Drizzlepac/HAP: Alignment problems using new gsss service

[stscijgbot-hstdp]

Issue HLA-1130 was created on JIRA by Rick White:

I've been sifting through the 2023-09-13 (gsss) and 2023-09-28 (gssstest) test results on dlhlalab1 looking for astrometry issues. The new gssstest service computes more accurate values for the Gaia RA and Dec errors for observations that are not near the Gaia reference epoch of 2016.

The great majority of the images have very similar astrometry. But when I compare the astrometry for the flt/flc images, I do find some disturbing cases where the gssstest results are significantly worse. Some of these make the HAP images unusable by introducing severe alignment errors between exposures or between filters.

I think these may be closely related to some of the problems being discussed in HLA-1112, which are thought possibly to be the result of the new weighting algorithm (which uses the errors) giving too much weight to the very brightest sources. That is why I have initially assigned this to Steve Goldman. I am putting in a separate ticket in case this is a different problem, but am hoping that when that problem is resolved, it also fixes these cases. So this ticket may be useful mainly as a guide to testing the fixed weights.

Comments below will give detailed lists of the visits found to have problems.

[stscijgbot-hstdp]

Comment by Rick White on JIRA:

I have extracted WCS information for 6,177 ipppssoot_flc/flc images (for 1,102 visits) that are in both the 2023-09-13 sample (using the old gsss service) and the 2023-09-28 sample (using the new testgsss service). I compared the astrometry for the two samples. Out of those images and visits, there are 26 visits that have exposure shifts by 0.25 arcsec or more. I examined all those exposures and found that most of them look OK in both samples. But there are 6 cases where the gssstest results are definitely worse than the gsss results (compared with only 1 where gssstest is better).

The table below lists the visits with problem, with a brief comment on the issue. The sep column is the shift between the gsss flt astrometry and the gssstest flt astrometry. ||ipppss||inst||sep(")||comment|| |j91wef|acs_wfc|1.2876|gssstest image bad (doubled images)| |j8fs70|acs_wfc|0.9488|gssstest image worse (misaligned exposures)| |id5d10|wfc3_uvis|0.3748|bad gssstest alignment, doubled stars, misaligned filters| |iboe1j|wfc3_uvis|0.3523|gssstest shift worse; weird image with 2 binned filters; dubious shift between filters| |icmx05|wfc3_uvis|0.2902|gssstest shift worse|

[stscijgbot-hstdp]

Comment by Rick White on JIRA:

The attached pair of images show the problem for the first visit, j91wef. There are 3 filters in this ACS/WFC visits: f850lp, f775w, and f606w. The images show the color display in the vicinity of a Gaia star (the Gaia position is marked by a circle, which is a bit hard to see). For the 2023-09-13 gsss version, the 3 filters are aligned. But for the 2023-09-28 gssstest version, the f775w image (green) is shifted compared with the other two filters. The misalignment is apparent both for the Gaia star and for the galaxy off on the right edge.

Somehow the new gssstest service has led to very different results for this image. Since only the errors have changed significantly in the gssstest service, and those errors are used to compute the weights, it seems likely that the weights are the culprit. !j91wef_2023-09-13_gsss.png|width=49%! !j91wef_2023-09-28_gssstest.png|width=49%!

[stscijgbot-hstdp]

Comment by Rick White on JIRA:

I have done a more thorough investigation of this issue in the test datasets. It looks like there are (rare) random differences in the alignment when the Gaia reference catalog changes. By "random", I mean that when there are differences, half the time the old version is better and half the time the new version is better. Since the changes in the catalog itself are very small, this indicates an undesirable sensitivity to modifications in the positions or errors of the Gaia sources.

I am still optimistic that improving the calculation of the weights in the fits will lead to better and more stable results. For aid in testing, I created an innerspace page with a table of problematic images.

