Source-peak to WPHOT-derived Position Biases |
Below we summarise systematic differences in the positions of sources determined using a flux-weighted centroid (or position of peak emission for symmetric profiles) and astrometrically calibrated positions from WPHOT (MDEX). These biases are statistically present in L1b frame products and manifest themselves as source peak--to--astrometric offsets in L3 coadd products. The latter are more discernable by eye due to the relatively higher signal-to-noise and spatial sampling.
The biases are band-dependent and have two consequences for co-adds:
(i) source peak emissions are misaligned across bands;
(ii) one will infer an astrometric solution different from
that represented by the WISE source catalog
if source positions are independently derived using the co-add WCS
and a peak-finding algorithm.
The following are trimmed-median differences in "multiframe-derived MDEX positions" - "flux-weighted source centroids" in the x-y system of L1b frames. See below for supporting plots. The trimmed-medians are shown as red dashed lines in the plots below. These results are consistent for ten co-add tiles over the sky. To get more sources in W3 and W4, results from four galactic center tiles were combined. Significant offsets are present in all bands with W4 having the largest.
Dx sigDx Dy sigDy √[Dx2+Dy2] W? (region) [asec] [asec] [asec] [asec] [asec] ------------------------------------------------------------------ W1 0229p545: -0.4407 0.0089 -0.1915 0.0101 0.4805 W2 0229p545: -0.2387 0.0076 0.0864 0.0064 0.2538 W3 gal center: -0.0097 0.0188 -0.2627 0.0181 0.2628 0229p545: -0.0855 0.0146 -0.2389 0.0143 0.2537 W4 gal center: 1.2107 0.0322 0.6644 0.0324 1.3810 0229p545: 0.9806 0.1740 0.7481 0.1587 1.2333 ------------------------------------------------------------------
From: fmasci.ipac.caltech.edu Subject: L1b peak-to-"truth" mismatch real Date: December 17, 2010 11:12:48 AM PST I reanalyzed three W1 L3a coadds and their corresponding L1b frames at different sky locations. Let me first summarize what the attached plots mean, then I'll jump to some conclusions/thoughts. You may have to view the attached montage in your PDF viewer to magnify it. Each row corresponds to the same L3a tile, where from left->right (as the plot titles say) we have: (1) 2mass-'flux centroid' differences in coadd x,y system; (2) these coadd x,y differences mapped into RA,Dec; (3) 2mass-'flux centroid' differences in *L1b* x,y system; (4) these L1b x,y differences mapped into RA,Dec. Three things to note: (1) the x,y deltas mapped into RA,Dec from either the coadd or L1b's agree (panels 2 and 4 in each row). I must be doing something right. (2) the averaged x,y deltas in the L1b coordinate system agree for all frames across these three tiles (plots in column 3 below). I looked at another two L3a tiles and their L1b 2mass-'peak flux' deltas also fall in the same place. This appears to be universal for at least W1. (3) it's reassuring that these statistics agree with my eyes when I compare a zoomed L1b frame frame against a coadd matched in WCS with 2MASS overlays. Let me know if you'd like to see jpegs of this. I didn't want to clutter this email. Conclusions/thoughts: - The peaks of source profiles in L1b frames are shifted relative to their possible "true" positions as determined by 2MASS (for at least W1). This for me is hard to fathom. I've never encountered anything like it. It explains why all coadders show the same biases on WISE data (awaic, montage, mopex). - I'm sure this has been done thousands of times, but we should check if there are any significant 2mass-mdex residuals for the specific coadds below, both at the frame and coadd level (for bright "dec-unbiased" sources). Howard may already be set up to do this. I want to be 100% sure there is no L1b WCS problem on a per-coadd basis. - Could there be an incestuous thing going on where WPRO fits are inexact for both sdex used for SFPRex and eventually mdex (e.g., inadvertent PSF flip, indexing error, or whatever)? Ignoring faint-source Dec biases, we all know mdex is astrometrically correct. This only means the WPRO position estimation was done consistently for both astrometric calibration and the final source catalog. That's all ok, but it hides the possibility that PSF templates were not aligned correctly. Perhaps a PSF-template jiggle in WPRO would get 2MASS positions to align with source peaks and continue to maintain good astrometry throughout. - Alternatively, every wcs library I'm using, including DS9 could be giving the wrong result. The analysis below (including coadds) used at least two different libraries, all of which understand the SIP convention. I doubt there are problems here.
