James Buchanan

Gather.town id
MLA18
Poster Title
Gaussian Process Identification of Galaxy Blends for LSST
Institution
Lawrence Livermore National Laboratory
Abstract (short summary)
The Vera C. Rubin Observatory, under construction, will undertake the ten-year Legacy Survey of Space and Time (LSST) beginning in 2023. A significant fraction of observed galaxies will overlap at least one other galaxy along the same line of sight, and so their images must be "deblended" to infer properties of the separate underlying galaxies. Commonly used deblenders rely on an initial estimate of the total number of galaxies participating in a given blend, and this estimate is traditionally made by counting up the number of intensity peaks in a smoothed image of the neighborhood around a blend. However, the reliability of this procedure for LSST images has not yet been comprehensively studied, and the method of peak counting does not naturally assign probabilities to its estimates. Both of these issues are addressed here, the first by constructing a realistic simulation of blends for evaluation, and the second by developing a novel classifier based on a Gaussian process model. This model is shown to have competitive performance compared to the standard peak counting method for identifying blends in i-band images.
Plain text (extended) Summary
A significant fraction of observed galaxies in the LSST will overlap at least one other galaxy along the same line of sight, and so their images must be "deblended" to infer properties of the separate underlying galaxies. Commonly used deblenders rely on an initial estimate of the total number of galaxies participating in a given blend. This estimate is traditionally made by counting up the number of intensity peaks in a smoothed image of the neighborhood around a blend. However, the reliability of this procedure for LSST images has not yet been comprehensively studied, and the method of peak counting does not naturally assign probabilities to its estimates. Both of these issues are addressed here, the first by constructing a realistic set of simulated blends for evaluation, and the second by developing a novel blend classifier based on a Gaussian process model. We show that the i-band blend identification accuracy of the Gaussian process model is competitive with the peak counting method. Furthermore, the Gaussian process model produces a direct estimate of blend probability, and we show that this estimate is generally reliable.
URL
https://www.linkedin.com/in/jjbuchanan