Photometric redshifts for S-PLUS using machine learning techniques

• Ano de defesa: 2019
• Autor(a) principal: Lima, Erik Vinicius Rodrigues de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado de máquina; Galáxias; Galaxies; Galaxy surveys; Levantamentos de galáxias.; Machine learning; Photometric redshift; Redshift fotométrico
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/14/14131/tde-14082020-174722/

This work focuses on the obtention of photometric redshifts of galaxies using the machine learning codes ANNz2, GPz, and Deep Learning models made with Keras. We take advantage of the great opportunity that the new multiband survey of the austral sky, called Southern Photometric Local Universe Survey (S-PLUS), provides due to the adoption of an unique filter system, composed of five broad-band filters and seven narrow-band filters. Besides the use of magnitudes, it is also possible to use non-photometric features with machine learning methods, such as object sizes, their full width at half maximum and their maximum surface brightness, in order to improve the results. This work used data from the S-PLUS Data Release 1, together with two other large projects, the Sloan Digital Sky Survey (SDSS) Data Release 15, and the unWISE catalogue of the Wide-field Infrared Survey Explorer (WISE), in the Stripe-82 region. Amongst the three algorithms compared in this work, the one which performs better overall is the deep-learning based method. The photometric redshifts obtained with this code have precision of 2.49% for galaxies with r-band magnitude between 16 and 21, bias of 0.4% and outlier fraction equal to 0.64%. When compared to the currently used method for photometric redshift determination in S-PLUS, the template fitting code BPZ, it is noticed that machine learning methods have higher accuracy, less bias and lower outlier fraction. An analysis regarding the probability distribution function is made, concluding that the machine learning algorithms present broader distributions when compared to the BPZ code.