Machine learning analysis of photometric data from the Dark Energy Survey

ABSTRACT In order to retrieve cosmological parameters from photometric surveys, we need to estimate the distribution of the photometric redshift in the sky with excellent accuracy. We use and apply three different machine learning methods to publicly available Dark Energy Survey data release 2 (DR2): (a) artificial neural network for photometric redshifts (ANNz2); (b) Gaussian processes for photometric redshifts (GPz); and (c) Keras, a deep learning application programming interface in python . We compare these different techniques applied to training data obtained from the Visible Multi-Object Spectrograph (VIMOS) Public Extragalactic Redshift Survey (VIPERS) survey. To deal with the incompleteness of the VIPERS catalogue, we use a space-partitioning data structure (K-d Tree) to estimate the reliability of the obtained photometric redshifts. We build a catalogue which is robust to the lack of training data in certain regions of colour space. We use the photometric data to create maps of overdensity as a function of the redshift distribution for more than 500 million galaxies. These maps split the sky into several onion-like redshift slices, which can be readily used for cosmological parameter estimation. On each angular slice, we create and present maps of the angular distribution of galaxies in that slice as well as an estimate of the redshift distribution, $n(z)$, related to the galaxy distribution of that slice, which is recovered from the redshift estimation methods. We achieve a subsample of DES galaxies, which are well matched to the VIPERS sample with an accuracy of the photometric redshifts with a $\sigma _{68}\sim 0.035$ and a catastrophic outlier rate of the order of 3 per cent.