Degradation analysis in the estimation of photometric redshifts from non-representative training sets

TL;DR

This study investigates how non-representative training sets affect photometric redshift estimation, comparing algorithms and training strategies to improve accuracy in real and simulated galaxy data.

Contribution

It demonstrates that training in color space reduces errors with non-representative data and compares the performance of ANNz2 and GPz algorithms under various conditions.

Findings

GPz slightly outperforms ANNz2 in complete training sets for single point estimates.

ANNz2 provides better results in deeper r-band cuts for full redshift distribution estimation.

The proposed Monte-Carlo-based estimator effectively matches spectroscopic galaxy distributions.

Abstract

We perform an analysis of photometric redshifts estimated by using a non-representative training sets in magnitude space. We use the ANNz2 and GPz algorithms to estimate the photometric redshift both in simulations as well as in real data from the Sloan Digital Sky Survey (DR12). We show that for the representative case, the results obtained by using both algorithms have the same quality, either using magnitudes or colours as input. In order to reduce the errors when estimating the redshifts with a non-representative training set, we perform the training in colour space. We estimate the quality of our results by using a mock catalogue which is split samples cuts in the $r$-band between $19.4< r< 20.8$. We obtain slightly better results with GPz on single point z-phot estimates in the complete training set case, however the photometric redshifts estimated with ANNz2 algorithm allows us to obtain mildly better results in deeper $r$-band cuts when estimating the full redshift distribution of the sample in the incomplete training set case. By using a cumulative distribution function and a Monte-Carlo process, we manage to define a photometric estimator which fits well the spectroscopic distribution of galaxies in the mock testing set, but with a larger scatter. To complete this work, we perform an analysis of the impact on the detection of clusters via density of galaxies in a field by using the photometric redshifts obtained with a non-representative training set.