Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Applied machine learning for radio galaxy classification and anomalous source detection
Thesis (MSc)--Stellenbosch University, 2024.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
Stellenbosch : Stellenbosch University
2025
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613970444058624 |
|---|---|
| access_status_str | Open Access |
| author | Brand, Kevin |
| author2 | Grobler, T. L. |
| author_browse | Brand, Kevin Grobler, T. L. |
| author_facet | Grobler, T. L. Brand, Kevin |
| author_sort | Brand, Kevin |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (MSc)--Stellenbosch University, 2024. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/131591 |
| institution | Stellenbosch University (South Africa) |
| language | English |
| last_indexed | 2026-06-10T12:44:36.436Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2025 |
| publishDateRange | 2025 |
| publishDateSort | 2025 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/131591 Applied machine learning for radio galaxy classification and anomalous source detection Brand, Kevin Grobler, T. L. Kleynhans, W. Stellenbosch University. Faculty of Science. Dept. of Computer Science. Radio galaxies Neural networks (Computer science) Pattern recognition systems Machine learning UCTD Thesis (MSc)--Stellenbosch University, 2024. ENGLISH ABSTRACT: The classification of radio sources and the identification of anomalous sources play a vital role in the development of the understanding regarding various cosmological processes, such as the formation and evolution of galaxies and how they interact with their local environments. As the new generation of radio telescopes — such as the square kilometre array (SKA) — come online, a massive influx is expected with respect to the number of observations of radio sources that will be generated. This increase makes the manual evaluation and classification of radio sources by experts infeasible. Approaches have been considered that enable the general public to assist with these classifications. However, it is not clear whether these approaches will be able to keep up with the growing rates at which radio telescopes produce observations. Thus, a growing body of literature is investigating whether these tasks can be automated by applying machine learning models instead. In this thesis we extended the work conducted in the literature by further investigating the automation of morphological classification and anomalous source detection. We investigated two adaptations when applying convolutional neural networks (CNNs) to morphological classification, as these models have been shown to be particularly useful in this regard. We investigated the impact of standardising source orientation prior to CNN training and found that it leads to improvements in classification performance. However — apart from faster training times — it provided no benefits when compared to rotational augmentation, with rotational augmentation leading to better classification results. We also investigated what the impact on morphological classification would be when introducing an additional objective to guide CNNs towards extracting source features that are known to be informative. Our results showed that feature guidance led to improvements in performance with respect to all four of the morphological classes in our dataset. Furthermore, we found that explicitly guiding CNNs to use these extracted features for classification led to better results than implicit guidance did. With respect to anomalous source detection, we extended the literature by further investigating the use of convolutional autoencoders (CAEs), which has previously been used for this purpose. We compared CAEs to conventional machine learning models and found that they consistently outperformed these models. In terms of architecture refinement, we also investigated how these CAEs could be adapted to improve their ability to find anomalous sources and found that increases in model depth, as well as the incorporation of a memory unit were beneficial. Furthermore, we found that certain reconstruction metrics were more performant when used as anomaly scores. Our final experiment consisted of applying conventional machine learning models to multiple anomaly scores and found that doing so could lead to further improvements in the anomaly detection performance of CAEs. AFRIKAANSE OPSOMMING: Die klassifikasie van radiobronne en die ontdekking van abnormale bronne speel ’n belangrike rol in die ontwikkeling van ons verstaan van verskeie kosmologiese prosesse, soos die vorming en evolusie van sterrestelsels, asook hoe hulle beïnvloed word deur hulle plaaslike omgewing. Soos die nuwe generasie van radioteleskope aanlyn kom, word ’n groot toename in die hoeveelheid radiobronwaarnemings verwag. Hierdie toename maak dit onprakties vir kenners om die radiobronne self te evalueer en te klassifiseer. Een moontlike oplossing behels die opleiding van die brëe publiek sodat hulle ook kan help met die klassifikasie van hierdie bronne. Dit is wel nie duidelik of so ’n oplossing sal kan byhou met die spoed waarteen die nuwe teleskope waarnemings gaan genereer nie. Dus is daar al hoe meer navorsing wat ondersoek of masjienleer gebruik kan word vir die outomatiese klassifikasie van radiobronne. In hierdie tesis brei ons die bestaande literatuur uit deur die outomatisering van morfologiese klassifikasie en abnormale bron ontdekking verder te ondersoek. Ons fokus spesifiek op twee aanpassings tot die gebruik van konvolusionele neurale netwerke (KNNe) vir morfologiese klassifikasie, aangesien die literatuur aandui dat hierdie modelle besonders nuttig is in hierdie hoedanigheid. Ons het ondersoek wat die impak is van die standaardisering van bronoriëntasie voordat KNNe afgerig word en het gevind dat dit ’n positiewe impak op hulle klassifikasievermoë gehad het. Ons het standaardisering ook met rotasionele aanvulling vergelyk en vasgestel dat aanvulling gelei het tot meer akkurate resultate as standaardisering. Standaardisering het wel minder afrigtyd benodig. Ons het ook ondersoek ingestel om uit te vind of daar enige voordele daarin lê om modelle te lei om sekere insiggewende eienskappe van radiobronne uit te soek tydens die klassifikasieproses. Hierdie ondersoek het aangedui dat sulke leiding wel ’n voordelige impak het op die akkuraatheid van die KNNe op al vier morfologiese klasse. Ons eksperiment het ook vasgestel dat die eksplisiete leiding van KNNe beter resultate getoon het as wat implisiete leiding het. Wat abnormale bron ontdekking betref, was konvolusionele outo-enkodeerders (KOEs) al vantevore in die literatuur gebruik. Ons het hierdie literatuur uitgebrei deur KOEs te vergelyk met konvensionele masjienleer modelle, en het gevind dat die KOEs deurlopend meer akkuraat was. Ons het ook ondersoek ingestel om vas te stel of die KOEs se argitektuur verbeter kan word vir abnormale bron ontdekking. Hier het ons gevind dat dieper modelle, asook die byvoeging van ’n geheue-eenheid voordelig was. Ons het ook vasgestel dat daar sekere abnormaliteitsmaatstawwe was wat beter kon evalueer of KOEs se afvoer aanduidend was van abnormale bronne. Laastens het ons konvensionele masjienleer modelle gebruik om verskeie abnormaliteitsmaatstawe te kombineer. Hierdie benadering het gelei tot ’n verdere verbetering in die KOEs se vermoë om abnormale bronne te identifiseer. Masters 2025-01-28T10:38:25Z 2025-01-28T10:38:25Z 2024-12 Thesis https://scholar.sun.ac.za/handle/10019.1/131591 en Stellenbosch University xxiii, 223 pages : illustrations application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Radio galaxies Neural networks (Computer science) Pattern recognition systems Machine learning UCTD Brand, Kevin Applied machine learning for radio galaxy classification and anomalous source detection |
| title | Applied machine learning for radio galaxy classification and anomalous source detection |
| title_full | Applied machine learning for radio galaxy classification and anomalous source detection |
| title_fullStr | Applied machine learning for radio galaxy classification and anomalous source detection |
| title_full_unstemmed | Applied machine learning for radio galaxy classification and anomalous source detection |
| title_short | Applied machine learning for radio galaxy classification and anomalous source detection |
| title_sort | applied machine learning for radio galaxy classification and anomalous source detection |
| topic | Radio galaxies Neural networks (Computer science) Pattern recognition systems Machine learning UCTD |
| url | https://scholar.sun.ac.za/handle/10019.1/131591 |
| work_keys_str_mv | AT brandkevin appliedmachinelearningforradiogalaxyclassificationandanomaloussourcedetection |