A Comparison of Six Photometric Redshift Methods Applied to 1.5 Million Luminous Red Galaxies

We present an updated version of MegaZ-LRG (Collister et al.,(2007)) with photometric redshifts derived with the neural network method, ANNz as well as five other publicly available photo-z codes (HyperZ, SDSS, Le PHARE, BPZ and ZEBRA) for ~1.5 million Luminous Red Galaxies (LRGs) in SDSS DR6. This allows us to identify how reliable codes are relative to each other if used as described in their public release. We compare and contrast the relative merits of each code using ~13000 spectroscopic redshifts from the 2SLAQ sample. We find that the performance of each code depends on the figure of merit used to assess it. As expected, the availability of a complete training set means that the training method performs best in the intermediate redshift bins where there are plenty of training objects. Codes such as Le PHARE, which use new observed templates perform best in the lower redshift bins. All codes produce reasonable photometric redshifts, the 1-sigma scatters ranging from 0.057 to 0.097 if averaged over the entire redshift range. We also perform tests to check whether a training set from a small region of the sky such as 2SLAQ produces biases if used to train over a larger area of the sky. We conclude that this is not likely to be a problem for future wide-field surveys. The complete photometric redshift catalogue including redshift estimates and errors on these from all six methods can be found at star.ucl.ac.uk 14 pages, 8 figures, submitted to MNRAS