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

Photometric redshifts (photo-$z$’s) will be crucial for studies of galaxy evolution, large-scale structure, and transients with the Nancy Grace Roman Space Telescope. Deep learning methods leverage pixel-level information from ground-based images to achieve the best photo-$z$’s for low-redshift galaxies, but their efficacy at higher redshifts with deep, space-based imaging remains largely untested. We used Hubble Space Telescope CANDELS optical and near-infrared imaging to evaluate fully-supervised, self-supervised, and semi-supervised deep learning photo-$z$ algorithms out to $z\sim3$. Compared to template-based and classical machine learning photometry methods, the fully-supervised and semi-supervised models achieved better performance. Our new semi-supervised model, PITA (Photo-$z$ Inference with a Triple-loss Algorithm), outperformed all others by learning from unlabeled and labeled data through a three-part loss function that incorporates images and colors for all objects as well as redshifts when available. PITA produces a latent space that varies smoothly in magnitude, color, and redshift, resulting in the best photo-$z$ performance even when the redshift training set was significantly reduced. In contrast, the self-supervised approach produced a latent space with significant color and redshift fluctuations that hindered photo-$z$ inference. Looking forward to Roman, we recommend using semi supervised deep learning to take full advantage of the information contained in the hundreds of millions of high-resolution images and color measurements, together with the limited redshift measurements available, to achieve the most accurate photo-$z$ estimates for both faint and bright sources.

💡 Research Summary

This paper investigates how deep‑learning techniques can improve photometric redshift (photo‑z) estimation for the upcoming Nancy Grace Roman Space Telescope, using Hubble Space Telescope CANDELS imaging as a realistic proxy for Roman’s high‑resolution, multi‑band data. The authors assemble a dataset of 99,505 galaxies with four HST filters (F606W, F814W, F125W, F160W), of which 20,624 have reliable redshift labels drawn from spectroscopic measurements, grism redshifts, and the COSMOS2020 many‑band photo‑z catalog. The remaining 78,881 objects are unlabeled and serve to test self‑ and semi‑supervised learning approaches.

Four families of methods are compared: (i) traditional template‑fitting codes (Le PHARE, BPZ, EAZY), (ii) classical machine‑learning regressors (random forest, XGBoost) that operate on catalog photometry, (iii) fully supervised convolutional neural networks (CNNs) that ingest the pixel data directly, and (iv) two novel deep‑learning strategies that exploit unlabeled data. The self‑supervised model uses a contrastive loss to align augmented views of the same galaxy, while the newly introduced semi‑supervised model, named PITA (Photo‑z Inference with a Triple‑loss Algorithm), combines three loss terms: (a) a redshift regression loss applied only to labeled objects, (b) an image reconstruction loss applied to all objects, and (c) a color‑consistency loss that forces the latent representation to respect the measured broadband colors. PITA’s architecture builds on a ResNet‑50 encoder, producing a 128‑dimensional latent vector that is trained jointly on images and catalog colors.

Performance is evaluated using the normalized median absolute deviation (σNMAD) and the outlier fraction (|Δz|/(1+z) > 0.15). Both the fully supervised CNN and PITA outperform the template‑fitting and classical ML baselines, achieving σNMAD ≈ 0.028 and outlier ≈ 7% on the full test set. PITA shows the strongest robustness: when the labeled training set is reduced to 10 % of its original size, σNMAD degrades by only ~10 % and the outlier rate remains below 8 %, whereas the fully supervised CNN’s error grows by >20 %. The self‑supervised model performs poorly in this regime, producing a latent space with abrupt color and redshift variations that translate into large photo‑z errors. Visualizations of the latent space reveal that PITA learns a smooth manifold where magnitude, color, and redshift vary continuously, facilitating reliable interpolation for faint, high‑z galaxies that lack spectroscopic labels.

The authors argue that Roman’s survey will produce hundreds of millions of high‑resolution images but only a relatively small spectroscopic training set, especially at the faint end (H ≈ 26 mag). Semi‑supervised deep learning, exemplified by PITA, can therefore extract the full information content of the images while leveraging the limited redshift labels, delivering more accurate photo‑z estimates for both bright and faint sources. They suggest integrating PITA‑like pipelines into Roman’s data processing, possibly extending the framework to incorporate additional filters, grism/prism redshifts, and active‑learning strategies that iteratively select the most informative galaxies for spectroscopic follow‑up. The study concludes that semi‑supervised deep learning offers a clear path to achieving the photometric redshift precision required for Roman’s cosmology and galaxy‑evolution science cases.