The PAU survey: Estimating galaxy photometry with deep learning

Abstract

With the dramatic rise in high-quality galaxy data expected from Euclid and Vera C. Rubin Observatory, there will be increasing demand for fast high-precision methods for measuring galaxy fluxes. These will be essential for inferring the redshifts of the galaxies. In this paper, we introduce Lumos, a deep learning method to measure photometry from galaxy images. Lumos builds on BKGnet, an algorithm to predict the background and its associated error, and predicts the background-subtracted flux probability density function. We have developed Lumos for data from the Physics of the Accelerating Universe Survey (PAUS), an imaging survey using a 40 narrow-band filter camera (PAUCam). PAUCam images are affected by scattered light, displaying a background noise pattern that can be predicted and corrected for. On average, Lumos increases the SNR of the observations by a factor of 2 compared to an aperture photometry algorithm. It also incorporates other advantages like robustness towards distorting artefacts, e.g. cosmic rays or scattered light, the ability of deblending and less sensitivity to uncertainties in the galaxy profile parameters used to infer the photometry. Indeed, the number of flagged photometry outlier observations is reduced from 10% to 2%, comparing to aperture photometry. Furthermore, with Lumos photometry, the photo-z scatter is reduced by ~10% with the Deepz machine learning photo-z code and the photo-z outlier rate by 20%. The photo-z improvement is lower than expected from the SNR increment, however currently the photometric calibration and outliers in the photometry seem to be its limiting factor.