Optimizing Deep Learning Photometric Redshifts for the Roman Space Telescope with HST/CANDELS

We developed an image-based deep learning photometric redshift (photo-z) model called PITA (Photo-z Inference with a Triple-loss Algorithm) that uses semi-supervised contrastive learning to improve photo-z predictions by leveraging images of galaxies both with and without pre-existing measured redshifts. We trained and tested our model on HST/CANDELS, where it outperformed photometry-based methods and other deep learning photo-z models. We provide suggestions and lessons learned for deep learning photo-z models as applied to Roman Space Telescope data.

HST/CANDELS cutout images | reference redshift dataset | PITA code | pre-trained model weights