Deep learning based photometric redshifts for the Kilo-Degree Survey Bright Galaxy Sample
Anjitha John William, Priyanka Jalan, Maciej Bilicki, Wojciech A. Hellwing
TL;DR
This work tackles the challenge of estimating photometric redshifts in large imaging surveys where spectroscopic redshifts are scarce. It introduces a multi-input deep learning framework that fuses 4-band KiDS images with 9-band KiDS+VIKING magnitudes, trained on GAMA spectroscopic redshifts, and employs an Inception CNN alongside an ordinary NN with a $\delta$-Huber loss. The authors demonstrate improved photo-$z$ performance over a prior ANNz2-based approach, reporting $\langle \Delta z \rangle \approx 0.001$ and $\mathrm{SMAD}(\Delta z) \approx 0.016$ on the KiDS-GAMA equatorial test set, with $R^2 > 0.92$ during validation. The approach showcases the potential of DL for precise redshift estimation in wide-field surveys and paves the way for applying the method to the full KiDS DR4 Bright Galaxy Sample and future KiDS DR5 analyses.
Abstract
In cosmological analyses, precise redshift determination remains pivotal for understanding cosmic evolution. However, with only a fraction of galaxies having spectroscopic redshifts (spec-$z$s), the challenge lies in estimating redshifts for a larger number. To address this, photometry-based redshift (photo-$z$) estimation, employing machine learning algorithms, is a viable solution. Identifying the limitations of previous methods, this study focuses on implementing deep learning (DL) techniques within the Kilo-Degree Survey (KiDS) Bright Galaxy Sample for more accurate photo-$z$ estimations. Comparing our new DL-based model against prior `shallow' neural networks, we showcase improvements in redshift accuracy. Our model gives mean photo-$z$ bias $\langle Δz\rangle= 10^{-3}$ and scatter $\mathrm{SMAD}(Δz)=0.016$, where $Δz = (z_\mathrm{phot}-z_\mathrm{spec})/(1+z_\mathrm{spec})$. This research highlights the promising role of DL in revolutionizing photo-$z$ estimation.