CARAD: Computer-aided Analysis of Radio Astronomy Data

Radio astronomy studies celestial objects by detecting their radio frequency emissions, enabling observation of phenomena invisible at other wavelengths. Next-generation telescopes generate unprecedented data volumes that challenge traditional analytical approaches. Manual inspection and classical statistical methods are increasingly inadequate for addressing the scale and complexity of these datasets, necessitating machine intelligence for automating critical tasks including morphological classification, image retrieval, and anomaly detection of radio sources. While deep learning techniques show considerable potential, their computational intensity, high data requirements, and lack of interpretability render them less sustainable for radio astronomical research. To address these limitations, this thesis proposes adapting trainable COSFIRE filters, which offer a lightweight, rotation-invariant, and interpretable framework for feature extraction that achieves performance comparable or superior to deep learning models while requiring significantly fewer computational resources. This work presents a comprehensive literature review identifying key gaps and opportunities in machine learning applications to radio astronomy, establishing the foundation for subsequent investigations. The thesis then demonstrates that the proposed COSFIRE-based morphological classification achieves superior performance compared to state-of-the-art deep learning approaches, attaining 93% accuracy while requiring approximately 20 times fewer computational operations than competing DenseNet-based methods. In addition, a content-based image retrieval system is developed using supervised hashing techniques combined with COSFIRE descriptors, enabling efficient similarity searches in large radio galaxy databases with 91% mean average precision. Finally, an innovative semi-supervised anomaly detection framework integrating COSFIRE descriptors with the Local Outlier Factor algorithm successfully identifies unusual radio galaxy morphologies, achieving a geometric mean of 79%, surpassing the performance of computationally intensive deep learning autoencoders. The COSFIRE methodology’s intrinsic rotation invariance eliminates the need for data augmentation, while its interpretable feature extraction process provides scientific transparency often lacking in deep learning approaches. Evaluated on benchmark datasets containing approximately 2,000 samples across multiple morphological classes (Compact, FRI, FRII, and Bent), the proposed frameworks consistently demonstrate both superior performance and remarkable computational efficiency.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Machetho, Stephen Ndung’u

Advisors dc:contributor.advisor

• De Villiers, Dirk
• Grobler, Trienko

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/136111

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/136111