[stscijgbot-hstdp]

Comment by Rick White on JIRA:

Here is the table of problematic images (with poor alignments or shifts) from the innerspace page (which has a bit more discussion). ||ipppss||inst||visit||filters||nfilters||nexp||sep(")||seprms(")||WCSType (drizzlepac main)||Refcats(GAIAeDR3, GSC242, 2MASS)||Cross-match magnitudes not correlated failure||comment||display gsss 2023-09-13||display gssstest 2023-09-28|| |iclg05|wfc3_uvis|hst_13698_05|f438w f606w f665n|3|9|0.0567|0.0476|undistorted not aligned|5, 155, 5|x|gssstest bad; exposures misaligned|gsss|gssstest| |icmx05|wfc3_uvis|hst_13804_05|f275w f336w|2|4|0.1451|0.1451|undistorted a priori solution based on GSC240|28, 244, 3| |gross filter misalignment for both|gsss|gssstest| |id1k4w|wfc3_uvis|hst_14178_4w|f814w|1|4|0.3526|0.0000|undistorted a posteriori solution aligned by visit to GSC242|16, 212, -| |gsss shift better|gsss|gssstest| |id1x09|wfc3_uvis|hst_14393_09|f280n f606w|2|8|0.0383|0.0614|undistorted a posteriori solution aligned by visit to GAIAEDR3|7194, -, -| |gssstest PSF worse|gsss|gssstest| |id5d10|wfc3_uvis|hst_14648_10|f275w f410m f621m|3|6|0.1761|0.1654|undistorted not aligned|1760, 2607, 576| |bad gssstest alignment, misaligned filters (note 2023-09-14_gssstest was ok]|gsss|gssstest| |idxm01|wfc3_ir|hst_15644_01|f110w|1|8|0.3574|0.1013|undistorted a posteriori solution aligned by visit to GSC242|16, 688, -| |gssstest scale change! gsss better; very worrisome|gsss|gssstest| |j8fs70|acs_wfc|hst_9401_70|f475w f850lp|2|5|0.1996|0.3747|undistorted a priori solution based on GSC240|14, 223, 8|x|gsss good, gssstest misaligned|gsss|gssstest| |j8i0aj|acs_wfc|hst_9488_aj|f475w f775w f850lp|3|9|0.0892|0.2301|undistorted a posteriori solution aligned image-by-image to GAIAEDR3|14, 188, 4|x|both bad with misalignments|gsss|gssstest| |j8za03|acs_wfc|hst_10127_03|f814w|1|9|0.0888|0.0450|undistorted a posteriori solution aligned by visit to GAIAEDR3|17, -, -| |gsss PSF def. better|gsss|gssstest| |j91wef|acs_wfc|hst_10339_ef|f606w f775w f850lp|3|6|1.0865|0.4673|undistorted a posteriori solution aligned by visit to GSC242|12, 134, -|x|gssstest misaligned filters|gsss|gssstest| |j92qj4|acs_wfc|hst_10340_j4|f775w f850lp|2|5|0.0717|0.0385|undistorted a posteriori solution aligned by visit to GSC242|13, 199, -|x|gssstest misaligned filters|gsss|gssstest| |jboa88|acs_wfc|hst_12440_88|f606w f814w|2|5|0.1244|0.1478|undistorted a posteriori solution aligned by visit to GAIAEDR3|20, -, - | |gsss good, gssstest misaligned|gsss|gssstest|

Column descriptions:

• {}ipppss{}, {}inst{}, and visit identify the observation
• filters lists the filters used in the visit
• nfilters is the number of filters
• nexp is the number of exposures
• sep is the offset in arcsec between the 2023-09-13 test (which used the older gsss service with slightly different positions and different ra and dec errors) and the 2023-09-28 test (which used the new gssstest service)
• seprms is the rms difference in the offsets for individual exposures in the visit.  A large seprms value may indicate that a single exposure had a large (different) shift in the two results sets.
• comment briefly describes the issue. Common issues include misaligned exposures (where exposures in the same filter are not correctly aligned, leading to bad pixel rejection or poor quality PSFs), filter misaligned (where images from different filters are not lined up, leading to double images in the detection/total image and poor quality catalogs), overall shift errors (the image does not match Gaia stars), etc.
• The two display columns are links that load the detection image (with the Gaia catalog overlaid) from the two different results samples to visually confirm the issue.

[stscijgbot-hstdp]

Comment by Steve Goldman on JIRA:

Thanks Rick White do you think that the best way forward is the tinker with a cap on the maximum brightness of a potential guide star, or to compare the new and old errors of the guide star sets to look for patterns? This is my current priority so please let me know if you have found anything else, or if you are working on anything related to this ticket. 

