Star-Galaxy classification with a novel neural network architecture

  • Myank Singhal

Student thesis: Master's ThesisMaster of Science by Research (MScR)


Euclid is an upcoming wide-area survey that aims to study the dark universe using baryon
acoustic oscillations (BAO) and weak gravitational lensing. To measure the effects of weak
gravitational lensing, we need to know the sensor’s response to high accuracy. A pure star
sample is required to create a point spread function model, which depends on the star’s stellar
class. We present statistical and machine learning methods to classify stars using photometric
measurements. In the first part of this work, we explore the use of SED fitting with stellar
models to classify stars into stellar classes. In the second part, we explore the use of machine
learning methods to identify stars. Finally, we introduce a novel classification framework that
uses an ensemble of supervised machine learning models capable of handling missing data,
uncertainty in photometric measurements and output a probability distribution function for its
classification. This work uses a dataset of 49,000 spectroscopically labelled objects. These objects
include photometry from SDSS, CFHTLS, KiDS, VISTA (VIKING and VIDEO) and WISE surveys
and cover the entire optical to mid-infrared wavelength space. We show that our classifier can
correctly classify objects up to an overall accuracy of 99.72% and F1 score of 99.0% on data with
uncertainty. We also test this classifier on a blind photometric sample of 30,000 objects with
photometry from KiDS, VIKING and WISE achieving an accuracy of 99.25% and F1 score of
97.9%. This novel classifier provides a framework comparable to other machine learning methods
in performance while also handling various constraints in observations like missing data and
uncertainty. This classifier framework can be used for the upcoming Euclid mission to classify
stars with high accuracy.
Date of Award2 Dec 2021
Original languageEnglish
Awarding Institution
  • University of Bristol
SupervisorMalcolm N Bremer (Supervisor) & Sotiria Fotopoulou (Supervisor)

Cite this