Identification and characterization of emission line objects in J-PAS using artificial neural network

Autor: Martínez-Solaeche, G.
Přispěvatelé: González Delgado, Rosa María, García Benito, Rubén, Universidad de Granada. Programa de Doctorado en Física y Ciencias del Espacio, González Delgado, Rosa M., García-Benito, Rubén
Jazyk: angličtina
Rok vydání: 2022
Předmět:
Popis: Tesis del Departamento de Astronomía Extragaláctica. Instituto de Astrofísica de Andalucía; Universidad de Granada. Programa de Doctorado en Física y Ciencias del Espacio. Leída el 19 de octubre del 2022, a las 12:00 h en el Salón de Actos del IAA.
[EN] In the years to come the Javalambre Physics of the Accelerating Universe Astrophysical Survey (J-PAS) will map ∼ 8000 deg2 of the northern sky in 56 colours (J-spectrum), providing an unprecedented amount of images of astronomical objects. Before arrival of JPCam to the Observatorio Astrof´ısico de Javalambre (OAJ), the J-PAS collaboration observed 1 deg2 of the AEGIS field with the JPAS- Pathfinder camera, using the same photometric system of J-PAS. More than 60 000 objects were detected in what is known as the miniJPAS survey. The main goal of this thesis is to identify and characterize emission line objects with J-PAS. In particular, we study emission line galaxies (ELG) and the properties that can be derived from the analysis of both the emission lines and the stellar populations. Furthermore, we dedicate one chapter to the detection of quasars. Unlike others photometric surveys that use few narrow band filters, the unique characteristics of J-PAS allows us to study these objects in a continuous range of redsfhit. For instance, we will be able to detect the emission lines of Hα or [O ii] in galaxies from 0 to z ∼ 0.35, and z ∼ 1, respectively. Similarly, the Lyα emission line of quasars will be detected from redshift 2.1 up to redshift 4. Traditional methods that measure the equivalent width (EW) of an emission line are generally based on the photometry contrast. Although, this methods gives a very good first approximation, it is limited in many ways. Firstly, there are emission lines such as Hα and [N ii] which are very close to each other in the spectrum. Therefore, they both contribute to the total observed flux in the filter, making difficult to disentangle the individual contribution of each emission line. This is particularly relevant in order to estimate the [N ii]/Hα ratio and determine the main ionization mechanisms of galaxies. What is more, in at least half of the observed galaxies by J-PAS the emission lines will fall in the middle of two adjacent filters. Consequently, measuring the EW is no longer feasible with the photometry contrast approach. In this thesis we developed new techniques based on machine learning (ML) in order to overcome these limitations. Unlike traditional methods, ML algorithms are able to find patterns in the data without making any empirical or theoretical assumptions. Nevertheless, large data sets are needed to train them efficiently. For this purpose, we generated mock J-PAS data, which are based on a collection of spectra from CALIFA, MaNGA, and SDSS. In chapter 3 we trained artificial neural networks (ANN) in order to predict from the generated syntethic J-PAS colors the EW of Hα, Hβ, [O iii], and [N ii] emission lines. Direct measurements of these lines were available in the catalogues for each spectrum. We showed that the minimum S/N that we need in the photometry to measure a line with an EW of 10 Å in Hα, Hβ, [N ii], and [O iii] is 5, 1.5, 3.5, and, 10 respectively. Instead, methods based on the photometry contrast need for the same EW a S/N in the photometry of at least 15.5. With a training set composed of CALIFA and MaNGA galaxies, we reached a precision of 0.092 and 0.078 dex in the [N ii]/Hα and [O iii]/Hβ ratios. Nevertheless, we found that there ratios are more difficult to constrain in galaxies hosting an active galactic nuclei (AGN). We also trained an ANN to distinguish between strong and week emission line galaxies (ELG). We proved that the regime of low emission (∼ 3 Å) can be explored in J-PAS. This is because ML algorithms are able to find much more complex relations between features, so even though we do not have enough sensitivity in the J-spectrum to distinguish galaxies with very low emission lines, the algorithms are able to find other patterns in the data to make this possible. As a proof of concept we applied our techniques to a sample of galaxies observed by miniJPAS in the redshift range 0 < z < 0.35. This is done in chapter 4. We showed that we are able to make a selection of emission line galaxies (ELG), distinguish AGNs from star forming galaxies based on the [N ii]/Hα and [O iii]/Hβ ratios, estimate the star formation rate (SFR) in galaxies throughout the flux of Hα, recover the star formation main sequence of galaxies or constrain the evolution of the cosmic star formation density up to redshift 0.35. Furthermore, our results derived from the properties of the emission lines are in agreement with the products obtained through the analysis of the stellar populations. For instance, we showed that blue (red) galaxies in miniJPAS are composed of a younger (older) stellar population and present stronger (weaker) emission lines. Finally, in chapter 5 we addressed the problem of source classification in order to distinguish between low redshift quasars, high redshift quasars, galaxies, and stars. We found that the main source of confusion appears between low redshift quasars and galaxies. This is because the host galaxy of low redshift quasars is sometimes bright enough to contribute to the observed spectrum. Thus, these objects present mixing features and consequently they are more difficult to classify. We paid special attention to the reliability of the ‘probabilities’ yield by the algorithms, something that is very often neglected in the community. In particular we investigated the effect of data augmentation via hybridisation. This technique consists in mixing the spectra from galaxies, quasars, and stars so as to generate hybrid objects with mixing probabilities. Unfortunately, we do not observe a global improvement in the performance of the algorithms. As a matter of fact, we observed that the ANN becomes under-confidence in their prediction. We believe this is likely due to the intrinsic nature of astronomical observations where errors are attached to observations, thus ‘hybridisation’ turns out to be a natural outcome as the S/N of the sources decreases. Although, the methods and techniques developed in this thesis are limited in some aspects, this work lays the foundations on which to study better the properties of emission line objects in J-PAS. As as soon as J-PAS begins to observe the sky, our methods will be tested in large sample of galaxies, thus it will be possible to improve them even further.
Databáze: OpenAIRE
Externí odkaz: