A statistical approach towards star-galaxy separation¶
This method uses measurable quantities, such as magnitudes, colours, and s/g classifiers, such as spread_model, along with their associated uncertainties,to estimate the probability that some source is a galaxy or a point source.Besides the measured quantities and their errors, the method uses some prior assumption about the distribution of galaxies and point sources as a function of these same quantities. The method does not need a training set.
The document A simple statistical approach for star-galaxy separation describes the method in its first 4 sections. Section 5 shows its results when applied to the S/G challenge sample. We show its performance in drawing both a galaxy and a stellar sample, using completeness vs. purity curves.
The ROC curve for the test sample is shown below. The solid line shows the result for our method, assuming spread_model_i=0.0015 as the dividing value between galaxies and stars, using galaxy number counts in i-band from Subaru (not exactly the same as DES filters) and stellar counts in magnitude and g-r vs r-i taken from AddStar simulations in the science portal. We assume, for simplicity, that galaxies are uniformly distributed in the same col-col space. The dotted (dashed) lines are based on simple cuts in class_star_i (spread_model_i). No cuts in magnitude were applied.
In Section 6 of the same document above we list some improvements that are planned.