The Good, the Bad and the Ugly: Statistical quality assessment of SZ detections

TL;DR

This paper compares three statistical methods—likelihood analysis, clustering, and neural networks—for classifying sources in astronomical catalogues based on their spectral energy distributions, demonstrating the neural network's superior performance.

Contribution

It introduces a comprehensive approach to classify astronomical sources using multiple statistical techniques, highlighting the effectiveness of neural networks in reliability assessment.

Findings

All three methods agree well in classification results.

Neural networks outperform other methods in separating reliable and unreliable sources.

The approach is applicable to future large-scale astronomical surveys.

Abstract

We examine three approaches to the problem of source classification in catalogues. Our goal is to determine the confidence with which the elements in these catalogues can be distinguished in populations on the basis of their spectral energy distribution (SED). Our analysis is based on the projection of the measurements onto a comprehensive SED model of the main signals in the considered range of frequencies. We first first consider likelihood analysis, which half way between supervised and unsupervised methods. Next, we investigate an unsupervised clustering technique. Finally, we consider a supervised classifier based on Artificial Neural Networks. We illustrate the approach and results using catalogues from various surveys. i.e., X-Rays (MCXC), optical (SDSS) and millimetric (Planck Sunyaev-Zeldovich (SZ)). We show that the results from the statistical classifications of the three methods are in very good agreement with each others, although the supervised neural network-based classification shows better performances allowing the best separation into populations of reliable and unreliable sources in catalogues. The latest method was applied to the SZ sources detected by the Planck satellite. It led to a classification assessing and thereby agreeing with the reliability assessment published in the Planck SZ catalogue. Our method could easily be applied to catalogues from future large survey such as SRG/eROSITA and Euclid.