Photometric Redshifts


The HSC photo-z team computed photometric redshifts using several independent codes for this public release.  As described in detail in the photo-z release paper, we first constructed a training sample by combining public spectroscopic redshifts, HST grism redshifts, and high-quality many-band photometric redshifts in COSMOS with weights to each objects to reproduce the color-magnitude distributions of the HSC objects.  We then performed a classical hold-out cross-validation or 5-fold cross-validation (validation technique depends on the code) to train and validate our codes.  Both the held-out test sample and COSMOS wide-depth stacks, in which the photometry is quasi-independent from the training sample, are used to evaluate the performance.  The trained codes are then run on the target catalogs for production.

Our catalog products are stored in the database and the schema browser gives the details of the catalogs.  As discussed in the photo-z release paper, we suggest that zbest should be used for point estimates and zrisk as an indicator of reliability.  In addition to the catalogs, we also release the redshift probability distribution functions in the fits format below.  Further description of our training procedure, data products, explanation of the statistics plots below can be found in this release note.


We have computed photo-z’s using several independent codes.  Here is a brief summary of our codes.

DEmP: Combination of nearest neighbor technique and polynomial (actually linear) fitting.  Redshift of each object is estimated using the 40 nearest neighbors in the nine-dimensional space (5 magnitude axes and g-r, r-i, i-z, z-y color axes using the PSF-matched aperture photometry) with a linear function.

NNPZ:  Another nearest neighbor code.  It searches for neighbors in the i, g-r, r-i, i-z, z-y magnitude/color space and PDF is constructed from the sum of neighbor’s redshifts with weights using the inverse of the Euclidian distance from the mag/color space.  The code uses the CModel photometry.

MLZ Self-Organizing Map from the machine-learning photo-z package by Carrasco Kind and Brunner 2014.  Both CModel and PSF-matched aperture photometry is used.

Franken-Z Combination of machine-learning and template fitting.  An ensemble of k-d trees with observed features of the training sample is used to search for nearest neighbors in the feature space.  PDFs are generated by stacking each neighbor’s redshift KDE weighted by likelihood. PSF-matched aperture photometry is used by the code.

Ephor:  (source: CModel version, PSF-matched version) Neural Network.  Photo-z’s computed with CModel and PSF-matched photometry are available in separate tables.

Mizuki:  Template fitting with Bayesian priors on physical properties of galaxies.  In addition to redshifts, physical parameters such as stellar mass and SFRs are computed.  The code uses CModel.

Statistics plots

Here are some basic statistics plots for the COSMOS wide-depth median seeing stack.  The accuracy here represents the average accuracy of our photo-z in the Wide layer.  Currently, the Deep layer is only slightly deeper than Wide and thus these plots will give you good guidance.  In UltraDeep, both the outlier rates and scatter are smaller by a factor of ~2.  More details can be found in the photo-z release paper (in prep).

Probability Distribution Functions

Our P(z) files are available for each field in the fits format (the file list).  Note these are massive files!  Expanding tar archives, you will get many fits files: one file for one tract. Tract numbers will be known from the file names.
Some important notes:

  1. One fits file for one tract.
  2. The 1st HDU contains P(z) and the 2nd HDU defines the redshift grids.
  3. Please do not use the header keywords to define the redshift grids.  Please always use the 2nd HDU.