| | 7 | |
| | 8 | This was a productive week with Micael Wood-Vasey visiting. He is investigating the differences between the IPP and photpipe photometry of sources in the photfest stack-stack diffs. He started by looking into the differences between psphot measurements and psphotForced measurements. He uncovered some differences triggered by differences in the background model. Based on this, as well as ongoing concerns that the background model is too sensitive to bright stars, I looked into tweaks that may help. The current processing uses 256 pixel superpixels without oversampling, using a subset of 10k measurements in the superpixel. I discovered that the fitted sky value in the superpixels had high outliers more frequently that expected, and tracked this down to a low-level bug in the way the sky histogram is fitted -- it uses a fit to the lower fraction of the (presumed) gaussian sky distribution to estimate the peak. This can cause problems if the histogram is too noisy, resulting in a poor constraint on the high end of that fit. I fixed this by requiring the full range fit to be greater than the lower half fit, or else the fit from the full range is used. In discussion with Michael, we concluded that the number of samples is also too small, introducing more scatter than we really want. Also, the fairly small size of the superpixels (64 arcsec) makes the measurement too sensitive to bright stars. We have decided to try 400pixel superpixels, with 2x oversampling (ie, the superpixels are spaced by 400 pixels, but sample a region 800 pixels wide), with at least 50k pixels used per measurement. |
| | 9 | |
| | 10 | I also explored the issue mentioned last week about regions where the sky is under subtracted (the model is unable to follow the shape of the sky sufficiently well). I tried to filter sources from this region on the basis of the flux in a small (psf-sized) aperture to that in a larger aperture. This seemed to work well for the image of interest, but in wider application it tends to exclude too many obvious galaxies. I tried an alternative of using the smoothed significance image in the peak culling process, and this seems to be a good balance: |
| | 11 | |
| | 12 | * [http://svn.pan-starrs.ifa.hawaii.edu/trac/ipp/wiki/Background_Oversubtraction Background subtraction problems] |
| | 13 | * [http://svn.pan-starrs.ifa.hawaii.edu/trac/ipp/wiki/PeakCullingInPSPHOT Peak Culling with the Smoothed Image] |
| | 14 | |
| | 15 | I also looked into a problem we've been having with some stacks in which the target PSF is surprisingly large (eg, 30 pix FWHM when inputs were limited to 8). This is caused by poor PSF models, especially in cases where the input warp has very few sources. I have added code to ppStack to examine the input set of warp PSFs and reject inputs based on 2 criteria: 1) if the PSF is larger than some user-defined max or 2) if the PSF is larger than some number of sigma of the input set. I checked on the impact of this by examining the statistics of warps used for the reference stacks generated to date, and this cut would help a number of those stacks. This issue also explains a concern raised by Nigel and Peter about the large range of PSFs in the MD04 refstacks. |
