Testing the transferability of machine learning techniques for determining photometric redshifts of galaxy catalogue populations

In this work, the random forest algorithm galpro is implemented to generate photometric redshift posteriors, and its performance when trained and then applied to data from another survey is investigated. The algorithm is initially calibrated using a truth data set compiled from the Dark Energy Spectroscopic Instrument (DESI) Legacy survey. We find that the testing and training data sets must have very similar redshift distributions, with the range of their photometric data overlapping by at least 90 per cent in the appropriate photometric bands in order for the training data to be applicable to the testing data. galpro is again trained using the DESI data set and then applied to a sample drawn from the Panoramic Survey Telescope and Rapid Response System survey, to explore whether galpro can be trained using a trusted data set and applied to an entirely new survey, albeit one that uses a different magnitude system for its photometric bands, thus requiring careful conversion of the measured magnitudes. The results of this further test indicate that galpro does not produce accurate photometric redshift posteriors for the new survey, even where the distribution of redshifts for the two data sets overlaps by over 90 per cent. We conclude that the photometric redshifts generated by galpro are not suitable for generating estimates of photometric redshifts and their posterior distribution functions when applied to an entirely new survey, particularly one that uses a different magnitude system. However, our results demonstrate that galpro is a useful tool for inferring photometric redshift estimates in the case where a spectroscopic galaxy survey is nearly complete, but missing some spectroscopic redshift values.