master-thesis

Master's research project by A.R. Offringa

Groningen, 2008

Project outputs:

Abstract

The work covers two connected topics: classifying astronomical data sets with several methods, and a new way to visualise a data set.

The thesis explains how this new visualisation technique is built — it displays information generated by the Learning Vector Quantisation method — and gives several examples. Using the idea of prototypes, the technique offers a fresh way to present the information inside a data set and is effective at spotting outliers and relationships. Possible improvements are suggested, and further research is needed before firm conclusions about its use can be drawn.

C4.5 decision trees, Learning Vector Quantisation (LVQ) and Self-Organising Maps (SOM) serve as the classification methods, and their performance and properties are compared. These classification and visualisation methods are applied to gain better insight into the astronomical ESO-LV data set — more than 15,000 galaxies labelled by morphological type — and into an unpublished set of 4,500 galaxies from 2dF-WFI observations, for which morphological types were determined a priori for 100 galaxies that were easy to classify by hand. Applying simulated annealing to the LVQ parameters markedly improves LVQ performance, and changes to the LVQ method resolve problems with unbalanced priors in the data set. Once optimised, LVQ and C4.5 perform almost equally well. The relevance of the features in the data set is examined and compared across the methods, which show some striking differences.

New features were proposed for detecting calibration errors in radio-frequency interferometry, with the goal of finding the calibration errors that create artifacts in images from the Low Frequency Array (LOFAR) telescope. Features based on shapelet decomposition and on variants of the images in Fourier space were tested for usefulness, but proved not discriminating enough.