Fixing Astrometry in SVA1 Finalcut

As laid out in Relative Astrometry in SVA1 Finalcut there are issues with the relative astrometry in SVA1 finalcut. These offsets appear to be the source of systematic biases in the WL shapes detected by Joe Zuntz and Erin Sheldon (plots to come).

Can the astrometry be fixed? Yes indeed!

The Problem

The initial SVA1 finalcut astrometry solved each exposure mosaic against UCAC-4, allowing a simple shift and rotation but otherwise fixing the distortion from the TPV solution in the ".ahead" file. However, this is an issue because (a) the TPV solution is different for different bands; (b) UCAC-4 has a limited number of stars reducing statistics. Using a co-solution of relative astrometry with DES stars should do better.

Note that the astrometric problems manifest themselves in two ways:

  1. There are systematic offsets in stars on a given CCD relative to the mean position of the stars
  2. There is increased RMS in these offsets.

I believe that the first problem is more serious...a shift of 200 mas is over half a pixel! But they are obviously both problems.


I (Eli) have compared 2 simple scamp-based solutions on the two tiles. These can be implemented in the short term, but are not as sophisticated as Gary's upcoming "all-singing all-dancing global astrometric solution".

CCD-by-CCD Solutions

One simple option is to solve the astrometry for each CCD, one at a time, allowing an offset and rotation on top of the default TPV distortions. Of course the problem is that there aren't enough UCAC-4 stars to do this. But there are at least 100 well-measured flag=0 DES stars on each CCD!

The easy solution is then to create a replacement catalog to UCAC-4 using DES star positions in the SVA1 coadd. (Alternatively, with a bit more work, the average position for each star from the initial solution can be computed by cross-matching all the finalcut catalogs. This is a bit of a pain, but of course is also being done by all the global calibrators). (It's also what scamp can do, but it's also easy to do this outside of scamp.)

So with this new "DES reference catalog" (warts and all...the initial biases may be baked in) each CCD can be solved with scamp. And voila! The advantage of this method is that it is really ridiculously fast, and it doesn't take a lot of memory.

Coadd-tile exposure solutions (as planned for Y1A1)

The "coadd_astrorefine" code is being tidied by Robert G. I've adapted the core from this:

  1. Find all the exposures that overlap a given coadd tile (this is much more area than the coadd!)
  2. Create a combined catalog for each of these exposures in all griz bands.
  3. Solve absolute and relative astrometry using scamp, setting INSTRUMENT constant and allowing tweaks of individual ccds.

This is a lot slower but does use essentially all the information that went into the ccd-by-ccd solution. It also fixes the relative positions of the ccds for all the exposures (in each band!) which is good. However, it doesn't know about warmups/cooldowns that can cause shifts in the ccd positions which may be a problem.

The current test call to scamp is:


Maybe some of these should be tweaked (see below).


These are the results for RA and Dec offsets for SVA1 (white points); coadd_astrorefine (blue); ccd-by-ccd (red)

The RA offsets are much improved, but the ccd-by-ccd solution brings the offsets down to essentially zero, while the coadd_astrorefine leaves some noise in the offsets.

The RMS values are also much improved for both coadd_astrorefine and ccd-by-ccd. Note that the astrorefine rms is comparable to the noise in the ra offsets. So by forcing the instrumental solution to be consistent, we're not able to move the ccds around by as much as we may wish, but this seems to be within the noise.

Dec offsets and rms look similar.


It may be that we can use either code, though it looks like the ccd-by-ccd is performing a bit better in terms of rms and a lot better in offset. (Though since the offset noise is consistent with the rms noise, it's not clear if this is a problem.)

In both cases there are ccds where the astrometric solution just is not very good (large offset or large noise). I will have to flag these ccds (which will involve matching all the stars over all of sva1 ... fun!) so that they are not used in the shape fitting.