Automated Source Extraction for the Next Generation of Neutral Hydrogen Surveys

Abstract

This thesis is a first step to develop the necessary tools to automatically extract and parameterize sources from future HI surveys with ASKAP, WSRT/Apertif, and SKA.<br /> The current approach to large-scale HI surveys, that is, automated source finding followed by manual classification and parametrization, is no longer feasible in light of the data volumes expected for future surveys. We use data from EBHIS to develop and test a completely automated source extraction pipeline for extragalactic HI surveys.<br /> We apply a 2D-1D wavelet de-noising technique to HI data and show that it is well adapted to the typical shapes of sources encountered in HI surveys. This technique allows to reliably extract sources even from data containing defects commonly encountered in single-dish HI surveys.<br /> Automating the task of false-positive rejection requires reliable parameters for all source candidates generated by the source-finding step. For this purpose, we develop a reliable, automated parametrization pipeline that combines time-tested algorithms with new approaches to baseline estimation, spectral filtering, and mask optimization. The accuracy of the algorithms is tested by performing extensive simulations. By comparison with the uncertainty estimates from HIPASS we show that our automated pipeline gives equal or better accuracy than manual parametrization.<br /> We implement the task of source classification using artificial neural networks using the automatically determined parameters of the source candidates as inputs. The viability of this approach is verified on a training data set comprised of parameters measured from simulated sources and false positives extracted from real EBHIS data. Since the number of true positives from real data is small compared to the number of false positives, we explore various methods of training artificial neural networks from imbalanced data sets. We show that the artificial neural networks trained in this way do not achieve sufficient completeness and reliability when applied to the source candidates detected from the extragalactic EBHIS survey.<br /> We use the trained artificial neural networks in a semi-supervised manner to compile the first extragalactic EBHIS source catalog. The use of artificial neural networks reduces the number of source candidates that require manual inspection by more than an order of magnitude. We compare the results from EBHIS to HIPASS and show that the number of sources in the compiled catalog is approximately half of the sources expected. The main reason for this detection inefficiency is identified to be mis-classification by the artificial neural networks. This is traced back to the limited training data set, which does not cover the parameter space of real detections sufficiently, and the similarity of true and false positives in the parameter space spanned by the measured parameters.<br /> We conclude that, while our automated source finding and parametrization algorithms perform satisfactorily, the classification of sources is the most challenging task for future HI surveys. Classification based on the measured source parameters does not provide sufficient discriminatory power and we propose to explore methods based on machine vision which learns features of real sources from the data directly.