80 likes | 196 Vues
This project focuses on implementing James Jee's "Stackfit" algorithm within the LSST software stack, ensuring consistency through direct comparisons of PSF models, stacks, and shape measurements. The initial implementation targets DLS data from the F2 field, involving creation of calibrated exposures, spatially-varying PSF models, and weighted coadds. We've successfully modeled PSFs for 10,000 objects and used LSST software for astrometric calibration and stacking, though some results require further refinement. Performance assessments show promising results using the Python-wrapped Mpfit code for parameter fitting.
E N D
Conversion of Stackfit to LSST software stack Status as of Feb 20, 2012
Goal of this projects • Implement James Jee’s “Stackfit” algorithm in the LSST software stack. • Demonstrate consistency between the LSST version and James’ implementation by direct comparison of PSF models, stacks, and shape measurements. • Initial implementation to be done on DLS data (from F2 field).
Key Components • Create calibrated exposure for each CCD image (247 CCD images in DLS F2p23). • Create spatially-varying PSF model for each CCD image. • Create weighted coadd of the entire field. • Create a source catalog from the coadd. • Create similarly weighted coadd of psfs for each object in the source catalog.
Key Components(cont’d) • Estimate the e1 and e2 components for each source in the coadd catalog. • The estimate is done by fitting a 7 parameter Gaussian or 9 parameter Sersic model. • The data is a cutout of each object with a square 4 * A, where A is the major axis dimension, estimated from moments. • The model is created by convolving a model matrix of the same size as the cutout with the stacked PSF for that object.
Status of PSF estimation • Have run the LSST psf modeling code on 10,000 objects from DLS F2p23. • The psfs models were used to create kernels at the centroids of bright objects. • The kernels were compared with the kernels from James’ psf catalogs by fitting to 2D elliptical Gaussian model • The difference in the estimates of σA and σB were almost uniformly smaller than the parameter errors.
Status of Calibration/Stacking • Used LSST software to make calibrated exposures from 30 DLS images. • The CCD astrometric calibration is the most important part. I used the DLS star catalog for calibration, I created a TAN-SIP correction. • The results were rather spotty – I concluded that we need to be able to combine multiple exposures from each CCD to model distortions more accurately. • The resulting stacks are not good enough for our purpose. However, I moved on ...
Status of shape measurement • Able to build stacks and create source catalogs. • Able to create psf stacks for each source. • Applied Minuit2 minimizer to the estimate a 2D Gaussian profile for each source. MiGrad does not converge very reliably. • Imported C translation of mpfit. This code is faster and converges more reliably.
Performance Assessment • Ran 1000 sources with footprints > 20 pixels for a 7 parameter Gaussian model. • The Python wrapped Mpfit code did the parameter fits in 191 s. • Ran 100 sources with footprints > 500 pixels. Mpfit took 65 seconds. • This was on a single core of an AMD Opteron 2427 2.2 GHz system with 32 GB of memory.