Photometric Redshifts for Hyper Suprime-Cam Subaru Strategic Program Data Release 1

TL;DR

This study evaluates multiple photo-z codes for the HSC-SSP PDR1 dataset, comparing empirical, neural-network, hybrid, SOM-based, NN-based, and template-fitting approaches across Deep/UltraDeep and Wide data. It introduces a risk-based best-point estimator and a unified PDF performance framework, showing peak accuracy in 0.2<z<1.5 with sigma(z)/(1+z) ~ 0.05 and outlier rates around 15% for i<25, and better metrics for i<24. The analysis highlights the importance of training data quality, PDF calibration, and depth/seeing effects, and provides publicly accessible photo-z products (point estimates and full PDFs). The results support broad scientific use for HSC-SSP while outlining avenues for improvement, including multi-code synthesis and clustering-based N(z) constraints.

Abstract

Photometric redshifts are a key component of many science objectives in the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP). In this paper, we describe and compare the codes used to compute photometric redshifts for HSC-SSP, how we calibrate them, and the typical accuracy we achieve with the HSC five-band photometry (grizy). We introduce a new point estimator based on an improved loss function and demonstrate that it works better than other commonly used estimators. We find that our photo-z's are most accurate at 0.2<~zphot<~1.5, where we can straddle the 4000A break. We achieve sigma(d_zphot/(1+zphot))~0.05 and an outlier rate of about 15% for galaxies down to i=25 within this redshift range. If we limit to a brighter sample of i<24, we achieve sigma~0.04 and ~8% outliers. Our photo-z's should thus enable many science cases for HSC-SSP. We also characterize the accuracy of our redshift probability distribution function (PDF) and discover that some codes over/under-estimate the redshift uncertainties, which have implications for N(z) reconstruction. Our photo-z products for the entire area in the Public Data Release 1 are publicly available, and both our catalog products (such as point estimates) and full PDFs can be retrieved from the data release site, https://hsc-release.mtk.nao.ac.jp/.

Figures (15)

• Figure 1: The original (dashed) and re-weighted (solid) normalized number densities of our training sample as a function of $grizy$ (left-to-right) magnitude (top) and error (bottom). Note that we use asinh magnitudes (i.e., Luptitudes) here. Our color-magnitude weights are able to effectively correct for biases in our original training sample to better mimic the HSC-SSP Wide data.
• Figure 2: The re-weighted, normalized redshift number density for our training sample. The full distribution is shown in solid black while the spec-$z$, g/prism-$z$, and many-band photo-$z$ components are shown in solid red, purple, and blue, respectively. The dashed lines show these same components re-normalized to the full sample in order to better highlight their differences. We can see that most of the substructure in the redshift distribution of our training sample comes from the many-band COSMOS photo-$z$'s, which also contribute almost all of our high-$z$ sources.
• Figure 3: Bias, $f_{outlier,conv}$ and $\sigma_{conv}$ plotted against $i$-band magnitude. The different panels are for different codes as indicated by the label on the top-left corner of each panel. The gray shades show $\pm0.01$ range, which will be useful for bias. The symbols are explained in the panels. Note that these plots are based on the COSMOS Wide-depth median stack and include objects in COSMOS only.
• Figure 5: Same as Fig. \ref{['fig:stat_mag']} but as a function of $z_{phot}$.
• Figure 7: Relationship between loss and other metrics. The symbols are color-coded according to the $i$-band magnitude cut applied. This is for MLZ using the Wide-depth median seeing catalog, but the other codes show similar trends.