Galaxy morphoto-Z with neural Networks (GaZNets)

I. Optimized accuracy and outlier fraction from imaging and photometry

¹ School of Astronomy and Space Science, University of Chinese Academy of Sciences, Beijing 100049, PR China
e-mail: liruiww@gmail.com
² National Astronomical Observatories, Chinese Academy of Sciences, 20A Datun Road, Chaoyang District, Beijing 100012, PR China
³ School of Physics and Astronomy, Sun Yat-sen University, Zhuhai Campus, 2 Daxue Road, Xiangzhou District, Zhuhai, PR China
e-mail: napolitano@mail.sysu.edu.cn
⁴ CSST Science Center for Guangdong-Hong Kong-Macau Great Bay Area, Zhuhai, 519082, PR China
⁵ Yunnan Observatories, Chinese Academy of Sciences, Kunming, 650011 Yunnan, PR China
⁶ INAF – Osservatorio Astronomico di Capodimonte, Salita Moiariello 16, 80131 Napoli, Italy
⁷ Center for Theoretical Physics, Polish Academy of Sciences, Al. Lotników 32/46, 02-668 Warsaw, Poland
⁸ INFN – Sezione di Napoli, via Cinthia 9, 80126 Napoli, Italy
⁹ INAF – Osservatorio Astronomico di Padova, via dell’Osservatorio 5, 35122 Padova, Italy

Received: 22 May 2022
Accepted: 16 July 2022

Abstract

Aims. In the era of large sky surveys, photometric redshifts (photo-z) represent crucial information for galaxy evolution and cosmology studies. In this work, we propose a new machine learning (ML) tool called Galaxy morphoto-Z with neural Networks (GaZNet-1), which uses both images and multi-band photometry measurements to predict galaxy redshifts, with accuracy, precision and outlier fraction superior to standard methods based on photometry only.

Methods. As a first application of this tool, we estimate photo-z for a sample of galaxies in the Kilo-Degree Survey (KiDS). GaZNet-1 is trained and tested on ∼140 000 galaxies collected from KiDS Data Release 4 (DR4), for which spectroscopic redshifts are available from different surveys. This sample is dominated by bright (MAG_AUTO < 21) and low-redshift (z <  0.8) systems; however, we could use ∼6500 galaxies in the range 0.8 <  z <  3 to effectively extend the training to higher redshift. The inputs are the r-band galaxy images plus the nine-band magnitudes and colors from the combined catalogs of optical photometry from KiDS and near-infrared photometry from the VISTA Kilo-degree Infrared survey.

Results. By combining the images and catalogs, GaZNet-1 can achieve extremely high precision in normalized median absolute deviation (NMAD = 0.014 for lower redshift and NMAD = 0.041 for higher redshift galaxies) and a low fraction of outliers (0.4% for lower and 1.27% for higher redshift galaxies). Compared to ML codes using only photometry as input, GaZNet-1 also shows a ∼10%−35% improvement in precision at different redshifts and a ∼45% reduction in the fraction of outliers. We finally discuss the finding that, by correctly separating galaxies from stars and active galactic nuclei, the overall photo-z outlier fraction of galaxies can be cut down to 0.3%.