| | 7 | |
| | 8 | I ran psphotStack on the 8-way MD04 stacks from last week. It took a couple of iterations to get reasonable answers. In the first round, the memory footprint exploded and crashed several of the machines used for the analysis. I discovered that the 'chisq' detection image (combination of all input filter images) was not correctly defined. I temporarily disabled this image as part of the detection process. This also uncovered an interesting memory waste : when psphotStack generates source entries for the undetected sources in the other filter images, the footprint analysis always uses the largest footprint as the container for all new footprints. This is excessive and can lead to large memory usage in some cases. I have not yet made a final fix on the latter problem, but by avoiding the chisq image for now we can work past it to continue testing. |
| | 9 | |
| | 10 | I addressed a number of other issues in the outputs from psphot, psphotStack, and ppStack. Specifically: I fixed the the calculated chisq values for all objects -- there were a few inconsistencies, but the worst was that the unweighted non-linear fits were making an ill-advised correction to the chisq to account for not using the weights in the fitting. I changed this to simply recalculate the chisq with the correct weights after the fit was completed. The chisq distribution for all sources now seems quite sensible (centered around 1.0 for most objects, though growing for the brighter stars where the systematic differences between the model and the data become important). |
| | 11 | |
| | 12 | I also fixed the output headers of the stacks to correctly set CELL.SATURATION, taking into account the image flux re-scaling. I also wrote up an analysis of the steps in ppStack which affect the flux scaling and zero points: [wiki:Stack]. I also added more information to the output stack headers: the names of all input images, their exposure times, airmasses, reported zero points and any flux scaling applied to them before combination. |
| | 13 | |
| | 14 | I looked into cases where the ppStack target PSF is oddly high. This appears to happen because the PSF model from one or another of the input warps is poor. Looking at this more closely, I believe that the pswarp psf model is being generated from sources whose positions are held fixed; any small error in the astrometry then couple badly into the resulting PSF model. |
| | 15 | |
| | 16 | Finally, I examined Celestial North Pole images which were failing in psphot in an interesting way: they were growing to 20GB and finding 10s of thousands of sources. The cause for this is clear: these images have very poor seeing and interesting variations in the background from clouds. The background model is not following the actual background structures perfectly, and leaving behind some large, low-surface brightness features in the background-subtracted image. When this image is smoothed with the large PSF for detection, these low-surface brightness regions show up as significant extended flux areas, and many sources are detected. Michael Wood-Vasey has suggested some additional filtering that could be used to check on the significance (as PSFs) of the sources found in these areas. |
