Comparing to Homogenized Coadds


Joe Mohr and Shantanu Desai have pointed out that there are some troubling differences in the star/galaxy separation when comparing coadds made with psf homogenized images (in the Munich reduction) to those without homogenization (in the SVA1 DESDM reduction).

In particular, the effect is strongest in the z-band star selection in the field of SPT-CLJ0539-6013 , which overlaps with DES tiles DES0538-5957 and DES0544-5957.

Mapping The Problem

The problem as posed on Shantanu's page is the following: 'What is the distribution of stars selected from the Munich homogenized coadds with SPREAD_MODEL<0.002 that have SPREAD_MODEL > 0.002 in SVA1?'. I have replicated this plot for the z-band star selection in this field (the strongest example of the problem), with an additional (?) cut of MAG_AUTO_I < 21.0:

There is a strong grid structure readily apparent. This is a problem, but what exactly is causing it, and how bad is it? Joe hypothesizes that it has something to do with discontinuities in the psf model. Sounds reasonable, but let's dig.

It is also instructive to look at the images:

This shows the comparison of a zoom of the Munich coadd to the SVA1 coadd, with the "bad stars" circled in red. The Munich coadd has significant striping visible, I believe because of the bug in fpack that turns 0 weights into non-zero weights (this was avoided in SVA1 by using the WEIGHT_THRESH option in swarp). It's also apparent that the bad stars are correlated with one particular stripe. Looking at the weight map (not shown), this is a region that is deeper than the other regions.

Looking at Spread_model in the non-homogenized coadds

One of the first things to note is that the spread_model locus is significantly wider in the non-homogenized coadds vs. the homogenized coadds (similarly, there is less scatter in the stellar locus; the effect on galaxy photometry is less certain). So while a cut of 0.002 for star selection is appropriate for the homogenized coadds, a cut closer to 0.003 is appropriate for the non-homogenized coadds.

In the above plot the white points are for the Munich homogenized coadds, with the tight spread_model locus. The red and green points are for SVA1, and the green points in particular are those with spread_model_z>0.002 "bad stars" (474 in total). Finally, the blue points are the 28 stars that are misclassified in SVA1 using the "modest" star/galaxy separator which uses (i-band) cuts appropriate for the non-homogenized images.

There are several take-aways here:

  1. The same spread_model cut is not appropriate for both homogenized and non-homogenized images, and the homogenized locus is (unsurprisingly) tighter
  2. Using appropriate cuts and star/galaxy selection there is no gross problem with SVA1 Gold star selection. Thus, "bad stars" is a bit of a misnomer.
  3. Within the broader non-homogenized spread_model locus there is definite spatial structure. Eg, the grid pattern shows up when you choose stars that are a little off 0 (>0.002) but still within the main locus.

Why the structure at all?

Looking at the weight map for the regions where the "bad stars" (see above) show up these strips are deeper regions. Why should this be? I looked back at the finalcut images and looked at the flux_radius in the individual epochs of all of the stars.

In particular, it was interesting to look at the maximum flux_radius_z of all the measurements in the individual images.

In the left panel is a histogram of all the stars (white) and the bad stars (red) as a function of maximum flux_radius_z. The right panel is a ratio of the two histograms. The stars are more likely to be (slightly) misclassified when one (or perhaps more) of their observations have bad seeing. And the (smoothly interpolated) psf model in the coadd is unable to handle the discontinuity where these bad regions are a problem, and thus these stars appear slightly fatter than the psf model and get tagged with a slightly larger spread_model value.

Conclusion (for now...)

It appears that psf discontinuities, exacerbated by poor seeing observations, are causing the spatial structure in spread_model. However, using appropriate star selection for non-homogenized images (such as the modest classifier, or others being worked on the s/g challenge) may not have such a problem.

On the other hand, this does expose some slight systematics in the psf modeling of the non-homogenized coadds, which will affect the photometry. The question is at what level is the photometry affected, which I will investigate next here: Comparing Photometry to Homogenized Coadds