[stscijgbot-hstdp]

Comment by Steve Goldman on JIRA:

After talking with Warren, he suggested tinkering with the HST point sources that are allowed to be used for matching. Specifically removing sources that are within a couple pixels of saturated (flagged) sources in order to remove sources that are labeled good, but are found in the wings of saturated PSFs. Over the next couple days, I will be looking at what criteria is used for vetting HST sources for alignment, but do you have any thoughts on this strategy for testing these visits?

[stscijgbot-hstdp]

Comment by Steve Goldman on JIRA:

Rick White, is the only difference between gsss and gssstest an update to the guide star catalog? And if that's the case, has a new catalog been added?

[stscijgbot-hstdp]

Comment by Steve Goldman on JIRA:

For J91wef, we have one filter with 4 dithers and then two filters without any dithers. It seems that if the code fails in finding relative alignment, then it abandons absolute alignment; and that seems to be the case here 

These are the resulting statuses for trying alignment on 'GAIAeDR3' with method 'relative' using fit geometry 'rscale'.

hst_10339_ef_acs_wfc_f606w_j91weft4_flc.fits , REFERENCE hst_10339_ef_acs_wfc_f606w_j91weft4_flc.fits , REFERENCE hst_10339_ef_acs_wfc_f850lp_j91wefti_flc.fits , FAILED: not enough matches hst_10339_ef_acs_wfc_f850lp_j91wefti_flc.fits , FAILED: not enough matches hst_10339_ef_acs_wfc_f850lp_j91weftc_flc.fits , SUCCESS hst_10339_ef_acs_wfc_f850lp_j91weftc_flc.fits , SUCCESS hst_10339_ef_acs_wfc_f850lp_j91weft9_flc.fits , SUCCESS hst_10339_ef_acs_wfc_f850lp_j91weft9_flc.fits , SUCCESS hst_10339_ef_acs_wfc_f850lp_j91weftf_flc.fits , SUCCESS hst_10339_ef_acs_wfc_f850lp_j91weftf_flc.fits , SUCCESS hst_10339_ef_acs_wfc_f775w_j91weft6_flc.fits , FAILED: not enough matches hst_10339_ef_acs_wfc_f775w_j91weft6_flc.fits , FAILED: not enough matches

 

Even though it succeeds in the third f850lp file, the first two failures result in a defacto a priori solution.

 

 

Is the previous WCS solution type easily accessible for this program in the previous gsss version?

[stscijgbot-hstdp]

Comment by Steve Goldman on JIRA:

Rick White Coming back to this. It looks like j91wef is being processed as expected. I'm going to be testing other individual datasets for the rest of this sprint to look for bugs, but please let me know if you have more information about the version of drizzlepac used for you two tests. That would help me identify how changes in the code may have affected the fitting. 

[stscijgbot-hstdp]

Comment by Steve Goldman on JIRA:

It looks like idxm01 is shifted and changes scale for both pipeline and SVM processing. It looks like an edge-case where multiple very bright sources (probably weighted far higher in the alignment now) are near the edges on multiple sides of the images. For some of the individual images, the bright sources are off the frame, but their diffraction spikes are not. It may be that these edge diffraction spikes are being picked up as sources and are weighted more highly than before, because they are bright. 

[stscijgbot-hstdp]

Comment by Steve Goldman on JIRA:

For j8za03, the difference between the images appears to be very minimal. I'm not sure there is a necessarily a problem with the alignment of the program, but there is a small shift that may be related to changes in the weighting of the alignment. 

[stscijgbot-hstdp]

Comment by Steve Goldman on JIRA:

id5d10 is centered on a source that is saturated in one of the three filters. The psf of the total image would be expected to change since the bright sources are weighted more highly in the alignment now. 

[stscijgbot-hstdp]

Comment by Steve Goldman on JIRA:

I've added a couple columns to the pinned table. Many of the programs have far fewer sources in GAIAeDR3 than in GSC242. Additionally several of the programs fail alignment due to a "cross-match magnitudes not correlated" failure.