Some Notes on Astrometry¶
As detailed in FPACKing exonerated as Bugs in swarp and ImageProc Corrupts Modeling, the source of the DETMODEL (and small aperture!) color problems was a 1+ pixel image registration error between the detection and measurement images. This led me to think: how good is the astrometric registration between bands, and does this have an effect on the photometry? And can we do better?
The current setup¶
Currently, each DECam mosaic image is registered to the UCAC star catalog independently during finalcut processing. However, while this absolute astrometric solution is good to a fraction of an arcsecond, it is not as good as can be achieved via relative photometry. Furthermore, there is no procedure in place to ensure good registration between coadd images, which is what's necessary to get unbiased low-scatter color measurements. Astrometric registration issues are discussed in docdb/6991 with regards to shape measurements for WL images, where 15 mas rms is required for relative astrometry between images in a stack. As I understand it, the intention is to generate a value-added astrometric solution outside of DESDM for the shape measurements. But what about coadd photometry?
A test field: DES0410-4957¶
I have started looking at a couple test fields including DES0410-4957 to examine image registration issues. All coadd processing was my own, since these tests need to remove the +/-1 pixel registration bug in the Y1C2 coadds.
- The first run uses the official finalcut "absolute" astrometry, where each coadd image was independently registered (with SCAMP) to the UCAC catalog.
- The second run uses a quick-and-dirty "relative" calibration. I took each CCD and registered it to a common frame derived from the i-band Y1C2 coadd catalog. Yes, this is circular and not practical.
The following figure shows the comparison of the bright star positions (using ALPHAWIN_J2000/DELTAWIN_J2000) for each coadd relative to the i band detection image. Red is the absolute astrometry run and black is the relative astrometry run:
The curves show the a fit of a 2d Gaussian, with the sigma shown in the figure legends. The g vs i band performance is especially poor, with an rms of 0.31 pixels. After implementing my simple "relative" astrometry the rms values are improved significantly.
The following figure shows the same comparison for bright galaxies:
In this case the improvement is not as significant, however the galaxy positions are much noisier so this isn't necessarily very informative.
Note that these tests do not show the spatial component of the mis-registration; we expect the large offset stars to be spatially correlated.
Implications for Photometry¶
As a first estimate of the photometric biases induced from the misregistration, I have taken a sample of bright (i<20) galaxies and plotted the detmodel color - aperture color (using a relatively large aperture; the detmodel color will be more susceptible to image misregistration) vs centroid offset:
The red lines are fit using the IDL LADFIT routine which is robust to outliers. It is apparent that a large amount of image misregistration can induce a few percent bias in the detmodel colors, which is definitely significant for red sequence cluster finding, especially since these biases should be spatially correlated. As it is, the problem may be worse or better than this, because the centroid offset for galaxies is very noisy (as seen above).
Conclusions and To-Do¶
The current absolute astrometric solutions, where each image does not know anything about the other images, can induce color biases of a few percent using DETMODEL or small aperture colors for galaxies. It is possible to do better, although how to integrate this into the pipeline is a question.
I would also like to explore the individual CCD astrometric offsets, which I have not done yet (only looked at the resulting coadds). If the registration is spatially correlated well enough and the density of stars is high enough then I can replace the delta pix for the galaxies with the delta pix of the nearest star. This may yield a better estimate of the color bias as a function of pixel offset.