Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing Galaxy Classification with U-Net Variational Autoencoders. II. JWST High Redshift Galaxy Sample

Sergey Mirzoyan

TL;DR

The paper addresses morphological classification of high-redshift galaxies in JWST data, which are noise-limited up to $z\sim 8$. It combines a U-Net Variational Autoencoder denoising stage with a symmetry-aware Group Convolutional Neural Network to perform binary disk vs non-disk classification on JWST data from the UNCOVER survey. It identifies 83 disk-like galaxies out of 292, with 70–80% located at $z>3$, suggesting disk structures may be more common in the early universe than previously thought. This work demonstrates a practical, generative-model–based preprocessing step that enhances morphological inference for upcoming deep-field surveys.

Abstract

Building on our previous work, we apply a U-Net Variational Autoencoder (VAE) framework to denoise galaxy images from the James Webb Space Telescope (JWST) and enhance morphological classification. This study focuses on galaxies observed up to redshift approximately at 8, capturing them at early evolutionary stages where their faintness and structural complexity pose challenges for the traditional classification methods. By mitigating observational noise, our approach enables the identification of morphological features, particularly in distinguishing between disk and non-disk galaxy types. We evaluate the denoising performance using standard image quality metrics and demonstrate that the enhanced images lead to improved classification accuracy across multiple deep learning models. Our analysis of a sample of 292 galaxies up to z=7.69 shows 83 galaxies classified as disk-like by the GCNN model with high confidence, of those approximately 70-80 % are of redshifts greater than 3. These findings suggest that disk-like structures can be prevalent in the early universe. The results highlight the potential of VAE-based denoising as a robust pre-processing step for analyzing high-redshift galaxy populations in ongoing astronomical surveys.

Enhancing Galaxy Classification with U-Net Variational Autoencoders. II. JWST High Redshift Galaxy Sample

TL;DR

The paper addresses morphological classification of high-redshift galaxies in JWST data, which are noise-limited up to . It combines a U-Net Variational Autoencoder denoising stage with a symmetry-aware Group Convolutional Neural Network to perform binary disk vs non-disk classification on JWST data from the UNCOVER survey. It identifies 83 disk-like galaxies out of 292, with 70–80% located at , suggesting disk structures may be more common in the early universe than previously thought. This work demonstrates a practical, generative-model–based preprocessing step that enhances morphological inference for upcoming deep-field surveys.

Abstract

Building on our previous work, we apply a U-Net Variational Autoencoder (VAE) framework to denoise galaxy images from the James Webb Space Telescope (JWST) and enhance morphological classification. This study focuses on galaxies observed up to redshift approximately at 8, capturing them at early evolutionary stages where their faintness and structural complexity pose challenges for the traditional classification methods. By mitigating observational noise, our approach enables the identification of morphological features, particularly in distinguishing between disk and non-disk galaxy types. We evaluate the denoising performance using standard image quality metrics and demonstrate that the enhanced images lead to improved classification accuracy across multiple deep learning models. Our analysis of a sample of 292 galaxies up to z=7.69 shows 83 galaxies classified as disk-like by the GCNN model with high confidence, of those approximately 70-80 % are of redshifts greater than 3. These findings suggest that disk-like structures can be prevalent in the early universe. The results highlight the potential of VAE-based denoising as a robust pre-processing step for analyzing high-redshift galaxy populations in ongoing astronomical surveys.

Paper Structure

This paper contains 5 sections, 3 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Dataset
  3. Method
  4. Analysis
  5. Conclusions

Figures (2)

  • Figure 1: Histogram showing the redshift distribution of 304 galaxies. The horizontal axis represents the redshift values, while the vertical one indicates the number of galaxies in each bin.
  • Figure 2: (a) Scatter plot of spectroscopic redshift versus predicted probability of disk-like morphology for 83 galaxies classified by the GCNN model. Each point represents a galaxy identified as disk-like, with predicted probabilities exceeding 0.5. The red ellipse marks a region of interest, extending from redshifts greater than 3 and encompassing relatively higher predicted probabilities, indicating a concentration of high-confidence disk-like galaxies at intermediate to high redshifts. (b) Mean predicted probability of disk-like morphology as a function of redshift, binned into 11 intervals. Error bars represent the 95% confidence interval of the mean in each bin, illustrating the variation and reliability of model predictions across redshift.