such that about 2% of all objects at any signal to noise are crazy offsets! Visual inspection shows that these are bad objects that are (a) in the halos of bright stars; (b) cosmic rays and satellites; (c) ghosts; (d) other defects. Plus, we can plot where the objects that have s2n_g>5 and dis>1" are, for a sample region of SPTE:

As can be seen below, these are largely in the vicinity of very bright stars. But that's not all (see below).

Therefore, I recommend we remove (at least) all objects with distance > 1" and s2n_g > 5.0. However, a geometry-based mask would be superior since it is much easier to simulate!

A 2MASS Based Stellar Mask

The current stellar mask is derived on an image-by-image basis based on the detection of a star in the image and with a mask radius based on the approximate size of the saturated star. However, diffraction spikes and stellar halos extend much further. Plus, we do not need to depend on detecting stars in the images ... we have good stellar catalogs, especially at the bright end! The Tycho catalog is very nice, but this is just the bright end (plus there is the Yale Bright Star Catalog for visually bright stars. This is another issue, that is largely taken care of by the blacklist.) The 2MASS stellar catalog is very useful because it has all the Tycho stars plus fainter ones that extend even into well-measured stars in DES (and they can be used for astrometry and SLR corrections.) The drawback is that the J band magnitudes are not quite our ideal band, so any radial scaling we derive will be a bit off for the bluest/reddest stars. (This will have to be addressed another time.)

I have compiled a nice version of the 2MASS star catalog here . This covers the full southern hemisphere with |b| > 20, and has J, H, and K magnitudes (and errors).

By eye, using the ngmix bad objects, I derived a scaling relation that will crop bad objects.

maskrad = -10 * J_M + 150  (arcseconds)
maskrad < 120  (ceiling)
remove all stars with maskrad < 30 

That is, the maximum maskrad for 2mass stars is set to 120" (2'), and stars that have mask radii less than 30" are not in the mask. (At this point, the star is not dominant for a pixelized mask with nside=4096, plus the original mangle mask contains much of this information.)

A simple fits table with ra/dec/maskrad for the stars in SVA1 is here

I then pixelize this mask with nside=4096 by cutting out pixels who's center is contained within the maskradius + 10" (via healpy.query_disc(inclusive=False)). (An alternative would be to cut all pixels that intersect with the maskradius, but this loses a lot more area.)

On the sample region from above, the circular stellar mask is in red, and the pixelized version in cyan. The pixelized version reduces the SVA1 Gold area from 254.4 deg^2 to 237.5 deg^2 (a reduction of 6.4% of the area).

The good region mask + stars is here:

A Windowed Bad Region Mask (or two)

But that's not the full story. There are more bad objects, and they still appear to be clustered (though the star mask gets rid of many of the clusters).

I have computed the density of bad objects in each healpix pixel (nside=2048), and computed the fraction of bad objects that are removed for any given area cut:

So we could remove 100% of the bad objects while removing 60% of the area. This doesn't sound so smart...

But a cut of ~4% of the area would get rid of 25% of the bad objects, which sounds like a reasonable trade-off to me. Because of overlap, this cuts an additional 3% of the area from the star-masked version. Alternatively, a more aggressive cut of 10% of the area (cutting an additional 8% of the area relative to the star mask) would get rid of 43% of the bad objects. The way these cuts look are as follows, with the magenta+blue showing the 10% cut and the blue showing the 5% cut.

My instinct would be to use the 4% + star cut, though I have both masks here for testing.

The masks are here:

Summary

I have produced two combined star + badregion masks (this is to replace the "goodreg" mask from Gold, and includes all of the masking already determined previously from concentrations of objects with crazy colors).

The files are here:

These files are also linked to from the Gold page. In addition, I have added another Gold file (to be added to database) with objects flagged according to whether they are bad objects/in bad regions.