Unsupervised self-organised mapping: a versatile empirical tool for object selection, classification and redshift estimation in large surveys

TL;DR

This paper demonstrates the use of unsupervised self-organised maps for visualising, classifying, and estimating redshifts in large astronomical survey data, showing competitive accuracy with existing methods.

Contribution

It introduces the application of SOMs as an empirical tool for object classification and photometric redshift estimation in large surveys, highlighting its potential for big data astronomy.

Findings

SOMs achieve photometric redshift accuracy of sigma(Dz)=0.03.

SOMs provide competitive results compared to established photo-z methods.

The technique is scalable for petabyte-scale astronomical datasets.

Abstract

We present an application of unsupervised machine learning - the self-organised map (SOM) - as a tool for visualising, exploring and mining the catalogues of large astronomical surveys. Self-organisation culminates in a low-resolution representation of the 'topology' of a parameter volume, and this can be exploited in various ways pertinent to astronomy. Using data from the Cosmological Evolution Survey (COSMOS), we demonstrate two key astronomical applications of the SOM: (i) object classification and selection, using the example of galaxies with active galactic nuclei as a demonstration, and (ii) photometric redshift estimation, illustrating how SOMs can be used as totally empirical predictive tools. With a training set of ~3800 galaxies with z_spec<1, we achieve photometric redshift accuracies competitive with other (mainly template fitting) techniques that use a similar number of photometric bands (sigma(Dz)=0.03 with a ~2% outlier rate when using u*-band to 8um photometry). We also test the SOM as a photo-z tool using the PHoto-z Accuracy Testing (PHAT) synthetic catalogue of Hildebrandt et al. (2010), which compares several different photo-z codes using a common input/training set. We find that the SOM can deliver accuracies that are competitive with many of the established template-fitting and empirical methods. This technique is not without clear limitations, which are discussed, but we suggest it could be a powerful tool in the era of extremely large - 'petabyte' - databases where efficient data-mining is a paramount concern.