| | 8 | I spent much of last week working on incorporating lessons from photfest into my branch of the IPP code, and attempting to address some related issue encountered along the way. I updated the measurement of the 'dipole parameters' (DIFF_FRATIO, etc) to NOT clip pixels with abs(S/N) < 3 (we mis-understood Armin's definition). Eddie Schlafly and I iterated further on the cosmic ray and galaxy separation issue. For the cosmic ray detection, we agree that a cosmic ray can be reliably identified if the minor axis of the 2nd moments is < about 0.85 pixel^2^, and if the Kron (or basic aperture mag, if needed) is < about -8.5 instrumental mags. For the star/galaxy separation issue, we divided this into two questions: (1) what is the a-priori probability that such-and-such an object is a PSF (in the absence of any knowledge about other kinds of shapes)? (2) given (1) and the distribution of galaxy shapes and sizes, what is the probability that a given source is a star or a galaxy? It looks like (1) can be defined quite well on the basis of the PSF - Kron magnitudes. (2) is much harder, depends on the distribution of galaxy shapes, and is probably outside of the scope of the IPP processing, and more relevant to the morphology analysis client. We can, however, make some initial estimates based on the comparison of GPC1 and the SDSS analysis of the Stripe 82 deep stack. Rather than worrying about getting (2) exactly right, we should be instead be answering the question: for which objects should we perform some further extended source analysis? Not all objects in this class need to have a 100% probability of being a galaxy. The Stripe 82 comparison gives us some guidance, for a specific ratio of stars / galaxies, or equivalently at a specific Galactic latitude. |
| | 10 | Relatedly, we have had reports that the psphot version that is used for the 'static sky' analysis (psphotStack) is not producing all of the outputs expected, including the extended source fits and the radial aperture photometry. I tracked down some configuration and software errors that were responsible these issues (failure to initialize the aperture sizes, default S/N limit for petrosians set quite high, etc). The more significant issue is the failure to generate all of the fits. In trying to understand this, I discovered that the convolution to a target PSF was resulting in a somewhat uglier output PSF than expected, and this was making the source detections go haywire. Although I can get better convolved stacks by tweaking the kernel options, I not entirely satisfied by the result. I've asked Mark Huber to help me with this by seeing how hotpants behaves on some specific image pair covering a wide range of seeing differences. |