From: fmasci.ipac.caltech.edu Subject: Re: L1b peak-to-"truth" mismatch real Date: December 17, 2010 3:23:18 PM PST Another idea we tossed around is the possibility that PSFs varied throughout the mission. To remind you, we're using a static set of PSF templates that were made in early March from early survey data. A mismatch to real point-source profiles could manifest itself as a shift between observed profile peak positions and fitted/refined WPRO positions. Well, below are the 2mass-'peak flux' L1b deltas for a coadd made from scans acquired beween Jan 19-23. The same deltas are seen as later in the mission, meaning PSFs did not vary enough in this context to cause the observed peak-to-astrometric position shifts. Frank
From: fmasci.ipac.caltech.edu Subject: Re: PSF coadd experiment Date: December 20, 2010 3:32:51 PM PST To: tim.ipac.caltech.edu Cc: roc.ipac.caltech.edu Frank: There is no co-adder problem. I speculate there is either a PSF fitting problem or a change in the real PSF shape compared to our static PSF templates that causes *L1b* source-intensity centroids (call it source peak emission if you like) to be offset from mdex and astrometric positions. Astrometry is perfect. We just need to tie source-peaks to mdex positions (including sdex for SFPRex) to make everything match and make L3 coadds happy. You won't learn anything from simply take existing L1b's and making a L3 using shifted psfs. The problem exists in L1b's and we're talking about sub-arcsec offsets in w1,w2,w3 as summarized on the webpage I sent earlier. It must be fixed at the source (very top of the chain). Tim: But I don't think anything about the L1b PSF's influences the L3's. Frank: Right, the native observed source profiles in L1b's have no influence on L3's. However, if these have changed compared to PSF-templates for WPRO. E.g., different structure around the peak, we'll get offsets between WPRO solutions and source peaks. Ken will need to examine what's going on. FYI, we see L1b biases in frames acquired in January - supposedly the data used to make our current PSF templates. This argues against a PSF change since the same biases are seen in all L1b's throughout the mission. Tim: The position refinement of the frames is fine, so the fact the old PSFs were used doesn't matter, does it? Frank: Yes, astrometry is fine, we just need to tune WPRO (or something in WPHOT) to get source peaks, mdex solutions and obviously astrometric references to line-up in L1b's. It's just impossible for any detector to give source "peaks" systematically offset from their "true" positions after astrometric calibration. Tim: Exactly. So it shouldn't matter what PSF was used to make the L2's or run wphot. They don't figure into it after sfprex is done. How am I thinking of this wrongly? Frank: It doesn't matter for the L2's but it matters for wphot, specifically wpro. E.g., if wpro systematically gives solutions offset from the source peak (e.g., centering or indexing problem during fitting of the template), this will affect sdex (that's fed into sfprex) and then eventually mdex. Astrometry will be fine, but the peak-to-mdex(or 2mass) bias will persist in all image products.
From: fmasci.ipac.caltech.edu Subject: Re: source peak to mdex biases Date: January 6, 2011 12:41:15 PM PST To: kam.ipac.caltech.edu Cc: jarrett, roc, tim, hlm, jwf Hi Ken, The problem exists in *L1b frames*, where the peak-to-mdex (or 2MASS position) offsets are the same (magnitude and direction) anywhere on the sky. This manifests as offsets of varying degree in L3 coadds due to varying scan angle. Astrometric solutions are correct for "a position" on each point source that does not happen to be the peak. This "wrong" position is different for each band (again, all in L1b space). Regards, Frank --- On Jan 6, 2011, at 12:30 PM, Ken Marsh wrote: Hi Frank, Interesting. I'll look at this in detail, but my first thought is this: Howard's position reconstruction solution implicitly compensates for the set of offsets between the nominal PSF origins and their corresponding true peaks (i.e. actual star location in PSF profile) over the 9x9 grid. That means that it will give the correct astrometric solution for profile-fitted source positions, but may give rise to small residual offsets in the coadds which assume a single isoplanatic PSF. Could this be the case? Ken. --- On Jan 6, 2011, at 12:30 PM, Ken Marsh wrote: On an L1b frame, the x,y position that comes out of wpro represents the position of the psf origin when the psf is aligned with the source on the image. Since the psf origin is, in general, not the source peak, doesn't this explain the constant offset? Ken.