Prediction and Inference Methods for Modern Astronomical Surveys

Abstract

Modern astronomical surveys study the sky at an unprecedented scale, often aiming to observe, catalog, and analyze millions to hundreds of billions of objects. Such large scale surveys require robust statistical tools to aid survey design, planning, and data analysis. This thesis introduces several novel prediction and inference methods applied to astronomical surveys. First, a target selection method based on the modeling of the number density of desired and undesired classes of objects is discussed in the context of emission line galaxy target selection for Dark Energy Spectroscopic Instrument survey. Second, an observational study to test the selection functions developed based on the aforementioned method and a convolutional neural network analysis of the two-dimensional spectra of the observed objects are presented. Lastly, a novel Bayesian point source inference method applied to noisy astronomical images is introduced. The method is based on trans-dimensional Riemannian Hamiltonian Monte Carlo algorithm, a type of Markov Chain Monte Carlo algorithm that allows a robust inference over convoluted posterior geometry that is difficult to explore with other inferential methods.