Photometric redshift estimation of quasars with fused features from photometric data and images

Abstract Redshift is a crucial parameter of quasars and performs a greatly important role in cosmological studies. This paper proposes a network called quasar photometric redshift (photo-z or zphoto) estimation network (Q-PreNet) that integrates images and photometric data to estimate the redshifts of quasars. To enhance the information richness, optical and infrared data, separately from Sloan Digital Sky Survey (SDSS) and Wide-field Infrared Survey Explorer (WISE), are used. In Q-PreNet, on the one hand, an image feature extraction network (IfeNet) is designed to obtain image features. On the other hand, magnitudes after extinction and their mutual differences are taken as the features of photometric data. Two features are then concatenated to form fused features. Finally, a regression network to estimate photo-z (RegNet-z) is proposed based on mixture density network (MDN), due to its ability to provide uncertainty information. To measure the uncertainty, two quantitative metrics are proposed. Experimental results show that the performance of Q-PreNet is superior: While using fused features, the proportion of samples with |Δz| = |(zspec − zphoto)/(1 + zspec)| (spectroscopic redshifts, spec-z or zspec) less than 0.15 can reach 86.3 per cent with the reduction of 8.15 per cent and 9.37 per cent, which is compared with separately using images and photometric data only. Compared with literatures, Q-PreNet offers a substantial improvement in the redshift estimation of quasars and this is significant for large-scale sky surveys.