Abstract

Contributed Talk - Splinter Learning

Wednesday, 23 September 2020, 10:05   (virtual room G)

Deep learning for the deblending of high-redshift galaxies

Alexandre Boucaud, Caroline Heneka, Emille E. O. Ishida, Rafael S. de Souza
APC Univ Paris Diderot, Univ Hamburg, Univ Clermont Auvergne, Shanghai Astron. Observatory CAS

Deep neural networks promise to provide photometry and shape measurements for galaxy surveys with high speed and accuracy. This is the more urgent insofar, as the new generation of deep photometric surveys requires unprecedentedly precise shape and photometry measurements of billions of galaxies to achieve their main science goals. At such depths, one major limiting factor is the blending of galaxies due to line-of-sight projection, with an expected fraction of blended galaxies of up to 50%. We therefore explore the use of deep neural networks to estimate the photometry of blended pairs of galaxies in monochrome space images, similar to the ones that will be delivered by the Euclid space telescope. Using an artificially blended sample of galaxies from the CANDELS survey, we train two different network models, based on standard CNN and U-Net architectures, to recover the photometry of the two galaxies, and optionally, their binary segmentation maps. Our networks can recover the original photometry of the galaxies with ~7% accuracy without any assumption on the galaxy shape. This represents an improvement of at least a factor of 4 compared to classical approaches.