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

¹ Center for Theoretical Physics, Polish Academy of Sciences, al. Lotników 32/46, 02-668 Warsaw, Poland
² National Centre for Nuclear Research (NCBJ), ul. Pasteura 7, 02-093 Warsaw, Poland
³ Division of Physics, Mathematics and Astronomy, California Institute of Technology, 1200 E California Blvd, Pasadena, CA 91125, USA

^⋆ Corresponding authors: anjithajm@cft.edu.pl, bilicki@cft.edu.pl

Received: 21 December 2024
Accepted: 5 May 2025

Abstract

We employed deep learning to improve the photometric redshifts (photo-zs) in the Kilo-Degree Survey Data Release 4 bright galaxy sample (KiDS-DR4 Bright). This dataset, used as foreground for KiDS lensing and clustering studies, is flux-limited to r < 20 mag with mean z = 0.23 and covers 1000 deg². Its photo-zs were previously derived with artificial neural networks from the ANNz2 package trained on the Galaxy And Mass Assembly (GAMA) spectroscopy. Here, we considerably improve on these previous redshift estimations by building a deep learning model, Hybrid-z, that combines an inception-based convolutional neural network operating on four-band KiDS images with an artificial neural network using nine-band magnitudes from KiDS+VIKING. The Hybrid-z framework provides state-of-the-art photo-zs for KiDS-Bright with negligible mean residuals of O(10⁻⁴) and scatter at a level of 0.014(1 + z) – representing a reduction of 20% compared to the previous nine-band derivations with ANNz2. Our photo-zs are robust and stable independently of galaxy magnitude, redshift, and color. In fact, for blue galaxies, which typically have more pronounced morphological features, Hybrid-z provides a larger improvement over ANNz2 than for red galaxies. We checked 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 found stable behavior and consistent improvement over ANNz2 throughout. Finally, we applied Hybrid-z trained on GAMA to the entire KiDS-Bright DR4 sample of 1.2 million galaxies. For these final predictions, we designed a method of smoothing the input redshift distribution of the training set in order 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 toward the best-possible photo-zs achievable with machine learning for any galaxy type for both the final KiDS-Bright DR5 data and for future deeper imaging, such as from the Legacy Survey of Space and Time.