The domain adaptation problem in photometric redshift estimation: A solution applied to the HSC Survey

Context. The multiband HSC-CLAUDS survey comprises several sky regions with varying observing conditions, only one of which, the COSMOS “Deep”, “Ultra Deep” and “Field” (UDF), offers extensive redshift coverage.

Aims. We aim to exploit a complete sample of labeled galaxies from the COSMOS UDF at i<25(z ≲ 5) to train a convolutional neural network (CNN) and infer more accurate photometric redshifts in the other regions than those currently available from SED-fitting methods.

Methods. To address the severe domain mismatch problem that we observed when applying the trained CNN to regions other than the COSMOS UDF, we developed an unsupervised adversarial domain adaptation network that we grafted onto the CNN. The method was validated by three tests: the predicted redshifts were compared to the spectroscopic redshifts that are available for limited samples of mostly bright galaxies; the predicted redshift distributions of the entire galaxy population of a given field in several intervals of magnitude were compared to those of the COSMOS UDF, assumed to be representative; and the redshifts predicted for a sample of galaxies selected by narrow-band filter observations sensitive to [OII] emitters at z ∼ 1.47 were compared to those of confirmed [OII] emission line galaxies.

Results. The results show successful domain adaptation: the network is able to transfer its redshift classification capability learned from the COSMOS UDF to other regions of HSC-CLAUDS. Accuracy varies depending on magnitude and redshift, following that of the labels we used, but far exceeds that of currently available photometric redshifts. The catalogs of CNN redshifts we inferred for the XMM, DEEP2, and ELAIS fields and for the remaining COSMOS region (∼ 4 million sources in total at i<25) are made public.