Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Optimization of Deep Learning Models for Radio Galaxy Classification

Philipp Denzel, Manuel Weiss, Elena Gavagnin, Frank-Peter Schilling

TL;DR

The paper evaluates whether widely used pretrained vision architectures can be effectively adapted to radio astronomy, specifically for classifying and detecting radio galaxies in the RGZ OD dataset. Using both CNNs (e.g., ResNet-50, EfficientNetV2-S) and transformers (ViT, EffFormer; DINO Swin for detection; YOLOv8 for detection), they show that data preprocessing—especially z-scale channel stacking—can yield performance close to or exceeding domain-specific approaches, without altering core architectures. Ensembling (e.g., 30 models) substantially reduces epistemic uncertainty and pushes top-1 accuracy beyond 90%, while parameter tuning and targeted augmentations further improve results (top-1 ≈ 89.7%–89.7% and top-2 ≈ 97–99% in some setups). The findings demonstrate the practicality of reusing standard CV models for radio astronomy with thoughtful data handling, with implications for upcoming SKAO, MWAA, and other surveys, and highlight remaining challenges in disentangling FR I/II classifications and potential label imperfections.

Abstract

Modern radio telescope surveys, capable of detecting billions of galaxies in wide-field surveys, have made manual morphological classification impracticable. This applies in particular when the Square Kilometre Array Observatory (SKAO) becomes operable in 2027, which is expected to close an important gap in our understanding of the Epoch of Reionization (EoR) and other areas of astrophysics. To this end, foreground objects, contaminants of the 21-cm signal, need to be identified and subtracted. Source finding and identification is thus an important albeit challenging task. We investigate the ability of AI and deep learning (DL) methods that have been previously trained on other data domains to localize and classify radio galaxies with minimal changes to their architectures. Various well-known pretrained neural network architectures for image classification and object detection are trained and fine-tuned and their performance is evaluated on a public radio galaxy dataset derived from the Radio Galaxy Zoo. A comparison between convolutional neural network (CNN)- and transformer-based algorithms is performed. The best performing architecture is systematically optimized and an uncertainty estimation is performed by means of an ensemble analysis. Radio source classification performance nearly comparable to the current leading customized models can be obtained using existing standard pretrained DL architectures, without modification and increase in complexity of the model architectures but rather adaptation of the data, by combining various transformations on replicated image channels. Using an ensemble of models can also further improve performance to over 90% accuracy, on par with top-performing models in the literature. The results can be transferred to other survey data, e.g. from the Murchison Wide-field Array (MWA), and in the future be used to study the EoR with the SKAO.

Optimization of Deep Learning Models for Radio Galaxy Classification

TL;DR

The paper evaluates whether widely used pretrained vision architectures can be effectively adapted to radio astronomy, specifically for classifying and detecting radio galaxies in the RGZ OD dataset. Using both CNNs (e.g., ResNet-50, EfficientNetV2-S) and transformers (ViT, EffFormer; DINO Swin for detection; YOLOv8 for detection), they show that data preprocessing—especially z-scale channel stacking—can yield performance close to or exceeding domain-specific approaches, without altering core architectures. Ensembling (e.g., 30 models) substantially reduces epistemic uncertainty and pushes top-1 accuracy beyond 90%, while parameter tuning and targeted augmentations further improve results (top-1 ≈ 89.7%–89.7% and top-2 ≈ 97–99% in some setups). The findings demonstrate the practicality of reusing standard CV models for radio astronomy with thoughtful data handling, with implications for upcoming SKAO, MWAA, and other surveys, and highlight remaining challenges in disentangling FR I/II classifications and potential label imperfections.

Abstract

Modern radio telescope surveys, capable of detecting billions of galaxies in wide-field surveys, have made manual morphological classification impracticable. This applies in particular when the Square Kilometre Array Observatory (SKAO) becomes operable in 2027, which is expected to close an important gap in our understanding of the Epoch of Reionization (EoR) and other areas of astrophysics. To this end, foreground objects, contaminants of the 21-cm signal, need to be identified and subtracted. Source finding and identification is thus an important albeit challenging task. We investigate the ability of AI and deep learning (DL) methods that have been previously trained on other data domains to localize and classify radio galaxies with minimal changes to their architectures. Various well-known pretrained neural network architectures for image classification and object detection are trained and fine-tuned and their performance is evaluated on a public radio galaxy dataset derived from the Radio Galaxy Zoo. A comparison between convolutional neural network (CNN)- and transformer-based algorithms is performed. The best performing architecture is systematically optimized and an uncertainty estimation is performed by means of an ensemble analysis. Radio source classification performance nearly comparable to the current leading customized models can be obtained using existing standard pretrained DL architectures, without modification and increase in complexity of the model architectures but rather adaptation of the data, by combining various transformations on replicated image channels. Using an ensemble of models can also further improve performance to over 90% accuracy, on par with top-performing models in the literature. The results can be transferred to other survey data, e.g. from the Murchison Wide-field Array (MWA), and in the future be used to study the EoR with the SKAO.
Paper Structure (13 sections, 3 figures, 3 tables)

This paper contains 13 sections, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Dataset
  5. Deep Learning Architectures
  6. Preprocessing
  7. Augmentation
  8. Results
  9. Classification Task
  10. Object Detection Task
  11. Model Uncertainty and Ensembling
  12. Parameter Tuning
  13. Summary and Outlook

Figures (3)

  • Figure 1: Example images from the RGZ-OD dataset Wu2019. The dataset is separated into classes in terms of number of peaks and number of components, where the former refers to the count of bright peaks that are detected, while the latter is defined as the number of discrete radio components. The resulting classes are named after the combination of the two as ${[\# \text{components}]\_[\# \text{peaks}]}$.
  • Figure 2: High variance examples (for each panel, left: Z-scaled, right: unscaled).
  • Figure 3: Visual comparison of Min-Max scale (left), Z-scale (middle) and ZMZStack (right) for three example images.