Photometric redshift estimation of galaxies in the P\lowercase{an}-STARRS 3$\pi$ survey- I. Methodology

TL;DR

This paper develops and tests Random Forest models for estimating galaxy redshifts using Pan-STARRS and infrared data, achieving high accuracy and low bias, especially at low redshifts.

Contribution

Introduces a combined photometric redshift estimation method using Random Forests trained on Pan-STARRS and infrared data, with improved performance over previous techniques.

Findings

Bias of 0.001 in redshift estimation

Standard deviation of 0.0225 in normalized redshift error

Outlier rate of 1.48% in test set

Abstract

We present a photometric redshift (photo-$z$) estimation technique for galaxies in the P\lowercase{an}-STARRS1 (PS1) $3\pi $ survey. Specifically, we train and test a regression and a classification Random-Forest (RF) models using photometric features (magnitudes, colors and moments of the radiation intensity) from the optical PS1 data release 2 (PS1-DR2) and from the AllWISE/unWISE infrared source catalogs. The classification RF model ($RF_{clas}$) has better performance in the local universe ($z\lesssim 0.1$), while the second one ($RF_{reg}$) is on average better for $0.1 \lesssim z\lesssim1$. We adopt as labels the spectroscopic redshift of the galaxies from the Sloan Digital Sky Survey (SDSS) data release 16 (SDSS-DR16). We find that the combination of AllWISE/unWISE and PS1-DR2 features leads to an average bias of $\overline{\Delta z_{norm}}=1\times 10^{-3}$, a standard deviation $\sigma(\Delta z_{norm})=0.0225$, (where $\Delta z_{norm} \equiv (z_{phot}-z_{spec})/(1+z_{spec})$), and an outlier rate of $P_0=1.48 \%$ in the test set for the $RF_{clas}$ model. In the low-redshift Universe ($z<0.1$) that is of primary interest to many astronomical transient studies, our model produces an error estimate on the inferred magnitude of an object of $\le$1 mag in 87\% of the test sample.