Photometric Redshift Estimation for Rubin Observatory Data Preview 1 with Redshift Assessment Infrastructure Layers (RAIL)

TL;DR

The study assesses photometric redshift estimation for Rubin DP1 data using the Redshift Assessment Infrastructure Layers (RAIL), applying eight algorithms (template-fitting and machine learning) to DP1 photometry cross-matched with ECDFS, DESI DR1, and Euclid crossmatches. It demonstrates that, with a representative training set, machine-learning methods achieve a bias below $| ext{E}[oldsymbol{ abla z}]| \,<\,0.005$, a scatter around $ ext{NMAD} \,\sim\,0.03$, and outlier fractions near $10\%$ for 6-band data, satisfying LSST Y1 requirements; including Euclid NIR data further improves high-$z$ performance ($z_{ m ref}>1.2$). The work validates the RAIL pipeline for Rubin photo-$z$ production and provides ensemble $n(z)$ estimates, while highlighting caveats related to training-set variance, non-detections, and the need for continued hyperparameter and flux-metric optimizations. The results establish a solid foundation for real Rubin analyses and inform future refinements to maximize photo-$z$ accuracy across redshift and magnitude ranges.

Abstract

We present the first systematic analysis of photometric redshifts (photo-z) estimated from the Rubin Observatory Data Preview 1 (DP1) data taken with the Legacy Survey of Space and Time (LSST) Commissioning Camera. Employing the Redshift Assessment Infrastructure Layers (RAIL) framework, we apply eight photo-z algorithms to the DP1 photometry, using deep ugrizy coverage in the Extended Chandra Deep Field South (ECDFS) field and griz data in the Rubin_SV_38_7 field. In the ECDFS field, we construct a reference catalog from spectroscopic redshift (spec-z), grism redshift (grism-z), and multiband photo-z for training and validating photo-z. Performance metrics of the photo-z are evaluated using spec-zs from ECDFS and Dark Energy Spectroscopic Instrument Data Release 1 samples. Across the algorithms, we achieve per-galaxy photo-z scatter of $σ_{\rm NMAD} \sim 0.03$ and outlier fractions around 10% in the 6-band data, with performance degrading at faint magnitudes and z>1.2. The overall bias and scatter of our machine-learning based photo-zs satisfy the LSST Y1 requirement. We also use our photo-z to infer the ensemble redshift distribution n(z). We study the photo-z improvement by including near-infrared photometry from the Euclid mission, and find that Euclid photometry improves photo-z at z>1.2. Our results validate the RAIL pipeline for Rubin photo-z production and demonstrate promising initial performance.

Figures (9)

• Figure 1: Training set in the ECDFS field. Top left: redshift distribution by methods to obtain redshift. Top right: redshift distribution by surveys. Bottom left: Scatter plot of the ECDFS reference catalog color-coded by redshift. Bottom right: scatter plot of the ECDFS reference catalog color-coded by the survey name. The 1-D distributions are normalized. Our reference galaxy catalog is constructed by a wide range of redshifts, and spans a wide range of redshifts.
• Figure 2: Corner plot of the training set color-magnitude-redshift distributions in the ECDFS field. Each panel shows $95\%$ confidence contours among redshift, adjacent-band colors ($u-g$, $g-r$, $r-i$, $i-z$, $z-y$), and the $i$-band magnitude. The blue contours represent galaxies with spec-$z$s, the green contours represent grism-$z$s, the red contours represent the multiband photo-$z$, while the black contours show all galaxies in the ECDFS field with i-band SNR greater than 5. The spec-$z$, grism-$z$ and photo-$z$ samples compose $53\%$, $34\%$, and $13\%$ of the entire training/testing galaxies, respectively. Our reference sample covers a reasonable color space of our ECDFS samples; however, in some color spaces, we still lack reference galaxies.
• Figure 3: Redshift and $i$-magnitude distribution for matched DESI objects. For the scatter plot and histograms, the BGS sample is shown in green, ELG in orange, and LRG in purple. The distribution of the total sample is shown in black in the outer histograms. The $i$-mag distribution peaks at $i=23$ and the redshift distribution at $z = 0.8$.
• Figure 4: Two-dimensional histograms comparing the mode of the photometric redshifts ($z_{\mathrm{phot}}$) to reference redshifts ($z_{\mathrm{ref}}$) for eight photo-$z$ algorithms applied to the DP1 test sample. Each panel shows one algorithm: FlexZBoost, kNN, CMNN, DNF, TPZ, GPz, BPZ, and LePhare. The grey dashed lines indicate the identity line ($z_{\mathrm{phot}}=z_{\mathrm{ref}}$) and the $|\Delta z| = 0.15$ lines. The color scale represents the number of objects in each bin on a logarithmic scale. These plots illustrate the overall agreement and outlier behavior of each algorithm across the redshift range $0 < z < 3$.
• Figure 5: Photo-$z$ performance metrics as a function of redshift (top row) and $i$-band magnitude (bottom row) for eight algorithms. Each panel shows three metrics: the photo-$z$ bias $\mathbb{E} [ \Delta z ]$, scatter $\sigma_{\rm NMAD}$, and outlier rate $\eta_{0.15}$. The grey shaded regions show the LSST Y1 and Y10 requirements on the mean and scatter of photo-$z$desc_srd. The photo-$z$ of all algorithms starts to deteriorate when $z>1.2$ and $i$-mag$>23$. We notice that some algorithms display significant bias at low and high redshift, and template-fitting methods have more bias across the magnitude range than empirical methods. We do not show statistics beyond $z_{\rm ref} = 2$ because of the limited number of reference galaxies.