Hybrid-z: Enhancing Kilo-Degree Survey bright galaxy sample photometric redshifts with deep learning

TL;DR

This paper introduces Hybrid-z, a deep learning model that significantly improves photometric redshift estimates for the KiDS-Bright galaxy sample by combining multi-band imaging data, outperforming previous methods and enabling more accurate cosmological analyses.

Contribution

The paper presents Hybrid-z, a novel deep learning framework that integrates four-band images with nine-band magnitudes to enhance photometric redshift accuracy for large galaxy surveys.

Findings

20% reduction in redshift scatter compared to previous methods

Stable performance across multiple spectroscopic datasets

Effective application to 1.2 million galaxies in KiDS-Bright DR4

Abstract

We employ deep learning (DL) to improve photometric redshifts (photo-$z$s) in the Kilo-Degree Survey Data Release 4 Bright galaxy sample (KiDS-Bright DR4). This dataset, used as a foreground for KiDS lensing and clustering studies, is flux-limited to $r<20$ mag with mean $z=0.23$ and covers 1000 deg$^2$. Its photo-$z$s were previously derived with artificial neural networks from the ANNz2 package, trained on the Galaxy And Mass Assembly (GAMA) spectroscopy. Here we considerably improve over these previous redshift estimations by building a DL model, Hybrid-z, which combines four-band KiDS images with nine-band magnitudes from KiDS+VIKING. The Hybrid-z framework provides photo-$z$s for KiDS-Bright, with negligible mean residuals of O($10^{-4}$) and scatter at the level of $0.014(1+z)$ -- reduction by 20% over the previous nine-band derivations with ANNz2. We check our photo-$z$ model performance on test data drawn from GAMA, as well as from other KiDS-overlapping wide-angle spectroscopic surveys, namely SDSS, 2dFLenS, and 2dFGRS. We find stable behavior and consistent improvement over ANNz2 throughout. We finally apply Hybrid-z trained on GAMA to the entire KiDS-Bright DR4 sample of 1.2 million galaxies. For these final predictions, we design a method of smoothing the input redshift distribution of the training set, to avoid propagation of features present in GAMA, related to its small sky area and large-scale structure imprint in its fields. Our work paves the way towards the best-possible photo-$z$s achievable with machine learning for any galaxy type both for the final KiDS-Bright DR5 data and for future deeper imaging, such as from the Legacy Survey of Space and Time.