Towards an LSST Platinum Sample: Using Machine Learning to Identify Galaxies with Highly Accurate Photo-z’s

The Legacy Survey of Space and Time will observe 25 billion galaxies, generating a key question: given this surfeit of statistics but only six photometric bands, can we select a “Platinum Sample” of galaxies that have highly accurate photometric redshifts to better probe large-scale structure? We investigate Extreme Emission Line galaxies (EELs), galaxies for which emission lines represent a large fraction of the total light detected in a broadband filter, as candidates for an LSST Platinum Sample due to the redshift constraints provided by their compact and luminous spectral features. We find that EELs for which one or more bright emission lines fall in single filters fail to produce highly accurate photometric redshifts. As an alternative approach, we have performed redshift fits on Subaru Strategic Program galaxies using Trees for Photo-Z (TPZ) with the goal of identifying galaxies with high-accuracy photo-z’s. We approximate the extrapolation of photo-z fits to galaxies without spectroscopic redshifts by forming training and test sets with similar differences in i-band flux and g-z color to that of the galaxies with and without spectroscopic redshifts. We find even with the known differences between the training and test sets, a neural network regressor applied to the TPZ fits is able to provide corrections that reduce the number of outliers by ~40%. Further, training a neural network classifier to identify galaxies with σ < 0.02 (1+z) using galaxy photometry, associated TPZ fits, and regressor corrections improves the standard deviation of the error distribution by a factor of several while keeping roughly a quarter of the sample.