Enhancing Photometric Redshift Catalogs Through Color-Space Analysis: Application to KiDS-Bright Galaxies

TL;DR

This paper introduces a method using self-organizing maps to analyze and improve the accuracy of photometric redshift catalogs by identifying poorly represented galaxies in spectroscopic training samples.

Contribution

The authors develop a SOM-based approach to assess and enhance the reliability of photometric redshift estimates by analyzing the representation of galaxies in color space.

Findings

GAMA under-represents faint and high-redshift galaxies in KiDS-Bright

Incorporating additional spectroscopic data helps identify suboptimal photo-$z$ estimates

Excluding poorly represented galaxies modestly reduces photo-$z$ scatter by less than 10

Abstract

We present a method to refine photometric redshift galaxy catalogs by comparing their color-space matching with overlapping spectroscopic calibration data. We focus on cases where photometric redshifts (photo-$z$) are estimated empirically. Identifying galaxies that are poorly represented in spectroscopic data is crucial, as their photo-$z$ may be unreliable due to extrapolation beyond the training sample. Our approach uses a self-organizing map (SOM) to project a multi-dimensional parameter space of magnitudes and colors onto a 2-D manifold, allowing us to analyze the resulting patterns as a function of various galaxy properties. Using SOM, we compare the Kilo-Degree Survey bright galaxy sample (KiDS-Bright), limited to $r<20$ mag, with various spectroscopic samples, including the Galaxy And Mass Assembly (GAMA). Our analysis reveals that GAMA under-represents KiDS-Bright at its faintest ($r\gtrsim19.5$) and highest-redshift ($z\gtrsim0.4$) ranges, however no strong trends in color or stellar mass. By incorporating additional spectroscopic data from the SDSS, 2dF, and early DESI, we identify SOM cells where photo-$z$ are estimated suboptimally. We derive a set of SOM-based criteria to refine the photometric sample and improve photo-$z$ statistics. For the KiDS-Bright sample, this improvement is modest: exclusion of the least represented 20% of the sample reduces photo-$z$ scatter by less than 10%. We conclude that GAMA, used for KiDS-Bright photo-$z$ training, is sufficiently representative for reliable redshift estimation across most of the color space. Future spectroscopic data from surveys such as DESI should be better suited for exploiting the full improvement potential of our method.