Probabilistic photo-z machine learning models for X-ray sky surveys

TL;DR

This paper develops probabilistic machine learning models using Random Forests to improve photo-z predictions for X-ray sources, leveraging multiple photometric surveys and a large training dataset, achieving near double the accuracy of previous methods.

Contribution

The authors introduce a novel Random Forest-based probabilistic photo-z model trained on a large quasar and galaxy sample, incorporating extinction and measurement uncertainties, for X-ray sky surveys.

Findings

Achieved photo-z accuracy of NMAD=0.034 on Stripe82X

Reduced catastrophic outliers to 8.8%

Nearly doubled the accuracy compared to previous methods

Abstract

Accurate photo-z measurements are important to construct a large-scale structure map of X-ray Universe in the ongoing SRG/eROSITA All-Sky Survey. We present machine learning Random Forest-based models for probabilistic photo-z predictions based on information from 4 large photometric surveys (SDSS, Pan-STARRS, DESI Legacy Imaging Survey, and WISE). Our models are trained on the large sample of $\approx$580000 quasars and galaxies selected from the SDSS DR14 spectral catalog and take into account Galactic extinction and uncertainties in photometric measurements for target objects. On the Stripe82X test sample we obtained photo-z accuracy for X-ray sources: $NMAD=0.034$ (normalized median absolute deviation) and $n_{>0.15}=0.088$ (catastrophic outliers fraction), which is almost $\sim2$ times better than best photo-z results available in the literature.