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

TL;DR

This work addresses the challenge of obtaining accurate photometric redshifts for the Roman Space Telescope's deep, high-redshift imaging by benchmarking deep learning approaches on HST CANDELS data. It compares a photometry-based baseline, a fully-supervised CNN, a self-supervised contrastive model, and introduces PITA, a semi-supervised framework that jointly optimizes contrastive, color, and redshift losses to leverage unlabeled data. PITA consistently yields the best photo-$z$ performance across metrics and magnitude ranges, with strong scalability as labeled data grows, while self-supervised pretraining alone underperforms at higher redshifts. The findings suggest that for Roman, semi-supervised, multi-task learning that exploits abundant unlabeled imaging data and incorporates domain-informed color and redshift objectives will maximize photo-$z$ accuracy, especially for label-scarce faint populations. These results provide practical guidance for Roman data processing pipelines and motivate further exploration of multi-modal, foundation-model–style approaches with careful task design.

Abstract

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.

Figures (13)

• Figure 1: Total system throughputs for HST/CANDELS filters (black) and effective areas for Roman $ZYJHFK$ filters (colored). The Roman HLWAS imaging will include all these bands in the deep tier, but only $YJH$ in the medium tier. These bands can be combined with Rubin $ugrizy$ optical photometry to obtain the best possible photo-$z$'s for Roman. In this context, the CANDELS F125W and F160W filters provide near-infrared wavelength coverage comparable to what is available for Roman's medium tier, while F606W and F814 provide optical coverage comparable to the redder Rubin bands, albeit at higher spatial resolution.
• Figure 2: Examples of cutout images for four different galaxies ordered by increasing magnitude in the F160W band. The RGB images are formed using F606W as blue, F125W as green, and F160W as red. For illustrative purposes the pixel values in these images are scaled with an inverse hyperbolic sine function. Galaxies up to $m_\mathrm{F160W}\sim{25}$ have well-measured pixel-to-pixel color variations; this information can potentially be leveraged by deep-learning algorithms.
• Figure 3: Left: Redshift distributions of the labeled samples (20,624 galaxies) split by spec-$z$'s, grism-$z$'s, and COSMOS2020 many-band photo-$z$'s. Right: $m_\text{F160W}$ distributions of the same labeled samples, in addition to the unlabeled photometric sample ($78,881$ galaxies). The photometric sample is deeper than most of the labeled data.
• Figure 4: Architecture of the fully-supervised redshift prediction model. A ConvNeXt CNN takes as input a four-band image and outputs a 1000-dimensional feature vector, which is then passed to an MLP that predicts a scalar redshift. This network is trained exclusively on galaxies with redshift labels.
• Figure 5: Illustration of the data augmentations used when training the fully-supervised algorithm. Starting with a $108\times108$ pixel image, we apply a horizontal flip with 50% probability, followed by a random rotation. Then, we jitter the image and crop to the central $64\times64$ pixels. Lastly, we add uncorrelated Gaussian noise. This figure illustrates the effects of the transformations on an RGB composite image, but during training the same transformations are applied on each of the four bands. These transformations make the network more robust and generalizable, and importantly they do not change the characteristics of the images which should provide information about redshift (e.g., color and morphology).